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