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 Node* generate_current_thread(Node* &tls_output);
156 Node* load_mirror_from_klass(Node* klass);
157 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path,
159 int offset);
160 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
161 RegionNode* region, int null_path) {
162 int offset = java_lang_Class::klass_offset_in_bytes();
163 return load_klass_from_mirror_common(mirror, never_see_null,
164 region, null_path,
165 offset);
166 }
167 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::array_klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* generate_access_flags_guard(Node* kls,
175 int modifier_mask, int modifier_bits,
176 RegionNode* region);
177 Node* generate_interface_guard(Node* kls, RegionNode* region);
178 Node* generate_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, false);
180 }
181 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, false, true);
183 }
184 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, false);
186 }
187 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, true, true);
189 }
190 Node* generate_array_guard_common(Node* kls, RegionNode* region,
191 bool obj_array, bool not_array);
192 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
193 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
194 bool is_virtual = false, bool is_static = false);
195 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, false, true);
197 }
198 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
199 return generate_method_call(method_id, true, false);
200 }
201 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
202
203 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
204 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
205 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
206 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
207 bool inline_string_indexOfChar();
208 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
209 bool inline_string_toBytesU();
210 bool inline_string_getCharsU();
211 bool inline_string_copy(bool compress);
212 bool inline_string_char_access(bool is_store);
213 Node* round_double_node(Node* n);
214 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
215 bool inline_math_native(vmIntrinsics::ID id);
216 bool inline_trig(vmIntrinsics::ID id);
217 bool inline_math(vmIntrinsics::ID id);
218 template <typename OverflowOp>
219 bool inline_math_overflow(Node* arg1, Node* arg2);
220 void inline_math_mathExact(Node* math, Node* test);
221 bool inline_math_addExactI(bool is_increment);
222 bool inline_math_addExactL(bool is_increment);
223 bool inline_math_multiplyExactI();
224 bool inline_math_multiplyExactL();
225 bool inline_math_negateExactI();
226 bool inline_math_negateExactL();
227 bool inline_math_subtractExactI(bool is_decrement);
228 bool inline_math_subtractExactL(bool is_decrement);
229 bool inline_pow();
230 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
231 bool inline_min_max(vmIntrinsics::ID id);
232 bool inline_notify(vmIntrinsics::ID id);
233 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
234 // This returns Type::AnyPtr, RawPtr, or OopPtr.
235 int classify_unsafe_addr(Node* &base, Node* &offset);
236 Node* make_unsafe_address(Node* base, Node* offset);
237 // Helper for inline_unsafe_access.
238 // Generates the guards that check whether the result of
239 // Unsafe.getObject should be recorded in an SATB log buffer.
240 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
241 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
242 static bool klass_needs_init_guard(Node* kls);
243 bool inline_unsafe_allocate();
244 bool inline_unsafe_copyMemory();
245 bool inline_native_currentThread();
246 #ifdef TRACE_HAVE_INTRINSICS
247 bool inline_native_classID();
248 bool inline_native_threadID();
249 #endif
250 bool inline_native_time_funcs(address method, const char* funcName);
251 bool inline_native_isInterrupted();
252 bool inline_native_Class_query(vmIntrinsics::ID id);
253 bool inline_native_subtype_check();
254
255 bool inline_native_newArray();
256 bool inline_native_getLength();
257 bool inline_array_copyOf(bool is_copyOfRange);
258 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
259 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
260 bool inline_native_clone(bool is_virtual);
261 bool inline_native_Reflection_getCallerClass();
262 // Helper function for inlining native object hash method
263 bool inline_native_hashcode(bool is_virtual, bool is_static);
264 bool inline_native_getClass();
265
266 // Helper functions for inlining arraycopy
267 bool inline_arraycopy();
268 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
269 RegionNode* slow_region);
270 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
271 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
272
273 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
274 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
275 bool inline_unsafe_ordered_store(BasicType type);
276 bool inline_unsafe_fence(vmIntrinsics::ID id);
277 bool inline_onspinwait();
278 bool inline_fp_conversions(vmIntrinsics::ID id);
279 bool inline_number_methods(vmIntrinsics::ID id);
280 bool inline_reference_get();
281 bool inline_Class_cast();
282 bool inline_aescrypt_Block(vmIntrinsics::ID id);
283 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
284 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
285 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
286 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
287 bool inline_ghash_processBlocks();
288 bool inline_sha_implCompress(vmIntrinsics::ID id);
289 bool inline_digestBase_implCompressMB(int predicate);
290 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
291 bool long_state, address stubAddr, const char *stubName,
292 Node* src_start, Node* ofs, Node* limit);
293 Node* get_state_from_sha_object(Node *sha_object);
294 Node* get_state_from_sha5_object(Node *sha_object);
295 Node* inline_digestBase_implCompressMB_predicate(int predicate);
296 bool inline_encodeISOArray();
297 bool inline_updateCRC32();
298 bool inline_updateBytesCRC32();
299 bool inline_updateByteBufferCRC32();
300 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
301 bool inline_updateBytesCRC32C();
302 bool inline_updateDirectByteBufferCRC32C();
303 bool inline_updateBytesAdler32();
304 bool inline_updateByteBufferAdler32();
305 bool inline_multiplyToLen();
306 bool inline_hasNegatives();
307 bool inline_squareToLen();
308 bool inline_mulAdd();
309 bool inline_montgomeryMultiply();
310 bool inline_montgomerySquare();
311
312 bool inline_profileBoolean();
313 bool inline_isCompileConstant();
314 };
315
316 //---------------------------make_vm_intrinsic----------------------------
317 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
318 vmIntrinsics::ID id = m->intrinsic_id();
319 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
320
321 if (!m->is_loaded()) {
322 // Do not attempt to inline unloaded methods.
323 return NULL;
324 }
325
326 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
327 bool is_available = false;
328
329 {
330 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
331 // the compiler must transition to '_thread_in_vm' state because both
332 // methods access VM-internal data.
333 VM_ENTRY_MARK;
334 methodHandle mh(THREAD, m->get_Method());
335 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
336 !C->directive()->is_intrinsic_disabled(mh) &&
337 !vmIntrinsics::is_disabled_by_flags(mh);
338
339 }
340
341 if (is_available) {
342 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
343 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
344 return new LibraryIntrinsic(m, is_virtual,
345 vmIntrinsics::predicates_needed(id),
346 vmIntrinsics::does_virtual_dispatch(id),
347 (vmIntrinsics::ID) id);
348 } else {
349 return NULL;
350 }
351 }
352
353 //----------------------register_library_intrinsics-----------------------
354 // Initialize this file's data structures, for each Compile instance.
355 void Compile::register_library_intrinsics() {
356 // Nothing to do here.
357 }
358
359 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
360 LibraryCallKit kit(jvms, this);
361 Compile* C = kit.C;
362 int nodes = C->unique();
363 #ifndef PRODUCT
364 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
365 char buf[1000];
366 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
367 tty->print_cr("Intrinsic %s", str);
368 }
369 #endif
370 ciMethod* callee = kit.callee();
371 const int bci = kit.bci();
372
373 // Try to inline the intrinsic.
374 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
375 kit.try_to_inline(_last_predicate)) {
376 if (C->print_intrinsics() || C->print_inlining()) {
377 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
378 }
379 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
380 if (C->log()) {
381 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
382 vmIntrinsics::name_at(intrinsic_id()),
383 (is_virtual() ? " virtual='1'" : ""),
384 C->unique() - nodes);
385 }
386 // Push the result from the inlined method onto the stack.
387 kit.push_result();
388 C->print_inlining_update(this);
389 return kit.transfer_exceptions_into_jvms();
390 }
391
392 // The intrinsic bailed out
393 if (C->print_intrinsics() || C->print_inlining()) {
394 if (jvms->has_method()) {
395 // Not a root compile.
396 const char* msg;
397 if (callee->intrinsic_candidate()) {
398 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
399 } else {
400 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
401 : "failed to inline (intrinsic), method not annotated";
402 }
403 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
404 } else {
405 // Root compile
406 tty->print("Did not generate intrinsic %s%s at bci:%d in",
407 vmIntrinsics::name_at(intrinsic_id()),
408 (is_virtual() ? " (virtual)" : ""), bci);
409 }
410 }
411 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
412 C->print_inlining_update(this);
413 return NULL;
414 }
415
416 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
417 LibraryCallKit kit(jvms, this);
418 Compile* C = kit.C;
419 int nodes = C->unique();
420 _last_predicate = predicate;
421 #ifndef PRODUCT
422 assert(is_predicated() && predicate < predicates_count(), "sanity");
423 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
424 char buf[1000];
425 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
426 tty->print_cr("Predicate for intrinsic %s", str);
427 }
428 #endif
429 ciMethod* callee = kit.callee();
430 const int bci = kit.bci();
431
432 Node* slow_ctl = kit.try_to_predicate(predicate);
433 if (!kit.failing()) {
434 if (C->print_intrinsics() || C->print_inlining()) {
435 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
436 }
437 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
438 if (C->log()) {
439 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
440 vmIntrinsics::name_at(intrinsic_id()),
441 (is_virtual() ? " virtual='1'" : ""),
442 C->unique() - nodes);
443 }
444 return slow_ctl; // Could be NULL if the check folds.
445 }
446
447 // The intrinsic bailed out
448 if (C->print_intrinsics() || C->print_inlining()) {
449 if (jvms->has_method()) {
450 // Not a root compile.
451 const char* msg = "failed to generate predicate for intrinsic";
452 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
453 } else {
454 // Root compile
455 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
456 vmIntrinsics::name_at(intrinsic_id()),
457 (is_virtual() ? " (virtual)" : ""), bci);
458 }
459 }
460 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
461 return NULL;
462 }
463
464 bool LibraryCallKit::try_to_inline(int predicate) {
465 // Handle symbolic names for otherwise undistinguished boolean switches:
466 const bool is_store = true;
467 const bool is_compress = true;
468 const bool is_native_ptr = true;
469 const bool is_static = true;
470 const bool is_volatile = true;
471
472 if (!jvms()->has_method()) {
473 // Root JVMState has a null method.
474 assert(map()->memory()->Opcode() == Op_Parm, "");
475 // Insert the memory aliasing node
476 set_all_memory(reset_memory());
477 }
478 assert(merged_memory(), "");
479
480
481 switch (intrinsic_id()) {
482 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
483 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
484 case vmIntrinsics::_getClass: return inline_native_getClass();
485
486 case vmIntrinsics::_dsin:
487 case vmIntrinsics::_dcos:
488 case vmIntrinsics::_dtan:
489 case vmIntrinsics::_dabs:
490 case vmIntrinsics::_datan2:
491 case vmIntrinsics::_dsqrt:
492 case vmIntrinsics::_dexp:
493 case vmIntrinsics::_dlog:
494 case vmIntrinsics::_dlog10:
495 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
496
497 case vmIntrinsics::_min:
498 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
499
500 case vmIntrinsics::_notify:
501 case vmIntrinsics::_notifyAll:
502 if (InlineNotify) {
503 return inline_notify(intrinsic_id());
504 }
505 return false;
506
507 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
508 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
509 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
510 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
511 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
512 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
513 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
514 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
515 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
516 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
517 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
518 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
519
520 case vmIntrinsics::_arraycopy: return inline_arraycopy();
521
522 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
523 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
524 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
525 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
526
527 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
528 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
529 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
530 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
531 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
532 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
533 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
534
535 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
536 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
537
538 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
539 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
540 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
541 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
542
543 case vmIntrinsics::_compressStringC:
544 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
545 case vmIntrinsics::_inflateStringC:
546 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
547
548 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
549 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
550 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
551 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
552 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
553 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
554 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
555 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
556 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
557 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
558 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
559 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
560 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
561 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
562 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
563 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
564 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
565 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
566
567 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
568 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
569 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
570 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
571 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
572 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
573 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
574 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
575
576 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
577 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
578 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
579 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
580 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
581 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
582 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
583 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
584
585 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
586 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
587 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
588 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
589 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
590 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
591 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
592 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
593 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
594
595 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
596 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
597 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
598 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
599 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
600 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
601 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
602 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
603 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
604
605 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
606 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
607 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
608 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
609
610 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
611 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
612 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
613 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
614
615 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
616 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
617 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
618
619 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
620 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
621 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
622
623 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
624 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
625 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
626 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
627 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
628
629 case vmIntrinsics::_loadFence:
630 case vmIntrinsics::_storeFence:
631 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
632
633 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
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::_clone: return inline_native_clone(intrinsic()->is_virtual());
654
655 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
656
657 case vmIntrinsics::_isInstance:
658 case vmIntrinsics::_getModifiers:
659 case vmIntrinsics::_isInterface:
660 case vmIntrinsics::_isArray:
661 case vmIntrinsics::_isPrimitive:
662 case vmIntrinsics::_getSuperclass:
663 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
664
665 case vmIntrinsics::_floatToRawIntBits:
666 case vmIntrinsics::_floatToIntBits:
667 case vmIntrinsics::_intBitsToFloat:
668 case vmIntrinsics::_doubleToRawLongBits:
669 case vmIntrinsics::_doubleToLongBits:
670 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
671
672 case vmIntrinsics::_numberOfLeadingZeros_i:
673 case vmIntrinsics::_numberOfLeadingZeros_l:
674 case vmIntrinsics::_numberOfTrailingZeros_i:
675 case vmIntrinsics::_numberOfTrailingZeros_l:
676 case vmIntrinsics::_bitCount_i:
677 case vmIntrinsics::_bitCount_l:
678 case vmIntrinsics::_reverseBytes_i:
679 case vmIntrinsics::_reverseBytes_l:
680 case vmIntrinsics::_reverseBytes_s:
681 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
682
683 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
684
685 case vmIntrinsics::_Reference_get: return inline_reference_get();
686
687 case vmIntrinsics::_Class_cast: return inline_Class_cast();
688
689 case vmIntrinsics::_aescrypt_encryptBlock:
690 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
691
692 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
693 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
694 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
695
696 case vmIntrinsics::_sha_implCompress:
697 case vmIntrinsics::_sha2_implCompress:
698 case vmIntrinsics::_sha5_implCompress:
699 return inline_sha_implCompress(intrinsic_id());
700
701 case vmIntrinsics::_digestBase_implCompressMB:
702 return inline_digestBase_implCompressMB(predicate);
703
704 case vmIntrinsics::_multiplyToLen:
705 return inline_multiplyToLen();
706
707 case vmIntrinsics::_squareToLen:
708 return inline_squareToLen();
709
710 case vmIntrinsics::_mulAdd:
711 return inline_mulAdd();
712
713 case vmIntrinsics::_montgomeryMultiply:
714 return inline_montgomeryMultiply();
715 case vmIntrinsics::_montgomerySquare:
716 return inline_montgomerySquare();
717
718 case vmIntrinsics::_ghash_processBlocks:
719 return inline_ghash_processBlocks();
720
721 case vmIntrinsics::_encodeISOArray:
722 case vmIntrinsics::_encodeByteISOArray:
723 return inline_encodeISOArray();
724
725 case vmIntrinsics::_updateCRC32:
726 return inline_updateCRC32();
727 case vmIntrinsics::_updateBytesCRC32:
728 return inline_updateBytesCRC32();
729 case vmIntrinsics::_updateByteBufferCRC32:
730 return inline_updateByteBufferCRC32();
731
732 case vmIntrinsics::_updateBytesCRC32C:
733 return inline_updateBytesCRC32C();
734 case vmIntrinsics::_updateDirectByteBufferCRC32C:
735 return inline_updateDirectByteBufferCRC32C();
736
737 case vmIntrinsics::_updateBytesAdler32:
738 return inline_updateBytesAdler32();
739 case vmIntrinsics::_updateByteBufferAdler32:
740 return inline_updateByteBufferAdler32();
741
742 case vmIntrinsics::_profileBoolean:
743 return inline_profileBoolean();
744 case vmIntrinsics::_isCompileConstant:
745 return inline_isCompileConstant();
746
747 case vmIntrinsics::_hasNegatives:
748 return inline_hasNegatives();
749
750 default:
751 // If you get here, it may be that someone has added a new intrinsic
752 // to the list in vmSymbols.hpp without implementing it here.
753 #ifndef PRODUCT
754 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
755 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
756 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
757 }
758 #endif
759 return false;
760 }
761 }
762
763 Node* LibraryCallKit::try_to_predicate(int predicate) {
764 if (!jvms()->has_method()) {
765 // Root JVMState has a null method.
766 assert(map()->memory()->Opcode() == Op_Parm, "");
767 // Insert the memory aliasing node
768 set_all_memory(reset_memory());
769 }
770 assert(merged_memory(), "");
771
772 switch (intrinsic_id()) {
773 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
774 return inline_cipherBlockChaining_AESCrypt_predicate(false);
775 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
776 return inline_cipherBlockChaining_AESCrypt_predicate(true);
777 case vmIntrinsics::_digestBase_implCompressMB:
778 return inline_digestBase_implCompressMB_predicate(predicate);
779
780 default:
781 // If you get here, it may be that someone has added a new intrinsic
782 // to the list in vmSymbols.hpp without implementing it here.
783 #ifndef PRODUCT
784 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
785 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
786 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
787 }
788 #endif
789 Node* slow_ctl = control();
790 set_control(top()); // No fast path instrinsic
791 return slow_ctl;
792 }
793 }
794
795 //------------------------------set_result-------------------------------
796 // Helper function for finishing intrinsics.
797 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
798 record_for_igvn(region);
799 set_control(_gvn.transform(region));
800 set_result( _gvn.transform(value));
801 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
802 }
803
804 //------------------------------generate_guard---------------------------
805 // Helper function for generating guarded fast-slow graph structures.
806 // The given 'test', if true, guards a slow path. If the test fails
807 // then a fast path can be taken. (We generally hope it fails.)
808 // In all cases, GraphKit::control() is updated to the fast path.
809 // The returned value represents the control for the slow path.
810 // The return value is never 'top'; it is either a valid control
811 // or NULL if it is obvious that the slow path can never be taken.
812 // Also, if region and the slow control are not NULL, the slow edge
813 // is appended to the region.
814 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
815 if (stopped()) {
816 // Already short circuited.
817 return NULL;
818 }
819
820 // Build an if node and its projections.
821 // If test is true we take the slow path, which we assume is uncommon.
822 if (_gvn.type(test) == TypeInt::ZERO) {
823 // The slow branch is never taken. No need to build this guard.
824 return NULL;
825 }
826
827 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
828
829 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
830 if (if_slow == top()) {
831 // The slow branch is never taken. No need to build this guard.
832 return NULL;
833 }
834
835 if (region != NULL)
836 region->add_req(if_slow);
837
838 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
839 set_control(if_fast);
840
841 return if_slow;
842 }
843
844 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
845 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
846 }
847 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
848 return generate_guard(test, region, PROB_FAIR);
849 }
850
851 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
852 Node* *pos_index) {
853 if (stopped())
854 return NULL; // already stopped
855 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
856 return NULL; // index is already adequately typed
857 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
858 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
859 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
860 if (is_neg != NULL && pos_index != NULL) {
861 // Emulate effect of Parse::adjust_map_after_if.
862 Node* ccast = new CastIINode(index, TypeInt::POS);
863 ccast->set_req(0, control());
864 (*pos_index) = _gvn.transform(ccast);
865 }
866 return is_neg;
867 }
868
869 // Make sure that 'position' is a valid limit index, in [0..length].
870 // There are two equivalent plans for checking this:
871 // A. (offset + copyLength) unsigned<= arrayLength
872 // B. offset <= (arrayLength - copyLength)
873 // We require that all of the values above, except for the sum and
874 // difference, are already known to be non-negative.
875 // Plan A is robust in the face of overflow, if offset and copyLength
876 // are both hugely positive.
877 //
878 // Plan B is less direct and intuitive, but it does not overflow at
879 // all, since the difference of two non-negatives is always
880 // representable. Whenever Java methods must perform the equivalent
881 // check they generally use Plan B instead of Plan A.
882 // For the moment we use Plan A.
883 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
884 Node* subseq_length,
885 Node* array_length,
886 RegionNode* region) {
887 if (stopped())
888 return NULL; // already stopped
889 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
890 if (zero_offset && subseq_length->eqv_uncast(array_length))
891 return NULL; // common case of whole-array copy
892 Node* last = subseq_length;
893 if (!zero_offset) // last += offset
894 last = _gvn.transform(new AddINode(last, offset));
895 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
896 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
897 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
898 return is_over;
899 }
900
901
902 //--------------------------generate_current_thread--------------------
903 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
904 ciKlass* thread_klass = env()->Thread_klass();
905 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
906 Node* thread = _gvn.transform(new ThreadLocalNode());
907 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
908 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
909 tls_output = thread;
910 return threadObj;
911 }
912
913
914 //------------------------------make_string_method_node------------------------
915 // Helper method for String intrinsic functions. This version is called with
916 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
917 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
918 // containing the lengths of str1 and str2.
919 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
920 Node* result = NULL;
921 switch (opcode) {
922 case Op_StrIndexOf:
923 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
924 str1_start, cnt1, str2_start, cnt2, ae);
925 break;
926 case Op_StrComp:
927 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
928 str1_start, cnt1, str2_start, cnt2, ae);
929 break;
930 case Op_StrEquals:
931 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
932 str1_start, str2_start, cnt1, ae);
933 break;
934 default:
935 ShouldNotReachHere();
936 return NULL;
937 }
938
939 // All these intrinsics have checks.
940 C->set_has_split_ifs(true); // Has chance for split-if optimization
941
942 return _gvn.transform(result);
943 }
944
945 //------------------------------inline_string_compareTo------------------------
946 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
947 Node* arg1 = argument(0);
948 Node* arg2 = argument(1);
949
950 // Get start addr and length of first argument
951 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
952 Node* arg1_cnt = load_array_length(arg1);
953
954 // Get start addr and length of second argument
955 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
956 Node* arg2_cnt = load_array_length(arg2);
957
958 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
959 set_result(result);
960 return true;
961 }
962
963 //------------------------------inline_string_equals------------------------
964 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
965 Node* arg1 = argument(0);
966 Node* arg2 = argument(1);
967
968 // paths (plus control) merge
969 RegionNode* region = new RegionNode(3);
970 Node* phi = new PhiNode(region, TypeInt::BOOL);
971
972 if (!stopped()) {
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 // Check for arg1_cnt != arg2_cnt
982 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
983 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
984 Node* if_ne = generate_slow_guard(bol, NULL);
985 if (if_ne != NULL) {
986 phi->init_req(2, intcon(0));
987 region->init_req(2, if_ne);
988 }
989
990 // Check for count == 0 is done by assembler code for StrEquals.
991
992 if (!stopped()) {
993 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
994 phi->init_req(1, equals);
995 region->init_req(1, control());
996 }
997 }
998
999 // post merge
1000 set_control(_gvn.transform(region));
1001 record_for_igvn(region);
1002
1003 set_result(_gvn.transform(phi));
1004 return true;
1005 }
1006
1007 //------------------------------inline_array_equals----------------------------
1008 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1009 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1010 Node* arg1 = argument(0);
1011 Node* arg2 = argument(1);
1012
1013 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1014 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1015 return true;
1016 }
1017
1018 //------------------------------inline_hasNegatives------------------------------
1019 bool LibraryCallKit::inline_hasNegatives() {
1020 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1021
1022 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1023 // no receiver since it is static method
1024 Node* ba = argument(0);
1025 Node* offset = argument(1);
1026 Node* len = argument(2);
1027
1028 RegionNode* bailout = new RegionNode(1);
1029 record_for_igvn(bailout);
1030
1031 // offset must not be negative.
1032 generate_negative_guard(offset, bailout);
1033
1034 // offset + length must not exceed length of ba.
1035 generate_limit_guard(offset, len, load_array_length(ba), bailout);
1036
1037 if (bailout->req() > 1) {
1038 PreserveJVMState pjvms(this);
1039 set_control(_gvn.transform(bailout));
1040 uncommon_trap(Deoptimization::Reason_intrinsic,
1041 Deoptimization::Action_maybe_recompile);
1042 }
1043 if (!stopped()) {
1044 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1045 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1046 set_result(_gvn.transform(result));
1047 }
1048 return true;
1049 }
1050
1051 //------------------------------inline_string_indexOf------------------------
1052 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1053 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1054 return false;
1055 }
1056 Node* src = argument(0);
1057 Node* tgt = argument(1);
1058
1059 // Make the merge point
1060 RegionNode* result_rgn = new RegionNode(4);
1061 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1062
1063 // Get start addr and length of source string
1064 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1065 Node* src_count = load_array_length(src);
1066
1067 // Get start addr and length of substring
1068 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1069 Node* tgt_count = load_array_length(tgt);
1070
1071 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1072 // Divide src size by 2 if String is UTF16 encoded
1073 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1074 }
1075 if (ae == StrIntrinsicNode::UU) {
1076 // Divide substring size by 2 if String is UTF16 encoded
1077 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1078 }
1079
1080 // Check for substr count > string count
1081 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1082 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1083 Node* if_gt = generate_slow_guard(bol, NULL);
1084 if (if_gt != NULL) {
1085 result_phi->init_req(2, intcon(-1));
1086 result_rgn->init_req(2, if_gt);
1087 }
1088
1089 if (!stopped()) {
1090 // Check for substr count == 0
1091 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1092 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1093 Node* if_zero = generate_slow_guard(bol, NULL);
1094 if (if_zero != NULL) {
1095 result_phi->init_req(3, intcon(0));
1096 result_rgn->init_req(3, if_zero);
1097 }
1098 }
1099
1100 if (!stopped()) {
1101 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1102 result_phi->init_req(1, result);
1103 result_rgn->init_req(1, control());
1104 }
1105 set_control(_gvn.transform(result_rgn));
1106 record_for_igvn(result_rgn);
1107 set_result(_gvn.transform(result_phi));
1108
1109 return true;
1110 }
1111
1112 //-----------------------------inline_string_indexOf-----------------------
1113 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1114 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1115 return false;
1116 }
1117 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1118 Node* src = argument(0); // byte[]
1119 Node* src_count = argument(1);
1120 Node* tgt = argument(2); // byte[]
1121 Node* tgt_count = argument(3);
1122 Node* from_index = argument(4);
1123
1124 // Java code which calls this method has range checks for from_index value.
1125 src_count = _gvn.transform(new SubINode(src_count, from_index));
1126
1127 // Multiply byte array index by 2 if String is UTF16 encoded
1128 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1129 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1130 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1131
1132 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1133
1134 // The result is index relative to from_index if substring was found, -1 otherwise.
1135 // Generate code which will fold into cmove.
1136 RegionNode* region = new RegionNode(3);
1137 Node* phi = new PhiNode(region, TypeInt::INT);
1138
1139 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1140 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1141
1142 Node* if_lt = generate_slow_guard(bol, NULL);
1143 if (if_lt != NULL) {
1144 // result == -1
1145 phi->init_req(2, result);
1146 region->init_req(2, if_lt);
1147 }
1148 if (!stopped()) {
1149 result = _gvn.transform(new AddINode(result, from_index));
1150 phi->init_req(1, result);
1151 region->init_req(1, control());
1152 }
1153
1154 set_control(_gvn.transform(region));
1155 record_for_igvn(region);
1156 set_result(_gvn.transform(phi));
1157
1158 return true;
1159 }
1160
1161 //-----------------------------inline_string_indexOfChar-----------------------
1162 bool LibraryCallKit::inline_string_indexOfChar() {
1163 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) {
1164 return false;
1165 }
1166 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1167 Node* src = argument(0); // byte[]
1168 Node* tgt = argument(1); // tgt is int ch
1169 Node* from_index = argument(2);
1170 Node* max = argument(3);
1171
1172 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1173 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1174
1175 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1176
1177 RegionNode* region = new RegionNode(3);
1178 Node* phi = new PhiNode(region, TypeInt::INT);
1179
1180 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1181 C->set_has_split_ifs(true); // Has chance for split-if optimization
1182 _gvn.transform(result);
1183
1184 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1185 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1186
1187 Node* if_lt = generate_slow_guard(bol, NULL);
1188 if (if_lt != NULL) {
1189 // result == -1
1190 phi->init_req(2, result);
1191 region->init_req(2, if_lt);
1192 }
1193 if (!stopped()) {
1194 result = _gvn.transform(new AddINode(result, from_index));
1195 phi->init_req(1, result);
1196 region->init_req(1, control());
1197 }
1198 set_control(_gvn.transform(region));
1199 record_for_igvn(region);
1200 set_result(_gvn.transform(phi));
1201
1202 return true;
1203 }
1204 //---------------------------inline_string_copy---------------------
1205 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1206 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1207 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1208 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1209 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1210 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1211 bool LibraryCallKit::inline_string_copy(bool compress) {
1212 int nargs = 5; // 2 oops, 3 ints
1213 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1214
1215 Node* src = argument(0);
1216 Node* src_offset = argument(1);
1217 Node* dst = argument(2);
1218 Node* dst_offset = argument(3);
1219 Node* length = argument(4);
1220
1221 // Check for allocation before we add nodes that would confuse
1222 // tightly_coupled_allocation()
1223 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1224
1225 // Figure out the size and type of the elements we will be copying.
1226 const Type* src_type = src->Value(&_gvn);
1227 const Type* dst_type = dst->Value(&_gvn);
1228 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1229 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1230 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1231 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1232 "Unsupported array types for inline_string_copy");
1233
1234 // Convert char[] offsets to byte[] offsets
1235 if (compress && src_elem == T_BYTE) {
1236 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1237 } else if (!compress && dst_elem == T_BYTE) {
1238 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1239 }
1240
1241 Node* src_start = array_element_address(src, src_offset, src_elem);
1242 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1243 // 'src_start' points to src array + scaled offset
1244 // 'dst_start' points to dst array + scaled offset
1245 Node* count = NULL;
1246 if (compress) {
1247 count = compress_string(src_start, dst_start, length);
1248 } else {
1249 inflate_string(src_start, dst_start, length);
1250 }
1251
1252 if (alloc != NULL) {
1253 if (alloc->maybe_set_complete(&_gvn)) {
1254 // "You break it, you buy it."
1255 InitializeNode* init = alloc->initialization();
1256 assert(init->is_complete(), "we just did this");
1257 init->set_complete_with_arraycopy();
1258 assert(dst->is_CheckCastPP(), "sanity");
1259 assert(dst->in(0)->in(0) == init, "dest pinned");
1260 }
1261 // Do not let stores that initialize this object be reordered with
1262 // a subsequent store that would make this object accessible by
1263 // other threads.
1264 // Record what AllocateNode this StoreStore protects so that
1265 // escape analysis can go from the MemBarStoreStoreNode to the
1266 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1267 // based on the escape status of the AllocateNode.
1268 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1269 }
1270 if (compress) {
1271 set_result(_gvn.transform(count));
1272 }
1273 return true;
1274 }
1275
1276 #ifdef _LP64
1277 #define XTOP ,top() /*additional argument*/
1278 #else //_LP64
1279 #define XTOP /*no additional argument*/
1280 #endif //_LP64
1281
1282 //------------------------inline_string_toBytesU--------------------------
1283 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1284 bool LibraryCallKit::inline_string_toBytesU() {
1285 // Get the arguments.
1286 Node* value = argument(0);
1287 Node* offset = argument(1);
1288 Node* length = argument(2);
1289
1290 Node* newcopy = NULL;
1291
1292 // Set the original stack and the reexecute bit for the interpreter to reexecute
1293 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1294 { PreserveReexecuteState preexecs(this);
1295 jvms()->set_should_reexecute(true);
1296
1297 // Check if a null path was taken unconditionally.
1298 value = null_check(value);
1299
1300 RegionNode* bailout = new RegionNode(1);
1301 record_for_igvn(bailout);
1302
1303 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1304 generate_negative_guard(length, bailout);
1305 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1306
1307 if (bailout->req() > 1) {
1308 PreserveJVMState pjvms(this);
1309 set_control(_gvn.transform(bailout));
1310 uncommon_trap(Deoptimization::Reason_intrinsic,
1311 Deoptimization::Action_maybe_recompile);
1312 }
1313 if (stopped()) return true;
1314
1315 // Range checks are done by caller.
1316
1317 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1318 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1319 newcopy = new_array(klass_node, size, 0); // no arguments to push
1320 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1321
1322 // Calculate starting addresses.
1323 Node* src_start = array_element_address(value, offset, T_CHAR);
1324 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1325
1326 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1327 const TypeInt* toffset = gvn().type(offset)->is_int();
1328 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1329
1330 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1331 const char* copyfunc_name = "arraycopy";
1332 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1333 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1334 OptoRuntime::fast_arraycopy_Type(),
1335 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1336 src_start, dst_start, ConvI2X(length) XTOP);
1337 // Do not let reads from the cloned object float above the arraycopy.
1338 if (alloc != NULL) {
1339 if (alloc->maybe_set_complete(&_gvn)) {
1340 // "You break it, you buy it."
1341 InitializeNode* init = alloc->initialization();
1342 assert(init->is_complete(), "we just did this");
1343 init->set_complete_with_arraycopy();
1344 assert(newcopy->is_CheckCastPP(), "sanity");
1345 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1346 }
1347 // Do not let stores that initialize this object be reordered with
1348 // a subsequent store that would make this object accessible by
1349 // other threads.
1350 // Record what AllocateNode this StoreStore protects so that
1351 // escape analysis can go from the MemBarStoreStoreNode to the
1352 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1353 // based on the escape status of the AllocateNode.
1354 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1355 } else {
1356 insert_mem_bar(Op_MemBarCPUOrder);
1357 }
1358 } // original reexecute is set back here
1359
1360 C->set_has_split_ifs(true); // Has chance for split-if optimization
1361 if (!stopped()) {
1362 set_result(newcopy);
1363 }
1364 return true;
1365 }
1366
1367 //------------------------inline_string_getCharsU--------------------------
1368 // public void StringUTF16.getChars(byte[] value, int srcBegin, int srcEnd, char dst[], int dstBegin)
1369 bool LibraryCallKit::inline_string_getCharsU() {
1370 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1371
1372 // Get the arguments.
1373 Node* value = argument(0);
1374 Node* src_begin = argument(1);
1375 Node* src_end = argument(2); // exclusive offset (i < src_end)
1376 Node* dst = argument(3);
1377 Node* dst_begin = argument(4);
1378
1379 // Check for allocation before we add nodes that would confuse
1380 // tightly_coupled_allocation()
1381 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1382
1383 // Check if a null path was taken unconditionally.
1384 value = null_check(value);
1385 dst = null_check(dst);
1386 if (stopped()) {
1387 return true;
1388 }
1389
1390 // Range checks are done by caller.
1391
1392 // Get length and convert char[] offset to byte[] offset
1393 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1394 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1395
1396 if (!stopped()) {
1397 // Calculate starting addresses.
1398 Node* src_start = array_element_address(value, src_begin, T_BYTE);
1399 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1400
1401 // Check if array addresses are aligned to HeapWordSize
1402 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1403 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1404 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1405 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1406
1407 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1408 const char* copyfunc_name = "arraycopy";
1409 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1410 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1411 OptoRuntime::fast_arraycopy_Type(),
1412 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1413 src_start, dst_start, ConvI2X(length) XTOP);
1414 // Do not let reads from the cloned object float above the arraycopy.
1415 if (alloc != NULL) {
1416 if (alloc->maybe_set_complete(&_gvn)) {
1417 // "You break it, you buy it."
1418 InitializeNode* init = alloc->initialization();
1419 assert(init->is_complete(), "we just did this");
1420 init->set_complete_with_arraycopy();
1421 assert(dst->is_CheckCastPP(), "sanity");
1422 assert(dst->in(0)->in(0) == init, "dest pinned");
1423 }
1424 // Do not let stores that initialize this object be reordered with
1425 // a subsequent store that would make this object accessible by
1426 // other threads.
1427 // Record what AllocateNode this StoreStore protects so that
1428 // escape analysis can go from the MemBarStoreStoreNode to the
1429 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1430 // based on the escape status of the AllocateNode.
1431 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1432 } else {
1433 insert_mem_bar(Op_MemBarCPUOrder);
1434 }
1435 }
1436
1437 C->set_has_split_ifs(true); // Has chance for split-if optimization
1438 return true;
1439 }
1440
1441 //----------------------inline_string_char_access----------------------------
1442 // Store/Load char to/from byte[] array.
1443 // static void StringUTF16.putChar(byte[] val, int index, int c)
1444 // static char StringUTF16.getChar(byte[] val, int index)
1445 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1446 Node* value = argument(0);
1447 Node* index = argument(1);
1448 Node* ch = is_store ? argument(2) : NULL;
1449
1450 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1451 // correctly requires matched array shapes.
1452 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1453 "sanity: byte[] and char[] bases agree");
1454 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1455 "sanity: byte[] and char[] scales agree");
1456
1457 Node* adr = array_element_address(value, index, T_CHAR);
1458 if (is_store) {
1459 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered);
1460 } else {
1461 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered);
1462 set_result(ch);
1463 }
1464 return true;
1465 }
1466
1467 //--------------------------round_double_node--------------------------------
1468 // Round a double node if necessary.
1469 Node* LibraryCallKit::round_double_node(Node* n) {
1470 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1471 n = _gvn.transform(new RoundDoubleNode(0, n));
1472 return n;
1473 }
1474
1475 //------------------------------inline_math-----------------------------------
1476 // public static double Math.abs(double)
1477 // public static double Math.sqrt(double)
1478 // public static double Math.log(double)
1479 // public static double Math.log10(double)
1480 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1481 Node* arg = round_double_node(argument(0));
1482 Node* n = NULL;
1483 switch (id) {
1484 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1485 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1486 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1487 default: fatal_unexpected_iid(id); break;
1488 }
1489 set_result(_gvn.transform(n));
1490 return true;
1491 }
1492
1493 //------------------------------inline_trig----------------------------------
1494 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1495 // argument reduction which will turn into a fast/slow diamond.
1496 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1497 Node* arg = round_double_node(argument(0));
1498 Node* n = NULL;
1499
1500 switch (id) {
1501 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1502 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1503 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1504 default: fatal_unexpected_iid(id); break;
1505 }
1506 n = _gvn.transform(n);
1507
1508 // Rounding required? Check for argument reduction!
1509 if (Matcher::strict_fp_requires_explicit_rounding) {
1510 static const double pi_4 = 0.7853981633974483;
1511 static const double neg_pi_4 = -0.7853981633974483;
1512 // pi/2 in 80-bit extended precision
1513 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1514 // -pi/2 in 80-bit extended precision
1515 // 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};
1516 // Cutoff value for using this argument reduction technique
1517 //static const double pi_2_minus_epsilon = 1.564660403643354;
1518 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1519
1520 // Pseudocode for sin:
1521 // if (x <= Math.PI / 4.0) {
1522 // if (x >= -Math.PI / 4.0) return fsin(x);
1523 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1524 // } else {
1525 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1526 // }
1527 // return StrictMath.sin(x);
1528
1529 // Pseudocode for cos:
1530 // if (x <= Math.PI / 4.0) {
1531 // if (x >= -Math.PI / 4.0) return fcos(x);
1532 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1533 // } else {
1534 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1535 // }
1536 // return StrictMath.cos(x);
1537
1538 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1539 // requires a special machine instruction to load it. Instead we'll try
1540 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1541 // probably do the math inside the SIN encoding.
1542
1543 // Make the merge point
1544 RegionNode* r = new RegionNode(3);
1545 Node* phi = new PhiNode(r, Type::DOUBLE);
1546
1547 // Flatten arg so we need only 1 test
1548 Node *abs = _gvn.transform(new AbsDNode(arg));
1549 // Node for PI/4 constant
1550 Node *pi4 = makecon(TypeD::make(pi_4));
1551 // Check PI/4 : abs(arg)
1552 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1553 // Check: If PI/4 < abs(arg) then go slow
1554 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1555 // Branch either way
1556 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1557 set_control(opt_iff(r,iff));
1558
1559 // Set fast path result
1560 phi->init_req(2, n);
1561
1562 // Slow path - non-blocking leaf call
1563 Node* call = NULL;
1564 switch (id) {
1565 case vmIntrinsics::_dsin:
1566 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1567 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1568 "Sin", NULL, arg, top());
1569 break;
1570 case vmIntrinsics::_dcos:
1571 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1572 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1573 "Cos", NULL, arg, top());
1574 break;
1575 case vmIntrinsics::_dtan:
1576 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1577 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1578 "Tan", NULL, arg, top());
1579 break;
1580 }
1581 assert(control()->in(0) == call, "");
1582 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1583 r->init_req(1, control());
1584 phi->init_req(1, slow_result);
1585
1586 // Post-merge
1587 set_control(_gvn.transform(r));
1588 record_for_igvn(r);
1589 n = _gvn.transform(phi);
1590
1591 C->set_has_split_ifs(true); // Has chance for split-if optimization
1592 }
1593 set_result(n);
1594 return true;
1595 }
1596
1597 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1598 //-------------------
1599 //result=(result.isNaN())? funcAddr():result;
1600 // Check: If isNaN() by checking result!=result? then either trap
1601 // or go to runtime
1602 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1603 // Build the boolean node
1604 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1605
1606 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1607 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1608 // The pow or exp intrinsic returned a NaN, which requires a call
1609 // to the runtime. Recompile with the runtime call.
1610 uncommon_trap(Deoptimization::Reason_intrinsic,
1611 Deoptimization::Action_make_not_entrant);
1612 }
1613 return result;
1614 } else {
1615 // If this inlining ever returned NaN in the past, we compile a call
1616 // to the runtime to properly handle corner cases
1617
1618 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1619 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1620 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1621
1622 if (!if_slow->is_top()) {
1623 RegionNode* result_region = new RegionNode(3);
1624 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1625
1626 result_region->init_req(1, if_fast);
1627 result_val->init_req(1, result);
1628
1629 set_control(if_slow);
1630
1631 const TypePtr* no_memory_effects = NULL;
1632 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1633 no_memory_effects,
1634 x, top(), y, y ? top() : NULL);
1635 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1636 #ifdef ASSERT
1637 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1638 assert(value_top == top(), "second value must be top");
1639 #endif
1640
1641 result_region->init_req(2, control());
1642 result_val->init_req(2, value);
1643 set_control(_gvn.transform(result_region));
1644 return _gvn.transform(result_val);
1645 } else {
1646 return result;
1647 }
1648 }
1649 }
1650
1651 //------------------------------inline_pow-------------------------------------
1652 // Inline power instructions, if possible.
1653 bool LibraryCallKit::inline_pow() {
1654 // Pseudocode for pow
1655 // if (y == 2) {
1656 // return x * x;
1657 // } else {
1658 // if (x <= 0.0) {
1659 // long longy = (long)y;
1660 // if ((double)longy == y) { // if y is long
1661 // if (y + 1 == y) longy = 0; // huge number: even
1662 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1663 // } else {
1664 // result = NaN;
1665 // }
1666 // } else {
1667 // result = DPow(x,y);
1668 // }
1669 // if (result != result)? {
1670 // result = uncommon_trap() or runtime_call();
1671 // }
1672 // return result;
1673 // }
1674
1675 Node* x = round_double_node(argument(0));
1676 Node* y = round_double_node(argument(2));
1677
1678 Node* result = NULL;
1679
1680 Node* const_two_node = makecon(TypeD::make(2.0));
1681 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1682 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1683 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1684 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1685 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1686
1687 RegionNode* region_node = new RegionNode(3);
1688 region_node->init_req(1, if_true);
1689
1690 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1691 // special case for x^y where y == 2, we can convert it to x * x
1692 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1693
1694 // set control to if_false since we will now process the false branch
1695 set_control(if_false);
1696
1697 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1698 // Short form: skip the fancy tests and just check for NaN result.
1699 result = _gvn.transform(new PowDNode(C, control(), x, y));
1700 } else {
1701 // If this inlining ever returned NaN in the past, include all
1702 // checks + call to the runtime.
1703
1704 // Set the merge point for If node with condition of (x <= 0.0)
1705 // There are four possible paths to region node and phi node
1706 RegionNode *r = new RegionNode(4);
1707 Node *phi = new PhiNode(r, Type::DOUBLE);
1708
1709 // Build the first if node: if (x <= 0.0)
1710 // Node for 0 constant
1711 Node *zeronode = makecon(TypeD::ZERO);
1712 // Check x:0
1713 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1714 // Check: If (x<=0) then go complex path
1715 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1716 // Branch either way
1717 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1718 // Fast path taken; set region slot 3
1719 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1720 r->init_req(3,fast_taken); // Capture fast-control
1721
1722 // Fast path not-taken, i.e. slow path
1723 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1724
1725 // Set fast path result
1726 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1727 phi->init_req(3, fast_result);
1728
1729 // Complex path
1730 // Build the second if node (if y is long)
1731 // Node for (long)y
1732 Node *longy = _gvn.transform(new ConvD2LNode(y));
1733 // Node for (double)((long) y)
1734 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1735 // Check (double)((long) y) : y
1736 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1737 // Check if (y isn't long) then go to slow path
1738
1739 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1740 // Branch either way
1741 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1742 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1743
1744 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1745
1746 // Calculate DPow(abs(x), y)*(1 & (long)y)
1747 // Node for constant 1
1748 Node *conone = longcon(1);
1749 // 1& (long)y
1750 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1751
1752 // A huge number is always even. Detect a huge number by checking
1753 // if y + 1 == y and set integer to be tested for parity to 0.
1754 // Required for corner case:
1755 // (long)9.223372036854776E18 = max_jlong
1756 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1757 // max_jlong is odd but 9.223372036854776E18 is even
1758 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1759 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1760 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1761 Node* correctedsign = NULL;
1762 if (ConditionalMoveLimit != 0) {
1763 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1764 } else {
1765 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1766 RegionNode *r = new RegionNode(3);
1767 Node *phi = new PhiNode(r, TypeLong::LONG);
1768 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1769 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1770 phi->init_req(1, signnode);
1771 phi->init_req(2, longcon(0));
1772 correctedsign = _gvn.transform(phi);
1773 ylong_path = _gvn.transform(r);
1774 record_for_igvn(r);
1775 }
1776
1777 // zero node
1778 Node *conzero = longcon(0);
1779 // Check (1&(long)y)==0?
1780 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1781 // Check if (1&(long)y)!=0?, if so the result is negative
1782 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1783 // abs(x)
1784 Node *absx=_gvn.transform(new AbsDNode(x));
1785 // abs(x)^y
1786 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1787 // -abs(x)^y
1788 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1789 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1790 Node *signresult = NULL;
1791 if (ConditionalMoveLimit != 0) {
1792 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1793 } else {
1794 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1795 RegionNode *r = new RegionNode(3);
1796 Node *phi = new PhiNode(r, Type::DOUBLE);
1797 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1798 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1799 phi->init_req(1, absxpowy);
1800 phi->init_req(2, negabsxpowy);
1801 signresult = _gvn.transform(phi);
1802 ylong_path = _gvn.transform(r);
1803 record_for_igvn(r);
1804 }
1805 // Set complex path fast result
1806 r->init_req(2, ylong_path);
1807 phi->init_req(2, signresult);
1808
1809 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1810 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1811 r->init_req(1,slow_path);
1812 phi->init_req(1,slow_result);
1813
1814 // Post merge
1815 set_control(_gvn.transform(r));
1816 record_for_igvn(r);
1817 result = _gvn.transform(phi);
1818 }
1819
1820 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1821
1822 // control from finish_pow_exp is now input to the region node
1823 region_node->set_req(2, control());
1824 // the result from finish_pow_exp is now input to the phi node
1825 phi_node->init_req(2, result);
1826 set_control(_gvn.transform(region_node));
1827 record_for_igvn(region_node);
1828 set_result(_gvn.transform(phi_node));
1829
1830 C->set_has_split_ifs(true); // Has chance for split-if optimization
1831 return true;
1832 }
1833
1834 //------------------------------runtime_math-----------------------------
1835 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1836 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1837 "must be (DD)D or (D)D type");
1838
1839 // Inputs
1840 Node* a = round_double_node(argument(0));
1841 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1842
1843 const TypePtr* no_memory_effects = NULL;
1844 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1845 no_memory_effects,
1846 a, top(), b, b ? top() : NULL);
1847 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1848 #ifdef ASSERT
1849 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1850 assert(value_top == top(), "second value must be top");
1851 #endif
1852
1853 set_result(value);
1854 return true;
1855 }
1856
1857 //------------------------------inline_math_native-----------------------------
1858 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1859 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1860 switch (id) {
1861 // These intrinsics are not properly supported on all hardware
1862 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1864 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1865 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1866 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1867 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1868
1869 case vmIntrinsics::_dlog:
1870 return StubRoutines::dlog() != NULL ?
1871 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1872 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1873 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1874 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1875
1876 // These intrinsics are supported on all hardware
1877 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1878 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1879
1880 case vmIntrinsics::_dexp:
1881 return StubRoutines::dexp() != NULL ?
1882 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1883 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1884 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1885 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1886 #undef FN_PTR
1887
1888 // These intrinsics are not yet correctly implemented
1889 case vmIntrinsics::_datan2:
1890 return false;
1891
1892 default:
1893 fatal_unexpected_iid(id);
1894 return false;
1895 }
1896 }
1897
1898 static bool is_simple_name(Node* n) {
1899 return (n->req() == 1 // constant
1900 || (n->is_Type() && n->as_Type()->type()->singleton())
1901 || n->is_Proj() // parameter or return value
1902 || n->is_Phi() // local of some sort
1903 );
1904 }
1905
1906 //----------------------------inline_notify-----------------------------------*
1907 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1908 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1909 address func;
1910 if (id == vmIntrinsics::_notify) {
1911 func = OptoRuntime::monitor_notify_Java();
1912 } else {
1913 func = OptoRuntime::monitor_notifyAll_Java();
1914 }
1915 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1916 make_slow_call_ex(call, env()->Throwable_klass(), false);
1917 return true;
1918 }
1919
1920
1921 //----------------------------inline_min_max-----------------------------------
1922 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1923 set_result(generate_min_max(id, argument(0), argument(1)));
1924 return true;
1925 }
1926
1927 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1928 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1929 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1930 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1931 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1932
1933 {
1934 PreserveJVMState pjvms(this);
1935 PreserveReexecuteState preexecs(this);
1936 jvms()->set_should_reexecute(true);
1937
1938 set_control(slow_path);
1939 set_i_o(i_o());
1940
1941 uncommon_trap(Deoptimization::Reason_intrinsic,
1942 Deoptimization::Action_none);
1943 }
1944
1945 set_control(fast_path);
1946 set_result(math);
1947 }
1948
1949 template <typename OverflowOp>
1950 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1951 typedef typename OverflowOp::MathOp MathOp;
1952
1953 MathOp* mathOp = new MathOp(arg1, arg2);
1954 Node* operation = _gvn.transform( mathOp );
1955 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1956 inline_math_mathExact(operation, ofcheck);
1957 return true;
1958 }
1959
1960 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1961 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1962 }
1963
1964 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1965 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1966 }
1967
1968 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1969 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1970 }
1971
1972 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1973 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1974 }
1975
1976 bool LibraryCallKit::inline_math_negateExactI() {
1977 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1978 }
1979
1980 bool LibraryCallKit::inline_math_negateExactL() {
1981 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1982 }
1983
1984 bool LibraryCallKit::inline_math_multiplyExactI() {
1985 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1986 }
1987
1988 bool LibraryCallKit::inline_math_multiplyExactL() {
1989 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1990 }
1991
1992 Node*
1993 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1994 // These are the candidate return value:
1995 Node* xvalue = x0;
1996 Node* yvalue = y0;
1997
1998 if (xvalue == yvalue) {
1999 return xvalue;
2000 }
2001
2002 bool want_max = (id == vmIntrinsics::_max);
2003
2004 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2005 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2006 if (txvalue == NULL || tyvalue == NULL) return top();
2007 // This is not really necessary, but it is consistent with a
2008 // hypothetical MaxINode::Value method:
2009 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2010
2011 // %%% This folding logic should (ideally) be in a different place.
2012 // Some should be inside IfNode, and there to be a more reliable
2013 // transformation of ?: style patterns into cmoves. We also want
2014 // more powerful optimizations around cmove and min/max.
2015
2016 // Try to find a dominating comparison of these guys.
2017 // It can simplify the index computation for Arrays.copyOf
2018 // and similar uses of System.arraycopy.
2019 // First, compute the normalized version of CmpI(x, y).
2020 int cmp_op = Op_CmpI;
2021 Node* xkey = xvalue;
2022 Node* ykey = yvalue;
2023 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2024 if (ideal_cmpxy->is_Cmp()) {
2025 // E.g., if we have CmpI(length - offset, count),
2026 // it might idealize to CmpI(length, count + offset)
2027 cmp_op = ideal_cmpxy->Opcode();
2028 xkey = ideal_cmpxy->in(1);
2029 ykey = ideal_cmpxy->in(2);
2030 }
2031
2032 // Start by locating any relevant comparisons.
2033 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2034 Node* cmpxy = NULL;
2035 Node* cmpyx = NULL;
2036 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2037 Node* cmp = start_from->fast_out(k);
2038 if (cmp->outcnt() > 0 && // must have prior uses
2039 cmp->in(0) == NULL && // must be context-independent
2040 cmp->Opcode() == cmp_op) { // right kind of compare
2041 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2042 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2043 }
2044 }
2045
2046 const int NCMPS = 2;
2047 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2048 int cmpn;
2049 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2050 if (cmps[cmpn] != NULL) break; // find a result
2051 }
2052 if (cmpn < NCMPS) {
2053 // Look for a dominating test that tells us the min and max.
2054 int depth = 0; // Limit search depth for speed
2055 Node* dom = control();
2056 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2057 if (++depth >= 100) break;
2058 Node* ifproj = dom;
2059 if (!ifproj->is_Proj()) continue;
2060 Node* iff = ifproj->in(0);
2061 if (!iff->is_If()) continue;
2062 Node* bol = iff->in(1);
2063 if (!bol->is_Bool()) continue;
2064 Node* cmp = bol->in(1);
2065 if (cmp == NULL) continue;
2066 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2067 if (cmps[cmpn] == cmp) break;
2068 if (cmpn == NCMPS) continue;
2069 BoolTest::mask btest = bol->as_Bool()->_test._test;
2070 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2071 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2072 // At this point, we know that 'x btest y' is true.
2073 switch (btest) {
2074 case BoolTest::eq:
2075 // They are proven equal, so we can collapse the min/max.
2076 // Either value is the answer. Choose the simpler.
2077 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2078 return yvalue;
2079 return xvalue;
2080 case BoolTest::lt: // x < y
2081 case BoolTest::le: // x <= y
2082 return (want_max ? yvalue : xvalue);
2083 case BoolTest::gt: // x > y
2084 case BoolTest::ge: // x >= y
2085 return (want_max ? xvalue : yvalue);
2086 }
2087 }
2088 }
2089
2090 // We failed to find a dominating test.
2091 // Let's pick a test that might GVN with prior tests.
2092 Node* best_bol = NULL;
2093 BoolTest::mask best_btest = BoolTest::illegal;
2094 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2095 Node* cmp = cmps[cmpn];
2096 if (cmp == NULL) continue;
2097 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2098 Node* bol = cmp->fast_out(j);
2099 if (!bol->is_Bool()) continue;
2100 BoolTest::mask btest = bol->as_Bool()->_test._test;
2101 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2102 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2103 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2104 best_bol = bol->as_Bool();
2105 best_btest = btest;
2106 }
2107 }
2108 }
2109
2110 Node* answer_if_true = NULL;
2111 Node* answer_if_false = NULL;
2112 switch (best_btest) {
2113 default:
2114 if (cmpxy == NULL)
2115 cmpxy = ideal_cmpxy;
2116 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2117 // and fall through:
2118 case BoolTest::lt: // x < y
2119 case BoolTest::le: // x <= y
2120 answer_if_true = (want_max ? yvalue : xvalue);
2121 answer_if_false = (want_max ? xvalue : yvalue);
2122 break;
2123 case BoolTest::gt: // x > y
2124 case BoolTest::ge: // x >= y
2125 answer_if_true = (want_max ? xvalue : yvalue);
2126 answer_if_false = (want_max ? yvalue : xvalue);
2127 break;
2128 }
2129
2130 jint hi, lo;
2131 if (want_max) {
2132 // We can sharpen the minimum.
2133 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2134 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2135 } else {
2136 // We can sharpen the maximum.
2137 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2138 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2139 }
2140
2141 // Use a flow-free graph structure, to avoid creating excess control edges
2142 // which could hinder other optimizations.
2143 // Since Math.min/max is often used with arraycopy, we want
2144 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2145 Node* cmov = CMoveNode::make(NULL, best_bol,
2146 answer_if_false, answer_if_true,
2147 TypeInt::make(lo, hi, widen));
2148
2149 return _gvn.transform(cmov);
2150
2151 /*
2152 // This is not as desirable as it may seem, since Min and Max
2153 // nodes do not have a full set of optimizations.
2154 // And they would interfere, anyway, with 'if' optimizations
2155 // and with CMoveI canonical forms.
2156 switch (id) {
2157 case vmIntrinsics::_min:
2158 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2159 case vmIntrinsics::_max:
2160 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2161 default:
2162 ShouldNotReachHere();
2163 }
2164 */
2165 }
2166
2167 inline int
2168 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2169 const TypePtr* base_type = TypePtr::NULL_PTR;
2170 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2171 if (base_type == NULL) {
2172 // Unknown type.
2173 return Type::AnyPtr;
2174 } else if (base_type == TypePtr::NULL_PTR) {
2175 // Since this is a NULL+long form, we have to switch to a rawptr.
2176 base = _gvn.transform(new CastX2PNode(offset));
2177 offset = MakeConX(0);
2178 return Type::RawPtr;
2179 } else if (base_type->base() == Type::RawPtr) {
2180 return Type::RawPtr;
2181 } else if (base_type->isa_oopptr()) {
2182 // Base is never null => always a heap address.
2183 if (base_type->ptr() == TypePtr::NotNull) {
2184 return Type::OopPtr;
2185 }
2186 // Offset is small => always a heap address.
2187 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2188 if (offset_type != NULL &&
2189 base_type->offset() == 0 && // (should always be?)
2190 offset_type->_lo >= 0 &&
2191 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2192 return Type::OopPtr;
2193 }
2194 // Otherwise, it might either be oop+off or NULL+addr.
2195 return Type::AnyPtr;
2196 } else {
2197 // No information:
2198 return Type::AnyPtr;
2199 }
2200 }
2201
2202 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2203 int kind = classify_unsafe_addr(base, offset);
2204 if (kind == Type::RawPtr) {
2205 return basic_plus_adr(top(), base, offset);
2206 } else {
2207 return basic_plus_adr(base, offset);
2208 }
2209 }
2210
2211 //--------------------------inline_number_methods-----------------------------
2212 // inline int Integer.numberOfLeadingZeros(int)
2213 // inline int Long.numberOfLeadingZeros(long)
2214 //
2215 // inline int Integer.numberOfTrailingZeros(int)
2216 // inline int Long.numberOfTrailingZeros(long)
2217 //
2218 // inline int Integer.bitCount(int)
2219 // inline int Long.bitCount(long)
2220 //
2221 // inline char Character.reverseBytes(char)
2222 // inline short Short.reverseBytes(short)
2223 // inline int Integer.reverseBytes(int)
2224 // inline long Long.reverseBytes(long)
2225 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2226 Node* arg = argument(0);
2227 Node* n = NULL;
2228 switch (id) {
2229 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2230 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2231 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2232 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2233 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2234 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2235 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2236 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2237 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2238 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2239 default: fatal_unexpected_iid(id); break;
2240 }
2241 set_result(_gvn.transform(n));
2242 return true;
2243 }
2244
2245 //----------------------------inline_unsafe_access----------------------------
2246
2247 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2248
2249 // Helper that guards and inserts a pre-barrier.
2250 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2251 Node* pre_val, bool need_mem_bar) {
2252 // We could be accessing the referent field of a reference object. If so, when G1
2253 // is enabled, we need to log the value in the referent field in an SATB buffer.
2254 // This routine performs some compile time filters and generates suitable
2255 // runtime filters that guard the pre-barrier code.
2256 // Also add memory barrier for non volatile load from the referent field
2257 // to prevent commoning of loads across safepoint.
2258 if (!UseG1GC && !need_mem_bar)
2259 return;
2260
2261 // Some compile time checks.
2262
2263 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2264 const TypeX* otype = offset->find_intptr_t_type();
2265 if (otype != NULL && otype->is_con() &&
2266 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2267 // Constant offset but not the reference_offset so just return
2268 return;
2269 }
2270
2271 // We only need to generate the runtime guards for instances.
2272 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2273 if (btype != NULL) {
2274 if (btype->isa_aryptr()) {
2275 // Array type so nothing to do
2276 return;
2277 }
2278
2279 const TypeInstPtr* itype = btype->isa_instptr();
2280 if (itype != NULL) {
2281 // Can the klass of base_oop be statically determined to be
2282 // _not_ a sub-class of Reference and _not_ Object?
2283 ciKlass* klass = itype->klass();
2284 if ( klass->is_loaded() &&
2285 !klass->is_subtype_of(env()->Reference_klass()) &&
2286 !env()->Object_klass()->is_subtype_of(klass)) {
2287 return;
2288 }
2289 }
2290 }
2291
2292 // The compile time filters did not reject base_oop/offset so
2293 // we need to generate the following runtime filters
2294 //
2295 // if (offset == java_lang_ref_Reference::_reference_offset) {
2296 // if (instance_of(base, java.lang.ref.Reference)) {
2297 // pre_barrier(_, pre_val, ...);
2298 // }
2299 // }
2300
2301 float likely = PROB_LIKELY( 0.999);
2302 float unlikely = PROB_UNLIKELY(0.999);
2303
2304 IdealKit ideal(this);
2305 #define __ ideal.
2306
2307 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2308
2309 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2310 // Update graphKit memory and control from IdealKit.
2311 sync_kit(ideal);
2312
2313 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2314 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2315
2316 // Update IdealKit memory and control from graphKit.
2317 __ sync_kit(this);
2318
2319 Node* one = __ ConI(1);
2320 // is_instof == 0 if base_oop == NULL
2321 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2322
2323 // Update graphKit from IdeakKit.
2324 sync_kit(ideal);
2325
2326 // Use the pre-barrier to record the value in the referent field
2327 pre_barrier(false /* do_load */,
2328 __ ctrl(),
2329 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2330 pre_val /* pre_val */,
2331 T_OBJECT);
2332 if (need_mem_bar) {
2333 // Add memory barrier to prevent commoning reads from this field
2334 // across safepoint since GC can change its value.
2335 insert_mem_bar(Op_MemBarCPUOrder);
2336 }
2337 // Update IdealKit from graphKit.
2338 __ sync_kit(this);
2339
2340 } __ end_if(); // _ref_type != ref_none
2341 } __ end_if(); // offset == referent_offset
2342
2343 // Final sync IdealKit and GraphKit.
2344 final_sync(ideal);
2345 #undef __
2346 }
2347
2348
2349 // Interpret Unsafe.fieldOffset cookies correctly:
2350 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2351
2352 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2353 // Attempt to infer a sharper value type from the offset and base type.
2354 ciKlass* sharpened_klass = NULL;
2355
2356 // See if it is an instance field, with an object type.
2357 if (alias_type->field() != NULL) {
2358 assert(!is_native_ptr, "native pointer op cannot use a java address");
2359 if (alias_type->field()->type()->is_klass()) {
2360 sharpened_klass = alias_type->field()->type()->as_klass();
2361 }
2362 }
2363
2364 // See if it is a narrow oop array.
2365 if (adr_type->isa_aryptr()) {
2366 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2367 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2368 if (elem_type != NULL) {
2369 sharpened_klass = elem_type->klass();
2370 }
2371 }
2372 }
2373
2374 // The sharpened class might be unloaded if there is no class loader
2375 // contraint in place.
2376 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2377 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2378
2379 #ifndef PRODUCT
2380 if (C->print_intrinsics() || C->print_inlining()) {
2381 tty->print(" from base type: "); adr_type->dump();
2382 tty->print(" sharpened value: "); tjp->dump();
2383 }
2384 #endif
2385 // Sharpen the value type.
2386 return tjp;
2387 }
2388 return NULL;
2389 }
2390
2391 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2392 if (callee()->is_static()) return false; // caller must have the capability!
2393
2394 #ifndef PRODUCT
2395 {
2396 ResourceMark rm;
2397 // Check the signatures.
2398 ciSignature* sig = callee()->signature();
2399 #ifdef ASSERT
2400 if (!is_store) {
2401 // Object getObject(Object base, int/long offset), etc.
2402 BasicType rtype = sig->return_type()->basic_type();
2403 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2404 rtype = T_ADDRESS; // it is really a C void*
2405 assert(rtype == type, "getter must return the expected value");
2406 if (!is_native_ptr) {
2407 assert(sig->count() == 2, "oop getter has 2 arguments");
2408 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2409 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2410 } else {
2411 assert(sig->count() == 1, "native getter has 1 argument");
2412 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2413 }
2414 } else {
2415 // void putObject(Object base, int/long offset, Object x), etc.
2416 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2417 if (!is_native_ptr) {
2418 assert(sig->count() == 3, "oop putter has 3 arguments");
2419 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2420 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2421 } else {
2422 assert(sig->count() == 2, "native putter has 2 arguments");
2423 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2424 }
2425 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2426 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2427 vtype = T_ADDRESS; // it is really a C void*
2428 assert(vtype == type, "putter must accept the expected value");
2429 }
2430 #endif // ASSERT
2431 }
2432 #endif //PRODUCT
2433
2434 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2435
2436 Node* receiver = argument(0); // type: oop
2437
2438 // Build address expression.
2439 Node* adr;
2440 Node* heap_base_oop = top();
2441 Node* offset = top();
2442 Node* val;
2443
2444 if (!is_native_ptr) {
2445 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2446 Node* base = argument(1); // type: oop
2447 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2448 offset = argument(2); // type: long
2449 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2450 // to be plain byte offsets, which are also the same as those accepted
2451 // by oopDesc::field_base.
2452 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2453 "fieldOffset must be byte-scaled");
2454 // 32-bit machines ignore the high half!
2455 offset = ConvL2X(offset);
2456 adr = make_unsafe_address(base, offset);
2457 heap_base_oop = base;
2458 val = is_store ? argument(4) : NULL;
2459 } else {
2460 Node* ptr = argument(1); // type: long
2461 ptr = ConvL2X(ptr); // adjust Java long to machine word
2462 adr = make_unsafe_address(NULL, ptr);
2463 val = is_store ? argument(3) : NULL;
2464 }
2465
2466 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2467
2468 // First guess at the value type.
2469 const Type *value_type = Type::get_const_basic_type(type);
2470
2471 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2472 // there was not enough information to nail it down.
2473 Compile::AliasType* alias_type = C->alias_type(adr_type);
2474 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2475
2476 // We will need memory barriers unless we can determine a unique
2477 // alias category for this reference. (Note: If for some reason
2478 // the barriers get omitted and the unsafe reference begins to "pollute"
2479 // the alias analysis of the rest of the graph, either Compile::can_alias
2480 // or Compile::must_alias will throw a diagnostic assert.)
2481 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2482
2483 // If we are reading the value of the referent field of a Reference
2484 // object (either by using Unsafe directly or through reflection)
2485 // then, if G1 is enabled, we need to record the referent in an
2486 // SATB log buffer using the pre-barrier mechanism.
2487 // Also we need to add memory barrier to prevent commoning reads
2488 // from this field across safepoint since GC can change its value.
2489 bool need_read_barrier = !is_native_ptr && !is_store &&
2490 offset != top() && heap_base_oop != top();
2491
2492 if (!is_store && type == T_OBJECT) {
2493 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2494 if (tjp != NULL) {
2495 value_type = tjp;
2496 }
2497 }
2498
2499 receiver = null_check(receiver);
2500 if (stopped()) {
2501 return true;
2502 }
2503 // Heap pointers get a null-check from the interpreter,
2504 // as a courtesy. However, this is not guaranteed by Unsafe,
2505 // and it is not possible to fully distinguish unintended nulls
2506 // from intended ones in this API.
2507
2508 if (is_volatile) {
2509 // We need to emit leading and trailing CPU membars (see below) in
2510 // addition to memory membars when is_volatile. This is a little
2511 // too strong, but avoids the need to insert per-alias-type
2512 // volatile membars (for stores; compare Parse::do_put_xxx), which
2513 // we cannot do effectively here because we probably only have a
2514 // rough approximation of type.
2515 need_mem_bar = true;
2516 // For Stores, place a memory ordering barrier now.
2517 if (is_store) {
2518 insert_mem_bar(Op_MemBarRelease);
2519 } else {
2520 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2521 insert_mem_bar(Op_MemBarVolatile);
2522 }
2523 }
2524 }
2525
2526 // Memory barrier to prevent normal and 'unsafe' accesses from
2527 // bypassing each other. Happens after null checks, so the
2528 // exception paths do not take memory state from the memory barrier,
2529 // so there's no problems making a strong assert about mixing users
2530 // of safe & unsafe memory.
2531 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2532
2533 if (!is_store) {
2534 Node* p = NULL;
2535 // Try to constant fold a load from a constant field
2536 ciField* field = alias_type->field();
2537 if (heap_base_oop != top() &&
2538 field != NULL && field->is_constant() && field->layout_type() == type) {
2539 // final or stable field
2540 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2541 if (con_type != NULL) {
2542 p = makecon(con_type);
2543 }
2544 }
2545 if (p == NULL) {
2546 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2547 // To be valid, unsafe loads may depend on other conditions than
2548 // the one that guards them: pin the Load node
2549 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2550 // load value
2551 switch (type) {
2552 case T_BOOLEAN:
2553 case T_CHAR:
2554 case T_BYTE:
2555 case T_SHORT:
2556 case T_INT:
2557 case T_LONG:
2558 case T_FLOAT:
2559 case T_DOUBLE:
2560 break;
2561 case T_OBJECT:
2562 if (need_read_barrier) {
2563 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2564 }
2565 break;
2566 case T_ADDRESS:
2567 // Cast to an int type.
2568 p = _gvn.transform(new CastP2XNode(NULL, p));
2569 p = ConvX2UL(p);
2570 break;
2571 default:
2572 fatal("unexpected type %d: %s", type, type2name(type));
2573 break;
2574 }
2575 }
2576 // The load node has the control of the preceding MemBarCPUOrder. All
2577 // following nodes will have the control of the MemBarCPUOrder inserted at
2578 // the end of this method. So, pushing the load onto the stack at a later
2579 // point is fine.
2580 set_result(p);
2581 } else {
2582 // place effect of store into memory
2583 switch (type) {
2584 case T_DOUBLE:
2585 val = dstore_rounding(val);
2586 break;
2587 case T_ADDRESS:
2588 // Repackage the long as a pointer.
2589 val = ConvL2X(val);
2590 val = _gvn.transform(new CastX2PNode(val));
2591 break;
2592 }
2593
2594 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2595 if (type != T_OBJECT ) {
2596 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2597 } else {
2598 // Possibly an oop being stored to Java heap or native memory
2599 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2600 // oop to Java heap.
2601 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2602 } else {
2603 // We can't tell at compile time if we are storing in the Java heap or outside
2604 // of it. So we need to emit code to conditionally do the proper type of
2605 // store.
2606
2607 IdealKit ideal(this);
2608 #define __ ideal.
2609 // QQQ who knows what probability is here??
2610 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2611 // Sync IdealKit and graphKit.
2612 sync_kit(ideal);
2613 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2614 // Update IdealKit memory.
2615 __ sync_kit(this);
2616 } __ else_(); {
2617 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2618 } __ end_if();
2619 // Final sync IdealKit and GraphKit.
2620 final_sync(ideal);
2621 #undef __
2622 }
2623 }
2624 }
2625
2626 if (is_volatile) {
2627 if (!is_store) {
2628 insert_mem_bar(Op_MemBarAcquire);
2629 } else {
2630 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2631 insert_mem_bar(Op_MemBarVolatile);
2632 }
2633 }
2634 }
2635
2636 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2637
2638 return true;
2639 }
2640
2641 //----------------------------inline_unsafe_load_store----------------------------
2642 // This method serves a couple of different customers (depending on LoadStoreKind):
2643 //
2644 // LS_cmpxchg:
2645 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2646 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2647 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2648 //
2649 // LS_xadd:
2650 // public int getAndAddInt( Object o, long offset, int delta)
2651 // public long getAndAddLong(Object o, long offset, long delta)
2652 //
2653 // LS_xchg:
2654 // int getAndSet(Object o, long offset, int newValue)
2655 // long getAndSet(Object o, long offset, long newValue)
2656 // Object getAndSet(Object o, long offset, Object newValue)
2657 //
2658 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2659 // This basic scheme here is the same as inline_unsafe_access, but
2660 // differs in enough details that combining them would make the code
2661 // overly confusing. (This is a true fact! I originally combined
2662 // them, but even I was confused by it!) As much code/comments as
2663 // possible are retained from inline_unsafe_access though to make
2664 // the correspondences clearer. - dl
2665
2666 if (callee()->is_static()) return false; // caller must have the capability!
2667
2668 #ifndef PRODUCT
2669 BasicType rtype;
2670 {
2671 ResourceMark rm;
2672 // Check the signatures.
2673 ciSignature* sig = callee()->signature();
2674 rtype = sig->return_type()->basic_type();
2675 if (kind == LS_xadd || kind == LS_xchg) {
2676 // Check the signatures.
2677 #ifdef ASSERT
2678 assert(rtype == type, "get and set must return the expected type");
2679 assert(sig->count() == 3, "get and set has 3 arguments");
2680 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2681 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2682 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2683 #endif // ASSERT
2684 } else if (kind == LS_cmpxchg) {
2685 // Check the signatures.
2686 #ifdef ASSERT
2687 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2688 assert(sig->count() == 4, "CAS has 4 arguments");
2689 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2690 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2691 #endif // ASSERT
2692 } else {
2693 ShouldNotReachHere();
2694 }
2695 }
2696 #endif //PRODUCT
2697
2698 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2699
2700 // Get arguments:
2701 Node* receiver = NULL;
2702 Node* base = NULL;
2703 Node* offset = NULL;
2704 Node* oldval = NULL;
2705 Node* newval = NULL;
2706 if (kind == LS_cmpxchg) {
2707 const bool two_slot_type = type2size[type] == 2;
2708 receiver = argument(0); // type: oop
2709 base = argument(1); // type: oop
2710 offset = argument(2); // type: long
2711 oldval = argument(4); // type: oop, int, or long
2712 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2713 } else if (kind == LS_xadd || kind == LS_xchg){
2714 receiver = argument(0); // type: oop
2715 base = argument(1); // type: oop
2716 offset = argument(2); // type: long
2717 oldval = NULL;
2718 newval = argument(4); // type: oop, int, or long
2719 }
2720
2721 // Null check receiver.
2722 receiver = null_check(receiver);
2723 if (stopped()) {
2724 return true;
2725 }
2726
2727 // Build field offset expression.
2728 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2729 // to be plain byte offsets, which are also the same as those accepted
2730 // by oopDesc::field_base.
2731 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2732 // 32-bit machines ignore the high half of long offsets
2733 offset = ConvL2X(offset);
2734 Node* adr = make_unsafe_address(base, offset);
2735 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2736
2737 // For CAS, unlike inline_unsafe_access, there seems no point in
2738 // trying to refine types. Just use the coarse types here.
2739 const Type *value_type = Type::get_const_basic_type(type);
2740 Compile::AliasType* alias_type = C->alias_type(adr_type);
2741 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2742
2743 if (kind == LS_xchg && type == T_OBJECT) {
2744 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2745 if (tjp != NULL) {
2746 value_type = tjp;
2747 }
2748 }
2749
2750 int alias_idx = C->get_alias_index(adr_type);
2751
2752 // Memory-model-wise, a LoadStore acts like a little synchronized
2753 // block, so needs barriers on each side. These don't translate
2754 // into actual barriers on most machines, but we still need rest of
2755 // compiler to respect ordering.
2756
2757 insert_mem_bar(Op_MemBarRelease);
2758 insert_mem_bar(Op_MemBarCPUOrder);
2759
2760 // 4984716: MemBars must be inserted before this
2761 // memory node in order to avoid a false
2762 // dependency which will confuse the scheduler.
2763 Node *mem = memory(alias_idx);
2764
2765 // For now, we handle only those cases that actually exist: ints,
2766 // longs, and Object. Adding others should be straightforward.
2767 Node* load_store = NULL;
2768 switch(type) {
2769 case T_INT:
2770 if (kind == LS_xadd) {
2771 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2772 } else if (kind == LS_xchg) {
2773 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2774 } else if (kind == LS_cmpxchg) {
2775 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2776 } else {
2777 ShouldNotReachHere();
2778 }
2779 break;
2780 case T_LONG:
2781 if (kind == LS_xadd) {
2782 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2783 } else if (kind == LS_xchg) {
2784 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2785 } else if (kind == LS_cmpxchg) {
2786 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2787 } else {
2788 ShouldNotReachHere();
2789 }
2790 break;
2791 case T_OBJECT:
2792 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2793 // could be delayed during Parse (for example, in adjust_map_after_if()).
2794 // Execute transformation here to avoid barrier generation in such case.
2795 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2796 newval = _gvn.makecon(TypePtr::NULL_PTR);
2797
2798 // Reference stores need a store barrier.
2799 if (kind == LS_xchg) {
2800 // If pre-barrier must execute before the oop store, old value will require do_load here.
2801 if (!can_move_pre_barrier()) {
2802 pre_barrier(true /* do_load*/,
2803 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2804 NULL /* pre_val*/,
2805 T_OBJECT);
2806 } // Else move pre_barrier to use load_store value, see below.
2807 } else if (kind == LS_cmpxchg) {
2808 // Same as for newval above:
2809 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2810 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2811 }
2812 // The only known value which might get overwritten is oldval.
2813 pre_barrier(false /* do_load */,
2814 control(), NULL, NULL, max_juint, NULL, NULL,
2815 oldval /* pre_val */,
2816 T_OBJECT);
2817 } else {
2818 ShouldNotReachHere();
2819 }
2820
2821 #ifdef _LP64
2822 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2823 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2824 if (kind == LS_xchg) {
2825 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2826 newval_enc, adr_type, value_type->make_narrowoop()));
2827 } else {
2828 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2829 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2830 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2831 newval_enc, oldval_enc));
2832 }
2833 } else
2834 #endif
2835 {
2836 if (kind == LS_xchg) {
2837 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2838 } else {
2839 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2840 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2841 }
2842 }
2843 if (kind == LS_cmpxchg) {
2844 // Emit the post barrier only when the actual store happened.
2845 // This makes sense to check only for compareAndSet that can fail to set the value.
2846 // CAS success path is marked more likely since we anticipate this is a performance
2847 // critical path, while CAS failure path can use the penalty for going through unlikely
2848 // path as backoff. Which is still better than doing a store barrier there.
2849 IdealKit ideal(this);
2850 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2851 sync_kit(ideal);
2852 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2853 ideal.sync_kit(this);
2854 } ideal.end_if();
2855 final_sync(ideal);
2856 } else {
2857 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2858 }
2859 break;
2860 default:
2861 fatal("unexpected type %d: %s", type, type2name(type));
2862 break;
2863 }
2864
2865 // SCMemProjNodes represent the memory state of a LoadStore. Their
2866 // main role is to prevent LoadStore nodes from being optimized away
2867 // when their results aren't used.
2868 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2869 set_memory(proj, alias_idx);
2870
2871 if (type == T_OBJECT && kind == LS_xchg) {
2872 #ifdef _LP64
2873 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2874 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2875 }
2876 #endif
2877 if (can_move_pre_barrier()) {
2878 // Don't need to load pre_val. The old value is returned by load_store.
2879 // The pre_barrier can execute after the xchg as long as no safepoint
2880 // gets inserted between them.
2881 pre_barrier(false /* do_load */,
2882 control(), NULL, NULL, max_juint, NULL, NULL,
2883 load_store /* pre_val */,
2884 T_OBJECT);
2885 }
2886 }
2887
2888 // Add the trailing membar surrounding the access
2889 insert_mem_bar(Op_MemBarCPUOrder);
2890 insert_mem_bar(Op_MemBarAcquire);
2891
2892 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2893 set_result(load_store);
2894 return true;
2895 }
2896
2897 //----------------------------inline_unsafe_ordered_store----------------------
2898 // public native void Unsafe.putOrderedObject(Object o, long offset, Object x);
2899 // public native void Unsafe.putOrderedInt(Object o, long offset, int x);
2900 // public native void Unsafe.putOrderedLong(Object o, long offset, long x);
2901 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2902 // This is another variant of inline_unsafe_access, differing in
2903 // that it always issues store-store ("release") barrier and ensures
2904 // store-atomicity (which only matters for "long").
2905
2906 if (callee()->is_static()) return false; // caller must have the capability!
2907
2908 #ifndef PRODUCT
2909 {
2910 ResourceMark rm;
2911 // Check the signatures.
2912 ciSignature* sig = callee()->signature();
2913 #ifdef ASSERT
2914 BasicType rtype = sig->return_type()->basic_type();
2915 assert(rtype == T_VOID, "must return void");
2916 assert(sig->count() == 3, "has 3 arguments");
2917 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2918 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2919 #endif // ASSERT
2920 }
2921 #endif //PRODUCT
2922
2923 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2924
2925 // Get arguments:
2926 Node* receiver = argument(0); // type: oop
2927 Node* base = argument(1); // type: oop
2928 Node* offset = argument(2); // type: long
2929 Node* val = argument(4); // type: oop, int, or long
2930
2931 // Null check receiver.
2932 receiver = null_check(receiver);
2933 if (stopped()) {
2934 return true;
2935 }
2936
2937 // Build field offset expression.
2938 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2939 // 32-bit machines ignore the high half of long offsets
2940 offset = ConvL2X(offset);
2941 Node* adr = make_unsafe_address(base, offset);
2942 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2943 const Type *value_type = Type::get_const_basic_type(type);
2944 Compile::AliasType* alias_type = C->alias_type(adr_type);
2945
2946 insert_mem_bar(Op_MemBarRelease);
2947 insert_mem_bar(Op_MemBarCPUOrder);
2948 // Ensure that the store is atomic for longs:
2949 const bool require_atomic_access = true;
2950 Node* store;
2951 if (type == T_OBJECT) // reference stores need a store barrier.
2952 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2953 else {
2954 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2955 }
2956 insert_mem_bar(Op_MemBarCPUOrder);
2957 return true;
2958 }
2959
2960 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2961 // Regardless of form, don't allow previous ld/st to move down,
2962 // then issue acquire, release, or volatile mem_bar.
2963 insert_mem_bar(Op_MemBarCPUOrder);
2964 switch(id) {
2965 case vmIntrinsics::_loadFence:
2966 insert_mem_bar(Op_LoadFence);
2967 return true;
2968 case vmIntrinsics::_storeFence:
2969 insert_mem_bar(Op_StoreFence);
2970 return true;
2971 case vmIntrinsics::_fullFence:
2972 insert_mem_bar(Op_MemBarVolatile);
2973 return true;
2974 default:
2975 fatal_unexpected_iid(id);
2976 return false;
2977 }
2978 }
2979
2980 bool LibraryCallKit::inline_onspinwait() {
2981 insert_mem_bar(Op_OnSpinWait);
2982 return true;
2983 }
2984
2985 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2986 if (!kls->is_Con()) {
2987 return true;
2988 }
2989 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2990 if (klsptr == NULL) {
2991 return true;
2992 }
2993 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2994 // don't need a guard for a klass that is already initialized
2995 return !ik->is_initialized();
2996 }
2997
2998 //----------------------------inline_unsafe_allocate---------------------------
2999 // public native Object Unsafe.allocateInstance(Class<?> cls);
3000 bool LibraryCallKit::inline_unsafe_allocate() {
3001 if (callee()->is_static()) return false; // caller must have the capability!
3002
3003 null_check_receiver(); // null-check, then ignore
3004 Node* cls = null_check(argument(1));
3005 if (stopped()) return true;
3006
3007 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3008 kls = null_check(kls);
3009 if (stopped()) return true; // argument was like int.class
3010
3011 Node* test = NULL;
3012 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3013 // Note: The argument might still be an illegal value like
3014 // Serializable.class or Object[].class. The runtime will handle it.
3015 // But we must make an explicit check for initialization.
3016 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3017 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3018 // can generate code to load it as unsigned byte.
3019 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3020 Node* bits = intcon(InstanceKlass::fully_initialized);
3021 test = _gvn.transform(new SubINode(inst, bits));
3022 // The 'test' is non-zero if we need to take a slow path.
3023 }
3024
3025 Node* obj = new_instance(kls, test);
3026 set_result(obj);
3027 return true;
3028 }
3029
3030 #ifdef TRACE_HAVE_INTRINSICS
3031 /*
3032 * oop -> myklass
3033 * myklass->trace_id |= USED
3034 * return myklass->trace_id & ~0x3
3035 */
3036 bool LibraryCallKit::inline_native_classID() {
3037 null_check_receiver(); // null-check, then ignore
3038 Node* cls = null_check(argument(1), T_OBJECT);
3039 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3040 kls = null_check(kls, T_OBJECT);
3041 ByteSize offset = TRACE_ID_OFFSET;
3042 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3043 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3044 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3045 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3046 Node* clsused = longcon(0x01l); // set the class bit
3047 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3048
3049 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3050 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3051 set_result(andl);
3052 return true;
3053 }
3054
3055 bool LibraryCallKit::inline_native_threadID() {
3056 Node* tls_ptr = NULL;
3057 Node* cur_thr = generate_current_thread(tls_ptr);
3058 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3059 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3060 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3061
3062 Node* threadid = NULL;
3063 size_t thread_id_size = OSThread::thread_id_size();
3064 if (thread_id_size == (size_t) BytesPerLong) {
3065 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3066 } else if (thread_id_size == (size_t) BytesPerInt) {
3067 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3068 } else {
3069 ShouldNotReachHere();
3070 }
3071 set_result(threadid);
3072 return true;
3073 }
3074 #endif
3075
3076 //------------------------inline_native_time_funcs--------------
3077 // inline code for System.currentTimeMillis() and System.nanoTime()
3078 // these have the same type and signature
3079 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3080 const TypeFunc* tf = OptoRuntime::void_long_Type();
3081 const TypePtr* no_memory_effects = NULL;
3082 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3083 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3084 #ifdef ASSERT
3085 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3086 assert(value_top == top(), "second value must be top");
3087 #endif
3088 set_result(value);
3089 return true;
3090 }
3091
3092 //------------------------inline_native_currentThread------------------
3093 bool LibraryCallKit::inline_native_currentThread() {
3094 Node* junk = NULL;
3095 set_result(generate_current_thread(junk));
3096 return true;
3097 }
3098
3099 //------------------------inline_native_isInterrupted------------------
3100 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3101 bool LibraryCallKit::inline_native_isInterrupted() {
3102 // Add a fast path to t.isInterrupted(clear_int):
3103 // (t == Thread.current() &&
3104 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3105 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3106 // So, in the common case that the interrupt bit is false,
3107 // we avoid making a call into the VM. Even if the interrupt bit
3108 // is true, if the clear_int argument is false, we avoid the VM call.
3109 // However, if the receiver is not currentThread, we must call the VM,
3110 // because there must be some locking done around the operation.
3111
3112 // We only go to the fast case code if we pass two guards.
3113 // Paths which do not pass are accumulated in the slow_region.
3114
3115 enum {
3116 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3117 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3118 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3119 PATH_LIMIT
3120 };
3121
3122 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3123 // out of the function.
3124 insert_mem_bar(Op_MemBarCPUOrder);
3125
3126 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3127 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3128
3129 RegionNode* slow_region = new RegionNode(1);
3130 record_for_igvn(slow_region);
3131
3132 // (a) Receiving thread must be the current thread.
3133 Node* rec_thr = argument(0);
3134 Node* tls_ptr = NULL;
3135 Node* cur_thr = generate_current_thread(tls_ptr);
3136 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3137 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3138
3139 generate_slow_guard(bol_thr, slow_region);
3140
3141 // (b) Interrupt bit on TLS must be false.
3142 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3143 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3144 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3145
3146 // Set the control input on the field _interrupted read to prevent it floating up.
3147 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3148 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3149 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3150
3151 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3152
3153 // First fast path: if (!TLS._interrupted) return false;
3154 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3155 result_rgn->init_req(no_int_result_path, false_bit);
3156 result_val->init_req(no_int_result_path, intcon(0));
3157
3158 // drop through to next case
3159 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3160
3161 #ifndef TARGET_OS_FAMILY_windows
3162 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3163 Node* clr_arg = argument(1);
3164 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3165 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3166 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3167
3168 // Second fast path: ... else if (!clear_int) return true;
3169 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3170 result_rgn->init_req(no_clear_result_path, false_arg);
3171 result_val->init_req(no_clear_result_path, intcon(1));
3172
3173 // drop through to next case
3174 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3175 #else
3176 // To return true on Windows you must read the _interrupted field
3177 // and check the the event state i.e. take the slow path.
3178 #endif // TARGET_OS_FAMILY_windows
3179
3180 // (d) Otherwise, go to the slow path.
3181 slow_region->add_req(control());
3182 set_control( _gvn.transform(slow_region));
3183
3184 if (stopped()) {
3185 // There is no slow path.
3186 result_rgn->init_req(slow_result_path, top());
3187 result_val->init_req(slow_result_path, top());
3188 } else {
3189 // non-virtual because it is a private non-static
3190 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3191
3192 Node* slow_val = set_results_for_java_call(slow_call);
3193 // this->control() comes from set_results_for_java_call
3194
3195 Node* fast_io = slow_call->in(TypeFunc::I_O);
3196 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3197
3198 // These two phis are pre-filled with copies of of the fast IO and Memory
3199 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3200 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3201
3202 result_rgn->init_req(slow_result_path, control());
3203 result_io ->init_req(slow_result_path, i_o());
3204 result_mem->init_req(slow_result_path, reset_memory());
3205 result_val->init_req(slow_result_path, slow_val);
3206
3207 set_all_memory(_gvn.transform(result_mem));
3208 set_i_o( _gvn.transform(result_io));
3209 }
3210
3211 C->set_has_split_ifs(true); // Has chance for split-if optimization
3212 set_result(result_rgn, result_val);
3213 return true;
3214 }
3215
3216 //---------------------------load_mirror_from_klass----------------------------
3217 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3218 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3219 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3220 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3221 }
3222
3223 //-----------------------load_klass_from_mirror_common-------------------------
3224 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3225 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3226 // and branch to the given path on the region.
3227 // If never_see_null, take an uncommon trap on null, so we can optimistically
3228 // compile for the non-null case.
3229 // If the region is NULL, force never_see_null = true.
3230 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3231 bool never_see_null,
3232 RegionNode* region,
3233 int null_path,
3234 int offset) {
3235 if (region == NULL) never_see_null = true;
3236 Node* p = basic_plus_adr(mirror, offset);
3237 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3238 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3239 Node* null_ctl = top();
3240 kls = null_check_oop(kls, &null_ctl, never_see_null);
3241 if (region != NULL) {
3242 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3243 region->init_req(null_path, null_ctl);
3244 } else {
3245 assert(null_ctl == top(), "no loose ends");
3246 }
3247 return kls;
3248 }
3249
3250 //--------------------(inline_native_Class_query helpers)---------------------
3251 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3252 // Fall through if (mods & mask) == bits, take the guard otherwise.
3253 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3254 // Branch around if the given klass has the given modifier bit set.
3255 // Like generate_guard, adds a new path onto the region.
3256 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3257 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3258 Node* mask = intcon(modifier_mask);
3259 Node* bits = intcon(modifier_bits);
3260 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3261 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3262 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3263 return generate_fair_guard(bol, region);
3264 }
3265 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3266 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3267 }
3268
3269 //-------------------------inline_native_Class_query-------------------
3270 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3271 const Type* return_type = TypeInt::BOOL;
3272 Node* prim_return_value = top(); // what happens if it's a primitive class?
3273 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3274 bool expect_prim = false; // most of these guys expect to work on refs
3275
3276 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3277
3278 Node* mirror = argument(0);
3279 Node* obj = top();
3280
3281 switch (id) {
3282 case vmIntrinsics::_isInstance:
3283 // nothing is an instance of a primitive type
3284 prim_return_value = intcon(0);
3285 obj = argument(1);
3286 break;
3287 case vmIntrinsics::_getModifiers:
3288 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3289 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3290 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3291 break;
3292 case vmIntrinsics::_isInterface:
3293 prim_return_value = intcon(0);
3294 break;
3295 case vmIntrinsics::_isArray:
3296 prim_return_value = intcon(0);
3297 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3298 break;
3299 case vmIntrinsics::_isPrimitive:
3300 prim_return_value = intcon(1);
3301 expect_prim = true; // obviously
3302 break;
3303 case vmIntrinsics::_getSuperclass:
3304 prim_return_value = null();
3305 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3306 break;
3307 case vmIntrinsics::_getClassAccessFlags:
3308 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3309 return_type = TypeInt::INT; // not bool! 6297094
3310 break;
3311 default:
3312 fatal_unexpected_iid(id);
3313 break;
3314 }
3315
3316 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3317 if (mirror_con == NULL) return false; // cannot happen?
3318
3319 #ifndef PRODUCT
3320 if (C->print_intrinsics() || C->print_inlining()) {
3321 ciType* k = mirror_con->java_mirror_type();
3322 if (k) {
3323 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3324 k->print_name();
3325 tty->cr();
3326 }
3327 }
3328 #endif
3329
3330 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3331 RegionNode* region = new RegionNode(PATH_LIMIT);
3332 record_for_igvn(region);
3333 PhiNode* phi = new PhiNode(region, return_type);
3334
3335 // The mirror will never be null of Reflection.getClassAccessFlags, however
3336 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3337 // if it is. See bug 4774291.
3338
3339 // For Reflection.getClassAccessFlags(), the null check occurs in
3340 // the wrong place; see inline_unsafe_access(), above, for a similar
3341 // situation.
3342 mirror = null_check(mirror);
3343 // If mirror or obj is dead, only null-path is taken.
3344 if (stopped()) return true;
3345
3346 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3347
3348 // Now load the mirror's klass metaobject, and null-check it.
3349 // Side-effects region with the control path if the klass is null.
3350 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3351 // If kls is null, we have a primitive mirror.
3352 phi->init_req(_prim_path, prim_return_value);
3353 if (stopped()) { set_result(region, phi); return true; }
3354 bool safe_for_replace = (region->in(_prim_path) == top());
3355
3356 Node* p; // handy temp
3357 Node* null_ctl;
3358
3359 // Now that we have the non-null klass, we can perform the real query.
3360 // For constant classes, the query will constant-fold in LoadNode::Value.
3361 Node* query_value = top();
3362 switch (id) {
3363 case vmIntrinsics::_isInstance:
3364 // nothing is an instance of a primitive type
3365 query_value = gen_instanceof(obj, kls, safe_for_replace);
3366 break;
3367
3368 case vmIntrinsics::_onSpinWait:
3369 break;
3370
3371 case vmIntrinsics::_getModifiers:
3372 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3373 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3374 break;
3375
3376 case vmIntrinsics::_isInterface:
3377 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3378 if (generate_interface_guard(kls, region) != NULL)
3379 // A guard was added. If the guard is taken, it was an interface.
3380 phi->add_req(intcon(1));
3381 // If we fall through, it's a plain class.
3382 query_value = intcon(0);
3383 break;
3384
3385 case vmIntrinsics::_isArray:
3386 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3387 if (generate_array_guard(kls, region) != NULL)
3388 // A guard was added. If the guard is taken, it was an array.
3389 phi->add_req(intcon(1));
3390 // If we fall through, it's a plain class.
3391 query_value = intcon(0);
3392 break;
3393
3394 case vmIntrinsics::_isPrimitive:
3395 query_value = intcon(0); // "normal" path produces false
3396 break;
3397
3398 case vmIntrinsics::_getSuperclass:
3399 // The rules here are somewhat unfortunate, but we can still do better
3400 // with random logic than with a JNI call.
3401 // Interfaces store null or Object as _super, but must report null.
3402 // Arrays store an intermediate super as _super, but must report Object.
3403 // Other types can report the actual _super.
3404 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3405 if (generate_interface_guard(kls, region) != NULL)
3406 // A guard was added. If the guard is taken, it was an interface.
3407 phi->add_req(null());
3408 if (generate_array_guard(kls, region) != NULL)
3409 // A guard was added. If the guard is taken, it was an array.
3410 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3411 // If we fall through, it's a plain class. Get its _super.
3412 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3413 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3414 null_ctl = top();
3415 kls = null_check_oop(kls, &null_ctl);
3416 if (null_ctl != top()) {
3417 // If the guard is taken, Object.superClass is null (both klass and mirror).
3418 region->add_req(null_ctl);
3419 phi ->add_req(null());
3420 }
3421 if (!stopped()) {
3422 query_value = load_mirror_from_klass(kls);
3423 }
3424 break;
3425
3426 case vmIntrinsics::_getClassAccessFlags:
3427 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3428 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3429 break;
3430
3431 default:
3432 fatal_unexpected_iid(id);
3433 break;
3434 }
3435
3436 // Fall-through is the normal case of a query to a real class.
3437 phi->init_req(1, query_value);
3438 region->init_req(1, control());
3439
3440 C->set_has_split_ifs(true); // Has chance for split-if optimization
3441 set_result(region, phi);
3442 return true;
3443 }
3444
3445 //-------------------------inline_Class_cast-------------------
3446 bool LibraryCallKit::inline_Class_cast() {
3447 Node* mirror = argument(0); // Class
3448 Node* obj = argument(1);
3449 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3450 if (mirror_con == NULL) {
3451 return false; // dead path (mirror->is_top()).
3452 }
3453 if (obj == NULL || obj->is_top()) {
3454 return false; // dead path
3455 }
3456 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3457
3458 // First, see if Class.cast() can be folded statically.
3459 // java_mirror_type() returns non-null for compile-time Class constants.
3460 ciType* tm = mirror_con->java_mirror_type();
3461 if (tm != NULL && tm->is_klass() &&
3462 tp != NULL && tp->klass() != NULL) {
3463 if (!tp->klass()->is_loaded()) {
3464 // Don't use intrinsic when class is not loaded.
3465 return false;
3466 } else {
3467 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3468 if (static_res == Compile::SSC_always_true) {
3469 // isInstance() is true - fold the code.
3470 set_result(obj);
3471 return true;
3472 } else if (static_res == Compile::SSC_always_false) {
3473 // Don't use intrinsic, have to throw ClassCastException.
3474 // If the reference is null, the non-intrinsic bytecode will
3475 // be optimized appropriately.
3476 return false;
3477 }
3478 }
3479 }
3480
3481 // Bailout intrinsic and do normal inlining if exception path is frequent.
3482 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3483 return false;
3484 }
3485
3486 // Generate dynamic checks.
3487 // Class.cast() is java implementation of _checkcast bytecode.
3488 // Do checkcast (Parse::do_checkcast()) optimizations here.
3489
3490 mirror = null_check(mirror);
3491 // If mirror is dead, only null-path is taken.
3492 if (stopped()) {
3493 return true;
3494 }
3495
3496 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3497 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3498 RegionNode* region = new RegionNode(PATH_LIMIT);
3499 record_for_igvn(region);
3500
3501 // Now load the mirror's klass metaobject, and null-check it.
3502 // If kls is null, we have a primitive mirror and
3503 // nothing is an instance of a primitive type.
3504 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3505
3506 Node* res = top();
3507 if (!stopped()) {
3508 Node* bad_type_ctrl = top();
3509 // Do checkcast optimizations.
3510 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3511 region->init_req(_bad_type_path, bad_type_ctrl);
3512 }
3513 if (region->in(_prim_path) != top() ||
3514 region->in(_bad_type_path) != top()) {
3515 // Let Interpreter throw ClassCastException.
3516 PreserveJVMState pjvms(this);
3517 set_control(_gvn.transform(region));
3518 uncommon_trap(Deoptimization::Reason_intrinsic,
3519 Deoptimization::Action_maybe_recompile);
3520 }
3521 if (!stopped()) {
3522 set_result(res);
3523 }
3524 return true;
3525 }
3526
3527
3528 //--------------------------inline_native_subtype_check------------------------
3529 // This intrinsic takes the JNI calls out of the heart of
3530 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3531 bool LibraryCallKit::inline_native_subtype_check() {
3532 // Pull both arguments off the stack.
3533 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3534 args[0] = argument(0);
3535 args[1] = argument(1);
3536 Node* klasses[2]; // corresponding Klasses: superk, subk
3537 klasses[0] = klasses[1] = top();
3538
3539 enum {
3540 // A full decision tree on {superc is prim, subc is prim}:
3541 _prim_0_path = 1, // {P,N} => false
3542 // {P,P} & superc!=subc => false
3543 _prim_same_path, // {P,P} & superc==subc => true
3544 _prim_1_path, // {N,P} => false
3545 _ref_subtype_path, // {N,N} & subtype check wins => true
3546 _both_ref_path, // {N,N} & subtype check loses => false
3547 PATH_LIMIT
3548 };
3549
3550 RegionNode* region = new RegionNode(PATH_LIMIT);
3551 Node* phi = new PhiNode(region, TypeInt::BOOL);
3552 record_for_igvn(region);
3553
3554 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3555 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3556 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3557
3558 // First null-check both mirrors and load each mirror's klass metaobject.
3559 int which_arg;
3560 for (which_arg = 0; which_arg <= 1; which_arg++) {
3561 Node* arg = args[which_arg];
3562 arg = null_check(arg);
3563 if (stopped()) break;
3564 args[which_arg] = arg;
3565
3566 Node* p = basic_plus_adr(arg, class_klass_offset);
3567 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3568 klasses[which_arg] = _gvn.transform(kls);
3569 }
3570
3571 // Having loaded both klasses, test each for null.
3572 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3573 for (which_arg = 0; which_arg <= 1; which_arg++) {
3574 Node* kls = klasses[which_arg];
3575 Node* null_ctl = top();
3576 kls = null_check_oop(kls, &null_ctl, never_see_null);
3577 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3578 region->init_req(prim_path, null_ctl);
3579 if (stopped()) break;
3580 klasses[which_arg] = kls;
3581 }
3582
3583 if (!stopped()) {
3584 // now we have two reference types, in klasses[0..1]
3585 Node* subk = klasses[1]; // the argument to isAssignableFrom
3586 Node* superk = klasses[0]; // the receiver
3587 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3588 // now we have a successful reference subtype check
3589 region->set_req(_ref_subtype_path, control());
3590 }
3591
3592 // If both operands are primitive (both klasses null), then
3593 // we must return true when they are identical primitives.
3594 // It is convenient to test this after the first null klass check.
3595 set_control(region->in(_prim_0_path)); // go back to first null check
3596 if (!stopped()) {
3597 // Since superc is primitive, make a guard for the superc==subc case.
3598 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3599 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3600 generate_guard(bol_eq, region, PROB_FAIR);
3601 if (region->req() == PATH_LIMIT+1) {
3602 // A guard was added. If the added guard is taken, superc==subc.
3603 region->swap_edges(PATH_LIMIT, _prim_same_path);
3604 region->del_req(PATH_LIMIT);
3605 }
3606 region->set_req(_prim_0_path, control()); // Not equal after all.
3607 }
3608
3609 // these are the only paths that produce 'true':
3610 phi->set_req(_prim_same_path, intcon(1));
3611 phi->set_req(_ref_subtype_path, intcon(1));
3612
3613 // pull together the cases:
3614 assert(region->req() == PATH_LIMIT, "sane region");
3615 for (uint i = 1; i < region->req(); i++) {
3616 Node* ctl = region->in(i);
3617 if (ctl == NULL || ctl == top()) {
3618 region->set_req(i, top());
3619 phi ->set_req(i, top());
3620 } else if (phi->in(i) == NULL) {
3621 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3622 }
3623 }
3624
3625 set_control(_gvn.transform(region));
3626 set_result(_gvn.transform(phi));
3627 return true;
3628 }
3629
3630 //---------------------generate_array_guard_common------------------------
3631 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3632 bool obj_array, bool not_array) {
3633
3634 if (stopped()) {
3635 return NULL;
3636 }
3637
3638 // If obj_array/non_array==false/false:
3639 // Branch around if the given klass is in fact an array (either obj or prim).
3640 // If obj_array/non_array==false/true:
3641 // Branch around if the given klass is not an array klass of any kind.
3642 // If obj_array/non_array==true/true:
3643 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3644 // If obj_array/non_array==true/false:
3645 // Branch around if the kls is an oop array (Object[] or subtype)
3646 //
3647 // Like generate_guard, adds a new path onto the region.
3648 jint layout_con = 0;
3649 Node* layout_val = get_layout_helper(kls, layout_con);
3650 if (layout_val == NULL) {
3651 bool query = (obj_array
3652 ? Klass::layout_helper_is_objArray(layout_con)
3653 : Klass::layout_helper_is_array(layout_con));
3654 if (query == not_array) {
3655 return NULL; // never a branch
3656 } else { // always a branch
3657 Node* always_branch = control();
3658 if (region != NULL)
3659 region->add_req(always_branch);
3660 set_control(top());
3661 return always_branch;
3662 }
3663 }
3664 // Now test the correct condition.
3665 jint nval = (obj_array
3666 ? ((jint)Klass::_lh_array_tag_type_value
3667 << Klass::_lh_array_tag_shift)
3668 : Klass::_lh_neutral_value);
3669 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3670 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3671 // invert the test if we are looking for a non-array
3672 if (not_array) btest = BoolTest(btest).negate();
3673 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3674 return generate_fair_guard(bol, region);
3675 }
3676
3677
3678 //-----------------------inline_native_newArray--------------------------
3679 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3680 bool LibraryCallKit::inline_native_newArray() {
3681 Node* mirror = argument(0);
3682 Node* count_val = argument(1);
3683
3684 mirror = null_check(mirror);
3685 // If mirror or obj is dead, only null-path is taken.
3686 if (stopped()) return true;
3687
3688 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3689 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3690 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3691 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3692 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3693
3694 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3695 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3696 result_reg, _slow_path);
3697 Node* normal_ctl = control();
3698 Node* no_array_ctl = result_reg->in(_slow_path);
3699
3700 // Generate code for the slow case. We make a call to newArray().
3701 set_control(no_array_ctl);
3702 if (!stopped()) {
3703 // Either the input type is void.class, or else the
3704 // array klass has not yet been cached. Either the
3705 // ensuing call will throw an exception, or else it
3706 // will cache the array klass for next time.
3707 PreserveJVMState pjvms(this);
3708 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3709 Node* slow_result = set_results_for_java_call(slow_call);
3710 // this->control() comes from set_results_for_java_call
3711 result_reg->set_req(_slow_path, control());
3712 result_val->set_req(_slow_path, slow_result);
3713 result_io ->set_req(_slow_path, i_o());
3714 result_mem->set_req(_slow_path, reset_memory());
3715 }
3716
3717 set_control(normal_ctl);
3718 if (!stopped()) {
3719 // Normal case: The array type has been cached in the java.lang.Class.
3720 // The following call works fine even if the array type is polymorphic.
3721 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3722 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3723 result_reg->init_req(_normal_path, control());
3724 result_val->init_req(_normal_path, obj);
3725 result_io ->init_req(_normal_path, i_o());
3726 result_mem->init_req(_normal_path, reset_memory());
3727 }
3728
3729 // Return the combined state.
3730 set_i_o( _gvn.transform(result_io) );
3731 set_all_memory( _gvn.transform(result_mem));
3732
3733 C->set_has_split_ifs(true); // Has chance for split-if optimization
3734 set_result(result_reg, result_val);
3735 return true;
3736 }
3737
3738 //----------------------inline_native_getLength--------------------------
3739 // public static native int java.lang.reflect.Array.getLength(Object array);
3740 bool LibraryCallKit::inline_native_getLength() {
3741 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3742
3743 Node* array = null_check(argument(0));
3744 // If array is dead, only null-path is taken.
3745 if (stopped()) return true;
3746
3747 // Deoptimize if it is a non-array.
3748 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3749
3750 if (non_array != NULL) {
3751 PreserveJVMState pjvms(this);
3752 set_control(non_array);
3753 uncommon_trap(Deoptimization::Reason_intrinsic,
3754 Deoptimization::Action_maybe_recompile);
3755 }
3756
3757 // If control is dead, only non-array-path is taken.
3758 if (stopped()) return true;
3759
3760 // The works fine even if the array type is polymorphic.
3761 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3762 Node* result = load_array_length(array);
3763
3764 C->set_has_split_ifs(true); // Has chance for split-if optimization
3765 set_result(result);
3766 return true;
3767 }
3768
3769 //------------------------inline_array_copyOf----------------------------
3770 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3771 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3772 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3773 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3774
3775 // Get the arguments.
3776 Node* original = argument(0);
3777 Node* start = is_copyOfRange? argument(1): intcon(0);
3778 Node* end = is_copyOfRange? argument(2): argument(1);
3779 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3780
3781 Node* newcopy = NULL;
3782
3783 // Set the original stack and the reexecute bit for the interpreter to reexecute
3784 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3785 { PreserveReexecuteState preexecs(this);
3786 jvms()->set_should_reexecute(true);
3787
3788 array_type_mirror = null_check(array_type_mirror);
3789 original = null_check(original);
3790
3791 // Check if a null path was taken unconditionally.
3792 if (stopped()) return true;
3793
3794 Node* orig_length = load_array_length(original);
3795
3796 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3797 klass_node = null_check(klass_node);
3798
3799 RegionNode* bailout = new RegionNode(1);
3800 record_for_igvn(bailout);
3801
3802 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3803 // Bail out if that is so.
3804 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3805 if (not_objArray != NULL) {
3806 // Improve the klass node's type from the new optimistic assumption:
3807 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3808 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3809 Node* cast = new CastPPNode(klass_node, akls);
3810 cast->init_req(0, control());
3811 klass_node = _gvn.transform(cast);
3812 }
3813
3814 // Bail out if either start or end is negative.
3815 generate_negative_guard(start, bailout, &start);
3816 generate_negative_guard(end, bailout, &end);
3817
3818 Node* length = end;
3819 if (_gvn.type(start) != TypeInt::ZERO) {
3820 length = _gvn.transform(new SubINode(end, start));
3821 }
3822
3823 // Bail out if length is negative.
3824 // Without this the new_array would throw
3825 // NegativeArraySizeException but IllegalArgumentException is what
3826 // should be thrown
3827 generate_negative_guard(length, bailout, &length);
3828
3829 if (bailout->req() > 1) {
3830 PreserveJVMState pjvms(this);
3831 set_control(_gvn.transform(bailout));
3832 uncommon_trap(Deoptimization::Reason_intrinsic,
3833 Deoptimization::Action_maybe_recompile);
3834 }
3835
3836 if (!stopped()) {
3837 // How many elements will we copy from the original?
3838 // The answer is MinI(orig_length - start, length).
3839 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3840 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3841
3842 // Generate a direct call to the right arraycopy function(s).
3843 // We know the copy is disjoint but we might not know if the
3844 // oop stores need checking.
3845 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3846 // This will fail a store-check if x contains any non-nulls.
3847
3848 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3849 // loads/stores but it is legal only if we're sure the
3850 // Arrays.copyOf would succeed. So we need all input arguments
3851 // to the copyOf to be validated, including that the copy to the
3852 // new array won't trigger an ArrayStoreException. That subtype
3853 // check can be optimized if we know something on the type of
3854 // the input array from type speculation.
3855 if (_gvn.type(klass_node)->singleton()) {
3856 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3857 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3858
3859 int test = C->static_subtype_check(superk, subk);
3860 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3861 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3862 if (t_original->speculative_type() != NULL) {
3863 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3864 }
3865 }
3866 }
3867
3868 bool validated = false;
3869 // Reason_class_check rather than Reason_intrinsic because we
3870 // want to intrinsify even if this traps.
3871 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3872 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3873 klass_node);
3874
3875 if (not_subtype_ctrl != top()) {
3876 PreserveJVMState pjvms(this);
3877 set_control(not_subtype_ctrl);
3878 uncommon_trap(Deoptimization::Reason_class_check,
3879 Deoptimization::Action_make_not_entrant);
3880 assert(stopped(), "Should be stopped");
3881 }
3882 validated = true;
3883 }
3884
3885 if (!stopped()) {
3886 newcopy = new_array(klass_node, length, 0); // no arguments to push
3887
3888 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3889 load_object_klass(original), klass_node);
3890 if (!is_copyOfRange) {
3891 ac->set_copyof(validated);
3892 } else {
3893 ac->set_copyofrange(validated);
3894 }
3895 Node* n = _gvn.transform(ac);
3896 if (n == ac) {
3897 ac->connect_outputs(this);
3898 } else {
3899 assert(validated, "shouldn't transform if all arguments not validated");
3900 set_all_memory(n);
3901 }
3902 }
3903 }
3904 } // original reexecute is set back here
3905
3906 C->set_has_split_ifs(true); // Has chance for split-if optimization
3907 if (!stopped()) {
3908 set_result(newcopy);
3909 }
3910 return true;
3911 }
3912
3913
3914 //----------------------generate_virtual_guard---------------------------
3915 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3916 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3917 RegionNode* slow_region) {
3918 ciMethod* method = callee();
3919 int vtable_index = method->vtable_index();
3920 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3921 "bad index %d", vtable_index);
3922 // Get the Method* out of the appropriate vtable entry.
3923 int entry_offset = (InstanceKlass::vtable_start_offset() +
3924 vtable_index*vtableEntry::size()) * wordSize +
3925 vtableEntry::method_offset_in_bytes();
3926 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3927 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3928
3929 // Compare the target method with the expected method (e.g., Object.hashCode).
3930 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3931
3932 Node* native_call = makecon(native_call_addr);
3933 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3934 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3935
3936 return generate_slow_guard(test_native, slow_region);
3937 }
3938
3939 //-----------------------generate_method_call----------------------------
3940 // Use generate_method_call to make a slow-call to the real
3941 // method if the fast path fails. An alternative would be to
3942 // use a stub like OptoRuntime::slow_arraycopy_Java.
3943 // This only works for expanding the current library call,
3944 // not another intrinsic. (E.g., don't use this for making an
3945 // arraycopy call inside of the copyOf intrinsic.)
3946 CallJavaNode*
3947 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3948 // When compiling the intrinsic method itself, do not use this technique.
3949 guarantee(callee() != C->method(), "cannot make slow-call to self");
3950
3951 ciMethod* method = callee();
3952 // ensure the JVMS we have will be correct for this call
3953 guarantee(method_id == method->intrinsic_id(), "must match");
3954
3955 const TypeFunc* tf = TypeFunc::make(method);
3956 CallJavaNode* slow_call;
3957 if (is_static) {
3958 assert(!is_virtual, "");
3959 slow_call = new CallStaticJavaNode(C, tf,
3960 SharedRuntime::get_resolve_static_call_stub(),
3961 method, bci());
3962 } else if (is_virtual) {
3963 null_check_receiver();
3964 int vtable_index = Method::invalid_vtable_index;
3965 if (UseInlineCaches) {
3966 // Suppress the vtable call
3967 } else {
3968 // hashCode and clone are not a miranda methods,
3969 // so the vtable index is fixed.
3970 // No need to use the linkResolver to get it.
3971 vtable_index = method->vtable_index();
3972 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3973 "bad index %d", vtable_index);
3974 }
3975 slow_call = new CallDynamicJavaNode(tf,
3976 SharedRuntime::get_resolve_virtual_call_stub(),
3977 method, vtable_index, bci());
3978 } else { // neither virtual nor static: opt_virtual
3979 null_check_receiver();
3980 slow_call = new CallStaticJavaNode(C, tf,
3981 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3982 method, bci());
3983 slow_call->set_optimized_virtual(true);
3984 }
3985 set_arguments_for_java_call(slow_call);
3986 set_edges_for_java_call(slow_call);
3987 return slow_call;
3988 }
3989
3990
3991 /**
3992 * Build special case code for calls to hashCode on an object. This call may
3993 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3994 * slightly different code.
3995 */
3996 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3997 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3998 assert(!(is_virtual && is_static), "either virtual, special, or static");
3999
4000 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4001
4002 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4003 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4004 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4005 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4006 Node* obj = NULL;
4007 if (!is_static) {
4008 // Check for hashing null object
4009 obj = null_check_receiver();
4010 if (stopped()) return true; // unconditionally null
4011 result_reg->init_req(_null_path, top());
4012 result_val->init_req(_null_path, top());
4013 } else {
4014 // Do a null check, and return zero if null.
4015 // System.identityHashCode(null) == 0
4016 obj = argument(0);
4017 Node* null_ctl = top();
4018 obj = null_check_oop(obj, &null_ctl);
4019 result_reg->init_req(_null_path, null_ctl);
4020 result_val->init_req(_null_path, _gvn.intcon(0));
4021 }
4022
4023 // Unconditionally null? Then return right away.
4024 if (stopped()) {
4025 set_control( result_reg->in(_null_path));
4026 if (!stopped())
4027 set_result(result_val->in(_null_path));
4028 return true;
4029 }
4030
4031 // We only go to the fast case code if we pass a number of guards. The
4032 // paths which do not pass are accumulated in the slow_region.
4033 RegionNode* slow_region = new RegionNode(1);
4034 record_for_igvn(slow_region);
4035
4036 // If this is a virtual call, we generate a funny guard. We pull out
4037 // the vtable entry corresponding to hashCode() from the target object.
4038 // If the target method which we are calling happens to be the native
4039 // Object hashCode() method, we pass the guard. We do not need this
4040 // guard for non-virtual calls -- the caller is known to be the native
4041 // Object hashCode().
4042 if (is_virtual) {
4043 // After null check, get the object's klass.
4044 Node* obj_klass = load_object_klass(obj);
4045 generate_virtual_guard(obj_klass, slow_region);
4046 }
4047
4048 // Get the header out of the object, use LoadMarkNode when available
4049 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4050 // The control of the load must be NULL. Otherwise, the load can move before
4051 // the null check after castPP removal.
4052 Node* no_ctrl = NULL;
4053 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4054
4055 // Test the header to see if it is unlocked.
4056 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4057 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4058 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4059 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4060 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4061
4062 generate_slow_guard(test_unlocked, slow_region);
4063
4064 // Get the hash value and check to see that it has been properly assigned.
4065 // We depend on hash_mask being at most 32 bits and avoid the use of
4066 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4067 // vm: see markOop.hpp.
4068 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4069 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4070 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4071 // This hack lets the hash bits live anywhere in the mark object now, as long
4072 // as the shift drops the relevant bits into the low 32 bits. Note that
4073 // Java spec says that HashCode is an int so there's no point in capturing
4074 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4075 hshifted_header = ConvX2I(hshifted_header);
4076 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4077
4078 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4079 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4080 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4081
4082 generate_slow_guard(test_assigned, slow_region);
4083
4084 Node* init_mem = reset_memory();
4085 // fill in the rest of the null path:
4086 result_io ->init_req(_null_path, i_o());
4087 result_mem->init_req(_null_path, init_mem);
4088
4089 result_val->init_req(_fast_path, hash_val);
4090 result_reg->init_req(_fast_path, control());
4091 result_io ->init_req(_fast_path, i_o());
4092 result_mem->init_req(_fast_path, init_mem);
4093
4094 // Generate code for the slow case. We make a call to hashCode().
4095 set_control(_gvn.transform(slow_region));
4096 if (!stopped()) {
4097 // No need for PreserveJVMState, because we're using up the present state.
4098 set_all_memory(init_mem);
4099 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4100 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4101 Node* slow_result = set_results_for_java_call(slow_call);
4102 // this->control() comes from set_results_for_java_call
4103 result_reg->init_req(_slow_path, control());
4104 result_val->init_req(_slow_path, slow_result);
4105 result_io ->set_req(_slow_path, i_o());
4106 result_mem ->set_req(_slow_path, reset_memory());
4107 }
4108
4109 // Return the combined state.
4110 set_i_o( _gvn.transform(result_io) );
4111 set_all_memory( _gvn.transform(result_mem));
4112
4113 set_result(result_reg, result_val);
4114 return true;
4115 }
4116
4117 //---------------------------inline_native_getClass----------------------------
4118 // public final native Class<?> java.lang.Object.getClass();
4119 //
4120 // Build special case code for calls to getClass on an object.
4121 bool LibraryCallKit::inline_native_getClass() {
4122 Node* obj = null_check_receiver();
4123 if (stopped()) return true;
4124 set_result(load_mirror_from_klass(load_object_klass(obj)));
4125 return true;
4126 }
4127
4128 //-----------------inline_native_Reflection_getCallerClass---------------------
4129 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4130 //
4131 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4132 //
4133 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4134 // in that it must skip particular security frames and checks for
4135 // caller sensitive methods.
4136 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4137 #ifndef PRODUCT
4138 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4139 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4140 }
4141 #endif
4142
4143 if (!jvms()->has_method()) {
4144 #ifndef PRODUCT
4145 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4146 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4147 }
4148 #endif
4149 return false;
4150 }
4151
4152 // Walk back up the JVM state to find the caller at the required
4153 // depth.
4154 JVMState* caller_jvms = jvms();
4155
4156 // Cf. JVM_GetCallerClass
4157 // NOTE: Start the loop at depth 1 because the current JVM state does
4158 // not include the Reflection.getCallerClass() frame.
4159 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4160 ciMethod* m = caller_jvms->method();
4161 switch (n) {
4162 case 0:
4163 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4164 break;
4165 case 1:
4166 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4167 if (!m->caller_sensitive()) {
4168 #ifndef PRODUCT
4169 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4170 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4171 }
4172 #endif
4173 return false; // bail-out; let JVM_GetCallerClass do the work
4174 }
4175 break;
4176 default:
4177 if (!m->is_ignored_by_security_stack_walk()) {
4178 // We have reached the desired frame; return the holder class.
4179 // Acquire method holder as java.lang.Class and push as constant.
4180 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4181 ciInstance* caller_mirror = caller_klass->java_mirror();
4182 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4183
4184 #ifndef PRODUCT
4185 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4186 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());
4187 tty->print_cr(" JVM state at this point:");
4188 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4189 ciMethod* m = jvms()->of_depth(i)->method();
4190 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4191 }
4192 }
4193 #endif
4194 return true;
4195 }
4196 break;
4197 }
4198 }
4199
4200 #ifndef PRODUCT
4201 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4202 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4203 tty->print_cr(" JVM state at this point:");
4204 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4205 ciMethod* m = jvms()->of_depth(i)->method();
4206 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4207 }
4208 }
4209 #endif
4210
4211 return false; // bail-out; let JVM_GetCallerClass do the work
4212 }
4213
4214 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4215 Node* arg = argument(0);
4216 Node* result = NULL;
4217
4218 switch (id) {
4219 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4220 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4221 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4222 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4223
4224 case vmIntrinsics::_doubleToLongBits: {
4225 // two paths (plus control) merge in a wood
4226 RegionNode *r = new RegionNode(3);
4227 Node *phi = new PhiNode(r, TypeLong::LONG);
4228
4229 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4230 // Build the boolean node
4231 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4232
4233 // Branch either way.
4234 // NaN case is less traveled, which makes all the difference.
4235 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4236 Node *opt_isnan = _gvn.transform(ifisnan);
4237 assert( opt_isnan->is_If(), "Expect an IfNode");
4238 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4239 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4240
4241 set_control(iftrue);
4242
4243 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4244 Node *slow_result = longcon(nan_bits); // return NaN
4245 phi->init_req(1, _gvn.transform( slow_result ));
4246 r->init_req(1, iftrue);
4247
4248 // Else fall through
4249 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4250 set_control(iffalse);
4251
4252 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4253 r->init_req(2, iffalse);
4254
4255 // Post merge
4256 set_control(_gvn.transform(r));
4257 record_for_igvn(r);
4258
4259 C->set_has_split_ifs(true); // Has chance for split-if optimization
4260 result = phi;
4261 assert(result->bottom_type()->isa_long(), "must be");
4262 break;
4263 }
4264
4265 case vmIntrinsics::_floatToIntBits: {
4266 // two paths (plus control) merge in a wood
4267 RegionNode *r = new RegionNode(3);
4268 Node *phi = new PhiNode(r, TypeInt::INT);
4269
4270 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4271 // Build the boolean node
4272 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4273
4274 // Branch either way.
4275 // NaN case is less traveled, which makes all the difference.
4276 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4277 Node *opt_isnan = _gvn.transform(ifisnan);
4278 assert( opt_isnan->is_If(), "Expect an IfNode");
4279 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4280 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4281
4282 set_control(iftrue);
4283
4284 static const jint nan_bits = 0x7fc00000;
4285 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4286 phi->init_req(1, _gvn.transform( slow_result ));
4287 r->init_req(1, iftrue);
4288
4289 // Else fall through
4290 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4291 set_control(iffalse);
4292
4293 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4294 r->init_req(2, iffalse);
4295
4296 // Post merge
4297 set_control(_gvn.transform(r));
4298 record_for_igvn(r);
4299
4300 C->set_has_split_ifs(true); // Has chance for split-if optimization
4301 result = phi;
4302 assert(result->bottom_type()->isa_int(), "must be");
4303 break;
4304 }
4305
4306 default:
4307 fatal_unexpected_iid(id);
4308 break;
4309 }
4310 set_result(_gvn.transform(result));
4311 return true;
4312 }
4313
4314 //----------------------inline_unsafe_copyMemory-------------------------
4315 // public native void Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4316 bool LibraryCallKit::inline_unsafe_copyMemory() {
4317 if (callee()->is_static()) return false; // caller must have the capability!
4318 null_check_receiver(); // null-check receiver
4319 if (stopped()) return true;
4320
4321 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4322
4323 Node* src_ptr = argument(1); // type: oop
4324 Node* src_off = ConvL2X(argument(2)); // type: long
4325 Node* dst_ptr = argument(4); // type: oop
4326 Node* dst_off = ConvL2X(argument(5)); // type: long
4327 Node* size = ConvL2X(argument(7)); // type: long
4328
4329 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4330 "fieldOffset must be byte-scaled");
4331
4332 Node* src = make_unsafe_address(src_ptr, src_off);
4333 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4334
4335 // Conservatively insert a memory barrier on all memory slices.
4336 // Do not let writes of the copy source or destination float below the copy.
4337 insert_mem_bar(Op_MemBarCPUOrder);
4338
4339 // Call it. Note that the length argument is not scaled.
4340 make_runtime_call(RC_LEAF|RC_NO_FP,
4341 OptoRuntime::fast_arraycopy_Type(),
4342 StubRoutines::unsafe_arraycopy(),
4343 "unsafe_arraycopy",
4344 TypeRawPtr::BOTTOM,
4345 src, dst, size XTOP);
4346
4347 // Do not let reads of the copy destination float above the copy.
4348 insert_mem_bar(Op_MemBarCPUOrder);
4349
4350 return true;
4351 }
4352
4353 //------------------------clone_coping-----------------------------------
4354 // Helper function for inline_native_clone.
4355 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4356 assert(obj_size != NULL, "");
4357 Node* raw_obj = alloc_obj->in(1);
4358 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4359
4360 AllocateNode* alloc = NULL;
4361 if (ReduceBulkZeroing) {
4362 // We will be completely responsible for initializing this object -
4363 // mark Initialize node as complete.
4364 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4365 // The object was just allocated - there should be no any stores!
4366 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4367 // Mark as complete_with_arraycopy so that on AllocateNode
4368 // expansion, we know this AllocateNode is initialized by an array
4369 // copy and a StoreStore barrier exists after the array copy.
4370 alloc->initialization()->set_complete_with_arraycopy();
4371 }
4372
4373 // Copy the fastest available way.
4374 // TODO: generate fields copies for small objects instead.
4375 Node* src = obj;
4376 Node* dest = alloc_obj;
4377 Node* size = _gvn.transform(obj_size);
4378
4379 // Exclude the header but include array length to copy by 8 bytes words.
4380 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4381 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4382 instanceOopDesc::base_offset_in_bytes();
4383 // base_off:
4384 // 8 - 32-bit VM
4385 // 12 - 64-bit VM, compressed klass
4386 // 16 - 64-bit VM, normal klass
4387 if (base_off % BytesPerLong != 0) {
4388 assert(UseCompressedClassPointers, "");
4389 if (is_array) {
4390 // Exclude length to copy by 8 bytes words.
4391 base_off += sizeof(int);
4392 } else {
4393 // Include klass to copy by 8 bytes words.
4394 base_off = instanceOopDesc::klass_offset_in_bytes();
4395 }
4396 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4397 }
4398 src = basic_plus_adr(src, base_off);
4399 dest = basic_plus_adr(dest, base_off);
4400
4401 // Compute the length also, if needed:
4402 Node* countx = size;
4403 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4404 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4405
4406 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4407
4408 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4409 ac->set_clonebasic();
4410 Node* n = _gvn.transform(ac);
4411 if (n == ac) {
4412 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4413 } else {
4414 set_all_memory(n);
4415 }
4416
4417 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4418 if (card_mark) {
4419 assert(!is_array, "");
4420 // Put in store barrier for any and all oops we are sticking
4421 // into this object. (We could avoid this if we could prove
4422 // that the object type contains no oop fields at all.)
4423 Node* no_particular_value = NULL;
4424 Node* no_particular_field = NULL;
4425 int raw_adr_idx = Compile::AliasIdxRaw;
4426 post_barrier(control(),
4427 memory(raw_adr_type),
4428 alloc_obj,
4429 no_particular_field,
4430 raw_adr_idx,
4431 no_particular_value,
4432 T_OBJECT,
4433 false);
4434 }
4435
4436 // Do not let reads from the cloned object float above the arraycopy.
4437 if (alloc != NULL) {
4438 // Do not let stores that initialize this object be reordered with
4439 // a subsequent store that would make this object accessible by
4440 // other threads.
4441 // Record what AllocateNode this StoreStore protects so that
4442 // escape analysis can go from the MemBarStoreStoreNode to the
4443 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4444 // based on the escape status of the AllocateNode.
4445 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4446 } else {
4447 insert_mem_bar(Op_MemBarCPUOrder);
4448 }
4449 }
4450
4451 //------------------------inline_native_clone----------------------------
4452 // protected native Object java.lang.Object.clone();
4453 //
4454 // Here are the simple edge cases:
4455 // null receiver => normal trap
4456 // virtual and clone was overridden => slow path to out-of-line clone
4457 // not cloneable or finalizer => slow path to out-of-line Object.clone
4458 //
4459 // The general case has two steps, allocation and copying.
4460 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4461 //
4462 // Copying also has two cases, oop arrays and everything else.
4463 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4464 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4465 //
4466 // These steps fold up nicely if and when the cloned object's klass
4467 // can be sharply typed as an object array, a type array, or an instance.
4468 //
4469 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4470 PhiNode* result_val;
4471
4472 // Set the reexecute bit for the interpreter to reexecute
4473 // the bytecode that invokes Object.clone if deoptimization happens.
4474 { PreserveReexecuteState preexecs(this);
4475 jvms()->set_should_reexecute(true);
4476
4477 Node* obj = null_check_receiver();
4478 if (stopped()) return true;
4479
4480 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4481
4482 // If we are going to clone an instance, we need its exact type to
4483 // know the number and types of fields to convert the clone to
4484 // loads/stores. Maybe a speculative type can help us.
4485 if (!obj_type->klass_is_exact() &&
4486 obj_type->speculative_type() != NULL &&
4487 obj_type->speculative_type()->is_instance_klass()) {
4488 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4489 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4490 !spec_ik->has_injected_fields()) {
4491 ciKlass* k = obj_type->klass();
4492 if (!k->is_instance_klass() ||
4493 k->as_instance_klass()->is_interface() ||
4494 k->as_instance_klass()->has_subklass()) {
4495 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4496 }
4497 }
4498 }
4499
4500 Node* obj_klass = load_object_klass(obj);
4501 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4502 const TypeOopPtr* toop = ((tklass != NULL)
4503 ? tklass->as_instance_type()
4504 : TypeInstPtr::NOTNULL);
4505
4506 // Conservatively insert a memory barrier on all memory slices.
4507 // Do not let writes into the original float below the clone.
4508 insert_mem_bar(Op_MemBarCPUOrder);
4509
4510 // paths into result_reg:
4511 enum {
4512 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4513 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4514 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4515 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4516 PATH_LIMIT
4517 };
4518 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4519 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4520 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4521 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4522 record_for_igvn(result_reg);
4523
4524 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4525 int raw_adr_idx = Compile::AliasIdxRaw;
4526
4527 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4528 if (array_ctl != NULL) {
4529 // It's an array.
4530 PreserveJVMState pjvms(this);
4531 set_control(array_ctl);
4532 Node* obj_length = load_array_length(obj);
4533 Node* obj_size = NULL;
4534 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4535
4536 if (!use_ReduceInitialCardMarks()) {
4537 // If it is an oop array, it requires very special treatment,
4538 // because card marking is required on each card of the array.
4539 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4540 if (is_obja != NULL) {
4541 PreserveJVMState pjvms2(this);
4542 set_control(is_obja);
4543 // Generate a direct call to the right arraycopy function(s).
4544 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4545 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4546 ac->set_cloneoop();
4547 Node* n = _gvn.transform(ac);
4548 assert(n == ac, "cannot disappear");
4549 ac->connect_outputs(this);
4550
4551 result_reg->init_req(_objArray_path, control());
4552 result_val->init_req(_objArray_path, alloc_obj);
4553 result_i_o ->set_req(_objArray_path, i_o());
4554 result_mem ->set_req(_objArray_path, reset_memory());
4555 }
4556 }
4557 // Otherwise, there are no card marks to worry about.
4558 // (We can dispense with card marks if we know the allocation
4559 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4560 // causes the non-eden paths to take compensating steps to
4561 // simulate a fresh allocation, so that no further
4562 // card marks are required in compiled code to initialize
4563 // the object.)
4564
4565 if (!stopped()) {
4566 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4567
4568 // Present the results of the copy.
4569 result_reg->init_req(_array_path, control());
4570 result_val->init_req(_array_path, alloc_obj);
4571 result_i_o ->set_req(_array_path, i_o());
4572 result_mem ->set_req(_array_path, reset_memory());
4573 }
4574 }
4575
4576 // We only go to the instance fast case code if we pass a number of guards.
4577 // The paths which do not pass are accumulated in the slow_region.
4578 RegionNode* slow_region = new RegionNode(1);
4579 record_for_igvn(slow_region);
4580 if (!stopped()) {
4581 // It's an instance (we did array above). Make the slow-path tests.
4582 // If this is a virtual call, we generate a funny guard. We grab
4583 // the vtable entry corresponding to clone() from the target object.
4584 // If the target method which we are calling happens to be the
4585 // Object clone() method, we pass the guard. We do not need this
4586 // guard for non-virtual calls; the caller is known to be the native
4587 // Object clone().
4588 if (is_virtual) {
4589 generate_virtual_guard(obj_klass, slow_region);
4590 }
4591
4592 // The object must be cloneable and must not have a finalizer.
4593 // Both of these conditions may be checked in a single test.
4594 // We could optimize the cloneable test further, but we don't care.
4595 generate_access_flags_guard(obj_klass,
4596 // Test both conditions:
4597 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4598 // Must be cloneable but not finalizer:
4599 JVM_ACC_IS_CLONEABLE,
4600 slow_region);
4601 }
4602
4603 if (!stopped()) {
4604 // It's an instance, and it passed the slow-path tests.
4605 PreserveJVMState pjvms(this);
4606 Node* obj_size = NULL;
4607 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4608 // is reexecuted if deoptimization occurs and there could be problems when merging
4609 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4610 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4611
4612 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4613
4614 // Present the results of the slow call.
4615 result_reg->init_req(_instance_path, control());
4616 result_val->init_req(_instance_path, alloc_obj);
4617 result_i_o ->set_req(_instance_path, i_o());
4618 result_mem ->set_req(_instance_path, reset_memory());
4619 }
4620
4621 // Generate code for the slow case. We make a call to clone().
4622 set_control(_gvn.transform(slow_region));
4623 if (!stopped()) {
4624 PreserveJVMState pjvms(this);
4625 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4626 Node* slow_result = set_results_for_java_call(slow_call);
4627 // this->control() comes from set_results_for_java_call
4628 result_reg->init_req(_slow_path, control());
4629 result_val->init_req(_slow_path, slow_result);
4630 result_i_o ->set_req(_slow_path, i_o());
4631 result_mem ->set_req(_slow_path, reset_memory());
4632 }
4633
4634 // Return the combined state.
4635 set_control( _gvn.transform(result_reg));
4636 set_i_o( _gvn.transform(result_i_o));
4637 set_all_memory( _gvn.transform(result_mem));
4638 } // original reexecute is set back here
4639
4640 set_result(_gvn.transform(result_val));
4641 return true;
4642 }
4643
4644 // If we have a tighly coupled allocation, the arraycopy may take care
4645 // of the array initialization. If one of the guards we insert between
4646 // the allocation and the arraycopy causes a deoptimization, an
4647 // unitialized array will escape the compiled method. To prevent that
4648 // we set the JVM state for uncommon traps between the allocation and
4649 // the arraycopy to the state before the allocation so, in case of
4650 // deoptimization, we'll reexecute the allocation and the
4651 // initialization.
4652 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4653 if (alloc != NULL) {
4654 ciMethod* trap_method = alloc->jvms()->method();
4655 int trap_bci = alloc->jvms()->bci();
4656
4657 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4658 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4659 // Make sure there's no store between the allocation and the
4660 // arraycopy otherwise visible side effects could be rexecuted
4661 // in case of deoptimization and cause incorrect execution.
4662 bool no_interfering_store = true;
4663 Node* mem = alloc->in(TypeFunc::Memory);
4664 if (mem->is_MergeMem()) {
4665 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4666 Node* n = mms.memory();
4667 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4668 assert(n->is_Store(), "what else?");
4669 no_interfering_store = false;
4670 break;
4671 }
4672 }
4673 } else {
4674 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4675 Node* n = mms.memory();
4676 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4677 assert(n->is_Store(), "what else?");
4678 no_interfering_store = false;
4679 break;
4680 }
4681 }
4682 }
4683
4684 if (no_interfering_store) {
4685 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4686 uint size = alloc->req();
4687 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4688 old_jvms->set_map(sfpt);
4689 for (uint i = 0; i < size; i++) {
4690 sfpt->init_req(i, alloc->in(i));
4691 }
4692 // re-push array length for deoptimization
4693 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4694 old_jvms->set_sp(old_jvms->sp()+1);
4695 old_jvms->set_monoff(old_jvms->monoff()+1);
4696 old_jvms->set_scloff(old_jvms->scloff()+1);
4697 old_jvms->set_endoff(old_jvms->endoff()+1);
4698 old_jvms->set_should_reexecute(true);
4699
4700 sfpt->set_i_o(map()->i_o());
4701 sfpt->set_memory(map()->memory());
4702 sfpt->set_control(map()->control());
4703
4704 JVMState* saved_jvms = jvms();
4705 saved_reexecute_sp = _reexecute_sp;
4706
4707 set_jvms(sfpt->jvms());
4708 _reexecute_sp = jvms()->sp();
4709
4710 return saved_jvms;
4711 }
4712 }
4713 }
4714 return NULL;
4715 }
4716
4717 // In case of a deoptimization, we restart execution at the
4718 // allocation, allocating a new array. We would leave an uninitialized
4719 // array in the heap that GCs wouldn't expect. Move the allocation
4720 // after the traps so we don't allocate the array if we
4721 // deoptimize. This is possible because tightly_coupled_allocation()
4722 // guarantees there's no observer of the allocated array at this point
4723 // and the control flow is simple enough.
4724 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4725 if (saved_jvms != NULL && !stopped()) {
4726 assert(alloc != NULL, "only with a tightly coupled allocation");
4727 // restore JVM state to the state at the arraycopy
4728 saved_jvms->map()->set_control(map()->control());
4729 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4730 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4731 // If we've improved the types of some nodes (null check) while
4732 // emitting the guards, propagate them to the current state
4733 map()->replaced_nodes().apply(saved_jvms->map());
4734 set_jvms(saved_jvms);
4735 _reexecute_sp = saved_reexecute_sp;
4736
4737 // Remove the allocation from above the guards
4738 CallProjections callprojs;
4739 alloc->extract_projections(&callprojs, true);
4740 InitializeNode* init = alloc->initialization();
4741 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4742 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4743 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4744 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4745
4746 // move the allocation here (after the guards)
4747 _gvn.hash_delete(alloc);
4748 alloc->set_req(TypeFunc::Control, control());
4749 alloc->set_req(TypeFunc::I_O, i_o());
4750 Node *mem = reset_memory();
4751 set_all_memory(mem);
4752 alloc->set_req(TypeFunc::Memory, mem);
4753 set_control(init->proj_out(TypeFunc::Control));
4754 set_i_o(callprojs.fallthrough_ioproj);
4755
4756 // Update memory as done in GraphKit::set_output_for_allocation()
4757 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4758 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4759 if (ary_type->isa_aryptr() && length_type != NULL) {
4760 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4761 }
4762 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4763 int elemidx = C->get_alias_index(telemref);
4764 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4765 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4766
4767 Node* allocx = _gvn.transform(alloc);
4768 assert(allocx == alloc, "where has the allocation gone?");
4769 assert(dest->is_CheckCastPP(), "not an allocation result?");
4770
4771 _gvn.hash_delete(dest);
4772 dest->set_req(0, control());
4773 Node* destx = _gvn.transform(dest);
4774 assert(destx == dest, "where has the allocation result gone?");
4775 }
4776 }
4777
4778
4779 //------------------------------inline_arraycopy-----------------------
4780 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4781 // Object dest, int destPos,
4782 // int length);
4783 bool LibraryCallKit::inline_arraycopy() {
4784 // Get the arguments.
4785 Node* src = argument(0); // type: oop
4786 Node* src_offset = argument(1); // type: int
4787 Node* dest = argument(2); // type: oop
4788 Node* dest_offset = argument(3); // type: int
4789 Node* length = argument(4); // type: int
4790
4791
4792 // Check for allocation before we add nodes that would confuse
4793 // tightly_coupled_allocation()
4794 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4795
4796 int saved_reexecute_sp = -1;
4797 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4798 // See arraycopy_restore_alloc_state() comment
4799 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4800 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4801 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4802 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4803
4804 // The following tests must be performed
4805 // (1) src and dest are arrays.
4806 // (2) src and dest arrays must have elements of the same BasicType
4807 // (3) src and dest must not be null.
4808 // (4) src_offset must not be negative.
4809 // (5) dest_offset must not be negative.
4810 // (6) length must not be negative.
4811 // (7) src_offset + length must not exceed length of src.
4812 // (8) dest_offset + length must not exceed length of dest.
4813 // (9) each element of an oop array must be assignable
4814
4815 // (3) src and dest must not be null.
4816 // always do this here because we need the JVM state for uncommon traps
4817 Node* null_ctl = top();
4818 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4819 assert(null_ctl->is_top(), "no null control here");
4820 dest = null_check(dest, T_ARRAY);
4821
4822 if (!can_emit_guards) {
4823 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4824 // guards but the arraycopy node could still take advantage of a
4825 // tightly allocated allocation. tightly_coupled_allocation() is
4826 // called again to make sure it takes the null check above into
4827 // account: the null check is mandatory and if it caused an
4828 // uncommon trap to be emitted then the allocation can't be
4829 // considered tightly coupled in this context.
4830 alloc = tightly_coupled_allocation(dest, NULL);
4831 }
4832
4833 bool validated = false;
4834
4835 const Type* src_type = _gvn.type(src);
4836 const Type* dest_type = _gvn.type(dest);
4837 const TypeAryPtr* top_src = src_type->isa_aryptr();
4838 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4839
4840 // Do we have the type of src?
4841 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4842 // Do we have the type of dest?
4843 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4844 // Is the type for src from speculation?
4845 bool src_spec = false;
4846 // Is the type for dest from speculation?
4847 bool dest_spec = false;
4848
4849 if ((!has_src || !has_dest) && can_emit_guards) {
4850 // We don't have sufficient type information, let's see if
4851 // speculative types can help. We need to have types for both src
4852 // and dest so that it pays off.
4853
4854 // Do we already have or could we have type information for src
4855 bool could_have_src = has_src;
4856 // Do we already have or could we have type information for dest
4857 bool could_have_dest = has_dest;
4858
4859 ciKlass* src_k = NULL;
4860 if (!has_src) {
4861 src_k = src_type->speculative_type_not_null();
4862 if (src_k != NULL && src_k->is_array_klass()) {
4863 could_have_src = true;
4864 }
4865 }
4866
4867 ciKlass* dest_k = NULL;
4868 if (!has_dest) {
4869 dest_k = dest_type->speculative_type_not_null();
4870 if (dest_k != NULL && dest_k->is_array_klass()) {
4871 could_have_dest = true;
4872 }
4873 }
4874
4875 if (could_have_src && could_have_dest) {
4876 // This is going to pay off so emit the required guards
4877 if (!has_src) {
4878 src = maybe_cast_profiled_obj(src, src_k, true);
4879 src_type = _gvn.type(src);
4880 top_src = src_type->isa_aryptr();
4881 has_src = (top_src != NULL && top_src->klass() != NULL);
4882 src_spec = true;
4883 }
4884 if (!has_dest) {
4885 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4886 dest_type = _gvn.type(dest);
4887 top_dest = dest_type->isa_aryptr();
4888 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4889 dest_spec = true;
4890 }
4891 }
4892 }
4893
4894 if (has_src && has_dest && can_emit_guards) {
4895 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4896 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4897 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4898 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4899
4900 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4901 // If both arrays are object arrays then having the exact types
4902 // for both will remove the need for a subtype check at runtime
4903 // before the call and may make it possible to pick a faster copy
4904 // routine (without a subtype check on every element)
4905 // Do we have the exact type of src?
4906 bool could_have_src = src_spec;
4907 // Do we have the exact type of dest?
4908 bool could_have_dest = dest_spec;
4909 ciKlass* src_k = top_src->klass();
4910 ciKlass* dest_k = top_dest->klass();
4911 if (!src_spec) {
4912 src_k = src_type->speculative_type_not_null();
4913 if (src_k != NULL && src_k->is_array_klass()) {
4914 could_have_src = true;
4915 }
4916 }
4917 if (!dest_spec) {
4918 dest_k = dest_type->speculative_type_not_null();
4919 if (dest_k != NULL && dest_k->is_array_klass()) {
4920 could_have_dest = true;
4921 }
4922 }
4923 if (could_have_src && could_have_dest) {
4924 // If we can have both exact types, emit the missing guards
4925 if (could_have_src && !src_spec) {
4926 src = maybe_cast_profiled_obj(src, src_k, true);
4927 }
4928 if (could_have_dest && !dest_spec) {
4929 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4930 }
4931 }
4932 }
4933 }
4934
4935 ciMethod* trap_method = method();
4936 int trap_bci = bci();
4937 if (saved_jvms != NULL) {
4938 trap_method = alloc->jvms()->method();
4939 trap_bci = alloc->jvms()->bci();
4940 }
4941
4942 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4943 can_emit_guards &&
4944 !src->is_top() && !dest->is_top()) {
4945 // validate arguments: enables transformation the ArrayCopyNode
4946 validated = true;
4947
4948 RegionNode* slow_region = new RegionNode(1);
4949 record_for_igvn(slow_region);
4950
4951 // (1) src and dest are arrays.
4952 generate_non_array_guard(load_object_klass(src), slow_region);
4953 generate_non_array_guard(load_object_klass(dest), slow_region);
4954
4955 // (2) src and dest arrays must have elements of the same BasicType
4956 // done at macro expansion or at Ideal transformation time
4957
4958 // (4) src_offset must not be negative.
4959 generate_negative_guard(src_offset, slow_region);
4960
4961 // (5) dest_offset must not be negative.
4962 generate_negative_guard(dest_offset, slow_region);
4963
4964 // (7) src_offset + length must not exceed length of src.
4965 generate_limit_guard(src_offset, length,
4966 load_array_length(src),
4967 slow_region);
4968
4969 // (8) dest_offset + length must not exceed length of dest.
4970 generate_limit_guard(dest_offset, length,
4971 load_array_length(dest),
4972 slow_region);
4973
4974 // (9) each element of an oop array must be assignable
4975 Node* src_klass = load_object_klass(src);
4976 Node* dest_klass = load_object_klass(dest);
4977 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4978
4979 if (not_subtype_ctrl != top()) {
4980 PreserveJVMState pjvms(this);
4981 set_control(not_subtype_ctrl);
4982 uncommon_trap(Deoptimization::Reason_intrinsic,
4983 Deoptimization::Action_make_not_entrant);
4984 assert(stopped(), "Should be stopped");
4985 }
4986 {
4987 PreserveJVMState pjvms(this);
4988 set_control(_gvn.transform(slow_region));
4989 uncommon_trap(Deoptimization::Reason_intrinsic,
4990 Deoptimization::Action_make_not_entrant);
4991 assert(stopped(), "Should be stopped");
4992 }
4993 }
4994
4995 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4996
4997 if (stopped()) {
4998 return true;
4999 }
5000
5001 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5002 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5003 // so the compiler has a chance to eliminate them: during macro expansion,
5004 // we have to set their control (CastPP nodes are eliminated).
5005 load_object_klass(src), load_object_klass(dest),
5006 load_array_length(src), load_array_length(dest));
5007
5008 ac->set_arraycopy(validated);
5009
5010 Node* n = _gvn.transform(ac);
5011 if (n == ac) {
5012 ac->connect_outputs(this);
5013 } else {
5014 assert(validated, "shouldn't transform if all arguments not validated");
5015 set_all_memory(n);
5016 }
5017
5018 return true;
5019 }
5020
5021
5022 // Helper function which determines if an arraycopy immediately follows
5023 // an allocation, with no intervening tests or other escapes for the object.
5024 AllocateArrayNode*
5025 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5026 RegionNode* slow_region) {
5027 if (stopped()) return NULL; // no fast path
5028 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5029
5030 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5031 if (alloc == NULL) return NULL;
5032
5033 Node* rawmem = memory(Compile::AliasIdxRaw);
5034 // Is the allocation's memory state untouched?
5035 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5036 // Bail out if there have been raw-memory effects since the allocation.
5037 // (Example: There might have been a call or safepoint.)
5038 return NULL;
5039 }
5040 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5041 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5042 return NULL;
5043 }
5044
5045 // There must be no unexpected observers of this allocation.
5046 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5047 Node* obs = ptr->fast_out(i);
5048 if (obs != this->map()) {
5049 return NULL;
5050 }
5051 }
5052
5053 // This arraycopy must unconditionally follow the allocation of the ptr.
5054 Node* alloc_ctl = ptr->in(0);
5055 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5056
5057 Node* ctl = control();
5058 while (ctl != alloc_ctl) {
5059 // There may be guards which feed into the slow_region.
5060 // Any other control flow means that we might not get a chance
5061 // to finish initializing the allocated object.
5062 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5063 IfNode* iff = ctl->in(0)->as_If();
5064 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5065 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5066 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5067 ctl = iff->in(0); // This test feeds the known slow_region.
5068 continue;
5069 }
5070 // One more try: Various low-level checks bottom out in
5071 // uncommon traps. If the debug-info of the trap omits
5072 // any reference to the allocation, as we've already
5073 // observed, then there can be no objection to the trap.
5074 bool found_trap = false;
5075 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5076 Node* obs = not_ctl->fast_out(j);
5077 if (obs->in(0) == not_ctl && obs->is_Call() &&
5078 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5079 found_trap = true; break;
5080 }
5081 }
5082 if (found_trap) {
5083 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5084 continue;
5085 }
5086 }
5087 return NULL;
5088 }
5089
5090 // If we get this far, we have an allocation which immediately
5091 // precedes the arraycopy, and we can take over zeroing the new object.
5092 // The arraycopy will finish the initialization, and provide
5093 // a new control state to which we will anchor the destination pointer.
5094
5095 return alloc;
5096 }
5097
5098 //-------------inline_encodeISOArray-----------------------------------
5099 // encode char[] to byte[] in ISO_8859_1
5100 bool LibraryCallKit::inline_encodeISOArray() {
5101 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5102 // no receiver since it is static method
5103 Node *src = argument(0);
5104 Node *src_offset = argument(1);
5105 Node *dst = argument(2);
5106 Node *dst_offset = argument(3);
5107 Node *length = argument(4);
5108
5109 const Type* src_type = src->Value(&_gvn);
5110 const Type* dst_type = dst->Value(&_gvn);
5111 const TypeAryPtr* top_src = src_type->isa_aryptr();
5112 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5113 if (top_src == NULL || top_src->klass() == NULL ||
5114 top_dest == NULL || top_dest->klass() == NULL) {
5115 // failed array check
5116 return false;
5117 }
5118
5119 // Figure out the size and type of the elements we will be copying.
5120 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5121 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5122 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5123 return false;
5124 }
5125
5126 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5127 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5128 // 'src_start' points to src array + scaled offset
5129 // 'dst_start' points to dst array + scaled offset
5130
5131 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5132 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5133 enc = _gvn.transform(enc);
5134 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5135 set_memory(res_mem, mtype);
5136 set_result(enc);
5137 return true;
5138 }
5139
5140 //-------------inline_multiplyToLen-----------------------------------
5141 bool LibraryCallKit::inline_multiplyToLen() {
5142 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5143
5144 address stubAddr = StubRoutines::multiplyToLen();
5145 if (stubAddr == NULL) {
5146 return false; // Intrinsic's stub is not implemented on this platform
5147 }
5148 const char* stubName = "multiplyToLen";
5149
5150 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5151
5152 // no receiver because it is a static method
5153 Node* x = argument(0);
5154 Node* xlen = argument(1);
5155 Node* y = argument(2);
5156 Node* ylen = argument(3);
5157 Node* z = argument(4);
5158
5159 const Type* x_type = x->Value(&_gvn);
5160 const Type* y_type = y->Value(&_gvn);
5161 const TypeAryPtr* top_x = x_type->isa_aryptr();
5162 const TypeAryPtr* top_y = y_type->isa_aryptr();
5163 if (top_x == NULL || top_x->klass() == NULL ||
5164 top_y == NULL || top_y->klass() == NULL) {
5165 // failed array check
5166 return false;
5167 }
5168
5169 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5170 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5171 if (x_elem != T_INT || y_elem != T_INT) {
5172 return false;
5173 }
5174
5175 // Set the original stack and the reexecute bit for the interpreter to reexecute
5176 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5177 // on the return from z array allocation in runtime.
5178 { PreserveReexecuteState preexecs(this);
5179 jvms()->set_should_reexecute(true);
5180
5181 Node* x_start = array_element_address(x, intcon(0), x_elem);
5182 Node* y_start = array_element_address(y, intcon(0), y_elem);
5183 // 'x_start' points to x array + scaled xlen
5184 // 'y_start' points to y array + scaled ylen
5185
5186 // Allocate the result array
5187 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5188 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5189 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5190
5191 IdealKit ideal(this);
5192
5193 #define __ ideal.
5194 Node* one = __ ConI(1);
5195 Node* zero = __ ConI(0);
5196 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5197 __ set(need_alloc, zero);
5198 __ set(z_alloc, z);
5199 __ if_then(z, BoolTest::eq, null()); {
5200 __ increment (need_alloc, one);
5201 } __ else_(); {
5202 // Update graphKit memory and control from IdealKit.
5203 sync_kit(ideal);
5204 Node* zlen_arg = load_array_length(z);
5205 // Update IdealKit memory and control from graphKit.
5206 __ sync_kit(this);
5207 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5208 __ increment (need_alloc, one);
5209 } __ end_if();
5210 } __ end_if();
5211
5212 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5213 // Update graphKit memory and control from IdealKit.
5214 sync_kit(ideal);
5215 Node * narr = new_array(klass_node, zlen, 1);
5216 // Update IdealKit memory and control from graphKit.
5217 __ sync_kit(this);
5218 __ set(z_alloc, narr);
5219 } __ end_if();
5220
5221 sync_kit(ideal);
5222 z = __ value(z_alloc);
5223 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5224 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5225 // Final sync IdealKit and GraphKit.
5226 final_sync(ideal);
5227 #undef __
5228
5229 Node* z_start = array_element_address(z, intcon(0), T_INT);
5230
5231 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5232 OptoRuntime::multiplyToLen_Type(),
5233 stubAddr, stubName, TypePtr::BOTTOM,
5234 x_start, xlen, y_start, ylen, z_start, zlen);
5235 } // original reexecute is set back here
5236
5237 C->set_has_split_ifs(true); // Has chance for split-if optimization
5238 set_result(z);
5239 return true;
5240 }
5241
5242 //-------------inline_squareToLen------------------------------------
5243 bool LibraryCallKit::inline_squareToLen() {
5244 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5245
5246 address stubAddr = StubRoutines::squareToLen();
5247 if (stubAddr == NULL) {
5248 return false; // Intrinsic's stub is not implemented on this platform
5249 }
5250 const char* stubName = "squareToLen";
5251
5252 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5253
5254 Node* x = argument(0);
5255 Node* len = argument(1);
5256 Node* z = argument(2);
5257 Node* zlen = argument(3);
5258
5259 const Type* x_type = x->Value(&_gvn);
5260 const Type* z_type = z->Value(&_gvn);
5261 const TypeAryPtr* top_x = x_type->isa_aryptr();
5262 const TypeAryPtr* top_z = z_type->isa_aryptr();
5263 if (top_x == NULL || top_x->klass() == NULL ||
5264 top_z == NULL || top_z->klass() == NULL) {
5265 // failed array check
5266 return false;
5267 }
5268
5269 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5270 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5271 if (x_elem != T_INT || z_elem != T_INT) {
5272 return false;
5273 }
5274
5275
5276 Node* x_start = array_element_address(x, intcon(0), x_elem);
5277 Node* z_start = array_element_address(z, intcon(0), z_elem);
5278
5279 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5280 OptoRuntime::squareToLen_Type(),
5281 stubAddr, stubName, TypePtr::BOTTOM,
5282 x_start, len, z_start, zlen);
5283
5284 set_result(z);
5285 return true;
5286 }
5287
5288 //-------------inline_mulAdd------------------------------------------
5289 bool LibraryCallKit::inline_mulAdd() {
5290 assert(UseMulAddIntrinsic, "not implemented on this platform");
5291
5292 address stubAddr = StubRoutines::mulAdd();
5293 if (stubAddr == NULL) {
5294 return false; // Intrinsic's stub is not implemented on this platform
5295 }
5296 const char* stubName = "mulAdd";
5297
5298 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5299
5300 Node* out = argument(0);
5301 Node* in = argument(1);
5302 Node* offset = argument(2);
5303 Node* len = argument(3);
5304 Node* k = argument(4);
5305
5306 const Type* out_type = out->Value(&_gvn);
5307 const Type* in_type = in->Value(&_gvn);
5308 const TypeAryPtr* top_out = out_type->isa_aryptr();
5309 const TypeAryPtr* top_in = in_type->isa_aryptr();
5310 if (top_out == NULL || top_out->klass() == NULL ||
5311 top_in == NULL || top_in->klass() == NULL) {
5312 // failed array check
5313 return false;
5314 }
5315
5316 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5317 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5318 if (out_elem != T_INT || in_elem != T_INT) {
5319 return false;
5320 }
5321
5322 Node* outlen = load_array_length(out);
5323 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5324 Node* out_start = array_element_address(out, intcon(0), out_elem);
5325 Node* in_start = array_element_address(in, intcon(0), in_elem);
5326
5327 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5328 OptoRuntime::mulAdd_Type(),
5329 stubAddr, stubName, TypePtr::BOTTOM,
5330 out_start,in_start, new_offset, len, k);
5331 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5332 set_result(result);
5333 return true;
5334 }
5335
5336 //-------------inline_montgomeryMultiply-----------------------------------
5337 bool LibraryCallKit::inline_montgomeryMultiply() {
5338 address stubAddr = StubRoutines::montgomeryMultiply();
5339 if (stubAddr == NULL) {
5340 return false; // Intrinsic's stub is not implemented on this platform
5341 }
5342
5343 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5344 const char* stubName = "montgomery_square";
5345
5346 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5347
5348 Node* a = argument(0);
5349 Node* b = argument(1);
5350 Node* n = argument(2);
5351 Node* len = argument(3);
5352 Node* inv = argument(4);
5353 Node* m = argument(6);
5354
5355 const Type* a_type = a->Value(&_gvn);
5356 const TypeAryPtr* top_a = a_type->isa_aryptr();
5357 const Type* b_type = b->Value(&_gvn);
5358 const TypeAryPtr* top_b = b_type->isa_aryptr();
5359 const Type* n_type = a->Value(&_gvn);
5360 const TypeAryPtr* top_n = n_type->isa_aryptr();
5361 const Type* m_type = a->Value(&_gvn);
5362 const TypeAryPtr* top_m = m_type->isa_aryptr();
5363 if (top_a == NULL || top_a->klass() == NULL ||
5364 top_b == NULL || top_b->klass() == NULL ||
5365 top_n == NULL || top_n->klass() == NULL ||
5366 top_m == NULL || top_m->klass() == NULL) {
5367 // failed array check
5368 return false;
5369 }
5370
5371 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5372 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5373 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5374 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5375 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5376 return false;
5377 }
5378
5379 // Make the call
5380 {
5381 Node* a_start = array_element_address(a, intcon(0), a_elem);
5382 Node* b_start = array_element_address(b, intcon(0), b_elem);
5383 Node* n_start = array_element_address(n, intcon(0), n_elem);
5384 Node* m_start = array_element_address(m, intcon(0), m_elem);
5385
5386 Node* call = make_runtime_call(RC_LEAF,
5387 OptoRuntime::montgomeryMultiply_Type(),
5388 stubAddr, stubName, TypePtr::BOTTOM,
5389 a_start, b_start, n_start, len, inv, top(),
5390 m_start);
5391 set_result(m);
5392 }
5393
5394 return true;
5395 }
5396
5397 bool LibraryCallKit::inline_montgomerySquare() {
5398 address stubAddr = StubRoutines::montgomerySquare();
5399 if (stubAddr == NULL) {
5400 return false; // Intrinsic's stub is not implemented on this platform
5401 }
5402
5403 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5404 const char* stubName = "montgomery_square";
5405
5406 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5407
5408 Node* a = argument(0);
5409 Node* n = argument(1);
5410 Node* len = argument(2);
5411 Node* inv = argument(3);
5412 Node* m = argument(5);
5413
5414 const Type* a_type = a->Value(&_gvn);
5415 const TypeAryPtr* top_a = a_type->isa_aryptr();
5416 const Type* n_type = a->Value(&_gvn);
5417 const TypeAryPtr* top_n = n_type->isa_aryptr();
5418 const Type* m_type = a->Value(&_gvn);
5419 const TypeAryPtr* top_m = m_type->isa_aryptr();
5420 if (top_a == NULL || top_a->klass() == NULL ||
5421 top_n == NULL || top_n->klass() == NULL ||
5422 top_m == NULL || top_m->klass() == NULL) {
5423 // failed array check
5424 return false;
5425 }
5426
5427 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5428 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5429 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5430 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5431 return false;
5432 }
5433
5434 // Make the call
5435 {
5436 Node* a_start = array_element_address(a, intcon(0), a_elem);
5437 Node* n_start = array_element_address(n, intcon(0), n_elem);
5438 Node* m_start = array_element_address(m, intcon(0), m_elem);
5439
5440 Node* call = make_runtime_call(RC_LEAF,
5441 OptoRuntime::montgomerySquare_Type(),
5442 stubAddr, stubName, TypePtr::BOTTOM,
5443 a_start, n_start, len, inv, top(),
5444 m_start);
5445 set_result(m);
5446 }
5447
5448 return true;
5449 }
5450
5451
5452 /**
5453 * Calculate CRC32 for byte.
5454 * int java.util.zip.CRC32.update(int crc, int b)
5455 */
5456 bool LibraryCallKit::inline_updateCRC32() {
5457 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5458 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5459 // no receiver since it is static method
5460 Node* crc = argument(0); // type: int
5461 Node* b = argument(1); // type: int
5462
5463 /*
5464 * int c = ~ crc;
5465 * b = timesXtoThe32[(b ^ c) & 0xFF];
5466 * b = b ^ (c >>> 8);
5467 * crc = ~b;
5468 */
5469
5470 Node* M1 = intcon(-1);
5471 crc = _gvn.transform(new XorINode(crc, M1));
5472 Node* result = _gvn.transform(new XorINode(crc, b));
5473 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5474
5475 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5476 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5477 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5478 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5479
5480 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5481 result = _gvn.transform(new XorINode(crc, result));
5482 result = _gvn.transform(new XorINode(result, M1));
5483 set_result(result);
5484 return true;
5485 }
5486
5487 /**
5488 * Calculate CRC32 for byte[] array.
5489 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5490 */
5491 bool LibraryCallKit::inline_updateBytesCRC32() {
5492 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5493 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5494 // no receiver since it is static method
5495 Node* crc = argument(0); // type: int
5496 Node* src = argument(1); // type: oop
5497 Node* offset = argument(2); // type: int
5498 Node* length = argument(3); // type: int
5499
5500 const Type* src_type = src->Value(&_gvn);
5501 const TypeAryPtr* top_src = src_type->isa_aryptr();
5502 if (top_src == NULL || top_src->klass() == NULL) {
5503 // failed array check
5504 return false;
5505 }
5506
5507 // Figure out the size and type of the elements we will be copying.
5508 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5509 if (src_elem != T_BYTE) {
5510 return false;
5511 }
5512
5513 // 'src_start' points to src array + scaled offset
5514 Node* src_start = array_element_address(src, offset, src_elem);
5515
5516 // We assume that range check is done by caller.
5517 // TODO: generate range check (offset+length < src.length) in debug VM.
5518
5519 // Call the stub.
5520 address stubAddr = StubRoutines::updateBytesCRC32();
5521 const char *stubName = "updateBytesCRC32";
5522
5523 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5524 stubAddr, stubName, TypePtr::BOTTOM,
5525 crc, src_start, length);
5526 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5527 set_result(result);
5528 return true;
5529 }
5530
5531 /**
5532 * Calculate CRC32 for ByteBuffer.
5533 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5534 */
5535 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5536 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5537 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5538 // no receiver since it is static method
5539 Node* crc = argument(0); // type: int
5540 Node* src = argument(1); // type: long
5541 Node* offset = argument(3); // type: int
5542 Node* length = argument(4); // type: int
5543
5544 src = ConvL2X(src); // adjust Java long to machine word
5545 Node* base = _gvn.transform(new CastX2PNode(src));
5546 offset = ConvI2X(offset);
5547
5548 // 'src_start' points to src array + scaled offset
5549 Node* src_start = basic_plus_adr(top(), base, offset);
5550
5551 // Call the stub.
5552 address stubAddr = StubRoutines::updateBytesCRC32();
5553 const char *stubName = "updateBytesCRC32";
5554
5555 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5556 stubAddr, stubName, TypePtr::BOTTOM,
5557 crc, src_start, length);
5558 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5559 set_result(result);
5560 return true;
5561 }
5562
5563 //------------------------------get_table_from_crc32c_class-----------------------
5564 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5565 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5566 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5567
5568 return table;
5569 }
5570
5571 //------------------------------inline_updateBytesCRC32C-----------------------
5572 //
5573 // Calculate CRC32C for byte[] array.
5574 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5575 //
5576 bool LibraryCallKit::inline_updateBytesCRC32C() {
5577 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5578 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5579 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5580 // no receiver since it is a static method
5581 Node* crc = argument(0); // type: int
5582 Node* src = argument(1); // type: oop
5583 Node* offset = argument(2); // type: int
5584 Node* end = argument(3); // type: int
5585
5586 Node* length = _gvn.transform(new SubINode(end, offset));
5587
5588 const Type* src_type = src->Value(&_gvn);
5589 const TypeAryPtr* top_src = src_type->isa_aryptr();
5590 if (top_src == NULL || top_src->klass() == NULL) {
5591 // failed array check
5592 return false;
5593 }
5594
5595 // Figure out the size and type of the elements we will be copying.
5596 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5597 if (src_elem != T_BYTE) {
5598 return false;
5599 }
5600
5601 // 'src_start' points to src array + scaled offset
5602 Node* src_start = array_element_address(src, offset, src_elem);
5603
5604 // static final int[] byteTable in class CRC32C
5605 Node* table = get_table_from_crc32c_class(callee()->holder());
5606 Node* table_start = array_element_address(table, intcon(0), T_INT);
5607
5608 // We assume that range check is done by caller.
5609 // TODO: generate range check (offset+length < src.length) in debug VM.
5610
5611 // Call the stub.
5612 address stubAddr = StubRoutines::updateBytesCRC32C();
5613 const char *stubName = "updateBytesCRC32C";
5614
5615 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5616 stubAddr, stubName, TypePtr::BOTTOM,
5617 crc, src_start, length, table_start);
5618 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5619 set_result(result);
5620 return true;
5621 }
5622
5623 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5624 //
5625 // Calculate CRC32C for DirectByteBuffer.
5626 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5627 //
5628 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5629 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5630 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5631 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5632 // no receiver since it is a static method
5633 Node* crc = argument(0); // type: int
5634 Node* src = argument(1); // type: long
5635 Node* offset = argument(3); // type: int
5636 Node* end = argument(4); // type: int
5637
5638 Node* length = _gvn.transform(new SubINode(end, offset));
5639
5640 src = ConvL2X(src); // adjust Java long to machine word
5641 Node* base = _gvn.transform(new CastX2PNode(src));
5642 offset = ConvI2X(offset);
5643
5644 // 'src_start' points to src array + scaled offset
5645 Node* src_start = basic_plus_adr(top(), base, offset);
5646
5647 // static final int[] byteTable in class CRC32C
5648 Node* table = get_table_from_crc32c_class(callee()->holder());
5649 Node* table_start = array_element_address(table, intcon(0), T_INT);
5650
5651 // Call the stub.
5652 address stubAddr = StubRoutines::updateBytesCRC32C();
5653 const char *stubName = "updateBytesCRC32C";
5654
5655 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5656 stubAddr, stubName, TypePtr::BOTTOM,
5657 crc, src_start, length, table_start);
5658 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5659 set_result(result);
5660 return true;
5661 }
5662
5663 //------------------------------inline_updateBytesAdler32----------------------
5664 //
5665 // Calculate Adler32 checksum for byte[] array.
5666 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5667 //
5668 bool LibraryCallKit::inline_updateBytesAdler32() {
5669 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5670 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5671 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5672 // no receiver since it is static method
5673 Node* crc = argument(0); // type: int
5674 Node* src = argument(1); // type: oop
5675 Node* offset = argument(2); // type: int
5676 Node* length = argument(3); // type: int
5677
5678 const Type* src_type = src->Value(&_gvn);
5679 const TypeAryPtr* top_src = src_type->isa_aryptr();
5680 if (top_src == NULL || top_src->klass() == NULL) {
5681 // failed array check
5682 return false;
5683 }
5684
5685 // Figure out the size and type of the elements we will be copying.
5686 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5687 if (src_elem != T_BYTE) {
5688 return false;
5689 }
5690
5691 // 'src_start' points to src array + scaled offset
5692 Node* src_start = array_element_address(src, offset, src_elem);
5693
5694 // We assume that range check is done by caller.
5695 // TODO: generate range check (offset+length < src.length) in debug VM.
5696
5697 // Call the stub.
5698 address stubAddr = StubRoutines::updateBytesAdler32();
5699 const char *stubName = "updateBytesAdler32";
5700
5701 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5702 stubAddr, stubName, TypePtr::BOTTOM,
5703 crc, src_start, length);
5704 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5705 set_result(result);
5706 return true;
5707 }
5708
5709 //------------------------------inline_updateByteBufferAdler32---------------
5710 //
5711 // Calculate Adler32 checksum for DirectByteBuffer.
5712 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5713 //
5714 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5715 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5716 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5717 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5718 // no receiver since it is static method
5719 Node* crc = argument(0); // type: int
5720 Node* src = argument(1); // type: long
5721 Node* offset = argument(3); // type: int
5722 Node* length = argument(4); // type: int
5723
5724 src = ConvL2X(src); // adjust Java long to machine word
5725 Node* base = _gvn.transform(new CastX2PNode(src));
5726 offset = ConvI2X(offset);
5727
5728 // 'src_start' points to src array + scaled offset
5729 Node* src_start = basic_plus_adr(top(), base, offset);
5730
5731 // Call the stub.
5732 address stubAddr = StubRoutines::updateBytesAdler32();
5733 const char *stubName = "updateBytesAdler32";
5734
5735 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5736 stubAddr, stubName, TypePtr::BOTTOM,
5737 crc, src_start, length);
5738
5739 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5740 set_result(result);
5741 return true;
5742 }
5743
5744 //----------------------------inline_reference_get----------------------------
5745 // public T java.lang.ref.Reference.get();
5746 bool LibraryCallKit::inline_reference_get() {
5747 const int referent_offset = java_lang_ref_Reference::referent_offset;
5748 guarantee(referent_offset > 0, "should have already been set");
5749
5750 // Get the argument:
5751 Node* reference_obj = null_check_receiver();
5752 if (stopped()) return true;
5753
5754 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5755
5756 ciInstanceKlass* klass = env()->Object_klass();
5757 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5758
5759 Node* no_ctrl = NULL;
5760 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5761
5762 // Use the pre-barrier to record the value in the referent field
5763 pre_barrier(false /* do_load */,
5764 control(),
5765 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5766 result /* pre_val */,
5767 T_OBJECT);
5768
5769 // Add memory barrier to prevent commoning reads from this field
5770 // across safepoint since GC can change its value.
5771 insert_mem_bar(Op_MemBarCPUOrder);
5772
5773 set_result(result);
5774 return true;
5775 }
5776
5777
5778 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5779 bool is_exact=true, bool is_static=false,
5780 ciInstanceKlass * fromKls=NULL) {
5781 if (fromKls == NULL) {
5782 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5783 assert(tinst != NULL, "obj is null");
5784 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5785 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5786 fromKls = tinst->klass()->as_instance_klass();
5787 } else {
5788 assert(is_static, "only for static field access");
5789 }
5790 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5791 ciSymbol::make(fieldTypeString),
5792 is_static);
5793
5794 assert (field != NULL, "undefined field");
5795 if (field == NULL) return (Node *) NULL;
5796
5797 if (is_static) {
5798 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5799 fromObj = makecon(tip);
5800 }
5801
5802 // Next code copied from Parse::do_get_xxx():
5803
5804 // Compute address and memory type.
5805 int offset = field->offset_in_bytes();
5806 bool is_vol = field->is_volatile();
5807 ciType* field_klass = field->type();
5808 assert(field_klass->is_loaded(), "should be loaded");
5809 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5810 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5811 BasicType bt = field->layout_type();
5812
5813 // Build the resultant type of the load
5814 const Type *type;
5815 if (bt == T_OBJECT) {
5816 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5817 } else {
5818 type = Type::get_const_basic_type(bt);
5819 }
5820
5821 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5822 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5823 }
5824 // Build the load.
5825 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5826 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5827 // If reference is volatile, prevent following memory ops from
5828 // floating up past the volatile read. Also prevents commoning
5829 // another volatile read.
5830 if (is_vol) {
5831 // Memory barrier includes bogus read of value to force load BEFORE membar
5832 insert_mem_bar(Op_MemBarAcquire, loadedField);
5833 }
5834 return loadedField;
5835 }
5836
5837
5838 //------------------------------inline_aescrypt_Block-----------------------
5839 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5840 address stubAddr = NULL;
5841 const char *stubName;
5842 assert(UseAES, "need AES instruction support");
5843
5844 switch(id) {
5845 case vmIntrinsics::_aescrypt_encryptBlock:
5846 stubAddr = StubRoutines::aescrypt_encryptBlock();
5847 stubName = "aescrypt_encryptBlock";
5848 break;
5849 case vmIntrinsics::_aescrypt_decryptBlock:
5850 stubAddr = StubRoutines::aescrypt_decryptBlock();
5851 stubName = "aescrypt_decryptBlock";
5852 break;
5853 }
5854 if (stubAddr == NULL) return false;
5855
5856 Node* aescrypt_object = argument(0);
5857 Node* src = argument(1);
5858 Node* src_offset = argument(2);
5859 Node* dest = argument(3);
5860 Node* dest_offset = argument(4);
5861
5862 // (1) src and dest are arrays.
5863 const Type* src_type = src->Value(&_gvn);
5864 const Type* dest_type = dest->Value(&_gvn);
5865 const TypeAryPtr* top_src = src_type->isa_aryptr();
5866 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5867 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5868
5869 // for the quick and dirty code we will skip all the checks.
5870 // we are just trying to get the call to be generated.
5871 Node* src_start = src;
5872 Node* dest_start = dest;
5873 if (src_offset != NULL || dest_offset != NULL) {
5874 assert(src_offset != NULL && dest_offset != NULL, "");
5875 src_start = array_element_address(src, src_offset, T_BYTE);
5876 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5877 }
5878
5879 // now need to get the start of its expanded key array
5880 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5881 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5882 if (k_start == NULL) return false;
5883
5884 if (Matcher::pass_original_key_for_aes()) {
5885 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5886 // compatibility issues between Java key expansion and SPARC crypto instructions
5887 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5888 if (original_k_start == NULL) return false;
5889
5890 // Call the stub.
5891 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5892 stubAddr, stubName, TypePtr::BOTTOM,
5893 src_start, dest_start, k_start, original_k_start);
5894 } else {
5895 // Call the stub.
5896 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5897 stubAddr, stubName, TypePtr::BOTTOM,
5898 src_start, dest_start, k_start);
5899 }
5900
5901 return true;
5902 }
5903
5904 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5905 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5906 address stubAddr = NULL;
5907 const char *stubName = NULL;
5908
5909 assert(UseAES, "need AES instruction support");
5910
5911 switch(id) {
5912 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5913 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5914 stubName = "cipherBlockChaining_encryptAESCrypt";
5915 break;
5916 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5917 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5918 stubName = "cipherBlockChaining_decryptAESCrypt";
5919 break;
5920 }
5921 if (stubAddr == NULL) return false;
5922
5923 Node* cipherBlockChaining_object = argument(0);
5924 Node* src = argument(1);
5925 Node* src_offset = argument(2);
5926 Node* len = argument(3);
5927 Node* dest = argument(4);
5928 Node* dest_offset = argument(5);
5929
5930 // (1) src and dest are arrays.
5931 const Type* src_type = src->Value(&_gvn);
5932 const Type* dest_type = dest->Value(&_gvn);
5933 const TypeAryPtr* top_src = src_type->isa_aryptr();
5934 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5935 assert (top_src != NULL && top_src->klass() != NULL
5936 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5937
5938 // checks are the responsibility of the caller
5939 Node* src_start = src;
5940 Node* dest_start = dest;
5941 if (src_offset != NULL || dest_offset != NULL) {
5942 assert(src_offset != NULL && dest_offset != NULL, "");
5943 src_start = array_element_address(src, src_offset, T_BYTE);
5944 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5945 }
5946
5947 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5948 // (because of the predicated logic executed earlier).
5949 // so we cast it here safely.
5950 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5951
5952 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5953 if (embeddedCipherObj == NULL) return false;
5954
5955 // cast it to what we know it will be at runtime
5956 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5957 assert(tinst != NULL, "CBC obj is null");
5958 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5959 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5960 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5961
5962 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5963 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5964 const TypeOopPtr* xtype = aklass->as_instance_type();
5965 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5966 aescrypt_object = _gvn.transform(aescrypt_object);
5967
5968 // we need to get the start of the aescrypt_object's expanded key array
5969 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5970 if (k_start == NULL) return false;
5971
5972 // similarly, get the start address of the r vector
5973 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5974 if (objRvec == NULL) return false;
5975 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5976
5977 Node* cbcCrypt;
5978 if (Matcher::pass_original_key_for_aes()) {
5979 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5980 // compatibility issues between Java key expansion and SPARC crypto instructions
5981 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5982 if (original_k_start == NULL) return false;
5983
5984 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5985 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5986 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5987 stubAddr, stubName, TypePtr::BOTTOM,
5988 src_start, dest_start, k_start, r_start, len, original_k_start);
5989 } else {
5990 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5991 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5992 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5993 stubAddr, stubName, TypePtr::BOTTOM,
5994 src_start, dest_start, k_start, r_start, len);
5995 }
5996
5997 // return cipher length (int)
5998 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5999 set_result(retvalue);
6000 return true;
6001 }
6002
6003 //------------------------------get_key_start_from_aescrypt_object-----------------------
6004 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6005 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6006 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6007 if (objAESCryptKey == NULL) return (Node *) NULL;
6008
6009 // now have the array, need to get the start address of the K array
6010 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6011 return k_start;
6012 }
6013
6014 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6015 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6016 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6017 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6018 if (objAESCryptKey == NULL) return (Node *) NULL;
6019
6020 // now have the array, need to get the start address of the lastKey array
6021 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6022 return original_k_start;
6023 }
6024
6025 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6026 // Return node representing slow path of predicate check.
6027 // the pseudo code we want to emulate with this predicate is:
6028 // for encryption:
6029 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6030 // for decryption:
6031 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6032 // note cipher==plain is more conservative than the original java code but that's OK
6033 //
6034 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6035 // The receiver was checked for NULL already.
6036 Node* objCBC = argument(0);
6037
6038 // Load embeddedCipher field of CipherBlockChaining object.
6039 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6040
6041 // get AESCrypt klass for instanceOf check
6042 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6043 // will have same classloader as CipherBlockChaining object
6044 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6045 assert(tinst != NULL, "CBCobj is null");
6046 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6047
6048 // we want to do an instanceof comparison against the AESCrypt class
6049 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6050 if (!klass_AESCrypt->is_loaded()) {
6051 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6052 Node* ctrl = control();
6053 set_control(top()); // no regular fast path
6054 return ctrl;
6055 }
6056 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6057
6058 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6059 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6060 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6061
6062 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6063
6064 // for encryption, we are done
6065 if (!decrypting)
6066 return instof_false; // even if it is NULL
6067
6068 // for decryption, we need to add a further check to avoid
6069 // taking the intrinsic path when cipher and plain are the same
6070 // see the original java code for why.
6071 RegionNode* region = new RegionNode(3);
6072 region->init_req(1, instof_false);
6073 Node* src = argument(1);
6074 Node* dest = argument(4);
6075 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6076 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6077 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6078 region->init_req(2, src_dest_conjoint);
6079
6080 record_for_igvn(region);
6081 return _gvn.transform(region);
6082 }
6083
6084 //------------------------------inline_ghash_processBlocks
6085 bool LibraryCallKit::inline_ghash_processBlocks() {
6086 address stubAddr;
6087 const char *stubName;
6088 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6089
6090 stubAddr = StubRoutines::ghash_processBlocks();
6091 stubName = "ghash_processBlocks";
6092
6093 Node* data = argument(0);
6094 Node* offset = argument(1);
6095 Node* len = argument(2);
6096 Node* state = argument(3);
6097 Node* subkeyH = argument(4);
6098
6099 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6100 assert(state_start, "state is NULL");
6101 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6102 assert(subkeyH_start, "subkeyH is NULL");
6103 Node* data_start = array_element_address(data, offset, T_BYTE);
6104 assert(data_start, "data is NULL");
6105
6106 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6107 OptoRuntime::ghash_processBlocks_Type(),
6108 stubAddr, stubName, TypePtr::BOTTOM,
6109 state_start, subkeyH_start, data_start, len);
6110 return true;
6111 }
6112
6113 //------------------------------inline_sha_implCompress-----------------------
6114 //
6115 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6116 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6117 //
6118 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6119 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6120 //
6121 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6122 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6123 //
6124 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6125 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6126
6127 Node* sha_obj = argument(0);
6128 Node* src = argument(1); // type oop
6129 Node* ofs = argument(2); // type int
6130
6131 const Type* src_type = src->Value(&_gvn);
6132 const TypeAryPtr* top_src = src_type->isa_aryptr();
6133 if (top_src == NULL || top_src->klass() == NULL) {
6134 // failed array check
6135 return false;
6136 }
6137 // Figure out the size and type of the elements we will be copying.
6138 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6139 if (src_elem != T_BYTE) {
6140 return false;
6141 }
6142 // 'src_start' points to src array + offset
6143 Node* src_start = array_element_address(src, ofs, src_elem);
6144 Node* state = NULL;
6145 address stubAddr;
6146 const char *stubName;
6147
6148 switch(id) {
6149 case vmIntrinsics::_sha_implCompress:
6150 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6151 state = get_state_from_sha_object(sha_obj);
6152 stubAddr = StubRoutines::sha1_implCompress();
6153 stubName = "sha1_implCompress";
6154 break;
6155 case vmIntrinsics::_sha2_implCompress:
6156 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6157 state = get_state_from_sha_object(sha_obj);
6158 stubAddr = StubRoutines::sha256_implCompress();
6159 stubName = "sha256_implCompress";
6160 break;
6161 case vmIntrinsics::_sha5_implCompress:
6162 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6163 state = get_state_from_sha5_object(sha_obj);
6164 stubAddr = StubRoutines::sha512_implCompress();
6165 stubName = "sha512_implCompress";
6166 break;
6167 default:
6168 fatal_unexpected_iid(id);
6169 return false;
6170 }
6171 if (state == NULL) return false;
6172
6173 // Call the stub.
6174 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6175 stubAddr, stubName, TypePtr::BOTTOM,
6176 src_start, state);
6177
6178 return true;
6179 }
6180
6181 //------------------------------inline_digestBase_implCompressMB-----------------------
6182 //
6183 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6184 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6185 //
6186 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6187 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6188 "need SHA1/SHA256/SHA512 instruction support");
6189 assert((uint)predicate < 3, "sanity");
6190 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6191
6192 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6193 Node* src = argument(1); // byte[] array
6194 Node* ofs = argument(2); // type int
6195 Node* limit = argument(3); // type int
6196
6197 const Type* src_type = src->Value(&_gvn);
6198 const TypeAryPtr* top_src = src_type->isa_aryptr();
6199 if (top_src == NULL || top_src->klass() == NULL) {
6200 // failed array check
6201 return false;
6202 }
6203 // Figure out the size and type of the elements we will be copying.
6204 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6205 if (src_elem != T_BYTE) {
6206 return false;
6207 }
6208 // 'src_start' points to src array + offset
6209 Node* src_start = array_element_address(src, ofs, src_elem);
6210
6211 const char* klass_SHA_name = NULL;
6212 const char* stub_name = NULL;
6213 address stub_addr = NULL;
6214 bool long_state = false;
6215
6216 switch (predicate) {
6217 case 0:
6218 if (UseSHA1Intrinsics) {
6219 klass_SHA_name = "sun/security/provider/SHA";
6220 stub_name = "sha1_implCompressMB";
6221 stub_addr = StubRoutines::sha1_implCompressMB();
6222 }
6223 break;
6224 case 1:
6225 if (UseSHA256Intrinsics) {
6226 klass_SHA_name = "sun/security/provider/SHA2";
6227 stub_name = "sha256_implCompressMB";
6228 stub_addr = StubRoutines::sha256_implCompressMB();
6229 }
6230 break;
6231 case 2:
6232 if (UseSHA512Intrinsics) {
6233 klass_SHA_name = "sun/security/provider/SHA5";
6234 stub_name = "sha512_implCompressMB";
6235 stub_addr = StubRoutines::sha512_implCompressMB();
6236 long_state = true;
6237 }
6238 break;
6239 default:
6240 fatal("unknown SHA intrinsic predicate: %d", predicate);
6241 }
6242 if (klass_SHA_name != NULL) {
6243 // get DigestBase klass to lookup for SHA klass
6244 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6245 assert(tinst != NULL, "digestBase_obj is not instance???");
6246 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6247
6248 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6249 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6250 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6251 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6252 }
6253 return false;
6254 }
6255 //------------------------------inline_sha_implCompressMB-----------------------
6256 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6257 bool long_state, address stubAddr, const char *stubName,
6258 Node* src_start, Node* ofs, Node* limit) {
6259 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6260 const TypeOopPtr* xtype = aklass->as_instance_type();
6261 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6262 sha_obj = _gvn.transform(sha_obj);
6263
6264 Node* state;
6265 if (long_state) {
6266 state = get_state_from_sha5_object(sha_obj);
6267 } else {
6268 state = get_state_from_sha_object(sha_obj);
6269 }
6270 if (state == NULL) return false;
6271
6272 // Call the stub.
6273 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6274 OptoRuntime::digestBase_implCompressMB_Type(),
6275 stubAddr, stubName, TypePtr::BOTTOM,
6276 src_start, state, ofs, limit);
6277 // return ofs (int)
6278 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6279 set_result(result);
6280
6281 return true;
6282 }
6283
6284 //------------------------------get_state_from_sha_object-----------------------
6285 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6286 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6287 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6288 if (sha_state == NULL) return (Node *) NULL;
6289
6290 // now have the array, need to get the start address of the state array
6291 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6292 return state;
6293 }
6294
6295 //------------------------------get_state_from_sha5_object-----------------------
6296 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6297 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6298 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6299 if (sha_state == NULL) return (Node *) NULL;
6300
6301 // now have the array, need to get the start address of the state array
6302 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6303 return state;
6304 }
6305
6306 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6307 // Return node representing slow path of predicate check.
6308 // the pseudo code we want to emulate with this predicate is:
6309 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6310 //
6311 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6312 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6313 "need SHA1/SHA256/SHA512 instruction support");
6314 assert((uint)predicate < 3, "sanity");
6315
6316 // The receiver was checked for NULL already.
6317 Node* digestBaseObj = argument(0);
6318
6319 // get DigestBase klass for instanceOf check
6320 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6321 assert(tinst != NULL, "digestBaseObj is null");
6322 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6323
6324 const char* klass_SHA_name = NULL;
6325 switch (predicate) {
6326 case 0:
6327 if (UseSHA1Intrinsics) {
6328 // we want to do an instanceof comparison against the SHA class
6329 klass_SHA_name = "sun/security/provider/SHA";
6330 }
6331 break;
6332 case 1:
6333 if (UseSHA256Intrinsics) {
6334 // we want to do an instanceof comparison against the SHA2 class
6335 klass_SHA_name = "sun/security/provider/SHA2";
6336 }
6337 break;
6338 case 2:
6339 if (UseSHA512Intrinsics) {
6340 // we want to do an instanceof comparison against the SHA5 class
6341 klass_SHA_name = "sun/security/provider/SHA5";
6342 }
6343 break;
6344 default:
6345 fatal("unknown SHA intrinsic predicate: %d", predicate);
6346 }
6347
6348 ciKlass* klass_SHA = NULL;
6349 if (klass_SHA_name != NULL) {
6350 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6351 }
6352 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6353 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6354 Node* ctrl = control();
6355 set_control(top()); // no intrinsic path
6356 return ctrl;
6357 }
6358 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6359
6360 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6361 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6362 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6363 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6364
6365 return instof_false; // even if it is NULL
6366 }
6367
6368 bool LibraryCallKit::inline_profileBoolean() {
6369 Node* counts = argument(1);
6370 const TypeAryPtr* ary = NULL;
6371 ciArray* aobj = NULL;
6372 if (counts->is_Con()
6373 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6374 && (aobj = ary->const_oop()->as_array()) != NULL
6375 && (aobj->length() == 2)) {
6376 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6377 jint false_cnt = aobj->element_value(0).as_int();
6378 jint true_cnt = aobj->element_value(1).as_int();
6379
6380 if (C->log() != NULL) {
6381 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6382 false_cnt, true_cnt);
6383 }
6384
6385 if (false_cnt + true_cnt == 0) {
6386 // According to profile, never executed.
6387 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6388 Deoptimization::Action_reinterpret);
6389 return true;
6390 }
6391
6392 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6393 // is a number of each value occurrences.
6394 Node* result = argument(0);
6395 if (false_cnt == 0 || true_cnt == 0) {
6396 // According to profile, one value has been never seen.
6397 int expected_val = (false_cnt == 0) ? 1 : 0;
6398
6399 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6400 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6401
6402 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6403 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6404 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6405
6406 { // Slow path: uncommon trap for never seen value and then reexecute
6407 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6408 // the value has been seen at least once.
6409 PreserveJVMState pjvms(this);
6410 PreserveReexecuteState preexecs(this);
6411 jvms()->set_should_reexecute(true);
6412
6413 set_control(slow_path);
6414 set_i_o(i_o());
6415
6416 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6417 Deoptimization::Action_reinterpret);
6418 }
6419 // The guard for never seen value enables sharpening of the result and
6420 // returning a constant. It allows to eliminate branches on the same value
6421 // later on.
6422 set_control(fast_path);
6423 result = intcon(expected_val);
6424 }
6425 // Stop profiling.
6426 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6427 // By replacing method body with profile data (represented as ProfileBooleanNode
6428 // on IR level) we effectively disable profiling.
6429 // It enables full speed execution once optimized code is generated.
6430 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6431 C->record_for_igvn(profile);
6432 set_result(profile);
6433 return true;
6434 } else {
6435 // Continue profiling.
6436 // Profile data isn't available at the moment. So, execute method's bytecode version.
6437 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6438 // is compiled and counters aren't available since corresponding MethodHandle
6439 // isn't a compile-time constant.
6440 return false;
6441 }
6442 }
6443
6444 bool LibraryCallKit::inline_isCompileConstant() {
6445 Node* n = argument(0);
6446 set_result(n->is_Con() ? intcon(1) : intcon(0));
6447 return true;
6448 }