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