1 /*
2 * Copyright (c) 1999, 2015, 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
205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
206 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
207 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
208 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
209 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
210 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
211 bool inline_string_indexOfChar();
212 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_toBytesU();
214 bool inline_string_getCharsU();
215 bool inline_string_copy(bool compress);
216 bool inline_string_char_access(bool is_store);
217 Node* round_double_node(Node* n);
218 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
219 bool inline_math_native(vmIntrinsics::ID id);
220 bool inline_trig(vmIntrinsics::ID id);
221 bool inline_math(vmIntrinsics::ID id);
222 template <typename OverflowOp>
223 bool inline_math_overflow(Node* arg1, Node* arg2);
224 void inline_math_mathExact(Node* math, Node* test);
225 bool inline_math_addExactI(bool is_increment);
226 bool inline_math_addExactL(bool is_increment);
227 bool inline_math_multiplyExactI();
228 bool inline_math_multiplyExactL();
229 bool inline_math_negateExactI();
230 bool inline_math_negateExactL();
231 bool inline_math_subtractExactI(bool is_decrement);
232 bool inline_math_subtractExactL(bool is_decrement);
233 bool inline_pow();
234 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
235 bool inline_min_max(vmIntrinsics::ID id);
236 bool inline_notify(vmIntrinsics::ID id);
237 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
238 // This returns Type::AnyPtr, RawPtr, or OopPtr.
239 int classify_unsafe_addr(Node* &base, Node* &offset);
240 Node* make_unsafe_address(Node* base, Node* offset);
241 // Helper for inline_unsafe_access.
242 // Generates the guards that check whether the result of
243 // Unsafe.getObject should be recorded in an SATB log buffer.
244 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
245 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
246 static bool klass_needs_init_guard(Node* kls);
247 bool inline_unsafe_allocate();
248 bool inline_unsafe_copyMemory();
249 bool inline_native_currentThread();
250 #ifdef TRACE_HAVE_INTRINSICS
251 bool inline_native_classID();
252 bool inline_native_threadID();
253 #endif
254 bool inline_native_time_funcs(address method, const char* funcName);
255 bool inline_native_isInterrupted();
256 bool inline_native_Class_query(vmIntrinsics::ID id);
257 bool inline_native_subtype_check();
258
259 bool inline_native_newArray();
260 bool inline_native_getLength();
261 bool inline_array_copyOf(bool is_copyOfRange);
262 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
263 bool inline_objects_checkIndex();
264 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
265 bool inline_native_clone(bool is_virtual);
266 bool inline_native_Reflection_getCallerClass();
267 // Helper function for inlining native object hash method
268 bool inline_native_hashcode(bool is_virtual, bool is_static);
269 bool inline_native_getClass();
270
271 // Helper functions for inlining arraycopy
272 bool inline_arraycopy();
273 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
274 RegionNode* slow_region);
275 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
276 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
277
278 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
279 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
280 bool inline_unsafe_ordered_store(BasicType type);
281 bool inline_unsafe_fence(vmIntrinsics::ID id);
282 bool inline_fp_conversions(vmIntrinsics::ID id);
283 bool inline_number_methods(vmIntrinsics::ID id);
284 bool inline_reference_get();
285 bool inline_Class_cast();
286 bool inline_aescrypt_Block(vmIntrinsics::ID id);
287 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
288 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
289 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
290 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
291 bool inline_ghash_processBlocks();
292 bool inline_sha_implCompress(vmIntrinsics::ID id);
293 bool inline_digestBase_implCompressMB(int predicate);
294 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
295 bool long_state, address stubAddr, const char *stubName,
296 Node* src_start, Node* ofs, Node* limit);
297 Node* get_state_from_sha_object(Node *sha_object);
298 Node* get_state_from_sha5_object(Node *sha_object);
299 Node* inline_digestBase_implCompressMB_predicate(int predicate);
300 bool inline_encodeISOArray();
301 bool inline_updateCRC32();
302 bool inline_updateBytesCRC32();
303 bool inline_updateByteBufferCRC32();
304 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
305 bool inline_updateBytesCRC32C();
306 bool inline_updateDirectByteBufferCRC32C();
307 bool inline_updateBytesAdler32();
308 bool inline_updateByteBufferAdler32();
309 bool inline_multiplyToLen();
310 bool inline_hasNegatives();
311 bool inline_squareToLen();
312 bool inline_mulAdd();
313 bool inline_montgomeryMultiply();
314 bool inline_montgomerySquare();
315
316 bool inline_profileBoolean();
317 bool inline_isCompileConstant();
318 };
319
320 //---------------------------make_vm_intrinsic----------------------------
321 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
322 vmIntrinsics::ID id = m->intrinsic_id();
323 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
324
325 if (!m->is_loaded()) {
326 // Do not attempt to inline unloaded methods.
327 return NULL;
328 }
329
330 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
331 bool is_available = false;
332
333 {
334 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
335 // the compiler must transition to '_thread_in_vm' state because both
336 // methods access VM-internal data.
337 VM_ENTRY_MARK;
338 methodHandle mh(THREAD, m->get_Method());
339 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
340 !C->directive()->is_intrinsic_disabled(mh) &&
341 !vmIntrinsics::is_disabled_by_flags(mh);
342
343 }
344
345 if (is_available) {
346 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
347 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
348 return new LibraryIntrinsic(m, is_virtual,
349 vmIntrinsics::predicates_needed(id),
350 vmIntrinsics::does_virtual_dispatch(id),
351 (vmIntrinsics::ID) id);
352 } else {
353 return NULL;
354 }
355 }
356
357 //----------------------register_library_intrinsics-----------------------
358 // Initialize this file's data structures, for each Compile instance.
359 void Compile::register_library_intrinsics() {
360 // Nothing to do here.
361 }
362
363 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
364 LibraryCallKit kit(jvms, this);
365 Compile* C = kit.C;
366 int nodes = C->unique();
367 #ifndef PRODUCT
368 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
369 char buf[1000];
370 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
371 tty->print_cr("Intrinsic %s", str);
372 }
373 #endif
374 ciMethod* callee = kit.callee();
375 const int bci = kit.bci();
376
377 // Try to inline the intrinsic.
378 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
379 kit.try_to_inline(_last_predicate)) {
380 if (C->print_intrinsics() || C->print_inlining()) {
381 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
382 }
383 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
384 if (C->log()) {
385 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
386 vmIntrinsics::name_at(intrinsic_id()),
387 (is_virtual() ? " virtual='1'" : ""),
388 C->unique() - nodes);
389 }
390 // Push the result from the inlined method onto the stack.
391 kit.push_result();
392 C->print_inlining_update(this);
393 return kit.transfer_exceptions_into_jvms();
394 }
395
396 // The intrinsic bailed out
397 if (C->print_intrinsics() || C->print_inlining()) {
398 if (jvms->has_method()) {
399 // Not a root compile.
400 const char* msg;
401 if (callee->intrinsic_candidate()) {
402 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
403 } else {
404 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
405 : "failed to inline (intrinsic), method not annotated";
406 }
407 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
408 } else {
409 // Root compile
410 tty->print("Did not generate intrinsic %s%s at bci:%d in",
411 vmIntrinsics::name_at(intrinsic_id()),
412 (is_virtual() ? " (virtual)" : ""), bci);
413 }
414 }
415 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
416 C->print_inlining_update(this);
417 return NULL;
418 }
419
420 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
421 LibraryCallKit kit(jvms, this);
422 Compile* C = kit.C;
423 int nodes = C->unique();
424 _last_predicate = predicate;
425 #ifndef PRODUCT
426 assert(is_predicated() && predicate < predicates_count(), "sanity");
427 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
428 char buf[1000];
429 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
430 tty->print_cr("Predicate for intrinsic %s", str);
431 }
432 #endif
433 ciMethod* callee = kit.callee();
434 const int bci = kit.bci();
435
436 Node* slow_ctl = kit.try_to_predicate(predicate);
437 if (!kit.failing()) {
438 if (C->print_intrinsics() || C->print_inlining()) {
439 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
440 }
441 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
442 if (C->log()) {
443 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
444 vmIntrinsics::name_at(intrinsic_id()),
445 (is_virtual() ? " virtual='1'" : ""),
446 C->unique() - nodes);
447 }
448 return slow_ctl; // Could be NULL if the check folds.
449 }
450
451 // The intrinsic bailed out
452 if (C->print_intrinsics() || C->print_inlining()) {
453 if (jvms->has_method()) {
454 // Not a root compile.
455 const char* msg = "failed to generate predicate for intrinsic";
456 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
457 } else {
458 // Root compile
459 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
460 vmIntrinsics::name_at(intrinsic_id()),
461 (is_virtual() ? " (virtual)" : ""), bci);
462 }
463 }
464 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
465 return NULL;
466 }
467
468 bool LibraryCallKit::try_to_inline(int predicate) {
469 // Handle symbolic names for otherwise undistinguished boolean switches:
470 const bool is_store = true;
471 const bool is_compress = true;
472 const bool is_native_ptr = true;
473 const bool is_static = true;
474 const bool is_volatile = true;
475
476 if (!jvms()->has_method()) {
477 // Root JVMState has a null method.
478 assert(map()->memory()->Opcode() == Op_Parm, "");
479 // Insert the memory aliasing node
480 set_all_memory(reset_memory());
481 }
482 assert(merged_memory(), "");
483
484
485 switch (intrinsic_id()) {
486 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
487 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
488 case vmIntrinsics::_getClass: return inline_native_getClass();
489
490 case vmIntrinsics::_dsin:
491 case vmIntrinsics::_dcos:
492 case vmIntrinsics::_dtan:
493 case vmIntrinsics::_dabs:
494 case vmIntrinsics::_datan2:
495 case vmIntrinsics::_dsqrt:
496 case vmIntrinsics::_dexp:
497 case vmIntrinsics::_dlog:
498 case vmIntrinsics::_dlog10:
499 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
500
501 case vmIntrinsics::_min:
502 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
503
504 case vmIntrinsics::_notify:
505 case vmIntrinsics::_notifyAll:
506 if (InlineNotify) {
507 return inline_notify(intrinsic_id());
508 }
509 return false;
510
511 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
512 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
513 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
514 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
515 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
516 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
517 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
518 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
519 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
520 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
521 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
522 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
523
524 case vmIntrinsics::_arraycopy: return inline_arraycopy();
525
526 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
527 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
528 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
529 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
530
531 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
532 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
533 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
534 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
535 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
536 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
537 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
538
539 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
540 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
541
542 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
543 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
544 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
545 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
546
547 case vmIntrinsics::_compressStringC:
548 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
549 case vmIntrinsics::_inflateStringC:
550 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
551
552 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
553 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
554 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
555 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
556 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
557 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
558 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
559 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
560 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
561 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
562 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
563 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
564 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
565 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
566 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
567 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
568 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
569 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
570
571 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
572 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
573 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
574 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
575 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
576 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
577 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
578 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
579
580 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
581 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
582 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
583 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
584 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
585 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
586 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
587 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
588
589 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
590 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
591 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
592 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
593 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
594 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
595 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
596 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
597 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
598
599 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
600 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
601 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
602 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
603 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
604 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
605 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
606 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
607 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
608
609 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true);
610 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true);
611 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true);
612 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true);
613
614 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true);
615 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true);
616 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true);
617 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true);
618
619 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
620 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
621 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
622
623 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
624 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
625 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
626
627 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
628 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
629 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
630 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
631 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
632
633 case vmIntrinsics::_loadFence:
634 case vmIntrinsics::_storeFence:
635 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
636
637 case vmIntrinsics::_currentThread: return inline_native_currentThread();
638 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
639
640 #ifdef TRACE_HAVE_INTRINSICS
641 case vmIntrinsics::_classID: return inline_native_classID();
642 case vmIntrinsics::_threadID: return inline_native_threadID();
643 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
644 #endif
645 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
646 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
647 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
648 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
649 case vmIntrinsics::_newArray: return inline_native_newArray();
650 case vmIntrinsics::_getLength: return inline_native_getLength();
651 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
652 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
653 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
654 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
655 case vmIntrinsics::_Objects_checkIndex: return inline_objects_checkIndex();
656 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
657
658 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
659
660 case vmIntrinsics::_isInstance:
661 case vmIntrinsics::_getModifiers:
662 case vmIntrinsics::_isInterface:
663 case vmIntrinsics::_isArray:
664 case vmIntrinsics::_isPrimitive:
665 case vmIntrinsics::_getSuperclass:
666 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
667
668 case vmIntrinsics::_floatToRawIntBits:
669 case vmIntrinsics::_floatToIntBits:
670 case vmIntrinsics::_intBitsToFloat:
671 case vmIntrinsics::_doubleToRawLongBits:
672 case vmIntrinsics::_doubleToLongBits:
673 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
674
675 case vmIntrinsics::_numberOfLeadingZeros_i:
676 case vmIntrinsics::_numberOfLeadingZeros_l:
677 case vmIntrinsics::_numberOfTrailingZeros_i:
678 case vmIntrinsics::_numberOfTrailingZeros_l:
679 case vmIntrinsics::_bitCount_i:
680 case vmIntrinsics::_bitCount_l:
681 case vmIntrinsics::_reverseBytes_i:
682 case vmIntrinsics::_reverseBytes_l:
683 case vmIntrinsics::_reverseBytes_s:
684 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
685
686 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
687
688 case vmIntrinsics::_Reference_get: return inline_reference_get();
689
690 case vmIntrinsics::_Class_cast: return inline_Class_cast();
691
692 case vmIntrinsics::_aescrypt_encryptBlock:
693 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
694
695 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
696 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
697 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
698
699 case vmIntrinsics::_sha_implCompress:
700 case vmIntrinsics::_sha2_implCompress:
701 case vmIntrinsics::_sha5_implCompress:
702 return inline_sha_implCompress(intrinsic_id());
703
704 case vmIntrinsics::_digestBase_implCompressMB:
705 return inline_digestBase_implCompressMB(predicate);
706
707 case vmIntrinsics::_multiplyToLen:
708 return inline_multiplyToLen();
709
710 case vmIntrinsics::_squareToLen:
711 return inline_squareToLen();
712
713 case vmIntrinsics::_mulAdd:
714 return inline_mulAdd();
715
716 case vmIntrinsics::_montgomeryMultiply:
717 return inline_montgomeryMultiply();
718 case vmIntrinsics::_montgomerySquare:
719 return inline_montgomerySquare();
720
721 case vmIntrinsics::_ghash_processBlocks:
722 return inline_ghash_processBlocks();
723
724 case vmIntrinsics::_encodeISOArray:
725 case vmIntrinsics::_encodeByteISOArray:
726 return inline_encodeISOArray();
727
728 case vmIntrinsics::_updateCRC32:
729 return inline_updateCRC32();
730 case vmIntrinsics::_updateBytesCRC32:
731 return inline_updateBytesCRC32();
732 case vmIntrinsics::_updateByteBufferCRC32:
733 return inline_updateByteBufferCRC32();
734
735 case vmIntrinsics::_updateBytesCRC32C:
736 return inline_updateBytesCRC32C();
737 case vmIntrinsics::_updateDirectByteBufferCRC32C:
738 return inline_updateDirectByteBufferCRC32C();
739
740 case vmIntrinsics::_updateBytesAdler32:
741 return inline_updateBytesAdler32();
742 case vmIntrinsics::_updateByteBufferAdler32:
743 return inline_updateByteBufferAdler32();
744
745 case vmIntrinsics::_profileBoolean:
746 return inline_profileBoolean();
747 case vmIntrinsics::_isCompileConstant:
748 return inline_isCompileConstant();
749
750 case vmIntrinsics::_hasNegatives:
751 return inline_hasNegatives();
752
753 default:
754 // If you get here, it may be that someone has added a new intrinsic
755 // to the list in vmSymbols.hpp without implementing it here.
756 #ifndef PRODUCT
757 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
758 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
759 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
760 }
761 #endif
762 return false;
763 }
764 }
765
766 Node* LibraryCallKit::try_to_predicate(int predicate) {
767 if (!jvms()->has_method()) {
768 // Root JVMState has a null method.
769 assert(map()->memory()->Opcode() == Op_Parm, "");
770 // Insert the memory aliasing node
771 set_all_memory(reset_memory());
772 }
773 assert(merged_memory(), "");
774
775 switch (intrinsic_id()) {
776 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
777 return inline_cipherBlockChaining_AESCrypt_predicate(false);
778 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
779 return inline_cipherBlockChaining_AESCrypt_predicate(true);
780 case vmIntrinsics::_digestBase_implCompressMB:
781 return inline_digestBase_implCompressMB_predicate(predicate);
782
783 default:
784 // If you get here, it may be that someone has added a new intrinsic
785 // to the list in vmSymbols.hpp without implementing it here.
786 #ifndef PRODUCT
787 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
788 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
789 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
790 }
791 #endif
792 Node* slow_ctl = control();
793 set_control(top()); // No fast path instrinsic
794 return slow_ctl;
795 }
796 }
797
798 //------------------------------set_result-------------------------------
799 // Helper function for finishing intrinsics.
800 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
801 record_for_igvn(region);
802 set_control(_gvn.transform(region));
803 set_result( _gvn.transform(value));
804 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
805 }
806
807 //------------------------------generate_guard---------------------------
808 // Helper function for generating guarded fast-slow graph structures.
809 // The given 'test', if true, guards a slow path. If the test fails
810 // then a fast path can be taken. (We generally hope it fails.)
811 // In all cases, GraphKit::control() is updated to the fast path.
812 // The returned value represents the control for the slow path.
813 // The return value is never 'top'; it is either a valid control
814 // or NULL if it is obvious that the slow path can never be taken.
815 // Also, if region and the slow control are not NULL, the slow edge
816 // is appended to the region.
817 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
818 if (stopped()) {
819 // Already short circuited.
820 return NULL;
821 }
822
823 // Build an if node and its projections.
824 // If test is true we take the slow path, which we assume is uncommon.
825 if (_gvn.type(test) == TypeInt::ZERO) {
826 // The slow branch is never taken. No need to build this guard.
827 return NULL;
828 }
829
830 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
831
832 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
833 if (if_slow == top()) {
834 // The slow branch is never taken. No need to build this guard.
835 return NULL;
836 }
837
838 if (region != NULL)
839 region->add_req(if_slow);
840
841 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
842 set_control(if_fast);
843
844 return if_slow;
845 }
846
847 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
848 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
849 }
850 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
851 return generate_guard(test, region, PROB_FAIR);
852 }
853
854 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
855 Node* *pos_index) {
856 if (stopped())
857 return NULL; // already stopped
858 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
859 return NULL; // index is already adequately typed
860 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
861 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
862 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
863 if (is_neg != NULL && pos_index != NULL) {
864 // Emulate effect of Parse::adjust_map_after_if.
865 Node* ccast = new CastIINode(index, TypeInt::POS);
866 ccast->set_req(0, control());
867 (*pos_index) = _gvn.transform(ccast);
868 }
869 return is_neg;
870 }
871
872 // Make sure that 'position' is a valid limit index, in [0..length].
873 // There are two equivalent plans for checking this:
874 // A. (offset + copyLength) unsigned<= arrayLength
875 // B. offset <= (arrayLength - copyLength)
876 // We require that all of the values above, except for the sum and
877 // difference, are already known to be non-negative.
878 // Plan A is robust in the face of overflow, if offset and copyLength
879 // are both hugely positive.
880 //
881 // Plan B is less direct and intuitive, but it does not overflow at
882 // all, since the difference of two non-negatives is always
883 // representable. Whenever Java methods must perform the equivalent
884 // check they generally use Plan B instead of Plan A.
885 // For the moment we use Plan A.
886 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
887 Node* subseq_length,
888 Node* array_length,
889 RegionNode* region) {
890 if (stopped())
891 return NULL; // already stopped
892 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
893 if (zero_offset && subseq_length->eqv_uncast(array_length))
894 return NULL; // common case of whole-array copy
895 Node* last = subseq_length;
896 if (!zero_offset) // last += offset
897 last = _gvn.transform(new AddINode(last, offset));
898 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
899 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
900 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
901 return is_over;
902 }
903
904 // Emit range checks for the given String.value byte array
905 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
906 if (stopped()) {
907 return; // already stopped
908 }
909 RegionNode* bailout = new RegionNode(1);
910 record_for_igvn(bailout);
911 if (char_count) {
912 // Convert char count to byte count
913 count = _gvn.transform(new LShiftINode(count, intcon(1)));
914 }
915
916 // Offset and count must not be negative
917 generate_negative_guard(offset, bailout);
918 generate_negative_guard(count, bailout);
919 // Offset + count must not exceed length of array
920 generate_limit_guard(offset, count, load_array_length(array), bailout);
921
922 if (bailout->req() > 1) {
923 PreserveJVMState pjvms(this);
924 set_control(_gvn.transform(bailout));
925 uncommon_trap(Deoptimization::Reason_intrinsic,
926 Deoptimization::Action_maybe_recompile);
927 }
928 }
929
930 //--------------------------generate_current_thread--------------------
931 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
932 ciKlass* thread_klass = env()->Thread_klass();
933 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
934 Node* thread = _gvn.transform(new ThreadLocalNode());
935 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
936 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
937 tls_output = thread;
938 return threadObj;
939 }
940
941
942 //------------------------------make_string_method_node------------------------
943 // Helper method for String intrinsic functions. This version is called with
944 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
945 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
946 // containing the lengths of str1 and str2.
947 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
948 Node* result = NULL;
949 switch (opcode) {
950 case Op_StrIndexOf:
951 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
952 str1_start, cnt1, str2_start, cnt2, ae);
953 break;
954 case Op_StrComp:
955 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
956 str1_start, cnt1, str2_start, cnt2, ae);
957 break;
958 case Op_StrEquals:
959 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
960 str1_start, str2_start, cnt1, ae);
961 break;
962 default:
963 ShouldNotReachHere();
964 return NULL;
965 }
966
967 // All these intrinsics have checks.
968 C->set_has_split_ifs(true); // Has chance for split-if optimization
969
970 return _gvn.transform(result);
971 }
972
973 //------------------------------inline_string_compareTo------------------------
974 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
975 Node* arg1 = argument(0);
976 Node* arg2 = argument(1);
977
978 // Get start addr and length of first argument
979 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
980 Node* arg1_cnt = load_array_length(arg1);
981
982 // Get start addr and length of second argument
983 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
984 Node* arg2_cnt = load_array_length(arg2);
985
986 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
987 set_result(result);
988 return true;
989 }
990
991 //------------------------------inline_string_equals------------------------
992 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
993 Node* arg1 = argument(0);
994 Node* arg2 = argument(1);
995
996 // paths (plus control) merge
997 RegionNode* region = new RegionNode(3);
998 Node* phi = new PhiNode(region, TypeInt::BOOL);
999
1000 if (!stopped()) {
1001 // Get start addr and length of first argument
1002 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1003 Node* arg1_cnt = load_array_length(arg1);
1004
1005 // Get start addr and length of second argument
1006 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1007 Node* arg2_cnt = load_array_length(arg2);
1008
1009 // Check for arg1_cnt != arg2_cnt
1010 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1011 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1012 Node* if_ne = generate_slow_guard(bol, NULL);
1013 if (if_ne != NULL) {
1014 phi->init_req(2, intcon(0));
1015 region->init_req(2, if_ne);
1016 }
1017
1018 // Check for count == 0 is done by assembler code for StrEquals.
1019
1020 if (!stopped()) {
1021 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1022 phi->init_req(1, equals);
1023 region->init_req(1, control());
1024 }
1025 }
1026
1027 // post merge
1028 set_control(_gvn.transform(region));
1029 record_for_igvn(region);
1030
1031 set_result(_gvn.transform(phi));
1032 return true;
1033 }
1034
1035 //------------------------------inline_array_equals----------------------------
1036 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1037 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1038 Node* arg1 = argument(0);
1039 Node* arg2 = argument(1);
1040
1041 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1042 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1043 return true;
1044 }
1045
1046 //------------------------------inline_hasNegatives------------------------------
1047 bool LibraryCallKit::inline_hasNegatives() {
1048 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1049 return false;
1050 }
1051
1052 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1053 // no receiver since it is static method
1054 Node* ba = argument(0);
1055 Node* offset = argument(1);
1056 Node* len = argument(2);
1057
1058 // Range checks
1059 generate_string_range_check(ba, offset, len, false);
1060 if (stopped()) {
1061 return true;
1062 }
1063 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1064 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1065 set_result(_gvn.transform(result));
1066 return true;
1067 }
1068
1069 bool LibraryCallKit::inline_objects_checkIndex() {
1070 Node* index = argument(0);
1071 Node* length = argument(1);
1072 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1073 return false;
1074 }
1075
1076 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1077 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1078
1079 {
1080 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1081 uncommon_trap(Deoptimization::Reason_intrinsic,
1082 Deoptimization::Action_make_not_entrant);
1083 }
1084
1085 if (stopped()) {
1086 return false;
1087 }
1088
1089 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1090 BoolTest::mask btest = BoolTest::lt;
1091 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1092 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1093 _gvn.set_type(rc, rc->Value(&_gvn));
1094 if (!rc_bool->is_Con()) {
1095 record_for_igvn(rc);
1096 }
1097 set_control(_gvn.transform(new IfTrueNode(rc)));
1098 {
1099 PreserveJVMState pjvms(this);
1100 set_control(_gvn.transform(new IfFalseNode(rc)));
1101 uncommon_trap(Deoptimization::Reason_range_check,
1102 Deoptimization::Action_make_not_entrant);
1103 }
1104
1105 if (stopped()) {
1106 return false;
1107 }
1108
1109 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1110 result->set_req(0, control());
1111 result = _gvn.transform(result);
1112 set_result(result);
1113 replace_in_map(index, result);
1114 return true;
1115 }
1116
1117 //------------------------------inline_string_indexOf------------------------
1118 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1119 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1120 return false;
1121 }
1122 Node* src = argument(0);
1123 Node* tgt = argument(1);
1124
1125 // Make the merge point
1126 RegionNode* result_rgn = new RegionNode(4);
1127 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1128
1129 // Get start addr and length of source string
1130 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1131 Node* src_count = load_array_length(src);
1132
1133 // Get start addr and length of substring
1134 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1135 Node* tgt_count = load_array_length(tgt);
1136
1137 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1138 // Divide src size by 2 if String is UTF16 encoded
1139 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1140 }
1141 if (ae == StrIntrinsicNode::UU) {
1142 // Divide substring size by 2 if String is UTF16 encoded
1143 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1144 }
1145
1146 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1147 if (result != NULL) {
1148 result_phi->init_req(3, result);
1149 result_rgn->init_req(3, control());
1150 }
1151 set_control(_gvn.transform(result_rgn));
1152 record_for_igvn(result_rgn);
1153 set_result(_gvn.transform(result_phi));
1154
1155 return true;
1156 }
1157
1158 //-----------------------------inline_string_indexOf-----------------------
1159 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1160 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1161 return false;
1162 }
1163 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1164 return false;
1165 }
1166 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1167 Node* src = argument(0); // byte[]
1168 Node* src_count = argument(1); // char count
1169 Node* tgt = argument(2); // byte[]
1170 Node* tgt_count = argument(3); // char count
1171 Node* from_index = argument(4); // char index
1172
1173 // Multiply byte array index by 2 if String is UTF16 encoded
1174 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1175 src_count = _gvn.transform(new SubINode(src_count, from_index));
1176 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1177 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1178
1179 // Range checks
1180 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1181 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1182 if (stopped()) {
1183 return true;
1184 }
1185
1186 RegionNode* region = new RegionNode(5);
1187 Node* phi = new PhiNode(region, TypeInt::INT);
1188
1189 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1190 if (result != NULL) {
1191 // The result is index relative to from_index if substring was found, -1 otherwise.
1192 // Generate code which will fold into cmove.
1193 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1194 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1195
1196 Node* if_lt = generate_slow_guard(bol, NULL);
1197 if (if_lt != NULL) {
1198 // result == -1
1199 phi->init_req(3, result);
1200 region->init_req(3, if_lt);
1201 }
1202 if (!stopped()) {
1203 result = _gvn.transform(new AddINode(result, from_index));
1204 phi->init_req(4, result);
1205 region->init_req(4, control());
1206 }
1207 }
1208
1209 set_control(_gvn.transform(region));
1210 record_for_igvn(region);
1211 set_result(_gvn.transform(phi));
1212
1213 return true;
1214 }
1215
1216 // Create StrIndexOfNode with fast path checks
1217 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1218 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1219 // Check for substr count > string count
1220 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1221 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1222 Node* if_gt = generate_slow_guard(bol, NULL);
1223 if (if_gt != NULL) {
1224 phi->init_req(1, intcon(-1));
1225 region->init_req(1, if_gt);
1226 }
1227 if (!stopped()) {
1228 // Check for substr count == 0
1229 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1230 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1231 Node* if_zero = generate_slow_guard(bol, NULL);
1232 if (if_zero != NULL) {
1233 phi->init_req(2, intcon(0));
1234 region->init_req(2, if_zero);
1235 }
1236 }
1237 if (!stopped()) {
1238 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1239 }
1240 return NULL;
1241 }
1242
1243 //-----------------------------inline_string_indexOfChar-----------------------
1244 bool LibraryCallKit::inline_string_indexOfChar() {
1245 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1246 return false;
1247 }
1248 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) {
1249 return false;
1250 }
1251 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1252 Node* src = argument(0); // byte[]
1253 Node* tgt = argument(1); // tgt is int ch
1254 Node* from_index = argument(2);
1255 Node* max = argument(3);
1256
1257 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1258 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1259 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1260
1261 // Range checks
1262 generate_string_range_check(src, src_offset, src_count, true);
1263 if (stopped()) {
1264 return true;
1265 }
1266
1267 RegionNode* region = new RegionNode(3);
1268 Node* phi = new PhiNode(region, TypeInt::INT);
1269
1270 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1271 C->set_has_split_ifs(true); // Has chance for split-if optimization
1272 _gvn.transform(result);
1273
1274 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1275 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1276
1277 Node* if_lt = generate_slow_guard(bol, NULL);
1278 if (if_lt != NULL) {
1279 // result == -1
1280 phi->init_req(2, result);
1281 region->init_req(2, if_lt);
1282 }
1283 if (!stopped()) {
1284 result = _gvn.transform(new AddINode(result, from_index));
1285 phi->init_req(1, result);
1286 region->init_req(1, control());
1287 }
1288 set_control(_gvn.transform(region));
1289 record_for_igvn(region);
1290 set_result(_gvn.transform(phi));
1291
1292 return true;
1293 }
1294 //---------------------------inline_string_copy---------------------
1295 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1296 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1297 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1298 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1299 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1300 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1301 bool LibraryCallKit::inline_string_copy(bool compress) {
1302 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1303 return false;
1304 }
1305 int nargs = 5; // 2 oops, 3 ints
1306 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1307
1308 Node* src = argument(0);
1309 Node* src_offset = argument(1);
1310 Node* dst = argument(2);
1311 Node* dst_offset = argument(3);
1312 Node* length = argument(4);
1313
1314 // Check for allocation before we add nodes that would confuse
1315 // tightly_coupled_allocation()
1316 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1317
1318 // Figure out the size and type of the elements we will be copying.
1319 const Type* src_type = src->Value(&_gvn);
1320 const Type* dst_type = dst->Value(&_gvn);
1321 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1322 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1323 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1324 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1325 "Unsupported array types for inline_string_copy");
1326
1327 // Range checks
1328 generate_string_range_check(src, src_offset, length, compress && src_elem == T_BYTE);
1329 generate_string_range_check(dst, dst_offset, length, !compress && dst_elem == T_BYTE);
1330 if (stopped()) {
1331 return true;
1332 }
1333
1334 // Convert char[] offsets to byte[] offsets
1335 if (compress && src_elem == T_BYTE) {
1336 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1337 } else if (!compress && dst_elem == T_BYTE) {
1338 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1339 }
1340
1341 Node* src_start = array_element_address(src, src_offset, src_elem);
1342 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1343 // 'src_start' points to src array + scaled offset
1344 // 'dst_start' points to dst array + scaled offset
1345 Node* count = NULL;
1346 if (compress) {
1347 count = compress_string(src_start, dst_start, length);
1348 } else {
1349 inflate_string(src_start, dst_start, length);
1350 }
1351
1352 if (alloc != NULL) {
1353 if (alloc->maybe_set_complete(&_gvn)) {
1354 // "You break it, you buy it."
1355 InitializeNode* init = alloc->initialization();
1356 assert(init->is_complete(), "we just did this");
1357 init->set_complete_with_arraycopy();
1358 assert(dst->is_CheckCastPP(), "sanity");
1359 assert(dst->in(0)->in(0) == init, "dest pinned");
1360 }
1361 // Do not let stores that initialize this object be reordered with
1362 // a subsequent store that would make this object accessible by
1363 // other threads.
1364 // Record what AllocateNode this StoreStore protects so that
1365 // escape analysis can go from the MemBarStoreStoreNode to the
1366 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1367 // based on the escape status of the AllocateNode.
1368 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1369 }
1370 if (compress) {
1371 set_result(_gvn.transform(count));
1372 }
1373 return true;
1374 }
1375
1376 #ifdef _LP64
1377 #define XTOP ,top() /*additional argument*/
1378 #else //_LP64
1379 #define XTOP /*no additional argument*/
1380 #endif //_LP64
1381
1382 //------------------------inline_string_toBytesU--------------------------
1383 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1384 bool LibraryCallKit::inline_string_toBytesU() {
1385 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1386 return false;
1387 }
1388 // Get the arguments.
1389 Node* value = argument(0);
1390 Node* offset = argument(1);
1391 Node* length = argument(2);
1392
1393 Node* newcopy = NULL;
1394
1395 // Set the original stack and the reexecute bit for the interpreter to reexecute
1396 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1397 { PreserveReexecuteState preexecs(this);
1398 jvms()->set_should_reexecute(true);
1399
1400 // Check if a null path was taken unconditionally.
1401 value = null_check(value);
1402
1403 RegionNode* bailout = new RegionNode(1);
1404 record_for_igvn(bailout);
1405
1406 // Range checks
1407 generate_negative_guard(offset, bailout);
1408 generate_negative_guard(length, bailout);
1409 generate_limit_guard(offset, length, load_array_length(value), bailout);
1410 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1411 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1412
1413 if (bailout->req() > 1) {
1414 PreserveJVMState pjvms(this);
1415 set_control(_gvn.transform(bailout));
1416 uncommon_trap(Deoptimization::Reason_intrinsic,
1417 Deoptimization::Action_maybe_recompile);
1418 }
1419 if (stopped()) {
1420 return true;
1421 }
1422
1423 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1424 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1425 newcopy = new_array(klass_node, size, 0); // no arguments to push
1426 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1427
1428 // Calculate starting addresses.
1429 Node* src_start = array_element_address(value, offset, T_CHAR);
1430 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1431
1432 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1433 const TypeInt* toffset = gvn().type(offset)->is_int();
1434 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1435
1436 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1437 const char* copyfunc_name = "arraycopy";
1438 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1439 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1440 OptoRuntime::fast_arraycopy_Type(),
1441 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1442 src_start, dst_start, ConvI2X(length) XTOP);
1443 // Do not let reads from the cloned object float above the arraycopy.
1444 if (alloc != NULL) {
1445 if (alloc->maybe_set_complete(&_gvn)) {
1446 // "You break it, you buy it."
1447 InitializeNode* init = alloc->initialization();
1448 assert(init->is_complete(), "we just did this");
1449 init->set_complete_with_arraycopy();
1450 assert(newcopy->is_CheckCastPP(), "sanity");
1451 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1452 }
1453 // Do not let stores that initialize this object be reordered with
1454 // a subsequent store that would make this object accessible by
1455 // other threads.
1456 // Record what AllocateNode this StoreStore protects so that
1457 // escape analysis can go from the MemBarStoreStoreNode to the
1458 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1459 // based on the escape status of the AllocateNode.
1460 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1461 } else {
1462 insert_mem_bar(Op_MemBarCPUOrder);
1463 }
1464 } // original reexecute is set back here
1465
1466 C->set_has_split_ifs(true); // Has chance for split-if optimization
1467 if (!stopped()) {
1468 set_result(newcopy);
1469 }
1470 return true;
1471 }
1472
1473 //------------------------inline_string_getCharsU--------------------------
1474 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1475 bool LibraryCallKit::inline_string_getCharsU() {
1476 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1477 return false;
1478 }
1479
1480 // Get the arguments.
1481 Node* src = argument(0);
1482 Node* src_begin = argument(1);
1483 Node* src_end = argument(2); // exclusive offset (i < src_end)
1484 Node* dst = argument(3);
1485 Node* dst_begin = argument(4);
1486
1487 // Check for allocation before we add nodes that would confuse
1488 // tightly_coupled_allocation()
1489 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1490
1491 // Check if a null path was taken unconditionally.
1492 src = null_check(src);
1493 dst = null_check(dst);
1494 if (stopped()) {
1495 return true;
1496 }
1497
1498 // Get length and convert char[] offset to byte[] offset
1499 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1500 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1501
1502 // Range checks
1503 generate_string_range_check(src, src_begin, length, true);
1504 generate_string_range_check(dst, dst_begin, length, false);
1505 if (stopped()) {
1506 return true;
1507 }
1508
1509 if (!stopped()) {
1510 // Calculate starting addresses.
1511 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1512 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1513
1514 // Check if array addresses are aligned to HeapWordSize
1515 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1516 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1517 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1518 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1519
1520 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1521 const char* copyfunc_name = "arraycopy";
1522 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1523 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1524 OptoRuntime::fast_arraycopy_Type(),
1525 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1526 src_start, dst_start, ConvI2X(length) XTOP);
1527 // Do not let reads from the cloned object float above the arraycopy.
1528 if (alloc != NULL) {
1529 if (alloc->maybe_set_complete(&_gvn)) {
1530 // "You break it, you buy it."
1531 InitializeNode* init = alloc->initialization();
1532 assert(init->is_complete(), "we just did this");
1533 init->set_complete_with_arraycopy();
1534 assert(dst->is_CheckCastPP(), "sanity");
1535 assert(dst->in(0)->in(0) == init, "dest pinned");
1536 }
1537 // Do not let stores that initialize this object be reordered with
1538 // a subsequent store that would make this object accessible by
1539 // other threads.
1540 // Record what AllocateNode this StoreStore protects so that
1541 // escape analysis can go from the MemBarStoreStoreNode to the
1542 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1543 // based on the escape status of the AllocateNode.
1544 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1545 } else {
1546 insert_mem_bar(Op_MemBarCPUOrder);
1547 }
1548 }
1549
1550 C->set_has_split_ifs(true); // Has chance for split-if optimization
1551 return true;
1552 }
1553
1554 //----------------------inline_string_char_access----------------------------
1555 // Store/Load char to/from byte[] array.
1556 // static void StringUTF16.putChar(byte[] val, int index, int c)
1557 // static char StringUTF16.getChar(byte[] val, int index)
1558 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1559 Node* value = argument(0);
1560 Node* index = argument(1);
1561 Node* ch = is_store ? argument(2) : NULL;
1562
1563 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1564 // correctly requires matched array shapes.
1565 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1566 "sanity: byte[] and char[] bases agree");
1567 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1568 "sanity: byte[] and char[] scales agree");
1569
1570 Node* adr = array_element_address(value, index, T_CHAR);
1571 if (is_store) {
1572 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1573 false, false, true /* mismatched */);
1574 } else {
1575 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered,
1576 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1577 set_result(ch);
1578 }
1579 return true;
1580 }
1581
1582 //--------------------------round_double_node--------------------------------
1583 // Round a double node if necessary.
1584 Node* LibraryCallKit::round_double_node(Node* n) {
1585 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1586 n = _gvn.transform(new RoundDoubleNode(0, n));
1587 return n;
1588 }
1589
1590 //------------------------------inline_math-----------------------------------
1591 // public static double Math.abs(double)
1592 // public static double Math.sqrt(double)
1593 // public static double Math.log(double)
1594 // public static double Math.log10(double)
1595 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1596 Node* arg = round_double_node(argument(0));
1597 Node* n = NULL;
1598 switch (id) {
1599 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1600 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1601 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1602 default: fatal_unexpected_iid(id); break;
1603 }
1604 set_result(_gvn.transform(n));
1605 return true;
1606 }
1607
1608 //------------------------------inline_trig----------------------------------
1609 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1610 // argument reduction which will turn into a fast/slow diamond.
1611 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1612 Node* arg = round_double_node(argument(0));
1613 Node* n = NULL;
1614
1615 switch (id) {
1616 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1617 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1618 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1619 default: fatal_unexpected_iid(id); break;
1620 }
1621 n = _gvn.transform(n);
1622
1623 // Rounding required? Check for argument reduction!
1624 if (Matcher::strict_fp_requires_explicit_rounding) {
1625 static const double pi_4 = 0.7853981633974483;
1626 static const double neg_pi_4 = -0.7853981633974483;
1627 // pi/2 in 80-bit extended precision
1628 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1629 // -pi/2 in 80-bit extended precision
1630 // 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};
1631 // Cutoff value for using this argument reduction technique
1632 //static const double pi_2_minus_epsilon = 1.564660403643354;
1633 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1634
1635 // Pseudocode for sin:
1636 // if (x <= Math.PI / 4.0) {
1637 // if (x >= -Math.PI / 4.0) return fsin(x);
1638 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1639 // } else {
1640 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1641 // }
1642 // return StrictMath.sin(x);
1643
1644 // Pseudocode for cos:
1645 // if (x <= Math.PI / 4.0) {
1646 // if (x >= -Math.PI / 4.0) return fcos(x);
1647 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1648 // } else {
1649 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1650 // }
1651 // return StrictMath.cos(x);
1652
1653 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1654 // requires a special machine instruction to load it. Instead we'll try
1655 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1656 // probably do the math inside the SIN encoding.
1657
1658 // Make the merge point
1659 RegionNode* r = new RegionNode(3);
1660 Node* phi = new PhiNode(r, Type::DOUBLE);
1661
1662 // Flatten arg so we need only 1 test
1663 Node *abs = _gvn.transform(new AbsDNode(arg));
1664 // Node for PI/4 constant
1665 Node *pi4 = makecon(TypeD::make(pi_4));
1666 // Check PI/4 : abs(arg)
1667 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1668 // Check: If PI/4 < abs(arg) then go slow
1669 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1670 // Branch either way
1671 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1672 set_control(opt_iff(r,iff));
1673
1674 // Set fast path result
1675 phi->init_req(2, n);
1676
1677 // Slow path - non-blocking leaf call
1678 Node* call = NULL;
1679 switch (id) {
1680 case vmIntrinsics::_dsin:
1681 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1682 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1683 "Sin", NULL, arg, top());
1684 break;
1685 case vmIntrinsics::_dcos:
1686 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1687 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1688 "Cos", NULL, arg, top());
1689 break;
1690 case vmIntrinsics::_dtan:
1691 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1692 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1693 "Tan", NULL, arg, top());
1694 break;
1695 }
1696 assert(control()->in(0) == call, "");
1697 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1698 r->init_req(1, control());
1699 phi->init_req(1, slow_result);
1700
1701 // Post-merge
1702 set_control(_gvn.transform(r));
1703 record_for_igvn(r);
1704 n = _gvn.transform(phi);
1705
1706 C->set_has_split_ifs(true); // Has chance for split-if optimization
1707 }
1708 set_result(n);
1709 return true;
1710 }
1711
1712 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1713 //-------------------
1714 //result=(result.isNaN())? funcAddr():result;
1715 // Check: If isNaN() by checking result!=result? then either trap
1716 // or go to runtime
1717 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1718 // Build the boolean node
1719 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1720
1721 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1722 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1723 // The pow or exp intrinsic returned a NaN, which requires a call
1724 // to the runtime. Recompile with the runtime call.
1725 uncommon_trap(Deoptimization::Reason_intrinsic,
1726 Deoptimization::Action_make_not_entrant);
1727 }
1728 return result;
1729 } else {
1730 // If this inlining ever returned NaN in the past, we compile a call
1731 // to the runtime to properly handle corner cases
1732
1733 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1734 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1735 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1736
1737 if (!if_slow->is_top()) {
1738 RegionNode* result_region = new RegionNode(3);
1739 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1740
1741 result_region->init_req(1, if_fast);
1742 result_val->init_req(1, result);
1743
1744 set_control(if_slow);
1745
1746 const TypePtr* no_memory_effects = NULL;
1747 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1748 no_memory_effects,
1749 x, top(), y, y ? top() : NULL);
1750 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1751 #ifdef ASSERT
1752 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1753 assert(value_top == top(), "second value must be top");
1754 #endif
1755
1756 result_region->init_req(2, control());
1757 result_val->init_req(2, value);
1758 set_control(_gvn.transform(result_region));
1759 return _gvn.transform(result_val);
1760 } else {
1761 return result;
1762 }
1763 }
1764 }
1765
1766 //------------------------------inline_pow-------------------------------------
1767 // Inline power instructions, if possible.
1768 bool LibraryCallKit::inline_pow() {
1769 // Pseudocode for pow
1770 // if (y == 2) {
1771 // return x * x;
1772 // } else {
1773 // if (x <= 0.0) {
1774 // long longy = (long)y;
1775 // if ((double)longy == y) { // if y is long
1776 // if (y + 1 == y) longy = 0; // huge number: even
1777 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1778 // } else {
1779 // result = NaN;
1780 // }
1781 // } else {
1782 // result = DPow(x,y);
1783 // }
1784 // if (result != result)? {
1785 // result = uncommon_trap() or runtime_call();
1786 // }
1787 // return result;
1788 // }
1789
1790 Node* x = round_double_node(argument(0));
1791 Node* y = round_double_node(argument(2));
1792
1793 Node* result = NULL;
1794
1795 Node* const_two_node = makecon(TypeD::make(2.0));
1796 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1797 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1798 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1799 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1800 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1801
1802 RegionNode* region_node = new RegionNode(3);
1803 region_node->init_req(1, if_true);
1804
1805 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1806 // special case for x^y where y == 2, we can convert it to x * x
1807 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1808
1809 // set control to if_false since we will now process the false branch
1810 set_control(if_false);
1811
1812 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1813 // Short form: skip the fancy tests and just check for NaN result.
1814 result = _gvn.transform(new PowDNode(C, control(), x, y));
1815 } else {
1816 // If this inlining ever returned NaN in the past, include all
1817 // checks + call to the runtime.
1818
1819 // Set the merge point for If node with condition of (x <= 0.0)
1820 // There are four possible paths to region node and phi node
1821 RegionNode *r = new RegionNode(4);
1822 Node *phi = new PhiNode(r, Type::DOUBLE);
1823
1824 // Build the first if node: if (x <= 0.0)
1825 // Node for 0 constant
1826 Node *zeronode = makecon(TypeD::ZERO);
1827 // Check x:0
1828 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1829 // Check: If (x<=0) then go complex path
1830 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1831 // Branch either way
1832 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1833 // Fast path taken; set region slot 3
1834 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1835 r->init_req(3,fast_taken); // Capture fast-control
1836
1837 // Fast path not-taken, i.e. slow path
1838 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1839
1840 // Set fast path result
1841 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1842 phi->init_req(3, fast_result);
1843
1844 // Complex path
1845 // Build the second if node (if y is long)
1846 // Node for (long)y
1847 Node *longy = _gvn.transform(new ConvD2LNode(y));
1848 // Node for (double)((long) y)
1849 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1850 // Check (double)((long) y) : y
1851 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1852 // Check if (y isn't long) then go to slow path
1853
1854 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1855 // Branch either way
1856 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1857 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1858
1859 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1860
1861 // Calculate DPow(abs(x), y)*(1 & (long)y)
1862 // Node for constant 1
1863 Node *conone = longcon(1);
1864 // 1& (long)y
1865 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1866
1867 // A huge number is always even. Detect a huge number by checking
1868 // if y + 1 == y and set integer to be tested for parity to 0.
1869 // Required for corner case:
1870 // (long)9.223372036854776E18 = max_jlong
1871 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1872 // max_jlong is odd but 9.223372036854776E18 is even
1873 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1874 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1875 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1876 Node* correctedsign = NULL;
1877 if (ConditionalMoveLimit != 0) {
1878 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1879 } else {
1880 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1881 RegionNode *r = new RegionNode(3);
1882 Node *phi = new PhiNode(r, TypeLong::LONG);
1883 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1884 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1885 phi->init_req(1, signnode);
1886 phi->init_req(2, longcon(0));
1887 correctedsign = _gvn.transform(phi);
1888 ylong_path = _gvn.transform(r);
1889 record_for_igvn(r);
1890 }
1891
1892 // zero node
1893 Node *conzero = longcon(0);
1894 // Check (1&(long)y)==0?
1895 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1896 // Check if (1&(long)y)!=0?, if so the result is negative
1897 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1898 // abs(x)
1899 Node *absx=_gvn.transform(new AbsDNode(x));
1900 // abs(x)^y
1901 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1902 // -abs(x)^y
1903 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1904 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1905 Node *signresult = NULL;
1906 if (ConditionalMoveLimit != 0) {
1907 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1908 } else {
1909 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1910 RegionNode *r = new RegionNode(3);
1911 Node *phi = new PhiNode(r, Type::DOUBLE);
1912 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1913 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1914 phi->init_req(1, absxpowy);
1915 phi->init_req(2, negabsxpowy);
1916 signresult = _gvn.transform(phi);
1917 ylong_path = _gvn.transform(r);
1918 record_for_igvn(r);
1919 }
1920 // Set complex path fast result
1921 r->init_req(2, ylong_path);
1922 phi->init_req(2, signresult);
1923
1924 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1925 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1926 r->init_req(1,slow_path);
1927 phi->init_req(1,slow_result);
1928
1929 // Post merge
1930 set_control(_gvn.transform(r));
1931 record_for_igvn(r);
1932 result = _gvn.transform(phi);
1933 }
1934
1935 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1936
1937 // control from finish_pow_exp is now input to the region node
1938 region_node->set_req(2, control());
1939 // the result from finish_pow_exp is now input to the phi node
1940 phi_node->init_req(2, result);
1941 set_control(_gvn.transform(region_node));
1942 record_for_igvn(region_node);
1943 set_result(_gvn.transform(phi_node));
1944
1945 C->set_has_split_ifs(true); // Has chance for split-if optimization
1946 return true;
1947 }
1948
1949 //------------------------------runtime_math-----------------------------
1950 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1951 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1952 "must be (DD)D or (D)D type");
1953
1954 // Inputs
1955 Node* a = round_double_node(argument(0));
1956 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1957
1958 const TypePtr* no_memory_effects = NULL;
1959 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1960 no_memory_effects,
1961 a, top(), b, b ? top() : NULL);
1962 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1963 #ifdef ASSERT
1964 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1965 assert(value_top == top(), "second value must be top");
1966 #endif
1967
1968 set_result(value);
1969 return true;
1970 }
1971
1972 //------------------------------inline_math_native-----------------------------
1973 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1974 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1975 switch (id) {
1976 // These intrinsics are not properly supported on all hardware
1977 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1978 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1979 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1980 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1981 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1982 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1983
1984 case vmIntrinsics::_dlog:
1985 return StubRoutines::dlog() != NULL ?
1986 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1987 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1988 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1989 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1990
1991 // These intrinsics are supported on all hardware
1992 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1993 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1994
1995 case vmIntrinsics::_dexp:
1996 return StubRoutines::dexp() != NULL ?
1997 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1998 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1999 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
2000 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
2001 #undef FN_PTR
2002
2003 // These intrinsics are not yet correctly implemented
2004 case vmIntrinsics::_datan2:
2005 return false;
2006
2007 default:
2008 fatal_unexpected_iid(id);
2009 return false;
2010 }
2011 }
2012
2013 static bool is_simple_name(Node* n) {
2014 return (n->req() == 1 // constant
2015 || (n->is_Type() && n->as_Type()->type()->singleton())
2016 || n->is_Proj() // parameter or return value
2017 || n->is_Phi() // local of some sort
2018 );
2019 }
2020
2021 //----------------------------inline_notify-----------------------------------*
2022 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
2023 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
2024 address func;
2025 if (id == vmIntrinsics::_notify) {
2026 func = OptoRuntime::monitor_notify_Java();
2027 } else {
2028 func = OptoRuntime::monitor_notifyAll_Java();
2029 }
2030 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
2031 make_slow_call_ex(call, env()->Throwable_klass(), false);
2032 return true;
2033 }
2034
2035
2036 //----------------------------inline_min_max-----------------------------------
2037 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2038 set_result(generate_min_max(id, argument(0), argument(1)));
2039 return true;
2040 }
2041
2042 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
2043 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
2044 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2045 Node* fast_path = _gvn.transform( new IfFalseNode(check));
2046 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
2047
2048 {
2049 PreserveJVMState pjvms(this);
2050 PreserveReexecuteState preexecs(this);
2051 jvms()->set_should_reexecute(true);
2052
2053 set_control(slow_path);
2054 set_i_o(i_o());
2055
2056 uncommon_trap(Deoptimization::Reason_intrinsic,
2057 Deoptimization::Action_none);
2058 }
2059
2060 set_control(fast_path);
2061 set_result(math);
2062 }
2063
2064 template <typename OverflowOp>
2065 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
2066 typedef typename OverflowOp::MathOp MathOp;
2067
2068 MathOp* mathOp = new MathOp(arg1, arg2);
2069 Node* operation = _gvn.transform( mathOp );
2070 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
2071 inline_math_mathExact(operation, ofcheck);
2072 return true;
2073 }
2074
2075 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2076 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2077 }
2078
2079 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2080 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2081 }
2082
2083 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2084 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2085 }
2086
2087 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2088 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2089 }
2090
2091 bool LibraryCallKit::inline_math_negateExactI() {
2092 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2093 }
2094
2095 bool LibraryCallKit::inline_math_negateExactL() {
2096 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2097 }
2098
2099 bool LibraryCallKit::inline_math_multiplyExactI() {
2100 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2101 }
2102
2103 bool LibraryCallKit::inline_math_multiplyExactL() {
2104 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2105 }
2106
2107 Node*
2108 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2109 // These are the candidate return value:
2110 Node* xvalue = x0;
2111 Node* yvalue = y0;
2112
2113 if (xvalue == yvalue) {
2114 return xvalue;
2115 }
2116
2117 bool want_max = (id == vmIntrinsics::_max);
2118
2119 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2120 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2121 if (txvalue == NULL || tyvalue == NULL) return top();
2122 // This is not really necessary, but it is consistent with a
2123 // hypothetical MaxINode::Value method:
2124 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2125
2126 // %%% This folding logic should (ideally) be in a different place.
2127 // Some should be inside IfNode, and there to be a more reliable
2128 // transformation of ?: style patterns into cmoves. We also want
2129 // more powerful optimizations around cmove and min/max.
2130
2131 // Try to find a dominating comparison of these guys.
2132 // It can simplify the index computation for Arrays.copyOf
2133 // and similar uses of System.arraycopy.
2134 // First, compute the normalized version of CmpI(x, y).
2135 int cmp_op = Op_CmpI;
2136 Node* xkey = xvalue;
2137 Node* ykey = yvalue;
2138 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2139 if (ideal_cmpxy->is_Cmp()) {
2140 // E.g., if we have CmpI(length - offset, count),
2141 // it might idealize to CmpI(length, count + offset)
2142 cmp_op = ideal_cmpxy->Opcode();
2143 xkey = ideal_cmpxy->in(1);
2144 ykey = ideal_cmpxy->in(2);
2145 }
2146
2147 // Start by locating any relevant comparisons.
2148 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2149 Node* cmpxy = NULL;
2150 Node* cmpyx = NULL;
2151 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2152 Node* cmp = start_from->fast_out(k);
2153 if (cmp->outcnt() > 0 && // must have prior uses
2154 cmp->in(0) == NULL && // must be context-independent
2155 cmp->Opcode() == cmp_op) { // right kind of compare
2156 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2157 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2158 }
2159 }
2160
2161 const int NCMPS = 2;
2162 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2163 int cmpn;
2164 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2165 if (cmps[cmpn] != NULL) break; // find a result
2166 }
2167 if (cmpn < NCMPS) {
2168 // Look for a dominating test that tells us the min and max.
2169 int depth = 0; // Limit search depth for speed
2170 Node* dom = control();
2171 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2172 if (++depth >= 100) break;
2173 Node* ifproj = dom;
2174 if (!ifproj->is_Proj()) continue;
2175 Node* iff = ifproj->in(0);
2176 if (!iff->is_If()) continue;
2177 Node* bol = iff->in(1);
2178 if (!bol->is_Bool()) continue;
2179 Node* cmp = bol->in(1);
2180 if (cmp == NULL) continue;
2181 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2182 if (cmps[cmpn] == cmp) break;
2183 if (cmpn == NCMPS) continue;
2184 BoolTest::mask btest = bol->as_Bool()->_test._test;
2185 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2186 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2187 // At this point, we know that 'x btest y' is true.
2188 switch (btest) {
2189 case BoolTest::eq:
2190 // They are proven equal, so we can collapse the min/max.
2191 // Either value is the answer. Choose the simpler.
2192 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2193 return yvalue;
2194 return xvalue;
2195 case BoolTest::lt: // x < y
2196 case BoolTest::le: // x <= y
2197 return (want_max ? yvalue : xvalue);
2198 case BoolTest::gt: // x > y
2199 case BoolTest::ge: // x >= y
2200 return (want_max ? xvalue : yvalue);
2201 }
2202 }
2203 }
2204
2205 // We failed to find a dominating test.
2206 // Let's pick a test that might GVN with prior tests.
2207 Node* best_bol = NULL;
2208 BoolTest::mask best_btest = BoolTest::illegal;
2209 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2210 Node* cmp = cmps[cmpn];
2211 if (cmp == NULL) continue;
2212 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2213 Node* bol = cmp->fast_out(j);
2214 if (!bol->is_Bool()) continue;
2215 BoolTest::mask btest = bol->as_Bool()->_test._test;
2216 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2217 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2218 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2219 best_bol = bol->as_Bool();
2220 best_btest = btest;
2221 }
2222 }
2223 }
2224
2225 Node* answer_if_true = NULL;
2226 Node* answer_if_false = NULL;
2227 switch (best_btest) {
2228 default:
2229 if (cmpxy == NULL)
2230 cmpxy = ideal_cmpxy;
2231 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2232 // and fall through:
2233 case BoolTest::lt: // x < y
2234 case BoolTest::le: // x <= y
2235 answer_if_true = (want_max ? yvalue : xvalue);
2236 answer_if_false = (want_max ? xvalue : yvalue);
2237 break;
2238 case BoolTest::gt: // x > y
2239 case BoolTest::ge: // x >= y
2240 answer_if_true = (want_max ? xvalue : yvalue);
2241 answer_if_false = (want_max ? yvalue : xvalue);
2242 break;
2243 }
2244
2245 jint hi, lo;
2246 if (want_max) {
2247 // We can sharpen the minimum.
2248 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2249 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2250 } else {
2251 // We can sharpen the maximum.
2252 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2253 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2254 }
2255
2256 // Use a flow-free graph structure, to avoid creating excess control edges
2257 // which could hinder other optimizations.
2258 // Since Math.min/max is often used with arraycopy, we want
2259 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2260 Node* cmov = CMoveNode::make(NULL, best_bol,
2261 answer_if_false, answer_if_true,
2262 TypeInt::make(lo, hi, widen));
2263
2264 return _gvn.transform(cmov);
2265
2266 /*
2267 // This is not as desirable as it may seem, since Min and Max
2268 // nodes do not have a full set of optimizations.
2269 // And they would interfere, anyway, with 'if' optimizations
2270 // and with CMoveI canonical forms.
2271 switch (id) {
2272 case vmIntrinsics::_min:
2273 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2274 case vmIntrinsics::_max:
2275 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2276 default:
2277 ShouldNotReachHere();
2278 }
2279 */
2280 }
2281
2282 inline int
2283 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2284 const TypePtr* base_type = TypePtr::NULL_PTR;
2285 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2286 if (base_type == NULL) {
2287 // Unknown type.
2288 return Type::AnyPtr;
2289 } else if (base_type == TypePtr::NULL_PTR) {
2290 // Since this is a NULL+long form, we have to switch to a rawptr.
2291 base = _gvn.transform(new CastX2PNode(offset));
2292 offset = MakeConX(0);
2293 return Type::RawPtr;
2294 } else if (base_type->base() == Type::RawPtr) {
2295 return Type::RawPtr;
2296 } else if (base_type->isa_oopptr()) {
2297 // Base is never null => always a heap address.
2298 if (base_type->ptr() == TypePtr::NotNull) {
2299 return Type::OopPtr;
2300 }
2301 // Offset is small => always a heap address.
2302 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2303 if (offset_type != NULL &&
2304 base_type->offset() == 0 && // (should always be?)
2305 offset_type->_lo >= 0 &&
2306 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2307 return Type::OopPtr;
2308 }
2309 // Otherwise, it might either be oop+off or NULL+addr.
2310 return Type::AnyPtr;
2311 } else {
2312 // No information:
2313 return Type::AnyPtr;
2314 }
2315 }
2316
2317 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2318 int kind = classify_unsafe_addr(base, offset);
2319 if (kind == Type::RawPtr) {
2320 return basic_plus_adr(top(), base, offset);
2321 } else {
2322 return basic_plus_adr(base, offset);
2323 }
2324 }
2325
2326 //--------------------------inline_number_methods-----------------------------
2327 // inline int Integer.numberOfLeadingZeros(int)
2328 // inline int Long.numberOfLeadingZeros(long)
2329 //
2330 // inline int Integer.numberOfTrailingZeros(int)
2331 // inline int Long.numberOfTrailingZeros(long)
2332 //
2333 // inline int Integer.bitCount(int)
2334 // inline int Long.bitCount(long)
2335 //
2336 // inline char Character.reverseBytes(char)
2337 // inline short Short.reverseBytes(short)
2338 // inline int Integer.reverseBytes(int)
2339 // inline long Long.reverseBytes(long)
2340 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2341 Node* arg = argument(0);
2342 Node* n = NULL;
2343 switch (id) {
2344 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2345 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2346 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2347 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2348 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2349 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2350 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2351 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2352 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2353 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2354 default: fatal_unexpected_iid(id); break;
2355 }
2356 set_result(_gvn.transform(n));
2357 return true;
2358 }
2359
2360 //----------------------------inline_unsafe_access----------------------------
2361
2362 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2363
2364 // Helper that guards and inserts a pre-barrier.
2365 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2366 Node* pre_val, bool need_mem_bar) {
2367 // We could be accessing the referent field of a reference object. If so, when G1
2368 // is enabled, we need to log the value in the referent field in an SATB buffer.
2369 // This routine performs some compile time filters and generates suitable
2370 // runtime filters that guard the pre-barrier code.
2371 // Also add memory barrier for non volatile load from the referent field
2372 // to prevent commoning of loads across safepoint.
2373 if (!UseG1GC && !need_mem_bar)
2374 return;
2375
2376 // Some compile time checks.
2377
2378 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2379 const TypeX* otype = offset->find_intptr_t_type();
2380 if (otype != NULL && otype->is_con() &&
2381 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2382 // Constant offset but not the reference_offset so just return
2383 return;
2384 }
2385
2386 // We only need to generate the runtime guards for instances.
2387 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2388 if (btype != NULL) {
2389 if (btype->isa_aryptr()) {
2390 // Array type so nothing to do
2391 return;
2392 }
2393
2394 const TypeInstPtr* itype = btype->isa_instptr();
2395 if (itype != NULL) {
2396 // Can the klass of base_oop be statically determined to be
2397 // _not_ a sub-class of Reference and _not_ Object?
2398 ciKlass* klass = itype->klass();
2399 if ( klass->is_loaded() &&
2400 !klass->is_subtype_of(env()->Reference_klass()) &&
2401 !env()->Object_klass()->is_subtype_of(klass)) {
2402 return;
2403 }
2404 }
2405 }
2406
2407 // The compile time filters did not reject base_oop/offset so
2408 // we need to generate the following runtime filters
2409 //
2410 // if (offset == java_lang_ref_Reference::_reference_offset) {
2411 // if (instance_of(base, java.lang.ref.Reference)) {
2412 // pre_barrier(_, pre_val, ...);
2413 // }
2414 // }
2415
2416 float likely = PROB_LIKELY( 0.999);
2417 float unlikely = PROB_UNLIKELY(0.999);
2418
2419 IdealKit ideal(this);
2420 #define __ ideal.
2421
2422 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2423
2424 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2425 // Update graphKit memory and control from IdealKit.
2426 sync_kit(ideal);
2427
2428 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2429 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2430
2431 // Update IdealKit memory and control from graphKit.
2432 __ sync_kit(this);
2433
2434 Node* one = __ ConI(1);
2435 // is_instof == 0 if base_oop == NULL
2436 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2437
2438 // Update graphKit from IdeakKit.
2439 sync_kit(ideal);
2440
2441 // Use the pre-barrier to record the value in the referent field
2442 pre_barrier(false /* do_load */,
2443 __ ctrl(),
2444 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2445 pre_val /* pre_val */,
2446 T_OBJECT);
2447 if (need_mem_bar) {
2448 // Add memory barrier to prevent commoning reads from this field
2449 // across safepoint since GC can change its value.
2450 insert_mem_bar(Op_MemBarCPUOrder);
2451 }
2452 // Update IdealKit from graphKit.
2453 __ sync_kit(this);
2454
2455 } __ end_if(); // _ref_type != ref_none
2456 } __ end_if(); // offset == referent_offset
2457
2458 // Final sync IdealKit and GraphKit.
2459 final_sync(ideal);
2460 #undef __
2461 }
2462
2463
2464 // Interpret Unsafe.fieldOffset cookies correctly:
2465 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2466
2467 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2468 // Attempt to infer a sharper value type from the offset and base type.
2469 ciKlass* sharpened_klass = NULL;
2470
2471 // See if it is an instance field, with an object type.
2472 if (alias_type->field() != NULL) {
2473 assert(!is_native_ptr, "native pointer op cannot use a java address");
2474 if (alias_type->field()->type()->is_klass()) {
2475 sharpened_klass = alias_type->field()->type()->as_klass();
2476 }
2477 }
2478
2479 // See if it is a narrow oop array.
2480 if (adr_type->isa_aryptr()) {
2481 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2482 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2483 if (elem_type != NULL) {
2484 sharpened_klass = elem_type->klass();
2485 }
2486 }
2487 }
2488
2489 // The sharpened class might be unloaded if there is no class loader
2490 // contraint in place.
2491 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2492 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2493
2494 #ifndef PRODUCT
2495 if (C->print_intrinsics() || C->print_inlining()) {
2496 tty->print(" from base type: "); adr_type->dump();
2497 tty->print(" sharpened value: "); tjp->dump();
2498 }
2499 #endif
2500 // Sharpen the value type.
2501 return tjp;
2502 }
2503 return NULL;
2504 }
2505
2506 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2507 if (callee()->is_static()) return false; // caller must have the capability!
2508
2509 #ifndef PRODUCT
2510 {
2511 ResourceMark rm;
2512 // Check the signatures.
2513 ciSignature* sig = callee()->signature();
2514 #ifdef ASSERT
2515 if (!is_store) {
2516 // Object getObject(Object base, int/long offset), etc.
2517 BasicType rtype = sig->return_type()->basic_type();
2518 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2519 rtype = T_ADDRESS; // it is really a C void*
2520 assert(rtype == type, "getter must return the expected value");
2521 if (!is_native_ptr) {
2522 assert(sig->count() == 2, "oop getter has 2 arguments");
2523 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2524 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2525 } else {
2526 assert(sig->count() == 1, "native getter has 1 argument");
2527 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2528 }
2529 } else {
2530 // void putObject(Object base, int/long offset, Object x), etc.
2531 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2532 if (!is_native_ptr) {
2533 assert(sig->count() == 3, "oop putter has 3 arguments");
2534 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2535 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2536 } else {
2537 assert(sig->count() == 2, "native putter has 2 arguments");
2538 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2539 }
2540 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2541 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2542 vtype = T_ADDRESS; // it is really a C void*
2543 assert(vtype == type, "putter must accept the expected value");
2544 }
2545 #endif // ASSERT
2546 }
2547 #endif //PRODUCT
2548
2549 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2550
2551 Node* receiver = argument(0); // type: oop
2552
2553 // Build address expression.
2554 Node* adr;
2555 Node* heap_base_oop = top();
2556 Node* offset = top();
2557 Node* val;
2558
2559 if (!is_native_ptr) {
2560 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2561 Node* base = argument(1); // type: oop
2562 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2563 offset = argument(2); // type: long
2564 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2565 // to be plain byte offsets, which are also the same as those accepted
2566 // by oopDesc::field_base.
2567 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2568 "fieldOffset must be byte-scaled");
2569 // 32-bit machines ignore the high half!
2570 offset = ConvL2X(offset);
2571 adr = make_unsafe_address(base, offset);
2572 heap_base_oop = base;
2573 val = is_store ? argument(4) : NULL;
2574 } else {
2575 Node* ptr = argument(1); // type: long
2576 ptr = ConvL2X(ptr); // adjust Java long to machine word
2577 adr = make_unsafe_address(NULL, ptr);
2578 val = is_store ? argument(3) : NULL;
2579 }
2580
2581 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2582
2583 // First guess at the value type.
2584 const Type *value_type = Type::get_const_basic_type(type);
2585
2586 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2587 // there was not enough information to nail it down.
2588 Compile::AliasType* alias_type = C->alias_type(adr_type);
2589 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2590
2591 // We will need memory barriers unless we can determine a unique
2592 // alias category for this reference. (Note: If for some reason
2593 // the barriers get omitted and the unsafe reference begins to "pollute"
2594 // the alias analysis of the rest of the graph, either Compile::can_alias
2595 // or Compile::must_alias will throw a diagnostic assert.)
2596 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2597
2598 // If we are reading the value of the referent field of a Reference
2599 // object (either by using Unsafe directly or through reflection)
2600 // then, if G1 is enabled, we need to record the referent in an
2601 // SATB log buffer using the pre-barrier mechanism.
2602 // Also we need to add memory barrier to prevent commoning reads
2603 // from this field across safepoint since GC can change its value.
2604 bool need_read_barrier = !is_native_ptr && !is_store &&
2605 offset != top() && heap_base_oop != top();
2606
2607 if (!is_store && type == T_OBJECT) {
2608 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2609 if (tjp != NULL) {
2610 value_type = tjp;
2611 }
2612 }
2613
2614 receiver = null_check(receiver);
2615 if (stopped()) {
2616 return true;
2617 }
2618 // Heap pointers get a null-check from the interpreter,
2619 // as a courtesy. However, this is not guaranteed by Unsafe,
2620 // and it is not possible to fully distinguish unintended nulls
2621 // from intended ones in this API.
2622
2623 if (is_volatile) {
2624 // We need to emit leading and trailing CPU membars (see below) in
2625 // addition to memory membars when is_volatile. This is a little
2626 // too strong, but avoids the need to insert per-alias-type
2627 // volatile membars (for stores; compare Parse::do_put_xxx), which
2628 // we cannot do effectively here because we probably only have a
2629 // rough approximation of type.
2630 need_mem_bar = true;
2631 // For Stores, place a memory ordering barrier now.
2632 if (is_store) {
2633 insert_mem_bar(Op_MemBarRelease);
2634 } else {
2635 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2636 insert_mem_bar(Op_MemBarVolatile);
2637 }
2638 }
2639 }
2640
2641 // Memory barrier to prevent normal and 'unsafe' accesses from
2642 // bypassing each other. Happens after null checks, so the
2643 // exception paths do not take memory state from the memory barrier,
2644 // so there's no problems making a strong assert about mixing users
2645 // of safe & unsafe memory.
2646 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2647
2648 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM ||
2649 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown");
2650 bool mismatched = false;
2651 if (alias_type->element() != NULL || alias_type->field() != NULL) {
2652 BasicType bt;
2653 if (alias_type->element() != NULL) {
2654 const Type* element = alias_type->element();
2655 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
2656 } else {
2657 bt = alias_type->field()->type()->basic_type();
2658 }
2659 if (bt == T_ARRAY) {
2660 // accessing an array field with getObject is not a mismatch
2661 bt = T_OBJECT;
2662 }
2663 if (bt != type) {
2664 mismatched = true;
2665 }
2666 }
2667 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2668
2669 if (!is_store) {
2670 Node* p = NULL;
2671 // Try to constant fold a load from a constant field
2672 ciField* field = alias_type->field();
2673 if (heap_base_oop != top() &&
2674 field != NULL && field->is_constant() && field->layout_type() == type) {
2675 // final or stable field
2676 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2677 if (con_type != NULL) {
2678 p = makecon(con_type);
2679 }
2680 }
2681 if (p == NULL) {
2682 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2683 // To be valid, unsafe loads may depend on other conditions than
2684 // the one that guards them: pin the Load node
2685 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2686 // load value
2687 switch (type) {
2688 case T_BOOLEAN:
2689 case T_CHAR:
2690 case T_BYTE:
2691 case T_SHORT:
2692 case T_INT:
2693 case T_LONG:
2694 case T_FLOAT:
2695 case T_DOUBLE:
2696 break;
2697 case T_OBJECT:
2698 if (need_read_barrier) {
2699 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2700 }
2701 break;
2702 case T_ADDRESS:
2703 // Cast to an int type.
2704 p = _gvn.transform(new CastP2XNode(NULL, p));
2705 p = ConvX2UL(p);
2706 break;
2707 default:
2708 fatal("unexpected type %d: %s", type, type2name(type));
2709 break;
2710 }
2711 }
2712 // The load node has the control of the preceding MemBarCPUOrder. All
2713 // following nodes will have the control of the MemBarCPUOrder inserted at
2714 // the end of this method. So, pushing the load onto the stack at a later
2715 // point is fine.
2716 set_result(p);
2717 } else {
2718 // place effect of store into memory
2719 switch (type) {
2720 case T_DOUBLE:
2721 val = dstore_rounding(val);
2722 break;
2723 case T_ADDRESS:
2724 // Repackage the long as a pointer.
2725 val = ConvL2X(val);
2726 val = _gvn.transform(new CastX2PNode(val));
2727 break;
2728 }
2729
2730 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2731 if (type != T_OBJECT ) {
2732 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2733 } else {
2734 // Possibly an oop being stored to Java heap or native memory
2735 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2736 // oop to Java heap.
2737 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2738 } else {
2739 // We can't tell at compile time if we are storing in the Java heap or outside
2740 // of it. So we need to emit code to conditionally do the proper type of
2741 // store.
2742
2743 IdealKit ideal(this);
2744 #define __ ideal.
2745 // QQQ who knows what probability is here??
2746 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2747 // Sync IdealKit and graphKit.
2748 sync_kit(ideal);
2749 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2750 // Update IdealKit memory.
2751 __ sync_kit(this);
2752 } __ else_(); {
2753 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched);
2754 } __ end_if();
2755 // Final sync IdealKit and GraphKit.
2756 final_sync(ideal);
2757 #undef __
2758 }
2759 }
2760 }
2761
2762 if (is_volatile) {
2763 if (!is_store) {
2764 insert_mem_bar(Op_MemBarAcquire);
2765 } else {
2766 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2767 insert_mem_bar(Op_MemBarVolatile);
2768 }
2769 }
2770 }
2771
2772 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2773
2774 return true;
2775 }
2776
2777 //----------------------------inline_unsafe_load_store----------------------------
2778 // This method serves a couple of different customers (depending on LoadStoreKind):
2779 //
2780 // LS_cmpxchg:
2781 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2782 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2783 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2784 //
2785 // LS_xadd:
2786 // public int getAndAddInt( Object o, long offset, int delta)
2787 // public long getAndAddLong(Object o, long offset, long delta)
2788 //
2789 // LS_xchg:
2790 // int getAndSet(Object o, long offset, int newValue)
2791 // long getAndSet(Object o, long offset, long newValue)
2792 // Object getAndSet(Object o, long offset, Object newValue)
2793 //
2794 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2795 // This basic scheme here is the same as inline_unsafe_access, but
2796 // differs in enough details that combining them would make the code
2797 // overly confusing. (This is a true fact! I originally combined
2798 // them, but even I was confused by it!) As much code/comments as
2799 // possible are retained from inline_unsafe_access though to make
2800 // the correspondences clearer. - dl
2801
2802 if (callee()->is_static()) return false; // caller must have the capability!
2803
2804 #ifndef PRODUCT
2805 BasicType rtype;
2806 {
2807 ResourceMark rm;
2808 // Check the signatures.
2809 ciSignature* sig = callee()->signature();
2810 rtype = sig->return_type()->basic_type();
2811 if (kind == LS_xadd || kind == LS_xchg) {
2812 // Check the signatures.
2813 #ifdef ASSERT
2814 assert(rtype == type, "get and set must return the expected type");
2815 assert(sig->count() == 3, "get and set has 3 arguments");
2816 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2817 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2818 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2819 #endif // ASSERT
2820 } else if (kind == LS_cmpxchg) {
2821 // Check the signatures.
2822 #ifdef ASSERT
2823 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2824 assert(sig->count() == 4, "CAS has 4 arguments");
2825 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2826 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2827 #endif // ASSERT
2828 } else {
2829 ShouldNotReachHere();
2830 }
2831 }
2832 #endif //PRODUCT
2833
2834 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2835
2836 // Get arguments:
2837 Node* receiver = NULL;
2838 Node* base = NULL;
2839 Node* offset = NULL;
2840 Node* oldval = NULL;
2841 Node* newval = NULL;
2842 if (kind == LS_cmpxchg) {
2843 const bool two_slot_type = type2size[type] == 2;
2844 receiver = argument(0); // type: oop
2845 base = argument(1); // type: oop
2846 offset = argument(2); // type: long
2847 oldval = argument(4); // type: oop, int, or long
2848 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2849 } else if (kind == LS_xadd || kind == LS_xchg){
2850 receiver = argument(0); // type: oop
2851 base = argument(1); // type: oop
2852 offset = argument(2); // type: long
2853 oldval = NULL;
2854 newval = argument(4); // type: oop, int, or long
2855 }
2856
2857 // Null check receiver.
2858 receiver = null_check(receiver);
2859 if (stopped()) {
2860 return true;
2861 }
2862
2863 // Build field offset expression.
2864 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2865 // to be plain byte offsets, which are also the same as those accepted
2866 // by oopDesc::field_base.
2867 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2868 // 32-bit machines ignore the high half of long offsets
2869 offset = ConvL2X(offset);
2870 Node* adr = make_unsafe_address(base, offset);
2871 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2872
2873 // For CAS, unlike inline_unsafe_access, there seems no point in
2874 // trying to refine types. Just use the coarse types here.
2875 const Type *value_type = Type::get_const_basic_type(type);
2876 Compile::AliasType* alias_type = C->alias_type(adr_type);
2877 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2878
2879 if (kind == LS_xchg && type == T_OBJECT) {
2880 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2881 if (tjp != NULL) {
2882 value_type = tjp;
2883 }
2884 }
2885
2886 int alias_idx = C->get_alias_index(adr_type);
2887
2888 // Memory-model-wise, a LoadStore acts like a little synchronized
2889 // block, so needs barriers on each side. These don't translate
2890 // into actual barriers on most machines, but we still need rest of
2891 // compiler to respect ordering.
2892
2893 insert_mem_bar(Op_MemBarRelease);
2894 insert_mem_bar(Op_MemBarCPUOrder);
2895
2896 // 4984716: MemBars must be inserted before this
2897 // memory node in order to avoid a false
2898 // dependency which will confuse the scheduler.
2899 Node *mem = memory(alias_idx);
2900
2901 // For now, we handle only those cases that actually exist: ints,
2902 // longs, and Object. Adding others should be straightforward.
2903 Node* load_store = NULL;
2904 switch(type) {
2905 case T_INT:
2906 if (kind == LS_xadd) {
2907 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2908 } else if (kind == LS_xchg) {
2909 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2910 } else if (kind == LS_cmpxchg) {
2911 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2912 } else {
2913 ShouldNotReachHere();
2914 }
2915 break;
2916 case T_LONG:
2917 if (kind == LS_xadd) {
2918 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2919 } else if (kind == LS_xchg) {
2920 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2921 } else if (kind == LS_cmpxchg) {
2922 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2923 } else {
2924 ShouldNotReachHere();
2925 }
2926 break;
2927 case T_OBJECT:
2928 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2929 // could be delayed during Parse (for example, in adjust_map_after_if()).
2930 // Execute transformation here to avoid barrier generation in such case.
2931 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2932 newval = _gvn.makecon(TypePtr::NULL_PTR);
2933
2934 // Reference stores need a store barrier.
2935 if (kind == LS_xchg) {
2936 // If pre-barrier must execute before the oop store, old value will require do_load here.
2937 if (!can_move_pre_barrier()) {
2938 pre_barrier(true /* do_load*/,
2939 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2940 NULL /* pre_val*/,
2941 T_OBJECT);
2942 } // Else move pre_barrier to use load_store value, see below.
2943 } else if (kind == LS_cmpxchg) {
2944 // Same as for newval above:
2945 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2946 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2947 }
2948 // The only known value which might get overwritten is oldval.
2949 pre_barrier(false /* do_load */,
2950 control(), NULL, NULL, max_juint, NULL, NULL,
2951 oldval /* pre_val */,
2952 T_OBJECT);
2953 } else {
2954 ShouldNotReachHere();
2955 }
2956
2957 #ifdef _LP64
2958 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2959 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2960 if (kind == LS_xchg) {
2961 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2962 newval_enc, adr_type, value_type->make_narrowoop()));
2963 } else {
2964 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2965 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2966 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2967 newval_enc, oldval_enc));
2968 }
2969 } else
2970 #endif
2971 {
2972 if (kind == LS_xchg) {
2973 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2974 } else {
2975 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2976 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2977 }
2978 }
2979 if (kind == LS_cmpxchg) {
2980 // Emit the post barrier only when the actual store happened.
2981 // This makes sense to check only for compareAndSet that can fail to set the value.
2982 // CAS success path is marked more likely since we anticipate this is a performance
2983 // critical path, while CAS failure path can use the penalty for going through unlikely
2984 // path as backoff. Which is still better than doing a store barrier there.
2985 IdealKit ideal(this);
2986 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2987 sync_kit(ideal);
2988 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2989 ideal.sync_kit(this);
2990 } ideal.end_if();
2991 final_sync(ideal);
2992 } else {
2993 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2994 }
2995 break;
2996 default:
2997 fatal("unexpected type %d: %s", type, type2name(type));
2998 break;
2999 }
3000
3001 // SCMemProjNodes represent the memory state of a LoadStore. Their
3002 // main role is to prevent LoadStore nodes from being optimized away
3003 // when their results aren't used.
3004 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3005 set_memory(proj, alias_idx);
3006
3007 if (type == T_OBJECT && kind == LS_xchg) {
3008 #ifdef _LP64
3009 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3010 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3011 }
3012 #endif
3013 if (can_move_pre_barrier()) {
3014 // Don't need to load pre_val. The old value is returned by load_store.
3015 // The pre_barrier can execute after the xchg as long as no safepoint
3016 // gets inserted between them.
3017 pre_barrier(false /* do_load */,
3018 control(), NULL, NULL, max_juint, NULL, NULL,
3019 load_store /* pre_val */,
3020 T_OBJECT);
3021 }
3022 }
3023
3024 // Add the trailing membar surrounding the access
3025 insert_mem_bar(Op_MemBarCPUOrder);
3026 insert_mem_bar(Op_MemBarAcquire);
3027
3028 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3029 set_result(load_store);
3030 return true;
3031 }
3032
3033 //----------------------------inline_unsafe_ordered_store----------------------
3034 // public native void Unsafe.putOrderedObject(Object o, long offset, Object x);
3035 // public native void Unsafe.putOrderedInt(Object o, long offset, int x);
3036 // public native void Unsafe.putOrderedLong(Object o, long offset, long x);
3037 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3038 // This is another variant of inline_unsafe_access, differing in
3039 // that it always issues store-store ("release") barrier and ensures
3040 // store-atomicity (which only matters for "long").
3041
3042 if (callee()->is_static()) return false; // caller must have the capability!
3043
3044 #ifndef PRODUCT
3045 {
3046 ResourceMark rm;
3047 // Check the signatures.
3048 ciSignature* sig = callee()->signature();
3049 #ifdef ASSERT
3050 BasicType rtype = sig->return_type()->basic_type();
3051 assert(rtype == T_VOID, "must return void");
3052 assert(sig->count() == 3, "has 3 arguments");
3053 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3054 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3055 #endif // ASSERT
3056 }
3057 #endif //PRODUCT
3058
3059 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3060
3061 // Get arguments:
3062 Node* receiver = argument(0); // type: oop
3063 Node* base = argument(1); // type: oop
3064 Node* offset = argument(2); // type: long
3065 Node* val = argument(4); // type: oop, int, or long
3066
3067 // Null check receiver.
3068 receiver = null_check(receiver);
3069 if (stopped()) {
3070 return true;
3071 }
3072
3073 // Build field offset expression.
3074 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3075 // 32-bit machines ignore the high half of long offsets
3076 offset = ConvL2X(offset);
3077 Node* adr = make_unsafe_address(base, offset);
3078 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3079 const Type *value_type = Type::get_const_basic_type(type);
3080 Compile::AliasType* alias_type = C->alias_type(adr_type);
3081
3082 insert_mem_bar(Op_MemBarRelease);
3083 insert_mem_bar(Op_MemBarCPUOrder);
3084 // Ensure that the store is atomic for longs:
3085 const bool require_atomic_access = true;
3086 Node* store;
3087 if (type == T_OBJECT) // reference stores need a store barrier.
3088 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3089 else {
3090 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3091 }
3092 insert_mem_bar(Op_MemBarCPUOrder);
3093 return true;
3094 }
3095
3096 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3097 // Regardless of form, don't allow previous ld/st to move down,
3098 // then issue acquire, release, or volatile mem_bar.
3099 insert_mem_bar(Op_MemBarCPUOrder);
3100 switch(id) {
3101 case vmIntrinsics::_loadFence:
3102 insert_mem_bar(Op_LoadFence);
3103 return true;
3104 case vmIntrinsics::_storeFence:
3105 insert_mem_bar(Op_StoreFence);
3106 return true;
3107 case vmIntrinsics::_fullFence:
3108 insert_mem_bar(Op_MemBarVolatile);
3109 return true;
3110 default:
3111 fatal_unexpected_iid(id);
3112 return false;
3113 }
3114 }
3115
3116 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3117 if (!kls->is_Con()) {
3118 return true;
3119 }
3120 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3121 if (klsptr == NULL) {
3122 return true;
3123 }
3124 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3125 // don't need a guard for a klass that is already initialized
3126 return !ik->is_initialized();
3127 }
3128
3129 //----------------------------inline_unsafe_allocate---------------------------
3130 // public native Object Unsafe.allocateInstance(Class<?> cls);
3131 bool LibraryCallKit::inline_unsafe_allocate() {
3132 if (callee()->is_static()) return false; // caller must have the capability!
3133
3134 null_check_receiver(); // null-check, then ignore
3135 Node* cls = null_check(argument(1));
3136 if (stopped()) return true;
3137
3138 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3139 kls = null_check(kls);
3140 if (stopped()) return true; // argument was like int.class
3141
3142 Node* test = NULL;
3143 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3144 // Note: The argument might still be an illegal value like
3145 // Serializable.class or Object[].class. The runtime will handle it.
3146 // But we must make an explicit check for initialization.
3147 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3148 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3149 // can generate code to load it as unsigned byte.
3150 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3151 Node* bits = intcon(InstanceKlass::fully_initialized);
3152 test = _gvn.transform(new SubINode(inst, bits));
3153 // The 'test' is non-zero if we need to take a slow path.
3154 }
3155
3156 Node* obj = new_instance(kls, test);
3157 set_result(obj);
3158 return true;
3159 }
3160
3161 #ifdef TRACE_HAVE_INTRINSICS
3162 /*
3163 * oop -> myklass
3164 * myklass->trace_id |= USED
3165 * return myklass->trace_id & ~0x3
3166 */
3167 bool LibraryCallKit::inline_native_classID() {
3168 null_check_receiver(); // null-check, then ignore
3169 Node* cls = null_check(argument(1), T_OBJECT);
3170 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3171 kls = null_check(kls, T_OBJECT);
3172 ByteSize offset = TRACE_ID_OFFSET;
3173 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3174 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3175 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3176 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3177 Node* clsused = longcon(0x01l); // set the class bit
3178 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3179
3180 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3181 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3182 set_result(andl);
3183 return true;
3184 }
3185
3186 bool LibraryCallKit::inline_native_threadID() {
3187 Node* tls_ptr = NULL;
3188 Node* cur_thr = generate_current_thread(tls_ptr);
3189 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3190 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3191 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3192
3193 Node* threadid = NULL;
3194 size_t thread_id_size = OSThread::thread_id_size();
3195 if (thread_id_size == (size_t) BytesPerLong) {
3196 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3197 } else if (thread_id_size == (size_t) BytesPerInt) {
3198 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3199 } else {
3200 ShouldNotReachHere();
3201 }
3202 set_result(threadid);
3203 return true;
3204 }
3205 #endif
3206
3207 //------------------------inline_native_time_funcs--------------
3208 // inline code for System.currentTimeMillis() and System.nanoTime()
3209 // these have the same type and signature
3210 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3211 const TypeFunc* tf = OptoRuntime::void_long_Type();
3212 const TypePtr* no_memory_effects = NULL;
3213 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3214 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3215 #ifdef ASSERT
3216 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3217 assert(value_top == top(), "second value must be top");
3218 #endif
3219 set_result(value);
3220 return true;
3221 }
3222
3223 //------------------------inline_native_currentThread------------------
3224 bool LibraryCallKit::inline_native_currentThread() {
3225 Node* junk = NULL;
3226 set_result(generate_current_thread(junk));
3227 return true;
3228 }
3229
3230 //------------------------inline_native_isInterrupted------------------
3231 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3232 bool LibraryCallKit::inline_native_isInterrupted() {
3233 // Add a fast path to t.isInterrupted(clear_int):
3234 // (t == Thread.current() &&
3235 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3236 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3237 // So, in the common case that the interrupt bit is false,
3238 // we avoid making a call into the VM. Even if the interrupt bit
3239 // is true, if the clear_int argument is false, we avoid the VM call.
3240 // However, if the receiver is not currentThread, we must call the VM,
3241 // because there must be some locking done around the operation.
3242
3243 // We only go to the fast case code if we pass two guards.
3244 // Paths which do not pass are accumulated in the slow_region.
3245
3246 enum {
3247 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3248 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3249 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3250 PATH_LIMIT
3251 };
3252
3253 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3254 // out of the function.
3255 insert_mem_bar(Op_MemBarCPUOrder);
3256
3257 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3258 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3259
3260 RegionNode* slow_region = new RegionNode(1);
3261 record_for_igvn(slow_region);
3262
3263 // (a) Receiving thread must be the current thread.
3264 Node* rec_thr = argument(0);
3265 Node* tls_ptr = NULL;
3266 Node* cur_thr = generate_current_thread(tls_ptr);
3267 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3268 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3269
3270 generate_slow_guard(bol_thr, slow_region);
3271
3272 // (b) Interrupt bit on TLS must be false.
3273 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3274 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3275 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3276
3277 // Set the control input on the field _interrupted read to prevent it floating up.
3278 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3279 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3280 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3281
3282 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3283
3284 // First fast path: if (!TLS._interrupted) return false;
3285 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3286 result_rgn->init_req(no_int_result_path, false_bit);
3287 result_val->init_req(no_int_result_path, intcon(0));
3288
3289 // drop through to next case
3290 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3291
3292 #ifndef TARGET_OS_FAMILY_windows
3293 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3294 Node* clr_arg = argument(1);
3295 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3296 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3297 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3298
3299 // Second fast path: ... else if (!clear_int) return true;
3300 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3301 result_rgn->init_req(no_clear_result_path, false_arg);
3302 result_val->init_req(no_clear_result_path, intcon(1));
3303
3304 // drop through to next case
3305 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3306 #else
3307 // To return true on Windows you must read the _interrupted field
3308 // and check the the event state i.e. take the slow path.
3309 #endif // TARGET_OS_FAMILY_windows
3310
3311 // (d) Otherwise, go to the slow path.
3312 slow_region->add_req(control());
3313 set_control( _gvn.transform(slow_region));
3314
3315 if (stopped()) {
3316 // There is no slow path.
3317 result_rgn->init_req(slow_result_path, top());
3318 result_val->init_req(slow_result_path, top());
3319 } else {
3320 // non-virtual because it is a private non-static
3321 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3322
3323 Node* slow_val = set_results_for_java_call(slow_call);
3324 // this->control() comes from set_results_for_java_call
3325
3326 Node* fast_io = slow_call->in(TypeFunc::I_O);
3327 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3328
3329 // These two phis are pre-filled with copies of of the fast IO and Memory
3330 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3331 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3332
3333 result_rgn->init_req(slow_result_path, control());
3334 result_io ->init_req(slow_result_path, i_o());
3335 result_mem->init_req(slow_result_path, reset_memory());
3336 result_val->init_req(slow_result_path, slow_val);
3337
3338 set_all_memory(_gvn.transform(result_mem));
3339 set_i_o( _gvn.transform(result_io));
3340 }
3341
3342 C->set_has_split_ifs(true); // Has chance for split-if optimization
3343 set_result(result_rgn, result_val);
3344 return true;
3345 }
3346
3347 //---------------------------load_mirror_from_klass----------------------------
3348 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3349 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3350 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3351 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3352 }
3353
3354 //-----------------------load_klass_from_mirror_common-------------------------
3355 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3356 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3357 // and branch to the given path on the region.
3358 // If never_see_null, take an uncommon trap on null, so we can optimistically
3359 // compile for the non-null case.
3360 // If the region is NULL, force never_see_null = true.
3361 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3362 bool never_see_null,
3363 RegionNode* region,
3364 int null_path,
3365 int offset) {
3366 if (region == NULL) never_see_null = true;
3367 Node* p = basic_plus_adr(mirror, offset);
3368 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3369 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3370 Node* null_ctl = top();
3371 kls = null_check_oop(kls, &null_ctl, never_see_null);
3372 if (region != NULL) {
3373 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3374 region->init_req(null_path, null_ctl);
3375 } else {
3376 assert(null_ctl == top(), "no loose ends");
3377 }
3378 return kls;
3379 }
3380
3381 //--------------------(inline_native_Class_query helpers)---------------------
3382 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3383 // Fall through if (mods & mask) == bits, take the guard otherwise.
3384 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3385 // Branch around if the given klass has the given modifier bit set.
3386 // Like generate_guard, adds a new path onto the region.
3387 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3388 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3389 Node* mask = intcon(modifier_mask);
3390 Node* bits = intcon(modifier_bits);
3391 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3392 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3393 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3394 return generate_fair_guard(bol, region);
3395 }
3396 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3397 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3398 }
3399
3400 //-------------------------inline_native_Class_query-------------------
3401 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3402 const Type* return_type = TypeInt::BOOL;
3403 Node* prim_return_value = top(); // what happens if it's a primitive class?
3404 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3405 bool expect_prim = false; // most of these guys expect to work on refs
3406
3407 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3408
3409 Node* mirror = argument(0);
3410 Node* obj = top();
3411
3412 switch (id) {
3413 case vmIntrinsics::_isInstance:
3414 // nothing is an instance of a primitive type
3415 prim_return_value = intcon(0);
3416 obj = argument(1);
3417 break;
3418 case vmIntrinsics::_getModifiers:
3419 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3420 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3421 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3422 break;
3423 case vmIntrinsics::_isInterface:
3424 prim_return_value = intcon(0);
3425 break;
3426 case vmIntrinsics::_isArray:
3427 prim_return_value = intcon(0);
3428 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3429 break;
3430 case vmIntrinsics::_isPrimitive:
3431 prim_return_value = intcon(1);
3432 expect_prim = true; // obviously
3433 break;
3434 case vmIntrinsics::_getSuperclass:
3435 prim_return_value = null();
3436 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3437 break;
3438 case vmIntrinsics::_getClassAccessFlags:
3439 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3440 return_type = TypeInt::INT; // not bool! 6297094
3441 break;
3442 default:
3443 fatal_unexpected_iid(id);
3444 break;
3445 }
3446
3447 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3448 if (mirror_con == NULL) return false; // cannot happen?
3449
3450 #ifndef PRODUCT
3451 if (C->print_intrinsics() || C->print_inlining()) {
3452 ciType* k = mirror_con->java_mirror_type();
3453 if (k) {
3454 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3455 k->print_name();
3456 tty->cr();
3457 }
3458 }
3459 #endif
3460
3461 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3462 RegionNode* region = new RegionNode(PATH_LIMIT);
3463 record_for_igvn(region);
3464 PhiNode* phi = new PhiNode(region, return_type);
3465
3466 // The mirror will never be null of Reflection.getClassAccessFlags, however
3467 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3468 // if it is. See bug 4774291.
3469
3470 // For Reflection.getClassAccessFlags(), the null check occurs in
3471 // the wrong place; see inline_unsafe_access(), above, for a similar
3472 // situation.
3473 mirror = null_check(mirror);
3474 // If mirror or obj is dead, only null-path is taken.
3475 if (stopped()) return true;
3476
3477 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3478
3479 // Now load the mirror's klass metaobject, and null-check it.
3480 // Side-effects region with the control path if the klass is null.
3481 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3482 // If kls is null, we have a primitive mirror.
3483 phi->init_req(_prim_path, prim_return_value);
3484 if (stopped()) { set_result(region, phi); return true; }
3485 bool safe_for_replace = (region->in(_prim_path) == top());
3486
3487 Node* p; // handy temp
3488 Node* null_ctl;
3489
3490 // Now that we have the non-null klass, we can perform the real query.
3491 // For constant classes, the query will constant-fold in LoadNode::Value.
3492 Node* query_value = top();
3493 switch (id) {
3494 case vmIntrinsics::_isInstance:
3495 // nothing is an instance of a primitive type
3496 query_value = gen_instanceof(obj, kls, safe_for_replace);
3497 break;
3498
3499 case vmIntrinsics::_getModifiers:
3500 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3501 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3502 break;
3503
3504 case vmIntrinsics::_isInterface:
3505 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3506 if (generate_interface_guard(kls, region) != NULL)
3507 // A guard was added. If the guard is taken, it was an interface.
3508 phi->add_req(intcon(1));
3509 // If we fall through, it's a plain class.
3510 query_value = intcon(0);
3511 break;
3512
3513 case vmIntrinsics::_isArray:
3514 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3515 if (generate_array_guard(kls, region) != NULL)
3516 // A guard was added. If the guard is taken, it was an array.
3517 phi->add_req(intcon(1));
3518 // If we fall through, it's a plain class.
3519 query_value = intcon(0);
3520 break;
3521
3522 case vmIntrinsics::_isPrimitive:
3523 query_value = intcon(0); // "normal" path produces false
3524 break;
3525
3526 case vmIntrinsics::_getSuperclass:
3527 // The rules here are somewhat unfortunate, but we can still do better
3528 // with random logic than with a JNI call.
3529 // Interfaces store null or Object as _super, but must report null.
3530 // Arrays store an intermediate super as _super, but must report Object.
3531 // Other types can report the actual _super.
3532 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3533 if (generate_interface_guard(kls, region) != NULL)
3534 // A guard was added. If the guard is taken, it was an interface.
3535 phi->add_req(null());
3536 if (generate_array_guard(kls, region) != NULL)
3537 // A guard was added. If the guard is taken, it was an array.
3538 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3539 // If we fall through, it's a plain class. Get its _super.
3540 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3541 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3542 null_ctl = top();
3543 kls = null_check_oop(kls, &null_ctl);
3544 if (null_ctl != top()) {
3545 // If the guard is taken, Object.superClass is null (both klass and mirror).
3546 region->add_req(null_ctl);
3547 phi ->add_req(null());
3548 }
3549 if (!stopped()) {
3550 query_value = load_mirror_from_klass(kls);
3551 }
3552 break;
3553
3554 case vmIntrinsics::_getClassAccessFlags:
3555 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3556 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3557 break;
3558
3559 default:
3560 fatal_unexpected_iid(id);
3561 break;
3562 }
3563
3564 // Fall-through is the normal case of a query to a real class.
3565 phi->init_req(1, query_value);
3566 region->init_req(1, control());
3567
3568 C->set_has_split_ifs(true); // Has chance for split-if optimization
3569 set_result(region, phi);
3570 return true;
3571 }
3572
3573 //-------------------------inline_Class_cast-------------------
3574 bool LibraryCallKit::inline_Class_cast() {
3575 Node* mirror = argument(0); // Class
3576 Node* obj = argument(1);
3577 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3578 if (mirror_con == NULL) {
3579 return false; // dead path (mirror->is_top()).
3580 }
3581 if (obj == NULL || obj->is_top()) {
3582 return false; // dead path
3583 }
3584 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3585
3586 // First, see if Class.cast() can be folded statically.
3587 // java_mirror_type() returns non-null for compile-time Class constants.
3588 ciType* tm = mirror_con->java_mirror_type();
3589 if (tm != NULL && tm->is_klass() &&
3590 tp != NULL && tp->klass() != NULL) {
3591 if (!tp->klass()->is_loaded()) {
3592 // Don't use intrinsic when class is not loaded.
3593 return false;
3594 } else {
3595 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3596 if (static_res == Compile::SSC_always_true) {
3597 // isInstance() is true - fold the code.
3598 set_result(obj);
3599 return true;
3600 } else if (static_res == Compile::SSC_always_false) {
3601 // Don't use intrinsic, have to throw ClassCastException.
3602 // If the reference is null, the non-intrinsic bytecode will
3603 // be optimized appropriately.
3604 return false;
3605 }
3606 }
3607 }
3608
3609 // Bailout intrinsic and do normal inlining if exception path is frequent.
3610 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3611 return false;
3612 }
3613
3614 // Generate dynamic checks.
3615 // Class.cast() is java implementation of _checkcast bytecode.
3616 // Do checkcast (Parse::do_checkcast()) optimizations here.
3617
3618 mirror = null_check(mirror);
3619 // If mirror is dead, only null-path is taken.
3620 if (stopped()) {
3621 return true;
3622 }
3623
3624 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3625 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3626 RegionNode* region = new RegionNode(PATH_LIMIT);
3627 record_for_igvn(region);
3628
3629 // Now load the mirror's klass metaobject, and null-check it.
3630 // If kls is null, we have a primitive mirror and
3631 // nothing is an instance of a primitive type.
3632 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3633
3634 Node* res = top();
3635 if (!stopped()) {
3636 Node* bad_type_ctrl = top();
3637 // Do checkcast optimizations.
3638 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3639 region->init_req(_bad_type_path, bad_type_ctrl);
3640 }
3641 if (region->in(_prim_path) != top() ||
3642 region->in(_bad_type_path) != top()) {
3643 // Let Interpreter throw ClassCastException.
3644 PreserveJVMState pjvms(this);
3645 set_control(_gvn.transform(region));
3646 uncommon_trap(Deoptimization::Reason_intrinsic,
3647 Deoptimization::Action_maybe_recompile);
3648 }
3649 if (!stopped()) {
3650 set_result(res);
3651 }
3652 return true;
3653 }
3654
3655
3656 //--------------------------inline_native_subtype_check------------------------
3657 // This intrinsic takes the JNI calls out of the heart of
3658 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3659 bool LibraryCallKit::inline_native_subtype_check() {
3660 // Pull both arguments off the stack.
3661 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3662 args[0] = argument(0);
3663 args[1] = argument(1);
3664 Node* klasses[2]; // corresponding Klasses: superk, subk
3665 klasses[0] = klasses[1] = top();
3666
3667 enum {
3668 // A full decision tree on {superc is prim, subc is prim}:
3669 _prim_0_path = 1, // {P,N} => false
3670 // {P,P} & superc!=subc => false
3671 _prim_same_path, // {P,P} & superc==subc => true
3672 _prim_1_path, // {N,P} => false
3673 _ref_subtype_path, // {N,N} & subtype check wins => true
3674 _both_ref_path, // {N,N} & subtype check loses => false
3675 PATH_LIMIT
3676 };
3677
3678 RegionNode* region = new RegionNode(PATH_LIMIT);
3679 Node* phi = new PhiNode(region, TypeInt::BOOL);
3680 record_for_igvn(region);
3681
3682 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3683 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3684 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3685
3686 // First null-check both mirrors and load each mirror's klass metaobject.
3687 int which_arg;
3688 for (which_arg = 0; which_arg <= 1; which_arg++) {
3689 Node* arg = args[which_arg];
3690 arg = null_check(arg);
3691 if (stopped()) break;
3692 args[which_arg] = arg;
3693
3694 Node* p = basic_plus_adr(arg, class_klass_offset);
3695 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3696 klasses[which_arg] = _gvn.transform(kls);
3697 }
3698
3699 // Having loaded both klasses, test each for null.
3700 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3701 for (which_arg = 0; which_arg <= 1; which_arg++) {
3702 Node* kls = klasses[which_arg];
3703 Node* null_ctl = top();
3704 kls = null_check_oop(kls, &null_ctl, never_see_null);
3705 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3706 region->init_req(prim_path, null_ctl);
3707 if (stopped()) break;
3708 klasses[which_arg] = kls;
3709 }
3710
3711 if (!stopped()) {
3712 // now we have two reference types, in klasses[0..1]
3713 Node* subk = klasses[1]; // the argument to isAssignableFrom
3714 Node* superk = klasses[0]; // the receiver
3715 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3716 // now we have a successful reference subtype check
3717 region->set_req(_ref_subtype_path, control());
3718 }
3719
3720 // If both operands are primitive (both klasses null), then
3721 // we must return true when they are identical primitives.
3722 // It is convenient to test this after the first null klass check.
3723 set_control(region->in(_prim_0_path)); // go back to first null check
3724 if (!stopped()) {
3725 // Since superc is primitive, make a guard for the superc==subc case.
3726 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3727 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3728 generate_guard(bol_eq, region, PROB_FAIR);
3729 if (region->req() == PATH_LIMIT+1) {
3730 // A guard was added. If the added guard is taken, superc==subc.
3731 region->swap_edges(PATH_LIMIT, _prim_same_path);
3732 region->del_req(PATH_LIMIT);
3733 }
3734 region->set_req(_prim_0_path, control()); // Not equal after all.
3735 }
3736
3737 // these are the only paths that produce 'true':
3738 phi->set_req(_prim_same_path, intcon(1));
3739 phi->set_req(_ref_subtype_path, intcon(1));
3740
3741 // pull together the cases:
3742 assert(region->req() == PATH_LIMIT, "sane region");
3743 for (uint i = 1; i < region->req(); i++) {
3744 Node* ctl = region->in(i);
3745 if (ctl == NULL || ctl == top()) {
3746 region->set_req(i, top());
3747 phi ->set_req(i, top());
3748 } else if (phi->in(i) == NULL) {
3749 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3750 }
3751 }
3752
3753 set_control(_gvn.transform(region));
3754 set_result(_gvn.transform(phi));
3755 return true;
3756 }
3757
3758 //---------------------generate_array_guard_common------------------------
3759 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3760 bool obj_array, bool not_array) {
3761
3762 if (stopped()) {
3763 return NULL;
3764 }
3765
3766 // If obj_array/non_array==false/false:
3767 // Branch around if the given klass is in fact an array (either obj or prim).
3768 // If obj_array/non_array==false/true:
3769 // Branch around if the given klass is not an array klass of any kind.
3770 // If obj_array/non_array==true/true:
3771 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3772 // If obj_array/non_array==true/false:
3773 // Branch around if the kls is an oop array (Object[] or subtype)
3774 //
3775 // Like generate_guard, adds a new path onto the region.
3776 jint layout_con = 0;
3777 Node* layout_val = get_layout_helper(kls, layout_con);
3778 if (layout_val == NULL) {
3779 bool query = (obj_array
3780 ? Klass::layout_helper_is_objArray(layout_con)
3781 : Klass::layout_helper_is_array(layout_con));
3782 if (query == not_array) {
3783 return NULL; // never a branch
3784 } else { // always a branch
3785 Node* always_branch = control();
3786 if (region != NULL)
3787 region->add_req(always_branch);
3788 set_control(top());
3789 return always_branch;
3790 }
3791 }
3792 // Now test the correct condition.
3793 jint nval = (obj_array
3794 ? ((jint)Klass::_lh_array_tag_type_value
3795 << Klass::_lh_array_tag_shift)
3796 : Klass::_lh_neutral_value);
3797 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3798 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3799 // invert the test if we are looking for a non-array
3800 if (not_array) btest = BoolTest(btest).negate();
3801 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3802 return generate_fair_guard(bol, region);
3803 }
3804
3805
3806 //-----------------------inline_native_newArray--------------------------
3807 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3808 bool LibraryCallKit::inline_native_newArray() {
3809 Node* mirror = argument(0);
3810 Node* count_val = argument(1);
3811
3812 mirror = null_check(mirror);
3813 // If mirror or obj is dead, only null-path is taken.
3814 if (stopped()) return true;
3815
3816 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3817 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3818 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3819 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3820 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3821
3822 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3823 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3824 result_reg, _slow_path);
3825 Node* normal_ctl = control();
3826 Node* no_array_ctl = result_reg->in(_slow_path);
3827
3828 // Generate code for the slow case. We make a call to newArray().
3829 set_control(no_array_ctl);
3830 if (!stopped()) {
3831 // Either the input type is void.class, or else the
3832 // array klass has not yet been cached. Either the
3833 // ensuing call will throw an exception, or else it
3834 // will cache the array klass for next time.
3835 PreserveJVMState pjvms(this);
3836 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3837 Node* slow_result = set_results_for_java_call(slow_call);
3838 // this->control() comes from set_results_for_java_call
3839 result_reg->set_req(_slow_path, control());
3840 result_val->set_req(_slow_path, slow_result);
3841 result_io ->set_req(_slow_path, i_o());
3842 result_mem->set_req(_slow_path, reset_memory());
3843 }
3844
3845 set_control(normal_ctl);
3846 if (!stopped()) {
3847 // Normal case: The array type has been cached in the java.lang.Class.
3848 // The following call works fine even if the array type is polymorphic.
3849 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3850 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3851 result_reg->init_req(_normal_path, control());
3852 result_val->init_req(_normal_path, obj);
3853 result_io ->init_req(_normal_path, i_o());
3854 result_mem->init_req(_normal_path, reset_memory());
3855 }
3856
3857 // Return the combined state.
3858 set_i_o( _gvn.transform(result_io) );
3859 set_all_memory( _gvn.transform(result_mem));
3860
3861 C->set_has_split_ifs(true); // Has chance for split-if optimization
3862 set_result(result_reg, result_val);
3863 return true;
3864 }
3865
3866 //----------------------inline_native_getLength--------------------------
3867 // public static native int java.lang.reflect.Array.getLength(Object array);
3868 bool LibraryCallKit::inline_native_getLength() {
3869 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3870
3871 Node* array = null_check(argument(0));
3872 // If array is dead, only null-path is taken.
3873 if (stopped()) return true;
3874
3875 // Deoptimize if it is a non-array.
3876 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3877
3878 if (non_array != NULL) {
3879 PreserveJVMState pjvms(this);
3880 set_control(non_array);
3881 uncommon_trap(Deoptimization::Reason_intrinsic,
3882 Deoptimization::Action_maybe_recompile);
3883 }
3884
3885 // If control is dead, only non-array-path is taken.
3886 if (stopped()) return true;
3887
3888 // The works fine even if the array type is polymorphic.
3889 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3890 Node* result = load_array_length(array);
3891
3892 C->set_has_split_ifs(true); // Has chance for split-if optimization
3893 set_result(result);
3894 return true;
3895 }
3896
3897 //------------------------inline_array_copyOf----------------------------
3898 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3899 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3900 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3901 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3902
3903 // Get the arguments.
3904 Node* original = argument(0);
3905 Node* start = is_copyOfRange? argument(1): intcon(0);
3906 Node* end = is_copyOfRange? argument(2): argument(1);
3907 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3908
3909 Node* newcopy = NULL;
3910
3911 // Set the original stack and the reexecute bit for the interpreter to reexecute
3912 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3913 { PreserveReexecuteState preexecs(this);
3914 jvms()->set_should_reexecute(true);
3915
3916 array_type_mirror = null_check(array_type_mirror);
3917 original = null_check(original);
3918
3919 // Check if a null path was taken unconditionally.
3920 if (stopped()) return true;
3921
3922 Node* orig_length = load_array_length(original);
3923
3924 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3925 klass_node = null_check(klass_node);
3926
3927 RegionNode* bailout = new RegionNode(1);
3928 record_for_igvn(bailout);
3929
3930 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3931 // Bail out if that is so.
3932 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3933 if (not_objArray != NULL) {
3934 // Improve the klass node's type from the new optimistic assumption:
3935 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3936 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3937 Node* cast = new CastPPNode(klass_node, akls);
3938 cast->init_req(0, control());
3939 klass_node = _gvn.transform(cast);
3940 }
3941
3942 // Bail out if either start or end is negative.
3943 generate_negative_guard(start, bailout, &start);
3944 generate_negative_guard(end, bailout, &end);
3945
3946 Node* length = end;
3947 if (_gvn.type(start) != TypeInt::ZERO) {
3948 length = _gvn.transform(new SubINode(end, start));
3949 }
3950
3951 // Bail out if length is negative.
3952 // Without this the new_array would throw
3953 // NegativeArraySizeException but IllegalArgumentException is what
3954 // should be thrown
3955 generate_negative_guard(length, bailout, &length);
3956
3957 if (bailout->req() > 1) {
3958 PreserveJVMState pjvms(this);
3959 set_control(_gvn.transform(bailout));
3960 uncommon_trap(Deoptimization::Reason_intrinsic,
3961 Deoptimization::Action_maybe_recompile);
3962 }
3963
3964 if (!stopped()) {
3965 // How many elements will we copy from the original?
3966 // The answer is MinI(orig_length - start, length).
3967 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3968 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3969
3970 // Generate a direct call to the right arraycopy function(s).
3971 // We know the copy is disjoint but we might not know if the
3972 // oop stores need checking.
3973 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3974 // This will fail a store-check if x contains any non-nulls.
3975
3976 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3977 // loads/stores but it is legal only if we're sure the
3978 // Arrays.copyOf would succeed. So we need all input arguments
3979 // to the copyOf to be validated, including that the copy to the
3980 // new array won't trigger an ArrayStoreException. That subtype
3981 // check can be optimized if we know something on the type of
3982 // the input array from type speculation.
3983 if (_gvn.type(klass_node)->singleton()) {
3984 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3985 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3986
3987 int test = C->static_subtype_check(superk, subk);
3988 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3989 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3990 if (t_original->speculative_type() != NULL) {
3991 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3992 }
3993 }
3994 }
3995
3996 bool validated = false;
3997 // Reason_class_check rather than Reason_intrinsic because we
3998 // want to intrinsify even if this traps.
3999 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4000 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4001 klass_node);
4002
4003 if (not_subtype_ctrl != top()) {
4004 PreserveJVMState pjvms(this);
4005 set_control(not_subtype_ctrl);
4006 uncommon_trap(Deoptimization::Reason_class_check,
4007 Deoptimization::Action_make_not_entrant);
4008 assert(stopped(), "Should be stopped");
4009 }
4010 validated = true;
4011 }
4012
4013 if (!stopped()) {
4014 newcopy = new_array(klass_node, length, 0); // no arguments to push
4015
4016 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
4017 load_object_klass(original), klass_node);
4018 if (!is_copyOfRange) {
4019 ac->set_copyof(validated);
4020 } else {
4021 ac->set_copyofrange(validated);
4022 }
4023 Node* n = _gvn.transform(ac);
4024 if (n == ac) {
4025 ac->connect_outputs(this);
4026 } else {
4027 assert(validated, "shouldn't transform if all arguments not validated");
4028 set_all_memory(n);
4029 }
4030 }
4031 }
4032 } // original reexecute is set back here
4033
4034 C->set_has_split_ifs(true); // Has chance for split-if optimization
4035 if (!stopped()) {
4036 set_result(newcopy);
4037 }
4038 return true;
4039 }
4040
4041
4042 //----------------------generate_virtual_guard---------------------------
4043 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4044 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4045 RegionNode* slow_region) {
4046 ciMethod* method = callee();
4047 int vtable_index = method->vtable_index();
4048 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4049 "bad index %d", vtable_index);
4050 // Get the Method* out of the appropriate vtable entry.
4051 int entry_offset = in_bytes(InstanceKlass::vtable_start_offset()) +
4052 vtable_index*vtableEntry::size_in_bytes() +
4053 vtableEntry::method_offset_in_bytes();
4054 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4055 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4056
4057 // Compare the target method with the expected method (e.g., Object.hashCode).
4058 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4059
4060 Node* native_call = makecon(native_call_addr);
4061 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4062 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4063
4064 return generate_slow_guard(test_native, slow_region);
4065 }
4066
4067 //-----------------------generate_method_call----------------------------
4068 // Use generate_method_call to make a slow-call to the real
4069 // method if the fast path fails. An alternative would be to
4070 // use a stub like OptoRuntime::slow_arraycopy_Java.
4071 // This only works for expanding the current library call,
4072 // not another intrinsic. (E.g., don't use this for making an
4073 // arraycopy call inside of the copyOf intrinsic.)
4074 CallJavaNode*
4075 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4076 // When compiling the intrinsic method itself, do not use this technique.
4077 guarantee(callee() != C->method(), "cannot make slow-call to self");
4078
4079 ciMethod* method = callee();
4080 // ensure the JVMS we have will be correct for this call
4081 guarantee(method_id == method->intrinsic_id(), "must match");
4082
4083 const TypeFunc* tf = TypeFunc::make(method);
4084 CallJavaNode* slow_call;
4085 if (is_static) {
4086 assert(!is_virtual, "");
4087 slow_call = new CallStaticJavaNode(C, tf,
4088 SharedRuntime::get_resolve_static_call_stub(),
4089 method, bci());
4090 } else if (is_virtual) {
4091 null_check_receiver();
4092 int vtable_index = Method::invalid_vtable_index;
4093 if (UseInlineCaches) {
4094 // Suppress the vtable call
4095 } else {
4096 // hashCode and clone are not a miranda methods,
4097 // so the vtable index is fixed.
4098 // No need to use the linkResolver to get it.
4099 vtable_index = method->vtable_index();
4100 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4101 "bad index %d", vtable_index);
4102 }
4103 slow_call = new CallDynamicJavaNode(tf,
4104 SharedRuntime::get_resolve_virtual_call_stub(),
4105 method, vtable_index, bci());
4106 } else { // neither virtual nor static: opt_virtual
4107 null_check_receiver();
4108 slow_call = new CallStaticJavaNode(C, tf,
4109 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4110 method, bci());
4111 slow_call->set_optimized_virtual(true);
4112 }
4113 set_arguments_for_java_call(slow_call);
4114 set_edges_for_java_call(slow_call);
4115 return slow_call;
4116 }
4117
4118
4119 /**
4120 * Build special case code for calls to hashCode on an object. This call may
4121 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4122 * slightly different code.
4123 */
4124 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4125 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4126 assert(!(is_virtual && is_static), "either virtual, special, or static");
4127
4128 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4129
4130 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4131 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4132 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4133 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4134 Node* obj = NULL;
4135 if (!is_static) {
4136 // Check for hashing null object
4137 obj = null_check_receiver();
4138 if (stopped()) return true; // unconditionally null
4139 result_reg->init_req(_null_path, top());
4140 result_val->init_req(_null_path, top());
4141 } else {
4142 // Do a null check, and return zero if null.
4143 // System.identityHashCode(null) == 0
4144 obj = argument(0);
4145 Node* null_ctl = top();
4146 obj = null_check_oop(obj, &null_ctl);
4147 result_reg->init_req(_null_path, null_ctl);
4148 result_val->init_req(_null_path, _gvn.intcon(0));
4149 }
4150
4151 // Unconditionally null? Then return right away.
4152 if (stopped()) {
4153 set_control( result_reg->in(_null_path));
4154 if (!stopped())
4155 set_result(result_val->in(_null_path));
4156 return true;
4157 }
4158
4159 // We only go to the fast case code if we pass a number of guards. The
4160 // paths which do not pass are accumulated in the slow_region.
4161 RegionNode* slow_region = new RegionNode(1);
4162 record_for_igvn(slow_region);
4163
4164 // If this is a virtual call, we generate a funny guard. We pull out
4165 // the vtable entry corresponding to hashCode() from the target object.
4166 // If the target method which we are calling happens to be the native
4167 // Object hashCode() method, we pass the guard. We do not need this
4168 // guard for non-virtual calls -- the caller is known to be the native
4169 // Object hashCode().
4170 if (is_virtual) {
4171 // After null check, get the object's klass.
4172 Node* obj_klass = load_object_klass(obj);
4173 generate_virtual_guard(obj_klass, slow_region);
4174 }
4175
4176 // Get the header out of the object, use LoadMarkNode when available
4177 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4178 // The control of the load must be NULL. Otherwise, the load can move before
4179 // the null check after castPP removal.
4180 Node* no_ctrl = NULL;
4181 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4182
4183 // Test the header to see if it is unlocked.
4184 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4185 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4186 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4187 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4188 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4189
4190 generate_slow_guard(test_unlocked, slow_region);
4191
4192 // Get the hash value and check to see that it has been properly assigned.
4193 // We depend on hash_mask being at most 32 bits and avoid the use of
4194 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4195 // vm: see markOop.hpp.
4196 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4197 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4198 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4199 // This hack lets the hash bits live anywhere in the mark object now, as long
4200 // as the shift drops the relevant bits into the low 32 bits. Note that
4201 // Java spec says that HashCode is an int so there's no point in capturing
4202 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4203 hshifted_header = ConvX2I(hshifted_header);
4204 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4205
4206 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4207 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4208 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4209
4210 generate_slow_guard(test_assigned, slow_region);
4211
4212 Node* init_mem = reset_memory();
4213 // fill in the rest of the null path:
4214 result_io ->init_req(_null_path, i_o());
4215 result_mem->init_req(_null_path, init_mem);
4216
4217 result_val->init_req(_fast_path, hash_val);
4218 result_reg->init_req(_fast_path, control());
4219 result_io ->init_req(_fast_path, i_o());
4220 result_mem->init_req(_fast_path, init_mem);
4221
4222 // Generate code for the slow case. We make a call to hashCode().
4223 set_control(_gvn.transform(slow_region));
4224 if (!stopped()) {
4225 // No need for PreserveJVMState, because we're using up the present state.
4226 set_all_memory(init_mem);
4227 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4228 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4229 Node* slow_result = set_results_for_java_call(slow_call);
4230 // this->control() comes from set_results_for_java_call
4231 result_reg->init_req(_slow_path, control());
4232 result_val->init_req(_slow_path, slow_result);
4233 result_io ->set_req(_slow_path, i_o());
4234 result_mem ->set_req(_slow_path, reset_memory());
4235 }
4236
4237 // Return the combined state.
4238 set_i_o( _gvn.transform(result_io) );
4239 set_all_memory( _gvn.transform(result_mem));
4240
4241 set_result(result_reg, result_val);
4242 return true;
4243 }
4244
4245 //---------------------------inline_native_getClass----------------------------
4246 // public final native Class<?> java.lang.Object.getClass();
4247 //
4248 // Build special case code for calls to getClass on an object.
4249 bool LibraryCallKit::inline_native_getClass() {
4250 Node* obj = null_check_receiver();
4251 if (stopped()) return true;
4252 set_result(load_mirror_from_klass(load_object_klass(obj)));
4253 return true;
4254 }
4255
4256 //-----------------inline_native_Reflection_getCallerClass---------------------
4257 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4258 //
4259 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4260 //
4261 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4262 // in that it must skip particular security frames and checks for
4263 // caller sensitive methods.
4264 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4265 #ifndef PRODUCT
4266 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4267 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4268 }
4269 #endif
4270
4271 if (!jvms()->has_method()) {
4272 #ifndef PRODUCT
4273 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4274 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4275 }
4276 #endif
4277 return false;
4278 }
4279
4280 // Walk back up the JVM state to find the caller at the required
4281 // depth.
4282 JVMState* caller_jvms = jvms();
4283
4284 // Cf. JVM_GetCallerClass
4285 // NOTE: Start the loop at depth 1 because the current JVM state does
4286 // not include the Reflection.getCallerClass() frame.
4287 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4288 ciMethod* m = caller_jvms->method();
4289 switch (n) {
4290 case 0:
4291 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4292 break;
4293 case 1:
4294 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4295 if (!m->caller_sensitive()) {
4296 #ifndef PRODUCT
4297 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4298 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4299 }
4300 #endif
4301 return false; // bail-out; let JVM_GetCallerClass do the work
4302 }
4303 break;
4304 default:
4305 if (!m->is_ignored_by_security_stack_walk()) {
4306 // We have reached the desired frame; return the holder class.
4307 // Acquire method holder as java.lang.Class and push as constant.
4308 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4309 ciInstance* caller_mirror = caller_klass->java_mirror();
4310 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4311
4312 #ifndef PRODUCT
4313 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4314 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());
4315 tty->print_cr(" JVM state at this point:");
4316 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4317 ciMethod* m = jvms()->of_depth(i)->method();
4318 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4319 }
4320 }
4321 #endif
4322 return true;
4323 }
4324 break;
4325 }
4326 }
4327
4328 #ifndef PRODUCT
4329 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4330 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4331 tty->print_cr(" JVM state at this point:");
4332 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4333 ciMethod* m = jvms()->of_depth(i)->method();
4334 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4335 }
4336 }
4337 #endif
4338
4339 return false; // bail-out; let JVM_GetCallerClass do the work
4340 }
4341
4342 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4343 Node* arg = argument(0);
4344 Node* result = NULL;
4345
4346 switch (id) {
4347 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4348 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4349 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4350 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4351
4352 case vmIntrinsics::_doubleToLongBits: {
4353 // two paths (plus control) merge in a wood
4354 RegionNode *r = new RegionNode(3);
4355 Node *phi = new PhiNode(r, TypeLong::LONG);
4356
4357 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4358 // Build the boolean node
4359 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4360
4361 // Branch either way.
4362 // NaN case is less traveled, which makes all the difference.
4363 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4364 Node *opt_isnan = _gvn.transform(ifisnan);
4365 assert( opt_isnan->is_If(), "Expect an IfNode");
4366 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4367 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4368
4369 set_control(iftrue);
4370
4371 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4372 Node *slow_result = longcon(nan_bits); // return NaN
4373 phi->init_req(1, _gvn.transform( slow_result ));
4374 r->init_req(1, iftrue);
4375
4376 // Else fall through
4377 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4378 set_control(iffalse);
4379
4380 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4381 r->init_req(2, iffalse);
4382
4383 // Post merge
4384 set_control(_gvn.transform(r));
4385 record_for_igvn(r);
4386
4387 C->set_has_split_ifs(true); // Has chance for split-if optimization
4388 result = phi;
4389 assert(result->bottom_type()->isa_long(), "must be");
4390 break;
4391 }
4392
4393 case vmIntrinsics::_floatToIntBits: {
4394 // two paths (plus control) merge in a wood
4395 RegionNode *r = new RegionNode(3);
4396 Node *phi = new PhiNode(r, TypeInt::INT);
4397
4398 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4399 // Build the boolean node
4400 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4401
4402 // Branch either way.
4403 // NaN case is less traveled, which makes all the difference.
4404 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4405 Node *opt_isnan = _gvn.transform(ifisnan);
4406 assert( opt_isnan->is_If(), "Expect an IfNode");
4407 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4408 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4409
4410 set_control(iftrue);
4411
4412 static const jint nan_bits = 0x7fc00000;
4413 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4414 phi->init_req(1, _gvn.transform( slow_result ));
4415 r->init_req(1, iftrue);
4416
4417 // Else fall through
4418 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4419 set_control(iffalse);
4420
4421 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4422 r->init_req(2, iffalse);
4423
4424 // Post merge
4425 set_control(_gvn.transform(r));
4426 record_for_igvn(r);
4427
4428 C->set_has_split_ifs(true); // Has chance for split-if optimization
4429 result = phi;
4430 assert(result->bottom_type()->isa_int(), "must be");
4431 break;
4432 }
4433
4434 default:
4435 fatal_unexpected_iid(id);
4436 break;
4437 }
4438 set_result(_gvn.transform(result));
4439 return true;
4440 }
4441
4442 //----------------------inline_unsafe_copyMemory-------------------------
4443 // public native void Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4444 bool LibraryCallKit::inline_unsafe_copyMemory() {
4445 if (callee()->is_static()) return false; // caller must have the capability!
4446 null_check_receiver(); // null-check receiver
4447 if (stopped()) return true;
4448
4449 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4450
4451 Node* src_ptr = argument(1); // type: oop
4452 Node* src_off = ConvL2X(argument(2)); // type: long
4453 Node* dst_ptr = argument(4); // type: oop
4454 Node* dst_off = ConvL2X(argument(5)); // type: long
4455 Node* size = ConvL2X(argument(7)); // type: long
4456
4457 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4458 "fieldOffset must be byte-scaled");
4459
4460 Node* src = make_unsafe_address(src_ptr, src_off);
4461 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4462
4463 // Conservatively insert a memory barrier on all memory slices.
4464 // Do not let writes of the copy source or destination float below the copy.
4465 insert_mem_bar(Op_MemBarCPUOrder);
4466
4467 // Call it. Note that the length argument is not scaled.
4468 make_runtime_call(RC_LEAF|RC_NO_FP,
4469 OptoRuntime::fast_arraycopy_Type(),
4470 StubRoutines::unsafe_arraycopy(),
4471 "unsafe_arraycopy",
4472 TypeRawPtr::BOTTOM,
4473 src, dst, size XTOP);
4474
4475 // Do not let reads of the copy destination float above the copy.
4476 insert_mem_bar(Op_MemBarCPUOrder);
4477
4478 return true;
4479 }
4480
4481 //------------------------clone_coping-----------------------------------
4482 // Helper function for inline_native_clone.
4483 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4484 assert(obj_size != NULL, "");
4485 Node* raw_obj = alloc_obj->in(1);
4486 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4487
4488 AllocateNode* alloc = NULL;
4489 if (ReduceBulkZeroing) {
4490 // We will be completely responsible for initializing this object -
4491 // mark Initialize node as complete.
4492 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4493 // The object was just allocated - there should be no any stores!
4494 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4495 // Mark as complete_with_arraycopy so that on AllocateNode
4496 // expansion, we know this AllocateNode is initialized by an array
4497 // copy and a StoreStore barrier exists after the array copy.
4498 alloc->initialization()->set_complete_with_arraycopy();
4499 }
4500
4501 // Copy the fastest available way.
4502 // TODO: generate fields copies for small objects instead.
4503 Node* src = obj;
4504 Node* dest = alloc_obj;
4505 Node* size = _gvn.transform(obj_size);
4506
4507 // Exclude the header but include array length to copy by 8 bytes words.
4508 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4509 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4510 instanceOopDesc::base_offset_in_bytes();
4511 // base_off:
4512 // 8 - 32-bit VM
4513 // 12 - 64-bit VM, compressed klass
4514 // 16 - 64-bit VM, normal klass
4515 if (base_off % BytesPerLong != 0) {
4516 assert(UseCompressedClassPointers, "");
4517 if (is_array) {
4518 // Exclude length to copy by 8 bytes words.
4519 base_off += sizeof(int);
4520 } else {
4521 // Include klass to copy by 8 bytes words.
4522 base_off = instanceOopDesc::klass_offset_in_bytes();
4523 }
4524 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4525 }
4526 src = basic_plus_adr(src, base_off);
4527 dest = basic_plus_adr(dest, base_off);
4528
4529 // Compute the length also, if needed:
4530 Node* countx = size;
4531 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4532 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4533
4534 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4535
4536 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4537 ac->set_clonebasic();
4538 Node* n = _gvn.transform(ac);
4539 if (n == ac) {
4540 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4541 } else {
4542 set_all_memory(n);
4543 }
4544
4545 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4546 if (card_mark) {
4547 assert(!is_array, "");
4548 // Put in store barrier for any and all oops we are sticking
4549 // into this object. (We could avoid this if we could prove
4550 // that the object type contains no oop fields at all.)
4551 Node* no_particular_value = NULL;
4552 Node* no_particular_field = NULL;
4553 int raw_adr_idx = Compile::AliasIdxRaw;
4554 post_barrier(control(),
4555 memory(raw_adr_type),
4556 alloc_obj,
4557 no_particular_field,
4558 raw_adr_idx,
4559 no_particular_value,
4560 T_OBJECT,
4561 false);
4562 }
4563
4564 // Do not let reads from the cloned object float above the arraycopy.
4565 if (alloc != NULL) {
4566 // Do not let stores that initialize this object be reordered with
4567 // a subsequent store that would make this object accessible by
4568 // other threads.
4569 // Record what AllocateNode this StoreStore protects so that
4570 // escape analysis can go from the MemBarStoreStoreNode to the
4571 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4572 // based on the escape status of the AllocateNode.
4573 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4574 } else {
4575 insert_mem_bar(Op_MemBarCPUOrder);
4576 }
4577 }
4578
4579 //------------------------inline_native_clone----------------------------
4580 // protected native Object java.lang.Object.clone();
4581 //
4582 // Here are the simple edge cases:
4583 // null receiver => normal trap
4584 // virtual and clone was overridden => slow path to out-of-line clone
4585 // not cloneable or finalizer => slow path to out-of-line Object.clone
4586 //
4587 // The general case has two steps, allocation and copying.
4588 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4589 //
4590 // Copying also has two cases, oop arrays and everything else.
4591 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4592 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4593 //
4594 // These steps fold up nicely if and when the cloned object's klass
4595 // can be sharply typed as an object array, a type array, or an instance.
4596 //
4597 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4598 PhiNode* result_val;
4599
4600 // Set the reexecute bit for the interpreter to reexecute
4601 // the bytecode that invokes Object.clone if deoptimization happens.
4602 { PreserveReexecuteState preexecs(this);
4603 jvms()->set_should_reexecute(true);
4604
4605 Node* obj = null_check_receiver();
4606 if (stopped()) return true;
4607
4608 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4609
4610 // If we are going to clone an instance, we need its exact type to
4611 // know the number and types of fields to convert the clone to
4612 // loads/stores. Maybe a speculative type can help us.
4613 if (!obj_type->klass_is_exact() &&
4614 obj_type->speculative_type() != NULL &&
4615 obj_type->speculative_type()->is_instance_klass()) {
4616 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4617 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4618 !spec_ik->has_injected_fields()) {
4619 ciKlass* k = obj_type->klass();
4620 if (!k->is_instance_klass() ||
4621 k->as_instance_klass()->is_interface() ||
4622 k->as_instance_klass()->has_subklass()) {
4623 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4624 }
4625 }
4626 }
4627
4628 Node* obj_klass = load_object_klass(obj);
4629 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4630 const TypeOopPtr* toop = ((tklass != NULL)
4631 ? tklass->as_instance_type()
4632 : TypeInstPtr::NOTNULL);
4633
4634 // Conservatively insert a memory barrier on all memory slices.
4635 // Do not let writes into the original float below the clone.
4636 insert_mem_bar(Op_MemBarCPUOrder);
4637
4638 // paths into result_reg:
4639 enum {
4640 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4641 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4642 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4643 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4644 PATH_LIMIT
4645 };
4646 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4647 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4648 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4649 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4650 record_for_igvn(result_reg);
4651
4652 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4653 int raw_adr_idx = Compile::AliasIdxRaw;
4654
4655 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4656 if (array_ctl != NULL) {
4657 // It's an array.
4658 PreserveJVMState pjvms(this);
4659 set_control(array_ctl);
4660 Node* obj_length = load_array_length(obj);
4661 Node* obj_size = NULL;
4662 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4663
4664 if (!use_ReduceInitialCardMarks()) {
4665 // If it is an oop array, it requires very special treatment,
4666 // because card marking is required on each card of the array.
4667 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4668 if (is_obja != NULL) {
4669 PreserveJVMState pjvms2(this);
4670 set_control(is_obja);
4671 // Generate a direct call to the right arraycopy function(s).
4672 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4673 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4674 ac->set_cloneoop();
4675 Node* n = _gvn.transform(ac);
4676 assert(n == ac, "cannot disappear");
4677 ac->connect_outputs(this);
4678
4679 result_reg->init_req(_objArray_path, control());
4680 result_val->init_req(_objArray_path, alloc_obj);
4681 result_i_o ->set_req(_objArray_path, i_o());
4682 result_mem ->set_req(_objArray_path, reset_memory());
4683 }
4684 }
4685 // Otherwise, there are no card marks to worry about.
4686 // (We can dispense with card marks if we know the allocation
4687 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4688 // causes the non-eden paths to take compensating steps to
4689 // simulate a fresh allocation, so that no further
4690 // card marks are required in compiled code to initialize
4691 // the object.)
4692
4693 if (!stopped()) {
4694 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4695
4696 // Present the results of the copy.
4697 result_reg->init_req(_array_path, control());
4698 result_val->init_req(_array_path, alloc_obj);
4699 result_i_o ->set_req(_array_path, i_o());
4700 result_mem ->set_req(_array_path, reset_memory());
4701 }
4702 }
4703
4704 // We only go to the instance fast case code if we pass a number of guards.
4705 // The paths which do not pass are accumulated in the slow_region.
4706 RegionNode* slow_region = new RegionNode(1);
4707 record_for_igvn(slow_region);
4708 if (!stopped()) {
4709 // It's an instance (we did array above). Make the slow-path tests.
4710 // If this is a virtual call, we generate a funny guard. We grab
4711 // the vtable entry corresponding to clone() from the target object.
4712 // If the target method which we are calling happens to be the
4713 // Object clone() method, we pass the guard. We do not need this
4714 // guard for non-virtual calls; the caller is known to be the native
4715 // Object clone().
4716 if (is_virtual) {
4717 generate_virtual_guard(obj_klass, slow_region);
4718 }
4719
4720 // The object must be cloneable and must not have a finalizer.
4721 // Both of these conditions may be checked in a single test.
4722 // We could optimize the cloneable test further, but we don't care.
4723 generate_access_flags_guard(obj_klass,
4724 // Test both conditions:
4725 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4726 // Must be cloneable but not finalizer:
4727 JVM_ACC_IS_CLONEABLE,
4728 slow_region);
4729 }
4730
4731 if (!stopped()) {
4732 // It's an instance, and it passed the slow-path tests.
4733 PreserveJVMState pjvms(this);
4734 Node* obj_size = NULL;
4735 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4736 // is reexecuted if deoptimization occurs and there could be problems when merging
4737 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4738 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4739
4740 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4741
4742 // Present the results of the slow call.
4743 result_reg->init_req(_instance_path, control());
4744 result_val->init_req(_instance_path, alloc_obj);
4745 result_i_o ->set_req(_instance_path, i_o());
4746 result_mem ->set_req(_instance_path, reset_memory());
4747 }
4748
4749 // Generate code for the slow case. We make a call to clone().
4750 set_control(_gvn.transform(slow_region));
4751 if (!stopped()) {
4752 PreserveJVMState pjvms(this);
4753 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4754 Node* slow_result = set_results_for_java_call(slow_call);
4755 // this->control() comes from set_results_for_java_call
4756 result_reg->init_req(_slow_path, control());
4757 result_val->init_req(_slow_path, slow_result);
4758 result_i_o ->set_req(_slow_path, i_o());
4759 result_mem ->set_req(_slow_path, reset_memory());
4760 }
4761
4762 // Return the combined state.
4763 set_control( _gvn.transform(result_reg));
4764 set_i_o( _gvn.transform(result_i_o));
4765 set_all_memory( _gvn.transform(result_mem));
4766 } // original reexecute is set back here
4767
4768 set_result(_gvn.transform(result_val));
4769 return true;
4770 }
4771
4772 // If we have a tighly coupled allocation, the arraycopy may take care
4773 // of the array initialization. If one of the guards we insert between
4774 // the allocation and the arraycopy causes a deoptimization, an
4775 // unitialized array will escape the compiled method. To prevent that
4776 // we set the JVM state for uncommon traps between the allocation and
4777 // the arraycopy to the state before the allocation so, in case of
4778 // deoptimization, we'll reexecute the allocation and the
4779 // initialization.
4780 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4781 if (alloc != NULL) {
4782 ciMethod* trap_method = alloc->jvms()->method();
4783 int trap_bci = alloc->jvms()->bci();
4784
4785 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4786 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4787 // Make sure there's no store between the allocation and the
4788 // arraycopy otherwise visible side effects could be rexecuted
4789 // in case of deoptimization and cause incorrect execution.
4790 bool no_interfering_store = true;
4791 Node* mem = alloc->in(TypeFunc::Memory);
4792 if (mem->is_MergeMem()) {
4793 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4794 Node* n = mms.memory();
4795 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4796 assert(n->is_Store(), "what else?");
4797 no_interfering_store = false;
4798 break;
4799 }
4800 }
4801 } else {
4802 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4803 Node* n = mms.memory();
4804 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4805 assert(n->is_Store(), "what else?");
4806 no_interfering_store = false;
4807 break;
4808 }
4809 }
4810 }
4811
4812 if (no_interfering_store) {
4813 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4814 uint size = alloc->req();
4815 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4816 old_jvms->set_map(sfpt);
4817 for (uint i = 0; i < size; i++) {
4818 sfpt->init_req(i, alloc->in(i));
4819 }
4820 // re-push array length for deoptimization
4821 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4822 old_jvms->set_sp(old_jvms->sp()+1);
4823 old_jvms->set_monoff(old_jvms->monoff()+1);
4824 old_jvms->set_scloff(old_jvms->scloff()+1);
4825 old_jvms->set_endoff(old_jvms->endoff()+1);
4826 old_jvms->set_should_reexecute(true);
4827
4828 sfpt->set_i_o(map()->i_o());
4829 sfpt->set_memory(map()->memory());
4830 sfpt->set_control(map()->control());
4831
4832 JVMState* saved_jvms = jvms();
4833 saved_reexecute_sp = _reexecute_sp;
4834
4835 set_jvms(sfpt->jvms());
4836 _reexecute_sp = jvms()->sp();
4837
4838 return saved_jvms;
4839 }
4840 }
4841 }
4842 return NULL;
4843 }
4844
4845 // In case of a deoptimization, we restart execution at the
4846 // allocation, allocating a new array. We would leave an uninitialized
4847 // array in the heap that GCs wouldn't expect. Move the allocation
4848 // after the traps so we don't allocate the array if we
4849 // deoptimize. This is possible because tightly_coupled_allocation()
4850 // guarantees there's no observer of the allocated array at this point
4851 // and the control flow is simple enough.
4852 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4853 if (saved_jvms != NULL && !stopped()) {
4854 assert(alloc != NULL, "only with a tightly coupled allocation");
4855 // restore JVM state to the state at the arraycopy
4856 saved_jvms->map()->set_control(map()->control());
4857 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4858 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4859 // If we've improved the types of some nodes (null check) while
4860 // emitting the guards, propagate them to the current state
4861 map()->replaced_nodes().apply(saved_jvms->map());
4862 set_jvms(saved_jvms);
4863 _reexecute_sp = saved_reexecute_sp;
4864
4865 // Remove the allocation from above the guards
4866 CallProjections callprojs;
4867 alloc->extract_projections(&callprojs, true);
4868 InitializeNode* init = alloc->initialization();
4869 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4870 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4871 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4872 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4873
4874 // move the allocation here (after the guards)
4875 _gvn.hash_delete(alloc);
4876 alloc->set_req(TypeFunc::Control, control());
4877 alloc->set_req(TypeFunc::I_O, i_o());
4878 Node *mem = reset_memory();
4879 set_all_memory(mem);
4880 alloc->set_req(TypeFunc::Memory, mem);
4881 set_control(init->proj_out(TypeFunc::Control));
4882 set_i_o(callprojs.fallthrough_ioproj);
4883
4884 // Update memory as done in GraphKit::set_output_for_allocation()
4885 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4886 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4887 if (ary_type->isa_aryptr() && length_type != NULL) {
4888 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4889 }
4890 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4891 int elemidx = C->get_alias_index(telemref);
4892 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4893 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4894
4895 Node* allocx = _gvn.transform(alloc);
4896 assert(allocx == alloc, "where has the allocation gone?");
4897 assert(dest->is_CheckCastPP(), "not an allocation result?");
4898
4899 _gvn.hash_delete(dest);
4900 dest->set_req(0, control());
4901 Node* destx = _gvn.transform(dest);
4902 assert(destx == dest, "where has the allocation result gone?");
4903 }
4904 }
4905
4906
4907 //------------------------------inline_arraycopy-----------------------
4908 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4909 // Object dest, int destPos,
4910 // int length);
4911 bool LibraryCallKit::inline_arraycopy() {
4912 // Get the arguments.
4913 Node* src = argument(0); // type: oop
4914 Node* src_offset = argument(1); // type: int
4915 Node* dest = argument(2); // type: oop
4916 Node* dest_offset = argument(3); // type: int
4917 Node* length = argument(4); // type: int
4918
4919
4920 // Check for allocation before we add nodes that would confuse
4921 // tightly_coupled_allocation()
4922 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4923
4924 int saved_reexecute_sp = -1;
4925 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4926 // See arraycopy_restore_alloc_state() comment
4927 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4928 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4929 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4930 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4931
4932 // The following tests must be performed
4933 // (1) src and dest are arrays.
4934 // (2) src and dest arrays must have elements of the same BasicType
4935 // (3) src and dest must not be null.
4936 // (4) src_offset must not be negative.
4937 // (5) dest_offset must not be negative.
4938 // (6) length must not be negative.
4939 // (7) src_offset + length must not exceed length of src.
4940 // (8) dest_offset + length must not exceed length of dest.
4941 // (9) each element of an oop array must be assignable
4942
4943 // (3) src and dest must not be null.
4944 // always do this here because we need the JVM state for uncommon traps
4945 Node* null_ctl = top();
4946 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4947 assert(null_ctl->is_top(), "no null control here");
4948 dest = null_check(dest, T_ARRAY);
4949
4950 if (!can_emit_guards) {
4951 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4952 // guards but the arraycopy node could still take advantage of a
4953 // tightly allocated allocation. tightly_coupled_allocation() is
4954 // called again to make sure it takes the null check above into
4955 // account: the null check is mandatory and if it caused an
4956 // uncommon trap to be emitted then the allocation can't be
4957 // considered tightly coupled in this context.
4958 alloc = tightly_coupled_allocation(dest, NULL);
4959 }
4960
4961 bool validated = false;
4962
4963 const Type* src_type = _gvn.type(src);
4964 const Type* dest_type = _gvn.type(dest);
4965 const TypeAryPtr* top_src = src_type->isa_aryptr();
4966 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4967
4968 // Do we have the type of src?
4969 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4970 // Do we have the type of dest?
4971 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4972 // Is the type for src from speculation?
4973 bool src_spec = false;
4974 // Is the type for dest from speculation?
4975 bool dest_spec = false;
4976
4977 if ((!has_src || !has_dest) && can_emit_guards) {
4978 // We don't have sufficient type information, let's see if
4979 // speculative types can help. We need to have types for both src
4980 // and dest so that it pays off.
4981
4982 // Do we already have or could we have type information for src
4983 bool could_have_src = has_src;
4984 // Do we already have or could we have type information for dest
4985 bool could_have_dest = has_dest;
4986
4987 ciKlass* src_k = NULL;
4988 if (!has_src) {
4989 src_k = src_type->speculative_type_not_null();
4990 if (src_k != NULL && src_k->is_array_klass()) {
4991 could_have_src = true;
4992 }
4993 }
4994
4995 ciKlass* dest_k = NULL;
4996 if (!has_dest) {
4997 dest_k = dest_type->speculative_type_not_null();
4998 if (dest_k != NULL && dest_k->is_array_klass()) {
4999 could_have_dest = true;
5000 }
5001 }
5002
5003 if (could_have_src && could_have_dest) {
5004 // This is going to pay off so emit the required guards
5005 if (!has_src) {
5006 src = maybe_cast_profiled_obj(src, src_k, true);
5007 src_type = _gvn.type(src);
5008 top_src = src_type->isa_aryptr();
5009 has_src = (top_src != NULL && top_src->klass() != NULL);
5010 src_spec = true;
5011 }
5012 if (!has_dest) {
5013 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5014 dest_type = _gvn.type(dest);
5015 top_dest = dest_type->isa_aryptr();
5016 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5017 dest_spec = true;
5018 }
5019 }
5020 }
5021
5022 if (has_src && has_dest && can_emit_guards) {
5023 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5024 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5025 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5026 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5027
5028 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5029 // If both arrays are object arrays then having the exact types
5030 // for both will remove the need for a subtype check at runtime
5031 // before the call and may make it possible to pick a faster copy
5032 // routine (without a subtype check on every element)
5033 // Do we have the exact type of src?
5034 bool could_have_src = src_spec;
5035 // Do we have the exact type of dest?
5036 bool could_have_dest = dest_spec;
5037 ciKlass* src_k = top_src->klass();
5038 ciKlass* dest_k = top_dest->klass();
5039 if (!src_spec) {
5040 src_k = src_type->speculative_type_not_null();
5041 if (src_k != NULL && src_k->is_array_klass()) {
5042 could_have_src = true;
5043 }
5044 }
5045 if (!dest_spec) {
5046 dest_k = dest_type->speculative_type_not_null();
5047 if (dest_k != NULL && dest_k->is_array_klass()) {
5048 could_have_dest = true;
5049 }
5050 }
5051 if (could_have_src && could_have_dest) {
5052 // If we can have both exact types, emit the missing guards
5053 if (could_have_src && !src_spec) {
5054 src = maybe_cast_profiled_obj(src, src_k, true);
5055 }
5056 if (could_have_dest && !dest_spec) {
5057 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5058 }
5059 }
5060 }
5061 }
5062
5063 ciMethod* trap_method = method();
5064 int trap_bci = bci();
5065 if (saved_jvms != NULL) {
5066 trap_method = alloc->jvms()->method();
5067 trap_bci = alloc->jvms()->bci();
5068 }
5069
5070 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5071 can_emit_guards &&
5072 !src->is_top() && !dest->is_top()) {
5073 // validate arguments: enables transformation the ArrayCopyNode
5074 validated = true;
5075
5076 RegionNode* slow_region = new RegionNode(1);
5077 record_for_igvn(slow_region);
5078
5079 // (1) src and dest are arrays.
5080 generate_non_array_guard(load_object_klass(src), slow_region);
5081 generate_non_array_guard(load_object_klass(dest), slow_region);
5082
5083 // (2) src and dest arrays must have elements of the same BasicType
5084 // done at macro expansion or at Ideal transformation time
5085
5086 // (4) src_offset must not be negative.
5087 generate_negative_guard(src_offset, slow_region);
5088
5089 // (5) dest_offset must not be negative.
5090 generate_negative_guard(dest_offset, slow_region);
5091
5092 // (7) src_offset + length must not exceed length of src.
5093 generate_limit_guard(src_offset, length,
5094 load_array_length(src),
5095 slow_region);
5096
5097 // (8) dest_offset + length must not exceed length of dest.
5098 generate_limit_guard(dest_offset, length,
5099 load_array_length(dest),
5100 slow_region);
5101
5102 // (9) each element of an oop array must be assignable
5103 Node* src_klass = load_object_klass(src);
5104 Node* dest_klass = load_object_klass(dest);
5105 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5106
5107 if (not_subtype_ctrl != top()) {
5108 PreserveJVMState pjvms(this);
5109 set_control(not_subtype_ctrl);
5110 uncommon_trap(Deoptimization::Reason_intrinsic,
5111 Deoptimization::Action_make_not_entrant);
5112 assert(stopped(), "Should be stopped");
5113 }
5114 {
5115 PreserveJVMState pjvms(this);
5116 set_control(_gvn.transform(slow_region));
5117 uncommon_trap(Deoptimization::Reason_intrinsic,
5118 Deoptimization::Action_make_not_entrant);
5119 assert(stopped(), "Should be stopped");
5120 }
5121 }
5122
5123 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5124
5125 if (stopped()) {
5126 return true;
5127 }
5128
5129 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5130 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5131 // so the compiler has a chance to eliminate them: during macro expansion,
5132 // we have to set their control (CastPP nodes are eliminated).
5133 load_object_klass(src), load_object_klass(dest),
5134 load_array_length(src), load_array_length(dest));
5135
5136 ac->set_arraycopy(validated);
5137
5138 Node* n = _gvn.transform(ac);
5139 if (n == ac) {
5140 ac->connect_outputs(this);
5141 } else {
5142 assert(validated, "shouldn't transform if all arguments not validated");
5143 set_all_memory(n);
5144 }
5145
5146 return true;
5147 }
5148
5149
5150 // Helper function which determines if an arraycopy immediately follows
5151 // an allocation, with no intervening tests or other escapes for the object.
5152 AllocateArrayNode*
5153 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5154 RegionNode* slow_region) {
5155 if (stopped()) return NULL; // no fast path
5156 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5157
5158 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5159 if (alloc == NULL) return NULL;
5160
5161 Node* rawmem = memory(Compile::AliasIdxRaw);
5162 // Is the allocation's memory state untouched?
5163 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5164 // Bail out if there have been raw-memory effects since the allocation.
5165 // (Example: There might have been a call or safepoint.)
5166 return NULL;
5167 }
5168 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5169 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5170 return NULL;
5171 }
5172
5173 // There must be no unexpected observers of this allocation.
5174 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5175 Node* obs = ptr->fast_out(i);
5176 if (obs != this->map()) {
5177 return NULL;
5178 }
5179 }
5180
5181 // This arraycopy must unconditionally follow the allocation of the ptr.
5182 Node* alloc_ctl = ptr->in(0);
5183 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5184
5185 Node* ctl = control();
5186 while (ctl != alloc_ctl) {
5187 // There may be guards which feed into the slow_region.
5188 // Any other control flow means that we might not get a chance
5189 // to finish initializing the allocated object.
5190 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5191 IfNode* iff = ctl->in(0)->as_If();
5192 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5193 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5194 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5195 ctl = iff->in(0); // This test feeds the known slow_region.
5196 continue;
5197 }
5198 // One more try: Various low-level checks bottom out in
5199 // uncommon traps. If the debug-info of the trap omits
5200 // any reference to the allocation, as we've already
5201 // observed, then there can be no objection to the trap.
5202 bool found_trap = false;
5203 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5204 Node* obs = not_ctl->fast_out(j);
5205 if (obs->in(0) == not_ctl && obs->is_Call() &&
5206 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5207 found_trap = true; break;
5208 }
5209 }
5210 if (found_trap) {
5211 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5212 continue;
5213 }
5214 }
5215 return NULL;
5216 }
5217
5218 // If we get this far, we have an allocation which immediately
5219 // precedes the arraycopy, and we can take over zeroing the new object.
5220 // The arraycopy will finish the initialization, and provide
5221 // a new control state to which we will anchor the destination pointer.
5222
5223 return alloc;
5224 }
5225
5226 //-------------inline_encodeISOArray-----------------------------------
5227 // encode char[] to byte[] in ISO_8859_1
5228 bool LibraryCallKit::inline_encodeISOArray() {
5229 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5230 // no receiver since it is static method
5231 Node *src = argument(0);
5232 Node *src_offset = argument(1);
5233 Node *dst = argument(2);
5234 Node *dst_offset = argument(3);
5235 Node *length = argument(4);
5236
5237 const Type* src_type = src->Value(&_gvn);
5238 const Type* dst_type = dst->Value(&_gvn);
5239 const TypeAryPtr* top_src = src_type->isa_aryptr();
5240 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5241 if (top_src == NULL || top_src->klass() == NULL ||
5242 top_dest == NULL || top_dest->klass() == NULL) {
5243 // failed array check
5244 return false;
5245 }
5246
5247 // Figure out the size and type of the elements we will be copying.
5248 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5249 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5250 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5251 return false;
5252 }
5253
5254 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5255 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5256 // 'src_start' points to src array + scaled offset
5257 // 'dst_start' points to dst array + scaled offset
5258
5259 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5260 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5261 enc = _gvn.transform(enc);
5262 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5263 set_memory(res_mem, mtype);
5264 set_result(enc);
5265 return true;
5266 }
5267
5268 //-------------inline_multiplyToLen-----------------------------------
5269 bool LibraryCallKit::inline_multiplyToLen() {
5270 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5271
5272 address stubAddr = StubRoutines::multiplyToLen();
5273 if (stubAddr == NULL) {
5274 return false; // Intrinsic's stub is not implemented on this platform
5275 }
5276 const char* stubName = "multiplyToLen";
5277
5278 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5279
5280 // no receiver because it is a static method
5281 Node* x = argument(0);
5282 Node* xlen = argument(1);
5283 Node* y = argument(2);
5284 Node* ylen = argument(3);
5285 Node* z = argument(4);
5286
5287 const Type* x_type = x->Value(&_gvn);
5288 const Type* y_type = y->Value(&_gvn);
5289 const TypeAryPtr* top_x = x_type->isa_aryptr();
5290 const TypeAryPtr* top_y = y_type->isa_aryptr();
5291 if (top_x == NULL || top_x->klass() == NULL ||
5292 top_y == NULL || top_y->klass() == NULL) {
5293 // failed array check
5294 return false;
5295 }
5296
5297 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5298 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5299 if (x_elem != T_INT || y_elem != T_INT) {
5300 return false;
5301 }
5302
5303 // Set the original stack and the reexecute bit for the interpreter to reexecute
5304 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5305 // on the return from z array allocation in runtime.
5306 { PreserveReexecuteState preexecs(this);
5307 jvms()->set_should_reexecute(true);
5308
5309 Node* x_start = array_element_address(x, intcon(0), x_elem);
5310 Node* y_start = array_element_address(y, intcon(0), y_elem);
5311 // 'x_start' points to x array + scaled xlen
5312 // 'y_start' points to y array + scaled ylen
5313
5314 // Allocate the result array
5315 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5316 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5317 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5318
5319 IdealKit ideal(this);
5320
5321 #define __ ideal.
5322 Node* one = __ ConI(1);
5323 Node* zero = __ ConI(0);
5324 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5325 __ set(need_alloc, zero);
5326 __ set(z_alloc, z);
5327 __ if_then(z, BoolTest::eq, null()); {
5328 __ increment (need_alloc, one);
5329 } __ else_(); {
5330 // Update graphKit memory and control from IdealKit.
5331 sync_kit(ideal);
5332 Node* zlen_arg = load_array_length(z);
5333 // Update IdealKit memory and control from graphKit.
5334 __ sync_kit(this);
5335 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5336 __ increment (need_alloc, one);
5337 } __ end_if();
5338 } __ end_if();
5339
5340 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5341 // Update graphKit memory and control from IdealKit.
5342 sync_kit(ideal);
5343 Node * narr = new_array(klass_node, zlen, 1);
5344 // Update IdealKit memory and control from graphKit.
5345 __ sync_kit(this);
5346 __ set(z_alloc, narr);
5347 } __ end_if();
5348
5349 sync_kit(ideal);
5350 z = __ value(z_alloc);
5351 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5352 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5353 // Final sync IdealKit and GraphKit.
5354 final_sync(ideal);
5355 #undef __
5356
5357 Node* z_start = array_element_address(z, intcon(0), T_INT);
5358
5359 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5360 OptoRuntime::multiplyToLen_Type(),
5361 stubAddr, stubName, TypePtr::BOTTOM,
5362 x_start, xlen, y_start, ylen, z_start, zlen);
5363 } // original reexecute is set back here
5364
5365 C->set_has_split_ifs(true); // Has chance for split-if optimization
5366 set_result(z);
5367 return true;
5368 }
5369
5370 //-------------inline_squareToLen------------------------------------
5371 bool LibraryCallKit::inline_squareToLen() {
5372 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5373
5374 address stubAddr = StubRoutines::squareToLen();
5375 if (stubAddr == NULL) {
5376 return false; // Intrinsic's stub is not implemented on this platform
5377 }
5378 const char* stubName = "squareToLen";
5379
5380 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5381
5382 Node* x = argument(0);
5383 Node* len = argument(1);
5384 Node* z = argument(2);
5385 Node* zlen = argument(3);
5386
5387 const Type* x_type = x->Value(&_gvn);
5388 const Type* z_type = z->Value(&_gvn);
5389 const TypeAryPtr* top_x = x_type->isa_aryptr();
5390 const TypeAryPtr* top_z = z_type->isa_aryptr();
5391 if (top_x == NULL || top_x->klass() == NULL ||
5392 top_z == NULL || top_z->klass() == NULL) {
5393 // failed array check
5394 return false;
5395 }
5396
5397 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5398 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5399 if (x_elem != T_INT || z_elem != T_INT) {
5400 return false;
5401 }
5402
5403
5404 Node* x_start = array_element_address(x, intcon(0), x_elem);
5405 Node* z_start = array_element_address(z, intcon(0), z_elem);
5406
5407 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5408 OptoRuntime::squareToLen_Type(),
5409 stubAddr, stubName, TypePtr::BOTTOM,
5410 x_start, len, z_start, zlen);
5411
5412 set_result(z);
5413 return true;
5414 }
5415
5416 //-------------inline_mulAdd------------------------------------------
5417 bool LibraryCallKit::inline_mulAdd() {
5418 assert(UseMulAddIntrinsic, "not implemented on this platform");
5419
5420 address stubAddr = StubRoutines::mulAdd();
5421 if (stubAddr == NULL) {
5422 return false; // Intrinsic's stub is not implemented on this platform
5423 }
5424 const char* stubName = "mulAdd";
5425
5426 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5427
5428 Node* out = argument(0);
5429 Node* in = argument(1);
5430 Node* offset = argument(2);
5431 Node* len = argument(3);
5432 Node* k = argument(4);
5433
5434 const Type* out_type = out->Value(&_gvn);
5435 const Type* in_type = in->Value(&_gvn);
5436 const TypeAryPtr* top_out = out_type->isa_aryptr();
5437 const TypeAryPtr* top_in = in_type->isa_aryptr();
5438 if (top_out == NULL || top_out->klass() == NULL ||
5439 top_in == NULL || top_in->klass() == NULL) {
5440 // failed array check
5441 return false;
5442 }
5443
5444 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5445 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5446 if (out_elem != T_INT || in_elem != T_INT) {
5447 return false;
5448 }
5449
5450 Node* outlen = load_array_length(out);
5451 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5452 Node* out_start = array_element_address(out, intcon(0), out_elem);
5453 Node* in_start = array_element_address(in, intcon(0), in_elem);
5454
5455 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5456 OptoRuntime::mulAdd_Type(),
5457 stubAddr, stubName, TypePtr::BOTTOM,
5458 out_start,in_start, new_offset, len, k);
5459 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5460 set_result(result);
5461 return true;
5462 }
5463
5464 //-------------inline_montgomeryMultiply-----------------------------------
5465 bool LibraryCallKit::inline_montgomeryMultiply() {
5466 address stubAddr = StubRoutines::montgomeryMultiply();
5467 if (stubAddr == NULL) {
5468 return false; // Intrinsic's stub is not implemented on this platform
5469 }
5470
5471 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5472 const char* stubName = "montgomery_square";
5473
5474 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5475
5476 Node* a = argument(0);
5477 Node* b = argument(1);
5478 Node* n = argument(2);
5479 Node* len = argument(3);
5480 Node* inv = argument(4);
5481 Node* m = argument(6);
5482
5483 const Type* a_type = a->Value(&_gvn);
5484 const TypeAryPtr* top_a = a_type->isa_aryptr();
5485 const Type* b_type = b->Value(&_gvn);
5486 const TypeAryPtr* top_b = b_type->isa_aryptr();
5487 const Type* n_type = a->Value(&_gvn);
5488 const TypeAryPtr* top_n = n_type->isa_aryptr();
5489 const Type* m_type = a->Value(&_gvn);
5490 const TypeAryPtr* top_m = m_type->isa_aryptr();
5491 if (top_a == NULL || top_a->klass() == NULL ||
5492 top_b == NULL || top_b->klass() == NULL ||
5493 top_n == NULL || top_n->klass() == NULL ||
5494 top_m == NULL || top_m->klass() == NULL) {
5495 // failed array check
5496 return false;
5497 }
5498
5499 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5500 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5501 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5502 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5503 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5504 return false;
5505 }
5506
5507 // Make the call
5508 {
5509 Node* a_start = array_element_address(a, intcon(0), a_elem);
5510 Node* b_start = array_element_address(b, intcon(0), b_elem);
5511 Node* n_start = array_element_address(n, intcon(0), n_elem);
5512 Node* m_start = array_element_address(m, intcon(0), m_elem);
5513
5514 Node* call = make_runtime_call(RC_LEAF,
5515 OptoRuntime::montgomeryMultiply_Type(),
5516 stubAddr, stubName, TypePtr::BOTTOM,
5517 a_start, b_start, n_start, len, inv, top(),
5518 m_start);
5519 set_result(m);
5520 }
5521
5522 return true;
5523 }
5524
5525 bool LibraryCallKit::inline_montgomerySquare() {
5526 address stubAddr = StubRoutines::montgomerySquare();
5527 if (stubAddr == NULL) {
5528 return false; // Intrinsic's stub is not implemented on this platform
5529 }
5530
5531 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5532 const char* stubName = "montgomery_square";
5533
5534 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5535
5536 Node* a = argument(0);
5537 Node* n = argument(1);
5538 Node* len = argument(2);
5539 Node* inv = argument(3);
5540 Node* m = argument(5);
5541
5542 const Type* a_type = a->Value(&_gvn);
5543 const TypeAryPtr* top_a = a_type->isa_aryptr();
5544 const Type* n_type = a->Value(&_gvn);
5545 const TypeAryPtr* top_n = n_type->isa_aryptr();
5546 const Type* m_type = a->Value(&_gvn);
5547 const TypeAryPtr* top_m = m_type->isa_aryptr();
5548 if (top_a == NULL || top_a->klass() == NULL ||
5549 top_n == NULL || top_n->klass() == NULL ||
5550 top_m == NULL || top_m->klass() == NULL) {
5551 // failed array check
5552 return false;
5553 }
5554
5555 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5556 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5557 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5558 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5559 return false;
5560 }
5561
5562 // Make the call
5563 {
5564 Node* a_start = array_element_address(a, intcon(0), a_elem);
5565 Node* n_start = array_element_address(n, intcon(0), n_elem);
5566 Node* m_start = array_element_address(m, intcon(0), m_elem);
5567
5568 Node* call = make_runtime_call(RC_LEAF,
5569 OptoRuntime::montgomerySquare_Type(),
5570 stubAddr, stubName, TypePtr::BOTTOM,
5571 a_start, n_start, len, inv, top(),
5572 m_start);
5573 set_result(m);
5574 }
5575
5576 return true;
5577 }
5578
5579
5580 /**
5581 * Calculate CRC32 for byte.
5582 * int java.util.zip.CRC32.update(int crc, int b)
5583 */
5584 bool LibraryCallKit::inline_updateCRC32() {
5585 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5586 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5587 // no receiver since it is static method
5588 Node* crc = argument(0); // type: int
5589 Node* b = argument(1); // type: int
5590
5591 /*
5592 * int c = ~ crc;
5593 * b = timesXtoThe32[(b ^ c) & 0xFF];
5594 * b = b ^ (c >>> 8);
5595 * crc = ~b;
5596 */
5597
5598 Node* M1 = intcon(-1);
5599 crc = _gvn.transform(new XorINode(crc, M1));
5600 Node* result = _gvn.transform(new XorINode(crc, b));
5601 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5602
5603 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5604 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5605 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5606 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5607
5608 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5609 result = _gvn.transform(new XorINode(crc, result));
5610 result = _gvn.transform(new XorINode(result, M1));
5611 set_result(result);
5612 return true;
5613 }
5614
5615 /**
5616 * Calculate CRC32 for byte[] array.
5617 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5618 */
5619 bool LibraryCallKit::inline_updateBytesCRC32() {
5620 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5621 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5622 // no receiver since it is static method
5623 Node* crc = argument(0); // type: int
5624 Node* src = argument(1); // type: oop
5625 Node* offset = argument(2); // type: int
5626 Node* length = argument(3); // type: int
5627
5628 const Type* src_type = src->Value(&_gvn);
5629 const TypeAryPtr* top_src = src_type->isa_aryptr();
5630 if (top_src == NULL || top_src->klass() == NULL) {
5631 // failed array check
5632 return false;
5633 }
5634
5635 // Figure out the size and type of the elements we will be copying.
5636 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5637 if (src_elem != T_BYTE) {
5638 return false;
5639 }
5640
5641 // 'src_start' points to src array + scaled offset
5642 Node* src_start = array_element_address(src, offset, src_elem);
5643
5644 // We assume that range check is done by caller.
5645 // TODO: generate range check (offset+length < src.length) in debug VM.
5646
5647 // Call the stub.
5648 address stubAddr = StubRoutines::updateBytesCRC32();
5649 const char *stubName = "updateBytesCRC32";
5650
5651 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5652 stubAddr, stubName, TypePtr::BOTTOM,
5653 crc, src_start, length);
5654 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5655 set_result(result);
5656 return true;
5657 }
5658
5659 /**
5660 * Calculate CRC32 for ByteBuffer.
5661 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5662 */
5663 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5664 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5665 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5666 // no receiver since it is static method
5667 Node* crc = argument(0); // type: int
5668 Node* src = argument(1); // type: long
5669 Node* offset = argument(3); // type: int
5670 Node* length = argument(4); // type: int
5671
5672 src = ConvL2X(src); // adjust Java long to machine word
5673 Node* base = _gvn.transform(new CastX2PNode(src));
5674 offset = ConvI2X(offset);
5675
5676 // 'src_start' points to src array + scaled offset
5677 Node* src_start = basic_plus_adr(top(), base, offset);
5678
5679 // Call the stub.
5680 address stubAddr = StubRoutines::updateBytesCRC32();
5681 const char *stubName = "updateBytesCRC32";
5682
5683 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5684 stubAddr, stubName, TypePtr::BOTTOM,
5685 crc, src_start, length);
5686 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5687 set_result(result);
5688 return true;
5689 }
5690
5691 //------------------------------get_table_from_crc32c_class-----------------------
5692 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5693 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5694 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5695
5696 return table;
5697 }
5698
5699 //------------------------------inline_updateBytesCRC32C-----------------------
5700 //
5701 // Calculate CRC32C for byte[] array.
5702 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5703 //
5704 bool LibraryCallKit::inline_updateBytesCRC32C() {
5705 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5706 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5707 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5708 // no receiver since it is a static method
5709 Node* crc = argument(0); // type: int
5710 Node* src = argument(1); // type: oop
5711 Node* offset = argument(2); // type: int
5712 Node* end = argument(3); // type: int
5713
5714 Node* length = _gvn.transform(new SubINode(end, offset));
5715
5716 const Type* src_type = src->Value(&_gvn);
5717 const TypeAryPtr* top_src = src_type->isa_aryptr();
5718 if (top_src == NULL || top_src->klass() == NULL) {
5719 // failed array check
5720 return false;
5721 }
5722
5723 // Figure out the size and type of the elements we will be copying.
5724 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5725 if (src_elem != T_BYTE) {
5726 return false;
5727 }
5728
5729 // 'src_start' points to src array + scaled offset
5730 Node* src_start = array_element_address(src, offset, src_elem);
5731
5732 // static final int[] byteTable in class CRC32C
5733 Node* table = get_table_from_crc32c_class(callee()->holder());
5734 Node* table_start = array_element_address(table, intcon(0), T_INT);
5735
5736 // We assume that range check is done by caller.
5737 // TODO: generate range check (offset+length < src.length) in debug VM.
5738
5739 // Call the stub.
5740 address stubAddr = StubRoutines::updateBytesCRC32C();
5741 const char *stubName = "updateBytesCRC32C";
5742
5743 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5744 stubAddr, stubName, TypePtr::BOTTOM,
5745 crc, src_start, length, table_start);
5746 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5747 set_result(result);
5748 return true;
5749 }
5750
5751 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5752 //
5753 // Calculate CRC32C for DirectByteBuffer.
5754 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5755 //
5756 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5757 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5758 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5759 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5760 // no receiver since it is a static method
5761 Node* crc = argument(0); // type: int
5762 Node* src = argument(1); // type: long
5763 Node* offset = argument(3); // type: int
5764 Node* end = argument(4); // type: int
5765
5766 Node* length = _gvn.transform(new SubINode(end, offset));
5767
5768 src = ConvL2X(src); // adjust Java long to machine word
5769 Node* base = _gvn.transform(new CastX2PNode(src));
5770 offset = ConvI2X(offset);
5771
5772 // 'src_start' points to src array + scaled offset
5773 Node* src_start = basic_plus_adr(top(), base, offset);
5774
5775 // static final int[] byteTable in class CRC32C
5776 Node* table = get_table_from_crc32c_class(callee()->holder());
5777 Node* table_start = array_element_address(table, intcon(0), T_INT);
5778
5779 // Call the stub.
5780 address stubAddr = StubRoutines::updateBytesCRC32C();
5781 const char *stubName = "updateBytesCRC32C";
5782
5783 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5784 stubAddr, stubName, TypePtr::BOTTOM,
5785 crc, src_start, length, table_start);
5786 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5787 set_result(result);
5788 return true;
5789 }
5790
5791 //------------------------------inline_updateBytesAdler32----------------------
5792 //
5793 // Calculate Adler32 checksum for byte[] array.
5794 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5795 //
5796 bool LibraryCallKit::inline_updateBytesAdler32() {
5797 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5798 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5799 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5800 // no receiver since it is static method
5801 Node* crc = argument(0); // type: int
5802 Node* src = argument(1); // type: oop
5803 Node* offset = argument(2); // type: int
5804 Node* length = argument(3); // type: int
5805
5806 const Type* src_type = src->Value(&_gvn);
5807 const TypeAryPtr* top_src = src_type->isa_aryptr();
5808 if (top_src == NULL || top_src->klass() == NULL) {
5809 // failed array check
5810 return false;
5811 }
5812
5813 // Figure out the size and type of the elements we will be copying.
5814 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5815 if (src_elem != T_BYTE) {
5816 return false;
5817 }
5818
5819 // 'src_start' points to src array + scaled offset
5820 Node* src_start = array_element_address(src, offset, src_elem);
5821
5822 // We assume that range check is done by caller.
5823 // TODO: generate range check (offset+length < src.length) in debug VM.
5824
5825 // Call the stub.
5826 address stubAddr = StubRoutines::updateBytesAdler32();
5827 const char *stubName = "updateBytesAdler32";
5828
5829 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5830 stubAddr, stubName, TypePtr::BOTTOM,
5831 crc, src_start, length);
5832 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5833 set_result(result);
5834 return true;
5835 }
5836
5837 //------------------------------inline_updateByteBufferAdler32---------------
5838 //
5839 // Calculate Adler32 checksum for DirectByteBuffer.
5840 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5841 //
5842 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5843 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5844 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5845 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5846 // no receiver since it is static method
5847 Node* crc = argument(0); // type: int
5848 Node* src = argument(1); // type: long
5849 Node* offset = argument(3); // type: int
5850 Node* length = argument(4); // type: int
5851
5852 src = ConvL2X(src); // adjust Java long to machine word
5853 Node* base = _gvn.transform(new CastX2PNode(src));
5854 offset = ConvI2X(offset);
5855
5856 // 'src_start' points to src array + scaled offset
5857 Node* src_start = basic_plus_adr(top(), base, offset);
5858
5859 // Call the stub.
5860 address stubAddr = StubRoutines::updateBytesAdler32();
5861 const char *stubName = "updateBytesAdler32";
5862
5863 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5864 stubAddr, stubName, TypePtr::BOTTOM,
5865 crc, src_start, length);
5866
5867 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5868 set_result(result);
5869 return true;
5870 }
5871
5872 //----------------------------inline_reference_get----------------------------
5873 // public T java.lang.ref.Reference.get();
5874 bool LibraryCallKit::inline_reference_get() {
5875 const int referent_offset = java_lang_ref_Reference::referent_offset;
5876 guarantee(referent_offset > 0, "should have already been set");
5877
5878 // Get the argument:
5879 Node* reference_obj = null_check_receiver();
5880 if (stopped()) return true;
5881
5882 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5883
5884 ciInstanceKlass* klass = env()->Object_klass();
5885 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5886
5887 Node* no_ctrl = NULL;
5888 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5889
5890 // Use the pre-barrier to record the value in the referent field
5891 pre_barrier(false /* do_load */,
5892 control(),
5893 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5894 result /* pre_val */,
5895 T_OBJECT);
5896
5897 // Add memory barrier to prevent commoning reads from this field
5898 // across safepoint since GC can change its value.
5899 insert_mem_bar(Op_MemBarCPUOrder);
5900
5901 set_result(result);
5902 return true;
5903 }
5904
5905
5906 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5907 bool is_exact=true, bool is_static=false,
5908 ciInstanceKlass * fromKls=NULL) {
5909 if (fromKls == NULL) {
5910 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5911 assert(tinst != NULL, "obj is null");
5912 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5913 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5914 fromKls = tinst->klass()->as_instance_klass();
5915 } else {
5916 assert(is_static, "only for static field access");
5917 }
5918 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5919 ciSymbol::make(fieldTypeString),
5920 is_static);
5921
5922 assert (field != NULL, "undefined field");
5923 if (field == NULL) return (Node *) NULL;
5924
5925 if (is_static) {
5926 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5927 fromObj = makecon(tip);
5928 }
5929
5930 // Next code copied from Parse::do_get_xxx():
5931
5932 // Compute address and memory type.
5933 int offset = field->offset_in_bytes();
5934 bool is_vol = field->is_volatile();
5935 ciType* field_klass = field->type();
5936 assert(field_klass->is_loaded(), "should be loaded");
5937 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5938 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5939 BasicType bt = field->layout_type();
5940
5941 // Build the resultant type of the load
5942 const Type *type;
5943 if (bt == T_OBJECT) {
5944 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5945 } else {
5946 type = Type::get_const_basic_type(bt);
5947 }
5948
5949 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5950 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5951 }
5952 // Build the load.
5953 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5954 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5955 // If reference is volatile, prevent following memory ops from
5956 // floating up past the volatile read. Also prevents commoning
5957 // another volatile read.
5958 if (is_vol) {
5959 // Memory barrier includes bogus read of value to force load BEFORE membar
5960 insert_mem_bar(Op_MemBarAcquire, loadedField);
5961 }
5962 return loadedField;
5963 }
5964
5965
5966 //------------------------------inline_aescrypt_Block-----------------------
5967 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5968 address stubAddr = NULL;
5969 const char *stubName;
5970 assert(UseAES, "need AES instruction support");
5971
5972 switch(id) {
5973 case vmIntrinsics::_aescrypt_encryptBlock:
5974 stubAddr = StubRoutines::aescrypt_encryptBlock();
5975 stubName = "aescrypt_encryptBlock";
5976 break;
5977 case vmIntrinsics::_aescrypt_decryptBlock:
5978 stubAddr = StubRoutines::aescrypt_decryptBlock();
5979 stubName = "aescrypt_decryptBlock";
5980 break;
5981 }
5982 if (stubAddr == NULL) return false;
5983
5984 Node* aescrypt_object = argument(0);
5985 Node* src = argument(1);
5986 Node* src_offset = argument(2);
5987 Node* dest = argument(3);
5988 Node* dest_offset = argument(4);
5989
5990 // (1) src and dest are arrays.
5991 const Type* src_type = src->Value(&_gvn);
5992 const Type* dest_type = dest->Value(&_gvn);
5993 const TypeAryPtr* top_src = src_type->isa_aryptr();
5994 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5995 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5996
5997 // for the quick and dirty code we will skip all the checks.
5998 // we are just trying to get the call to be generated.
5999 Node* src_start = src;
6000 Node* dest_start = dest;
6001 if (src_offset != NULL || dest_offset != NULL) {
6002 assert(src_offset != NULL && dest_offset != NULL, "");
6003 src_start = array_element_address(src, src_offset, T_BYTE);
6004 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6005 }
6006
6007 // now need to get the start of its expanded key array
6008 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6009 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6010 if (k_start == NULL) return false;
6011
6012 if (Matcher::pass_original_key_for_aes()) {
6013 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6014 // compatibility issues between Java key expansion and SPARC crypto instructions
6015 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6016 if (original_k_start == NULL) return false;
6017
6018 // Call the stub.
6019 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6020 stubAddr, stubName, TypePtr::BOTTOM,
6021 src_start, dest_start, k_start, original_k_start);
6022 } else {
6023 // Call the stub.
6024 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6025 stubAddr, stubName, TypePtr::BOTTOM,
6026 src_start, dest_start, k_start);
6027 }
6028
6029 return true;
6030 }
6031
6032 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6033 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6034 address stubAddr = NULL;
6035 const char *stubName = NULL;
6036
6037 assert(UseAES, "need AES instruction support");
6038
6039 switch(id) {
6040 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6041 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6042 stubName = "cipherBlockChaining_encryptAESCrypt";
6043 break;
6044 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6045 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6046 stubName = "cipherBlockChaining_decryptAESCrypt";
6047 break;
6048 }
6049 if (stubAddr == NULL) return false;
6050
6051 Node* cipherBlockChaining_object = argument(0);
6052 Node* src = argument(1);
6053 Node* src_offset = argument(2);
6054 Node* len = argument(3);
6055 Node* dest = argument(4);
6056 Node* dest_offset = argument(5);
6057
6058 // (1) src and dest are arrays.
6059 const Type* src_type = src->Value(&_gvn);
6060 const Type* dest_type = dest->Value(&_gvn);
6061 const TypeAryPtr* top_src = src_type->isa_aryptr();
6062 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6063 assert (top_src != NULL && top_src->klass() != NULL
6064 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6065
6066 // checks are the responsibility of the caller
6067 Node* src_start = src;
6068 Node* dest_start = dest;
6069 if (src_offset != NULL || dest_offset != NULL) {
6070 assert(src_offset != NULL && dest_offset != NULL, "");
6071 src_start = array_element_address(src, src_offset, T_BYTE);
6072 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6073 }
6074
6075 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6076 // (because of the predicated logic executed earlier).
6077 // so we cast it here safely.
6078 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6079
6080 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6081 if (embeddedCipherObj == NULL) return false;
6082
6083 // cast it to what we know it will be at runtime
6084 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6085 assert(tinst != NULL, "CBC obj is null");
6086 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6087 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6088 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6089
6090 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6091 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6092 const TypeOopPtr* xtype = aklass->as_instance_type();
6093 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6094 aescrypt_object = _gvn.transform(aescrypt_object);
6095
6096 // we need to get the start of the aescrypt_object's expanded key array
6097 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6098 if (k_start == NULL) return false;
6099
6100 // similarly, get the start address of the r vector
6101 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6102 if (objRvec == NULL) return false;
6103 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6104
6105 Node* cbcCrypt;
6106 if (Matcher::pass_original_key_for_aes()) {
6107 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6108 // compatibility issues between Java key expansion and SPARC crypto instructions
6109 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6110 if (original_k_start == NULL) return false;
6111
6112 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6113 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6114 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6115 stubAddr, stubName, TypePtr::BOTTOM,
6116 src_start, dest_start, k_start, r_start, len, original_k_start);
6117 } else {
6118 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6119 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6120 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6121 stubAddr, stubName, TypePtr::BOTTOM,
6122 src_start, dest_start, k_start, r_start, len);
6123 }
6124
6125 // return cipher length (int)
6126 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6127 set_result(retvalue);
6128 return true;
6129 }
6130
6131 //------------------------------get_key_start_from_aescrypt_object-----------------------
6132 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6133 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6134 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6135 if (objAESCryptKey == NULL) return (Node *) NULL;
6136
6137 // now have the array, need to get the start address of the K array
6138 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6139 return k_start;
6140 }
6141
6142 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6143 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6144 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6145 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6146 if (objAESCryptKey == NULL) return (Node *) NULL;
6147
6148 // now have the array, need to get the start address of the lastKey array
6149 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6150 return original_k_start;
6151 }
6152
6153 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6154 // Return node representing slow path of predicate check.
6155 // the pseudo code we want to emulate with this predicate is:
6156 // for encryption:
6157 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6158 // for decryption:
6159 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6160 // note cipher==plain is more conservative than the original java code but that's OK
6161 //
6162 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6163 // The receiver was checked for NULL already.
6164 Node* objCBC = argument(0);
6165
6166 // Load embeddedCipher field of CipherBlockChaining object.
6167 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6168
6169 // get AESCrypt klass for instanceOf check
6170 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6171 // will have same classloader as CipherBlockChaining object
6172 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6173 assert(tinst != NULL, "CBCobj is null");
6174 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6175
6176 // we want to do an instanceof comparison against the AESCrypt class
6177 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6178 if (!klass_AESCrypt->is_loaded()) {
6179 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6180 Node* ctrl = control();
6181 set_control(top()); // no regular fast path
6182 return ctrl;
6183 }
6184 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6185
6186 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6187 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6188 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6189
6190 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6191
6192 // for encryption, we are done
6193 if (!decrypting)
6194 return instof_false; // even if it is NULL
6195
6196 // for decryption, we need to add a further check to avoid
6197 // taking the intrinsic path when cipher and plain are the same
6198 // see the original java code for why.
6199 RegionNode* region = new RegionNode(3);
6200 region->init_req(1, instof_false);
6201 Node* src = argument(1);
6202 Node* dest = argument(4);
6203 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6204 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6205 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6206 region->init_req(2, src_dest_conjoint);
6207
6208 record_for_igvn(region);
6209 return _gvn.transform(region);
6210 }
6211
6212 //------------------------------inline_ghash_processBlocks
6213 bool LibraryCallKit::inline_ghash_processBlocks() {
6214 address stubAddr;
6215 const char *stubName;
6216 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6217
6218 stubAddr = StubRoutines::ghash_processBlocks();
6219 stubName = "ghash_processBlocks";
6220
6221 Node* data = argument(0);
6222 Node* offset = argument(1);
6223 Node* len = argument(2);
6224 Node* state = argument(3);
6225 Node* subkeyH = argument(4);
6226
6227 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6228 assert(state_start, "state is NULL");
6229 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6230 assert(subkeyH_start, "subkeyH is NULL");
6231 Node* data_start = array_element_address(data, offset, T_BYTE);
6232 assert(data_start, "data is NULL");
6233
6234 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6235 OptoRuntime::ghash_processBlocks_Type(),
6236 stubAddr, stubName, TypePtr::BOTTOM,
6237 state_start, subkeyH_start, data_start, len);
6238 return true;
6239 }
6240
6241 //------------------------------inline_sha_implCompress-----------------------
6242 //
6243 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6244 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6245 //
6246 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6247 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6248 //
6249 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6250 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6251 //
6252 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6253 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6254
6255 Node* sha_obj = argument(0);
6256 Node* src = argument(1); // type oop
6257 Node* ofs = argument(2); // type int
6258
6259 const Type* src_type = src->Value(&_gvn);
6260 const TypeAryPtr* top_src = src_type->isa_aryptr();
6261 if (top_src == NULL || top_src->klass() == NULL) {
6262 // failed array check
6263 return false;
6264 }
6265 // Figure out the size and type of the elements we will be copying.
6266 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6267 if (src_elem != T_BYTE) {
6268 return false;
6269 }
6270 // 'src_start' points to src array + offset
6271 Node* src_start = array_element_address(src, ofs, src_elem);
6272 Node* state = NULL;
6273 address stubAddr;
6274 const char *stubName;
6275
6276 switch(id) {
6277 case vmIntrinsics::_sha_implCompress:
6278 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6279 state = get_state_from_sha_object(sha_obj);
6280 stubAddr = StubRoutines::sha1_implCompress();
6281 stubName = "sha1_implCompress";
6282 break;
6283 case vmIntrinsics::_sha2_implCompress:
6284 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6285 state = get_state_from_sha_object(sha_obj);
6286 stubAddr = StubRoutines::sha256_implCompress();
6287 stubName = "sha256_implCompress";
6288 break;
6289 case vmIntrinsics::_sha5_implCompress:
6290 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6291 state = get_state_from_sha5_object(sha_obj);
6292 stubAddr = StubRoutines::sha512_implCompress();
6293 stubName = "sha512_implCompress";
6294 break;
6295 default:
6296 fatal_unexpected_iid(id);
6297 return false;
6298 }
6299 if (state == NULL) return false;
6300
6301 // Call the stub.
6302 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6303 stubAddr, stubName, TypePtr::BOTTOM,
6304 src_start, state);
6305
6306 return true;
6307 }
6308
6309 //------------------------------inline_digestBase_implCompressMB-----------------------
6310 //
6311 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6312 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6313 //
6314 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6315 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6316 "need SHA1/SHA256/SHA512 instruction support");
6317 assert((uint)predicate < 3, "sanity");
6318 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6319
6320 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6321 Node* src = argument(1); // byte[] array
6322 Node* ofs = argument(2); // type int
6323 Node* limit = argument(3); // type int
6324
6325 const Type* src_type = src->Value(&_gvn);
6326 const TypeAryPtr* top_src = src_type->isa_aryptr();
6327 if (top_src == NULL || top_src->klass() == NULL) {
6328 // failed array check
6329 return false;
6330 }
6331 // Figure out the size and type of the elements we will be copying.
6332 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6333 if (src_elem != T_BYTE) {
6334 return false;
6335 }
6336 // 'src_start' points to src array + offset
6337 Node* src_start = array_element_address(src, ofs, src_elem);
6338
6339 const char* klass_SHA_name = NULL;
6340 const char* stub_name = NULL;
6341 address stub_addr = NULL;
6342 bool long_state = false;
6343
6344 switch (predicate) {
6345 case 0:
6346 if (UseSHA1Intrinsics) {
6347 klass_SHA_name = "sun/security/provider/SHA";
6348 stub_name = "sha1_implCompressMB";
6349 stub_addr = StubRoutines::sha1_implCompressMB();
6350 }
6351 break;
6352 case 1:
6353 if (UseSHA256Intrinsics) {
6354 klass_SHA_name = "sun/security/provider/SHA2";
6355 stub_name = "sha256_implCompressMB";
6356 stub_addr = StubRoutines::sha256_implCompressMB();
6357 }
6358 break;
6359 case 2:
6360 if (UseSHA512Intrinsics) {
6361 klass_SHA_name = "sun/security/provider/SHA5";
6362 stub_name = "sha512_implCompressMB";
6363 stub_addr = StubRoutines::sha512_implCompressMB();
6364 long_state = true;
6365 }
6366 break;
6367 default:
6368 fatal("unknown SHA intrinsic predicate: %d", predicate);
6369 }
6370 if (klass_SHA_name != NULL) {
6371 // get DigestBase klass to lookup for SHA klass
6372 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6373 assert(tinst != NULL, "digestBase_obj is not instance???");
6374 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6375
6376 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6377 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6378 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6379 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6380 }
6381 return false;
6382 }
6383 //------------------------------inline_sha_implCompressMB-----------------------
6384 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6385 bool long_state, address stubAddr, const char *stubName,
6386 Node* src_start, Node* ofs, Node* limit) {
6387 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6388 const TypeOopPtr* xtype = aklass->as_instance_type();
6389 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6390 sha_obj = _gvn.transform(sha_obj);
6391
6392 Node* state;
6393 if (long_state) {
6394 state = get_state_from_sha5_object(sha_obj);
6395 } else {
6396 state = get_state_from_sha_object(sha_obj);
6397 }
6398 if (state == NULL) return false;
6399
6400 // Call the stub.
6401 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6402 OptoRuntime::digestBase_implCompressMB_Type(),
6403 stubAddr, stubName, TypePtr::BOTTOM,
6404 src_start, state, ofs, limit);
6405 // return ofs (int)
6406 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6407 set_result(result);
6408
6409 return true;
6410 }
6411
6412 //------------------------------get_state_from_sha_object-----------------------
6413 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6414 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6415 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6416 if (sha_state == NULL) return (Node *) NULL;
6417
6418 // now have the array, need to get the start address of the state array
6419 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6420 return state;
6421 }
6422
6423 //------------------------------get_state_from_sha5_object-----------------------
6424 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6425 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6426 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6427 if (sha_state == NULL) return (Node *) NULL;
6428
6429 // now have the array, need to get the start address of the state array
6430 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6431 return state;
6432 }
6433
6434 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6435 // Return node representing slow path of predicate check.
6436 // the pseudo code we want to emulate with this predicate is:
6437 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6438 //
6439 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6440 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6441 "need SHA1/SHA256/SHA512 instruction support");
6442 assert((uint)predicate < 3, "sanity");
6443
6444 // The receiver was checked for NULL already.
6445 Node* digestBaseObj = argument(0);
6446
6447 // get DigestBase klass for instanceOf check
6448 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6449 assert(tinst != NULL, "digestBaseObj is null");
6450 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6451
6452 const char* klass_SHA_name = NULL;
6453 switch (predicate) {
6454 case 0:
6455 if (UseSHA1Intrinsics) {
6456 // we want to do an instanceof comparison against the SHA class
6457 klass_SHA_name = "sun/security/provider/SHA";
6458 }
6459 break;
6460 case 1:
6461 if (UseSHA256Intrinsics) {
6462 // we want to do an instanceof comparison against the SHA2 class
6463 klass_SHA_name = "sun/security/provider/SHA2";
6464 }
6465 break;
6466 case 2:
6467 if (UseSHA512Intrinsics) {
6468 // we want to do an instanceof comparison against the SHA5 class
6469 klass_SHA_name = "sun/security/provider/SHA5";
6470 }
6471 break;
6472 default:
6473 fatal("unknown SHA intrinsic predicate: %d", predicate);
6474 }
6475
6476 ciKlass* klass_SHA = NULL;
6477 if (klass_SHA_name != NULL) {
6478 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6479 }
6480 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6481 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6482 Node* ctrl = control();
6483 set_control(top()); // no intrinsic path
6484 return ctrl;
6485 }
6486 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6487
6488 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6489 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6490 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6491 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6492
6493 return instof_false; // even if it is NULL
6494 }
6495
6496 bool LibraryCallKit::inline_profileBoolean() {
6497 Node* counts = argument(1);
6498 const TypeAryPtr* ary = NULL;
6499 ciArray* aobj = NULL;
6500 if (counts->is_Con()
6501 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6502 && (aobj = ary->const_oop()->as_array()) != NULL
6503 && (aobj->length() == 2)) {
6504 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6505 jint false_cnt = aobj->element_value(0).as_int();
6506 jint true_cnt = aobj->element_value(1).as_int();
6507
6508 if (C->log() != NULL) {
6509 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6510 false_cnt, true_cnt);
6511 }
6512
6513 if (false_cnt + true_cnt == 0) {
6514 // According to profile, never executed.
6515 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6516 Deoptimization::Action_reinterpret);
6517 return true;
6518 }
6519
6520 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6521 // is a number of each value occurrences.
6522 Node* result = argument(0);
6523 if (false_cnt == 0 || true_cnt == 0) {
6524 // According to profile, one value has been never seen.
6525 int expected_val = (false_cnt == 0) ? 1 : 0;
6526
6527 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6528 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6529
6530 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6531 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6532 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6533
6534 { // Slow path: uncommon trap for never seen value and then reexecute
6535 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6536 // the value has been seen at least once.
6537 PreserveJVMState pjvms(this);
6538 PreserveReexecuteState preexecs(this);
6539 jvms()->set_should_reexecute(true);
6540
6541 set_control(slow_path);
6542 set_i_o(i_o());
6543
6544 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6545 Deoptimization::Action_reinterpret);
6546 }
6547 // The guard for never seen value enables sharpening of the result and
6548 // returning a constant. It allows to eliminate branches on the same value
6549 // later on.
6550 set_control(fast_path);
6551 result = intcon(expected_val);
6552 }
6553 // Stop profiling.
6554 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6555 // By replacing method body with profile data (represented as ProfileBooleanNode
6556 // on IR level) we effectively disable profiling.
6557 // It enables full speed execution once optimized code is generated.
6558 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6559 C->record_for_igvn(profile);
6560 set_result(profile);
6561 return true;
6562 } else {
6563 // Continue profiling.
6564 // Profile data isn't available at the moment. So, execute method's bytecode version.
6565 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6566 // is compiled and counters aren't available since corresponding MethodHandle
6567 // isn't a compile-time constant.
6568 return false;
6569 }
6570 }
6571
6572 bool LibraryCallKit::inline_isCompileConstant() {
6573 Node* n = argument(0);
6574 set_result(n->is_Con() ? intcon(1) : intcon(0));
6575 return true;
6576 }