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