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 "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callGenerator.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/countbitsnode.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/mathexactnode.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/narrowptrnode.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/parse.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "prims/nativeLookup.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "trace/traceMacros.hpp"
53
54 class LibraryIntrinsic : public InlineCallGenerator {
55 // Extend the set of intrinsics known to the runtime:
56 public:
57 private:
58 bool _is_virtual;
59 bool _does_virtual_dispatch;
60 int8_t _predicates_count; // Intrinsic is predicated by several conditions
61 int8_t _last_predicate; // Last generated predicate
62 vmIntrinsics::ID _intrinsic_id;
63
64 public:
65 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
66 : InlineCallGenerator(m),
67 _is_virtual(is_virtual),
68 _does_virtual_dispatch(does_virtual_dispatch),
69 _predicates_count((int8_t)predicates_count),
70 _last_predicate((int8_t)-1),
71 _intrinsic_id(id)
72 {
73 }
74 virtual bool is_intrinsic() const { return true; }
75 virtual bool is_virtual() const { return _is_virtual; }
76 virtual bool is_predicated() const { return _predicates_count > 0; }
77 virtual int predicates_count() const { return _predicates_count; }
78 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
79 virtual JVMState* generate(JVMState* jvms);
80 virtual Node* generate_predicate(JVMState* jvms, int predicate);
81 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
82 };
83
84
85 // Local helper class for LibraryIntrinsic:
86 class LibraryCallKit : public GraphKit {
87 private:
88 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
89 Node* _result; // the result node, if any
90 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
91
92 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
93
94 public:
95 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
96 : GraphKit(jvms),
97 _intrinsic(intrinsic),
98 _result(NULL)
99 {
100 // Check if this is a root compile. In that case we don't have a caller.
101 if (!jvms->has_method()) {
102 _reexecute_sp = sp();
103 } else {
104 // Find out how many arguments the interpreter needs when deoptimizing
105 // and save the stack pointer value so it can used by uncommon_trap.
106 // We find the argument count by looking at the declared signature.
107 bool ignored_will_link;
108 ciSignature* declared_signature = NULL;
109 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
110 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
111 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
112 }
113 }
114
115 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
116
117 ciMethod* caller() const { return jvms()->method(); }
118 int bci() const { return jvms()->bci(); }
119 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
120 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
121 ciMethod* callee() const { return _intrinsic->method(); }
122
123 bool try_to_inline(int predicate);
124 Node* try_to_predicate(int predicate);
125
126 void push_result() {
127 // Push the result onto the stack.
128 if (!stopped() && result() != NULL) {
129 BasicType bt = result()->bottom_type()->basic_type();
130 push_node(bt, result());
131 }
132 }
133
134 private:
135 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
136 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
137 }
138
139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
140 void set_result(RegionNode* region, PhiNode* value);
141 Node* result() { return _result; }
142
143 virtual int reexecute_sp() { return _reexecute_sp; }
144
145 // Helper functions to inline natives
146 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
147 Node* generate_slow_guard(Node* test, RegionNode* region);
148 Node* generate_fair_guard(Node* test, RegionNode* region);
149 Node* generate_negative_guard(Node* index, RegionNode* region,
150 // resulting CastII of index:
151 Node* *pos_index = NULL);
152 Node* generate_limit_guard(Node* offset, Node* subseq_length,
153 Node* array_length,
154 RegionNode* region);
155 void generate_string_range_check(Node* array, Node* offset,
156 Node* length, bool char_count);
157 Node* generate_current_thread(Node* &tls_output);
158 Node* load_mirror_from_klass(Node* klass);
159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path,
161 int offset);
162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
163 RegionNode* region, int null_path) {
164 int offset = java_lang_Class::klass_offset_in_bytes();
165 return load_klass_from_mirror_common(mirror, never_see_null,
166 region, null_path,
167 offset);
168 }
169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
170 RegionNode* region, int null_path) {
171 int offset = java_lang_Class::array_klass_offset_in_bytes();
172 return load_klass_from_mirror_common(mirror, never_see_null,
173 region, null_path,
174 offset);
175 }
176 Node* generate_access_flags_guard(Node* kls,
177 int modifier_mask, int modifier_bits,
178 RegionNode* region);
179 Node* generate_interface_guard(Node* kls, RegionNode* region);
180 Node* generate_array_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, false, false);
182 }
183 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, false, true);
185 }
186 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, true, false);
188 }
189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, true, true);
191 }
192 Node* generate_array_guard_common(Node* kls, RegionNode* region,
193 bool obj_array, bool not_array);
194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
196 bool is_virtual = false, bool is_static = false);
197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
198 return generate_method_call(method_id, false, true);
199 }
200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
201 return generate_method_call(method_id, true, false);
202 }
203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
204 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
205
206 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
207 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
208 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
209 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
210 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
211 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_indexOfChar();
213 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_toBytesU();
215 bool inline_string_getCharsU();
216 bool inline_string_copy(bool compress);
217 bool inline_string_char_access(bool is_store);
218 Node* round_double_node(Node* n);
219 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
220 bool inline_math_native(vmIntrinsics::ID id);
221 bool inline_trig(vmIntrinsics::ID id);
222 bool inline_math(vmIntrinsics::ID id);
223 template <typename OverflowOp>
224 bool inline_math_overflow(Node* arg1, Node* arg2);
225 void inline_math_mathExact(Node* math, Node* test);
226 bool inline_math_addExactI(bool is_increment);
227 bool inline_math_addExactL(bool is_increment);
228 bool inline_math_multiplyExactI();
229 bool inline_math_multiplyExactL();
230 bool inline_math_negateExactI();
231 bool inline_math_negateExactL();
232 bool inline_math_subtractExactI(bool is_decrement);
233 bool inline_math_subtractExactL(bool is_decrement);
234 bool inline_min_max(vmIntrinsics::ID id);
235 bool inline_notify(vmIntrinsics::ID id);
236 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
237 // This returns Type::AnyPtr, RawPtr, or OopPtr.
238 int classify_unsafe_addr(Node* &base, Node* &offset);
239 Node* make_unsafe_address(Node* base, Node* offset);
240 // Helper for inline_unsafe_access.
241 // Generates the guards that check whether the result of
242 // Unsafe.getObject should be recorded in an SATB log buffer.
243 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
244 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
245 static bool klass_needs_init_guard(Node* kls);
246 bool inline_unsafe_allocate();
247 bool inline_unsafe_copyMemory();
248 bool inline_native_currentThread();
249 #ifdef TRACE_HAVE_INTRINSICS
250 bool inline_native_classID();
251 bool inline_native_threadID();
252 #endif
253 bool inline_native_time_funcs(address method, const char* funcName);
254 bool inline_native_isInterrupted();
255 bool inline_native_Class_query(vmIntrinsics::ID id);
256 bool inline_native_subtype_check();
257
258 bool inline_native_newArray();
259 bool inline_native_getLength();
260 bool inline_array_copyOf(bool is_copyOfRange);
261 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
262 bool inline_objects_checkIndex();
263 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
264 bool inline_native_clone(bool is_virtual);
265 bool inline_native_Reflection_getCallerClass();
266 // Helper function for inlining native object hash method
267 bool inline_native_hashcode(bool is_virtual, bool is_static);
268 bool inline_native_getClass();
269
270 // Helper functions for inlining arraycopy
271 bool inline_arraycopy();
272 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
273 RegionNode* slow_region);
274 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
275 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
276
277 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
278 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
279 bool inline_unsafe_ordered_store(BasicType type);
280 bool inline_unsafe_fence(vmIntrinsics::ID id);
281 bool inline_fp_conversions(vmIntrinsics::ID id);
282 bool inline_number_methods(vmIntrinsics::ID id);
283 bool inline_reference_get();
284 bool inline_Class_cast();
285 bool inline_aescrypt_Block(vmIntrinsics::ID id);
286 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
287 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
288 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
289 Node* inline_counterMode_AESCrypt_predicate();
290 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
291 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
292 bool inline_ghash_processBlocks();
293 bool inline_sha_implCompress(vmIntrinsics::ID id);
294 bool inline_digestBase_implCompressMB(int predicate);
295 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
296 bool long_state, address stubAddr, const char *stubName,
297 Node* src_start, Node* ofs, Node* limit);
298 Node* get_state_from_sha_object(Node *sha_object);
299 Node* get_state_from_sha5_object(Node *sha_object);
300 Node* inline_digestBase_implCompressMB_predicate(int predicate);
301 bool inline_encodeISOArray();
302 bool inline_updateCRC32();
303 bool inline_updateBytesCRC32();
304 bool inline_updateByteBufferCRC32();
305 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
306 bool inline_updateBytesCRC32C();
307 bool inline_updateDirectByteBufferCRC32C();
308 bool inline_updateBytesAdler32();
309 bool inline_updateByteBufferAdler32();
310 bool inline_multiplyToLen();
311 bool inline_hasNegatives();
312 bool inline_squareToLen();
313 bool inline_mulAdd();
314 bool inline_montgomeryMultiply();
315 bool inline_montgomerySquare();
316 bool inline_vectorizedMismatch();
317
318 bool inline_profileBoolean();
319 bool inline_isCompileConstant();
320
321 bool inline_deoptimize();
322 };
323
324 //---------------------------make_vm_intrinsic----------------------------
325 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
326 vmIntrinsics::ID id = m->intrinsic_id();
327 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
328
329 if (!m->is_loaded()) {
330 // Do not attempt to inline unloaded methods.
331 return NULL;
332 }
333
334 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
335 bool is_available = false;
336
337 {
338 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
339 // the compiler must transition to '_thread_in_vm' state because both
340 // methods access VM-internal data.
341 VM_ENTRY_MARK;
342 methodHandle mh(THREAD, m->get_Method());
343 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
344 !C->directive()->is_intrinsic_disabled(mh) &&
345 !vmIntrinsics::is_disabled_by_flags(mh);
346
347 }
348
349 if (is_available) {
350 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
351 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
352 return new LibraryIntrinsic(m, is_virtual,
353 vmIntrinsics::predicates_needed(id),
354 vmIntrinsics::does_virtual_dispatch(id),
355 (vmIntrinsics::ID) id);
356 } else {
357 return NULL;
358 }
359 }
360
361 //----------------------register_library_intrinsics-----------------------
362 // Initialize this file's data structures, for each Compile instance.
363 void Compile::register_library_intrinsics() {
364 // Nothing to do here.
365 }
366
367 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
368 LibraryCallKit kit(jvms, this);
369 Compile* C = kit.C;
370 int nodes = C->unique();
371 #ifndef PRODUCT
372 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
373 char buf[1000];
374 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
375 tty->print_cr("Intrinsic %s", str);
376 }
377 #endif
378 ciMethod* callee = kit.callee();
379 const int bci = kit.bci();
380
381 // Try to inline the intrinsic.
382 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
383 kit.try_to_inline(_last_predicate)) {
384 if (C->print_intrinsics() || C->print_inlining()) {
385 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
386 }
387 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
388 if (C->log()) {
389 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
390 vmIntrinsics::name_at(intrinsic_id()),
391 (is_virtual() ? " virtual='1'" : ""),
392 C->unique() - nodes);
393 }
394 // Push the result from the inlined method onto the stack.
395 kit.push_result();
396 C->print_inlining_update(this);
397 return kit.transfer_exceptions_into_jvms();
398 }
399
400 // The intrinsic bailed out
401 if (C->print_intrinsics() || C->print_inlining()) {
402 if (jvms->has_method()) {
403 // Not a root compile.
404 const char* msg;
405 if (callee->intrinsic_candidate()) {
406 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
407 } else {
408 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
409 : "failed to inline (intrinsic), method not annotated";
410 }
411 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
412 } else {
413 // Root compile
414 tty->print("Did not generate intrinsic %s%s at bci:%d in",
415 vmIntrinsics::name_at(intrinsic_id()),
416 (is_virtual() ? " (virtual)" : ""), bci);
417 }
418 }
419 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
420 C->print_inlining_update(this);
421 return NULL;
422 }
423
424 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
425 LibraryCallKit kit(jvms, this);
426 Compile* C = kit.C;
427 int nodes = C->unique();
428 _last_predicate = predicate;
429 #ifndef PRODUCT
430 assert(is_predicated() && predicate < predicates_count(), "sanity");
431 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
432 char buf[1000];
433 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
434 tty->print_cr("Predicate for intrinsic %s", str);
435 }
436 #endif
437 ciMethod* callee = kit.callee();
438 const int bci = kit.bci();
439
440 Node* slow_ctl = kit.try_to_predicate(predicate);
441 if (!kit.failing()) {
442 if (C->print_intrinsics() || C->print_inlining()) {
443 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
444 }
445 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
446 if (C->log()) {
447 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
448 vmIntrinsics::name_at(intrinsic_id()),
449 (is_virtual() ? " virtual='1'" : ""),
450 C->unique() - nodes);
451 }
452 return slow_ctl; // Could be NULL if the check folds.
453 }
454
455 // The intrinsic bailed out
456 if (C->print_intrinsics() || C->print_inlining()) {
457 if (jvms->has_method()) {
458 // Not a root compile.
459 const char* msg = "failed to generate predicate for intrinsic";
460 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
461 } else {
462 // Root compile
463 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
464 vmIntrinsics::name_at(intrinsic_id()),
465 (is_virtual() ? " (virtual)" : ""), bci);
466 }
467 }
468 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
469 return NULL;
470 }
471
472 bool LibraryCallKit::try_to_inline(int predicate) {
473 // Handle symbolic names for otherwise undistinguished boolean switches:
474 const bool is_store = true;
475 const bool is_compress = true;
476 const bool is_native_ptr = true;
477 const bool is_static = true;
478 const bool is_volatile = true;
479
480 if (!jvms()->has_method()) {
481 // Root JVMState has a null method.
482 assert(map()->memory()->Opcode() == Op_Parm, "");
483 // Insert the memory aliasing node
484 set_all_memory(reset_memory());
485 }
486 assert(merged_memory(), "");
487
488
489 switch (intrinsic_id()) {
490 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
491 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
492 case vmIntrinsics::_getClass: return inline_native_getClass();
493
494 case vmIntrinsics::_dsin:
495 case vmIntrinsics::_dcos:
496 case vmIntrinsics::_dtan:
497 case vmIntrinsics::_dabs:
498 case vmIntrinsics::_datan2:
499 case vmIntrinsics::_dsqrt:
500 case vmIntrinsics::_dexp:
501 case vmIntrinsics::_dlog:
502 case vmIntrinsics::_dlog10:
503 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
504
505 case vmIntrinsics::_min:
506 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
507
508 case vmIntrinsics::_notify:
509 case vmIntrinsics::_notifyAll:
510 if (InlineNotify) {
511 return inline_notify(intrinsic_id());
512 }
513 return false;
514
515 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
516 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
517 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
518 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
519 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
520 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
521 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
522 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
523 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
524 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
525 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
526 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
527
528 case vmIntrinsics::_arraycopy: return inline_arraycopy();
529
530 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
531 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
532 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
533 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
534
535 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
536 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
537 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
538 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
539 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
540 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
541 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
542
543 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
544 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
545
546 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
547 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
548 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
549 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
550
551 case vmIntrinsics::_compressStringC:
552 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
553 case vmIntrinsics::_inflateStringC:
554 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
555
556 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
557 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
558 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
559 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
560 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
561 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
562 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
563 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
564 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
565 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
566 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
567 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
568 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
569 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
570 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
571 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
572 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
573 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
574
575 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
576 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
577 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
578 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
579 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
580 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
581 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
582 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
583
584 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
585 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
586 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
587 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
588 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
589 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
590 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
591 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
592
593 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
594 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
595 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
596 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
597 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
598 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
599 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
600 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
601 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
602
603 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
604 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
605 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
606 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
607 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
608 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
609 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
610 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
611 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
612
613 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true);
614 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true);
615 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true);
616 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true);
617
618 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true);
619 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true);
620 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true);
621 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true);
622
623 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
624 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
625 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
626
627 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
628 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
629 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
630
631 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
632 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
633 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
634 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
635 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
636
637 case vmIntrinsics::_loadFence:
638 case vmIntrinsics::_storeFence:
639 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
640
641 case vmIntrinsics::_currentThread: return inline_native_currentThread();
642 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
643
644 #ifdef TRACE_HAVE_INTRINSICS
645 case vmIntrinsics::_classID: return inline_native_classID();
646 case vmIntrinsics::_threadID: return inline_native_threadID();
647 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
648 #endif
649 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
650 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
651 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
652 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
653 case vmIntrinsics::_newArray: return inline_native_newArray();
654 case vmIntrinsics::_getLength: return inline_native_getLength();
655 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
656 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
657 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
658 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
659 case vmIntrinsics::_Objects_checkIndex: return inline_objects_checkIndex();
660 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
661
662 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
663
664 case vmIntrinsics::_isInstance:
665 case vmIntrinsics::_getModifiers:
666 case vmIntrinsics::_isInterface:
667 case vmIntrinsics::_isArray:
668 case vmIntrinsics::_isPrimitive:
669 case vmIntrinsics::_getSuperclass:
670 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
671
672 case vmIntrinsics::_floatToRawIntBits:
673 case vmIntrinsics::_floatToIntBits:
674 case vmIntrinsics::_intBitsToFloat:
675 case vmIntrinsics::_doubleToRawLongBits:
676 case vmIntrinsics::_doubleToLongBits:
677 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
678
679 case vmIntrinsics::_numberOfLeadingZeros_i:
680 case vmIntrinsics::_numberOfLeadingZeros_l:
681 case vmIntrinsics::_numberOfTrailingZeros_i:
682 case vmIntrinsics::_numberOfTrailingZeros_l:
683 case vmIntrinsics::_bitCount_i:
684 case vmIntrinsics::_bitCount_l:
685 case vmIntrinsics::_reverseBytes_i:
686 case vmIntrinsics::_reverseBytes_l:
687 case vmIntrinsics::_reverseBytes_s:
688 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
689
690 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
691
692 case vmIntrinsics::_Reference_get: return inline_reference_get();
693
694 case vmIntrinsics::_Class_cast: return inline_Class_cast();
695
696 case vmIntrinsics::_aescrypt_encryptBlock:
697 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
698
699 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
700 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
701 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
702
703 case vmIntrinsics::_counterMode_AESCrypt:
704 return inline_counterMode_AESCrypt(intrinsic_id());
705
706 case vmIntrinsics::_sha_implCompress:
707 case vmIntrinsics::_sha2_implCompress:
708 case vmIntrinsics::_sha5_implCompress:
709 return inline_sha_implCompress(intrinsic_id());
710
711 case vmIntrinsics::_digestBase_implCompressMB:
712 return inline_digestBase_implCompressMB(predicate);
713
714 case vmIntrinsics::_multiplyToLen:
715 return inline_multiplyToLen();
716
717 case vmIntrinsics::_squareToLen:
718 return inline_squareToLen();
719
720 case vmIntrinsics::_mulAdd:
721 return inline_mulAdd();
722
723 case vmIntrinsics::_montgomeryMultiply:
724 return inline_montgomeryMultiply();
725 case vmIntrinsics::_montgomerySquare:
726 return inline_montgomerySquare();
727
728 case vmIntrinsics::_vectorizedMismatch:
729 return inline_vectorizedMismatch();
730
731 case vmIntrinsics::_ghash_processBlocks:
732 return inline_ghash_processBlocks();
733
734 case vmIntrinsics::_encodeISOArray:
735 case vmIntrinsics::_encodeByteISOArray:
736 return inline_encodeISOArray();
737
738 case vmIntrinsics::_updateCRC32:
739 return inline_updateCRC32();
740 case vmIntrinsics::_updateBytesCRC32:
741 return inline_updateBytesCRC32();
742 case vmIntrinsics::_updateByteBufferCRC32:
743 return inline_updateByteBufferCRC32();
744
745 case vmIntrinsics::_updateBytesCRC32C:
746 return inline_updateBytesCRC32C();
747 case vmIntrinsics::_updateDirectByteBufferCRC32C:
748 return inline_updateDirectByteBufferCRC32C();
749
750 case vmIntrinsics::_updateBytesAdler32:
751 return inline_updateBytesAdler32();
752 case vmIntrinsics::_updateByteBufferAdler32:
753 return inline_updateByteBufferAdler32();
754
755 case vmIntrinsics::_profileBoolean:
756 return inline_profileBoolean();
757 case vmIntrinsics::_isCompileConstant:
758 return inline_isCompileConstant();
759
760 case vmIntrinsics::_hasNegatives:
761 return inline_hasNegatives();
762
763 case vmIntrinsics::_deoptimize:
764 return inline_deoptimize();
765
766 default:
767 // If you get here, it may be that someone has added a new intrinsic
768 // to the list in vmSymbols.hpp without implementing it here.
769 #ifndef PRODUCT
770 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
771 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
772 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
773 }
774 #endif
775 return false;
776 }
777 }
778
779 Node* LibraryCallKit::try_to_predicate(int predicate) {
780 if (!jvms()->has_method()) {
781 // Root JVMState has a null method.
782 assert(map()->memory()->Opcode() == Op_Parm, "");
783 // Insert the memory aliasing node
784 set_all_memory(reset_memory());
785 }
786 assert(merged_memory(), "");
787
788 switch (intrinsic_id()) {
789 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
790 return inline_cipherBlockChaining_AESCrypt_predicate(false);
791 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
792 return inline_cipherBlockChaining_AESCrypt_predicate(true);
793 case vmIntrinsics::_counterMode_AESCrypt:
794 return inline_counterMode_AESCrypt_predicate();
795 case vmIntrinsics::_digestBase_implCompressMB:
796 return inline_digestBase_implCompressMB_predicate(predicate);
797
798 default:
799 // If you get here, it may be that someone has added a new intrinsic
800 // to the list in vmSymbols.hpp without implementing it here.
801 #ifndef PRODUCT
802 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
803 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
804 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
805 }
806 #endif
807 Node* slow_ctl = control();
808 set_control(top()); // No fast path instrinsic
809 return slow_ctl;
810 }
811 }
812
813 //------------------------------set_result-------------------------------
814 // Helper function for finishing intrinsics.
815 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
816 record_for_igvn(region);
817 set_control(_gvn.transform(region));
818 set_result( _gvn.transform(value));
819 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
820 }
821
822 //------------------------------generate_guard---------------------------
823 // Helper function for generating guarded fast-slow graph structures.
824 // The given 'test', if true, guards a slow path. If the test fails
825 // then a fast path can be taken. (We generally hope it fails.)
826 // In all cases, GraphKit::control() is updated to the fast path.
827 // The returned value represents the control for the slow path.
828 // The return value is never 'top'; it is either a valid control
829 // or NULL if it is obvious that the slow path can never be taken.
830 // Also, if region and the slow control are not NULL, the slow edge
831 // is appended to the region.
832 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
833 if (stopped()) {
834 // Already short circuited.
835 return NULL;
836 }
837
838 // Build an if node and its projections.
839 // If test is true we take the slow path, which we assume is uncommon.
840 if (_gvn.type(test) == TypeInt::ZERO) {
841 // The slow branch is never taken. No need to build this guard.
842 return NULL;
843 }
844
845 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
846
847 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
848 if (if_slow == top()) {
849 // The slow branch is never taken. No need to build this guard.
850 return NULL;
851 }
852
853 if (region != NULL)
854 region->add_req(if_slow);
855
856 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
857 set_control(if_fast);
858
859 return if_slow;
860 }
861
862 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
863 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
864 }
865 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
866 return generate_guard(test, region, PROB_FAIR);
867 }
868
869 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
870 Node* *pos_index) {
871 if (stopped())
872 return NULL; // already stopped
873 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
874 return NULL; // index is already adequately typed
875 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
876 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
877 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
878 if (is_neg != NULL && pos_index != NULL) {
879 // Emulate effect of Parse::adjust_map_after_if.
880 Node* ccast = new CastIINode(index, TypeInt::POS);
881 ccast->set_req(0, control());
882 (*pos_index) = _gvn.transform(ccast);
883 }
884 return is_neg;
885 }
886
887 // Make sure that 'position' is a valid limit index, in [0..length].
888 // There are two equivalent plans for checking this:
889 // A. (offset + copyLength) unsigned<= arrayLength
890 // B. offset <= (arrayLength - copyLength)
891 // We require that all of the values above, except for the sum and
892 // difference, are already known to be non-negative.
893 // Plan A is robust in the face of overflow, if offset and copyLength
894 // are both hugely positive.
895 //
896 // Plan B is less direct and intuitive, but it does not overflow at
897 // all, since the difference of two non-negatives is always
898 // representable. Whenever Java methods must perform the equivalent
899 // check they generally use Plan B instead of Plan A.
900 // For the moment we use Plan A.
901 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
902 Node* subseq_length,
903 Node* array_length,
904 RegionNode* region) {
905 if (stopped())
906 return NULL; // already stopped
907 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
908 if (zero_offset && subseq_length->eqv_uncast(array_length))
909 return NULL; // common case of whole-array copy
910 Node* last = subseq_length;
911 if (!zero_offset) // last += offset
912 last = _gvn.transform(new AddINode(last, offset));
913 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
914 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
915 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
916 return is_over;
917 }
918
919 // Emit range checks for the given String.value byte array
920 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
921 if (stopped()) {
922 return; // already stopped
923 }
924 RegionNode* bailout = new RegionNode(1);
925 record_for_igvn(bailout);
926 if (char_count) {
927 // Convert char count to byte count
928 count = _gvn.transform(new LShiftINode(count, intcon(1)));
929 }
930
931 // Offset and count must not be negative
932 generate_negative_guard(offset, bailout);
933 generate_negative_guard(count, bailout);
934 // Offset + count must not exceed length of array
935 generate_limit_guard(offset, count, load_array_length(array), bailout);
936
937 if (bailout->req() > 1) {
938 PreserveJVMState pjvms(this);
939 set_control(_gvn.transform(bailout));
940 uncommon_trap(Deoptimization::Reason_intrinsic,
941 Deoptimization::Action_maybe_recompile);
942 }
943 }
944
945 //--------------------------generate_current_thread--------------------
946 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
947 ciKlass* thread_klass = env()->Thread_klass();
948 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
949 Node* thread = _gvn.transform(new ThreadLocalNode());
950 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
951 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
952 tls_output = thread;
953 return threadObj;
954 }
955
956
957 //------------------------------make_string_method_node------------------------
958 // Helper method for String intrinsic functions. This version is called with
959 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
960 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
961 // containing the lengths of str1 and str2.
962 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
963 Node* result = NULL;
964 switch (opcode) {
965 case Op_StrIndexOf:
966 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
967 str1_start, cnt1, str2_start, cnt2, ae);
968 break;
969 case Op_StrComp:
970 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
971 str1_start, cnt1, str2_start, cnt2, ae);
972 break;
973 case Op_StrEquals:
974 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
975 str1_start, str2_start, cnt1, ae);
976 break;
977 default:
978 ShouldNotReachHere();
979 return NULL;
980 }
981
982 // All these intrinsics have checks.
983 C->set_has_split_ifs(true); // Has chance for split-if optimization
984
985 return _gvn.transform(result);
986 }
987
988 //------------------------------inline_string_compareTo------------------------
989 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
990 Node* arg1 = argument(0);
991 Node* arg2 = argument(1);
992
993 // Get start addr and length of first argument
994 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
995 Node* arg1_cnt = load_array_length(arg1);
996
997 // Get start addr and length of second argument
998 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
999 Node* arg2_cnt = load_array_length(arg2);
1000
1001 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1002 set_result(result);
1003 return true;
1004 }
1005
1006 //------------------------------inline_string_equals------------------------
1007 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1008 Node* arg1 = argument(0);
1009 Node* arg2 = argument(1);
1010
1011 // paths (plus control) merge
1012 RegionNode* region = new RegionNode(3);
1013 Node* phi = new PhiNode(region, TypeInt::BOOL);
1014
1015 if (!stopped()) {
1016 // Get start addr and length of first argument
1017 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1018 Node* arg1_cnt = load_array_length(arg1);
1019
1020 // Get start addr and length of second argument
1021 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1022 Node* arg2_cnt = load_array_length(arg2);
1023
1024 // Check for arg1_cnt != arg2_cnt
1025 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1026 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1027 Node* if_ne = generate_slow_guard(bol, NULL);
1028 if (if_ne != NULL) {
1029 phi->init_req(2, intcon(0));
1030 region->init_req(2, if_ne);
1031 }
1032
1033 // Check for count == 0 is done by assembler code for StrEquals.
1034
1035 if (!stopped()) {
1036 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1037 phi->init_req(1, equals);
1038 region->init_req(1, control());
1039 }
1040 }
1041
1042 // post merge
1043 set_control(_gvn.transform(region));
1044 record_for_igvn(region);
1045
1046 set_result(_gvn.transform(phi));
1047 return true;
1048 }
1049
1050 //------------------------------inline_array_equals----------------------------
1051 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1052 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1053 Node* arg1 = argument(0);
1054 Node* arg2 = argument(1);
1055
1056 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1057 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1058 return true;
1059 }
1060
1061 //------------------------------inline_hasNegatives------------------------------
1062 bool LibraryCallKit::inline_hasNegatives() {
1063 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1064 return false;
1065 }
1066
1067 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1068 // no receiver since it is static method
1069 Node* ba = argument(0);
1070 Node* offset = argument(1);
1071 Node* len = argument(2);
1072
1073 // Range checks
1074 generate_string_range_check(ba, offset, len, false);
1075 if (stopped()) {
1076 return true;
1077 }
1078 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1079 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1080 set_result(_gvn.transform(result));
1081 return true;
1082 }
1083
1084 bool LibraryCallKit::inline_objects_checkIndex() {
1085 Node* index = argument(0);
1086 Node* length = argument(1);
1087 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1088 return false;
1089 }
1090
1091 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1092 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1093
1094 {
1095 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1096 uncommon_trap(Deoptimization::Reason_intrinsic,
1097 Deoptimization::Action_make_not_entrant);
1098 }
1099
1100 if (stopped()) {
1101 return false;
1102 }
1103
1104 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1105 BoolTest::mask btest = BoolTest::lt;
1106 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1107 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1108 _gvn.set_type(rc, rc->Value(&_gvn));
1109 if (!rc_bool->is_Con()) {
1110 record_for_igvn(rc);
1111 }
1112 set_control(_gvn.transform(new IfTrueNode(rc)));
1113 {
1114 PreserveJVMState pjvms(this);
1115 set_control(_gvn.transform(new IfFalseNode(rc)));
1116 uncommon_trap(Deoptimization::Reason_range_check,
1117 Deoptimization::Action_make_not_entrant);
1118 }
1119
1120 if (stopped()) {
1121 return false;
1122 }
1123
1124 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1125 result->set_req(0, control());
1126 result = _gvn.transform(result);
1127 set_result(result);
1128 replace_in_map(index, result);
1129 return true;
1130 }
1131
1132 //------------------------------inline_string_indexOf------------------------
1133 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1134 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1135 return false;
1136 }
1137 Node* src = argument(0);
1138 Node* tgt = argument(1);
1139
1140 // Make the merge point
1141 RegionNode* result_rgn = new RegionNode(4);
1142 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1143
1144 // Get start addr and length of source string
1145 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1146 Node* src_count = load_array_length(src);
1147
1148 // Get start addr and length of substring
1149 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1150 Node* tgt_count = load_array_length(tgt);
1151
1152 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1153 // Divide src size by 2 if String is UTF16 encoded
1154 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1155 }
1156 if (ae == StrIntrinsicNode::UU) {
1157 // Divide substring size by 2 if String is UTF16 encoded
1158 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1159 }
1160
1161 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1162 if (result != NULL) {
1163 result_phi->init_req(3, result);
1164 result_rgn->init_req(3, control());
1165 }
1166 set_control(_gvn.transform(result_rgn));
1167 record_for_igvn(result_rgn);
1168 set_result(_gvn.transform(result_phi));
1169
1170 return true;
1171 }
1172
1173 //-----------------------------inline_string_indexOf-----------------------
1174 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1175 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1176 return false;
1177 }
1178 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1179 return false;
1180 }
1181 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1182 Node* src = argument(0); // byte[]
1183 Node* src_count = argument(1); // char count
1184 Node* tgt = argument(2); // byte[]
1185 Node* tgt_count = argument(3); // char count
1186 Node* from_index = argument(4); // char index
1187
1188 // Multiply byte array index by 2 if String is UTF16 encoded
1189 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1190 src_count = _gvn.transform(new SubINode(src_count, from_index));
1191 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1192 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1193
1194 // Range checks
1195 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1196 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1197 if (stopped()) {
1198 return true;
1199 }
1200
1201 RegionNode* region = new RegionNode(5);
1202 Node* phi = new PhiNode(region, TypeInt::INT);
1203
1204 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1205 if (result != NULL) {
1206 // The result is index relative to from_index if substring was found, -1 otherwise.
1207 // Generate code which will fold into cmove.
1208 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1209 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1210
1211 Node* if_lt = generate_slow_guard(bol, NULL);
1212 if (if_lt != NULL) {
1213 // result == -1
1214 phi->init_req(3, result);
1215 region->init_req(3, if_lt);
1216 }
1217 if (!stopped()) {
1218 result = _gvn.transform(new AddINode(result, from_index));
1219 phi->init_req(4, result);
1220 region->init_req(4, control());
1221 }
1222 }
1223
1224 set_control(_gvn.transform(region));
1225 record_for_igvn(region);
1226 set_result(_gvn.transform(phi));
1227
1228 return true;
1229 }
1230
1231 // Create StrIndexOfNode with fast path checks
1232 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1233 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1234 // Check for substr count > string count
1235 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1236 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1237 Node* if_gt = generate_slow_guard(bol, NULL);
1238 if (if_gt != NULL) {
1239 phi->init_req(1, intcon(-1));
1240 region->init_req(1, if_gt);
1241 }
1242 if (!stopped()) {
1243 // Check for substr count == 0
1244 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1245 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1246 Node* if_zero = generate_slow_guard(bol, NULL);
1247 if (if_zero != NULL) {
1248 phi->init_req(2, intcon(0));
1249 region->init_req(2, if_zero);
1250 }
1251 }
1252 if (!stopped()) {
1253 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1254 }
1255 return NULL;
1256 }
1257
1258 //-----------------------------inline_string_indexOfChar-----------------------
1259 bool LibraryCallKit::inline_string_indexOfChar() {
1260 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1261 return false;
1262 }
1263 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) {
1264 return false;
1265 }
1266 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1267 Node* src = argument(0); // byte[]
1268 Node* tgt = argument(1); // tgt is int ch
1269 Node* from_index = argument(2);
1270 Node* max = argument(3);
1271
1272 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1273 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1274 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1275
1276 // Range checks
1277 generate_string_range_check(src, src_offset, src_count, true);
1278 if (stopped()) {
1279 return true;
1280 }
1281
1282 RegionNode* region = new RegionNode(3);
1283 Node* phi = new PhiNode(region, TypeInt::INT);
1284
1285 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1286 C->set_has_split_ifs(true); // Has chance for split-if optimization
1287 _gvn.transform(result);
1288
1289 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1290 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1291
1292 Node* if_lt = generate_slow_guard(bol, NULL);
1293 if (if_lt != NULL) {
1294 // result == -1
1295 phi->init_req(2, result);
1296 region->init_req(2, if_lt);
1297 }
1298 if (!stopped()) {
1299 result = _gvn.transform(new AddINode(result, from_index));
1300 phi->init_req(1, result);
1301 region->init_req(1, control());
1302 }
1303 set_control(_gvn.transform(region));
1304 record_for_igvn(region);
1305 set_result(_gvn.transform(phi));
1306
1307 return true;
1308 }
1309 //---------------------------inline_string_copy---------------------
1310 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1311 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1312 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1313 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1314 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1315 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1316 bool LibraryCallKit::inline_string_copy(bool compress) {
1317 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1318 return false;
1319 }
1320 int nargs = 5; // 2 oops, 3 ints
1321 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1322
1323 Node* src = argument(0);
1324 Node* src_offset = argument(1);
1325 Node* dst = argument(2);
1326 Node* dst_offset = argument(3);
1327 Node* length = argument(4);
1328
1329 // Check for allocation before we add nodes that would confuse
1330 // tightly_coupled_allocation()
1331 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1332
1333 // Figure out the size and type of the elements we will be copying.
1334 const Type* src_type = src->Value(&_gvn);
1335 const Type* dst_type = dst->Value(&_gvn);
1336 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1337 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1338 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1339 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1340 "Unsupported array types for inline_string_copy");
1341
1342 // Range checks
1343 generate_string_range_check(src, src_offset, length, compress && src_elem == T_BYTE);
1344 generate_string_range_check(dst, dst_offset, length, !compress && dst_elem == T_BYTE);
1345 if (stopped()) {
1346 return true;
1347 }
1348
1349 // Convert char[] offsets to byte[] offsets
1350 if (compress && src_elem == T_BYTE) {
1351 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1352 } else if (!compress && dst_elem == T_BYTE) {
1353 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1354 }
1355
1356 Node* src_start = array_element_address(src, src_offset, src_elem);
1357 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1358 // 'src_start' points to src array + scaled offset
1359 // 'dst_start' points to dst array + scaled offset
1360 Node* count = NULL;
1361 if (compress) {
1362 count = compress_string(src_start, dst_start, length);
1363 } else {
1364 inflate_string(src_start, dst_start, length);
1365 }
1366
1367 if (alloc != NULL) {
1368 if (alloc->maybe_set_complete(&_gvn)) {
1369 // "You break it, you buy it."
1370 InitializeNode* init = alloc->initialization();
1371 assert(init->is_complete(), "we just did this");
1372 init->set_complete_with_arraycopy();
1373 assert(dst->is_CheckCastPP(), "sanity");
1374 assert(dst->in(0)->in(0) == init, "dest pinned");
1375 }
1376 // Do not let stores that initialize this object be reordered with
1377 // a subsequent store that would make this object accessible by
1378 // other threads.
1379 // Record what AllocateNode this StoreStore protects so that
1380 // escape analysis can go from the MemBarStoreStoreNode to the
1381 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1382 // based on the escape status of the AllocateNode.
1383 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1384 }
1385 if (compress) {
1386 set_result(_gvn.transform(count));
1387 }
1388 return true;
1389 }
1390
1391 #ifdef _LP64
1392 #define XTOP ,top() /*additional argument*/
1393 #else //_LP64
1394 #define XTOP /*no additional argument*/
1395 #endif //_LP64
1396
1397 //------------------------inline_string_toBytesU--------------------------
1398 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1399 bool LibraryCallKit::inline_string_toBytesU() {
1400 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1401 return false;
1402 }
1403 // Get the arguments.
1404 Node* value = argument(0);
1405 Node* offset = argument(1);
1406 Node* length = argument(2);
1407
1408 Node* newcopy = NULL;
1409
1410 // Set the original stack and the reexecute bit for the interpreter to reexecute
1411 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1412 { PreserveReexecuteState preexecs(this);
1413 jvms()->set_should_reexecute(true);
1414
1415 // Check if a null path was taken unconditionally.
1416 value = null_check(value);
1417
1418 RegionNode* bailout = new RegionNode(1);
1419 record_for_igvn(bailout);
1420
1421 // Range checks
1422 generate_negative_guard(offset, bailout);
1423 generate_negative_guard(length, bailout);
1424 generate_limit_guard(offset, length, load_array_length(value), bailout);
1425 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1426 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1427
1428 if (bailout->req() > 1) {
1429 PreserveJVMState pjvms(this);
1430 set_control(_gvn.transform(bailout));
1431 uncommon_trap(Deoptimization::Reason_intrinsic,
1432 Deoptimization::Action_maybe_recompile);
1433 }
1434 if (stopped()) {
1435 return true;
1436 }
1437
1438 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1439 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1440 newcopy = new_array(klass_node, size, 0); // no arguments to push
1441 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1442
1443 // Calculate starting addresses.
1444 Node* src_start = array_element_address(value, offset, T_CHAR);
1445 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1446
1447 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1448 const TypeInt* toffset = gvn().type(offset)->is_int();
1449 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1450
1451 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1452 const char* copyfunc_name = "arraycopy";
1453 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1454 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1455 OptoRuntime::fast_arraycopy_Type(),
1456 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1457 src_start, dst_start, ConvI2X(length) XTOP);
1458 // Do not let reads from the cloned object float above the arraycopy.
1459 if (alloc != NULL) {
1460 if (alloc->maybe_set_complete(&_gvn)) {
1461 // "You break it, you buy it."
1462 InitializeNode* init = alloc->initialization();
1463 assert(init->is_complete(), "we just did this");
1464 init->set_complete_with_arraycopy();
1465 assert(newcopy->is_CheckCastPP(), "sanity");
1466 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1467 }
1468 // Do not let stores that initialize this object be reordered with
1469 // a subsequent store that would make this object accessible by
1470 // other threads.
1471 // Record what AllocateNode this StoreStore protects so that
1472 // escape analysis can go from the MemBarStoreStoreNode to the
1473 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1474 // based on the escape status of the AllocateNode.
1475 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1476 } else {
1477 insert_mem_bar(Op_MemBarCPUOrder);
1478 }
1479 } // original reexecute is set back here
1480
1481 C->set_has_split_ifs(true); // Has chance for split-if optimization
1482 if (!stopped()) {
1483 set_result(newcopy);
1484 }
1485 return true;
1486 }
1487
1488 //------------------------inline_string_getCharsU--------------------------
1489 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1490 bool LibraryCallKit::inline_string_getCharsU() {
1491 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1492 return false;
1493 }
1494
1495 // Get the arguments.
1496 Node* src = argument(0);
1497 Node* src_begin = argument(1);
1498 Node* src_end = argument(2); // exclusive offset (i < src_end)
1499 Node* dst = argument(3);
1500 Node* dst_begin = argument(4);
1501
1502 // Check for allocation before we add nodes that would confuse
1503 // tightly_coupled_allocation()
1504 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1505
1506 // Check if a null path was taken unconditionally.
1507 src = null_check(src);
1508 dst = null_check(dst);
1509 if (stopped()) {
1510 return true;
1511 }
1512
1513 // Get length and convert char[] offset to byte[] offset
1514 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1515 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1516
1517 // Range checks
1518 generate_string_range_check(src, src_begin, length, true);
1519 generate_string_range_check(dst, dst_begin, length, false);
1520 if (stopped()) {
1521 return true;
1522 }
1523
1524 if (!stopped()) {
1525 // Calculate starting addresses.
1526 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1527 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1528
1529 // Check if array addresses are aligned to HeapWordSize
1530 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1531 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1532 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1533 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1534
1535 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1536 const char* copyfunc_name = "arraycopy";
1537 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1538 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1539 OptoRuntime::fast_arraycopy_Type(),
1540 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1541 src_start, dst_start, ConvI2X(length) XTOP);
1542 // Do not let reads from the cloned object float above the arraycopy.
1543 if (alloc != NULL) {
1544 if (alloc->maybe_set_complete(&_gvn)) {
1545 // "You break it, you buy it."
1546 InitializeNode* init = alloc->initialization();
1547 assert(init->is_complete(), "we just did this");
1548 init->set_complete_with_arraycopy();
1549 assert(dst->is_CheckCastPP(), "sanity");
1550 assert(dst->in(0)->in(0) == init, "dest pinned");
1551 }
1552 // Do not let stores that initialize this object be reordered with
1553 // a subsequent store that would make this object accessible by
1554 // other threads.
1555 // Record what AllocateNode this StoreStore protects so that
1556 // escape analysis can go from the MemBarStoreStoreNode to the
1557 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1558 // based on the escape status of the AllocateNode.
1559 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1560 } else {
1561 insert_mem_bar(Op_MemBarCPUOrder);
1562 }
1563 }
1564
1565 C->set_has_split_ifs(true); // Has chance for split-if optimization
1566 return true;
1567 }
1568
1569 //----------------------inline_string_char_access----------------------------
1570 // Store/Load char to/from byte[] array.
1571 // static void StringUTF16.putChar(byte[] val, int index, int c)
1572 // static char StringUTF16.getChar(byte[] val, int index)
1573 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1574 Node* value = argument(0);
1575 Node* index = argument(1);
1576 Node* ch = is_store ? argument(2) : NULL;
1577
1578 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1579 // correctly requires matched array shapes.
1580 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1581 "sanity: byte[] and char[] bases agree");
1582 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1583 "sanity: byte[] and char[] scales agree");
1584
1585 Node* adr = array_element_address(value, index, T_CHAR);
1586 if (is_store) {
1587 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1588 false, false, true /* mismatched */);
1589 } else {
1590 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered,
1591 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1592 set_result(ch);
1593 }
1594 return true;
1595 }
1596
1597 //--------------------------round_double_node--------------------------------
1598 // Round a double node if necessary.
1599 Node* LibraryCallKit::round_double_node(Node* n) {
1600 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1601 n = _gvn.transform(new RoundDoubleNode(0, n));
1602 return n;
1603 }
1604
1605 //------------------------------inline_math-----------------------------------
1606 // public static double Math.abs(double)
1607 // public static double Math.sqrt(double)
1608 // public static double Math.log(double)
1609 // public static double Math.log10(double)
1610 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1611 Node* arg = round_double_node(argument(0));
1612 Node* n = NULL;
1613 switch (id) {
1614 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1615 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1616 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1617 default: fatal_unexpected_iid(id); break;
1618 }
1619 set_result(_gvn.transform(n));
1620 return true;
1621 }
1622
1623 //------------------------------inline_trig----------------------------------
1624 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1625 // argument reduction which will turn into a fast/slow diamond.
1626 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1627 Node* arg = round_double_node(argument(0));
1628 Node* n = NULL;
1629
1630 switch (id) {
1631 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1632 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1633 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1634 default: fatal_unexpected_iid(id); break;
1635 }
1636 n = _gvn.transform(n);
1637
1638 // Rounding required? Check for argument reduction!
1639 if (Matcher::strict_fp_requires_explicit_rounding) {
1640 static const double pi_4 = 0.7853981633974483;
1641 static const double neg_pi_4 = -0.7853981633974483;
1642 // pi/2 in 80-bit extended precision
1643 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1644 // -pi/2 in 80-bit extended precision
1645 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1646 // Cutoff value for using this argument reduction technique
1647 //static const double pi_2_minus_epsilon = 1.564660403643354;
1648 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1649
1650 // Pseudocode for sin:
1651 // if (x <= Math.PI / 4.0) {
1652 // if (x >= -Math.PI / 4.0) return fsin(x);
1653 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1654 // } else {
1655 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1656 // }
1657 // return StrictMath.sin(x);
1658
1659 // Pseudocode for cos:
1660 // if (x <= Math.PI / 4.0) {
1661 // if (x >= -Math.PI / 4.0) return fcos(x);
1662 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1663 // } else {
1664 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1665 // }
1666 // return StrictMath.cos(x);
1667
1668 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1669 // requires a special machine instruction to load it. Instead we'll try
1670 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1671 // probably do the math inside the SIN encoding.
1672
1673 // Make the merge point
1674 RegionNode* r = new RegionNode(3);
1675 Node* phi = new PhiNode(r, Type::DOUBLE);
1676
1677 // Flatten arg so we need only 1 test
1678 Node *abs = _gvn.transform(new AbsDNode(arg));
1679 // Node for PI/4 constant
1680 Node *pi4 = makecon(TypeD::make(pi_4));
1681 // Check PI/4 : abs(arg)
1682 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1683 // Check: If PI/4 < abs(arg) then go slow
1684 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1685 // Branch either way
1686 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1687 set_control(opt_iff(r,iff));
1688
1689 // Set fast path result
1690 phi->init_req(2, n);
1691
1692 // Slow path - non-blocking leaf call
1693 Node* call = NULL;
1694 switch (id) {
1695 case vmIntrinsics::_dsin:
1696 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1697 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1698 "Sin", NULL, arg, top());
1699 break;
1700 case vmIntrinsics::_dcos:
1701 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1702 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1703 "Cos", NULL, arg, top());
1704 break;
1705 case vmIntrinsics::_dtan:
1706 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1707 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1708 "Tan", NULL, arg, top());
1709 break;
1710 }
1711 assert(control()->in(0) == call, "");
1712 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1713 r->init_req(1, control());
1714 phi->init_req(1, slow_result);
1715
1716 // Post-merge
1717 set_control(_gvn.transform(r));
1718 record_for_igvn(r);
1719 n = _gvn.transform(phi);
1720
1721 C->set_has_split_ifs(true); // Has chance for split-if optimization
1722 }
1723 set_result(n);
1724 return true;
1725 }
1726
1727 //------------------------------runtime_math-----------------------------
1728 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1729 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1730 "must be (DD)D or (D)D type");
1731
1732 // Inputs
1733 Node* a = round_double_node(argument(0));
1734 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1735
1736 const TypePtr* no_memory_effects = NULL;
1737 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1738 no_memory_effects,
1739 a, top(), b, b ? top() : NULL);
1740 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1741 #ifdef ASSERT
1742 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1743 assert(value_top == top(), "second value must be top");
1744 #endif
1745
1746 set_result(value);
1747 return true;
1748 }
1749
1750 //------------------------------inline_math_native-----------------------------
1751 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1752 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1753 switch (id) {
1754 // These intrinsics are not properly supported on all hardware
1755 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1756 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1757 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1758 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1759 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1760 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1761
1762 case vmIntrinsics::_dlog:
1763 return StubRoutines::dlog() != NULL ?
1764 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1765 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1766 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1767 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1768
1769 // These intrinsics are supported on all hardware
1770 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1771 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1772
1773 case vmIntrinsics::_dexp:
1774 return StubRoutines::dexp() != NULL ?
1775 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1776 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1777 case vmIntrinsics::_dpow:
1778 return StubRoutines::dpow() != NULL ?
1779 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1780 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1781 #undef FN_PTR
1782
1783 // These intrinsics are not yet correctly implemented
1784 case vmIntrinsics::_datan2:
1785 return false;
1786
1787 default:
1788 fatal_unexpected_iid(id);
1789 return false;
1790 }
1791 }
1792
1793 static bool is_simple_name(Node* n) {
1794 return (n->req() == 1 // constant
1795 || (n->is_Type() && n->as_Type()->type()->singleton())
1796 || n->is_Proj() // parameter or return value
1797 || n->is_Phi() // local of some sort
1798 );
1799 }
1800
1801 //----------------------------inline_notify-----------------------------------*
1802 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1803 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1804 address func;
1805 if (id == vmIntrinsics::_notify) {
1806 func = OptoRuntime::monitor_notify_Java();
1807 } else {
1808 func = OptoRuntime::monitor_notifyAll_Java();
1809 }
1810 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1811 make_slow_call_ex(call, env()->Throwable_klass(), false);
1812 return true;
1813 }
1814
1815
1816 //----------------------------inline_min_max-----------------------------------
1817 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1818 set_result(generate_min_max(id, argument(0), argument(1)));
1819 return true;
1820 }
1821
1822 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1823 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1824 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1825 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1826 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1827
1828 {
1829 PreserveJVMState pjvms(this);
1830 PreserveReexecuteState preexecs(this);
1831 jvms()->set_should_reexecute(true);
1832
1833 set_control(slow_path);
1834 set_i_o(i_o());
1835
1836 uncommon_trap(Deoptimization::Reason_intrinsic,
1837 Deoptimization::Action_none);
1838 }
1839
1840 set_control(fast_path);
1841 set_result(math);
1842 }
1843
1844 template <typename OverflowOp>
1845 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1846 typedef typename OverflowOp::MathOp MathOp;
1847
1848 MathOp* mathOp = new MathOp(arg1, arg2);
1849 Node* operation = _gvn.transform( mathOp );
1850 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1851 inline_math_mathExact(operation, ofcheck);
1852 return true;
1853 }
1854
1855 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1856 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1857 }
1858
1859 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1860 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1861 }
1862
1863 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1864 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1865 }
1866
1867 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1868 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1869 }
1870
1871 bool LibraryCallKit::inline_math_negateExactI() {
1872 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1873 }
1874
1875 bool LibraryCallKit::inline_math_negateExactL() {
1876 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1877 }
1878
1879 bool LibraryCallKit::inline_math_multiplyExactI() {
1880 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1881 }
1882
1883 bool LibraryCallKit::inline_math_multiplyExactL() {
1884 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1885 }
1886
1887 Node*
1888 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1889 // These are the candidate return value:
1890 Node* xvalue = x0;
1891 Node* yvalue = y0;
1892
1893 if (xvalue == yvalue) {
1894 return xvalue;
1895 }
1896
1897 bool want_max = (id == vmIntrinsics::_max);
1898
1899 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1900 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1901 if (txvalue == NULL || tyvalue == NULL) return top();
1902 // This is not really necessary, but it is consistent with a
1903 // hypothetical MaxINode::Value method:
1904 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1905
1906 // %%% This folding logic should (ideally) be in a different place.
1907 // Some should be inside IfNode, and there to be a more reliable
1908 // transformation of ?: style patterns into cmoves. We also want
1909 // more powerful optimizations around cmove and min/max.
1910
1911 // Try to find a dominating comparison of these guys.
1912 // It can simplify the index computation for Arrays.copyOf
1913 // and similar uses of System.arraycopy.
1914 // First, compute the normalized version of CmpI(x, y).
1915 int cmp_op = Op_CmpI;
1916 Node* xkey = xvalue;
1917 Node* ykey = yvalue;
1918 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1919 if (ideal_cmpxy->is_Cmp()) {
1920 // E.g., if we have CmpI(length - offset, count),
1921 // it might idealize to CmpI(length, count + offset)
1922 cmp_op = ideal_cmpxy->Opcode();
1923 xkey = ideal_cmpxy->in(1);
1924 ykey = ideal_cmpxy->in(2);
1925 }
1926
1927 // Start by locating any relevant comparisons.
1928 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1929 Node* cmpxy = NULL;
1930 Node* cmpyx = NULL;
1931 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1932 Node* cmp = start_from->fast_out(k);
1933 if (cmp->outcnt() > 0 && // must have prior uses
1934 cmp->in(0) == NULL && // must be context-independent
1935 cmp->Opcode() == cmp_op) { // right kind of compare
1936 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1937 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1938 }
1939 }
1940
1941 const int NCMPS = 2;
1942 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1943 int cmpn;
1944 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1945 if (cmps[cmpn] != NULL) break; // find a result
1946 }
1947 if (cmpn < NCMPS) {
1948 // Look for a dominating test that tells us the min and max.
1949 int depth = 0; // Limit search depth for speed
1950 Node* dom = control();
1951 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1952 if (++depth >= 100) break;
1953 Node* ifproj = dom;
1954 if (!ifproj->is_Proj()) continue;
1955 Node* iff = ifproj->in(0);
1956 if (!iff->is_If()) continue;
1957 Node* bol = iff->in(1);
1958 if (!bol->is_Bool()) continue;
1959 Node* cmp = bol->in(1);
1960 if (cmp == NULL) continue;
1961 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1962 if (cmps[cmpn] == cmp) break;
1963 if (cmpn == NCMPS) continue;
1964 BoolTest::mask btest = bol->as_Bool()->_test._test;
1965 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1966 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1967 // At this point, we know that 'x btest y' is true.
1968 switch (btest) {
1969 case BoolTest::eq:
1970 // They are proven equal, so we can collapse the min/max.
1971 // Either value is the answer. Choose the simpler.
1972 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1973 return yvalue;
1974 return xvalue;
1975 case BoolTest::lt: // x < y
1976 case BoolTest::le: // x <= y
1977 return (want_max ? yvalue : xvalue);
1978 case BoolTest::gt: // x > y
1979 case BoolTest::ge: // x >= y
1980 return (want_max ? xvalue : yvalue);
1981 }
1982 }
1983 }
1984
1985 // We failed to find a dominating test.
1986 // Let's pick a test that might GVN with prior tests.
1987 Node* best_bol = NULL;
1988 BoolTest::mask best_btest = BoolTest::illegal;
1989 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1990 Node* cmp = cmps[cmpn];
1991 if (cmp == NULL) continue;
1992 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1993 Node* bol = cmp->fast_out(j);
1994 if (!bol->is_Bool()) continue;
1995 BoolTest::mask btest = bol->as_Bool()->_test._test;
1996 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1997 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1998 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1999 best_bol = bol->as_Bool();
2000 best_btest = btest;
2001 }
2002 }
2003 }
2004
2005 Node* answer_if_true = NULL;
2006 Node* answer_if_false = NULL;
2007 switch (best_btest) {
2008 default:
2009 if (cmpxy == NULL)
2010 cmpxy = ideal_cmpxy;
2011 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2012 // and fall through:
2013 case BoolTest::lt: // x < y
2014 case BoolTest::le: // x <= y
2015 answer_if_true = (want_max ? yvalue : xvalue);
2016 answer_if_false = (want_max ? xvalue : yvalue);
2017 break;
2018 case BoolTest::gt: // x > y
2019 case BoolTest::ge: // x >= y
2020 answer_if_true = (want_max ? xvalue : yvalue);
2021 answer_if_false = (want_max ? yvalue : xvalue);
2022 break;
2023 }
2024
2025 jint hi, lo;
2026 if (want_max) {
2027 // We can sharpen the minimum.
2028 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2029 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2030 } else {
2031 // We can sharpen the maximum.
2032 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2033 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2034 }
2035
2036 // Use a flow-free graph structure, to avoid creating excess control edges
2037 // which could hinder other optimizations.
2038 // Since Math.min/max is often used with arraycopy, we want
2039 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2040 Node* cmov = CMoveNode::make(NULL, best_bol,
2041 answer_if_false, answer_if_true,
2042 TypeInt::make(lo, hi, widen));
2043
2044 return _gvn.transform(cmov);
2045
2046 /*
2047 // This is not as desirable as it may seem, since Min and Max
2048 // nodes do not have a full set of optimizations.
2049 // And they would interfere, anyway, with 'if' optimizations
2050 // and with CMoveI canonical forms.
2051 switch (id) {
2052 case vmIntrinsics::_min:
2053 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2054 case vmIntrinsics::_max:
2055 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2056 default:
2057 ShouldNotReachHere();
2058 }
2059 */
2060 }
2061
2062 inline int
2063 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2064 const TypePtr* base_type = TypePtr::NULL_PTR;
2065 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2066 if (base_type == NULL) {
2067 // Unknown type.
2068 return Type::AnyPtr;
2069 } else if (base_type == TypePtr::NULL_PTR) {
2070 // Since this is a NULL+long form, we have to switch to a rawptr.
2071 base = _gvn.transform(new CastX2PNode(offset));
2072 offset = MakeConX(0);
2073 return Type::RawPtr;
2074 } else if (base_type->base() == Type::RawPtr) {
2075 return Type::RawPtr;
2076 } else if (base_type->isa_oopptr()) {
2077 // Base is never null => always a heap address.
2078 if (base_type->ptr() == TypePtr::NotNull) {
2079 return Type::OopPtr;
2080 }
2081 // Offset is small => always a heap address.
2082 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2083 if (offset_type != NULL &&
2084 base_type->offset() == 0 && // (should always be?)
2085 offset_type->_lo >= 0 &&
2086 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2087 return Type::OopPtr;
2088 }
2089 // Otherwise, it might either be oop+off or NULL+addr.
2090 return Type::AnyPtr;
2091 } else {
2092 // No information:
2093 return Type::AnyPtr;
2094 }
2095 }
2096
2097 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2098 int kind = classify_unsafe_addr(base, offset);
2099 if (kind == Type::RawPtr) {
2100 return basic_plus_adr(top(), base, offset);
2101 } else {
2102 return basic_plus_adr(base, offset);
2103 }
2104 }
2105
2106 //--------------------------inline_number_methods-----------------------------
2107 // inline int Integer.numberOfLeadingZeros(int)
2108 // inline int Long.numberOfLeadingZeros(long)
2109 //
2110 // inline int Integer.numberOfTrailingZeros(int)
2111 // inline int Long.numberOfTrailingZeros(long)
2112 //
2113 // inline int Integer.bitCount(int)
2114 // inline int Long.bitCount(long)
2115 //
2116 // inline char Character.reverseBytes(char)
2117 // inline short Short.reverseBytes(short)
2118 // inline int Integer.reverseBytes(int)
2119 // inline long Long.reverseBytes(long)
2120 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2121 Node* arg = argument(0);
2122 Node* n = NULL;
2123 switch (id) {
2124 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2125 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2126 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2127 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2128 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2129 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2130 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2131 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2132 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2133 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2134 default: fatal_unexpected_iid(id); break;
2135 }
2136 set_result(_gvn.transform(n));
2137 return true;
2138 }
2139
2140 //----------------------------inline_unsafe_access----------------------------
2141
2142 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2143
2144 // Helper that guards and inserts a pre-barrier.
2145 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2146 Node* pre_val, bool need_mem_bar) {
2147 // We could be accessing the referent field of a reference object. If so, when G1
2148 // is enabled, we need to log the value in the referent field in an SATB buffer.
2149 // This routine performs some compile time filters and generates suitable
2150 // runtime filters that guard the pre-barrier code.
2151 // Also add memory barrier for non volatile load from the referent field
2152 // to prevent commoning of loads across safepoint.
2153 if (!UseG1GC && !need_mem_bar)
2154 return;
2155
2156 // Some compile time checks.
2157
2158 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2159 const TypeX* otype = offset->find_intptr_t_type();
2160 if (otype != NULL && otype->is_con() &&
2161 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2162 // Constant offset but not the reference_offset so just return
2163 return;
2164 }
2165
2166 // We only need to generate the runtime guards for instances.
2167 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2168 if (btype != NULL) {
2169 if (btype->isa_aryptr()) {
2170 // Array type so nothing to do
2171 return;
2172 }
2173
2174 const TypeInstPtr* itype = btype->isa_instptr();
2175 if (itype != NULL) {
2176 // Can the klass of base_oop be statically determined to be
2177 // _not_ a sub-class of Reference and _not_ Object?
2178 ciKlass* klass = itype->klass();
2179 if ( klass->is_loaded() &&
2180 !klass->is_subtype_of(env()->Reference_klass()) &&
2181 !env()->Object_klass()->is_subtype_of(klass)) {
2182 return;
2183 }
2184 }
2185 }
2186
2187 // The compile time filters did not reject base_oop/offset so
2188 // we need to generate the following runtime filters
2189 //
2190 // if (offset == java_lang_ref_Reference::_reference_offset) {
2191 // if (instance_of(base, java.lang.ref.Reference)) {
2192 // pre_barrier(_, pre_val, ...);
2193 // }
2194 // }
2195
2196 float likely = PROB_LIKELY( 0.999);
2197 float unlikely = PROB_UNLIKELY(0.999);
2198
2199 IdealKit ideal(this);
2200 #define __ ideal.
2201
2202 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2203
2204 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2205 // Update graphKit memory and control from IdealKit.
2206 sync_kit(ideal);
2207
2208 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2209 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2210
2211 // Update IdealKit memory and control from graphKit.
2212 __ sync_kit(this);
2213
2214 Node* one = __ ConI(1);
2215 // is_instof == 0 if base_oop == NULL
2216 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2217
2218 // Update graphKit from IdeakKit.
2219 sync_kit(ideal);
2220
2221 // Use the pre-barrier to record the value in the referent field
2222 pre_barrier(false /* do_load */,
2223 __ ctrl(),
2224 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2225 pre_val /* pre_val */,
2226 T_OBJECT);
2227 if (need_mem_bar) {
2228 // Add memory barrier to prevent commoning reads from this field
2229 // across safepoint since GC can change its value.
2230 insert_mem_bar(Op_MemBarCPUOrder);
2231 }
2232 // Update IdealKit from graphKit.
2233 __ sync_kit(this);
2234
2235 } __ end_if(); // _ref_type != ref_none
2236 } __ end_if(); // offset == referent_offset
2237
2238 // Final sync IdealKit and GraphKit.
2239 final_sync(ideal);
2240 #undef __
2241 }
2242
2243
2244 // Interpret Unsafe.fieldOffset cookies correctly:
2245 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2246
2247 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2248 // Attempt to infer a sharper value type from the offset and base type.
2249 ciKlass* sharpened_klass = NULL;
2250
2251 // See if it is an instance field, with an object type.
2252 if (alias_type->field() != NULL) {
2253 assert(!is_native_ptr, "native pointer op cannot use a java address");
2254 if (alias_type->field()->type()->is_klass()) {
2255 sharpened_klass = alias_type->field()->type()->as_klass();
2256 }
2257 }
2258
2259 // See if it is a narrow oop array.
2260 if (adr_type->isa_aryptr()) {
2261 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2262 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2263 if (elem_type != NULL) {
2264 sharpened_klass = elem_type->klass();
2265 }
2266 }
2267 }
2268
2269 // The sharpened class might be unloaded if there is no class loader
2270 // contraint in place.
2271 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2272 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2273
2274 #ifndef PRODUCT
2275 if (C->print_intrinsics() || C->print_inlining()) {
2276 tty->print(" from base type: "); adr_type->dump();
2277 tty->print(" sharpened value: "); tjp->dump();
2278 }
2279 #endif
2280 // Sharpen the value type.
2281 return tjp;
2282 }
2283 return NULL;
2284 }
2285
2286 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2287 if (callee()->is_static()) return false; // caller must have the capability!
2288
2289 #ifndef PRODUCT
2290 {
2291 ResourceMark rm;
2292 // Check the signatures.
2293 ciSignature* sig = callee()->signature();
2294 #ifdef ASSERT
2295 if (!is_store) {
2296 // Object getObject(Object base, int/long offset), etc.
2297 BasicType rtype = sig->return_type()->basic_type();
2298 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2299 rtype = T_ADDRESS; // it is really a C void*
2300 assert(rtype == type, "getter must return the expected value");
2301 if (!is_native_ptr) {
2302 assert(sig->count() == 2, "oop getter has 2 arguments");
2303 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2304 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2305 } else {
2306 assert(sig->count() == 1, "native getter has 1 argument");
2307 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2308 }
2309 } else {
2310 // void putObject(Object base, int/long offset, Object x), etc.
2311 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2312 if (!is_native_ptr) {
2313 assert(sig->count() == 3, "oop putter has 3 arguments");
2314 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2315 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2316 } else {
2317 assert(sig->count() == 2, "native putter has 2 arguments");
2318 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2319 }
2320 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2321 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2322 vtype = T_ADDRESS; // it is really a C void*
2323 assert(vtype == type, "putter must accept the expected value");
2324 }
2325 #endif // ASSERT
2326 }
2327 #endif //PRODUCT
2328
2329 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2330
2331 Node* receiver = argument(0); // type: oop
2332
2333 // Build address expression.
2334 Node* adr;
2335 Node* heap_base_oop = top();
2336 Node* offset = top();
2337 Node* val;
2338
2339 if (!is_native_ptr) {
2340 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2341 Node* base = argument(1); // type: oop
2342 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2343 offset = argument(2); // type: long
2344 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2345 // to be plain byte offsets, which are also the same as those accepted
2346 // by oopDesc::field_base.
2347 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2348 "fieldOffset must be byte-scaled");
2349 // 32-bit machines ignore the high half!
2350 offset = ConvL2X(offset);
2351 adr = make_unsafe_address(base, offset);
2352 heap_base_oop = base;
2353 val = is_store ? argument(4) : NULL;
2354 } else {
2355 Node* ptr = argument(1); // type: long
2356 ptr = ConvL2X(ptr); // adjust Java long to machine word
2357 adr = make_unsafe_address(NULL, ptr);
2358 val = is_store ? argument(3) : NULL;
2359 }
2360
2361 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2362
2363 // First guess at the value type.
2364 const Type *value_type = Type::get_const_basic_type(type);
2365
2366 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2367 // there was not enough information to nail it down.
2368 Compile::AliasType* alias_type = C->alias_type(adr_type);
2369 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2370
2371 // We will need memory barriers unless we can determine a unique
2372 // alias category for this reference. (Note: If for some reason
2373 // the barriers get omitted and the unsafe reference begins to "pollute"
2374 // the alias analysis of the rest of the graph, either Compile::can_alias
2375 // or Compile::must_alias will throw a diagnostic assert.)
2376 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2377
2378 // If we are reading the value of the referent field of a Reference
2379 // object (either by using Unsafe directly or through reflection)
2380 // then, if G1 is enabled, we need to record the referent in an
2381 // SATB log buffer using the pre-barrier mechanism.
2382 // Also we need to add memory barrier to prevent commoning reads
2383 // from this field across safepoint since GC can change its value.
2384 bool need_read_barrier = !is_native_ptr && !is_store &&
2385 offset != top() && heap_base_oop != top();
2386
2387 if (!is_store && type == T_OBJECT) {
2388 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2389 if (tjp != NULL) {
2390 value_type = tjp;
2391 }
2392 }
2393
2394 receiver = null_check(receiver);
2395 if (stopped()) {
2396 return true;
2397 }
2398 // Heap pointers get a null-check from the interpreter,
2399 // as a courtesy. However, this is not guaranteed by Unsafe,
2400 // and it is not possible to fully distinguish unintended nulls
2401 // from intended ones in this API.
2402
2403 if (is_volatile) {
2404 // We need to emit leading and trailing CPU membars (see below) in
2405 // addition to memory membars when is_volatile. This is a little
2406 // too strong, but avoids the need to insert per-alias-type
2407 // volatile membars (for stores; compare Parse::do_put_xxx), which
2408 // we cannot do effectively here because we probably only have a
2409 // rough approximation of type.
2410 need_mem_bar = true;
2411 // For Stores, place a memory ordering barrier now.
2412 if (is_store) {
2413 insert_mem_bar(Op_MemBarRelease);
2414 } else {
2415 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2416 insert_mem_bar(Op_MemBarVolatile);
2417 }
2418 }
2419 }
2420
2421 // Memory barrier to prevent normal and 'unsafe' accesses from
2422 // bypassing each other. Happens after null checks, so the
2423 // exception paths do not take memory state from the memory barrier,
2424 // so there's no problems making a strong assert about mixing users
2425 // of safe & unsafe memory.
2426 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2427
2428 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM ||
2429 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown");
2430 bool mismatched = false;
2431 if (alias_type->element() != NULL || alias_type->field() != NULL) {
2432 BasicType bt;
2433 if (alias_type->element() != NULL) {
2434 const Type* element = alias_type->element();
2435 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
2436 } else {
2437 bt = alias_type->field()->type()->basic_type();
2438 }
2439 if (bt == T_ARRAY) {
2440 // accessing an array field with getObject is not a mismatch
2441 bt = T_OBJECT;
2442 }
2443 if (bt != type) {
2444 mismatched = true;
2445 }
2446 }
2447 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2448
2449 if (!is_store) {
2450 Node* p = NULL;
2451 // Try to constant fold a load from a constant field
2452 ciField* field = alias_type->field();
2453 if (heap_base_oop != top() &&
2454 field != NULL && field->is_constant() && field->layout_type() == type) {
2455 // final or stable field
2456 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2457 if (con_type != NULL) {
2458 p = makecon(con_type);
2459 }
2460 }
2461 if (p == NULL) {
2462 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2463 // To be valid, unsafe loads may depend on other conditions than
2464 // the one that guards them: pin the Load node
2465 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2466 // load value
2467 switch (type) {
2468 case T_BOOLEAN:
2469 case T_CHAR:
2470 case T_BYTE:
2471 case T_SHORT:
2472 case T_INT:
2473 case T_LONG:
2474 case T_FLOAT:
2475 case T_DOUBLE:
2476 break;
2477 case T_OBJECT:
2478 if (need_read_barrier) {
2479 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2480 }
2481 break;
2482 case T_ADDRESS:
2483 // Cast to an int type.
2484 p = _gvn.transform(new CastP2XNode(NULL, p));
2485 p = ConvX2UL(p);
2486 break;
2487 default:
2488 fatal("unexpected type %d: %s", type, type2name(type));
2489 break;
2490 }
2491 }
2492 // The load node has the control of the preceding MemBarCPUOrder. All
2493 // following nodes will have the control of the MemBarCPUOrder inserted at
2494 // the end of this method. So, pushing the load onto the stack at a later
2495 // point is fine.
2496 set_result(p);
2497 } else {
2498 // place effect of store into memory
2499 switch (type) {
2500 case T_DOUBLE:
2501 val = dstore_rounding(val);
2502 break;
2503 case T_ADDRESS:
2504 // Repackage the long as a pointer.
2505 val = ConvL2X(val);
2506 val = _gvn.transform(new CastX2PNode(val));
2507 break;
2508 }
2509
2510 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2511 if (type != T_OBJECT ) {
2512 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2513 } else {
2514 // Possibly an oop being stored to Java heap or native memory
2515 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2516 // oop to Java heap.
2517 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2518 } else {
2519 // We can't tell at compile time if we are storing in the Java heap or outside
2520 // of it. So we need to emit code to conditionally do the proper type of
2521 // store.
2522
2523 IdealKit ideal(this);
2524 #define __ ideal.
2525 // QQQ who knows what probability is here??
2526 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2527 // Sync IdealKit and graphKit.
2528 sync_kit(ideal);
2529 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2530 // Update IdealKit memory.
2531 __ sync_kit(this);
2532 } __ else_(); {
2533 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched);
2534 } __ end_if();
2535 // Final sync IdealKit and GraphKit.
2536 final_sync(ideal);
2537 #undef __
2538 }
2539 }
2540 }
2541
2542 if (is_volatile) {
2543 if (!is_store) {
2544 insert_mem_bar(Op_MemBarAcquire);
2545 } else {
2546 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2547 insert_mem_bar(Op_MemBarVolatile);
2548 }
2549 }
2550 }
2551
2552 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2553
2554 return true;
2555 }
2556
2557 //----------------------------inline_unsafe_load_store----------------------------
2558 // This method serves a couple of different customers (depending on LoadStoreKind):
2559 //
2560 // LS_cmpxchg:
2561 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2562 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2563 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2564 //
2565 // LS_xadd:
2566 // public int getAndAddInt( Object o, long offset, int delta)
2567 // public long getAndAddLong(Object o, long offset, long delta)
2568 //
2569 // LS_xchg:
2570 // int getAndSet(Object o, long offset, int newValue)
2571 // long getAndSet(Object o, long offset, long newValue)
2572 // Object getAndSet(Object o, long offset, Object newValue)
2573 //
2574 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2575 // This basic scheme here is the same as inline_unsafe_access, but
2576 // differs in enough details that combining them would make the code
2577 // overly confusing. (This is a true fact! I originally combined
2578 // them, but even I was confused by it!) As much code/comments as
2579 // possible are retained from inline_unsafe_access though to make
2580 // the correspondences clearer. - dl
2581
2582 if (callee()->is_static()) return false; // caller must have the capability!
2583
2584 #ifndef PRODUCT
2585 BasicType rtype;
2586 {
2587 ResourceMark rm;
2588 // Check the signatures.
2589 ciSignature* sig = callee()->signature();
2590 rtype = sig->return_type()->basic_type();
2591 if (kind == LS_xadd || kind == LS_xchg) {
2592 // Check the signatures.
2593 #ifdef ASSERT
2594 assert(rtype == type, "get and set must return the expected type");
2595 assert(sig->count() == 3, "get and set has 3 arguments");
2596 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2597 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2598 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2599 #endif // ASSERT
2600 } else if (kind == LS_cmpxchg) {
2601 // Check the signatures.
2602 #ifdef ASSERT
2603 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2604 assert(sig->count() == 4, "CAS has 4 arguments");
2605 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2606 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2607 #endif // ASSERT
2608 } else {
2609 ShouldNotReachHere();
2610 }
2611 }
2612 #endif //PRODUCT
2613
2614 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2615
2616 // Get arguments:
2617 Node* receiver = NULL;
2618 Node* base = NULL;
2619 Node* offset = NULL;
2620 Node* oldval = NULL;
2621 Node* newval = NULL;
2622 if (kind == LS_cmpxchg) {
2623 const bool two_slot_type = type2size[type] == 2;
2624 receiver = argument(0); // type: oop
2625 base = argument(1); // type: oop
2626 offset = argument(2); // type: long
2627 oldval = argument(4); // type: oop, int, or long
2628 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2629 } else if (kind == LS_xadd || kind == LS_xchg){
2630 receiver = argument(0); // type: oop
2631 base = argument(1); // type: oop
2632 offset = argument(2); // type: long
2633 oldval = NULL;
2634 newval = argument(4); // type: oop, int, or long
2635 }
2636
2637 // Null check receiver.
2638 receiver = null_check(receiver);
2639 if (stopped()) {
2640 return true;
2641 }
2642
2643 // Build field offset expression.
2644 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2645 // to be plain byte offsets, which are also the same as those accepted
2646 // by oopDesc::field_base.
2647 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2648 // 32-bit machines ignore the high half of long offsets
2649 offset = ConvL2X(offset);
2650 Node* adr = make_unsafe_address(base, offset);
2651 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2652
2653 // For CAS, unlike inline_unsafe_access, there seems no point in
2654 // trying to refine types. Just use the coarse types here.
2655 const Type *value_type = Type::get_const_basic_type(type);
2656 Compile::AliasType* alias_type = C->alias_type(adr_type);
2657 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2658
2659 if (kind == LS_xchg && type == T_OBJECT) {
2660 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2661 if (tjp != NULL) {
2662 value_type = tjp;
2663 }
2664 }
2665
2666 int alias_idx = C->get_alias_index(adr_type);
2667
2668 // Memory-model-wise, a LoadStore acts like a little synchronized
2669 // block, so needs barriers on each side. These don't translate
2670 // into actual barriers on most machines, but we still need rest of
2671 // compiler to respect ordering.
2672
2673 insert_mem_bar(Op_MemBarRelease);
2674 insert_mem_bar(Op_MemBarCPUOrder);
2675
2676 // 4984716: MemBars must be inserted before this
2677 // memory node in order to avoid a false
2678 // dependency which will confuse the scheduler.
2679 Node *mem = memory(alias_idx);
2680
2681 // For now, we handle only those cases that actually exist: ints,
2682 // longs, and Object. Adding others should be straightforward.
2683 Node* load_store = NULL;
2684 switch(type) {
2685 case T_INT:
2686 if (kind == LS_xadd) {
2687 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2688 } else if (kind == LS_xchg) {
2689 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2690 } else if (kind == LS_cmpxchg) {
2691 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2692 } else {
2693 ShouldNotReachHere();
2694 }
2695 break;
2696 case T_LONG:
2697 if (kind == LS_xadd) {
2698 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2699 } else if (kind == LS_xchg) {
2700 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2701 } else if (kind == LS_cmpxchg) {
2702 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2703 } else {
2704 ShouldNotReachHere();
2705 }
2706 break;
2707 case T_OBJECT:
2708 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2709 // could be delayed during Parse (for example, in adjust_map_after_if()).
2710 // Execute transformation here to avoid barrier generation in such case.
2711 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2712 newval = _gvn.makecon(TypePtr::NULL_PTR);
2713
2714 // Reference stores need a store barrier.
2715 if (kind == LS_xchg) {
2716 // If pre-barrier must execute before the oop store, old value will require do_load here.
2717 if (!can_move_pre_barrier()) {
2718 pre_barrier(true /* do_load*/,
2719 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2720 NULL /* pre_val*/,
2721 T_OBJECT);
2722 } // Else move pre_barrier to use load_store value, see below.
2723 } else if (kind == LS_cmpxchg) {
2724 // Same as for newval above:
2725 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2726 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2727 }
2728 // The only known value which might get overwritten is oldval.
2729 pre_barrier(false /* do_load */,
2730 control(), NULL, NULL, max_juint, NULL, NULL,
2731 oldval /* pre_val */,
2732 T_OBJECT);
2733 } else {
2734 ShouldNotReachHere();
2735 }
2736
2737 #ifdef _LP64
2738 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2739 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2740 if (kind == LS_xchg) {
2741 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2742 newval_enc, adr_type, value_type->make_narrowoop()));
2743 } else {
2744 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2745 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2746 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2747 newval_enc, oldval_enc));
2748 }
2749 } else
2750 #endif
2751 {
2752 if (kind == LS_xchg) {
2753 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2754 } else {
2755 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2756 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2757 }
2758 }
2759 if (kind == LS_cmpxchg) {
2760 // Emit the post barrier only when the actual store happened.
2761 // This makes sense to check only for compareAndSet that can fail to set the value.
2762 // CAS success path is marked more likely since we anticipate this is a performance
2763 // critical path, while CAS failure path can use the penalty for going through unlikely
2764 // path as backoff. Which is still better than doing a store barrier there.
2765 IdealKit ideal(this);
2766 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2767 sync_kit(ideal);
2768 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2769 ideal.sync_kit(this);
2770 } ideal.end_if();
2771 final_sync(ideal);
2772 } else {
2773 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2774 }
2775 break;
2776 default:
2777 fatal("unexpected type %d: %s", type, type2name(type));
2778 break;
2779 }
2780
2781 // SCMemProjNodes represent the memory state of a LoadStore. Their
2782 // main role is to prevent LoadStore nodes from being optimized away
2783 // when their results aren't used.
2784 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2785 set_memory(proj, alias_idx);
2786
2787 if (type == T_OBJECT && kind == LS_xchg) {
2788 #ifdef _LP64
2789 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2790 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2791 }
2792 #endif
2793 if (can_move_pre_barrier()) {
2794 // Don't need to load pre_val. The old value is returned by load_store.
2795 // The pre_barrier can execute after the xchg as long as no safepoint
2796 // gets inserted between them.
2797 pre_barrier(false /* do_load */,
2798 control(), NULL, NULL, max_juint, NULL, NULL,
2799 load_store /* pre_val */,
2800 T_OBJECT);
2801 }
2802 }
2803
2804 // Add the trailing membar surrounding the access
2805 insert_mem_bar(Op_MemBarCPUOrder);
2806 insert_mem_bar(Op_MemBarAcquire);
2807
2808 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2809 set_result(load_store);
2810 return true;
2811 }
2812
2813 //----------------------------inline_unsafe_ordered_store----------------------
2814 // public native void Unsafe.putOrderedObject(Object o, long offset, Object x);
2815 // public native void Unsafe.putOrderedInt(Object o, long offset, int x);
2816 // public native void Unsafe.putOrderedLong(Object o, long offset, long x);
2817 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2818 // This is another variant of inline_unsafe_access, differing in
2819 // that it always issues store-store ("release") barrier and ensures
2820 // store-atomicity (which only matters for "long").
2821
2822 if (callee()->is_static()) return false; // caller must have the capability!
2823
2824 #ifndef PRODUCT
2825 {
2826 ResourceMark rm;
2827 // Check the signatures.
2828 ciSignature* sig = callee()->signature();
2829 #ifdef ASSERT
2830 BasicType rtype = sig->return_type()->basic_type();
2831 assert(rtype == T_VOID, "must return void");
2832 assert(sig->count() == 3, "has 3 arguments");
2833 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2834 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2835 #endif // ASSERT
2836 }
2837 #endif //PRODUCT
2838
2839 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2840
2841 // Get arguments:
2842 Node* receiver = argument(0); // type: oop
2843 Node* base = argument(1); // type: oop
2844 Node* offset = argument(2); // type: long
2845 Node* val = argument(4); // type: oop, int, or long
2846
2847 // Null check receiver.
2848 receiver = null_check(receiver);
2849 if (stopped()) {
2850 return true;
2851 }
2852
2853 // Build field offset expression.
2854 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2855 // 32-bit machines ignore the high half of long offsets
2856 offset = ConvL2X(offset);
2857 Node* adr = make_unsafe_address(base, offset);
2858 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2859 const Type *value_type = Type::get_const_basic_type(type);
2860 Compile::AliasType* alias_type = C->alias_type(adr_type);
2861
2862 insert_mem_bar(Op_MemBarRelease);
2863 insert_mem_bar(Op_MemBarCPUOrder);
2864 // Ensure that the store is atomic for longs:
2865 const bool require_atomic_access = true;
2866 Node* store;
2867 if (type == T_OBJECT) // reference stores need a store barrier.
2868 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2869 else {
2870 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2871 }
2872 insert_mem_bar(Op_MemBarCPUOrder);
2873 return true;
2874 }
2875
2876 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2877 // Regardless of form, don't allow previous ld/st to move down,
2878 // then issue acquire, release, or volatile mem_bar.
2879 insert_mem_bar(Op_MemBarCPUOrder);
2880 switch(id) {
2881 case vmIntrinsics::_loadFence:
2882 insert_mem_bar(Op_LoadFence);
2883 return true;
2884 case vmIntrinsics::_storeFence:
2885 insert_mem_bar(Op_StoreFence);
2886 return true;
2887 case vmIntrinsics::_fullFence:
2888 insert_mem_bar(Op_MemBarVolatile);
2889 return true;
2890 default:
2891 fatal_unexpected_iid(id);
2892 return false;
2893 }
2894 }
2895
2896 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2897 if (!kls->is_Con()) {
2898 return true;
2899 }
2900 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2901 if (klsptr == NULL) {
2902 return true;
2903 }
2904 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2905 // don't need a guard for a klass that is already initialized
2906 return !ik->is_initialized();
2907 }
2908
2909 //----------------------------inline_unsafe_allocate---------------------------
2910 // public native Object Unsafe.allocateInstance(Class<?> cls);
2911 bool LibraryCallKit::inline_unsafe_allocate() {
2912 if (callee()->is_static()) return false; // caller must have the capability!
2913
2914 null_check_receiver(); // null-check, then ignore
2915 Node* cls = null_check(argument(1));
2916 if (stopped()) return true;
2917
2918 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2919 kls = null_check(kls);
2920 if (stopped()) return true; // argument was like int.class
2921
2922 Node* test = NULL;
2923 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2924 // Note: The argument might still be an illegal value like
2925 // Serializable.class or Object[].class. The runtime will handle it.
2926 // But we must make an explicit check for initialization.
2927 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2928 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2929 // can generate code to load it as unsigned byte.
2930 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2931 Node* bits = intcon(InstanceKlass::fully_initialized);
2932 test = _gvn.transform(new SubINode(inst, bits));
2933 // The 'test' is non-zero if we need to take a slow path.
2934 }
2935
2936 Node* obj = new_instance(kls, test);
2937 set_result(obj);
2938 return true;
2939 }
2940
2941 #ifdef TRACE_HAVE_INTRINSICS
2942 /*
2943 * oop -> myklass
2944 * myklass->trace_id |= USED
2945 * return myklass->trace_id & ~0x3
2946 */
2947 bool LibraryCallKit::inline_native_classID() {
2948 null_check_receiver(); // null-check, then ignore
2949 Node* cls = null_check(argument(1), T_OBJECT);
2950 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2951 kls = null_check(kls, T_OBJECT);
2952 ByteSize offset = TRACE_ID_OFFSET;
2953 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2954 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2955 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2956 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2957 Node* clsused = longcon(0x01l); // set the class bit
2958 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2959
2960 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2961 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2962 set_result(andl);
2963 return true;
2964 }
2965
2966 bool LibraryCallKit::inline_native_threadID() {
2967 Node* tls_ptr = NULL;
2968 Node* cur_thr = generate_current_thread(tls_ptr);
2969 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2970 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2971 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2972
2973 Node* threadid = NULL;
2974 size_t thread_id_size = OSThread::thread_id_size();
2975 if (thread_id_size == (size_t) BytesPerLong) {
2976 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2977 } else if (thread_id_size == (size_t) BytesPerInt) {
2978 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2979 } else {
2980 ShouldNotReachHere();
2981 }
2982 set_result(threadid);
2983 return true;
2984 }
2985 #endif
2986
2987 //------------------------inline_native_time_funcs--------------
2988 // inline code for System.currentTimeMillis() and System.nanoTime()
2989 // these have the same type and signature
2990 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2991 const TypeFunc* tf = OptoRuntime::void_long_Type();
2992 const TypePtr* no_memory_effects = NULL;
2993 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2994 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2995 #ifdef ASSERT
2996 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2997 assert(value_top == top(), "second value must be top");
2998 #endif
2999 set_result(value);
3000 return true;
3001 }
3002
3003 //------------------------inline_native_currentThread------------------
3004 bool LibraryCallKit::inline_native_currentThread() {
3005 Node* junk = NULL;
3006 set_result(generate_current_thread(junk));
3007 return true;
3008 }
3009
3010 //------------------------inline_native_isInterrupted------------------
3011 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3012 bool LibraryCallKit::inline_native_isInterrupted() {
3013 // Add a fast path to t.isInterrupted(clear_int):
3014 // (t == Thread.current() &&
3015 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3016 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3017 // So, in the common case that the interrupt bit is false,
3018 // we avoid making a call into the VM. Even if the interrupt bit
3019 // is true, if the clear_int argument is false, we avoid the VM call.
3020 // However, if the receiver is not currentThread, we must call the VM,
3021 // because there must be some locking done around the operation.
3022
3023 // We only go to the fast case code if we pass two guards.
3024 // Paths which do not pass are accumulated in the slow_region.
3025
3026 enum {
3027 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3028 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3029 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3030 PATH_LIMIT
3031 };
3032
3033 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3034 // out of the function.
3035 insert_mem_bar(Op_MemBarCPUOrder);
3036
3037 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3038 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3039
3040 RegionNode* slow_region = new RegionNode(1);
3041 record_for_igvn(slow_region);
3042
3043 // (a) Receiving thread must be the current thread.
3044 Node* rec_thr = argument(0);
3045 Node* tls_ptr = NULL;
3046 Node* cur_thr = generate_current_thread(tls_ptr);
3047 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3048 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3049
3050 generate_slow_guard(bol_thr, slow_region);
3051
3052 // (b) Interrupt bit on TLS must be false.
3053 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3054 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3055 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3056
3057 // Set the control input on the field _interrupted read to prevent it floating up.
3058 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3059 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3060 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3061
3062 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3063
3064 // First fast path: if (!TLS._interrupted) return false;
3065 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3066 result_rgn->init_req(no_int_result_path, false_bit);
3067 result_val->init_req(no_int_result_path, intcon(0));
3068
3069 // drop through to next case
3070 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3071
3072 #ifndef TARGET_OS_FAMILY_windows
3073 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3074 Node* clr_arg = argument(1);
3075 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3076 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3077 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3078
3079 // Second fast path: ... else if (!clear_int) return true;
3080 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3081 result_rgn->init_req(no_clear_result_path, false_arg);
3082 result_val->init_req(no_clear_result_path, intcon(1));
3083
3084 // drop through to next case
3085 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3086 #else
3087 // To return true on Windows you must read the _interrupted field
3088 // and check the the event state i.e. take the slow path.
3089 #endif // TARGET_OS_FAMILY_windows
3090
3091 // (d) Otherwise, go to the slow path.
3092 slow_region->add_req(control());
3093 set_control( _gvn.transform(slow_region));
3094
3095 if (stopped()) {
3096 // There is no slow path.
3097 result_rgn->init_req(slow_result_path, top());
3098 result_val->init_req(slow_result_path, top());
3099 } else {
3100 // non-virtual because it is a private non-static
3101 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3102
3103 Node* slow_val = set_results_for_java_call(slow_call);
3104 // this->control() comes from set_results_for_java_call
3105
3106 Node* fast_io = slow_call->in(TypeFunc::I_O);
3107 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3108
3109 // These two phis are pre-filled with copies of of the fast IO and Memory
3110 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3111 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3112
3113 result_rgn->init_req(slow_result_path, control());
3114 result_io ->init_req(slow_result_path, i_o());
3115 result_mem->init_req(slow_result_path, reset_memory());
3116 result_val->init_req(slow_result_path, slow_val);
3117
3118 set_all_memory(_gvn.transform(result_mem));
3119 set_i_o( _gvn.transform(result_io));
3120 }
3121
3122 C->set_has_split_ifs(true); // Has chance for split-if optimization
3123 set_result(result_rgn, result_val);
3124 return true;
3125 }
3126
3127 //---------------------------load_mirror_from_klass----------------------------
3128 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3129 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3130 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3131 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3132 }
3133
3134 //-----------------------load_klass_from_mirror_common-------------------------
3135 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3136 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3137 // and branch to the given path on the region.
3138 // If never_see_null, take an uncommon trap on null, so we can optimistically
3139 // compile for the non-null case.
3140 // If the region is NULL, force never_see_null = true.
3141 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3142 bool never_see_null,
3143 RegionNode* region,
3144 int null_path,
3145 int offset) {
3146 if (region == NULL) never_see_null = true;
3147 Node* p = basic_plus_adr(mirror, offset);
3148 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3149 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3150 Node* null_ctl = top();
3151 kls = null_check_oop(kls, &null_ctl, never_see_null);
3152 if (region != NULL) {
3153 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3154 region->init_req(null_path, null_ctl);
3155 } else {
3156 assert(null_ctl == top(), "no loose ends");
3157 }
3158 return kls;
3159 }
3160
3161 //--------------------(inline_native_Class_query helpers)---------------------
3162 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3163 // Fall through if (mods & mask) == bits, take the guard otherwise.
3164 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3165 // Branch around if the given klass has the given modifier bit set.
3166 // Like generate_guard, adds a new path onto the region.
3167 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3168 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3169 Node* mask = intcon(modifier_mask);
3170 Node* bits = intcon(modifier_bits);
3171 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3172 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3173 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3174 return generate_fair_guard(bol, region);
3175 }
3176 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3177 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3178 }
3179
3180 //-------------------------inline_native_Class_query-------------------
3181 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3182 const Type* return_type = TypeInt::BOOL;
3183 Node* prim_return_value = top(); // what happens if it's a primitive class?
3184 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3185 bool expect_prim = false; // most of these guys expect to work on refs
3186
3187 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3188
3189 Node* mirror = argument(0);
3190 Node* obj = top();
3191
3192 switch (id) {
3193 case vmIntrinsics::_isInstance:
3194 // nothing is an instance of a primitive type
3195 prim_return_value = intcon(0);
3196 obj = argument(1);
3197 break;
3198 case vmIntrinsics::_getModifiers:
3199 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3200 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3201 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3202 break;
3203 case vmIntrinsics::_isInterface:
3204 prim_return_value = intcon(0);
3205 break;
3206 case vmIntrinsics::_isArray:
3207 prim_return_value = intcon(0);
3208 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3209 break;
3210 case vmIntrinsics::_isPrimitive:
3211 prim_return_value = intcon(1);
3212 expect_prim = true; // obviously
3213 break;
3214 case vmIntrinsics::_getSuperclass:
3215 prim_return_value = null();
3216 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3217 break;
3218 case vmIntrinsics::_getClassAccessFlags:
3219 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3220 return_type = TypeInt::INT; // not bool! 6297094
3221 break;
3222 default:
3223 fatal_unexpected_iid(id);
3224 break;
3225 }
3226
3227 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3228 if (mirror_con == NULL) return false; // cannot happen?
3229
3230 #ifndef PRODUCT
3231 if (C->print_intrinsics() || C->print_inlining()) {
3232 ciType* k = mirror_con->java_mirror_type();
3233 if (k) {
3234 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3235 k->print_name();
3236 tty->cr();
3237 }
3238 }
3239 #endif
3240
3241 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3242 RegionNode* region = new RegionNode(PATH_LIMIT);
3243 record_for_igvn(region);
3244 PhiNode* phi = new PhiNode(region, return_type);
3245
3246 // The mirror will never be null of Reflection.getClassAccessFlags, however
3247 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3248 // if it is. See bug 4774291.
3249
3250 // For Reflection.getClassAccessFlags(), the null check occurs in
3251 // the wrong place; see inline_unsafe_access(), above, for a similar
3252 // situation.
3253 mirror = null_check(mirror);
3254 // If mirror or obj is dead, only null-path is taken.
3255 if (stopped()) return true;
3256
3257 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3258
3259 // Now load the mirror's klass metaobject, and null-check it.
3260 // Side-effects region with the control path if the klass is null.
3261 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3262 // If kls is null, we have a primitive mirror.
3263 phi->init_req(_prim_path, prim_return_value);
3264 if (stopped()) { set_result(region, phi); return true; }
3265 bool safe_for_replace = (region->in(_prim_path) == top());
3266
3267 Node* p; // handy temp
3268 Node* null_ctl;
3269
3270 // Now that we have the non-null klass, we can perform the real query.
3271 // For constant classes, the query will constant-fold in LoadNode::Value.
3272 Node* query_value = top();
3273 switch (id) {
3274 case vmIntrinsics::_isInstance:
3275 // nothing is an instance of a primitive type
3276 query_value = gen_instanceof(obj, kls, safe_for_replace);
3277 break;
3278
3279 case vmIntrinsics::_getModifiers:
3280 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3281 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3282 break;
3283
3284 case vmIntrinsics::_isInterface:
3285 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3286 if (generate_interface_guard(kls, region) != NULL)
3287 // A guard was added. If the guard is taken, it was an interface.
3288 phi->add_req(intcon(1));
3289 // If we fall through, it's a plain class.
3290 query_value = intcon(0);
3291 break;
3292
3293 case vmIntrinsics::_isArray:
3294 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3295 if (generate_array_guard(kls, region) != NULL)
3296 // A guard was added. If the guard is taken, it was an array.
3297 phi->add_req(intcon(1));
3298 // If we fall through, it's a plain class.
3299 query_value = intcon(0);
3300 break;
3301
3302 case vmIntrinsics::_isPrimitive:
3303 query_value = intcon(0); // "normal" path produces false
3304 break;
3305
3306 case vmIntrinsics::_getSuperclass:
3307 // The rules here are somewhat unfortunate, but we can still do better
3308 // with random logic than with a JNI call.
3309 // Interfaces store null or Object as _super, but must report null.
3310 // Arrays store an intermediate super as _super, but must report Object.
3311 // Other types can report the actual _super.
3312 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3313 if (generate_interface_guard(kls, region) != NULL)
3314 // A guard was added. If the guard is taken, it was an interface.
3315 phi->add_req(null());
3316 if (generate_array_guard(kls, region) != NULL)
3317 // A guard was added. If the guard is taken, it was an array.
3318 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3319 // If we fall through, it's a plain class. Get its _super.
3320 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3321 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3322 null_ctl = top();
3323 kls = null_check_oop(kls, &null_ctl);
3324 if (null_ctl != top()) {
3325 // If the guard is taken, Object.superClass is null (both klass and mirror).
3326 region->add_req(null_ctl);
3327 phi ->add_req(null());
3328 }
3329 if (!stopped()) {
3330 query_value = load_mirror_from_klass(kls);
3331 }
3332 break;
3333
3334 case vmIntrinsics::_getClassAccessFlags:
3335 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3336 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3337 break;
3338
3339 default:
3340 fatal_unexpected_iid(id);
3341 break;
3342 }
3343
3344 // Fall-through is the normal case of a query to a real class.
3345 phi->init_req(1, query_value);
3346 region->init_req(1, control());
3347
3348 C->set_has_split_ifs(true); // Has chance for split-if optimization
3349 set_result(region, phi);
3350 return true;
3351 }
3352
3353 //-------------------------inline_Class_cast-------------------
3354 bool LibraryCallKit::inline_Class_cast() {
3355 Node* mirror = argument(0); // Class
3356 Node* obj = argument(1);
3357 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3358 if (mirror_con == NULL) {
3359 return false; // dead path (mirror->is_top()).
3360 }
3361 if (obj == NULL || obj->is_top()) {
3362 return false; // dead path
3363 }
3364 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3365
3366 // First, see if Class.cast() can be folded statically.
3367 // java_mirror_type() returns non-null for compile-time Class constants.
3368 ciType* tm = mirror_con->java_mirror_type();
3369 if (tm != NULL && tm->is_klass() &&
3370 tp != NULL && tp->klass() != NULL) {
3371 if (!tp->klass()->is_loaded()) {
3372 // Don't use intrinsic when class is not loaded.
3373 return false;
3374 } else {
3375 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3376 if (static_res == Compile::SSC_always_true) {
3377 // isInstance() is true - fold the code.
3378 set_result(obj);
3379 return true;
3380 } else if (static_res == Compile::SSC_always_false) {
3381 // Don't use intrinsic, have to throw ClassCastException.
3382 // If the reference is null, the non-intrinsic bytecode will
3383 // be optimized appropriately.
3384 return false;
3385 }
3386 }
3387 }
3388
3389 // Bailout intrinsic and do normal inlining if exception path is frequent.
3390 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3391 return false;
3392 }
3393
3394 // Generate dynamic checks.
3395 // Class.cast() is java implementation of _checkcast bytecode.
3396 // Do checkcast (Parse::do_checkcast()) optimizations here.
3397
3398 mirror = null_check(mirror);
3399 // If mirror is dead, only null-path is taken.
3400 if (stopped()) {
3401 return true;
3402 }
3403
3404 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3405 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3406 RegionNode* region = new RegionNode(PATH_LIMIT);
3407 record_for_igvn(region);
3408
3409 // Now load the mirror's klass metaobject, and null-check it.
3410 // If kls is null, we have a primitive mirror and
3411 // nothing is an instance of a primitive type.
3412 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3413
3414 Node* res = top();
3415 if (!stopped()) {
3416 Node* bad_type_ctrl = top();
3417 // Do checkcast optimizations.
3418 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3419 region->init_req(_bad_type_path, bad_type_ctrl);
3420 }
3421 if (region->in(_prim_path) != top() ||
3422 region->in(_bad_type_path) != top()) {
3423 // Let Interpreter throw ClassCastException.
3424 PreserveJVMState pjvms(this);
3425 set_control(_gvn.transform(region));
3426 uncommon_trap(Deoptimization::Reason_intrinsic,
3427 Deoptimization::Action_maybe_recompile);
3428 }
3429 if (!stopped()) {
3430 set_result(res);
3431 }
3432 return true;
3433 }
3434
3435
3436 //--------------------------inline_native_subtype_check------------------------
3437 // This intrinsic takes the JNI calls out of the heart of
3438 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3439 bool LibraryCallKit::inline_native_subtype_check() {
3440 // Pull both arguments off the stack.
3441 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3442 args[0] = argument(0);
3443 args[1] = argument(1);
3444 Node* klasses[2]; // corresponding Klasses: superk, subk
3445 klasses[0] = klasses[1] = top();
3446
3447 enum {
3448 // A full decision tree on {superc is prim, subc is prim}:
3449 _prim_0_path = 1, // {P,N} => false
3450 // {P,P} & superc!=subc => false
3451 _prim_same_path, // {P,P} & superc==subc => true
3452 _prim_1_path, // {N,P} => false
3453 _ref_subtype_path, // {N,N} & subtype check wins => true
3454 _both_ref_path, // {N,N} & subtype check loses => false
3455 PATH_LIMIT
3456 };
3457
3458 RegionNode* region = new RegionNode(PATH_LIMIT);
3459 Node* phi = new PhiNode(region, TypeInt::BOOL);
3460 record_for_igvn(region);
3461
3462 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3463 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3464 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3465
3466 // First null-check both mirrors and load each mirror's klass metaobject.
3467 int which_arg;
3468 for (which_arg = 0; which_arg <= 1; which_arg++) {
3469 Node* arg = args[which_arg];
3470 arg = null_check(arg);
3471 if (stopped()) break;
3472 args[which_arg] = arg;
3473
3474 Node* p = basic_plus_adr(arg, class_klass_offset);
3475 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3476 klasses[which_arg] = _gvn.transform(kls);
3477 }
3478
3479 // Having loaded both klasses, test each for null.
3480 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3481 for (which_arg = 0; which_arg <= 1; which_arg++) {
3482 Node* kls = klasses[which_arg];
3483 Node* null_ctl = top();
3484 kls = null_check_oop(kls, &null_ctl, never_see_null);
3485 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3486 region->init_req(prim_path, null_ctl);
3487 if (stopped()) break;
3488 klasses[which_arg] = kls;
3489 }
3490
3491 if (!stopped()) {
3492 // now we have two reference types, in klasses[0..1]
3493 Node* subk = klasses[1]; // the argument to isAssignableFrom
3494 Node* superk = klasses[0]; // the receiver
3495 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3496 // now we have a successful reference subtype check
3497 region->set_req(_ref_subtype_path, control());
3498 }
3499
3500 // If both operands are primitive (both klasses null), then
3501 // we must return true when they are identical primitives.
3502 // It is convenient to test this after the first null klass check.
3503 set_control(region->in(_prim_0_path)); // go back to first null check
3504 if (!stopped()) {
3505 // Since superc is primitive, make a guard for the superc==subc case.
3506 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3507 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3508 generate_guard(bol_eq, region, PROB_FAIR);
3509 if (region->req() == PATH_LIMIT+1) {
3510 // A guard was added. If the added guard is taken, superc==subc.
3511 region->swap_edges(PATH_LIMIT, _prim_same_path);
3512 region->del_req(PATH_LIMIT);
3513 }
3514 region->set_req(_prim_0_path, control()); // Not equal after all.
3515 }
3516
3517 // these are the only paths that produce 'true':
3518 phi->set_req(_prim_same_path, intcon(1));
3519 phi->set_req(_ref_subtype_path, intcon(1));
3520
3521 // pull together the cases:
3522 assert(region->req() == PATH_LIMIT, "sane region");
3523 for (uint i = 1; i < region->req(); i++) {
3524 Node* ctl = region->in(i);
3525 if (ctl == NULL || ctl == top()) {
3526 region->set_req(i, top());
3527 phi ->set_req(i, top());
3528 } else if (phi->in(i) == NULL) {
3529 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3530 }
3531 }
3532
3533 set_control(_gvn.transform(region));
3534 set_result(_gvn.transform(phi));
3535 return true;
3536 }
3537
3538 //---------------------generate_array_guard_common------------------------
3539 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3540 bool obj_array, bool not_array) {
3541
3542 if (stopped()) {
3543 return NULL;
3544 }
3545
3546 // If obj_array/non_array==false/false:
3547 // Branch around if the given klass is in fact an array (either obj or prim).
3548 // If obj_array/non_array==false/true:
3549 // Branch around if the given klass is not an array klass of any kind.
3550 // If obj_array/non_array==true/true:
3551 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3552 // If obj_array/non_array==true/false:
3553 // Branch around if the kls is an oop array (Object[] or subtype)
3554 //
3555 // Like generate_guard, adds a new path onto the region.
3556 jint layout_con = 0;
3557 Node* layout_val = get_layout_helper(kls, layout_con);
3558 if (layout_val == NULL) {
3559 bool query = (obj_array
3560 ? Klass::layout_helper_is_objArray(layout_con)
3561 : Klass::layout_helper_is_array(layout_con));
3562 if (query == not_array) {
3563 return NULL; // never a branch
3564 } else { // always a branch
3565 Node* always_branch = control();
3566 if (region != NULL)
3567 region->add_req(always_branch);
3568 set_control(top());
3569 return always_branch;
3570 }
3571 }
3572 // Now test the correct condition.
3573 jint nval = (obj_array
3574 ? ((jint)Klass::_lh_array_tag_type_value
3575 << Klass::_lh_array_tag_shift)
3576 : Klass::_lh_neutral_value);
3577 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3578 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3579 // invert the test if we are looking for a non-array
3580 if (not_array) btest = BoolTest(btest).negate();
3581 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3582 return generate_fair_guard(bol, region);
3583 }
3584
3585
3586 //-----------------------inline_native_newArray--------------------------
3587 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3588 bool LibraryCallKit::inline_native_newArray() {
3589 Node* mirror = argument(0);
3590 Node* count_val = argument(1);
3591
3592 mirror = null_check(mirror);
3593 // If mirror or obj is dead, only null-path is taken.
3594 if (stopped()) return true;
3595
3596 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3597 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3598 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3599 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3600 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3601
3602 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3603 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3604 result_reg, _slow_path);
3605 Node* normal_ctl = control();
3606 Node* no_array_ctl = result_reg->in(_slow_path);
3607
3608 // Generate code for the slow case. We make a call to newArray().
3609 set_control(no_array_ctl);
3610 if (!stopped()) {
3611 // Either the input type is void.class, or else the
3612 // array klass has not yet been cached. Either the
3613 // ensuing call will throw an exception, or else it
3614 // will cache the array klass for next time.
3615 PreserveJVMState pjvms(this);
3616 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3617 Node* slow_result = set_results_for_java_call(slow_call);
3618 // this->control() comes from set_results_for_java_call
3619 result_reg->set_req(_slow_path, control());
3620 result_val->set_req(_slow_path, slow_result);
3621 result_io ->set_req(_slow_path, i_o());
3622 result_mem->set_req(_slow_path, reset_memory());
3623 }
3624
3625 set_control(normal_ctl);
3626 if (!stopped()) {
3627 // Normal case: The array type has been cached in the java.lang.Class.
3628 // The following call works fine even if the array type is polymorphic.
3629 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3630 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3631 result_reg->init_req(_normal_path, control());
3632 result_val->init_req(_normal_path, obj);
3633 result_io ->init_req(_normal_path, i_o());
3634 result_mem->init_req(_normal_path, reset_memory());
3635 }
3636
3637 // Return the combined state.
3638 set_i_o( _gvn.transform(result_io) );
3639 set_all_memory( _gvn.transform(result_mem));
3640
3641 C->set_has_split_ifs(true); // Has chance for split-if optimization
3642 set_result(result_reg, result_val);
3643 return true;
3644 }
3645
3646 //----------------------inline_native_getLength--------------------------
3647 // public static native int java.lang.reflect.Array.getLength(Object array);
3648 bool LibraryCallKit::inline_native_getLength() {
3649 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3650
3651 Node* array = null_check(argument(0));
3652 // If array is dead, only null-path is taken.
3653 if (stopped()) return true;
3654
3655 // Deoptimize if it is a non-array.
3656 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3657
3658 if (non_array != NULL) {
3659 PreserveJVMState pjvms(this);
3660 set_control(non_array);
3661 uncommon_trap(Deoptimization::Reason_intrinsic,
3662 Deoptimization::Action_maybe_recompile);
3663 }
3664
3665 // If control is dead, only non-array-path is taken.
3666 if (stopped()) return true;
3667
3668 // The works fine even if the array type is polymorphic.
3669 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3670 Node* result = load_array_length(array);
3671
3672 C->set_has_split_ifs(true); // Has chance for split-if optimization
3673 set_result(result);
3674 return true;
3675 }
3676
3677 //------------------------inline_array_copyOf----------------------------
3678 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3679 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3680 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3681 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3682
3683 // Get the arguments.
3684 Node* original = argument(0);
3685 Node* start = is_copyOfRange? argument(1): intcon(0);
3686 Node* end = is_copyOfRange? argument(2): argument(1);
3687 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3688
3689 Node* newcopy = NULL;
3690
3691 // Set the original stack and the reexecute bit for the interpreter to reexecute
3692 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3693 { PreserveReexecuteState preexecs(this);
3694 jvms()->set_should_reexecute(true);
3695
3696 array_type_mirror = null_check(array_type_mirror);
3697 original = null_check(original);
3698
3699 // Check if a null path was taken unconditionally.
3700 if (stopped()) return true;
3701
3702 Node* orig_length = load_array_length(original);
3703
3704 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3705 klass_node = null_check(klass_node);
3706
3707 RegionNode* bailout = new RegionNode(1);
3708 record_for_igvn(bailout);
3709
3710 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3711 // Bail out if that is so.
3712 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3713 if (not_objArray != NULL) {
3714 // Improve the klass node's type from the new optimistic assumption:
3715 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3716 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3717 Node* cast = new CastPPNode(klass_node, akls);
3718 cast->init_req(0, control());
3719 klass_node = _gvn.transform(cast);
3720 }
3721
3722 // Bail out if either start or end is negative.
3723 generate_negative_guard(start, bailout, &start);
3724 generate_negative_guard(end, bailout, &end);
3725
3726 Node* length = end;
3727 if (_gvn.type(start) != TypeInt::ZERO) {
3728 length = _gvn.transform(new SubINode(end, start));
3729 }
3730
3731 // Bail out if length is negative.
3732 // Without this the new_array would throw
3733 // NegativeArraySizeException but IllegalArgumentException is what
3734 // should be thrown
3735 generate_negative_guard(length, bailout, &length);
3736
3737 if (bailout->req() > 1) {
3738 PreserveJVMState pjvms(this);
3739 set_control(_gvn.transform(bailout));
3740 uncommon_trap(Deoptimization::Reason_intrinsic,
3741 Deoptimization::Action_maybe_recompile);
3742 }
3743
3744 if (!stopped()) {
3745 // How many elements will we copy from the original?
3746 // The answer is MinI(orig_length - start, length).
3747 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3748 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3749
3750 // Generate a direct call to the right arraycopy function(s).
3751 // We know the copy is disjoint but we might not know if the
3752 // oop stores need checking.
3753 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3754 // This will fail a store-check if x contains any non-nulls.
3755
3756 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3757 // loads/stores but it is legal only if we're sure the
3758 // Arrays.copyOf would succeed. So we need all input arguments
3759 // to the copyOf to be validated, including that the copy to the
3760 // new array won't trigger an ArrayStoreException. That subtype
3761 // check can be optimized if we know something on the type of
3762 // the input array from type speculation.
3763 if (_gvn.type(klass_node)->singleton()) {
3764 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3765 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3766
3767 int test = C->static_subtype_check(superk, subk);
3768 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3769 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3770 if (t_original->speculative_type() != NULL) {
3771 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3772 }
3773 }
3774 }
3775
3776 bool validated = false;
3777 // Reason_class_check rather than Reason_intrinsic because we
3778 // want to intrinsify even if this traps.
3779 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3780 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3781 klass_node);
3782
3783 if (not_subtype_ctrl != top()) {
3784 PreserveJVMState pjvms(this);
3785 set_control(not_subtype_ctrl);
3786 uncommon_trap(Deoptimization::Reason_class_check,
3787 Deoptimization::Action_make_not_entrant);
3788 assert(stopped(), "Should be stopped");
3789 }
3790 validated = true;
3791 }
3792
3793 if (!stopped()) {
3794 newcopy = new_array(klass_node, length, 0); // no arguments to push
3795
3796 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3797 load_object_klass(original), klass_node);
3798 if (!is_copyOfRange) {
3799 ac->set_copyof(validated);
3800 } else {
3801 ac->set_copyofrange(validated);
3802 }
3803 Node* n = _gvn.transform(ac);
3804 if (n == ac) {
3805 ac->connect_outputs(this);
3806 } else {
3807 assert(validated, "shouldn't transform if all arguments not validated");
3808 set_all_memory(n);
3809 }
3810 }
3811 }
3812 } // original reexecute is set back here
3813
3814 C->set_has_split_ifs(true); // Has chance for split-if optimization
3815 if (!stopped()) {
3816 set_result(newcopy);
3817 }
3818 return true;
3819 }
3820
3821
3822 //----------------------generate_virtual_guard---------------------------
3823 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3824 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3825 RegionNode* slow_region) {
3826 ciMethod* method = callee();
3827 int vtable_index = method->vtable_index();
3828 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3829 "bad index %d", vtable_index);
3830 // Get the Method* out of the appropriate vtable entry.
3831 int entry_offset = in_bytes(InstanceKlass::vtable_start_offset()) +
3832 vtable_index*vtableEntry::size_in_bytes() +
3833 vtableEntry::method_offset_in_bytes();
3834 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3835 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3836
3837 // Compare the target method with the expected method (e.g., Object.hashCode).
3838 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3839
3840 Node* native_call = makecon(native_call_addr);
3841 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3842 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3843
3844 return generate_slow_guard(test_native, slow_region);
3845 }
3846
3847 //-----------------------generate_method_call----------------------------
3848 // Use generate_method_call to make a slow-call to the real
3849 // method if the fast path fails. An alternative would be to
3850 // use a stub like OptoRuntime::slow_arraycopy_Java.
3851 // This only works for expanding the current library call,
3852 // not another intrinsic. (E.g., don't use this for making an
3853 // arraycopy call inside of the copyOf intrinsic.)
3854 CallJavaNode*
3855 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3856 // When compiling the intrinsic method itself, do not use this technique.
3857 guarantee(callee() != C->method(), "cannot make slow-call to self");
3858
3859 ciMethod* method = callee();
3860 // ensure the JVMS we have will be correct for this call
3861 guarantee(method_id == method->intrinsic_id(), "must match");
3862
3863 const TypeFunc* tf = TypeFunc::make(method);
3864 CallJavaNode* slow_call;
3865 if (is_static) {
3866 assert(!is_virtual, "");
3867 slow_call = new CallStaticJavaNode(C, tf,
3868 SharedRuntime::get_resolve_static_call_stub(),
3869 method, bci());
3870 } else if (is_virtual) {
3871 null_check_receiver();
3872 int vtable_index = Method::invalid_vtable_index;
3873 if (UseInlineCaches) {
3874 // Suppress the vtable call
3875 } else {
3876 // hashCode and clone are not a miranda methods,
3877 // so the vtable index is fixed.
3878 // No need to use the linkResolver to get it.
3879 vtable_index = method->vtable_index();
3880 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3881 "bad index %d", vtable_index);
3882 }
3883 slow_call = new CallDynamicJavaNode(tf,
3884 SharedRuntime::get_resolve_virtual_call_stub(),
3885 method, vtable_index, bci());
3886 } else { // neither virtual nor static: opt_virtual
3887 null_check_receiver();
3888 slow_call = new CallStaticJavaNode(C, tf,
3889 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3890 method, bci());
3891 slow_call->set_optimized_virtual(true);
3892 }
3893 set_arguments_for_java_call(slow_call);
3894 set_edges_for_java_call(slow_call);
3895 return slow_call;
3896 }
3897
3898
3899 /**
3900 * Build special case code for calls to hashCode on an object. This call may
3901 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3902 * slightly different code.
3903 */
3904 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3905 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3906 assert(!(is_virtual && is_static), "either virtual, special, or static");
3907
3908 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3909
3910 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3911 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3912 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3913 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3914 Node* obj = NULL;
3915 if (!is_static) {
3916 // Check for hashing null object
3917 obj = null_check_receiver();
3918 if (stopped()) return true; // unconditionally null
3919 result_reg->init_req(_null_path, top());
3920 result_val->init_req(_null_path, top());
3921 } else {
3922 // Do a null check, and return zero if null.
3923 // System.identityHashCode(null) == 0
3924 obj = argument(0);
3925 Node* null_ctl = top();
3926 obj = null_check_oop(obj, &null_ctl);
3927 result_reg->init_req(_null_path, null_ctl);
3928 result_val->init_req(_null_path, _gvn.intcon(0));
3929 }
3930
3931 // Unconditionally null? Then return right away.
3932 if (stopped()) {
3933 set_control( result_reg->in(_null_path));
3934 if (!stopped())
3935 set_result(result_val->in(_null_path));
3936 return true;
3937 }
3938
3939 // We only go to the fast case code if we pass a number of guards. The
3940 // paths which do not pass are accumulated in the slow_region.
3941 RegionNode* slow_region = new RegionNode(1);
3942 record_for_igvn(slow_region);
3943
3944 // If this is a virtual call, we generate a funny guard. We pull out
3945 // the vtable entry corresponding to hashCode() from the target object.
3946 // If the target method which we are calling happens to be the native
3947 // Object hashCode() method, we pass the guard. We do not need this
3948 // guard for non-virtual calls -- the caller is known to be the native
3949 // Object hashCode().
3950 if (is_virtual) {
3951 // After null check, get the object's klass.
3952 Node* obj_klass = load_object_klass(obj);
3953 generate_virtual_guard(obj_klass, slow_region);
3954 }
3955
3956 // Get the header out of the object, use LoadMarkNode when available
3957 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3958 // The control of the load must be NULL. Otherwise, the load can move before
3959 // the null check after castPP removal.
3960 Node* no_ctrl = NULL;
3961 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3962
3963 // Test the header to see if it is unlocked.
3964 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3965 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3966 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3967 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3968 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3969
3970 generate_slow_guard(test_unlocked, slow_region);
3971
3972 // Get the hash value and check to see that it has been properly assigned.
3973 // We depend on hash_mask being at most 32 bits and avoid the use of
3974 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3975 // vm: see markOop.hpp.
3976 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3977 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3978 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3979 // This hack lets the hash bits live anywhere in the mark object now, as long
3980 // as the shift drops the relevant bits into the low 32 bits. Note that
3981 // Java spec says that HashCode is an int so there's no point in capturing
3982 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3983 hshifted_header = ConvX2I(hshifted_header);
3984 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3985
3986 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3987 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3988 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3989
3990 generate_slow_guard(test_assigned, slow_region);
3991
3992 Node* init_mem = reset_memory();
3993 // fill in the rest of the null path:
3994 result_io ->init_req(_null_path, i_o());
3995 result_mem->init_req(_null_path, init_mem);
3996
3997 result_val->init_req(_fast_path, hash_val);
3998 result_reg->init_req(_fast_path, control());
3999 result_io ->init_req(_fast_path, i_o());
4000 result_mem->init_req(_fast_path, init_mem);
4001
4002 // Generate code for the slow case. We make a call to hashCode().
4003 set_control(_gvn.transform(slow_region));
4004 if (!stopped()) {
4005 // No need for PreserveJVMState, because we're using up the present state.
4006 set_all_memory(init_mem);
4007 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4008 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4009 Node* slow_result = set_results_for_java_call(slow_call);
4010 // this->control() comes from set_results_for_java_call
4011 result_reg->init_req(_slow_path, control());
4012 result_val->init_req(_slow_path, slow_result);
4013 result_io ->set_req(_slow_path, i_o());
4014 result_mem ->set_req(_slow_path, reset_memory());
4015 }
4016
4017 // Return the combined state.
4018 set_i_o( _gvn.transform(result_io) );
4019 set_all_memory( _gvn.transform(result_mem));
4020
4021 set_result(result_reg, result_val);
4022 return true;
4023 }
4024
4025 //---------------------------inline_native_getClass----------------------------
4026 // public final native Class<?> java.lang.Object.getClass();
4027 //
4028 // Build special case code for calls to getClass on an object.
4029 bool LibraryCallKit::inline_native_getClass() {
4030 Node* obj = null_check_receiver();
4031 if (stopped()) return true;
4032 set_result(load_mirror_from_klass(load_object_klass(obj)));
4033 return true;
4034 }
4035
4036 //-----------------inline_native_Reflection_getCallerClass---------------------
4037 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4038 //
4039 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4040 //
4041 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4042 // in that it must skip particular security frames and checks for
4043 // caller sensitive methods.
4044 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4045 #ifndef PRODUCT
4046 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4047 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4048 }
4049 #endif
4050
4051 if (!jvms()->has_method()) {
4052 #ifndef PRODUCT
4053 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4054 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4055 }
4056 #endif
4057 return false;
4058 }
4059
4060 // Walk back up the JVM state to find the caller at the required
4061 // depth.
4062 JVMState* caller_jvms = jvms();
4063
4064 // Cf. JVM_GetCallerClass
4065 // NOTE: Start the loop at depth 1 because the current JVM state does
4066 // not include the Reflection.getCallerClass() frame.
4067 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4068 ciMethod* m = caller_jvms->method();
4069 switch (n) {
4070 case 0:
4071 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4072 break;
4073 case 1:
4074 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4075 if (!m->caller_sensitive()) {
4076 #ifndef PRODUCT
4077 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4078 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4079 }
4080 #endif
4081 return false; // bail-out; let JVM_GetCallerClass do the work
4082 }
4083 break;
4084 default:
4085 if (!m->is_ignored_by_security_stack_walk()) {
4086 // We have reached the desired frame; return the holder class.
4087 // Acquire method holder as java.lang.Class and push as constant.
4088 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4089 ciInstance* caller_mirror = caller_klass->java_mirror();
4090 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4091
4092 #ifndef PRODUCT
4093 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4094 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());
4095 tty->print_cr(" JVM state at this point:");
4096 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4097 ciMethod* m = jvms()->of_depth(i)->method();
4098 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4099 }
4100 }
4101 #endif
4102 return true;
4103 }
4104 break;
4105 }
4106 }
4107
4108 #ifndef PRODUCT
4109 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4110 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4111 tty->print_cr(" JVM state at this point:");
4112 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4113 ciMethod* m = jvms()->of_depth(i)->method();
4114 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4115 }
4116 }
4117 #endif
4118
4119 return false; // bail-out; let JVM_GetCallerClass do the work
4120 }
4121
4122 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4123 Node* arg = argument(0);
4124 Node* result = NULL;
4125
4126 switch (id) {
4127 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4128 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4129 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4130 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4131
4132 case vmIntrinsics::_doubleToLongBits: {
4133 // two paths (plus control) merge in a wood
4134 RegionNode *r = new RegionNode(3);
4135 Node *phi = new PhiNode(r, TypeLong::LONG);
4136
4137 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4138 // Build the boolean node
4139 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4140
4141 // Branch either way.
4142 // NaN case is less traveled, which makes all the difference.
4143 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4144 Node *opt_isnan = _gvn.transform(ifisnan);
4145 assert( opt_isnan->is_If(), "Expect an IfNode");
4146 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4147 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4148
4149 set_control(iftrue);
4150
4151 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4152 Node *slow_result = longcon(nan_bits); // return NaN
4153 phi->init_req(1, _gvn.transform( slow_result ));
4154 r->init_req(1, iftrue);
4155
4156 // Else fall through
4157 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4158 set_control(iffalse);
4159
4160 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4161 r->init_req(2, iffalse);
4162
4163 // Post merge
4164 set_control(_gvn.transform(r));
4165 record_for_igvn(r);
4166
4167 C->set_has_split_ifs(true); // Has chance for split-if optimization
4168 result = phi;
4169 assert(result->bottom_type()->isa_long(), "must be");
4170 break;
4171 }
4172
4173 case vmIntrinsics::_floatToIntBits: {
4174 // two paths (plus control) merge in a wood
4175 RegionNode *r = new RegionNode(3);
4176 Node *phi = new PhiNode(r, TypeInt::INT);
4177
4178 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4179 // Build the boolean node
4180 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4181
4182 // Branch either way.
4183 // NaN case is less traveled, which makes all the difference.
4184 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4185 Node *opt_isnan = _gvn.transform(ifisnan);
4186 assert( opt_isnan->is_If(), "Expect an IfNode");
4187 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4188 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4189
4190 set_control(iftrue);
4191
4192 static const jint nan_bits = 0x7fc00000;
4193 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4194 phi->init_req(1, _gvn.transform( slow_result ));
4195 r->init_req(1, iftrue);
4196
4197 // Else fall through
4198 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4199 set_control(iffalse);
4200
4201 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4202 r->init_req(2, iffalse);
4203
4204 // Post merge
4205 set_control(_gvn.transform(r));
4206 record_for_igvn(r);
4207
4208 C->set_has_split_ifs(true); // Has chance for split-if optimization
4209 result = phi;
4210 assert(result->bottom_type()->isa_int(), "must be");
4211 break;
4212 }
4213
4214 default:
4215 fatal_unexpected_iid(id);
4216 break;
4217 }
4218 set_result(_gvn.transform(result));
4219 return true;
4220 }
4221
4222 //----------------------inline_unsafe_copyMemory-------------------------
4223 // public native void Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4224 bool LibraryCallKit::inline_unsafe_copyMemory() {
4225 if (callee()->is_static()) return false; // caller must have the capability!
4226 null_check_receiver(); // null-check receiver
4227 if (stopped()) return true;
4228
4229 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4230
4231 Node* src_ptr = argument(1); // type: oop
4232 Node* src_off = ConvL2X(argument(2)); // type: long
4233 Node* dst_ptr = argument(4); // type: oop
4234 Node* dst_off = ConvL2X(argument(5)); // type: long
4235 Node* size = ConvL2X(argument(7)); // type: long
4236
4237 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4238 "fieldOffset must be byte-scaled");
4239
4240 Node* src = make_unsafe_address(src_ptr, src_off);
4241 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4242
4243 // Conservatively insert a memory barrier on all memory slices.
4244 // Do not let writes of the copy source or destination float below the copy.
4245 insert_mem_bar(Op_MemBarCPUOrder);
4246
4247 // Call it. Note that the length argument is not scaled.
4248 make_runtime_call(RC_LEAF|RC_NO_FP,
4249 OptoRuntime::fast_arraycopy_Type(),
4250 StubRoutines::unsafe_arraycopy(),
4251 "unsafe_arraycopy",
4252 TypeRawPtr::BOTTOM,
4253 src, dst, size XTOP);
4254
4255 // Do not let reads of the copy destination float above the copy.
4256 insert_mem_bar(Op_MemBarCPUOrder);
4257
4258 return true;
4259 }
4260
4261 //------------------------clone_coping-----------------------------------
4262 // Helper function for inline_native_clone.
4263 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4264 assert(obj_size != NULL, "");
4265 Node* raw_obj = alloc_obj->in(1);
4266 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4267
4268 AllocateNode* alloc = NULL;
4269 if (ReduceBulkZeroing) {
4270 // We will be completely responsible for initializing this object -
4271 // mark Initialize node as complete.
4272 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4273 // The object was just allocated - there should be no any stores!
4274 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4275 // Mark as complete_with_arraycopy so that on AllocateNode
4276 // expansion, we know this AllocateNode is initialized by an array
4277 // copy and a StoreStore barrier exists after the array copy.
4278 alloc->initialization()->set_complete_with_arraycopy();
4279 }
4280
4281 // Copy the fastest available way.
4282 // TODO: generate fields copies for small objects instead.
4283 Node* src = obj;
4284 Node* dest = alloc_obj;
4285 Node* size = _gvn.transform(obj_size);
4286
4287 // Exclude the header but include array length to copy by 8 bytes words.
4288 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4289 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4290 instanceOopDesc::base_offset_in_bytes();
4291 // base_off:
4292 // 8 - 32-bit VM
4293 // 12 - 64-bit VM, compressed klass
4294 // 16 - 64-bit VM, normal klass
4295 if (base_off % BytesPerLong != 0) {
4296 assert(UseCompressedClassPointers, "");
4297 if (is_array) {
4298 // Exclude length to copy by 8 bytes words.
4299 base_off += sizeof(int);
4300 } else {
4301 // Include klass to copy by 8 bytes words.
4302 base_off = instanceOopDesc::klass_offset_in_bytes();
4303 }
4304 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4305 }
4306 src = basic_plus_adr(src, base_off);
4307 dest = basic_plus_adr(dest, base_off);
4308
4309 // Compute the length also, if needed:
4310 Node* countx = size;
4311 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4312 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4313
4314 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4315
4316 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4317 ac->set_clonebasic();
4318 Node* n = _gvn.transform(ac);
4319 if (n == ac) {
4320 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4321 } else {
4322 set_all_memory(n);
4323 }
4324
4325 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4326 if (card_mark) {
4327 assert(!is_array, "");
4328 // Put in store barrier for any and all oops we are sticking
4329 // into this object. (We could avoid this if we could prove
4330 // that the object type contains no oop fields at all.)
4331 Node* no_particular_value = NULL;
4332 Node* no_particular_field = NULL;
4333 int raw_adr_idx = Compile::AliasIdxRaw;
4334 post_barrier(control(),
4335 memory(raw_adr_type),
4336 alloc_obj,
4337 no_particular_field,
4338 raw_adr_idx,
4339 no_particular_value,
4340 T_OBJECT,
4341 false);
4342 }
4343
4344 // Do not let reads from the cloned object float above the arraycopy.
4345 if (alloc != NULL) {
4346 // Do not let stores that initialize this object be reordered with
4347 // a subsequent store that would make this object accessible by
4348 // other threads.
4349 // Record what AllocateNode this StoreStore protects so that
4350 // escape analysis can go from the MemBarStoreStoreNode to the
4351 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4352 // based on the escape status of the AllocateNode.
4353 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4354 } else {
4355 insert_mem_bar(Op_MemBarCPUOrder);
4356 }
4357 }
4358
4359 //------------------------inline_native_clone----------------------------
4360 // protected native Object java.lang.Object.clone();
4361 //
4362 // Here are the simple edge cases:
4363 // null receiver => normal trap
4364 // virtual and clone was overridden => slow path to out-of-line clone
4365 // not cloneable or finalizer => slow path to out-of-line Object.clone
4366 //
4367 // The general case has two steps, allocation and copying.
4368 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4369 //
4370 // Copying also has two cases, oop arrays and everything else.
4371 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4372 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4373 //
4374 // These steps fold up nicely if and when the cloned object's klass
4375 // can be sharply typed as an object array, a type array, or an instance.
4376 //
4377 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4378 PhiNode* result_val;
4379
4380 // Set the reexecute bit for the interpreter to reexecute
4381 // the bytecode that invokes Object.clone if deoptimization happens.
4382 { PreserveReexecuteState preexecs(this);
4383 jvms()->set_should_reexecute(true);
4384
4385 Node* obj = null_check_receiver();
4386 if (stopped()) return true;
4387
4388 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4389
4390 // If we are going to clone an instance, we need its exact type to
4391 // know the number and types of fields to convert the clone to
4392 // loads/stores. Maybe a speculative type can help us.
4393 if (!obj_type->klass_is_exact() &&
4394 obj_type->speculative_type() != NULL &&
4395 obj_type->speculative_type()->is_instance_klass()) {
4396 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4397 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4398 !spec_ik->has_injected_fields()) {
4399 ciKlass* k = obj_type->klass();
4400 if (!k->is_instance_klass() ||
4401 k->as_instance_klass()->is_interface() ||
4402 k->as_instance_klass()->has_subklass()) {
4403 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4404 }
4405 }
4406 }
4407
4408 Node* obj_klass = load_object_klass(obj);
4409 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4410 const TypeOopPtr* toop = ((tklass != NULL)
4411 ? tklass->as_instance_type()
4412 : TypeInstPtr::NOTNULL);
4413
4414 // Conservatively insert a memory barrier on all memory slices.
4415 // Do not let writes into the original float below the clone.
4416 insert_mem_bar(Op_MemBarCPUOrder);
4417
4418 // paths into result_reg:
4419 enum {
4420 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4421 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4422 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4423 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4424 PATH_LIMIT
4425 };
4426 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4427 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4428 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4429 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4430 record_for_igvn(result_reg);
4431
4432 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4433 int raw_adr_idx = Compile::AliasIdxRaw;
4434
4435 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4436 if (array_ctl != NULL) {
4437 // It's an array.
4438 PreserveJVMState pjvms(this);
4439 set_control(array_ctl);
4440 Node* obj_length = load_array_length(obj);
4441 Node* obj_size = NULL;
4442 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4443
4444 if (!use_ReduceInitialCardMarks()) {
4445 // If it is an oop array, it requires very special treatment,
4446 // because card marking is required on each card of the array.
4447 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4448 if (is_obja != NULL) {
4449 PreserveJVMState pjvms2(this);
4450 set_control(is_obja);
4451 // Generate a direct call to the right arraycopy function(s).
4452 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4453 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4454 ac->set_cloneoop();
4455 Node* n = _gvn.transform(ac);
4456 assert(n == ac, "cannot disappear");
4457 ac->connect_outputs(this);
4458
4459 result_reg->init_req(_objArray_path, control());
4460 result_val->init_req(_objArray_path, alloc_obj);
4461 result_i_o ->set_req(_objArray_path, i_o());
4462 result_mem ->set_req(_objArray_path, reset_memory());
4463 }
4464 }
4465 // Otherwise, there are no card marks to worry about.
4466 // (We can dispense with card marks if we know the allocation
4467 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4468 // causes the non-eden paths to take compensating steps to
4469 // simulate a fresh allocation, so that no further
4470 // card marks are required in compiled code to initialize
4471 // the object.)
4472
4473 if (!stopped()) {
4474 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4475
4476 // Present the results of the copy.
4477 result_reg->init_req(_array_path, control());
4478 result_val->init_req(_array_path, alloc_obj);
4479 result_i_o ->set_req(_array_path, i_o());
4480 result_mem ->set_req(_array_path, reset_memory());
4481 }
4482 }
4483
4484 // We only go to the instance fast case code if we pass a number of guards.
4485 // The paths which do not pass are accumulated in the slow_region.
4486 RegionNode* slow_region = new RegionNode(1);
4487 record_for_igvn(slow_region);
4488 if (!stopped()) {
4489 // It's an instance (we did array above). Make the slow-path tests.
4490 // If this is a virtual call, we generate a funny guard. We grab
4491 // the vtable entry corresponding to clone() from the target object.
4492 // If the target method which we are calling happens to be the
4493 // Object clone() method, we pass the guard. We do not need this
4494 // guard for non-virtual calls; the caller is known to be the native
4495 // Object clone().
4496 if (is_virtual) {
4497 generate_virtual_guard(obj_klass, slow_region);
4498 }
4499
4500 // The object must be cloneable and must not have a finalizer.
4501 // Both of these conditions may be checked in a single test.
4502 // We could optimize the cloneable test further, but we don't care.
4503 generate_access_flags_guard(obj_klass,
4504 // Test both conditions:
4505 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4506 // Must be cloneable but not finalizer:
4507 JVM_ACC_IS_CLONEABLE,
4508 slow_region);
4509 }
4510
4511 if (!stopped()) {
4512 // It's an instance, and it passed the slow-path tests.
4513 PreserveJVMState pjvms(this);
4514 Node* obj_size = NULL;
4515 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4516 // is reexecuted if deoptimization occurs and there could be problems when merging
4517 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4518 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4519
4520 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4521
4522 // Present the results of the slow call.
4523 result_reg->init_req(_instance_path, control());
4524 result_val->init_req(_instance_path, alloc_obj);
4525 result_i_o ->set_req(_instance_path, i_o());
4526 result_mem ->set_req(_instance_path, reset_memory());
4527 }
4528
4529 // Generate code for the slow case. We make a call to clone().
4530 set_control(_gvn.transform(slow_region));
4531 if (!stopped()) {
4532 PreserveJVMState pjvms(this);
4533 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4534 Node* slow_result = set_results_for_java_call(slow_call);
4535 // this->control() comes from set_results_for_java_call
4536 result_reg->init_req(_slow_path, control());
4537 result_val->init_req(_slow_path, slow_result);
4538 result_i_o ->set_req(_slow_path, i_o());
4539 result_mem ->set_req(_slow_path, reset_memory());
4540 }
4541
4542 // Return the combined state.
4543 set_control( _gvn.transform(result_reg));
4544 set_i_o( _gvn.transform(result_i_o));
4545 set_all_memory( _gvn.transform(result_mem));
4546 } // original reexecute is set back here
4547
4548 set_result(_gvn.transform(result_val));
4549 return true;
4550 }
4551
4552 // If we have a tighly coupled allocation, the arraycopy may take care
4553 // of the array initialization. If one of the guards we insert between
4554 // the allocation and the arraycopy causes a deoptimization, an
4555 // unitialized array will escape the compiled method. To prevent that
4556 // we set the JVM state for uncommon traps between the allocation and
4557 // the arraycopy to the state before the allocation so, in case of
4558 // deoptimization, we'll reexecute the allocation and the
4559 // initialization.
4560 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4561 if (alloc != NULL) {
4562 ciMethod* trap_method = alloc->jvms()->method();
4563 int trap_bci = alloc->jvms()->bci();
4564
4565 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4566 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4567 // Make sure there's no store between the allocation and the
4568 // arraycopy otherwise visible side effects could be rexecuted
4569 // in case of deoptimization and cause incorrect execution.
4570 bool no_interfering_store = true;
4571 Node* mem = alloc->in(TypeFunc::Memory);
4572 if (mem->is_MergeMem()) {
4573 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4574 Node* n = mms.memory();
4575 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4576 assert(n->is_Store(), "what else?");
4577 no_interfering_store = false;
4578 break;
4579 }
4580 }
4581 } else {
4582 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4583 Node* n = mms.memory();
4584 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4585 assert(n->is_Store(), "what else?");
4586 no_interfering_store = false;
4587 break;
4588 }
4589 }
4590 }
4591
4592 if (no_interfering_store) {
4593 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4594 uint size = alloc->req();
4595 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4596 old_jvms->set_map(sfpt);
4597 for (uint i = 0; i < size; i++) {
4598 sfpt->init_req(i, alloc->in(i));
4599 }
4600 // re-push array length for deoptimization
4601 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4602 old_jvms->set_sp(old_jvms->sp()+1);
4603 old_jvms->set_monoff(old_jvms->monoff()+1);
4604 old_jvms->set_scloff(old_jvms->scloff()+1);
4605 old_jvms->set_endoff(old_jvms->endoff()+1);
4606 old_jvms->set_should_reexecute(true);
4607
4608 sfpt->set_i_o(map()->i_o());
4609 sfpt->set_memory(map()->memory());
4610 sfpt->set_control(map()->control());
4611
4612 JVMState* saved_jvms = jvms();
4613 saved_reexecute_sp = _reexecute_sp;
4614
4615 set_jvms(sfpt->jvms());
4616 _reexecute_sp = jvms()->sp();
4617
4618 return saved_jvms;
4619 }
4620 }
4621 }
4622 return NULL;
4623 }
4624
4625 // In case of a deoptimization, we restart execution at the
4626 // allocation, allocating a new array. We would leave an uninitialized
4627 // array in the heap that GCs wouldn't expect. Move the allocation
4628 // after the traps so we don't allocate the array if we
4629 // deoptimize. This is possible because tightly_coupled_allocation()
4630 // guarantees there's no observer of the allocated array at this point
4631 // and the control flow is simple enough.
4632 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4633 if (saved_jvms != NULL && !stopped()) {
4634 assert(alloc != NULL, "only with a tightly coupled allocation");
4635 // restore JVM state to the state at the arraycopy
4636 saved_jvms->map()->set_control(map()->control());
4637 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4638 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4639 // If we've improved the types of some nodes (null check) while
4640 // emitting the guards, propagate them to the current state
4641 map()->replaced_nodes().apply(saved_jvms->map());
4642 set_jvms(saved_jvms);
4643 _reexecute_sp = saved_reexecute_sp;
4644
4645 // Remove the allocation from above the guards
4646 CallProjections callprojs;
4647 alloc->extract_projections(&callprojs, true);
4648 InitializeNode* init = alloc->initialization();
4649 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4650 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4651 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4652 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4653
4654 // move the allocation here (after the guards)
4655 _gvn.hash_delete(alloc);
4656 alloc->set_req(TypeFunc::Control, control());
4657 alloc->set_req(TypeFunc::I_O, i_o());
4658 Node *mem = reset_memory();
4659 set_all_memory(mem);
4660 alloc->set_req(TypeFunc::Memory, mem);
4661 set_control(init->proj_out(TypeFunc::Control));
4662 set_i_o(callprojs.fallthrough_ioproj);
4663
4664 // Update memory as done in GraphKit::set_output_for_allocation()
4665 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4666 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4667 if (ary_type->isa_aryptr() && length_type != NULL) {
4668 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4669 }
4670 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4671 int elemidx = C->get_alias_index(telemref);
4672 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4673 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4674
4675 Node* allocx = _gvn.transform(alloc);
4676 assert(allocx == alloc, "where has the allocation gone?");
4677 assert(dest->is_CheckCastPP(), "not an allocation result?");
4678
4679 _gvn.hash_delete(dest);
4680 dest->set_req(0, control());
4681 Node* destx = _gvn.transform(dest);
4682 assert(destx == dest, "where has the allocation result gone?");
4683 }
4684 }
4685
4686
4687 //------------------------------inline_arraycopy-----------------------
4688 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4689 // Object dest, int destPos,
4690 // int length);
4691 bool LibraryCallKit::inline_arraycopy() {
4692 // Get the arguments.
4693 Node* src = argument(0); // type: oop
4694 Node* src_offset = argument(1); // type: int
4695 Node* dest = argument(2); // type: oop
4696 Node* dest_offset = argument(3); // type: int
4697 Node* length = argument(4); // type: int
4698
4699
4700 // Check for allocation before we add nodes that would confuse
4701 // tightly_coupled_allocation()
4702 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4703
4704 int saved_reexecute_sp = -1;
4705 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4706 // See arraycopy_restore_alloc_state() comment
4707 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4708 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4709 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4710 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4711
4712 // The following tests must be performed
4713 // (1) src and dest are arrays.
4714 // (2) src and dest arrays must have elements of the same BasicType
4715 // (3) src and dest must not be null.
4716 // (4) src_offset must not be negative.
4717 // (5) dest_offset must not be negative.
4718 // (6) length must not be negative.
4719 // (7) src_offset + length must not exceed length of src.
4720 // (8) dest_offset + length must not exceed length of dest.
4721 // (9) each element of an oop array must be assignable
4722
4723 // (3) src and dest must not be null.
4724 // always do this here because we need the JVM state for uncommon traps
4725 Node* null_ctl = top();
4726 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4727 assert(null_ctl->is_top(), "no null control here");
4728 dest = null_check(dest, T_ARRAY);
4729
4730 if (!can_emit_guards) {
4731 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4732 // guards but the arraycopy node could still take advantage of a
4733 // tightly allocated allocation. tightly_coupled_allocation() is
4734 // called again to make sure it takes the null check above into
4735 // account: the null check is mandatory and if it caused an
4736 // uncommon trap to be emitted then the allocation can't be
4737 // considered tightly coupled in this context.
4738 alloc = tightly_coupled_allocation(dest, NULL);
4739 }
4740
4741 bool validated = false;
4742
4743 const Type* src_type = _gvn.type(src);
4744 const Type* dest_type = _gvn.type(dest);
4745 const TypeAryPtr* top_src = src_type->isa_aryptr();
4746 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4747
4748 // Do we have the type of src?
4749 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4750 // Do we have the type of dest?
4751 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4752 // Is the type for src from speculation?
4753 bool src_spec = false;
4754 // Is the type for dest from speculation?
4755 bool dest_spec = false;
4756
4757 if ((!has_src || !has_dest) && can_emit_guards) {
4758 // We don't have sufficient type information, let's see if
4759 // speculative types can help. We need to have types for both src
4760 // and dest so that it pays off.
4761
4762 // Do we already have or could we have type information for src
4763 bool could_have_src = has_src;
4764 // Do we already have or could we have type information for dest
4765 bool could_have_dest = has_dest;
4766
4767 ciKlass* src_k = NULL;
4768 if (!has_src) {
4769 src_k = src_type->speculative_type_not_null();
4770 if (src_k != NULL && src_k->is_array_klass()) {
4771 could_have_src = true;
4772 }
4773 }
4774
4775 ciKlass* dest_k = NULL;
4776 if (!has_dest) {
4777 dest_k = dest_type->speculative_type_not_null();
4778 if (dest_k != NULL && dest_k->is_array_klass()) {
4779 could_have_dest = true;
4780 }
4781 }
4782
4783 if (could_have_src && could_have_dest) {
4784 // This is going to pay off so emit the required guards
4785 if (!has_src) {
4786 src = maybe_cast_profiled_obj(src, src_k, true);
4787 src_type = _gvn.type(src);
4788 top_src = src_type->isa_aryptr();
4789 has_src = (top_src != NULL && top_src->klass() != NULL);
4790 src_spec = true;
4791 }
4792 if (!has_dest) {
4793 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4794 dest_type = _gvn.type(dest);
4795 top_dest = dest_type->isa_aryptr();
4796 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4797 dest_spec = true;
4798 }
4799 }
4800 }
4801
4802 if (has_src && has_dest && can_emit_guards) {
4803 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4804 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4805 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4806 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4807
4808 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4809 // If both arrays are object arrays then having the exact types
4810 // for both will remove the need for a subtype check at runtime
4811 // before the call and may make it possible to pick a faster copy
4812 // routine (without a subtype check on every element)
4813 // Do we have the exact type of src?
4814 bool could_have_src = src_spec;
4815 // Do we have the exact type of dest?
4816 bool could_have_dest = dest_spec;
4817 ciKlass* src_k = top_src->klass();
4818 ciKlass* dest_k = top_dest->klass();
4819 if (!src_spec) {
4820 src_k = src_type->speculative_type_not_null();
4821 if (src_k != NULL && src_k->is_array_klass()) {
4822 could_have_src = true;
4823 }
4824 }
4825 if (!dest_spec) {
4826 dest_k = dest_type->speculative_type_not_null();
4827 if (dest_k != NULL && dest_k->is_array_klass()) {
4828 could_have_dest = true;
4829 }
4830 }
4831 if (could_have_src && could_have_dest) {
4832 // If we can have both exact types, emit the missing guards
4833 if (could_have_src && !src_spec) {
4834 src = maybe_cast_profiled_obj(src, src_k, true);
4835 }
4836 if (could_have_dest && !dest_spec) {
4837 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4838 }
4839 }
4840 }
4841 }
4842
4843 ciMethod* trap_method = method();
4844 int trap_bci = bci();
4845 if (saved_jvms != NULL) {
4846 trap_method = alloc->jvms()->method();
4847 trap_bci = alloc->jvms()->bci();
4848 }
4849
4850 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4851 can_emit_guards &&
4852 !src->is_top() && !dest->is_top()) {
4853 // validate arguments: enables transformation the ArrayCopyNode
4854 validated = true;
4855
4856 RegionNode* slow_region = new RegionNode(1);
4857 record_for_igvn(slow_region);
4858
4859 // (1) src and dest are arrays.
4860 generate_non_array_guard(load_object_klass(src), slow_region);
4861 generate_non_array_guard(load_object_klass(dest), slow_region);
4862
4863 // (2) src and dest arrays must have elements of the same BasicType
4864 // done at macro expansion or at Ideal transformation time
4865
4866 // (4) src_offset must not be negative.
4867 generate_negative_guard(src_offset, slow_region);
4868
4869 // (5) dest_offset must not be negative.
4870 generate_negative_guard(dest_offset, slow_region);
4871
4872 // (7) src_offset + length must not exceed length of src.
4873 generate_limit_guard(src_offset, length,
4874 load_array_length(src),
4875 slow_region);
4876
4877 // (8) dest_offset + length must not exceed length of dest.
4878 generate_limit_guard(dest_offset, length,
4879 load_array_length(dest),
4880 slow_region);
4881
4882 // (9) each element of an oop array must be assignable
4883 Node* src_klass = load_object_klass(src);
4884 Node* dest_klass = load_object_klass(dest);
4885 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4886
4887 if (not_subtype_ctrl != top()) {
4888 PreserveJVMState pjvms(this);
4889 set_control(not_subtype_ctrl);
4890 uncommon_trap(Deoptimization::Reason_intrinsic,
4891 Deoptimization::Action_make_not_entrant);
4892 assert(stopped(), "Should be stopped");
4893 }
4894 {
4895 PreserveJVMState pjvms(this);
4896 set_control(_gvn.transform(slow_region));
4897 uncommon_trap(Deoptimization::Reason_intrinsic,
4898 Deoptimization::Action_make_not_entrant);
4899 assert(stopped(), "Should be stopped");
4900 }
4901 }
4902
4903 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4904
4905 if (stopped()) {
4906 return true;
4907 }
4908
4909 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4910 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4911 // so the compiler has a chance to eliminate them: during macro expansion,
4912 // we have to set their control (CastPP nodes are eliminated).
4913 load_object_klass(src), load_object_klass(dest),
4914 load_array_length(src), load_array_length(dest));
4915
4916 ac->set_arraycopy(validated);
4917
4918 Node* n = _gvn.transform(ac);
4919 if (n == ac) {
4920 ac->connect_outputs(this);
4921 } else {
4922 assert(validated, "shouldn't transform if all arguments not validated");
4923 set_all_memory(n);
4924 }
4925
4926 return true;
4927 }
4928
4929
4930 // Helper function which determines if an arraycopy immediately follows
4931 // an allocation, with no intervening tests or other escapes for the object.
4932 AllocateArrayNode*
4933 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4934 RegionNode* slow_region) {
4935 if (stopped()) return NULL; // no fast path
4936 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4937
4938 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4939 if (alloc == NULL) return NULL;
4940
4941 Node* rawmem = memory(Compile::AliasIdxRaw);
4942 // Is the allocation's memory state untouched?
4943 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4944 // Bail out if there have been raw-memory effects since the allocation.
4945 // (Example: There might have been a call or safepoint.)
4946 return NULL;
4947 }
4948 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4949 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4950 return NULL;
4951 }
4952
4953 // There must be no unexpected observers of this allocation.
4954 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4955 Node* obs = ptr->fast_out(i);
4956 if (obs != this->map()) {
4957 return NULL;
4958 }
4959 }
4960
4961 // This arraycopy must unconditionally follow the allocation of the ptr.
4962 Node* alloc_ctl = ptr->in(0);
4963 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4964
4965 Node* ctl = control();
4966 while (ctl != alloc_ctl) {
4967 // There may be guards which feed into the slow_region.
4968 // Any other control flow means that we might not get a chance
4969 // to finish initializing the allocated object.
4970 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4971 IfNode* iff = ctl->in(0)->as_If();
4972 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4973 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4974 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4975 ctl = iff->in(0); // This test feeds the known slow_region.
4976 continue;
4977 }
4978 // One more try: Various low-level checks bottom out in
4979 // uncommon traps. If the debug-info of the trap omits
4980 // any reference to the allocation, as we've already
4981 // observed, then there can be no objection to the trap.
4982 bool found_trap = false;
4983 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4984 Node* obs = not_ctl->fast_out(j);
4985 if (obs->in(0) == not_ctl && obs->is_Call() &&
4986 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4987 found_trap = true; break;
4988 }
4989 }
4990 if (found_trap) {
4991 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4992 continue;
4993 }
4994 }
4995 return NULL;
4996 }
4997
4998 // If we get this far, we have an allocation which immediately
4999 // precedes the arraycopy, and we can take over zeroing the new object.
5000 // The arraycopy will finish the initialization, and provide
5001 // a new control state to which we will anchor the destination pointer.
5002
5003 return alloc;
5004 }
5005
5006 //-------------inline_encodeISOArray-----------------------------------
5007 // encode char[] to byte[] in ISO_8859_1
5008 bool LibraryCallKit::inline_encodeISOArray() {
5009 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5010 // no receiver since it is static method
5011 Node *src = argument(0);
5012 Node *src_offset = argument(1);
5013 Node *dst = argument(2);
5014 Node *dst_offset = argument(3);
5015 Node *length = argument(4);
5016
5017 const Type* src_type = src->Value(&_gvn);
5018 const Type* dst_type = dst->Value(&_gvn);
5019 const TypeAryPtr* top_src = src_type->isa_aryptr();
5020 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5021 if (top_src == NULL || top_src->klass() == NULL ||
5022 top_dest == NULL || top_dest->klass() == NULL) {
5023 // failed array check
5024 return false;
5025 }
5026
5027 // Figure out the size and type of the elements we will be copying.
5028 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5029 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5030 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5031 return false;
5032 }
5033
5034 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5035 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5036 // 'src_start' points to src array + scaled offset
5037 // 'dst_start' points to dst array + scaled offset
5038
5039 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5040 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5041 enc = _gvn.transform(enc);
5042 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5043 set_memory(res_mem, mtype);
5044 set_result(enc);
5045 return true;
5046 }
5047
5048 //-------------inline_multiplyToLen-----------------------------------
5049 bool LibraryCallKit::inline_multiplyToLen() {
5050 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5051
5052 address stubAddr = StubRoutines::multiplyToLen();
5053 if (stubAddr == NULL) {
5054 return false; // Intrinsic's stub is not implemented on this platform
5055 }
5056 const char* stubName = "multiplyToLen";
5057
5058 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5059
5060 // no receiver because it is a static method
5061 Node* x = argument(0);
5062 Node* xlen = argument(1);
5063 Node* y = argument(2);
5064 Node* ylen = argument(3);
5065 Node* z = argument(4);
5066
5067 const Type* x_type = x->Value(&_gvn);
5068 const Type* y_type = y->Value(&_gvn);
5069 const TypeAryPtr* top_x = x_type->isa_aryptr();
5070 const TypeAryPtr* top_y = y_type->isa_aryptr();
5071 if (top_x == NULL || top_x->klass() == NULL ||
5072 top_y == NULL || top_y->klass() == NULL) {
5073 // failed array check
5074 return false;
5075 }
5076
5077 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5078 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5079 if (x_elem != T_INT || y_elem != T_INT) {
5080 return false;
5081 }
5082
5083 // Set the original stack and the reexecute bit for the interpreter to reexecute
5084 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5085 // on the return from z array allocation in runtime.
5086 { PreserveReexecuteState preexecs(this);
5087 jvms()->set_should_reexecute(true);
5088
5089 Node* x_start = array_element_address(x, intcon(0), x_elem);
5090 Node* y_start = array_element_address(y, intcon(0), y_elem);
5091 // 'x_start' points to x array + scaled xlen
5092 // 'y_start' points to y array + scaled ylen
5093
5094 // Allocate the result array
5095 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5096 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5097 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5098
5099 IdealKit ideal(this);
5100
5101 #define __ ideal.
5102 Node* one = __ ConI(1);
5103 Node* zero = __ ConI(0);
5104 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5105 __ set(need_alloc, zero);
5106 __ set(z_alloc, z);
5107 __ if_then(z, BoolTest::eq, null()); {
5108 __ increment (need_alloc, one);
5109 } __ else_(); {
5110 // Update graphKit memory and control from IdealKit.
5111 sync_kit(ideal);
5112 Node* zlen_arg = load_array_length(z);
5113 // Update IdealKit memory and control from graphKit.
5114 __ sync_kit(this);
5115 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5116 __ increment (need_alloc, one);
5117 } __ end_if();
5118 } __ end_if();
5119
5120 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5121 // Update graphKit memory and control from IdealKit.
5122 sync_kit(ideal);
5123 Node * narr = new_array(klass_node, zlen, 1);
5124 // Update IdealKit memory and control from graphKit.
5125 __ sync_kit(this);
5126 __ set(z_alloc, narr);
5127 } __ end_if();
5128
5129 sync_kit(ideal);
5130 z = __ value(z_alloc);
5131 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5132 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5133 // Final sync IdealKit and GraphKit.
5134 final_sync(ideal);
5135 #undef __
5136
5137 Node* z_start = array_element_address(z, intcon(0), T_INT);
5138
5139 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5140 OptoRuntime::multiplyToLen_Type(),
5141 stubAddr, stubName, TypePtr::BOTTOM,
5142 x_start, xlen, y_start, ylen, z_start, zlen);
5143 } // original reexecute is set back here
5144
5145 C->set_has_split_ifs(true); // Has chance for split-if optimization
5146 set_result(z);
5147 return true;
5148 }
5149
5150 //-------------inline_squareToLen------------------------------------
5151 bool LibraryCallKit::inline_squareToLen() {
5152 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5153
5154 address stubAddr = StubRoutines::squareToLen();
5155 if (stubAddr == NULL) {
5156 return false; // Intrinsic's stub is not implemented on this platform
5157 }
5158 const char* stubName = "squareToLen";
5159
5160 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5161
5162 Node* x = argument(0);
5163 Node* len = argument(1);
5164 Node* z = argument(2);
5165 Node* zlen = argument(3);
5166
5167 const Type* x_type = x->Value(&_gvn);
5168 const Type* z_type = z->Value(&_gvn);
5169 const TypeAryPtr* top_x = x_type->isa_aryptr();
5170 const TypeAryPtr* top_z = z_type->isa_aryptr();
5171 if (top_x == NULL || top_x->klass() == NULL ||
5172 top_z == NULL || top_z->klass() == NULL) {
5173 // failed array check
5174 return false;
5175 }
5176
5177 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5178 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5179 if (x_elem != T_INT || z_elem != T_INT) {
5180 return false;
5181 }
5182
5183
5184 Node* x_start = array_element_address(x, intcon(0), x_elem);
5185 Node* z_start = array_element_address(z, intcon(0), z_elem);
5186
5187 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5188 OptoRuntime::squareToLen_Type(),
5189 stubAddr, stubName, TypePtr::BOTTOM,
5190 x_start, len, z_start, zlen);
5191
5192 set_result(z);
5193 return true;
5194 }
5195
5196 //-------------inline_mulAdd------------------------------------------
5197 bool LibraryCallKit::inline_mulAdd() {
5198 assert(UseMulAddIntrinsic, "not implemented on this platform");
5199
5200 address stubAddr = StubRoutines::mulAdd();
5201 if (stubAddr == NULL) {
5202 return false; // Intrinsic's stub is not implemented on this platform
5203 }
5204 const char* stubName = "mulAdd";
5205
5206 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5207
5208 Node* out = argument(0);
5209 Node* in = argument(1);
5210 Node* offset = argument(2);
5211 Node* len = argument(3);
5212 Node* k = argument(4);
5213
5214 const Type* out_type = out->Value(&_gvn);
5215 const Type* in_type = in->Value(&_gvn);
5216 const TypeAryPtr* top_out = out_type->isa_aryptr();
5217 const TypeAryPtr* top_in = in_type->isa_aryptr();
5218 if (top_out == NULL || top_out->klass() == NULL ||
5219 top_in == NULL || top_in->klass() == NULL) {
5220 // failed array check
5221 return false;
5222 }
5223
5224 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5225 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5226 if (out_elem != T_INT || in_elem != T_INT) {
5227 return false;
5228 }
5229
5230 Node* outlen = load_array_length(out);
5231 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5232 Node* out_start = array_element_address(out, intcon(0), out_elem);
5233 Node* in_start = array_element_address(in, intcon(0), in_elem);
5234
5235 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5236 OptoRuntime::mulAdd_Type(),
5237 stubAddr, stubName, TypePtr::BOTTOM,
5238 out_start,in_start, new_offset, len, k);
5239 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5240 set_result(result);
5241 return true;
5242 }
5243
5244 //-------------inline_montgomeryMultiply-----------------------------------
5245 bool LibraryCallKit::inline_montgomeryMultiply() {
5246 address stubAddr = StubRoutines::montgomeryMultiply();
5247 if (stubAddr == NULL) {
5248 return false; // Intrinsic's stub is not implemented on this platform
5249 }
5250
5251 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5252 const char* stubName = "montgomery_square";
5253
5254 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5255
5256 Node* a = argument(0);
5257 Node* b = argument(1);
5258 Node* n = argument(2);
5259 Node* len = argument(3);
5260 Node* inv = argument(4);
5261 Node* m = argument(6);
5262
5263 const Type* a_type = a->Value(&_gvn);
5264 const TypeAryPtr* top_a = a_type->isa_aryptr();
5265 const Type* b_type = b->Value(&_gvn);
5266 const TypeAryPtr* top_b = b_type->isa_aryptr();
5267 const Type* n_type = a->Value(&_gvn);
5268 const TypeAryPtr* top_n = n_type->isa_aryptr();
5269 const Type* m_type = a->Value(&_gvn);
5270 const TypeAryPtr* top_m = m_type->isa_aryptr();
5271 if (top_a == NULL || top_a->klass() == NULL ||
5272 top_b == NULL || top_b->klass() == NULL ||
5273 top_n == NULL || top_n->klass() == NULL ||
5274 top_m == NULL || top_m->klass() == NULL) {
5275 // failed array check
5276 return false;
5277 }
5278
5279 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5280 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5281 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5282 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5283 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5284 return false;
5285 }
5286
5287 // Make the call
5288 {
5289 Node* a_start = array_element_address(a, intcon(0), a_elem);
5290 Node* b_start = array_element_address(b, intcon(0), b_elem);
5291 Node* n_start = array_element_address(n, intcon(0), n_elem);
5292 Node* m_start = array_element_address(m, intcon(0), m_elem);
5293
5294 Node* call = make_runtime_call(RC_LEAF,
5295 OptoRuntime::montgomeryMultiply_Type(),
5296 stubAddr, stubName, TypePtr::BOTTOM,
5297 a_start, b_start, n_start, len, inv, top(),
5298 m_start);
5299 set_result(m);
5300 }
5301
5302 return true;
5303 }
5304
5305 bool LibraryCallKit::inline_montgomerySquare() {
5306 address stubAddr = StubRoutines::montgomerySquare();
5307 if (stubAddr == NULL) {
5308 return false; // Intrinsic's stub is not implemented on this platform
5309 }
5310
5311 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5312 const char* stubName = "montgomery_square";
5313
5314 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5315
5316 Node* a = argument(0);
5317 Node* n = argument(1);
5318 Node* len = argument(2);
5319 Node* inv = argument(3);
5320 Node* m = argument(5);
5321
5322 const Type* a_type = a->Value(&_gvn);
5323 const TypeAryPtr* top_a = a_type->isa_aryptr();
5324 const Type* n_type = a->Value(&_gvn);
5325 const TypeAryPtr* top_n = n_type->isa_aryptr();
5326 const Type* m_type = a->Value(&_gvn);
5327 const TypeAryPtr* top_m = m_type->isa_aryptr();
5328 if (top_a == NULL || top_a->klass() == NULL ||
5329 top_n == NULL || top_n->klass() == NULL ||
5330 top_m == NULL || top_m->klass() == NULL) {
5331 // failed array check
5332 return false;
5333 }
5334
5335 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5336 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5337 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5338 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5339 return false;
5340 }
5341
5342 // Make the call
5343 {
5344 Node* a_start = array_element_address(a, intcon(0), a_elem);
5345 Node* n_start = array_element_address(n, intcon(0), n_elem);
5346 Node* m_start = array_element_address(m, intcon(0), m_elem);
5347
5348 Node* call = make_runtime_call(RC_LEAF,
5349 OptoRuntime::montgomerySquare_Type(),
5350 stubAddr, stubName, TypePtr::BOTTOM,
5351 a_start, n_start, len, inv, top(),
5352 m_start);
5353 set_result(m);
5354 }
5355
5356 return true;
5357 }
5358
5359 //-------------inline_vectorizedMismatch------------------------------
5360 bool LibraryCallKit::inline_vectorizedMismatch() {
5361 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5362
5363 address stubAddr = StubRoutines::vectorizedMismatch();
5364 if (stubAddr == NULL) {
5365 return false; // Intrinsic's stub is not implemented on this platform
5366 }
5367 const char* stubName = "vectorizedMismatch";
5368 int size_l = callee()->signature()->size();
5369 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5370
5371 Node* obja = argument(0);
5372 Node* aoffset = argument(1);
5373 Node* objb = argument(3);
5374 Node* boffset = argument(4);
5375 Node* length = argument(6);
5376 Node* scale = argument(7);
5377
5378 const Type* a_type = obja->Value(&_gvn);
5379 const Type* b_type = objb->Value(&_gvn);
5380 const TypeAryPtr* top_a = a_type->isa_aryptr();
5381 const TypeAryPtr* top_b = b_type->isa_aryptr();
5382 if (top_a == NULL || top_a->klass() == NULL ||
5383 top_b == NULL || top_b->klass() == NULL) {
5384 // failed array check
5385 return false;
5386 }
5387
5388 Node* call;
5389 jvms()->set_should_reexecute(true);
5390
5391 Node* obja_adr = make_unsafe_address(obja, aoffset);
5392 Node* objb_adr = make_unsafe_address(objb, boffset);
5393
5394 call = make_runtime_call(RC_LEAF,
5395 OptoRuntime::vectorizedMismatch_Type(),
5396 stubAddr, stubName, TypePtr::BOTTOM,
5397 obja_adr, objb_adr, length, scale);
5398
5399 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5400 set_result(result);
5401 return true;
5402 }
5403
5404 /**
5405 * Calculate CRC32 for byte.
5406 * int java.util.zip.CRC32.update(int crc, int b)
5407 */
5408 bool LibraryCallKit::inline_updateCRC32() {
5409 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5410 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5411 // no receiver since it is static method
5412 Node* crc = argument(0); // type: int
5413 Node* b = argument(1); // type: int
5414
5415 /*
5416 * int c = ~ crc;
5417 * b = timesXtoThe32[(b ^ c) & 0xFF];
5418 * b = b ^ (c >>> 8);
5419 * crc = ~b;
5420 */
5421
5422 Node* M1 = intcon(-1);
5423 crc = _gvn.transform(new XorINode(crc, M1));
5424 Node* result = _gvn.transform(new XorINode(crc, b));
5425 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5426
5427 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5428 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5429 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5430 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5431
5432 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5433 result = _gvn.transform(new XorINode(crc, result));
5434 result = _gvn.transform(new XorINode(result, M1));
5435 set_result(result);
5436 return true;
5437 }
5438
5439 /**
5440 * Calculate CRC32 for byte[] array.
5441 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5442 */
5443 bool LibraryCallKit::inline_updateBytesCRC32() {
5444 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5445 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5446 // no receiver since it is static method
5447 Node* crc = argument(0); // type: int
5448 Node* src = argument(1); // type: oop
5449 Node* offset = argument(2); // type: int
5450 Node* length = argument(3); // type: int
5451
5452 const Type* src_type = src->Value(&_gvn);
5453 const TypeAryPtr* top_src = src_type->isa_aryptr();
5454 if (top_src == NULL || top_src->klass() == NULL) {
5455 // failed array check
5456 return false;
5457 }
5458
5459 // Figure out the size and type of the elements we will be copying.
5460 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5461 if (src_elem != T_BYTE) {
5462 return false;
5463 }
5464
5465 // 'src_start' points to src array + scaled offset
5466 Node* src_start = array_element_address(src, offset, src_elem);
5467
5468 // We assume that range check is done by caller.
5469 // TODO: generate range check (offset+length < src.length) in debug VM.
5470
5471 // Call the stub.
5472 address stubAddr = StubRoutines::updateBytesCRC32();
5473 const char *stubName = "updateBytesCRC32";
5474
5475 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5476 stubAddr, stubName, TypePtr::BOTTOM,
5477 crc, src_start, length);
5478 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5479 set_result(result);
5480 return true;
5481 }
5482
5483 /**
5484 * Calculate CRC32 for ByteBuffer.
5485 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5486 */
5487 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5488 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5489 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5490 // no receiver since it is static method
5491 Node* crc = argument(0); // type: int
5492 Node* src = argument(1); // type: long
5493 Node* offset = argument(3); // type: int
5494 Node* length = argument(4); // type: int
5495
5496 src = ConvL2X(src); // adjust Java long to machine word
5497 Node* base = _gvn.transform(new CastX2PNode(src));
5498 offset = ConvI2X(offset);
5499
5500 // 'src_start' points to src array + scaled offset
5501 Node* src_start = basic_plus_adr(top(), base, offset);
5502
5503 // Call the stub.
5504 address stubAddr = StubRoutines::updateBytesCRC32();
5505 const char *stubName = "updateBytesCRC32";
5506
5507 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5508 stubAddr, stubName, TypePtr::BOTTOM,
5509 crc, src_start, length);
5510 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5511 set_result(result);
5512 return true;
5513 }
5514
5515 //------------------------------get_table_from_crc32c_class-----------------------
5516 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5517 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5518 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5519
5520 return table;
5521 }
5522
5523 //------------------------------inline_updateBytesCRC32C-----------------------
5524 //
5525 // Calculate CRC32C for byte[] array.
5526 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5527 //
5528 bool LibraryCallKit::inline_updateBytesCRC32C() {
5529 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5530 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5531 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5532 // no receiver since it is a static method
5533 Node* crc = argument(0); // type: int
5534 Node* src = argument(1); // type: oop
5535 Node* offset = argument(2); // type: int
5536 Node* end = argument(3); // type: int
5537
5538 Node* length = _gvn.transform(new SubINode(end, offset));
5539
5540 const Type* src_type = src->Value(&_gvn);
5541 const TypeAryPtr* top_src = src_type->isa_aryptr();
5542 if (top_src == NULL || top_src->klass() == NULL) {
5543 // failed array check
5544 return false;
5545 }
5546
5547 // Figure out the size and type of the elements we will be copying.
5548 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5549 if (src_elem != T_BYTE) {
5550 return false;
5551 }
5552
5553 // 'src_start' points to src array + scaled offset
5554 Node* src_start = array_element_address(src, offset, src_elem);
5555
5556 // static final int[] byteTable in class CRC32C
5557 Node* table = get_table_from_crc32c_class(callee()->holder());
5558 Node* table_start = array_element_address(table, intcon(0), T_INT);
5559
5560 // We assume that range check is done by caller.
5561 // TODO: generate range check (offset+length < src.length) in debug VM.
5562
5563 // Call the stub.
5564 address stubAddr = StubRoutines::updateBytesCRC32C();
5565 const char *stubName = "updateBytesCRC32C";
5566
5567 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5568 stubAddr, stubName, TypePtr::BOTTOM,
5569 crc, src_start, length, table_start);
5570 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5571 set_result(result);
5572 return true;
5573 }
5574
5575 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5576 //
5577 // Calculate CRC32C for DirectByteBuffer.
5578 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5579 //
5580 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5581 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5582 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5583 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5584 // no receiver since it is a static method
5585 Node* crc = argument(0); // type: int
5586 Node* src = argument(1); // type: long
5587 Node* offset = argument(3); // type: int
5588 Node* end = argument(4); // type: int
5589
5590 Node* length = _gvn.transform(new SubINode(end, offset));
5591
5592 src = ConvL2X(src); // adjust Java long to machine word
5593 Node* base = _gvn.transform(new CastX2PNode(src));
5594 offset = ConvI2X(offset);
5595
5596 // 'src_start' points to src array + scaled offset
5597 Node* src_start = basic_plus_adr(top(), base, offset);
5598
5599 // static final int[] byteTable in class CRC32C
5600 Node* table = get_table_from_crc32c_class(callee()->holder());
5601 Node* table_start = array_element_address(table, intcon(0), T_INT);
5602
5603 // Call the stub.
5604 address stubAddr = StubRoutines::updateBytesCRC32C();
5605 const char *stubName = "updateBytesCRC32C";
5606
5607 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5608 stubAddr, stubName, TypePtr::BOTTOM,
5609 crc, src_start, length, table_start);
5610 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5611 set_result(result);
5612 return true;
5613 }
5614
5615 //------------------------------inline_updateBytesAdler32----------------------
5616 //
5617 // Calculate Adler32 checksum for byte[] array.
5618 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5619 //
5620 bool LibraryCallKit::inline_updateBytesAdler32() {
5621 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5622 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5623 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5624 // no receiver since it is static method
5625 Node* crc = argument(0); // type: int
5626 Node* src = argument(1); // type: oop
5627 Node* offset = argument(2); // type: int
5628 Node* length = argument(3); // type: int
5629
5630 const Type* src_type = src->Value(&_gvn);
5631 const TypeAryPtr* top_src = src_type->isa_aryptr();
5632 if (top_src == NULL || top_src->klass() == NULL) {
5633 // failed array check
5634 return false;
5635 }
5636
5637 // Figure out the size and type of the elements we will be copying.
5638 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5639 if (src_elem != T_BYTE) {
5640 return false;
5641 }
5642
5643 // 'src_start' points to src array + scaled offset
5644 Node* src_start = array_element_address(src, offset, src_elem);
5645
5646 // We assume that range check is done by caller.
5647 // TODO: generate range check (offset+length < src.length) in debug VM.
5648
5649 // Call the stub.
5650 address stubAddr = StubRoutines::updateBytesAdler32();
5651 const char *stubName = "updateBytesAdler32";
5652
5653 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5654 stubAddr, stubName, TypePtr::BOTTOM,
5655 crc, src_start, length);
5656 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5657 set_result(result);
5658 return true;
5659 }
5660
5661 //------------------------------inline_updateByteBufferAdler32---------------
5662 //
5663 // Calculate Adler32 checksum for DirectByteBuffer.
5664 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5665 //
5666 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5667 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5668 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5669 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5670 // no receiver since it is static method
5671 Node* crc = argument(0); // type: int
5672 Node* src = argument(1); // type: long
5673 Node* offset = argument(3); // type: int
5674 Node* length = argument(4); // type: int
5675
5676 src = ConvL2X(src); // adjust Java long to machine word
5677 Node* base = _gvn.transform(new CastX2PNode(src));
5678 offset = ConvI2X(offset);
5679
5680 // 'src_start' points to src array + scaled offset
5681 Node* src_start = basic_plus_adr(top(), base, offset);
5682
5683 // Call the stub.
5684 address stubAddr = StubRoutines::updateBytesAdler32();
5685 const char *stubName = "updateBytesAdler32";
5686
5687 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5688 stubAddr, stubName, TypePtr::BOTTOM,
5689 crc, src_start, length);
5690
5691 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5692 set_result(result);
5693 return true;
5694 }
5695
5696 //----------------------------inline_reference_get----------------------------
5697 // public T java.lang.ref.Reference.get();
5698 bool LibraryCallKit::inline_reference_get() {
5699 const int referent_offset = java_lang_ref_Reference::referent_offset;
5700 guarantee(referent_offset > 0, "should have already been set");
5701
5702 // Get the argument:
5703 Node* reference_obj = null_check_receiver();
5704 if (stopped()) return true;
5705
5706 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5707
5708 ciInstanceKlass* klass = env()->Object_klass();
5709 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5710
5711 Node* no_ctrl = NULL;
5712 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5713
5714 // Use the pre-barrier to record the value in the referent field
5715 pre_barrier(false /* do_load */,
5716 control(),
5717 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5718 result /* pre_val */,
5719 T_OBJECT);
5720
5721 // Add memory barrier to prevent commoning reads from this field
5722 // across safepoint since GC can change its value.
5723 insert_mem_bar(Op_MemBarCPUOrder);
5724
5725 set_result(result);
5726 return true;
5727 }
5728
5729
5730 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5731 bool is_exact=true, bool is_static=false,
5732 ciInstanceKlass * fromKls=NULL) {
5733 if (fromKls == NULL) {
5734 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5735 assert(tinst != NULL, "obj is null");
5736 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5737 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5738 fromKls = tinst->klass()->as_instance_klass();
5739 } else {
5740 assert(is_static, "only for static field access");
5741 }
5742 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5743 ciSymbol::make(fieldTypeString),
5744 is_static);
5745
5746 assert (field != NULL, "undefined field");
5747 if (field == NULL) return (Node *) NULL;
5748
5749 if (is_static) {
5750 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5751 fromObj = makecon(tip);
5752 }
5753
5754 // Next code copied from Parse::do_get_xxx():
5755
5756 // Compute address and memory type.
5757 int offset = field->offset_in_bytes();
5758 bool is_vol = field->is_volatile();
5759 ciType* field_klass = field->type();
5760 assert(field_klass->is_loaded(), "should be loaded");
5761 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5762 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5763 BasicType bt = field->layout_type();
5764
5765 // Build the resultant type of the load
5766 const Type *type;
5767 if (bt == T_OBJECT) {
5768 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5769 } else {
5770 type = Type::get_const_basic_type(bt);
5771 }
5772
5773 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5774 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5775 }
5776 // Build the load.
5777 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5778 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5779 // If reference is volatile, prevent following memory ops from
5780 // floating up past the volatile read. Also prevents commoning
5781 // another volatile read.
5782 if (is_vol) {
5783 // Memory barrier includes bogus read of value to force load BEFORE membar
5784 insert_mem_bar(Op_MemBarAcquire, loadedField);
5785 }
5786 return loadedField;
5787 }
5788
5789 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5790 bool is_exact = true, bool is_static = false,
5791 ciInstanceKlass * fromKls = NULL) {
5792 if (fromKls == NULL) {
5793 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5794 assert(tinst != NULL, "obj is null");
5795 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5796 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5797 fromKls = tinst->klass()->as_instance_klass();
5798 }
5799 else {
5800 assert(is_static, "only for static field access");
5801 }
5802 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5803 ciSymbol::make(fieldTypeString),
5804 is_static);
5805
5806 assert(field != NULL, "undefined field");
5807 assert(!field->is_volatile(), "not defined for volatile fields");
5808
5809 if (is_static) {
5810 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5811 fromObj = makecon(tip);
5812 }
5813
5814 // Next code copied from Parse::do_get_xxx():
5815
5816 // Compute address and memory type.
5817 int offset = field->offset_in_bytes();
5818 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5819
5820 return adr;
5821 }
5822
5823 //------------------------------inline_aescrypt_Block-----------------------
5824 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5825 address stubAddr = NULL;
5826 const char *stubName;
5827 assert(UseAES, "need AES instruction support");
5828
5829 switch(id) {
5830 case vmIntrinsics::_aescrypt_encryptBlock:
5831 stubAddr = StubRoutines::aescrypt_encryptBlock();
5832 stubName = "aescrypt_encryptBlock";
5833 break;
5834 case vmIntrinsics::_aescrypt_decryptBlock:
5835 stubAddr = StubRoutines::aescrypt_decryptBlock();
5836 stubName = "aescrypt_decryptBlock";
5837 break;
5838 }
5839 if (stubAddr == NULL) return false;
5840
5841 Node* aescrypt_object = argument(0);
5842 Node* src = argument(1);
5843 Node* src_offset = argument(2);
5844 Node* dest = argument(3);
5845 Node* dest_offset = argument(4);
5846
5847 // (1) src and dest are arrays.
5848 const Type* src_type = src->Value(&_gvn);
5849 const Type* dest_type = dest->Value(&_gvn);
5850 const TypeAryPtr* top_src = src_type->isa_aryptr();
5851 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5852 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5853
5854 // for the quick and dirty code we will skip all the checks.
5855 // we are just trying to get the call to be generated.
5856 Node* src_start = src;
5857 Node* dest_start = dest;
5858 if (src_offset != NULL || dest_offset != NULL) {
5859 assert(src_offset != NULL && dest_offset != NULL, "");
5860 src_start = array_element_address(src, src_offset, T_BYTE);
5861 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5862 }
5863
5864 // now need to get the start of its expanded key array
5865 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5866 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5867 if (k_start == NULL) return false;
5868
5869 if (Matcher::pass_original_key_for_aes()) {
5870 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5871 // compatibility issues between Java key expansion and SPARC crypto instructions
5872 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5873 if (original_k_start == NULL) return false;
5874
5875 // Call the stub.
5876 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5877 stubAddr, stubName, TypePtr::BOTTOM,
5878 src_start, dest_start, k_start, original_k_start);
5879 } else {
5880 // Call the stub.
5881 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5882 stubAddr, stubName, TypePtr::BOTTOM,
5883 src_start, dest_start, k_start);
5884 }
5885
5886 return true;
5887 }
5888
5889 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5890 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5891 address stubAddr = NULL;
5892 const char *stubName = NULL;
5893
5894 assert(UseAES, "need AES instruction support");
5895
5896 switch(id) {
5897 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5898 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5899 stubName = "cipherBlockChaining_encryptAESCrypt";
5900 break;
5901 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5902 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5903 stubName = "cipherBlockChaining_decryptAESCrypt";
5904 break;
5905 }
5906 if (stubAddr == NULL) return false;
5907
5908 Node* cipherBlockChaining_object = argument(0);
5909 Node* src = argument(1);
5910 Node* src_offset = argument(2);
5911 Node* len = argument(3);
5912 Node* dest = argument(4);
5913 Node* dest_offset = argument(5);
5914
5915 // (1) src and dest are arrays.
5916 const Type* src_type = src->Value(&_gvn);
5917 const Type* dest_type = dest->Value(&_gvn);
5918 const TypeAryPtr* top_src = src_type->isa_aryptr();
5919 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5920 assert (top_src != NULL && top_src->klass() != NULL
5921 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5922
5923 // checks are the responsibility of the caller
5924 Node* src_start = src;
5925 Node* dest_start = dest;
5926 if (src_offset != NULL || dest_offset != NULL) {
5927 assert(src_offset != NULL && dest_offset != NULL, "");
5928 src_start = array_element_address(src, src_offset, T_BYTE);
5929 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5930 }
5931
5932 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5933 // (because of the predicated logic executed earlier).
5934 // so we cast it here safely.
5935 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5936
5937 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5938 if (embeddedCipherObj == NULL) return false;
5939
5940 // cast it to what we know it will be at runtime
5941 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5942 assert(tinst != NULL, "CBC obj is null");
5943 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5944 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5945 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5946
5947 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5948 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5949 const TypeOopPtr* xtype = aklass->as_instance_type();
5950 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5951 aescrypt_object = _gvn.transform(aescrypt_object);
5952
5953 // we need to get the start of the aescrypt_object's expanded key array
5954 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5955 if (k_start == NULL) return false;
5956
5957 // similarly, get the start address of the r vector
5958 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5959 if (objRvec == NULL) return false;
5960 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5961
5962 Node* cbcCrypt;
5963 if (Matcher::pass_original_key_for_aes()) {
5964 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5965 // compatibility issues between Java key expansion and SPARC crypto instructions
5966 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5967 if (original_k_start == NULL) return false;
5968
5969 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5970 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5971 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5972 stubAddr, stubName, TypePtr::BOTTOM,
5973 src_start, dest_start, k_start, r_start, len, original_k_start);
5974 } else {
5975 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5976 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5977 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5978 stubAddr, stubName, TypePtr::BOTTOM,
5979 src_start, dest_start, k_start, r_start, len);
5980 }
5981
5982 // return cipher length (int)
5983 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5984 set_result(retvalue);
5985 return true;
5986 }
5987
5988 //------------------------------inline_counterMode_AESCrypt-----------------------
5989 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
5990 assert(UseAES, "need AES instruction support");
5991 if (!UseAESCTRIntrinsics) return false;
5992
5993 address stubAddr = NULL;
5994 const char *stubName = NULL;
5995 if (id == vmIntrinsics::_counterMode_AESCrypt) {
5996 stubAddr = StubRoutines::counterMode_AESCrypt();
5997 stubName = "counterMode_AESCrypt";
5998 }
5999 if (stubAddr == NULL) return false;
6000
6001 Node* counterMode_object = argument(0);
6002 Node* src = argument(1);
6003 Node* src_offset = argument(2);
6004 Node* len = argument(3);
6005 Node* dest = argument(4);
6006 Node* dest_offset = argument(5);
6007
6008 // (1) src and dest are arrays.
6009 const Type* src_type = src->Value(&_gvn);
6010 const Type* dest_type = dest->Value(&_gvn);
6011 const TypeAryPtr* top_src = src_type->isa_aryptr();
6012 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6013 assert(top_src != NULL && top_src->klass() != NULL &&
6014 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6015
6016 // checks are the responsibility of the caller
6017 Node* src_start = src;
6018 Node* dest_start = dest;
6019 if (src_offset != NULL || dest_offset != NULL) {
6020 assert(src_offset != NULL && dest_offset != NULL, "");
6021 src_start = array_element_address(src, src_offset, T_BYTE);
6022 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6023 }
6024
6025 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6026 // (because of the predicated logic executed earlier).
6027 // so we cast it here safely.
6028 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6029 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6030 if (embeddedCipherObj == NULL) return false;
6031 // cast it to what we know it will be at runtime
6032 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6033 assert(tinst != NULL, "CTR obj is null");
6034 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6035 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6036 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6037 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6038 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6039 const TypeOopPtr* xtype = aklass->as_instance_type();
6040 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6041 aescrypt_object = _gvn.transform(aescrypt_object);
6042 // we need to get the start of the aescrypt_object's expanded key array
6043 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6044 if (k_start == NULL) return false;
6045 // similarly, get the start address of the r vector
6046 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6047 if (obj_counter == NULL) return false;
6048 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6049
6050 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6051 if (saved_encCounter == NULL) return false;
6052 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6053 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6054
6055 Node* ctrCrypt;
6056 if (Matcher::pass_original_key_for_aes()) {
6057 // no SPARC version for AES/CTR intrinsics now.
6058 return false;
6059 }
6060 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6061 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6062 OptoRuntime::counterMode_aescrypt_Type(),
6063 stubAddr, stubName, TypePtr::BOTTOM,
6064 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6065
6066 // return cipher length (int)
6067 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6068 set_result(retvalue);
6069 return true;
6070 }
6071
6072 //------------------------------get_key_start_from_aescrypt_object-----------------------
6073 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6074 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6075 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6076 if (objAESCryptKey == NULL) return (Node *) NULL;
6077
6078 // now have the array, need to get the start address of the K array
6079 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6080 return k_start;
6081 }
6082
6083 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6084 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6085 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6086 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6087 if (objAESCryptKey == NULL) return (Node *) NULL;
6088
6089 // now have the array, need to get the start address of the lastKey array
6090 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6091 return original_k_start;
6092 }
6093
6094 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6095 // Return node representing slow path of predicate check.
6096 // the pseudo code we want to emulate with this predicate is:
6097 // for encryption:
6098 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6099 // for decryption:
6100 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6101 // note cipher==plain is more conservative than the original java code but that's OK
6102 //
6103 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6104 // The receiver was checked for NULL already.
6105 Node* objCBC = argument(0);
6106
6107 // Load embeddedCipher field of CipherBlockChaining object.
6108 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6109
6110 // get AESCrypt klass for instanceOf check
6111 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6112 // will have same classloader as CipherBlockChaining object
6113 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6114 assert(tinst != NULL, "CBCobj is null");
6115 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6116
6117 // we want to do an instanceof comparison against the AESCrypt class
6118 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6119 if (!klass_AESCrypt->is_loaded()) {
6120 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6121 Node* ctrl = control();
6122 set_control(top()); // no regular fast path
6123 return ctrl;
6124 }
6125 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6126
6127 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6128 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6129 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6130
6131 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6132
6133 // for encryption, we are done
6134 if (!decrypting)
6135 return instof_false; // even if it is NULL
6136
6137 // for decryption, we need to add a further check to avoid
6138 // taking the intrinsic path when cipher and plain are the same
6139 // see the original java code for why.
6140 RegionNode* region = new RegionNode(3);
6141 region->init_req(1, instof_false);
6142 Node* src = argument(1);
6143 Node* dest = argument(4);
6144 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6145 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6146 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6147 region->init_req(2, src_dest_conjoint);
6148
6149 record_for_igvn(region);
6150 return _gvn.transform(region);
6151 }
6152
6153 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6154 // Return node representing slow path of predicate check.
6155 // the pseudo code we want to emulate with this predicate is:
6156 // for encryption:
6157 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6158 // for decryption:
6159 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6160 // note cipher==plain is more conservative than the original java code but that's OK
6161 //
6162
6163 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6164 // The receiver was checked for NULL already.
6165 Node* objCTR = argument(0);
6166
6167 // Load embeddedCipher field of CipherBlockChaining object.
6168 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6169
6170 // get AESCrypt klass for instanceOf check
6171 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6172 // will have same classloader as CipherBlockChaining object
6173 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6174 assert(tinst != NULL, "CTRobj is null");
6175 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6176
6177 // we want to do an instanceof comparison against the AESCrypt class
6178 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6179 if (!klass_AESCrypt->is_loaded()) {
6180 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6181 Node* ctrl = control();
6182 set_control(top()); // no regular fast path
6183 return ctrl;
6184 }
6185
6186 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6187 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6188 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6189 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6190 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6191
6192 return instof_false; // even if it is NULL
6193 }
6194
6195 //------------------------------inline_ghash_processBlocks
6196 bool LibraryCallKit::inline_ghash_processBlocks() {
6197 address stubAddr;
6198 const char *stubName;
6199 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6200
6201 stubAddr = StubRoutines::ghash_processBlocks();
6202 stubName = "ghash_processBlocks";
6203
6204 Node* data = argument(0);
6205 Node* offset = argument(1);
6206 Node* len = argument(2);
6207 Node* state = argument(3);
6208 Node* subkeyH = argument(4);
6209
6210 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6211 assert(state_start, "state is NULL");
6212 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6213 assert(subkeyH_start, "subkeyH is NULL");
6214 Node* data_start = array_element_address(data, offset, T_BYTE);
6215 assert(data_start, "data is NULL");
6216
6217 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6218 OptoRuntime::ghash_processBlocks_Type(),
6219 stubAddr, stubName, TypePtr::BOTTOM,
6220 state_start, subkeyH_start, data_start, len);
6221 return true;
6222 }
6223
6224 //------------------------------inline_sha_implCompress-----------------------
6225 //
6226 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6227 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6228 //
6229 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6230 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6231 //
6232 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6233 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6234 //
6235 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6236 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6237
6238 Node* sha_obj = argument(0);
6239 Node* src = argument(1); // type oop
6240 Node* ofs = argument(2); // type int
6241
6242 const Type* src_type = src->Value(&_gvn);
6243 const TypeAryPtr* top_src = src_type->isa_aryptr();
6244 if (top_src == NULL || top_src->klass() == NULL) {
6245 // failed array check
6246 return false;
6247 }
6248 // Figure out the size and type of the elements we will be copying.
6249 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6250 if (src_elem != T_BYTE) {
6251 return false;
6252 }
6253 // 'src_start' points to src array + offset
6254 Node* src_start = array_element_address(src, ofs, src_elem);
6255 Node* state = NULL;
6256 address stubAddr;
6257 const char *stubName;
6258
6259 switch(id) {
6260 case vmIntrinsics::_sha_implCompress:
6261 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6262 state = get_state_from_sha_object(sha_obj);
6263 stubAddr = StubRoutines::sha1_implCompress();
6264 stubName = "sha1_implCompress";
6265 break;
6266 case vmIntrinsics::_sha2_implCompress:
6267 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6268 state = get_state_from_sha_object(sha_obj);
6269 stubAddr = StubRoutines::sha256_implCompress();
6270 stubName = "sha256_implCompress";
6271 break;
6272 case vmIntrinsics::_sha5_implCompress:
6273 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6274 state = get_state_from_sha5_object(sha_obj);
6275 stubAddr = StubRoutines::sha512_implCompress();
6276 stubName = "sha512_implCompress";
6277 break;
6278 default:
6279 fatal_unexpected_iid(id);
6280 return false;
6281 }
6282 if (state == NULL) return false;
6283
6284 // Call the stub.
6285 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6286 stubAddr, stubName, TypePtr::BOTTOM,
6287 src_start, state);
6288
6289 return true;
6290 }
6291
6292 //------------------------------inline_digestBase_implCompressMB-----------------------
6293 //
6294 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6295 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6296 //
6297 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6298 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6299 "need SHA1/SHA256/SHA512 instruction support");
6300 assert((uint)predicate < 3, "sanity");
6301 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6302
6303 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6304 Node* src = argument(1); // byte[] array
6305 Node* ofs = argument(2); // type int
6306 Node* limit = argument(3); // type int
6307
6308 const Type* src_type = src->Value(&_gvn);
6309 const TypeAryPtr* top_src = src_type->isa_aryptr();
6310 if (top_src == NULL || top_src->klass() == NULL) {
6311 // failed array check
6312 return false;
6313 }
6314 // Figure out the size and type of the elements we will be copying.
6315 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6316 if (src_elem != T_BYTE) {
6317 return false;
6318 }
6319 // 'src_start' points to src array + offset
6320 Node* src_start = array_element_address(src, ofs, src_elem);
6321
6322 const char* klass_SHA_name = NULL;
6323 const char* stub_name = NULL;
6324 address stub_addr = NULL;
6325 bool long_state = false;
6326
6327 switch (predicate) {
6328 case 0:
6329 if (UseSHA1Intrinsics) {
6330 klass_SHA_name = "sun/security/provider/SHA";
6331 stub_name = "sha1_implCompressMB";
6332 stub_addr = StubRoutines::sha1_implCompressMB();
6333 }
6334 break;
6335 case 1:
6336 if (UseSHA256Intrinsics) {
6337 klass_SHA_name = "sun/security/provider/SHA2";
6338 stub_name = "sha256_implCompressMB";
6339 stub_addr = StubRoutines::sha256_implCompressMB();
6340 }
6341 break;
6342 case 2:
6343 if (UseSHA512Intrinsics) {
6344 klass_SHA_name = "sun/security/provider/SHA5";
6345 stub_name = "sha512_implCompressMB";
6346 stub_addr = StubRoutines::sha512_implCompressMB();
6347 long_state = true;
6348 }
6349 break;
6350 default:
6351 fatal("unknown SHA intrinsic predicate: %d", predicate);
6352 }
6353 if (klass_SHA_name != NULL) {
6354 // get DigestBase klass to lookup for SHA klass
6355 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6356 assert(tinst != NULL, "digestBase_obj is not instance???");
6357 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6358
6359 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6360 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6361 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6362 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6363 }
6364 return false;
6365 }
6366 //------------------------------inline_sha_implCompressMB-----------------------
6367 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6368 bool long_state, address stubAddr, const char *stubName,
6369 Node* src_start, Node* ofs, Node* limit) {
6370 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6371 const TypeOopPtr* xtype = aklass->as_instance_type();
6372 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6373 sha_obj = _gvn.transform(sha_obj);
6374
6375 Node* state;
6376 if (long_state) {
6377 state = get_state_from_sha5_object(sha_obj);
6378 } else {
6379 state = get_state_from_sha_object(sha_obj);
6380 }
6381 if (state == NULL) return false;
6382
6383 // Call the stub.
6384 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6385 OptoRuntime::digestBase_implCompressMB_Type(),
6386 stubAddr, stubName, TypePtr::BOTTOM,
6387 src_start, state, ofs, limit);
6388 // return ofs (int)
6389 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6390 set_result(result);
6391
6392 return true;
6393 }
6394
6395 //------------------------------get_state_from_sha_object-----------------------
6396 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6397 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6398 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6399 if (sha_state == NULL) return (Node *) NULL;
6400
6401 // now have the array, need to get the start address of the state array
6402 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6403 return state;
6404 }
6405
6406 //------------------------------get_state_from_sha5_object-----------------------
6407 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6408 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6409 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6410 if (sha_state == NULL) return (Node *) NULL;
6411
6412 // now have the array, need to get the start address of the state array
6413 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6414 return state;
6415 }
6416
6417 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6418 // Return node representing slow path of predicate check.
6419 // the pseudo code we want to emulate with this predicate is:
6420 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6421 //
6422 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6423 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6424 "need SHA1/SHA256/SHA512 instruction support");
6425 assert((uint)predicate < 3, "sanity");
6426
6427 // The receiver was checked for NULL already.
6428 Node* digestBaseObj = argument(0);
6429
6430 // get DigestBase klass for instanceOf check
6431 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6432 assert(tinst != NULL, "digestBaseObj is null");
6433 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6434
6435 const char* klass_SHA_name = NULL;
6436 switch (predicate) {
6437 case 0:
6438 if (UseSHA1Intrinsics) {
6439 // we want to do an instanceof comparison against the SHA class
6440 klass_SHA_name = "sun/security/provider/SHA";
6441 }
6442 break;
6443 case 1:
6444 if (UseSHA256Intrinsics) {
6445 // we want to do an instanceof comparison against the SHA2 class
6446 klass_SHA_name = "sun/security/provider/SHA2";
6447 }
6448 break;
6449 case 2:
6450 if (UseSHA512Intrinsics) {
6451 // we want to do an instanceof comparison against the SHA5 class
6452 klass_SHA_name = "sun/security/provider/SHA5";
6453 }
6454 break;
6455 default:
6456 fatal("unknown SHA intrinsic predicate: %d", predicate);
6457 }
6458
6459 ciKlass* klass_SHA = NULL;
6460 if (klass_SHA_name != NULL) {
6461 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6462 }
6463 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6464 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6465 Node* ctrl = control();
6466 set_control(top()); // no intrinsic path
6467 return ctrl;
6468 }
6469 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6470
6471 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6472 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6473 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6474 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6475
6476 return instof_false; // even if it is NULL
6477 }
6478
6479 bool LibraryCallKit::inline_profileBoolean() {
6480 Node* counts = argument(1);
6481 const TypeAryPtr* ary = NULL;
6482 ciArray* aobj = NULL;
6483 if (counts->is_Con()
6484 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6485 && (aobj = ary->const_oop()->as_array()) != NULL
6486 && (aobj->length() == 2)) {
6487 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6488 jint false_cnt = aobj->element_value(0).as_int();
6489 jint true_cnt = aobj->element_value(1).as_int();
6490
6491 if (C->log() != NULL) {
6492 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6493 false_cnt, true_cnt);
6494 }
6495
6496 if (false_cnt + true_cnt == 0) {
6497 // According to profile, never executed.
6498 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6499 Deoptimization::Action_reinterpret);
6500 return true;
6501 }
6502
6503 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6504 // is a number of each value occurrences.
6505 Node* result = argument(0);
6506 if (false_cnt == 0 || true_cnt == 0) {
6507 // According to profile, one value has been never seen.
6508 int expected_val = (false_cnt == 0) ? 1 : 0;
6509
6510 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6511 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6512
6513 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6514 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6515 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6516
6517 { // Slow path: uncommon trap for never seen value and then reexecute
6518 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6519 // the value has been seen at least once.
6520 PreserveJVMState pjvms(this);
6521 PreserveReexecuteState preexecs(this);
6522 jvms()->set_should_reexecute(true);
6523
6524 set_control(slow_path);
6525 set_i_o(i_o());
6526
6527 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6528 Deoptimization::Action_reinterpret);
6529 }
6530 // The guard for never seen value enables sharpening of the result and
6531 // returning a constant. It allows to eliminate branches on the same value
6532 // later on.
6533 set_control(fast_path);
6534 result = intcon(expected_val);
6535 }
6536 // Stop profiling.
6537 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6538 // By replacing method body with profile data (represented as ProfileBooleanNode
6539 // on IR level) we effectively disable profiling.
6540 // It enables full speed execution once optimized code is generated.
6541 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6542 C->record_for_igvn(profile);
6543 set_result(profile);
6544 return true;
6545 } else {
6546 // Continue profiling.
6547 // Profile data isn't available at the moment. So, execute method's bytecode version.
6548 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6549 // is compiled and counters aren't available since corresponding MethodHandle
6550 // isn't a compile-time constant.
6551 return false;
6552 }
6553 }
6554
6555 bool LibraryCallKit::inline_isCompileConstant() {
6556 Node* n = argument(0);
6557 set_result(n->is_Con() ? intcon(1) : intcon(0));
6558 return true;
6559 }
6560
6561 bool LibraryCallKit::inline_deoptimize() {
6562 assert(WhiteBoxAPI, "");
6563 PreserveReexecuteState preexecs(this);
6564 jvms()->set_should_reexecute(false);
6565 uncommon_trap(Deoptimization::Reason_intrinsic,
6566 Deoptimization::Action_none);
6567 return true;
6568 }