1 /*
2 * Copyright (c) 1999, 2017, 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 "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "prims/unsafe.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #ifdef TRACE_HAVE_INTRINSICS
56 #include "trace/traceMacros.hpp"
57 #endif
58
59 class LibraryIntrinsic : public InlineCallGenerator {
60 // Extend the set of intrinsics known to the runtime:
61 public:
62 private:
63 bool _is_virtual;
64 bool _does_virtual_dispatch;
65 int8_t _predicates_count; // Intrinsic is predicated by several conditions
66 int8_t _last_predicate; // Last generated predicate
67 vmIntrinsics::ID _intrinsic_id;
68
69 public:
70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
71 : InlineCallGenerator(m),
72 _is_virtual(is_virtual),
73 _does_virtual_dispatch(does_virtual_dispatch),
74 _predicates_count((int8_t)predicates_count),
75 _last_predicate((int8_t)-1),
76 _intrinsic_id(id)
77 {
78 }
79 virtual bool is_intrinsic() const { return true; }
80 virtual bool is_virtual() const { return _is_virtual; }
81 virtual bool is_predicated() const { return _predicates_count > 0; }
82 virtual int predicates_count() const { return _predicates_count; }
83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
84 virtual JVMState* generate(JVMState* jvms);
85 virtual Node* generate_predicate(JVMState* jvms, int predicate);
86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
87 };
88
89
90 // Local helper class for LibraryIntrinsic:
91 class LibraryCallKit : public GraphKit {
92 private:
93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
94 Node* _result; // the result node, if any
95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
96
97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
98
99 public:
100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
101 : GraphKit(jvms),
102 _intrinsic(intrinsic),
103 _result(NULL)
104 {
105 // Check if this is a root compile. In that case we don't have a caller.
106 if (!jvms->has_method()) {
107 _reexecute_sp = sp();
108 } else {
109 // Find out how many arguments the interpreter needs when deoptimizing
110 // and save the stack pointer value so it can used by uncommon_trap.
111 // We find the argument count by looking at the declared signature.
112 bool ignored_will_link;
113 ciSignature* declared_signature = NULL;
114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
117 }
118 }
119
120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
121
122 ciMethod* caller() const { return jvms()->method(); }
123 int bci() const { return jvms()->bci(); }
124 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
126 ciMethod* callee() const { return _intrinsic->method(); }
127
128 bool try_to_inline(int predicate);
129 Node* try_to_predicate(int predicate);
130
131 void push_result() {
132 // Push the result onto the stack.
133 if (!stopped() && result() != NULL) {
134 BasicType bt = result()->bottom_type()->basic_type();
135 push_node(bt, result());
136 }
137 }
138
139 private:
140 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
142 }
143
144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
145 void set_result(RegionNode* region, PhiNode* value);
146 Node* result() { return _result; }
147
148 virtual int reexecute_sp() { return _reexecute_sp; }
149
150 // Helper functions to inline natives
151 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
152 Node* generate_slow_guard(Node* test, RegionNode* region);
153 Node* generate_fair_guard(Node* test, RegionNode* region);
154 Node* generate_negative_guard(Node* index, RegionNode* region,
155 // resulting CastII of index:
156 Node* *pos_index = NULL);
157 Node* generate_limit_guard(Node* offset, Node* subseq_length,
158 Node* array_length,
159 RegionNode* region);
160 void generate_string_range_check(Node* array, Node* offset,
161 Node* length, bool char_count);
162 Node* generate_current_thread(Node* &tls_output);
163 Node* load_mirror_from_klass(Node* klass);
164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path,
166 int offset);
167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
175 RegionNode* region, int null_path) {
176 int offset = java_lang_Class::array_klass_offset_in_bytes();
177 return load_klass_from_mirror_common(mirror, never_see_null,
178 region, null_path,
179 offset);
180 }
181 Node* generate_access_flags_guard(Node* kls,
182 int modifier_mask, int modifier_bits,
183 RegionNode* region);
184 Node* generate_interface_guard(Node* kls, RegionNode* region);
185 Node* generate_array_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, false, false);
187 }
188 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
189 return generate_array_guard_common(kls, region, false, true);
190 }
191 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
192 return generate_array_guard_common(kls, region, true, false);
193 }
194 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
195 return generate_array_guard_common(kls, region, true, true);
196 }
197 Node* generate_array_guard_common(Node* kls, RegionNode* region,
198 bool obj_array, bool not_array);
199 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
200 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
201 bool is_virtual = false, bool is_static = false);
202 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
203 return generate_method_call(method_id, false, true);
204 }
205 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
206 return generate_method_call(method_id, true, false);
207 }
208 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210
211 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
215 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
216 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_indexOfChar();
218 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_toBytesU();
220 bool inline_string_getCharsU();
221 bool inline_string_copy(bool compress);
222 bool inline_string_char_access(bool is_store);
223 Node* round_double_node(Node* n);
224 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_math_native(vmIntrinsics::ID id);
226 bool inline_math(vmIntrinsics::ID id);
227 template <typename OverflowOp>
228 bool inline_math_overflow(Node* arg1, Node* arg2);
229 void inline_math_mathExact(Node* math, Node* test);
230 bool inline_math_addExactI(bool is_increment);
231 bool inline_math_addExactL(bool is_increment);
232 bool inline_math_multiplyExactI();
233 bool inline_math_multiplyExactL();
234 bool inline_math_negateExactI();
235 bool inline_math_negateExactL();
236 bool inline_math_subtractExactI(bool is_decrement);
237 bool inline_math_subtractExactL(bool is_decrement);
238 bool inline_min_max(vmIntrinsics::ID id);
239 bool inline_notify(vmIntrinsics::ID id);
240 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
241 // This returns Type::AnyPtr, RawPtr, or OopPtr.
242 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
243 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
244 // Helper for inline_unsafe_access.
245 // Generates the guards that check whether the result of
246 // Unsafe.getObject should be recorded in an SATB log buffer.
247 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
248
249 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
250 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
251 static bool klass_needs_init_guard(Node* kls);
252 bool inline_unsafe_allocate();
253 bool inline_unsafe_newArray(bool uninitialized);
254 bool inline_unsafe_copyMemory();
255 bool inline_native_currentThread();
256
257 bool inline_native_time_funcs(address method, const char* funcName);
258 #ifdef TRACE_HAVE_INTRINSICS
259 bool inline_native_classID();
260 bool inline_native_getBufferWriter();
261 #endif
262 bool inline_native_isInterrupted();
263 bool inline_native_Class_query(vmIntrinsics::ID id);
264 bool inline_native_subtype_check();
265 bool inline_native_getLength();
266 bool inline_array_copyOf(bool is_copyOfRange);
267 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
268 bool inline_preconditions_checkIndex();
269 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
270 bool inline_native_clone(bool is_virtual);
271 bool inline_native_Reflection_getCallerClass();
272 // Helper function for inlining native object hash method
273 bool inline_native_hashcode(bool is_virtual, bool is_static);
274 bool inline_native_getClass();
275
276 // Helper functions for inlining arraycopy
277 bool inline_arraycopy();
278 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
279 RegionNode* slow_region);
280 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
281 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
282 uint new_idx);
283
284 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
285 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
286 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
288 bool inline_unsafe_fence(vmIntrinsics::ID id);
289 bool inline_onspinwait();
290 bool inline_fp_conversions(vmIntrinsics::ID id);
291 bool inline_number_methods(vmIntrinsics::ID id);
292 bool inline_reference_get();
293 bool inline_Class_cast();
294 bool inline_aescrypt_Block(vmIntrinsics::ID id);
295 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
296 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
297 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
298 Node* inline_counterMode_AESCrypt_predicate();
299 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
300 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
301 bool inline_ghash_processBlocks();
302 bool inline_sha_implCompress(vmIntrinsics::ID id);
303 bool inline_digestBase_implCompressMB(int predicate);
304 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
305 bool long_state, address stubAddr, const char *stubName,
306 Node* src_start, Node* ofs, Node* limit);
307 Node* get_state_from_sha_object(Node *sha_object);
308 Node* get_state_from_sha5_object(Node *sha_object);
309 Node* inline_digestBase_implCompressMB_predicate(int predicate);
310 bool inline_encodeISOArray();
311 bool inline_updateCRC32();
312 bool inline_updateBytesCRC32();
313 bool inline_updateByteBufferCRC32();
314 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
315 bool inline_updateBytesCRC32C();
316 bool inline_updateDirectByteBufferCRC32C();
317 bool inline_updateBytesAdler32();
318 bool inline_updateByteBufferAdler32();
319 bool inline_multiplyToLen();
320 bool inline_hasNegatives();
321 bool inline_squareToLen();
322 bool inline_mulAdd();
323 bool inline_montgomeryMultiply();
324 bool inline_montgomerySquare();
325 bool inline_vectorizedMismatch();
326 bool inline_fma(vmIntrinsics::ID id);
327
328 bool inline_profileBoolean();
329 bool inline_isCompileConstant();
330 };
331
332 //---------------------------make_vm_intrinsic----------------------------
333 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
334 vmIntrinsics::ID id = m->intrinsic_id();
335 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
336
337 if (!m->is_loaded()) {
338 // Do not attempt to inline unloaded methods.
339 return NULL;
340 }
341
342 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
343 bool is_available = false;
344
345 {
346 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
347 // the compiler must transition to '_thread_in_vm' state because both
348 // methods access VM-internal data.
349 VM_ENTRY_MARK;
350 methodHandle mh(THREAD, m->get_Method());
351 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
352 !C->directive()->is_intrinsic_disabled(mh) &&
353 !vmIntrinsics::is_disabled_by_flags(mh);
354
355 }
356
357 if (is_available) {
358 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
359 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
360 return new LibraryIntrinsic(m, is_virtual,
361 vmIntrinsics::predicates_needed(id),
362 vmIntrinsics::does_virtual_dispatch(id),
363 (vmIntrinsics::ID) id);
364 } else {
365 return NULL;
366 }
367 }
368
369 //----------------------register_library_intrinsics-----------------------
370 // Initialize this file's data structures, for each Compile instance.
371 void Compile::register_library_intrinsics() {
372 // Nothing to do here.
373 }
374
375 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
376 LibraryCallKit kit(jvms, this);
377 Compile* C = kit.C;
378 int nodes = C->unique();
379 #ifndef PRODUCT
380 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
381 char buf[1000];
382 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
383 tty->print_cr("Intrinsic %s", str);
384 }
385 #endif
386 ciMethod* callee = kit.callee();
387 const int bci = kit.bci();
388
389 // Try to inline the intrinsic.
390 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
391 kit.try_to_inline(_last_predicate)) {
392 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
393 : "(intrinsic)";
394 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
395 if (C->print_intrinsics() || C->print_inlining()) {
396 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
397 }
398 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
399 if (C->log()) {
400 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
401 vmIntrinsics::name_at(intrinsic_id()),
402 (is_virtual() ? " virtual='1'" : ""),
403 C->unique() - nodes);
404 }
405 // Push the result from the inlined method onto the stack.
406 kit.push_result();
407 C->print_inlining_update(this);
408 return kit.transfer_exceptions_into_jvms();
409 }
410
411 // The intrinsic bailed out
412 if (jvms->has_method()) {
413 // Not a root compile.
414 const char* msg;
415 if (callee->intrinsic_candidate()) {
416 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
417 } else {
418 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
419 : "failed to inline (intrinsic), method not annotated";
420 }
421 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
422 if (C->print_intrinsics() || C->print_inlining()) {
423 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
424 }
425 } else {
426 // Root compile
427 ResourceMark rm;
428 stringStream msg_stream;
429 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
430 vmIntrinsics::name_at(intrinsic_id()),
431 is_virtual() ? " (virtual)" : "", bci);
432 const char *msg = msg_stream.as_string();
433 log_debug(jit, inlining)("%s", msg);
434 if (C->print_intrinsics() || C->print_inlining()) {
435 tty->print("%s", msg);
436 }
437 }
438 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
439 C->print_inlining_update(this);
440 return NULL;
441 }
442
443 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
444 LibraryCallKit kit(jvms, this);
445 Compile* C = kit.C;
446 int nodes = C->unique();
447 _last_predicate = predicate;
448 #ifndef PRODUCT
449 assert(is_predicated() && predicate < predicates_count(), "sanity");
450 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
451 char buf[1000];
452 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
453 tty->print_cr("Predicate for intrinsic %s", str);
454 }
455 #endif
456 ciMethod* callee = kit.callee();
457 const int bci = kit.bci();
458
459 Node* slow_ctl = kit.try_to_predicate(predicate);
460 if (!kit.failing()) {
461 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
462 : "(intrinsic, predicate)";
463 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
464 if (C->print_intrinsics() || C->print_inlining()) {
465 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
466 }
467 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
468 if (C->log()) {
469 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
470 vmIntrinsics::name_at(intrinsic_id()),
471 (is_virtual() ? " virtual='1'" : ""),
472 C->unique() - nodes);
473 }
474 return slow_ctl; // Could be NULL if the check folds.
475 }
476
477 // The intrinsic bailed out
478 if (jvms->has_method()) {
479 // Not a root compile.
480 const char* msg = "failed to generate predicate for intrinsic";
481 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
482 if (C->print_intrinsics() || C->print_inlining()) {
483 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
484 }
485 } else {
486 // Root compile
487 ResourceMark rm;
488 stringStream msg_stream;
489 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
490 vmIntrinsics::name_at(intrinsic_id()),
491 is_virtual() ? " (virtual)" : "", bci);
492 const char *msg = msg_stream.as_string();
493 log_debug(jit, inlining)("%s", msg);
494 if (C->print_intrinsics() || C->print_inlining()) {
495 C->print_inlining_stream()->print("%s", msg);
496 }
497 }
498 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
499 return NULL;
500 }
501
502 bool LibraryCallKit::try_to_inline(int predicate) {
503 // Handle symbolic names for otherwise undistinguished boolean switches:
504 const bool is_store = true;
505 const bool is_compress = true;
506 const bool is_static = true;
507 const bool is_volatile = true;
508
509 if (!jvms()->has_method()) {
510 // Root JVMState has a null method.
511 assert(map()->memory()->Opcode() == Op_Parm, "");
512 // Insert the memory aliasing node
513 set_all_memory(reset_memory());
514 }
515 assert(merged_memory(), "");
516
517
518 switch (intrinsic_id()) {
519 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
520 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
521 case vmIntrinsics::_getClass: return inline_native_getClass();
522
523 case vmIntrinsics::_dsin:
524 case vmIntrinsics::_dcos:
525 case vmIntrinsics::_dtan:
526 case vmIntrinsics::_dabs:
527 case vmIntrinsics::_datan2:
528 case vmIntrinsics::_dsqrt:
529 case vmIntrinsics::_dexp:
530 case vmIntrinsics::_dlog:
531 case vmIntrinsics::_dlog10:
532 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
533
534 case vmIntrinsics::_min:
535 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
536
537 case vmIntrinsics::_notify:
538 case vmIntrinsics::_notifyAll:
539 if (InlineNotify) {
540 return inline_notify(intrinsic_id());
541 }
542 return false;
543
544 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
545 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
546 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
547 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
548 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
549 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
550 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
551 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
552 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
553 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
554 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
555 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
556
557 case vmIntrinsics::_arraycopy: return inline_arraycopy();
558
559 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
560 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
561 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
562 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
563
564 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
565 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
566 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
567 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
568 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
569 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
570 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
571
572 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
573 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
574
575 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
576 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
577 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
578 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
579
580 case vmIntrinsics::_compressStringC:
581 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
582 case vmIntrinsics::_inflateStringC:
583 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
584
585 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
586 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
587 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
588 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
589 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
590 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
591 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
592 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
593 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
594
595 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
596 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
597 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
598 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
599 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
600 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
601 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
602 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
603 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
604
605 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
606 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
607 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
608 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
609 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
610 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
611 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
612 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
613 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
614
615 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
616 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
617 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
618 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
619 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
620 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
621 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
622 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
623 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
624
625 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
626 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
627 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
628 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
629
630 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
631 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
632 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
633 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
634
635 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
636 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
637 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
638 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
639 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
640 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
641 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
642 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
643 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
644
645 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
646 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
647 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
648 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
649 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
650 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
651 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
652 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
653 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
654
655 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
656 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
657 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
658 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
659 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
660 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
661 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
662 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
663 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
664
665 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
666 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
667 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
668 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
669 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
670 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
671 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
672 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
673 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
674
675 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
676 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
677 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
678 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
679 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
680
681 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
682 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
683 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
684 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
685 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
686 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
687 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
688 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
689 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
690 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
691 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
692 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
693 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
694 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
695 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
696 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
697 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
698 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
699 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
700 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
701
702 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
703 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
704 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
705 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
706 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
707 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
708 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
709 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
710 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
711 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
712 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
713 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
714 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
715 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
716 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
717
718 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
719 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
720 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
721 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
722
723 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
724 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
725 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
726 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
727 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
728
729 case vmIntrinsics::_loadFence:
730 case vmIntrinsics::_storeFence:
731 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
732
733 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
734
735 case vmIntrinsics::_currentThread: return inline_native_currentThread();
736 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
737
738 #ifdef TRACE_HAVE_INTRINSICS
739 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
740 case vmIntrinsics::_getClassId: return inline_native_classID();
741 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
742 #endif
743 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
744 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
745 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
746 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
747 case vmIntrinsics::_getLength: return inline_native_getLength();
748 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
749 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
750 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
751 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
752 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
753 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
754
755 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
756 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
757
758 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
759
760 case vmIntrinsics::_isInstance:
761 case vmIntrinsics::_getModifiers:
762 case vmIntrinsics::_isInterface:
763 case vmIntrinsics::_isArray:
764 case vmIntrinsics::_isPrimitive:
765 case vmIntrinsics::_getSuperclass:
766 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
767
768 case vmIntrinsics::_floatToRawIntBits:
769 case vmIntrinsics::_floatToIntBits:
770 case vmIntrinsics::_intBitsToFloat:
771 case vmIntrinsics::_doubleToRawLongBits:
772 case vmIntrinsics::_doubleToLongBits:
773 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
774
775 case vmIntrinsics::_numberOfLeadingZeros_i:
776 case vmIntrinsics::_numberOfLeadingZeros_l:
777 case vmIntrinsics::_numberOfTrailingZeros_i:
778 case vmIntrinsics::_numberOfTrailingZeros_l:
779 case vmIntrinsics::_bitCount_i:
780 case vmIntrinsics::_bitCount_l:
781 case vmIntrinsics::_reverseBytes_i:
782 case vmIntrinsics::_reverseBytes_l:
783 case vmIntrinsics::_reverseBytes_s:
784 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
785
786 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
787
788 case vmIntrinsics::_Reference_get: return inline_reference_get();
789
790 case vmIntrinsics::_Class_cast: return inline_Class_cast();
791
792 case vmIntrinsics::_aescrypt_encryptBlock:
793 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
794
795 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
796 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
797 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
798
799 case vmIntrinsics::_counterMode_AESCrypt:
800 return inline_counterMode_AESCrypt(intrinsic_id());
801
802 case vmIntrinsics::_sha_implCompress:
803 case vmIntrinsics::_sha2_implCompress:
804 case vmIntrinsics::_sha5_implCompress:
805 return inline_sha_implCompress(intrinsic_id());
806
807 case vmIntrinsics::_digestBase_implCompressMB:
808 return inline_digestBase_implCompressMB(predicate);
809
810 case vmIntrinsics::_multiplyToLen:
811 return inline_multiplyToLen();
812
813 case vmIntrinsics::_squareToLen:
814 return inline_squareToLen();
815
816 case vmIntrinsics::_mulAdd:
817 return inline_mulAdd();
818
819 case vmIntrinsics::_montgomeryMultiply:
820 return inline_montgomeryMultiply();
821 case vmIntrinsics::_montgomerySquare:
822 return inline_montgomerySquare();
823
824 case vmIntrinsics::_vectorizedMismatch:
825 return inline_vectorizedMismatch();
826
827 case vmIntrinsics::_ghash_processBlocks:
828 return inline_ghash_processBlocks();
829
830 case vmIntrinsics::_encodeISOArray:
831 case vmIntrinsics::_encodeByteISOArray:
832 return inline_encodeISOArray();
833
834 case vmIntrinsics::_updateCRC32:
835 return inline_updateCRC32();
836 case vmIntrinsics::_updateBytesCRC32:
837 return inline_updateBytesCRC32();
838 case vmIntrinsics::_updateByteBufferCRC32:
839 return inline_updateByteBufferCRC32();
840
841 case vmIntrinsics::_updateBytesCRC32C:
842 return inline_updateBytesCRC32C();
843 case vmIntrinsics::_updateDirectByteBufferCRC32C:
844 return inline_updateDirectByteBufferCRC32C();
845
846 case vmIntrinsics::_updateBytesAdler32:
847 return inline_updateBytesAdler32();
848 case vmIntrinsics::_updateByteBufferAdler32:
849 return inline_updateByteBufferAdler32();
850
851 case vmIntrinsics::_profileBoolean:
852 return inline_profileBoolean();
853 case vmIntrinsics::_isCompileConstant:
854 return inline_isCompileConstant();
855
856 case vmIntrinsics::_hasNegatives:
857 return inline_hasNegatives();
858
859 case vmIntrinsics::_fmaD:
860 case vmIntrinsics::_fmaF:
861 return inline_fma(intrinsic_id());
862
863 default:
864 // If you get here, it may be that someone has added a new intrinsic
865 // to the list in vmSymbols.hpp without implementing it here.
866 #ifndef PRODUCT
867 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
868 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
869 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
870 }
871 #endif
872 return false;
873 }
874 }
875
876 Node* LibraryCallKit::try_to_predicate(int predicate) {
877 if (!jvms()->has_method()) {
878 // Root JVMState has a null method.
879 assert(map()->memory()->Opcode() == Op_Parm, "");
880 // Insert the memory aliasing node
881 set_all_memory(reset_memory());
882 }
883 assert(merged_memory(), "");
884
885 switch (intrinsic_id()) {
886 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
887 return inline_cipherBlockChaining_AESCrypt_predicate(false);
888 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
889 return inline_cipherBlockChaining_AESCrypt_predicate(true);
890 case vmIntrinsics::_counterMode_AESCrypt:
891 return inline_counterMode_AESCrypt_predicate();
892 case vmIntrinsics::_digestBase_implCompressMB:
893 return inline_digestBase_implCompressMB_predicate(predicate);
894
895 default:
896 // If you get here, it may be that someone has added a new intrinsic
897 // to the list in vmSymbols.hpp without implementing it here.
898 #ifndef PRODUCT
899 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
900 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
901 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
902 }
903 #endif
904 Node* slow_ctl = control();
905 set_control(top()); // No fast path instrinsic
906 return slow_ctl;
907 }
908 }
909
910 //------------------------------set_result-------------------------------
911 // Helper function for finishing intrinsics.
912 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
913 record_for_igvn(region);
914 set_control(_gvn.transform(region));
915 set_result( _gvn.transform(value));
916 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
917 }
918
919 //------------------------------generate_guard---------------------------
920 // Helper function for generating guarded fast-slow graph structures.
921 // The given 'test', if true, guards a slow path. If the test fails
922 // then a fast path can be taken. (We generally hope it fails.)
923 // In all cases, GraphKit::control() is updated to the fast path.
924 // The returned value represents the control for the slow path.
925 // The return value is never 'top'; it is either a valid control
926 // or NULL if it is obvious that the slow path can never be taken.
927 // Also, if region and the slow control are not NULL, the slow edge
928 // is appended to the region.
929 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
930 if (stopped()) {
931 // Already short circuited.
932 return NULL;
933 }
934
935 // Build an if node and its projections.
936 // If test is true we take the slow path, which we assume is uncommon.
937 if (_gvn.type(test) == TypeInt::ZERO) {
938 // The slow branch is never taken. No need to build this guard.
939 return NULL;
940 }
941
942 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
943
944 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
945 if (if_slow == top()) {
946 // The slow branch is never taken. No need to build this guard.
947 return NULL;
948 }
949
950 if (region != NULL)
951 region->add_req(if_slow);
952
953 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
954 set_control(if_fast);
955
956 return if_slow;
957 }
958
959 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
960 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
961 }
962 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
963 return generate_guard(test, region, PROB_FAIR);
964 }
965
966 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
967 Node* *pos_index) {
968 if (stopped())
969 return NULL; // already stopped
970 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
971 return NULL; // index is already adequately typed
972 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
973 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
974 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
975 if (is_neg != NULL && pos_index != NULL) {
976 // Emulate effect of Parse::adjust_map_after_if.
977 Node* ccast = new CastIINode(index, TypeInt::POS);
978 ccast->set_req(0, control());
979 (*pos_index) = _gvn.transform(ccast);
980 }
981 return is_neg;
982 }
983
984 // Make sure that 'position' is a valid limit index, in [0..length].
985 // There are two equivalent plans for checking this:
986 // A. (offset + copyLength) unsigned<= arrayLength
987 // B. offset <= (arrayLength - copyLength)
988 // We require that all of the values above, except for the sum and
989 // difference, are already known to be non-negative.
990 // Plan A is robust in the face of overflow, if offset and copyLength
991 // are both hugely positive.
992 //
993 // Plan B is less direct and intuitive, but it does not overflow at
994 // all, since the difference of two non-negatives is always
995 // representable. Whenever Java methods must perform the equivalent
996 // check they generally use Plan B instead of Plan A.
997 // For the moment we use Plan A.
998 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
999 Node* subseq_length,
1000 Node* array_length,
1001 RegionNode* region) {
1002 if (stopped())
1003 return NULL; // already stopped
1004 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1005 if (zero_offset && subseq_length->eqv_uncast(array_length))
1006 return NULL; // common case of whole-array copy
1007 Node* last = subseq_length;
1008 if (!zero_offset) // last += offset
1009 last = _gvn.transform(new AddINode(last, offset));
1010 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1011 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1012 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1013 return is_over;
1014 }
1015
1016 // Emit range checks for the given String.value byte array
1017 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1018 if (stopped()) {
1019 return; // already stopped
1020 }
1021 RegionNode* bailout = new RegionNode(1);
1022 record_for_igvn(bailout);
1023 if (char_count) {
1024 // Convert char count to byte count
1025 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1026 }
1027
1028 // Offset and count must not be negative
1029 generate_negative_guard(offset, bailout);
1030 generate_negative_guard(count, bailout);
1031 // Offset + count must not exceed length of array
1032 generate_limit_guard(offset, count, load_array_length(array), 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 }
1041
1042 //--------------------------generate_current_thread--------------------
1043 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1044 ciKlass* thread_klass = env()->Thread_klass();
1045 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1046 Node* thread = _gvn.transform(new ThreadLocalNode());
1047 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1048 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1049 tls_output = thread;
1050 return threadObj;
1051 }
1052
1053
1054 //------------------------------make_string_method_node------------------------
1055 // Helper method for String intrinsic functions. This version is called with
1056 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1057 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1058 // containing the lengths of str1 and str2.
1059 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1060 Node* result = NULL;
1061 switch (opcode) {
1062 case Op_StrIndexOf:
1063 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1064 str1_start, cnt1, str2_start, cnt2, ae);
1065 break;
1066 case Op_StrComp:
1067 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1068 str1_start, cnt1, str2_start, cnt2, ae);
1069 break;
1070 case Op_StrEquals:
1071 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1072 // Use the constant length if there is one because optimized match rule may exist.
1073 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1074 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1075 break;
1076 default:
1077 ShouldNotReachHere();
1078 return NULL;
1079 }
1080
1081 // All these intrinsics have checks.
1082 C->set_has_split_ifs(true); // Has chance for split-if optimization
1083
1084 return _gvn.transform(result);
1085 }
1086
1087 //------------------------------inline_string_compareTo------------------------
1088 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1089 Node* arg1 = argument(0);
1090 Node* arg2 = argument(1);
1091
1092 // Get start addr and length of first argument
1093 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1094 Node* arg1_cnt = load_array_length(arg1);
1095
1096 // Get start addr and length of second argument
1097 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1098 Node* arg2_cnt = load_array_length(arg2);
1099
1100 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1101 set_result(result);
1102 return true;
1103 }
1104
1105 //------------------------------inline_string_equals------------------------
1106 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1107 Node* arg1 = argument(0);
1108 Node* arg2 = argument(1);
1109
1110 // paths (plus control) merge
1111 RegionNode* region = new RegionNode(3);
1112 Node* phi = new PhiNode(region, TypeInt::BOOL);
1113
1114 if (!stopped()) {
1115 // Get start addr and length of first argument
1116 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1117 Node* arg1_cnt = load_array_length(arg1);
1118
1119 // Get start addr and length of second argument
1120 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1121 Node* arg2_cnt = load_array_length(arg2);
1122
1123 // Check for arg1_cnt != arg2_cnt
1124 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1125 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1126 Node* if_ne = generate_slow_guard(bol, NULL);
1127 if (if_ne != NULL) {
1128 phi->init_req(2, intcon(0));
1129 region->init_req(2, if_ne);
1130 }
1131
1132 // Check for count == 0 is done by assembler code for StrEquals.
1133
1134 if (!stopped()) {
1135 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1136 phi->init_req(1, equals);
1137 region->init_req(1, control());
1138 }
1139 }
1140
1141 // post merge
1142 set_control(_gvn.transform(region));
1143 record_for_igvn(region);
1144
1145 set_result(_gvn.transform(phi));
1146 return true;
1147 }
1148
1149 //------------------------------inline_array_equals----------------------------
1150 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1151 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1152 Node* arg1 = argument(0);
1153 Node* arg2 = argument(1);
1154
1155 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1156 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1157 return true;
1158 }
1159
1160 //------------------------------inline_hasNegatives------------------------------
1161 bool LibraryCallKit::inline_hasNegatives() {
1162 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1163 return false;
1164 }
1165
1166 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1167 // no receiver since it is static method
1168 Node* ba = argument(0);
1169 Node* offset = argument(1);
1170 Node* len = argument(2);
1171
1172 // Range checks
1173 generate_string_range_check(ba, offset, len, false);
1174 if (stopped()) {
1175 return true;
1176 }
1177 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1178 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1179 set_result(_gvn.transform(result));
1180 return true;
1181 }
1182
1183 bool LibraryCallKit::inline_preconditions_checkIndex() {
1184 Node* index = argument(0);
1185 Node* length = argument(1);
1186 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1187 return false;
1188 }
1189
1190 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1191 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1192
1193 {
1194 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1195 uncommon_trap(Deoptimization::Reason_intrinsic,
1196 Deoptimization::Action_make_not_entrant);
1197 }
1198
1199 if (stopped()) {
1200 return false;
1201 }
1202
1203 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1204 BoolTest::mask btest = BoolTest::lt;
1205 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1206 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1207 _gvn.set_type(rc, rc->Value(&_gvn));
1208 if (!rc_bool->is_Con()) {
1209 record_for_igvn(rc);
1210 }
1211 set_control(_gvn.transform(new IfTrueNode(rc)));
1212 {
1213 PreserveJVMState pjvms(this);
1214 set_control(_gvn.transform(new IfFalseNode(rc)));
1215 uncommon_trap(Deoptimization::Reason_range_check,
1216 Deoptimization::Action_make_not_entrant);
1217 }
1218
1219 if (stopped()) {
1220 return false;
1221 }
1222
1223 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1224 result->set_req(0, control());
1225 result = _gvn.transform(result);
1226 set_result(result);
1227 replace_in_map(index, result);
1228 return true;
1229 }
1230
1231 //------------------------------inline_string_indexOf------------------------
1232 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1233 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1234 return false;
1235 }
1236 Node* src = argument(0);
1237 Node* tgt = argument(1);
1238
1239 // Make the merge point
1240 RegionNode* result_rgn = new RegionNode(4);
1241 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1242
1243 // Get start addr and length of source string
1244 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1245 Node* src_count = load_array_length(src);
1246
1247 // Get start addr and length of substring
1248 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1249 Node* tgt_count = load_array_length(tgt);
1250
1251 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1252 // Divide src size by 2 if String is UTF16 encoded
1253 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1254 }
1255 if (ae == StrIntrinsicNode::UU) {
1256 // Divide substring size by 2 if String is UTF16 encoded
1257 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1258 }
1259
1260 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1261 if (result != NULL) {
1262 result_phi->init_req(3, result);
1263 result_rgn->init_req(3, control());
1264 }
1265 set_control(_gvn.transform(result_rgn));
1266 record_for_igvn(result_rgn);
1267 set_result(_gvn.transform(result_phi));
1268
1269 return true;
1270 }
1271
1272 //-----------------------------inline_string_indexOf-----------------------
1273 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1274 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1275 return false;
1276 }
1277 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1278 return false;
1279 }
1280 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1281 Node* src = argument(0); // byte[]
1282 Node* src_count = argument(1); // char count
1283 Node* tgt = argument(2); // byte[]
1284 Node* tgt_count = argument(3); // char count
1285 Node* from_index = argument(4); // char index
1286
1287 // Multiply byte array index by 2 if String is UTF16 encoded
1288 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1289 src_count = _gvn.transform(new SubINode(src_count, from_index));
1290 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1291 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1292
1293 // Range checks
1294 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1295 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1296 if (stopped()) {
1297 return true;
1298 }
1299
1300 RegionNode* region = new RegionNode(5);
1301 Node* phi = new PhiNode(region, TypeInt::INT);
1302
1303 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1304 if (result != NULL) {
1305 // The result is index relative to from_index if substring was found, -1 otherwise.
1306 // Generate code which will fold into cmove.
1307 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1308 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1309
1310 Node* if_lt = generate_slow_guard(bol, NULL);
1311 if (if_lt != NULL) {
1312 // result == -1
1313 phi->init_req(3, result);
1314 region->init_req(3, if_lt);
1315 }
1316 if (!stopped()) {
1317 result = _gvn.transform(new AddINode(result, from_index));
1318 phi->init_req(4, result);
1319 region->init_req(4, control());
1320 }
1321 }
1322
1323 set_control(_gvn.transform(region));
1324 record_for_igvn(region);
1325 set_result(_gvn.transform(phi));
1326
1327 return true;
1328 }
1329
1330 // Create StrIndexOfNode with fast path checks
1331 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1332 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1333 // Check for substr count > string count
1334 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1335 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1336 Node* if_gt = generate_slow_guard(bol, NULL);
1337 if (if_gt != NULL) {
1338 phi->init_req(1, intcon(-1));
1339 region->init_req(1, if_gt);
1340 }
1341 if (!stopped()) {
1342 // Check for substr count == 0
1343 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1344 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1345 Node* if_zero = generate_slow_guard(bol, NULL);
1346 if (if_zero != NULL) {
1347 phi->init_req(2, intcon(0));
1348 region->init_req(2, if_zero);
1349 }
1350 }
1351 if (!stopped()) {
1352 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1353 }
1354 return NULL;
1355 }
1356
1357 //-----------------------------inline_string_indexOfChar-----------------------
1358 bool LibraryCallKit::inline_string_indexOfChar() {
1359 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1360 return false;
1361 }
1362 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1363 return false;
1364 }
1365 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1366 Node* src = argument(0); // byte[]
1367 Node* tgt = argument(1); // tgt is int ch
1368 Node* from_index = argument(2);
1369 Node* max = argument(3);
1370
1371 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1372 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1373 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1374
1375 // Range checks
1376 generate_string_range_check(src, src_offset, src_count, true);
1377 if (stopped()) {
1378 return true;
1379 }
1380
1381 RegionNode* region = new RegionNode(3);
1382 Node* phi = new PhiNode(region, TypeInt::INT);
1383
1384 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1385 C->set_has_split_ifs(true); // Has chance for split-if optimization
1386 _gvn.transform(result);
1387
1388 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1389 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1390
1391 Node* if_lt = generate_slow_guard(bol, NULL);
1392 if (if_lt != NULL) {
1393 // result == -1
1394 phi->init_req(2, result);
1395 region->init_req(2, if_lt);
1396 }
1397 if (!stopped()) {
1398 result = _gvn.transform(new AddINode(result, from_index));
1399 phi->init_req(1, result);
1400 region->init_req(1, control());
1401 }
1402 set_control(_gvn.transform(region));
1403 record_for_igvn(region);
1404 set_result(_gvn.transform(phi));
1405
1406 return true;
1407 }
1408 //---------------------------inline_string_copy---------------------
1409 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1410 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1411 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1412 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1413 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1414 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1415 bool LibraryCallKit::inline_string_copy(bool compress) {
1416 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1417 return false;
1418 }
1419 int nargs = 5; // 2 oops, 3 ints
1420 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1421
1422 Node* src = argument(0);
1423 Node* src_offset = argument(1);
1424 Node* dst = argument(2);
1425 Node* dst_offset = argument(3);
1426 Node* length = argument(4);
1427
1428 // Check for allocation before we add nodes that would confuse
1429 // tightly_coupled_allocation()
1430 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1431
1432 // Figure out the size and type of the elements we will be copying.
1433 const Type* src_type = src->Value(&_gvn);
1434 const Type* dst_type = dst->Value(&_gvn);
1435 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1436 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1437 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1438 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1439 "Unsupported array types for inline_string_copy");
1440
1441 // Convert char[] offsets to byte[] offsets
1442 bool convert_src = (compress && src_elem == T_BYTE);
1443 bool convert_dst = (!compress && dst_elem == T_BYTE);
1444 if (convert_src) {
1445 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1446 } else if (convert_dst) {
1447 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1448 }
1449
1450 // Range checks
1451 generate_string_range_check(src, src_offset, length, convert_src);
1452 generate_string_range_check(dst, dst_offset, length, convert_dst);
1453 if (stopped()) {
1454 return true;
1455 }
1456
1457 Node* src_start = array_element_address(src, src_offset, src_elem);
1458 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1459 // 'src_start' points to src array + scaled offset
1460 // 'dst_start' points to dst array + scaled offset
1461 Node* count = NULL;
1462 if (compress) {
1463 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1464 } else {
1465 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1466 }
1467
1468 if (alloc != NULL) {
1469 if (alloc->maybe_set_complete(&_gvn)) {
1470 // "You break it, you buy it."
1471 InitializeNode* init = alloc->initialization();
1472 assert(init->is_complete(), "we just did this");
1473 init->set_complete_with_arraycopy();
1474 assert(dst->is_CheckCastPP(), "sanity");
1475 assert(dst->in(0)->in(0) == init, "dest pinned");
1476 }
1477 // Do not let stores that initialize this object be reordered with
1478 // a subsequent store that would make this object accessible by
1479 // other threads.
1480 // Record what AllocateNode this StoreStore protects so that
1481 // escape analysis can go from the MemBarStoreStoreNode to the
1482 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1483 // based on the escape status of the AllocateNode.
1484 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1485 }
1486 if (compress) {
1487 set_result(_gvn.transform(count));
1488 }
1489 return true;
1490 }
1491
1492 #ifdef _LP64
1493 #define XTOP ,top() /*additional argument*/
1494 #else //_LP64
1495 #define XTOP /*no additional argument*/
1496 #endif //_LP64
1497
1498 //------------------------inline_string_toBytesU--------------------------
1499 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1500 bool LibraryCallKit::inline_string_toBytesU() {
1501 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1502 return false;
1503 }
1504 // Get the arguments.
1505 Node* value = argument(0);
1506 Node* offset = argument(1);
1507 Node* length = argument(2);
1508
1509 Node* newcopy = NULL;
1510
1511 // Set the original stack and the reexecute bit for the interpreter to reexecute
1512 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1513 { PreserveReexecuteState preexecs(this);
1514 jvms()->set_should_reexecute(true);
1515
1516 // Check if a null path was taken unconditionally.
1517 value = null_check(value);
1518
1519 RegionNode* bailout = new RegionNode(1);
1520 record_for_igvn(bailout);
1521
1522 // Range checks
1523 generate_negative_guard(offset, bailout);
1524 generate_negative_guard(length, bailout);
1525 generate_limit_guard(offset, length, load_array_length(value), bailout);
1526 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1527 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1528
1529 if (bailout->req() > 1) {
1530 PreserveJVMState pjvms(this);
1531 set_control(_gvn.transform(bailout));
1532 uncommon_trap(Deoptimization::Reason_intrinsic,
1533 Deoptimization::Action_maybe_recompile);
1534 }
1535 if (stopped()) {
1536 return true;
1537 }
1538
1539 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1540 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1541 newcopy = new_array(klass_node, size, 0); // no arguments to push
1542 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1543
1544 // Calculate starting addresses.
1545 Node* src_start = array_element_address(value, offset, T_CHAR);
1546 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1547
1548 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1549 const TypeInt* toffset = gvn().type(offset)->is_int();
1550 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1551
1552 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1553 const char* copyfunc_name = "arraycopy";
1554 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1555 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1556 OptoRuntime::fast_arraycopy_Type(),
1557 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1558 src_start, dst_start, ConvI2X(length) XTOP);
1559 // Do not let reads from the cloned object float above the arraycopy.
1560 if (alloc != NULL) {
1561 if (alloc->maybe_set_complete(&_gvn)) {
1562 // "You break it, you buy it."
1563 InitializeNode* init = alloc->initialization();
1564 assert(init->is_complete(), "we just did this");
1565 init->set_complete_with_arraycopy();
1566 assert(newcopy->is_CheckCastPP(), "sanity");
1567 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1568 }
1569 // Do not let stores that initialize this object be reordered with
1570 // a subsequent store that would make this object accessible by
1571 // other threads.
1572 // Record what AllocateNode this StoreStore protects so that
1573 // escape analysis can go from the MemBarStoreStoreNode to the
1574 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1575 // based on the escape status of the AllocateNode.
1576 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1577 } else {
1578 insert_mem_bar(Op_MemBarCPUOrder);
1579 }
1580 } // original reexecute is set back here
1581
1582 C->set_has_split_ifs(true); // Has chance for split-if optimization
1583 if (!stopped()) {
1584 set_result(newcopy);
1585 }
1586 return true;
1587 }
1588
1589 //------------------------inline_string_getCharsU--------------------------
1590 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1591 bool LibraryCallKit::inline_string_getCharsU() {
1592 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1593 return false;
1594 }
1595
1596 // Get the arguments.
1597 Node* src = argument(0);
1598 Node* src_begin = argument(1);
1599 Node* src_end = argument(2); // exclusive offset (i < src_end)
1600 Node* dst = argument(3);
1601 Node* dst_begin = argument(4);
1602
1603 // Check for allocation before we add nodes that would confuse
1604 // tightly_coupled_allocation()
1605 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1606
1607 // Check if a null path was taken unconditionally.
1608 src = null_check(src);
1609 dst = null_check(dst);
1610 if (stopped()) {
1611 return true;
1612 }
1613
1614 // Get length and convert char[] offset to byte[] offset
1615 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1616 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1617
1618 // Range checks
1619 generate_string_range_check(src, src_begin, length, true);
1620 generate_string_range_check(dst, dst_begin, length, false);
1621 if (stopped()) {
1622 return true;
1623 }
1624
1625 if (!stopped()) {
1626 // Calculate starting addresses.
1627 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1628 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1629
1630 // Check if array addresses are aligned to HeapWordSize
1631 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1632 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1633 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1634 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1635
1636 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1637 const char* copyfunc_name = "arraycopy";
1638 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1639 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1640 OptoRuntime::fast_arraycopy_Type(),
1641 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1642 src_start, dst_start, ConvI2X(length) XTOP);
1643 // Do not let reads from the cloned object float above the arraycopy.
1644 if (alloc != NULL) {
1645 if (alloc->maybe_set_complete(&_gvn)) {
1646 // "You break it, you buy it."
1647 InitializeNode* init = alloc->initialization();
1648 assert(init->is_complete(), "we just did this");
1649 init->set_complete_with_arraycopy();
1650 assert(dst->is_CheckCastPP(), "sanity");
1651 assert(dst->in(0)->in(0) == init, "dest pinned");
1652 }
1653 // Do not let stores that initialize this object be reordered with
1654 // a subsequent store that would make this object accessible by
1655 // other threads.
1656 // Record what AllocateNode this StoreStore protects so that
1657 // escape analysis can go from the MemBarStoreStoreNode to the
1658 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1659 // based on the escape status of the AllocateNode.
1660 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1661 } else {
1662 insert_mem_bar(Op_MemBarCPUOrder);
1663 }
1664 }
1665
1666 C->set_has_split_ifs(true); // Has chance for split-if optimization
1667 return true;
1668 }
1669
1670 //----------------------inline_string_char_access----------------------------
1671 // Store/Load char to/from byte[] array.
1672 // static void StringUTF16.putChar(byte[] val, int index, int c)
1673 // static char StringUTF16.getChar(byte[] val, int index)
1674 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1675 Node* value = argument(0);
1676 Node* index = argument(1);
1677 Node* ch = is_store ? argument(2) : NULL;
1678
1679 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1680 // correctly requires matched array shapes.
1681 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1682 "sanity: byte[] and char[] bases agree");
1683 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1684 "sanity: byte[] and char[] scales agree");
1685
1686 // Bail when getChar over constants is requested: constant folding would
1687 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1688 // Java method would constant fold nicely instead.
1689 if (!is_store && value->is_Con() && index->is_Con()) {
1690 return false;
1691 }
1692
1693 Node* adr = array_element_address(value, index, T_CHAR);
1694 if (adr->is_top()) {
1695 return false;
1696 }
1697 if (is_store) {
1698 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1699 false, false, true /* mismatched */);
1700 } else {
1701 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1702 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1703 set_result(ch);
1704 }
1705 return true;
1706 }
1707
1708 //--------------------------round_double_node--------------------------------
1709 // Round a double node if necessary.
1710 Node* LibraryCallKit::round_double_node(Node* n) {
1711 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1712 n = _gvn.transform(new RoundDoubleNode(0, n));
1713 return n;
1714 }
1715
1716 //------------------------------inline_math-----------------------------------
1717 // public static double Math.abs(double)
1718 // public static double Math.sqrt(double)
1719 // public static double Math.log(double)
1720 // public static double Math.log10(double)
1721 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1722 Node* arg = round_double_node(argument(0));
1723 Node* n = NULL;
1724 switch (id) {
1725 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1726 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1727 default: fatal_unexpected_iid(id); break;
1728 }
1729 set_result(_gvn.transform(n));
1730 return true;
1731 }
1732
1733 //------------------------------runtime_math-----------------------------
1734 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1735 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1736 "must be (DD)D or (D)D type");
1737
1738 // Inputs
1739 Node* a = round_double_node(argument(0));
1740 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1741
1742 const TypePtr* no_memory_effects = NULL;
1743 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1744 no_memory_effects,
1745 a, top(), b, b ? top() : NULL);
1746 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1747 #ifdef ASSERT
1748 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1749 assert(value_top == top(), "second value must be top");
1750 #endif
1751
1752 set_result(value);
1753 return true;
1754 }
1755
1756 //------------------------------inline_math_native-----------------------------
1757 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1758 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1759 switch (id) {
1760 // These intrinsics are not properly supported on all hardware
1761 case vmIntrinsics::_dsin:
1762 return StubRoutines::dsin() != NULL ?
1763 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1764 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1765 case vmIntrinsics::_dcos:
1766 return StubRoutines::dcos() != NULL ?
1767 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1768 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1769 case vmIntrinsics::_dtan:
1770 return StubRoutines::dtan() != NULL ?
1771 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1772 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1773 case vmIntrinsics::_dlog:
1774 return StubRoutines::dlog() != NULL ?
1775 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1776 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1777 case vmIntrinsics::_dlog10:
1778 return StubRoutines::dlog10() != NULL ?
1779 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1780 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1781
1782 // These intrinsics are supported on all hardware
1783 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1784 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1785
1786 case vmIntrinsics::_dexp:
1787 return StubRoutines::dexp() != NULL ?
1788 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1789 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1790 case vmIntrinsics::_dpow:
1791 return StubRoutines::dpow() != NULL ?
1792 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1793 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1794 #undef FN_PTR
1795
1796 // These intrinsics are not yet correctly implemented
1797 case vmIntrinsics::_datan2:
1798 return false;
1799
1800 default:
1801 fatal_unexpected_iid(id);
1802 return false;
1803 }
1804 }
1805
1806 static bool is_simple_name(Node* n) {
1807 return (n->req() == 1 // constant
1808 || (n->is_Type() && n->as_Type()->type()->singleton())
1809 || n->is_Proj() // parameter or return value
1810 || n->is_Phi() // local of some sort
1811 );
1812 }
1813
1814 //----------------------------inline_notify-----------------------------------*
1815 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1816 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1817 address func;
1818 if (id == vmIntrinsics::_notify) {
1819 func = OptoRuntime::monitor_notify_Java();
1820 } else {
1821 func = OptoRuntime::monitor_notifyAll_Java();
1822 }
1823 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1824 make_slow_call_ex(call, env()->Throwable_klass(), false);
1825 return true;
1826 }
1827
1828
1829 //----------------------------inline_min_max-----------------------------------
1830 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1831 set_result(generate_min_max(id, argument(0), argument(1)));
1832 return true;
1833 }
1834
1835 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1836 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1837 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1838 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1839 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1840
1841 {
1842 PreserveJVMState pjvms(this);
1843 PreserveReexecuteState preexecs(this);
1844 jvms()->set_should_reexecute(true);
1845
1846 set_control(slow_path);
1847 set_i_o(i_o());
1848
1849 uncommon_trap(Deoptimization::Reason_intrinsic,
1850 Deoptimization::Action_none);
1851 }
1852
1853 set_control(fast_path);
1854 set_result(math);
1855 }
1856
1857 template <typename OverflowOp>
1858 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1859 typedef typename OverflowOp::MathOp MathOp;
1860
1861 MathOp* mathOp = new MathOp(arg1, arg2);
1862 Node* operation = _gvn.transform( mathOp );
1863 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1864 inline_math_mathExact(operation, ofcheck);
1865 return true;
1866 }
1867
1868 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1869 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1870 }
1871
1872 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1873 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1874 }
1875
1876 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1877 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1878 }
1879
1880 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1881 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1882 }
1883
1884 bool LibraryCallKit::inline_math_negateExactI() {
1885 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1886 }
1887
1888 bool LibraryCallKit::inline_math_negateExactL() {
1889 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1890 }
1891
1892 bool LibraryCallKit::inline_math_multiplyExactI() {
1893 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1894 }
1895
1896 bool LibraryCallKit::inline_math_multiplyExactL() {
1897 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1898 }
1899
1900 Node*
1901 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1902 // These are the candidate return value:
1903 Node* xvalue = x0;
1904 Node* yvalue = y0;
1905
1906 if (xvalue == yvalue) {
1907 return xvalue;
1908 }
1909
1910 bool want_max = (id == vmIntrinsics::_max);
1911
1912 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1913 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1914 if (txvalue == NULL || tyvalue == NULL) return top();
1915 // This is not really necessary, but it is consistent with a
1916 // hypothetical MaxINode::Value method:
1917 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1918
1919 // %%% This folding logic should (ideally) be in a different place.
1920 // Some should be inside IfNode, and there to be a more reliable
1921 // transformation of ?: style patterns into cmoves. We also want
1922 // more powerful optimizations around cmove and min/max.
1923
1924 // Try to find a dominating comparison of these guys.
1925 // It can simplify the index computation for Arrays.copyOf
1926 // and similar uses of System.arraycopy.
1927 // First, compute the normalized version of CmpI(x, y).
1928 int cmp_op = Op_CmpI;
1929 Node* xkey = xvalue;
1930 Node* ykey = yvalue;
1931 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1932 if (ideal_cmpxy->is_Cmp()) {
1933 // E.g., if we have CmpI(length - offset, count),
1934 // it might idealize to CmpI(length, count + offset)
1935 cmp_op = ideal_cmpxy->Opcode();
1936 xkey = ideal_cmpxy->in(1);
1937 ykey = ideal_cmpxy->in(2);
1938 }
1939
1940 // Start by locating any relevant comparisons.
1941 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1942 Node* cmpxy = NULL;
1943 Node* cmpyx = NULL;
1944 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1945 Node* cmp = start_from->fast_out(k);
1946 if (cmp->outcnt() > 0 && // must have prior uses
1947 cmp->in(0) == NULL && // must be context-independent
1948 cmp->Opcode() == cmp_op) { // right kind of compare
1949 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1950 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1951 }
1952 }
1953
1954 const int NCMPS = 2;
1955 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1956 int cmpn;
1957 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1958 if (cmps[cmpn] != NULL) break; // find a result
1959 }
1960 if (cmpn < NCMPS) {
1961 // Look for a dominating test that tells us the min and max.
1962 int depth = 0; // Limit search depth for speed
1963 Node* dom = control();
1964 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1965 if (++depth >= 100) break;
1966 Node* ifproj = dom;
1967 if (!ifproj->is_Proj()) continue;
1968 Node* iff = ifproj->in(0);
1969 if (!iff->is_If()) continue;
1970 Node* bol = iff->in(1);
1971 if (!bol->is_Bool()) continue;
1972 Node* cmp = bol->in(1);
1973 if (cmp == NULL) continue;
1974 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1975 if (cmps[cmpn] == cmp) break;
1976 if (cmpn == NCMPS) continue;
1977 BoolTest::mask btest = bol->as_Bool()->_test._test;
1978 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1979 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1980 // At this point, we know that 'x btest y' is true.
1981 switch (btest) {
1982 case BoolTest::eq:
1983 // They are proven equal, so we can collapse the min/max.
1984 // Either value is the answer. Choose the simpler.
1985 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1986 return yvalue;
1987 return xvalue;
1988 case BoolTest::lt: // x < y
1989 case BoolTest::le: // x <= y
1990 return (want_max ? yvalue : xvalue);
1991 case BoolTest::gt: // x > y
1992 case BoolTest::ge: // x >= y
1993 return (want_max ? xvalue : yvalue);
1994 }
1995 }
1996 }
1997
1998 // We failed to find a dominating test.
1999 // Let's pick a test that might GVN with prior tests.
2000 Node* best_bol = NULL;
2001 BoolTest::mask best_btest = BoolTest::illegal;
2002 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2003 Node* cmp = cmps[cmpn];
2004 if (cmp == NULL) continue;
2005 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2006 Node* bol = cmp->fast_out(j);
2007 if (!bol->is_Bool()) continue;
2008 BoolTest::mask btest = bol->as_Bool()->_test._test;
2009 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2010 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2011 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2012 best_bol = bol->as_Bool();
2013 best_btest = btest;
2014 }
2015 }
2016 }
2017
2018 Node* answer_if_true = NULL;
2019 Node* answer_if_false = NULL;
2020 switch (best_btest) {
2021 default:
2022 if (cmpxy == NULL)
2023 cmpxy = ideal_cmpxy;
2024 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2025 // and fall through:
2026 case BoolTest::lt: // x < y
2027 case BoolTest::le: // x <= y
2028 answer_if_true = (want_max ? yvalue : xvalue);
2029 answer_if_false = (want_max ? xvalue : yvalue);
2030 break;
2031 case BoolTest::gt: // x > y
2032 case BoolTest::ge: // x >= y
2033 answer_if_true = (want_max ? xvalue : yvalue);
2034 answer_if_false = (want_max ? yvalue : xvalue);
2035 break;
2036 }
2037
2038 jint hi, lo;
2039 if (want_max) {
2040 // We can sharpen the minimum.
2041 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2042 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2043 } else {
2044 // We can sharpen the maximum.
2045 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2046 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2047 }
2048
2049 // Use a flow-free graph structure, to avoid creating excess control edges
2050 // which could hinder other optimizations.
2051 // Since Math.min/max is often used with arraycopy, we want
2052 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2053 Node* cmov = CMoveNode::make(NULL, best_bol,
2054 answer_if_false, answer_if_true,
2055 TypeInt::make(lo, hi, widen));
2056
2057 return _gvn.transform(cmov);
2058
2059 /*
2060 // This is not as desirable as it may seem, since Min and Max
2061 // nodes do not have a full set of optimizations.
2062 // And they would interfere, anyway, with 'if' optimizations
2063 // and with CMoveI canonical forms.
2064 switch (id) {
2065 case vmIntrinsics::_min:
2066 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2067 case vmIntrinsics::_max:
2068 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2069 default:
2070 ShouldNotReachHere();
2071 }
2072 */
2073 }
2074
2075 inline int
2076 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2077 const TypePtr* base_type = TypePtr::NULL_PTR;
2078 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2079 if (base_type == NULL) {
2080 // Unknown type.
2081 return Type::AnyPtr;
2082 } else if (base_type == TypePtr::NULL_PTR) {
2083 // Since this is a NULL+long form, we have to switch to a rawptr.
2084 base = _gvn.transform(new CastX2PNode(offset));
2085 offset = MakeConX(0);
2086 return Type::RawPtr;
2087 } else if (base_type->base() == Type::RawPtr) {
2088 return Type::RawPtr;
2089 } else if (base_type->isa_oopptr()) {
2090 // Base is never null => always a heap address.
2091 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2092 return Type::OopPtr;
2093 }
2094 // Offset is small => always a heap address.
2095 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2096 if (offset_type != NULL &&
2097 base_type->offset() == 0 && // (should always be?)
2098 offset_type->_lo >= 0 &&
2099 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2100 return Type::OopPtr;
2101 } else if (type == T_OBJECT) {
2102 // off heap access to an oop doesn't make any sense. Has to be on
2103 // heap.
2104 return Type::OopPtr;
2105 }
2106 // Otherwise, it might either be oop+off or NULL+addr.
2107 return Type::AnyPtr;
2108 } else {
2109 // No information:
2110 return Type::AnyPtr;
2111 }
2112 }
2113
2114 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2115 Node* uncasted_base = base;
2116 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2117 if (kind == Type::RawPtr) {
2118 return basic_plus_adr(top(), uncasted_base, offset);
2119 } else if (kind == Type::AnyPtr) {
2120 assert(base == uncasted_base, "unexpected base change");
2121 if (can_cast) {
2122 if (!_gvn.type(base)->speculative_maybe_null() &&
2123 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2124 // According to profiling, this access is always on
2125 // heap. Casting the base to not null and thus avoiding membars
2126 // around the access should allow better optimizations
2127 Node* null_ctl = top();
2128 base = null_check_oop(base, &null_ctl, true, true, true);
2129 assert(null_ctl->is_top(), "no null control here");
2130 return basic_plus_adr(base, offset);
2131 } else if (_gvn.type(base)->speculative_always_null() &&
2132 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2133 // According to profiling, this access is always off
2134 // heap.
2135 base = null_assert(base);
2136 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2137 offset = MakeConX(0);
2138 return basic_plus_adr(top(), raw_base, offset);
2139 }
2140 }
2141 // We don't know if it's an on heap or off heap access. Fall back
2142 // to raw memory access.
2143 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2144 return basic_plus_adr(top(), raw, offset);
2145 } else {
2146 assert(base == uncasted_base, "unexpected base change");
2147 // We know it's an on heap access so base can't be null
2148 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2149 base = must_be_not_null(base, true);
2150 }
2151 return basic_plus_adr(base, offset);
2152 }
2153 }
2154
2155 //--------------------------inline_number_methods-----------------------------
2156 // inline int Integer.numberOfLeadingZeros(int)
2157 // inline int Long.numberOfLeadingZeros(long)
2158 //
2159 // inline int Integer.numberOfTrailingZeros(int)
2160 // inline int Long.numberOfTrailingZeros(long)
2161 //
2162 // inline int Integer.bitCount(int)
2163 // inline int Long.bitCount(long)
2164 //
2165 // inline char Character.reverseBytes(char)
2166 // inline short Short.reverseBytes(short)
2167 // inline int Integer.reverseBytes(int)
2168 // inline long Long.reverseBytes(long)
2169 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2170 Node* arg = argument(0);
2171 Node* n = NULL;
2172 switch (id) {
2173 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2174 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2175 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2176 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2177 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2178 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2179 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2180 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2181 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2182 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2183 default: fatal_unexpected_iid(id); break;
2184 }
2185 set_result(_gvn.transform(n));
2186 return true;
2187 }
2188
2189 //----------------------------inline_unsafe_access----------------------------
2190
2191 // Helper that guards and inserts a pre-barrier.
2192 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2193 Node* pre_val, bool need_mem_bar) {
2194 // We could be accessing the referent field of a reference object. If so, when G1
2195 // is enabled, we need to log the value in the referent field in an SATB buffer.
2196 // This routine performs some compile time filters and generates suitable
2197 // runtime filters that guard the pre-barrier code.
2198 // Also add memory barrier for non volatile load from the referent field
2199 // to prevent commoning of loads across safepoint.
2200 if (!UseG1GC && !need_mem_bar)
2201 return;
2202
2203 // Some compile time checks.
2204
2205 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2206 const TypeX* otype = offset->find_intptr_t_type();
2207 if (otype != NULL && otype->is_con() &&
2208 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2209 // Constant offset but not the reference_offset so just return
2210 return;
2211 }
2212
2213 // We only need to generate the runtime guards for instances.
2214 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2215 if (btype != NULL) {
2216 if (btype->isa_aryptr()) {
2217 // Array type so nothing to do
2218 return;
2219 }
2220
2221 const TypeInstPtr* itype = btype->isa_instptr();
2222 if (itype != NULL) {
2223 // Can the klass of base_oop be statically determined to be
2224 // _not_ a sub-class of Reference and _not_ Object?
2225 ciKlass* klass = itype->klass();
2226 if ( klass->is_loaded() &&
2227 !klass->is_subtype_of(env()->Reference_klass()) &&
2228 !env()->Object_klass()->is_subtype_of(klass)) {
2229 return;
2230 }
2231 }
2232 }
2233
2234 // The compile time filters did not reject base_oop/offset so
2235 // we need to generate the following runtime filters
2236 //
2237 // if (offset == java_lang_ref_Reference::_reference_offset) {
2238 // if (instance_of(base, java.lang.ref.Reference)) {
2239 // pre_barrier(_, pre_val, ...);
2240 // }
2241 // }
2242
2243 float likely = PROB_LIKELY( 0.999);
2244 float unlikely = PROB_UNLIKELY(0.999);
2245
2246 IdealKit ideal(this);
2247 #define __ ideal.
2248
2249 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2250
2251 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2252 // Update graphKit memory and control from IdealKit.
2253 sync_kit(ideal);
2254
2255 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2256 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2257
2258 // Update IdealKit memory and control from graphKit.
2259 __ sync_kit(this);
2260
2261 Node* one = __ ConI(1);
2262 // is_instof == 0 if base_oop == NULL
2263 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2264
2265 // Update graphKit from IdeakKit.
2266 sync_kit(ideal);
2267
2268 // Use the pre-barrier to record the value in the referent field
2269 pre_barrier(false /* do_load */,
2270 __ ctrl(),
2271 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2272 pre_val /* pre_val */,
2273 T_OBJECT);
2274 if (need_mem_bar) {
2275 // Add memory barrier to prevent commoning reads from this field
2276 // across safepoint since GC can change its value.
2277 insert_mem_bar(Op_MemBarCPUOrder);
2278 }
2279 // Update IdealKit from graphKit.
2280 __ sync_kit(this);
2281
2282 } __ end_if(); // _ref_type != ref_none
2283 } __ end_if(); // offset == referent_offset
2284
2285 // Final sync IdealKit and GraphKit.
2286 final_sync(ideal);
2287 #undef __
2288 }
2289
2290
2291 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2292 // Attempt to infer a sharper value type from the offset and base type.
2293 ciKlass* sharpened_klass = NULL;
2294
2295 // See if it is an instance field, with an object type.
2296 if (alias_type->field() != NULL) {
2297 if (alias_type->field()->type()->is_klass()) {
2298 sharpened_klass = alias_type->field()->type()->as_klass();
2299 }
2300 }
2301
2302 // See if it is a narrow oop array.
2303 if (adr_type->isa_aryptr()) {
2304 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2305 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2306 if (elem_type != NULL) {
2307 sharpened_klass = elem_type->klass();
2308 }
2309 }
2310 }
2311
2312 // The sharpened class might be unloaded if there is no class loader
2313 // contraint in place.
2314 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2315 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2316
2317 #ifndef PRODUCT
2318 if (C->print_intrinsics() || C->print_inlining()) {
2319 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2320 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2321 }
2322 #endif
2323 // Sharpen the value type.
2324 return tjp;
2325 }
2326 return NULL;
2327 }
2328
2329 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2330 if (callee()->is_static()) return false; // caller must have the capability!
2331 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2332 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2333 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2334
2335 #ifndef PRODUCT
2336 {
2337 ResourceMark rm;
2338 // Check the signatures.
2339 ciSignature* sig = callee()->signature();
2340 #ifdef ASSERT
2341 if (!is_store) {
2342 // Object getObject(Object base, int/long offset), etc.
2343 BasicType rtype = sig->return_type()->basic_type();
2344 assert(rtype == type, "getter must return the expected value");
2345 assert(sig->count() == 2, "oop getter has 2 arguments");
2346 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2347 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2348 } else {
2349 // void putObject(Object base, int/long offset, Object x), etc.
2350 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2351 assert(sig->count() == 3, "oop putter has 3 arguments");
2352 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2353 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2354 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2355 assert(vtype == type, "putter must accept the expected value");
2356 }
2357 #endif // ASSERT
2358 }
2359 #endif //PRODUCT
2360
2361 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2362
2363 Node* receiver = argument(0); // type: oop
2364
2365 // Build address expression.
2366 Node* adr;
2367 Node* heap_base_oop = top();
2368 Node* offset = top();
2369 Node* val;
2370
2371 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2372 Node* base = argument(1); // type: oop
2373 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2374 offset = argument(2); // type: long
2375 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2376 // to be plain byte offsets, which are also the same as those accepted
2377 // by oopDesc::field_base.
2378 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2379 "fieldOffset must be byte-scaled");
2380 // 32-bit machines ignore the high half!
2381 offset = ConvL2X(offset);
2382 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2383
2384 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2385 heap_base_oop = base;
2386 } else if (type == T_OBJECT) {
2387 return false; // off-heap oop accesses are not supported
2388 }
2389
2390 // Can base be NULL? Otherwise, always on-heap access.
2391 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2392
2393 val = is_store ? argument(4) : NULL;
2394
2395 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2396
2397 // Try to categorize the address.
2398 Compile::AliasType* alias_type = C->alias_type(adr_type);
2399 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2400
2401 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2402 alias_type->adr_type() == TypeAryPtr::RANGE) {
2403 return false; // not supported
2404 }
2405
2406 bool mismatched = false;
2407 BasicType bt = alias_type->basic_type();
2408 if (bt != T_ILLEGAL) {
2409 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2410 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2411 // Alias type doesn't differentiate between byte[] and boolean[]).
2412 // Use address type to get the element type.
2413 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2414 }
2415 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2416 // accessing an array field with getObject is not a mismatch
2417 bt = T_OBJECT;
2418 }
2419 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2420 // Don't intrinsify mismatched object accesses
2421 return false;
2422 }
2423 mismatched = (bt != type);
2424 } else if (alias_type->adr_type()->isa_oopptr()) {
2425 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2426 }
2427
2428 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2429
2430 // First guess at the value type.
2431 const Type *value_type = Type::get_const_basic_type(type);
2432
2433 // We will need memory barriers unless we can determine a unique
2434 // alias category for this reference. (Note: If for some reason
2435 // the barriers get omitted and the unsafe reference begins to "pollute"
2436 // the alias analysis of the rest of the graph, either Compile::can_alias
2437 // or Compile::must_alias will throw a diagnostic assert.)
2438 bool need_mem_bar = false;
2439 switch (kind) {
2440 case Relaxed:
2441 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2442 break;
2443 case Opaque:
2444 // Opaque uses CPUOrder membars for protection against code movement.
2445 case Acquire:
2446 case Release:
2447 case Volatile:
2448 need_mem_bar = true;
2449 break;
2450 default:
2451 ShouldNotReachHere();
2452 }
2453
2454 // Some accesses require access atomicity for all types, notably longs and doubles.
2455 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2456 bool requires_atomic_access = false;
2457 switch (kind) {
2458 case Relaxed:
2459 requires_atomic_access = AlwaysAtomicAccesses;
2460 break;
2461 case Opaque:
2462 // Opaque accesses are atomic.
2463 case Acquire:
2464 case Release:
2465 case Volatile:
2466 requires_atomic_access = true;
2467 break;
2468 default:
2469 ShouldNotReachHere();
2470 }
2471
2472 // Figure out the memory ordering.
2473 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2474 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2475
2476 // If we are reading the value of the referent field of a Reference
2477 // object (either by using Unsafe directly or through reflection)
2478 // then, if G1 is enabled, we need to record the referent in an
2479 // SATB log buffer using the pre-barrier mechanism.
2480 // Also we need to add memory barrier to prevent commoning reads
2481 // from this field across safepoint since GC can change its value.
2482 bool need_read_barrier = !is_store &&
2483 offset != top() && heap_base_oop != top();
2484
2485 if (!is_store && type == T_OBJECT) {
2486 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2487 if (tjp != NULL) {
2488 value_type = tjp;
2489 }
2490 }
2491
2492 receiver = null_check(receiver);
2493 if (stopped()) {
2494 return true;
2495 }
2496 // Heap pointers get a null-check from the interpreter,
2497 // as a courtesy. However, this is not guaranteed by Unsafe,
2498 // and it is not possible to fully distinguish unintended nulls
2499 // from intended ones in this API.
2500
2501 // We need to emit leading and trailing CPU membars (see below) in
2502 // addition to memory membars for special access modes. This is a little
2503 // too strong, but avoids the need to insert per-alias-type
2504 // volatile membars (for stores; compare Parse::do_put_xxx), which
2505 // we cannot do effectively here because we probably only have a
2506 // rough approximation of type.
2507
2508 switch(kind) {
2509 case Relaxed:
2510 case Opaque:
2511 case Acquire:
2512 break;
2513 case Release:
2514 case Volatile:
2515 if (is_store) {
2516 insert_mem_bar(Op_MemBarRelease);
2517 } else {
2518 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2519 insert_mem_bar(Op_MemBarVolatile);
2520 }
2521 }
2522 break;
2523 default:
2524 ShouldNotReachHere();
2525 }
2526
2527 // Memory barrier to prevent normal and 'unsafe' accesses from
2528 // bypassing each other. Happens after null checks, so the
2529 // exception paths do not take memory state from the memory barrier,
2530 // so there's no problems making a strong assert about mixing users
2531 // of safe & unsafe memory.
2532 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2533
2534 if (!is_store) {
2535 Node* p = NULL;
2536 // Try to constant fold a load from a constant field
2537 ciField* field = alias_type->field();
2538 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2539 // final or stable field
2540 p = make_constant_from_field(field, heap_base_oop);
2541 }
2542 if (p == NULL) {
2543 // To be valid, unsafe loads may depend on other conditions than
2544 // the one that guards them: pin the Load node
2545 LoadNode::ControlDependency dep = LoadNode::Pinned;
2546 Node* ctrl = control();
2547 if (adr_type->isa_instptr()) {
2548 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2549 intptr_t offset = Type::OffsetBot;
2550 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2551 if (offset >= 0) {
2552 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2553 if (offset < s) {
2554 // Guaranteed to be a valid access, no need to pin it
2555 dep = LoadNode::DependsOnlyOnTest;
2556 ctrl = NULL;
2557 }
2558 }
2559 }
2560 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2561 // load value
2562 switch (type) {
2563 case T_BOOLEAN:
2564 {
2565 // Normalize the value returned by getBoolean in the following cases
2566 if (mismatched ||
2567 heap_base_oop == top() || // - heap_base_oop is NULL or
2568 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2569 // and the unsafe access is made to large offset
2570 // (i.e., larger than the maximum offset necessary for any
2571 // field access)
2572 ) {
2573 IdealKit ideal = IdealKit(this);
2574 #define __ ideal.
2575 IdealVariable normalized_result(ideal);
2576 __ declarations_done();
2577 __ set(normalized_result, p);
2578 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2579 __ set(normalized_result, ideal.ConI(1));
2580 ideal.end_if();
2581 final_sync(ideal);
2582 p = __ value(normalized_result);
2583 #undef __
2584 }
2585 }
2586 case T_CHAR:
2587 case T_BYTE:
2588 case T_SHORT:
2589 case T_INT:
2590 case T_LONG:
2591 case T_FLOAT:
2592 case T_DOUBLE:
2593 break;
2594 case T_OBJECT:
2595 if (need_read_barrier) {
2596 // We do not require a mem bar inside pre_barrier if need_mem_bar
2597 // is set: the barriers would be emitted by us.
2598 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2599 }
2600 break;
2601 case T_ADDRESS:
2602 // Cast to an int type.
2603 p = _gvn.transform(new CastP2XNode(NULL, p));
2604 p = ConvX2UL(p);
2605 break;
2606 default:
2607 fatal("unexpected type %d: %s", type, type2name(type));
2608 break;
2609 }
2610 }
2611 // The load node has the control of the preceding MemBarCPUOrder. All
2612 // following nodes will have the control of the MemBarCPUOrder inserted at
2613 // the end of this method. So, pushing the load onto the stack at a later
2614 // point is fine.
2615 set_result(p);
2616 } else {
2617 // place effect of store into memory
2618 switch (type) {
2619 case T_DOUBLE:
2620 val = dstore_rounding(val);
2621 break;
2622 case T_ADDRESS:
2623 // Repackage the long as a pointer.
2624 val = ConvL2X(val);
2625 val = _gvn.transform(new CastX2PNode(val));
2626 break;
2627 }
2628
2629 if (type == T_OBJECT) {
2630 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2631 } else {
2632 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2633 }
2634 }
2635
2636 switch(kind) {
2637 case Relaxed:
2638 case Opaque:
2639 case Release:
2640 break;
2641 case Acquire:
2642 case Volatile:
2643 if (!is_store) {
2644 insert_mem_bar(Op_MemBarAcquire);
2645 } else {
2646 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2647 insert_mem_bar(Op_MemBarVolatile);
2648 }
2649 }
2650 break;
2651 default:
2652 ShouldNotReachHere();
2653 }
2654
2655 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2656
2657 return true;
2658 }
2659
2660 //----------------------------inline_unsafe_load_store----------------------------
2661 // This method serves a couple of different customers (depending on LoadStoreKind):
2662 //
2663 // LS_cmp_swap:
2664 //
2665 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2666 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2667 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2668 //
2669 // LS_cmp_swap_weak:
2670 //
2671 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2672 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2673 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2674 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2675 //
2676 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2677 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2678 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2679 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2680 //
2681 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2682 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2683 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2684 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2685 //
2686 // LS_cmp_exchange:
2687 //
2688 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2689 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2690 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2691 //
2692 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2693 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2694 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2695 //
2696 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2697 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2698 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2699 //
2700 // LS_get_add:
2701 //
2702 // int getAndAddInt( Object o, long offset, int delta)
2703 // long getAndAddLong(Object o, long offset, long delta)
2704 //
2705 // LS_get_set:
2706 //
2707 // int getAndSet(Object o, long offset, int newValue)
2708 // long getAndSet(Object o, long offset, long newValue)
2709 // Object getAndSet(Object o, long offset, Object newValue)
2710 //
2711 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2712 // This basic scheme here is the same as inline_unsafe_access, but
2713 // differs in enough details that combining them would make the code
2714 // overly confusing. (This is a true fact! I originally combined
2715 // them, but even I was confused by it!) As much code/comments as
2716 // possible are retained from inline_unsafe_access though to make
2717 // the correspondences clearer. - dl
2718
2719 if (callee()->is_static()) return false; // caller must have the capability!
2720
2721 #ifndef PRODUCT
2722 BasicType rtype;
2723 {
2724 ResourceMark rm;
2725 // Check the signatures.
2726 ciSignature* sig = callee()->signature();
2727 rtype = sig->return_type()->basic_type();
2728 switch(kind) {
2729 case LS_get_add:
2730 case LS_get_set: {
2731 // Check the signatures.
2732 #ifdef ASSERT
2733 assert(rtype == type, "get and set must return the expected type");
2734 assert(sig->count() == 3, "get and set has 3 arguments");
2735 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2736 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2737 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2738 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2739 #endif // ASSERT
2740 break;
2741 }
2742 case LS_cmp_swap:
2743 case LS_cmp_swap_weak: {
2744 // Check the signatures.
2745 #ifdef ASSERT
2746 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2747 assert(sig->count() == 4, "CAS has 4 arguments");
2748 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2749 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2750 #endif // ASSERT
2751 break;
2752 }
2753 case LS_cmp_exchange: {
2754 // Check the signatures.
2755 #ifdef ASSERT
2756 assert(rtype == type, "CAS must return the expected type");
2757 assert(sig->count() == 4, "CAS has 4 arguments");
2758 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2759 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2760 #endif // ASSERT
2761 break;
2762 }
2763 default:
2764 ShouldNotReachHere();
2765 }
2766 }
2767 #endif //PRODUCT
2768
2769 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2770
2771 // Get arguments:
2772 Node* receiver = NULL;
2773 Node* base = NULL;
2774 Node* offset = NULL;
2775 Node* oldval = NULL;
2776 Node* newval = NULL;
2777 switch(kind) {
2778 case LS_cmp_swap:
2779 case LS_cmp_swap_weak:
2780 case LS_cmp_exchange: {
2781 const bool two_slot_type = type2size[type] == 2;
2782 receiver = argument(0); // type: oop
2783 base = argument(1); // type: oop
2784 offset = argument(2); // type: long
2785 oldval = argument(4); // type: oop, int, or long
2786 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2787 break;
2788 }
2789 case LS_get_add:
2790 case LS_get_set: {
2791 receiver = argument(0); // type: oop
2792 base = argument(1); // type: oop
2793 offset = argument(2); // type: long
2794 oldval = NULL;
2795 newval = argument(4); // type: oop, int, or long
2796 break;
2797 }
2798 default:
2799 ShouldNotReachHere();
2800 }
2801
2802 // Build field offset expression.
2803 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2804 // to be plain byte offsets, which are also the same as those accepted
2805 // by oopDesc::field_base.
2806 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2807 // 32-bit machines ignore the high half of long offsets
2808 offset = ConvL2X(offset);
2809 Node* adr = make_unsafe_address(base, offset, type, false);
2810 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2811
2812 Compile::AliasType* alias_type = C->alias_type(adr_type);
2813 BasicType bt = alias_type->basic_type();
2814 if (bt != T_ILLEGAL &&
2815 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2816 // Don't intrinsify mismatched object accesses.
2817 return false;
2818 }
2819
2820 // For CAS, unlike inline_unsafe_access, there seems no point in
2821 // trying to refine types. Just use the coarse types here.
2822 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2823 const Type *value_type = Type::get_const_basic_type(type);
2824
2825 switch (kind) {
2826 case LS_get_set:
2827 case LS_cmp_exchange: {
2828 if (type == T_OBJECT) {
2829 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2830 if (tjp != NULL) {
2831 value_type = tjp;
2832 }
2833 }
2834 break;
2835 }
2836 case LS_cmp_swap:
2837 case LS_cmp_swap_weak:
2838 case LS_get_add:
2839 break;
2840 default:
2841 ShouldNotReachHere();
2842 }
2843
2844 // Null check receiver.
2845 receiver = null_check(receiver);
2846 if (stopped()) {
2847 return true;
2848 }
2849
2850 int alias_idx = C->get_alias_index(adr_type);
2851
2852 // Memory-model-wise, a LoadStore acts like a little synchronized
2853 // block, so needs barriers on each side. These don't translate
2854 // into actual barriers on most machines, but we still need rest of
2855 // compiler to respect ordering.
2856
2857 switch (access_kind) {
2858 case Relaxed:
2859 case Acquire:
2860 break;
2861 case Release:
2862 insert_mem_bar(Op_MemBarRelease);
2863 break;
2864 case Volatile:
2865 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2866 insert_mem_bar(Op_MemBarVolatile);
2867 } else {
2868 insert_mem_bar(Op_MemBarRelease);
2869 }
2870 break;
2871 default:
2872 ShouldNotReachHere();
2873 }
2874 insert_mem_bar(Op_MemBarCPUOrder);
2875
2876 // Figure out the memory ordering.
2877 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2878
2879 // 4984716: MemBars must be inserted before this
2880 // memory node in order to avoid a false
2881 // dependency which will confuse the scheduler.
2882 Node *mem = memory(alias_idx);
2883
2884 // For now, we handle only those cases that actually exist: ints,
2885 // longs, and Object. Adding others should be straightforward.
2886 Node* load_store = NULL;
2887 switch(type) {
2888 case T_BYTE:
2889 switch(kind) {
2890 case LS_get_add:
2891 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2892 break;
2893 case LS_get_set:
2894 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2895 break;
2896 case LS_cmp_swap_weak:
2897 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2898 break;
2899 case LS_cmp_swap:
2900 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2901 break;
2902 case LS_cmp_exchange:
2903 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2904 break;
2905 default:
2906 ShouldNotReachHere();
2907 }
2908 break;
2909 case T_SHORT:
2910 switch(kind) {
2911 case LS_get_add:
2912 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2913 break;
2914 case LS_get_set:
2915 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2916 break;
2917 case LS_cmp_swap_weak:
2918 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2919 break;
2920 case LS_cmp_swap:
2921 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2922 break;
2923 case LS_cmp_exchange:
2924 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2925 break;
2926 default:
2927 ShouldNotReachHere();
2928 }
2929 break;
2930 case T_INT:
2931 switch(kind) {
2932 case LS_get_add:
2933 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2934 break;
2935 case LS_get_set:
2936 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2937 break;
2938 case LS_cmp_swap_weak:
2939 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2940 break;
2941 case LS_cmp_swap:
2942 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2943 break;
2944 case LS_cmp_exchange:
2945 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2946 break;
2947 default:
2948 ShouldNotReachHere();
2949 }
2950 break;
2951 case T_LONG:
2952 switch(kind) {
2953 case LS_get_add:
2954 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2955 break;
2956 case LS_get_set:
2957 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2958 break;
2959 case LS_cmp_swap_weak:
2960 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2961 break;
2962 case LS_cmp_swap:
2963 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2964 break;
2965 case LS_cmp_exchange:
2966 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2967 break;
2968 default:
2969 ShouldNotReachHere();
2970 }
2971 break;
2972 case T_OBJECT:
2973 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2974 // could be delayed during Parse (for example, in adjust_map_after_if()).
2975 // Execute transformation here to avoid barrier generation in such case.
2976 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2977 newval = _gvn.makecon(TypePtr::NULL_PTR);
2978
2979 // Reference stores need a store barrier.
2980 switch(kind) {
2981 case LS_get_set: {
2982 // If pre-barrier must execute before the oop store, old value will require do_load here.
2983 if (!can_move_pre_barrier()) {
2984 pre_barrier(true /* do_load*/,
2985 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2986 NULL /* pre_val*/,
2987 T_OBJECT);
2988 } // Else move pre_barrier to use load_store value, see below.
2989 break;
2990 }
2991 case LS_cmp_swap_weak:
2992 case LS_cmp_swap:
2993 case LS_cmp_exchange: {
2994 // Same as for newval above:
2995 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2996 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2997 }
2998 // The only known value which might get overwritten is oldval.
2999 pre_barrier(false /* do_load */,
3000 control(), NULL, NULL, max_juint, NULL, NULL,
3001 oldval /* pre_val */,
3002 T_OBJECT);
3003 break;
3004 }
3005 default:
3006 ShouldNotReachHere();
3007 }
3008
3009 #ifdef _LP64
3010 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3011 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3012
3013 switch(kind) {
3014 case LS_get_set:
3015 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3016 break;
3017 case LS_cmp_swap_weak: {
3018 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3019 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3020 break;
3021 }
3022 case LS_cmp_swap: {
3023 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3024 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3025 break;
3026 }
3027 case LS_cmp_exchange: {
3028 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3029 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3030 break;
3031 }
3032 default:
3033 ShouldNotReachHere();
3034 }
3035 } else
3036 #endif
3037 switch (kind) {
3038 case LS_get_set:
3039 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3040 break;
3041 case LS_cmp_swap_weak:
3042 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3043 break;
3044 case LS_cmp_swap:
3045 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3046 break;
3047 case LS_cmp_exchange:
3048 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3049 break;
3050 default:
3051 ShouldNotReachHere();
3052 }
3053
3054 // Emit the post barrier only when the actual store happened. This makes sense
3055 // to check only for LS_cmp_* that can fail to set the value.
3056 // LS_cmp_exchange does not produce any branches by default, so there is no
3057 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3058 // CompareAndExchange and move branches here, it would make sense to conditionalize
3059 // post_barriers for LS_cmp_exchange as well.
3060 //
3061 // CAS success path is marked more likely since we anticipate this is a performance
3062 // critical path, while CAS failure path can use the penalty for going through unlikely
3063 // path as backoff. Which is still better than doing a store barrier there.
3064 switch (kind) {
3065 case LS_get_set:
3066 case LS_cmp_exchange: {
3067 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3068 break;
3069 }
3070 case LS_cmp_swap_weak:
3071 case LS_cmp_swap: {
3072 IdealKit ideal(this);
3073 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3074 sync_kit(ideal);
3075 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3076 ideal.sync_kit(this);
3077 } ideal.end_if();
3078 final_sync(ideal);
3079 break;
3080 }
3081 default:
3082 ShouldNotReachHere();
3083 }
3084 break;
3085 default:
3086 fatal("unexpected type %d: %s", type, type2name(type));
3087 break;
3088 }
3089
3090 // SCMemProjNodes represent the memory state of a LoadStore. Their
3091 // main role is to prevent LoadStore nodes from being optimized away
3092 // when their results aren't used.
3093 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3094 set_memory(proj, alias_idx);
3095
3096 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3097 #ifdef _LP64
3098 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3099 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3100 }
3101 #endif
3102 if (can_move_pre_barrier() && kind == LS_get_set) {
3103 // Don't need to load pre_val. The old value is returned by load_store.
3104 // The pre_barrier can execute after the xchg as long as no safepoint
3105 // gets inserted between them.
3106 pre_barrier(false /* do_load */,
3107 control(), NULL, NULL, max_juint, NULL, NULL,
3108 load_store /* pre_val */,
3109 T_OBJECT);
3110 }
3111 }
3112
3113 // Add the trailing membar surrounding the access
3114 insert_mem_bar(Op_MemBarCPUOrder);
3115
3116 switch (access_kind) {
3117 case Relaxed:
3118 case Release:
3119 break; // do nothing
3120 case Acquire:
3121 case Volatile:
3122 insert_mem_bar(Op_MemBarAcquire);
3123 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3124 break;
3125 default:
3126 ShouldNotReachHere();
3127 }
3128
3129 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3130 set_result(load_store);
3131 return true;
3132 }
3133
3134 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3135 MemNode::MemOrd mo = MemNode::unset;
3136 switch(kind) {
3137 case Opaque:
3138 case Relaxed: mo = MemNode::unordered; break;
3139 case Acquire: mo = MemNode::acquire; break;
3140 case Release: mo = MemNode::release; break;
3141 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3142 default:
3143 ShouldNotReachHere();
3144 }
3145 guarantee(mo != MemNode::unset, "Should select memory ordering");
3146 return mo;
3147 }
3148
3149 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3150 MemNode::MemOrd mo = MemNode::unset;
3151 switch(kind) {
3152 case Opaque:
3153 case Relaxed: mo = MemNode::unordered; break;
3154 case Acquire: mo = MemNode::acquire; break;
3155 case Release: mo = MemNode::release; break;
3156 case Volatile: mo = MemNode::seqcst; break;
3157 default:
3158 ShouldNotReachHere();
3159 }
3160 guarantee(mo != MemNode::unset, "Should select memory ordering");
3161 return mo;
3162 }
3163
3164 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3165 // Regardless of form, don't allow previous ld/st to move down,
3166 // then issue acquire, release, or volatile mem_bar.
3167 insert_mem_bar(Op_MemBarCPUOrder);
3168 switch(id) {
3169 case vmIntrinsics::_loadFence:
3170 insert_mem_bar(Op_LoadFence);
3171 return true;
3172 case vmIntrinsics::_storeFence:
3173 insert_mem_bar(Op_StoreFence);
3174 return true;
3175 case vmIntrinsics::_fullFence:
3176 insert_mem_bar(Op_MemBarVolatile);
3177 return true;
3178 default:
3179 fatal_unexpected_iid(id);
3180 return false;
3181 }
3182 }
3183
3184 bool LibraryCallKit::inline_onspinwait() {
3185 insert_mem_bar(Op_OnSpinWait);
3186 return true;
3187 }
3188
3189 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3190 if (!kls->is_Con()) {
3191 return true;
3192 }
3193 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3194 if (klsptr == NULL) {
3195 return true;
3196 }
3197 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3198 // don't need a guard for a klass that is already initialized
3199 return !ik->is_initialized();
3200 }
3201
3202 //----------------------------inline_unsafe_allocate---------------------------
3203 // public native Object Unsafe.allocateInstance(Class<?> cls);
3204 bool LibraryCallKit::inline_unsafe_allocate() {
3205 if (callee()->is_static()) return false; // caller must have the capability!
3206
3207 null_check_receiver(); // null-check, then ignore
3208 Node* cls = null_check(argument(1));
3209 if (stopped()) return true;
3210
3211 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3212 kls = null_check(kls);
3213 if (stopped()) return true; // argument was like int.class
3214
3215 Node* test = NULL;
3216 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3217 // Note: The argument might still be an illegal value like
3218 // Serializable.class or Object[].class. The runtime will handle it.
3219 // But we must make an explicit check for initialization.
3220 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3221 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3222 // can generate code to load it as unsigned byte.
3223 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3224 Node* bits = intcon(InstanceKlass::fully_initialized);
3225 test = _gvn.transform(new SubINode(inst, bits));
3226 // The 'test' is non-zero if we need to take a slow path.
3227 }
3228
3229 Node* obj = new_instance(kls, test);
3230 set_result(obj);
3231 return true;
3232 }
3233
3234 //------------------------inline_native_time_funcs--------------
3235 // inline code for System.currentTimeMillis() and System.nanoTime()
3236 // these have the same type and signature
3237 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3238 const TypeFunc* tf = OptoRuntime::void_long_Type();
3239 const TypePtr* no_memory_effects = NULL;
3240 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3241 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3242 #ifdef ASSERT
3243 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3244 assert(value_top == top(), "second value must be top");
3245 #endif
3246 set_result(value);
3247 return true;
3248 }
3249
3250 #ifdef TRACE_HAVE_INTRINSICS
3251
3252 /*
3253 * oop -> myklass
3254 * myklass->trace_id |= USED
3255 * return myklass->trace_id & ~0x3
3256 */
3257 bool LibraryCallKit::inline_native_classID() {
3258 Node* cls = null_check(argument(0), T_OBJECT);
3259 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3260 kls = null_check(kls, T_OBJECT);
3261
3262 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3263 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3264 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3265
3266 Node* clsused = longcon(0x01l); // set the class bit
3267 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3268 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3269 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3270
3271 #ifdef TRACE_ID_META_BITS
3272 Node* mbits = longcon(~TRACE_ID_META_BITS);
3273 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3274 #endif
3275 #ifdef TRACE_ID_CLASS_SHIFT
3276 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3277 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3278 #endif
3279
3280 set_result(tvalue);
3281 return true;
3282
3283 }
3284
3285 bool LibraryCallKit::inline_native_getBufferWriter() {
3286 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3287
3288 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3289 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3290 );
3291
3292 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3293
3294 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3295 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3296
3297 IfNode* iff_jobj_null =
3298 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3299
3300 enum { _normal_path = 1,
3301 _null_path = 2,
3302 PATH_LIMIT };
3303
3304 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3305 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3306
3307 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3308 result_rgn->init_req(_null_path, jobj_is_null);
3309 result_val->init_req(_null_path, null());
3310
3311 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3312 result_rgn->init_req(_normal_path, jobj_is_not_null);
3313
3314 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3315 result_val->init_req(_normal_path, res);
3316
3317 set_result(result_rgn, result_val);
3318
3319 return true;
3320 }
3321
3322 #endif
3323
3324 //------------------------inline_native_currentThread------------------
3325 bool LibraryCallKit::inline_native_currentThread() {
3326 Node* junk = NULL;
3327 set_result(generate_current_thread(junk));
3328 return true;
3329 }
3330
3331 //------------------------inline_native_isInterrupted------------------
3332 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3333 bool LibraryCallKit::inline_native_isInterrupted() {
3334 // Add a fast path to t.isInterrupted(clear_int):
3335 // (t == Thread.current() &&
3336 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3337 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3338 // So, in the common case that the interrupt bit is false,
3339 // we avoid making a call into the VM. Even if the interrupt bit
3340 // is true, if the clear_int argument is false, we avoid the VM call.
3341 // However, if the receiver is not currentThread, we must call the VM,
3342 // because there must be some locking done around the operation.
3343
3344 // We only go to the fast case code if we pass two guards.
3345 // Paths which do not pass are accumulated in the slow_region.
3346
3347 enum {
3348 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3349 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3350 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3351 PATH_LIMIT
3352 };
3353
3354 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3355 // out of the function.
3356 insert_mem_bar(Op_MemBarCPUOrder);
3357
3358 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3359 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3360
3361 RegionNode* slow_region = new RegionNode(1);
3362 record_for_igvn(slow_region);
3363
3364 // (a) Receiving thread must be the current thread.
3365 Node* rec_thr = argument(0);
3366 Node* tls_ptr = NULL;
3367 Node* cur_thr = generate_current_thread(tls_ptr);
3368 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3369 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3370
3371 generate_slow_guard(bol_thr, slow_region);
3372
3373 // (b) Interrupt bit on TLS must be false.
3374 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3375 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3376 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3377
3378 // Set the control input on the field _interrupted read to prevent it floating up.
3379 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3380 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3381 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3382
3383 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3384
3385 // First fast path: if (!TLS._interrupted) return false;
3386 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3387 result_rgn->init_req(no_int_result_path, false_bit);
3388 result_val->init_req(no_int_result_path, intcon(0));
3389
3390 // drop through to next case
3391 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3392
3393 #ifndef _WINDOWS
3394 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3395 Node* clr_arg = argument(1);
3396 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3397 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3398 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3399
3400 // Second fast path: ... else if (!clear_int) return true;
3401 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3402 result_rgn->init_req(no_clear_result_path, false_arg);
3403 result_val->init_req(no_clear_result_path, intcon(1));
3404
3405 // drop through to next case
3406 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3407 #else
3408 // To return true on Windows you must read the _interrupted field
3409 // and check the event state i.e. take the slow path.
3410 #endif // _WINDOWS
3411
3412 // (d) Otherwise, go to the slow path.
3413 slow_region->add_req(control());
3414 set_control( _gvn.transform(slow_region));
3415
3416 if (stopped()) {
3417 // There is no slow path.
3418 result_rgn->init_req(slow_result_path, top());
3419 result_val->init_req(slow_result_path, top());
3420 } else {
3421 // non-virtual because it is a private non-static
3422 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3423
3424 Node* slow_val = set_results_for_java_call(slow_call);
3425 // this->control() comes from set_results_for_java_call
3426
3427 Node* fast_io = slow_call->in(TypeFunc::I_O);
3428 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3429
3430 // These two phis are pre-filled with copies of of the fast IO and Memory
3431 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3432 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3433
3434 result_rgn->init_req(slow_result_path, control());
3435 result_io ->init_req(slow_result_path, i_o());
3436 result_mem->init_req(slow_result_path, reset_memory());
3437 result_val->init_req(slow_result_path, slow_val);
3438
3439 set_all_memory(_gvn.transform(result_mem));
3440 set_i_o( _gvn.transform(result_io));
3441 }
3442
3443 C->set_has_split_ifs(true); // Has chance for split-if optimization
3444 set_result(result_rgn, result_val);
3445 return true;
3446 }
3447
3448 //---------------------------load_mirror_from_klass----------------------------
3449 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3450 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3451 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3452 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3453 }
3454
3455 //-----------------------load_klass_from_mirror_common-------------------------
3456 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3457 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3458 // and branch to the given path on the region.
3459 // If never_see_null, take an uncommon trap on null, so we can optimistically
3460 // compile for the non-null case.
3461 // If the region is NULL, force never_see_null = true.
3462 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3463 bool never_see_null,
3464 RegionNode* region,
3465 int null_path,
3466 int offset) {
3467 if (region == NULL) never_see_null = true;
3468 Node* p = basic_plus_adr(mirror, offset);
3469 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3470 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3471 Node* null_ctl = top();
3472 kls = null_check_oop(kls, &null_ctl, never_see_null);
3473 if (region != NULL) {
3474 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3475 region->init_req(null_path, null_ctl);
3476 } else {
3477 assert(null_ctl == top(), "no loose ends");
3478 }
3479 return kls;
3480 }
3481
3482 //--------------------(inline_native_Class_query helpers)---------------------
3483 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3484 // Fall through if (mods & mask) == bits, take the guard otherwise.
3485 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3486 // Branch around if the given klass has the given modifier bit set.
3487 // Like generate_guard, adds a new path onto the region.
3488 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3489 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3490 Node* mask = intcon(modifier_mask);
3491 Node* bits = intcon(modifier_bits);
3492 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3493 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3494 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3495 return generate_fair_guard(bol, region);
3496 }
3497 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3498 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3499 }
3500
3501 //-------------------------inline_native_Class_query-------------------
3502 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3503 const Type* return_type = TypeInt::BOOL;
3504 Node* prim_return_value = top(); // what happens if it's a primitive class?
3505 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3506 bool expect_prim = false; // most of these guys expect to work on refs
3507
3508 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3509
3510 Node* mirror = argument(0);
3511 Node* obj = top();
3512
3513 switch (id) {
3514 case vmIntrinsics::_isInstance:
3515 // nothing is an instance of a primitive type
3516 prim_return_value = intcon(0);
3517 obj = argument(1);
3518 break;
3519 case vmIntrinsics::_getModifiers:
3520 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3521 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3522 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3523 break;
3524 case vmIntrinsics::_isInterface:
3525 prim_return_value = intcon(0);
3526 break;
3527 case vmIntrinsics::_isArray:
3528 prim_return_value = intcon(0);
3529 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3530 break;
3531 case vmIntrinsics::_isPrimitive:
3532 prim_return_value = intcon(1);
3533 expect_prim = true; // obviously
3534 break;
3535 case vmIntrinsics::_getSuperclass:
3536 prim_return_value = null();
3537 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3538 break;
3539 case vmIntrinsics::_getClassAccessFlags:
3540 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3541 return_type = TypeInt::INT; // not bool! 6297094
3542 break;
3543 default:
3544 fatal_unexpected_iid(id);
3545 break;
3546 }
3547
3548 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3549 if (mirror_con == NULL) return false; // cannot happen?
3550
3551 #ifndef PRODUCT
3552 if (C->print_intrinsics() || C->print_inlining()) {
3553 ciType* k = mirror_con->java_mirror_type();
3554 if (k) {
3555 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3556 k->print_name();
3557 tty->cr();
3558 }
3559 }
3560 #endif
3561
3562 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3563 RegionNode* region = new RegionNode(PATH_LIMIT);
3564 record_for_igvn(region);
3565 PhiNode* phi = new PhiNode(region, return_type);
3566
3567 // The mirror will never be null of Reflection.getClassAccessFlags, however
3568 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3569 // if it is. See bug 4774291.
3570
3571 // For Reflection.getClassAccessFlags(), the null check occurs in
3572 // the wrong place; see inline_unsafe_access(), above, for a similar
3573 // situation.
3574 mirror = null_check(mirror);
3575 // If mirror or obj is dead, only null-path is taken.
3576 if (stopped()) return true;
3577
3578 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3579
3580 // Now load the mirror's klass metaobject, and null-check it.
3581 // Side-effects region with the control path if the klass is null.
3582 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3583 // If kls is null, we have a primitive mirror.
3584 phi->init_req(_prim_path, prim_return_value);
3585 if (stopped()) { set_result(region, phi); return true; }
3586 bool safe_for_replace = (region->in(_prim_path) == top());
3587
3588 Node* p; // handy temp
3589 Node* null_ctl;
3590
3591 // Now that we have the non-null klass, we can perform the real query.
3592 // For constant classes, the query will constant-fold in LoadNode::Value.
3593 Node* query_value = top();
3594 switch (id) {
3595 case vmIntrinsics::_isInstance:
3596 // nothing is an instance of a primitive type
3597 query_value = gen_instanceof(obj, kls, safe_for_replace);
3598 break;
3599
3600 case vmIntrinsics::_getModifiers:
3601 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3602 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3603 break;
3604
3605 case vmIntrinsics::_isInterface:
3606 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3607 if (generate_interface_guard(kls, region) != NULL)
3608 // A guard was added. If the guard is taken, it was an interface.
3609 phi->add_req(intcon(1));
3610 // If we fall through, it's a plain class.
3611 query_value = intcon(0);
3612 break;
3613
3614 case vmIntrinsics::_isArray:
3615 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3616 if (generate_array_guard(kls, region) != NULL)
3617 // A guard was added. If the guard is taken, it was an array.
3618 phi->add_req(intcon(1));
3619 // If we fall through, it's a plain class.
3620 query_value = intcon(0);
3621 break;
3622
3623 case vmIntrinsics::_isPrimitive:
3624 query_value = intcon(0); // "normal" path produces false
3625 break;
3626
3627 case vmIntrinsics::_getSuperclass:
3628 // The rules here are somewhat unfortunate, but we can still do better
3629 // with random logic than with a JNI call.
3630 // Interfaces store null or Object as _super, but must report null.
3631 // Arrays store an intermediate super as _super, but must report Object.
3632 // Other types can report the actual _super.
3633 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3634 if (generate_interface_guard(kls, region) != NULL)
3635 // A guard was added. If the guard is taken, it was an interface.
3636 phi->add_req(null());
3637 if (generate_array_guard(kls, region) != NULL)
3638 // A guard was added. If the guard is taken, it was an array.
3639 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3640 // If we fall through, it's a plain class. Get its _super.
3641 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3642 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3643 null_ctl = top();
3644 kls = null_check_oop(kls, &null_ctl);
3645 if (null_ctl != top()) {
3646 // If the guard is taken, Object.superClass is null (both klass and mirror).
3647 region->add_req(null_ctl);
3648 phi ->add_req(null());
3649 }
3650 if (!stopped()) {
3651 query_value = load_mirror_from_klass(kls);
3652 }
3653 break;
3654
3655 case vmIntrinsics::_getClassAccessFlags:
3656 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3657 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3658 break;
3659
3660 default:
3661 fatal_unexpected_iid(id);
3662 break;
3663 }
3664
3665 // Fall-through is the normal case of a query to a real class.
3666 phi->init_req(1, query_value);
3667 region->init_req(1, control());
3668
3669 C->set_has_split_ifs(true); // Has chance for split-if optimization
3670 set_result(region, phi);
3671 return true;
3672 }
3673
3674 //-------------------------inline_Class_cast-------------------
3675 bool LibraryCallKit::inline_Class_cast() {
3676 Node* mirror = argument(0); // Class
3677 Node* obj = argument(1);
3678 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3679 if (mirror_con == NULL) {
3680 return false; // dead path (mirror->is_top()).
3681 }
3682 if (obj == NULL || obj->is_top()) {
3683 return false; // dead path
3684 }
3685 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3686
3687 // First, see if Class.cast() can be folded statically.
3688 // java_mirror_type() returns non-null for compile-time Class constants.
3689 ciType* tm = mirror_con->java_mirror_type();
3690 if (tm != NULL && tm->is_klass() &&
3691 tp != NULL && tp->klass() != NULL) {
3692 if (!tp->klass()->is_loaded()) {
3693 // Don't use intrinsic when class is not loaded.
3694 return false;
3695 } else {
3696 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3697 if (static_res == Compile::SSC_always_true) {
3698 // isInstance() is true - fold the code.
3699 set_result(obj);
3700 return true;
3701 } else if (static_res == Compile::SSC_always_false) {
3702 // Don't use intrinsic, have to throw ClassCastException.
3703 // If the reference is null, the non-intrinsic bytecode will
3704 // be optimized appropriately.
3705 return false;
3706 }
3707 }
3708 }
3709
3710 // Bailout intrinsic and do normal inlining if exception path is frequent.
3711 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3712 return false;
3713 }
3714
3715 // Generate dynamic checks.
3716 // Class.cast() is java implementation of _checkcast bytecode.
3717 // Do checkcast (Parse::do_checkcast()) optimizations here.
3718
3719 mirror = null_check(mirror);
3720 // If mirror is dead, only null-path is taken.
3721 if (stopped()) {
3722 return true;
3723 }
3724
3725 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3726 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3727 RegionNode* region = new RegionNode(PATH_LIMIT);
3728 record_for_igvn(region);
3729
3730 // Now load the mirror's klass metaobject, and null-check it.
3731 // If kls is null, we have a primitive mirror and
3732 // nothing is an instance of a primitive type.
3733 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3734
3735 Node* res = top();
3736 if (!stopped()) {
3737 Node* bad_type_ctrl = top();
3738 // Do checkcast optimizations.
3739 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3740 region->init_req(_bad_type_path, bad_type_ctrl);
3741 }
3742 if (region->in(_prim_path) != top() ||
3743 region->in(_bad_type_path) != top()) {
3744 // Let Interpreter throw ClassCastException.
3745 PreserveJVMState pjvms(this);
3746 set_control(_gvn.transform(region));
3747 uncommon_trap(Deoptimization::Reason_intrinsic,
3748 Deoptimization::Action_maybe_recompile);
3749 }
3750 if (!stopped()) {
3751 set_result(res);
3752 }
3753 return true;
3754 }
3755
3756
3757 //--------------------------inline_native_subtype_check------------------------
3758 // This intrinsic takes the JNI calls out of the heart of
3759 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3760 bool LibraryCallKit::inline_native_subtype_check() {
3761 // Pull both arguments off the stack.
3762 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3763 args[0] = argument(0);
3764 args[1] = argument(1);
3765 Node* klasses[2]; // corresponding Klasses: superk, subk
3766 klasses[0] = klasses[1] = top();
3767
3768 enum {
3769 // A full decision tree on {superc is prim, subc is prim}:
3770 _prim_0_path = 1, // {P,N} => false
3771 // {P,P} & superc!=subc => false
3772 _prim_same_path, // {P,P} & superc==subc => true
3773 _prim_1_path, // {N,P} => false
3774 _ref_subtype_path, // {N,N} & subtype check wins => true
3775 _both_ref_path, // {N,N} & subtype check loses => false
3776 PATH_LIMIT
3777 };
3778
3779 RegionNode* region = new RegionNode(PATH_LIMIT);
3780 Node* phi = new PhiNode(region, TypeInt::BOOL);
3781 record_for_igvn(region);
3782
3783 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3784 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3785 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3786
3787 // First null-check both mirrors and load each mirror's klass metaobject.
3788 int which_arg;
3789 for (which_arg = 0; which_arg <= 1; which_arg++) {
3790 Node* arg = args[which_arg];
3791 arg = null_check(arg);
3792 if (stopped()) break;
3793 args[which_arg] = arg;
3794
3795 Node* p = basic_plus_adr(arg, class_klass_offset);
3796 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3797 klasses[which_arg] = _gvn.transform(kls);
3798 }
3799
3800 // Having loaded both klasses, test each for null.
3801 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3802 for (which_arg = 0; which_arg <= 1; which_arg++) {
3803 Node* kls = klasses[which_arg];
3804 Node* null_ctl = top();
3805 kls = null_check_oop(kls, &null_ctl, never_see_null);
3806 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3807 region->init_req(prim_path, null_ctl);
3808 if (stopped()) break;
3809 klasses[which_arg] = kls;
3810 }
3811
3812 if (!stopped()) {
3813 // now we have two reference types, in klasses[0..1]
3814 Node* subk = klasses[1]; // the argument to isAssignableFrom
3815 Node* superk = klasses[0]; // the receiver
3816 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3817 // now we have a successful reference subtype check
3818 region->set_req(_ref_subtype_path, control());
3819 }
3820
3821 // If both operands are primitive (both klasses null), then
3822 // we must return true when they are identical primitives.
3823 // It is convenient to test this after the first null klass check.
3824 set_control(region->in(_prim_0_path)); // go back to first null check
3825 if (!stopped()) {
3826 // Since superc is primitive, make a guard for the superc==subc case.
3827 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3828 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3829 generate_guard(bol_eq, region, PROB_FAIR);
3830 if (region->req() == PATH_LIMIT+1) {
3831 // A guard was added. If the added guard is taken, superc==subc.
3832 region->swap_edges(PATH_LIMIT, _prim_same_path);
3833 region->del_req(PATH_LIMIT);
3834 }
3835 region->set_req(_prim_0_path, control()); // Not equal after all.
3836 }
3837
3838 // these are the only paths that produce 'true':
3839 phi->set_req(_prim_same_path, intcon(1));
3840 phi->set_req(_ref_subtype_path, intcon(1));
3841
3842 // pull together the cases:
3843 assert(region->req() == PATH_LIMIT, "sane region");
3844 for (uint i = 1; i < region->req(); i++) {
3845 Node* ctl = region->in(i);
3846 if (ctl == NULL || ctl == top()) {
3847 region->set_req(i, top());
3848 phi ->set_req(i, top());
3849 } else if (phi->in(i) == NULL) {
3850 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3851 }
3852 }
3853
3854 set_control(_gvn.transform(region));
3855 set_result(_gvn.transform(phi));
3856 return true;
3857 }
3858
3859 //---------------------generate_array_guard_common------------------------
3860 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3861 bool obj_array, bool not_array) {
3862
3863 if (stopped()) {
3864 return NULL;
3865 }
3866
3867 // If obj_array/non_array==false/false:
3868 // Branch around if the given klass is in fact an array (either obj or prim).
3869 // If obj_array/non_array==false/true:
3870 // Branch around if the given klass is not an array klass of any kind.
3871 // If obj_array/non_array==true/true:
3872 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3873 // If obj_array/non_array==true/false:
3874 // Branch around if the kls is an oop array (Object[] or subtype)
3875 //
3876 // Like generate_guard, adds a new path onto the region.
3877 jint layout_con = 0;
3878 Node* layout_val = get_layout_helper(kls, layout_con);
3879 if (layout_val == NULL) {
3880 bool query = (obj_array
3881 ? Klass::layout_helper_is_objArray(layout_con)
3882 : Klass::layout_helper_is_array(layout_con));
3883 if (query == not_array) {
3884 return NULL; // never a branch
3885 } else { // always a branch
3886 Node* always_branch = control();
3887 if (region != NULL)
3888 region->add_req(always_branch);
3889 set_control(top());
3890 return always_branch;
3891 }
3892 }
3893 // Now test the correct condition.
3894 jint nval = (obj_array
3895 ? (jint)(Klass::_lh_array_tag_type_value
3896 << Klass::_lh_array_tag_shift)
3897 : Klass::_lh_neutral_value);
3898 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3899 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3900 // invert the test if we are looking for a non-array
3901 if (not_array) btest = BoolTest(btest).negate();
3902 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3903 return generate_fair_guard(bol, region);
3904 }
3905
3906
3907 //-----------------------inline_native_newArray--------------------------
3908 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3909 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3910 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3911 Node* mirror;
3912 Node* count_val;
3913 if (uninitialized) {
3914 mirror = argument(1);
3915 count_val = argument(2);
3916 } else {
3917 mirror = argument(0);
3918 count_val = argument(1);
3919 }
3920
3921 mirror = null_check(mirror);
3922 // If mirror or obj is dead, only null-path is taken.
3923 if (stopped()) return true;
3924
3925 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3926 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3927 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3928 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3929 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3930
3931 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3932 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3933 result_reg, _slow_path);
3934 Node* normal_ctl = control();
3935 Node* no_array_ctl = result_reg->in(_slow_path);
3936
3937 // Generate code for the slow case. We make a call to newArray().
3938 set_control(no_array_ctl);
3939 if (!stopped()) {
3940 // Either the input type is void.class, or else the
3941 // array klass has not yet been cached. Either the
3942 // ensuing call will throw an exception, or else it
3943 // will cache the array klass for next time.
3944 PreserveJVMState pjvms(this);
3945 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3946 Node* slow_result = set_results_for_java_call(slow_call);
3947 // this->control() comes from set_results_for_java_call
3948 result_reg->set_req(_slow_path, control());
3949 result_val->set_req(_slow_path, slow_result);
3950 result_io ->set_req(_slow_path, i_o());
3951 result_mem->set_req(_slow_path, reset_memory());
3952 }
3953
3954 set_control(normal_ctl);
3955 if (!stopped()) {
3956 // Normal case: The array type has been cached in the java.lang.Class.
3957 // The following call works fine even if the array type is polymorphic.
3958 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3959 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3960 result_reg->init_req(_normal_path, control());
3961 result_val->init_req(_normal_path, obj);
3962 result_io ->init_req(_normal_path, i_o());
3963 result_mem->init_req(_normal_path, reset_memory());
3964
3965 if (uninitialized) {
3966 // Mark the allocation so that zeroing is skipped
3967 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3968 alloc->maybe_set_complete(&_gvn);
3969 }
3970 }
3971
3972 // Return the combined state.
3973 set_i_o( _gvn.transform(result_io) );
3974 set_all_memory( _gvn.transform(result_mem));
3975
3976 C->set_has_split_ifs(true); // Has chance for split-if optimization
3977 set_result(result_reg, result_val);
3978 return true;
3979 }
3980
3981 //----------------------inline_native_getLength--------------------------
3982 // public static native int java.lang.reflect.Array.getLength(Object array);
3983 bool LibraryCallKit::inline_native_getLength() {
3984 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3985
3986 Node* array = null_check(argument(0));
3987 // If array is dead, only null-path is taken.
3988 if (stopped()) return true;
3989
3990 // Deoptimize if it is a non-array.
3991 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3992
3993 if (non_array != NULL) {
3994 PreserveJVMState pjvms(this);
3995 set_control(non_array);
3996 uncommon_trap(Deoptimization::Reason_intrinsic,
3997 Deoptimization::Action_maybe_recompile);
3998 }
3999
4000 // If control is dead, only non-array-path is taken.
4001 if (stopped()) return true;
4002
4003 // The works fine even if the array type is polymorphic.
4004 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4005 Node* result = load_array_length(array);
4006
4007 C->set_has_split_ifs(true); // Has chance for split-if optimization
4008 set_result(result);
4009 return true;
4010 }
4011
4012 //------------------------inline_array_copyOf----------------------------
4013 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4014 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4015 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4016 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4017
4018 // Get the arguments.
4019 Node* original = argument(0);
4020 Node* start = is_copyOfRange? argument(1): intcon(0);
4021 Node* end = is_copyOfRange? argument(2): argument(1);
4022 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4023
4024 Node* newcopy = NULL;
4025
4026 // Set the original stack and the reexecute bit for the interpreter to reexecute
4027 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4028 { PreserveReexecuteState preexecs(this);
4029 jvms()->set_should_reexecute(true);
4030
4031 array_type_mirror = null_check(array_type_mirror);
4032 original = null_check(original);
4033
4034 // Check if a null path was taken unconditionally.
4035 if (stopped()) return true;
4036
4037 Node* orig_length = load_array_length(original);
4038
4039 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4040 klass_node = null_check(klass_node);
4041
4042 RegionNode* bailout = new RegionNode(1);
4043 record_for_igvn(bailout);
4044
4045 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4046 // Bail out if that is so.
4047 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4048 if (not_objArray != NULL) {
4049 // Improve the klass node's type from the new optimistic assumption:
4050 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4051 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4052 Node* cast = new CastPPNode(klass_node, akls);
4053 cast->init_req(0, control());
4054 klass_node = _gvn.transform(cast);
4055 }
4056
4057 // Bail out if either start or end is negative.
4058 generate_negative_guard(start, bailout, &start);
4059 generate_negative_guard(end, bailout, &end);
4060
4061 Node* length = end;
4062 if (_gvn.type(start) != TypeInt::ZERO) {
4063 length = _gvn.transform(new SubINode(end, start));
4064 }
4065
4066 // Bail out if length is negative.
4067 // Without this the new_array would throw
4068 // NegativeArraySizeException but IllegalArgumentException is what
4069 // should be thrown
4070 generate_negative_guard(length, bailout, &length);
4071
4072 if (bailout->req() > 1) {
4073 PreserveJVMState pjvms(this);
4074 set_control(_gvn.transform(bailout));
4075 uncommon_trap(Deoptimization::Reason_intrinsic,
4076 Deoptimization::Action_maybe_recompile);
4077 }
4078
4079 if (!stopped()) {
4080 // How many elements will we copy from the original?
4081 // The answer is MinI(orig_length - start, length).
4082 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4083 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4084
4085 // Generate a direct call to the right arraycopy function(s).
4086 // We know the copy is disjoint but we might not know if the
4087 // oop stores need checking.
4088 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4089 // This will fail a store-check if x contains any non-nulls.
4090
4091 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4092 // loads/stores but it is legal only if we're sure the
4093 // Arrays.copyOf would succeed. So we need all input arguments
4094 // to the copyOf to be validated, including that the copy to the
4095 // new array won't trigger an ArrayStoreException. That subtype
4096 // check can be optimized if we know something on the type of
4097 // the input array from type speculation.
4098 if (_gvn.type(klass_node)->singleton()) {
4099 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4100 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4101
4102 int test = C->static_subtype_check(superk, subk);
4103 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4104 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4105 if (t_original->speculative_type() != NULL) {
4106 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4107 }
4108 }
4109 }
4110
4111 bool validated = false;
4112 // Reason_class_check rather than Reason_intrinsic because we
4113 // want to intrinsify even if this traps.
4114 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4115 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4116 klass_node);
4117
4118 if (not_subtype_ctrl != top()) {
4119 PreserveJVMState pjvms(this);
4120 set_control(not_subtype_ctrl);
4121 uncommon_trap(Deoptimization::Reason_class_check,
4122 Deoptimization::Action_make_not_entrant);
4123 assert(stopped(), "Should be stopped");
4124 }
4125 validated = true;
4126 }
4127
4128 if (!stopped()) {
4129 newcopy = new_array(klass_node, length, 0); // no arguments to push
4130
4131 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4132 load_object_klass(original), klass_node);
4133 if (!is_copyOfRange) {
4134 ac->set_copyof(validated);
4135 } else {
4136 ac->set_copyofrange(validated);
4137 }
4138 Node* n = _gvn.transform(ac);
4139 if (n == ac) {
4140 ac->connect_outputs(this);
4141 } else {
4142 assert(validated, "shouldn't transform if all arguments not validated");
4143 set_all_memory(n);
4144 }
4145 }
4146 }
4147 } // original reexecute is set back here
4148
4149 C->set_has_split_ifs(true); // Has chance for split-if optimization
4150 if (!stopped()) {
4151 set_result(newcopy);
4152 }
4153 return true;
4154 }
4155
4156
4157 //----------------------generate_virtual_guard---------------------------
4158 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4159 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4160 RegionNode* slow_region) {
4161 ciMethod* method = callee();
4162 int vtable_index = method->vtable_index();
4163 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4164 "bad index %d", vtable_index);
4165 // Get the Method* out of the appropriate vtable entry.
4166 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4167 vtable_index*vtableEntry::size_in_bytes() +
4168 vtableEntry::method_offset_in_bytes();
4169 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4170 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4171
4172 // Compare the target method with the expected method (e.g., Object.hashCode).
4173 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4174
4175 Node* native_call = makecon(native_call_addr);
4176 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4177 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4178
4179 return generate_slow_guard(test_native, slow_region);
4180 }
4181
4182 //-----------------------generate_method_call----------------------------
4183 // Use generate_method_call to make a slow-call to the real
4184 // method if the fast path fails. An alternative would be to
4185 // use a stub like OptoRuntime::slow_arraycopy_Java.
4186 // This only works for expanding the current library call,
4187 // not another intrinsic. (E.g., don't use this for making an
4188 // arraycopy call inside of the copyOf intrinsic.)
4189 CallJavaNode*
4190 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4191 // When compiling the intrinsic method itself, do not use this technique.
4192 guarantee(callee() != C->method(), "cannot make slow-call to self");
4193
4194 ciMethod* method = callee();
4195 // ensure the JVMS we have will be correct for this call
4196 guarantee(method_id == method->intrinsic_id(), "must match");
4197
4198 const TypeFunc* tf = TypeFunc::make(method);
4199 CallJavaNode* slow_call;
4200 if (is_static) {
4201 assert(!is_virtual, "");
4202 slow_call = new CallStaticJavaNode(C, tf,
4203 SharedRuntime::get_resolve_static_call_stub(),
4204 method, bci());
4205 } else if (is_virtual) {
4206 null_check_receiver();
4207 int vtable_index = Method::invalid_vtable_index;
4208 if (UseInlineCaches) {
4209 // Suppress the vtable call
4210 } else {
4211 // hashCode and clone are not a miranda methods,
4212 // so the vtable index is fixed.
4213 // No need to use the linkResolver to get it.
4214 vtable_index = method->vtable_index();
4215 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4216 "bad index %d", vtable_index);
4217 }
4218 slow_call = new CallDynamicJavaNode(tf,
4219 SharedRuntime::get_resolve_virtual_call_stub(),
4220 method, vtable_index, bci());
4221 } else { // neither virtual nor static: opt_virtual
4222 null_check_receiver();
4223 slow_call = new CallStaticJavaNode(C, tf,
4224 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4225 method, bci());
4226 slow_call->set_optimized_virtual(true);
4227 }
4228 set_arguments_for_java_call(slow_call);
4229 set_edges_for_java_call(slow_call);
4230 return slow_call;
4231 }
4232
4233
4234 /**
4235 * Build special case code for calls to hashCode on an object. This call may
4236 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4237 * slightly different code.
4238 */
4239 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4240 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4241 assert(!(is_virtual && is_static), "either virtual, special, or static");
4242
4243 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4244
4245 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4246 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4247 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4248 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4249 Node* obj = NULL;
4250 if (!is_static) {
4251 // Check for hashing null object
4252 obj = null_check_receiver();
4253 if (stopped()) return true; // unconditionally null
4254 result_reg->init_req(_null_path, top());
4255 result_val->init_req(_null_path, top());
4256 } else {
4257 // Do a null check, and return zero if null.
4258 // System.identityHashCode(null) == 0
4259 obj = argument(0);
4260 Node* null_ctl = top();
4261 obj = null_check_oop(obj, &null_ctl);
4262 result_reg->init_req(_null_path, null_ctl);
4263 result_val->init_req(_null_path, _gvn.intcon(0));
4264 }
4265
4266 // Unconditionally null? Then return right away.
4267 if (stopped()) {
4268 set_control( result_reg->in(_null_path));
4269 if (!stopped())
4270 set_result(result_val->in(_null_path));
4271 return true;
4272 }
4273
4274 // We only go to the fast case code if we pass a number of guards. The
4275 // paths which do not pass are accumulated in the slow_region.
4276 RegionNode* slow_region = new RegionNode(1);
4277 record_for_igvn(slow_region);
4278
4279 // If this is a virtual call, we generate a funny guard. We pull out
4280 // the vtable entry corresponding to hashCode() from the target object.
4281 // If the target method which we are calling happens to be the native
4282 // Object hashCode() method, we pass the guard. We do not need this
4283 // guard for non-virtual calls -- the caller is known to be the native
4284 // Object hashCode().
4285 if (is_virtual) {
4286 // After null check, get the object's klass.
4287 Node* obj_klass = load_object_klass(obj);
4288 generate_virtual_guard(obj_klass, slow_region);
4289 }
4290
4291 // Get the header out of the object, use LoadMarkNode when available
4292 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4293 // The control of the load must be NULL. Otherwise, the load can move before
4294 // the null check after castPP removal.
4295 Node* no_ctrl = NULL;
4296 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4297
4298 // Test the header to see if it is unlocked.
4299 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4300 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4301 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4302 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4303 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4304
4305 generate_slow_guard(test_unlocked, slow_region);
4306
4307 // Get the hash value and check to see that it has been properly assigned.
4308 // We depend on hash_mask being at most 32 bits and avoid the use of
4309 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4310 // vm: see markOop.hpp.
4311 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4312 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4313 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4314 // This hack lets the hash bits live anywhere in the mark object now, as long
4315 // as the shift drops the relevant bits into the low 32 bits. Note that
4316 // Java spec says that HashCode is an int so there's no point in capturing
4317 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4318 hshifted_header = ConvX2I(hshifted_header);
4319 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4320
4321 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4322 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4323 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4324
4325 generate_slow_guard(test_assigned, slow_region);
4326
4327 Node* init_mem = reset_memory();
4328 // fill in the rest of the null path:
4329 result_io ->init_req(_null_path, i_o());
4330 result_mem->init_req(_null_path, init_mem);
4331
4332 result_val->init_req(_fast_path, hash_val);
4333 result_reg->init_req(_fast_path, control());
4334 result_io ->init_req(_fast_path, i_o());
4335 result_mem->init_req(_fast_path, init_mem);
4336
4337 // Generate code for the slow case. We make a call to hashCode().
4338 set_control(_gvn.transform(slow_region));
4339 if (!stopped()) {
4340 // No need for PreserveJVMState, because we're using up the present state.
4341 set_all_memory(init_mem);
4342 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4343 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4344 Node* slow_result = set_results_for_java_call(slow_call);
4345 // this->control() comes from set_results_for_java_call
4346 result_reg->init_req(_slow_path, control());
4347 result_val->init_req(_slow_path, slow_result);
4348 result_io ->set_req(_slow_path, i_o());
4349 result_mem ->set_req(_slow_path, reset_memory());
4350 }
4351
4352 // Return the combined state.
4353 set_i_o( _gvn.transform(result_io) );
4354 set_all_memory( _gvn.transform(result_mem));
4355
4356 set_result(result_reg, result_val);
4357 return true;
4358 }
4359
4360 //---------------------------inline_native_getClass----------------------------
4361 // public final native Class<?> java.lang.Object.getClass();
4362 //
4363 // Build special case code for calls to getClass on an object.
4364 bool LibraryCallKit::inline_native_getClass() {
4365 Node* obj = null_check_receiver();
4366 if (stopped()) return true;
4367 set_result(load_mirror_from_klass(load_object_klass(obj)));
4368 return true;
4369 }
4370
4371 //-----------------inline_native_Reflection_getCallerClass---------------------
4372 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4373 //
4374 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4375 //
4376 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4377 // in that it must skip particular security frames and checks for
4378 // caller sensitive methods.
4379 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4380 #ifndef PRODUCT
4381 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4382 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4383 }
4384 #endif
4385
4386 if (!jvms()->has_method()) {
4387 #ifndef PRODUCT
4388 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4389 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4390 }
4391 #endif
4392 return false;
4393 }
4394
4395 // Walk back up the JVM state to find the caller at the required
4396 // depth.
4397 JVMState* caller_jvms = jvms();
4398
4399 // Cf. JVM_GetCallerClass
4400 // NOTE: Start the loop at depth 1 because the current JVM state does
4401 // not include the Reflection.getCallerClass() frame.
4402 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4403 ciMethod* m = caller_jvms->method();
4404 switch (n) {
4405 case 0:
4406 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4407 break;
4408 case 1:
4409 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4410 if (!m->caller_sensitive()) {
4411 #ifndef PRODUCT
4412 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4413 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4414 }
4415 #endif
4416 return false; // bail-out; let JVM_GetCallerClass do the work
4417 }
4418 break;
4419 default:
4420 if (!m->is_ignored_by_security_stack_walk()) {
4421 // We have reached the desired frame; return the holder class.
4422 // Acquire method holder as java.lang.Class and push as constant.
4423 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4424 ciInstance* caller_mirror = caller_klass->java_mirror();
4425 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4426
4427 #ifndef PRODUCT
4428 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4429 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());
4430 tty->print_cr(" JVM state at this point:");
4431 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4432 ciMethod* m = jvms()->of_depth(i)->method();
4433 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4434 }
4435 }
4436 #endif
4437 return true;
4438 }
4439 break;
4440 }
4441 }
4442
4443 #ifndef PRODUCT
4444 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4445 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4446 tty->print_cr(" JVM state at this point:");
4447 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4448 ciMethod* m = jvms()->of_depth(i)->method();
4449 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4450 }
4451 }
4452 #endif
4453
4454 return false; // bail-out; let JVM_GetCallerClass do the work
4455 }
4456
4457 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4458 Node* arg = argument(0);
4459 Node* result = NULL;
4460
4461 switch (id) {
4462 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4463 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4464 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4465 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4466
4467 case vmIntrinsics::_doubleToLongBits: {
4468 // two paths (plus control) merge in a wood
4469 RegionNode *r = new RegionNode(3);
4470 Node *phi = new PhiNode(r, TypeLong::LONG);
4471
4472 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4473 // Build the boolean node
4474 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4475
4476 // Branch either way.
4477 // NaN case is less traveled, which makes all the difference.
4478 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4479 Node *opt_isnan = _gvn.transform(ifisnan);
4480 assert( opt_isnan->is_If(), "Expect an IfNode");
4481 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4482 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4483
4484 set_control(iftrue);
4485
4486 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4487 Node *slow_result = longcon(nan_bits); // return NaN
4488 phi->init_req(1, _gvn.transform( slow_result ));
4489 r->init_req(1, iftrue);
4490
4491 // Else fall through
4492 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4493 set_control(iffalse);
4494
4495 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4496 r->init_req(2, iffalse);
4497
4498 // Post merge
4499 set_control(_gvn.transform(r));
4500 record_for_igvn(r);
4501
4502 C->set_has_split_ifs(true); // Has chance for split-if optimization
4503 result = phi;
4504 assert(result->bottom_type()->isa_long(), "must be");
4505 break;
4506 }
4507
4508 case vmIntrinsics::_floatToIntBits: {
4509 // two paths (plus control) merge in a wood
4510 RegionNode *r = new RegionNode(3);
4511 Node *phi = new PhiNode(r, TypeInt::INT);
4512
4513 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4514 // Build the boolean node
4515 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4516
4517 // Branch either way.
4518 // NaN case is less traveled, which makes all the difference.
4519 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4520 Node *opt_isnan = _gvn.transform(ifisnan);
4521 assert( opt_isnan->is_If(), "Expect an IfNode");
4522 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4523 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4524
4525 set_control(iftrue);
4526
4527 static const jint nan_bits = 0x7fc00000;
4528 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4529 phi->init_req(1, _gvn.transform( slow_result ));
4530 r->init_req(1, iftrue);
4531
4532 // Else fall through
4533 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4534 set_control(iffalse);
4535
4536 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4537 r->init_req(2, iffalse);
4538
4539 // Post merge
4540 set_control(_gvn.transform(r));
4541 record_for_igvn(r);
4542
4543 C->set_has_split_ifs(true); // Has chance for split-if optimization
4544 result = phi;
4545 assert(result->bottom_type()->isa_int(), "must be");
4546 break;
4547 }
4548
4549 default:
4550 fatal_unexpected_iid(id);
4551 break;
4552 }
4553 set_result(_gvn.transform(result));
4554 return true;
4555 }
4556
4557 //----------------------inline_unsafe_copyMemory-------------------------
4558 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4559 bool LibraryCallKit::inline_unsafe_copyMemory() {
4560 if (callee()->is_static()) return false; // caller must have the capability!
4561 null_check_receiver(); // null-check receiver
4562 if (stopped()) return true;
4563
4564 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4565
4566 Node* src_ptr = argument(1); // type: oop
4567 Node* src_off = ConvL2X(argument(2)); // type: long
4568 Node* dst_ptr = argument(4); // type: oop
4569 Node* dst_off = ConvL2X(argument(5)); // type: long
4570 Node* size = ConvL2X(argument(7)); // type: long
4571
4572 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4573 "fieldOffset must be byte-scaled");
4574
4575 Node* src = make_unsafe_address(src_ptr, src_off);
4576 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4577
4578 // Conservatively insert a memory barrier on all memory slices.
4579 // Do not let writes of the copy source or destination float below the copy.
4580 insert_mem_bar(Op_MemBarCPUOrder);
4581
4582 // Call it. Note that the length argument is not scaled.
4583 make_runtime_call(RC_LEAF|RC_NO_FP,
4584 OptoRuntime::fast_arraycopy_Type(),
4585 StubRoutines::unsafe_arraycopy(),
4586 "unsafe_arraycopy",
4587 TypeRawPtr::BOTTOM,
4588 src, dst, size XTOP);
4589
4590 // Do not let reads of the copy destination float above the copy.
4591 insert_mem_bar(Op_MemBarCPUOrder);
4592
4593 return true;
4594 }
4595
4596 //------------------------clone_coping-----------------------------------
4597 // Helper function for inline_native_clone.
4598 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4599 assert(obj_size != NULL, "");
4600 Node* raw_obj = alloc_obj->in(1);
4601 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4602
4603 AllocateNode* alloc = NULL;
4604 if (ReduceBulkZeroing) {
4605 // We will be completely responsible for initializing this object -
4606 // mark Initialize node as complete.
4607 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4608 // The object was just allocated - there should be no any stores!
4609 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4610 // Mark as complete_with_arraycopy so that on AllocateNode
4611 // expansion, we know this AllocateNode is initialized by an array
4612 // copy and a StoreStore barrier exists after the array copy.
4613 alloc->initialization()->set_complete_with_arraycopy();
4614 }
4615
4616 // Copy the fastest available way.
4617 // TODO: generate fields copies for small objects instead.
4618 Node* src = obj;
4619 Node* dest = alloc_obj;
4620 Node* size = _gvn.transform(obj_size);
4621
4622 // Exclude the header but include array length to copy by 8 bytes words.
4623 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4624 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4625 instanceOopDesc::base_offset_in_bytes();
4626 // base_off:
4627 // 8 - 32-bit VM
4628 // 12 - 64-bit VM, compressed klass
4629 // 16 - 64-bit VM, normal klass
4630 if (base_off % BytesPerLong != 0) {
4631 assert(UseCompressedClassPointers, "");
4632 if (is_array) {
4633 // Exclude length to copy by 8 bytes words.
4634 base_off += sizeof(int);
4635 } else {
4636 // Include klass to copy by 8 bytes words.
4637 base_off = instanceOopDesc::klass_offset_in_bytes();
4638 }
4639 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4640 }
4641 src = basic_plus_adr(src, base_off);
4642 dest = basic_plus_adr(dest, base_off);
4643
4644 // Compute the length also, if needed:
4645 Node* countx = size;
4646 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4647 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4648
4649 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4650
4651 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4652 ac->set_clonebasic();
4653 Node* n = _gvn.transform(ac);
4654 if (n == ac) {
4655 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4656 } else {
4657 set_all_memory(n);
4658 }
4659
4660 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4661 if (card_mark) {
4662 assert(!is_array, "");
4663 // Put in store barrier for any and all oops we are sticking
4664 // into this object. (We could avoid this if we could prove
4665 // that the object type contains no oop fields at all.)
4666 Node* no_particular_value = NULL;
4667 Node* no_particular_field = NULL;
4668 int raw_adr_idx = Compile::AliasIdxRaw;
4669 post_barrier(control(),
4670 memory(raw_adr_type),
4671 alloc_obj,
4672 no_particular_field,
4673 raw_adr_idx,
4674 no_particular_value,
4675 T_OBJECT,
4676 false);
4677 }
4678
4679 // Do not let reads from the cloned object float above the arraycopy.
4680 if (alloc != NULL) {
4681 // Do not let stores that initialize this object be reordered with
4682 // a subsequent store that would make this object accessible by
4683 // other threads.
4684 // Record what AllocateNode this StoreStore protects so that
4685 // escape analysis can go from the MemBarStoreStoreNode to the
4686 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4687 // based on the escape status of the AllocateNode.
4688 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4689 } else {
4690 insert_mem_bar(Op_MemBarCPUOrder);
4691 }
4692 }
4693
4694 //------------------------inline_native_clone----------------------------
4695 // protected native Object java.lang.Object.clone();
4696 //
4697 // Here are the simple edge cases:
4698 // null receiver => normal trap
4699 // virtual and clone was overridden => slow path to out-of-line clone
4700 // not cloneable or finalizer => slow path to out-of-line Object.clone
4701 //
4702 // The general case has two steps, allocation and copying.
4703 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4704 //
4705 // Copying also has two cases, oop arrays and everything else.
4706 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4707 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4708 //
4709 // These steps fold up nicely if and when the cloned object's klass
4710 // can be sharply typed as an object array, a type array, or an instance.
4711 //
4712 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4713 PhiNode* result_val;
4714
4715 // Set the reexecute bit for the interpreter to reexecute
4716 // the bytecode that invokes Object.clone if deoptimization happens.
4717 { PreserveReexecuteState preexecs(this);
4718 jvms()->set_should_reexecute(true);
4719
4720 Node* obj = null_check_receiver();
4721 if (stopped()) return true;
4722
4723 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4724
4725 // If we are going to clone an instance, we need its exact type to
4726 // know the number and types of fields to convert the clone to
4727 // loads/stores. Maybe a speculative type can help us.
4728 if (!obj_type->klass_is_exact() &&
4729 obj_type->speculative_type() != NULL &&
4730 obj_type->speculative_type()->is_instance_klass()) {
4731 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4732 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4733 !spec_ik->has_injected_fields()) {
4734 ciKlass* k = obj_type->klass();
4735 if (!k->is_instance_klass() ||
4736 k->as_instance_klass()->is_interface() ||
4737 k->as_instance_klass()->has_subklass()) {
4738 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4739 }
4740 }
4741 }
4742
4743 Node* obj_klass = load_object_klass(obj);
4744 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4745 const TypeOopPtr* toop = ((tklass != NULL)
4746 ? tklass->as_instance_type()
4747 : TypeInstPtr::NOTNULL);
4748
4749 // Conservatively insert a memory barrier on all memory slices.
4750 // Do not let writes into the original float below the clone.
4751 insert_mem_bar(Op_MemBarCPUOrder);
4752
4753 // paths into result_reg:
4754 enum {
4755 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4756 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4757 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4758 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4759 PATH_LIMIT
4760 };
4761 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4762 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4763 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4764 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4765 record_for_igvn(result_reg);
4766
4767 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4768 int raw_adr_idx = Compile::AliasIdxRaw;
4769
4770 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4771 if (array_ctl != NULL) {
4772 // It's an array.
4773 PreserveJVMState pjvms(this);
4774 set_control(array_ctl);
4775 Node* obj_length = load_array_length(obj);
4776 Node* obj_size = NULL;
4777 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4778
4779 if (!use_ReduceInitialCardMarks()) {
4780 // If it is an oop array, it requires very special treatment,
4781 // because card marking is required on each card of the array.
4782 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4783 if (is_obja != NULL) {
4784 PreserveJVMState pjvms2(this);
4785 set_control(is_obja);
4786 // Generate a direct call to the right arraycopy function(s).
4787 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4788 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4789 ac->set_cloneoop();
4790 Node* n = _gvn.transform(ac);
4791 assert(n == ac, "cannot disappear");
4792 ac->connect_outputs(this);
4793
4794 result_reg->init_req(_objArray_path, control());
4795 result_val->init_req(_objArray_path, alloc_obj);
4796 result_i_o ->set_req(_objArray_path, i_o());
4797 result_mem ->set_req(_objArray_path, reset_memory());
4798 }
4799 }
4800 // Otherwise, there are no card marks to worry about.
4801 // (We can dispense with card marks if we know the allocation
4802 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4803 // causes the non-eden paths to take compensating steps to
4804 // simulate a fresh allocation, so that no further
4805 // card marks are required in compiled code to initialize
4806 // the object.)
4807
4808 if (!stopped()) {
4809 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4810
4811 // Present the results of the copy.
4812 result_reg->init_req(_array_path, control());
4813 result_val->init_req(_array_path, alloc_obj);
4814 result_i_o ->set_req(_array_path, i_o());
4815 result_mem ->set_req(_array_path, reset_memory());
4816 }
4817 }
4818
4819 // We only go to the instance fast case code if we pass a number of guards.
4820 // The paths which do not pass are accumulated in the slow_region.
4821 RegionNode* slow_region = new RegionNode(1);
4822 record_for_igvn(slow_region);
4823 if (!stopped()) {
4824 // It's an instance (we did array above). Make the slow-path tests.
4825 // If this is a virtual call, we generate a funny guard. We grab
4826 // the vtable entry corresponding to clone() from the target object.
4827 // If the target method which we are calling happens to be the
4828 // Object clone() method, we pass the guard. We do not need this
4829 // guard for non-virtual calls; the caller is known to be the native
4830 // Object clone().
4831 if (is_virtual) {
4832 generate_virtual_guard(obj_klass, slow_region);
4833 }
4834
4835 // The object must be easily cloneable and must not have a finalizer.
4836 // Both of these conditions may be checked in a single test.
4837 // We could optimize the test further, but we don't care.
4838 generate_access_flags_guard(obj_klass,
4839 // Test both conditions:
4840 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4841 // Must be cloneable but not finalizer:
4842 JVM_ACC_IS_CLONEABLE_FAST,
4843 slow_region);
4844 }
4845
4846 if (!stopped()) {
4847 // It's an instance, and it passed the slow-path tests.
4848 PreserveJVMState pjvms(this);
4849 Node* obj_size = NULL;
4850 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4851 // is reexecuted if deoptimization occurs and there could be problems when merging
4852 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4853 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4854
4855 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4856
4857 // Present the results of the slow call.
4858 result_reg->init_req(_instance_path, control());
4859 result_val->init_req(_instance_path, alloc_obj);
4860 result_i_o ->set_req(_instance_path, i_o());
4861 result_mem ->set_req(_instance_path, reset_memory());
4862 }
4863
4864 // Generate code for the slow case. We make a call to clone().
4865 set_control(_gvn.transform(slow_region));
4866 if (!stopped()) {
4867 PreserveJVMState pjvms(this);
4868 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4869 Node* slow_result = set_results_for_java_call(slow_call);
4870 // this->control() comes from set_results_for_java_call
4871 result_reg->init_req(_slow_path, control());
4872 result_val->init_req(_slow_path, slow_result);
4873 result_i_o ->set_req(_slow_path, i_o());
4874 result_mem ->set_req(_slow_path, reset_memory());
4875 }
4876
4877 // Return the combined state.
4878 set_control( _gvn.transform(result_reg));
4879 set_i_o( _gvn.transform(result_i_o));
4880 set_all_memory( _gvn.transform(result_mem));
4881 } // original reexecute is set back here
4882
4883 set_result(_gvn.transform(result_val));
4884 return true;
4885 }
4886
4887 // If we have a tighly coupled allocation, the arraycopy may take care
4888 // of the array initialization. If one of the guards we insert between
4889 // the allocation and the arraycopy causes a deoptimization, an
4890 // unitialized array will escape the compiled method. To prevent that
4891 // we set the JVM state for uncommon traps between the allocation and
4892 // the arraycopy to the state before the allocation so, in case of
4893 // deoptimization, we'll reexecute the allocation and the
4894 // initialization.
4895 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4896 if (alloc != NULL) {
4897 ciMethod* trap_method = alloc->jvms()->method();
4898 int trap_bci = alloc->jvms()->bci();
4899
4900 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4901 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4902 // Make sure there's no store between the allocation and the
4903 // arraycopy otherwise visible side effects could be rexecuted
4904 // in case of deoptimization and cause incorrect execution.
4905 bool no_interfering_store = true;
4906 Node* mem = alloc->in(TypeFunc::Memory);
4907 if (mem->is_MergeMem()) {
4908 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4909 Node* n = mms.memory();
4910 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4911 assert(n->is_Store(), "what else?");
4912 no_interfering_store = false;
4913 break;
4914 }
4915 }
4916 } else {
4917 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4918 Node* n = mms.memory();
4919 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4920 assert(n->is_Store(), "what else?");
4921 no_interfering_store = false;
4922 break;
4923 }
4924 }
4925 }
4926
4927 if (no_interfering_store) {
4928 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4929 uint size = alloc->req();
4930 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4931 old_jvms->set_map(sfpt);
4932 for (uint i = 0; i < size; i++) {
4933 sfpt->init_req(i, alloc->in(i));
4934 }
4935 // re-push array length for deoptimization
4936 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4937 old_jvms->set_sp(old_jvms->sp()+1);
4938 old_jvms->set_monoff(old_jvms->monoff()+1);
4939 old_jvms->set_scloff(old_jvms->scloff()+1);
4940 old_jvms->set_endoff(old_jvms->endoff()+1);
4941 old_jvms->set_should_reexecute(true);
4942
4943 sfpt->set_i_o(map()->i_o());
4944 sfpt->set_memory(map()->memory());
4945 sfpt->set_control(map()->control());
4946
4947 JVMState* saved_jvms = jvms();
4948 saved_reexecute_sp = _reexecute_sp;
4949
4950 set_jvms(sfpt->jvms());
4951 _reexecute_sp = jvms()->sp();
4952
4953 return saved_jvms;
4954 }
4955 }
4956 }
4957 return NULL;
4958 }
4959
4960 // In case of a deoptimization, we restart execution at the
4961 // allocation, allocating a new array. We would leave an uninitialized
4962 // array in the heap that GCs wouldn't expect. Move the allocation
4963 // after the traps so we don't allocate the array if we
4964 // deoptimize. This is possible because tightly_coupled_allocation()
4965 // guarantees there's no observer of the allocated array at this point
4966 // and the control flow is simple enough.
4967 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4968 int saved_reexecute_sp, uint new_idx) {
4969 if (saved_jvms != NULL && !stopped()) {
4970 assert(alloc != NULL, "only with a tightly coupled allocation");
4971 // restore JVM state to the state at the arraycopy
4972 saved_jvms->map()->set_control(map()->control());
4973 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4974 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4975 // If we've improved the types of some nodes (null check) while
4976 // emitting the guards, propagate them to the current state
4977 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4978 set_jvms(saved_jvms);
4979 _reexecute_sp = saved_reexecute_sp;
4980
4981 // Remove the allocation from above the guards
4982 CallProjections callprojs;
4983 alloc->extract_projections(&callprojs, true);
4984 InitializeNode* init = alloc->initialization();
4985 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4986 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4987 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4988 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4989
4990 // move the allocation here (after the guards)
4991 _gvn.hash_delete(alloc);
4992 alloc->set_req(TypeFunc::Control, control());
4993 alloc->set_req(TypeFunc::I_O, i_o());
4994 Node *mem = reset_memory();
4995 set_all_memory(mem);
4996 alloc->set_req(TypeFunc::Memory, mem);
4997 set_control(init->proj_out(TypeFunc::Control));
4998 set_i_o(callprojs.fallthrough_ioproj);
4999
5000 // Update memory as done in GraphKit::set_output_for_allocation()
5001 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5002 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5003 if (ary_type->isa_aryptr() && length_type != NULL) {
5004 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5005 }
5006 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5007 int elemidx = C->get_alias_index(telemref);
5008 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
5009 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
5010
5011 Node* allocx = _gvn.transform(alloc);
5012 assert(allocx == alloc, "where has the allocation gone?");
5013 assert(dest->is_CheckCastPP(), "not an allocation result?");
5014
5015 _gvn.hash_delete(dest);
5016 dest->set_req(0, control());
5017 Node* destx = _gvn.transform(dest);
5018 assert(destx == dest, "where has the allocation result gone?");
5019 }
5020 }
5021
5022
5023 //------------------------------inline_arraycopy-----------------------
5024 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5025 // Object dest, int destPos,
5026 // int length);
5027 bool LibraryCallKit::inline_arraycopy() {
5028 // Get the arguments.
5029 Node* src = argument(0); // type: oop
5030 Node* src_offset = argument(1); // type: int
5031 Node* dest = argument(2); // type: oop
5032 Node* dest_offset = argument(3); // type: int
5033 Node* length = argument(4); // type: int
5034
5035 uint new_idx = C->unique();
5036
5037 // Check for allocation before we add nodes that would confuse
5038 // tightly_coupled_allocation()
5039 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5040
5041 int saved_reexecute_sp = -1;
5042 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5043 // See arraycopy_restore_alloc_state() comment
5044 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5045 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5046 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5047 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5048
5049 // The following tests must be performed
5050 // (1) src and dest are arrays.
5051 // (2) src and dest arrays must have elements of the same BasicType
5052 // (3) src and dest must not be null.
5053 // (4) src_offset must not be negative.
5054 // (5) dest_offset must not be negative.
5055 // (6) length must not be negative.
5056 // (7) src_offset + length must not exceed length of src.
5057 // (8) dest_offset + length must not exceed length of dest.
5058 // (9) each element of an oop array must be assignable
5059
5060 // (3) src and dest must not be null.
5061 // always do this here because we need the JVM state for uncommon traps
5062 Node* null_ctl = top();
5063 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5064 assert(null_ctl->is_top(), "no null control here");
5065 dest = null_check(dest, T_ARRAY);
5066
5067 if (!can_emit_guards) {
5068 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5069 // guards but the arraycopy node could still take advantage of a
5070 // tightly allocated allocation. tightly_coupled_allocation() is
5071 // called again to make sure it takes the null check above into
5072 // account: the null check is mandatory and if it caused an
5073 // uncommon trap to be emitted then the allocation can't be
5074 // considered tightly coupled in this context.
5075 alloc = tightly_coupled_allocation(dest, NULL);
5076 }
5077
5078 bool validated = false;
5079
5080 const Type* src_type = _gvn.type(src);
5081 const Type* dest_type = _gvn.type(dest);
5082 const TypeAryPtr* top_src = src_type->isa_aryptr();
5083 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5084
5085 // Do we have the type of src?
5086 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5087 // Do we have the type of dest?
5088 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5089 // Is the type for src from speculation?
5090 bool src_spec = false;
5091 // Is the type for dest from speculation?
5092 bool dest_spec = false;
5093
5094 if ((!has_src || !has_dest) && can_emit_guards) {
5095 // We don't have sufficient type information, let's see if
5096 // speculative types can help. We need to have types for both src
5097 // and dest so that it pays off.
5098
5099 // Do we already have or could we have type information for src
5100 bool could_have_src = has_src;
5101 // Do we already have or could we have type information for dest
5102 bool could_have_dest = has_dest;
5103
5104 ciKlass* src_k = NULL;
5105 if (!has_src) {
5106 src_k = src_type->speculative_type_not_null();
5107 if (src_k != NULL && src_k->is_array_klass()) {
5108 could_have_src = true;
5109 }
5110 }
5111
5112 ciKlass* dest_k = NULL;
5113 if (!has_dest) {
5114 dest_k = dest_type->speculative_type_not_null();
5115 if (dest_k != NULL && dest_k->is_array_klass()) {
5116 could_have_dest = true;
5117 }
5118 }
5119
5120 if (could_have_src && could_have_dest) {
5121 // This is going to pay off so emit the required guards
5122 if (!has_src) {
5123 src = maybe_cast_profiled_obj(src, src_k, true);
5124 src_type = _gvn.type(src);
5125 top_src = src_type->isa_aryptr();
5126 has_src = (top_src != NULL && top_src->klass() != NULL);
5127 src_spec = true;
5128 }
5129 if (!has_dest) {
5130 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5131 dest_type = _gvn.type(dest);
5132 top_dest = dest_type->isa_aryptr();
5133 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5134 dest_spec = true;
5135 }
5136 }
5137 }
5138
5139 if (has_src && has_dest && can_emit_guards) {
5140 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5141 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5142 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5143 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5144
5145 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5146 // If both arrays are object arrays then having the exact types
5147 // for both will remove the need for a subtype check at runtime
5148 // before the call and may make it possible to pick a faster copy
5149 // routine (without a subtype check on every element)
5150 // Do we have the exact type of src?
5151 bool could_have_src = src_spec;
5152 // Do we have the exact type of dest?
5153 bool could_have_dest = dest_spec;
5154 ciKlass* src_k = top_src->klass();
5155 ciKlass* dest_k = top_dest->klass();
5156 if (!src_spec) {
5157 src_k = src_type->speculative_type_not_null();
5158 if (src_k != NULL && src_k->is_array_klass()) {
5159 could_have_src = true;
5160 }
5161 }
5162 if (!dest_spec) {
5163 dest_k = dest_type->speculative_type_not_null();
5164 if (dest_k != NULL && dest_k->is_array_klass()) {
5165 could_have_dest = true;
5166 }
5167 }
5168 if (could_have_src && could_have_dest) {
5169 // If we can have both exact types, emit the missing guards
5170 if (could_have_src && !src_spec) {
5171 src = maybe_cast_profiled_obj(src, src_k, true);
5172 }
5173 if (could_have_dest && !dest_spec) {
5174 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5175 }
5176 }
5177 }
5178 }
5179
5180 ciMethod* trap_method = method();
5181 int trap_bci = bci();
5182 if (saved_jvms != NULL) {
5183 trap_method = alloc->jvms()->method();
5184 trap_bci = alloc->jvms()->bci();
5185 }
5186
5187 bool negative_length_guard_generated = false;
5188
5189 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5190 can_emit_guards &&
5191 !src->is_top() && !dest->is_top()) {
5192 // validate arguments: enables transformation the ArrayCopyNode
5193 validated = true;
5194
5195 RegionNode* slow_region = new RegionNode(1);
5196 record_for_igvn(slow_region);
5197
5198 // (1) src and dest are arrays.
5199 generate_non_array_guard(load_object_klass(src), slow_region);
5200 generate_non_array_guard(load_object_klass(dest), slow_region);
5201
5202 // (2) src and dest arrays must have elements of the same BasicType
5203 // done at macro expansion or at Ideal transformation time
5204
5205 // (4) src_offset must not be negative.
5206 generate_negative_guard(src_offset, slow_region);
5207
5208 // (5) dest_offset must not be negative.
5209 generate_negative_guard(dest_offset, slow_region);
5210
5211 // (7) src_offset + length must not exceed length of src.
5212 generate_limit_guard(src_offset, length,
5213 load_array_length(src),
5214 slow_region);
5215
5216 // (8) dest_offset + length must not exceed length of dest.
5217 generate_limit_guard(dest_offset, length,
5218 load_array_length(dest),
5219 slow_region);
5220
5221 // (6) length must not be negative.
5222 // This is also checked in generate_arraycopy() during macro expansion, but
5223 // we also have to check it here for the case where the ArrayCopyNode will
5224 // be eliminated by Escape Analysis.
5225 if (EliminateAllocations) {
5226 generate_negative_guard(length, slow_region);
5227 negative_length_guard_generated = true;
5228 }
5229
5230 // (9) each element of an oop array must be assignable
5231 Node* src_klass = load_object_klass(src);
5232 Node* dest_klass = load_object_klass(dest);
5233 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5234
5235 if (not_subtype_ctrl != top()) {
5236 PreserveJVMState pjvms(this);
5237 set_control(not_subtype_ctrl);
5238 uncommon_trap(Deoptimization::Reason_intrinsic,
5239 Deoptimization::Action_make_not_entrant);
5240 assert(stopped(), "Should be stopped");
5241 }
5242 {
5243 PreserveJVMState pjvms(this);
5244 set_control(_gvn.transform(slow_region));
5245 uncommon_trap(Deoptimization::Reason_intrinsic,
5246 Deoptimization::Action_make_not_entrant);
5247 assert(stopped(), "Should be stopped");
5248 }
5249 }
5250
5251 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5252
5253 if (stopped()) {
5254 return true;
5255 }
5256
5257 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5258 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5259 // so the compiler has a chance to eliminate them: during macro expansion,
5260 // we have to set their control (CastPP nodes are eliminated).
5261 load_object_klass(src), load_object_klass(dest),
5262 load_array_length(src), load_array_length(dest));
5263
5264 ac->set_arraycopy(validated);
5265
5266 Node* n = _gvn.transform(ac);
5267 if (n == ac) {
5268 ac->connect_outputs(this);
5269 } else {
5270 assert(validated, "shouldn't transform if all arguments not validated");
5271 set_all_memory(n);
5272 }
5273
5274 return true;
5275 }
5276
5277
5278 // Helper function which determines if an arraycopy immediately follows
5279 // an allocation, with no intervening tests or other escapes for the object.
5280 AllocateArrayNode*
5281 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5282 RegionNode* slow_region) {
5283 if (stopped()) return NULL; // no fast path
5284 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5285
5286 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5287 if (alloc == NULL) return NULL;
5288
5289 Node* rawmem = memory(Compile::AliasIdxRaw);
5290 // Is the allocation's memory state untouched?
5291 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5292 // Bail out if there have been raw-memory effects since the allocation.
5293 // (Example: There might have been a call or safepoint.)
5294 return NULL;
5295 }
5296 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5297 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5298 return NULL;
5299 }
5300
5301 // There must be no unexpected observers of this allocation.
5302 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5303 Node* obs = ptr->fast_out(i);
5304 if (obs != this->map()) {
5305 return NULL;
5306 }
5307 }
5308
5309 // This arraycopy must unconditionally follow the allocation of the ptr.
5310 Node* alloc_ctl = ptr->in(0);
5311 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5312
5313 Node* ctl = control();
5314 while (ctl != alloc_ctl) {
5315 // There may be guards which feed into the slow_region.
5316 // Any other control flow means that we might not get a chance
5317 // to finish initializing the allocated object.
5318 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5319 IfNode* iff = ctl->in(0)->as_If();
5320 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5321 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5322 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5323 ctl = iff->in(0); // This test feeds the known slow_region.
5324 continue;
5325 }
5326 // One more try: Various low-level checks bottom out in
5327 // uncommon traps. If the debug-info of the trap omits
5328 // any reference to the allocation, as we've already
5329 // observed, then there can be no objection to the trap.
5330 bool found_trap = false;
5331 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5332 Node* obs = not_ctl->fast_out(j);
5333 if (obs->in(0) == not_ctl && obs->is_Call() &&
5334 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5335 found_trap = true; break;
5336 }
5337 }
5338 if (found_trap) {
5339 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5340 continue;
5341 }
5342 }
5343 return NULL;
5344 }
5345
5346 // If we get this far, we have an allocation which immediately
5347 // precedes the arraycopy, and we can take over zeroing the new object.
5348 // The arraycopy will finish the initialization, and provide
5349 // a new control state to which we will anchor the destination pointer.
5350
5351 return alloc;
5352 }
5353
5354 //-------------inline_encodeISOArray-----------------------------------
5355 // encode char[] to byte[] in ISO_8859_1
5356 bool LibraryCallKit::inline_encodeISOArray() {
5357 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5358 // no receiver since it is static method
5359 Node *src = argument(0);
5360 Node *src_offset = argument(1);
5361 Node *dst = argument(2);
5362 Node *dst_offset = argument(3);
5363 Node *length = argument(4);
5364
5365 const Type* src_type = src->Value(&_gvn);
5366 const Type* dst_type = dst->Value(&_gvn);
5367 const TypeAryPtr* top_src = src_type->isa_aryptr();
5368 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5369 if (top_src == NULL || top_src->klass() == NULL ||
5370 top_dest == NULL || top_dest->klass() == NULL) {
5371 // failed array check
5372 return false;
5373 }
5374
5375 // Figure out the size and type of the elements we will be copying.
5376 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5377 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5378 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5379 return false;
5380 }
5381
5382 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5383 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5384 // 'src_start' points to src array + scaled offset
5385 // 'dst_start' points to dst array + scaled offset
5386
5387 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5388 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5389 enc = _gvn.transform(enc);
5390 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5391 set_memory(res_mem, mtype);
5392 set_result(enc);
5393 return true;
5394 }
5395
5396 //-------------inline_multiplyToLen-----------------------------------
5397 bool LibraryCallKit::inline_multiplyToLen() {
5398 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5399
5400 address stubAddr = StubRoutines::multiplyToLen();
5401 if (stubAddr == NULL) {
5402 return false; // Intrinsic's stub is not implemented on this platform
5403 }
5404 const char* stubName = "multiplyToLen";
5405
5406 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5407
5408 // no receiver because it is a static method
5409 Node* x = argument(0);
5410 Node* xlen = argument(1);
5411 Node* y = argument(2);
5412 Node* ylen = argument(3);
5413 Node* z = argument(4);
5414
5415 const Type* x_type = x->Value(&_gvn);
5416 const Type* y_type = y->Value(&_gvn);
5417 const TypeAryPtr* top_x = x_type->isa_aryptr();
5418 const TypeAryPtr* top_y = y_type->isa_aryptr();
5419 if (top_x == NULL || top_x->klass() == NULL ||
5420 top_y == NULL || top_y->klass() == NULL) {
5421 // failed array check
5422 return false;
5423 }
5424
5425 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5426 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5427 if (x_elem != T_INT || y_elem != T_INT) {
5428 return false;
5429 }
5430
5431 // Set the original stack and the reexecute bit for the interpreter to reexecute
5432 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5433 // on the return from z array allocation in runtime.
5434 { PreserveReexecuteState preexecs(this);
5435 jvms()->set_should_reexecute(true);
5436
5437 Node* x_start = array_element_address(x, intcon(0), x_elem);
5438 Node* y_start = array_element_address(y, intcon(0), y_elem);
5439 // 'x_start' points to x array + scaled xlen
5440 // 'y_start' points to y array + scaled ylen
5441
5442 // Allocate the result array
5443 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5444 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5445 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5446
5447 IdealKit ideal(this);
5448
5449 #define __ ideal.
5450 Node* one = __ ConI(1);
5451 Node* zero = __ ConI(0);
5452 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5453 __ set(need_alloc, zero);
5454 __ set(z_alloc, z);
5455 __ if_then(z, BoolTest::eq, null()); {
5456 __ increment (need_alloc, one);
5457 } __ else_(); {
5458 // Update graphKit memory and control from IdealKit.
5459 sync_kit(ideal);
5460 Node* zlen_arg = load_array_length(z);
5461 // Update IdealKit memory and control from graphKit.
5462 __ sync_kit(this);
5463 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5464 __ increment (need_alloc, one);
5465 } __ end_if();
5466 } __ end_if();
5467
5468 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5469 // Update graphKit memory and control from IdealKit.
5470 sync_kit(ideal);
5471 Node * narr = new_array(klass_node, zlen, 1);
5472 // Update IdealKit memory and control from graphKit.
5473 __ sync_kit(this);
5474 __ set(z_alloc, narr);
5475 } __ end_if();
5476
5477 sync_kit(ideal);
5478 z = __ value(z_alloc);
5479 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5480 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5481 // Final sync IdealKit and GraphKit.
5482 final_sync(ideal);
5483 #undef __
5484
5485 Node* z_start = array_element_address(z, intcon(0), T_INT);
5486
5487 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5488 OptoRuntime::multiplyToLen_Type(),
5489 stubAddr, stubName, TypePtr::BOTTOM,
5490 x_start, xlen, y_start, ylen, z_start, zlen);
5491 } // original reexecute is set back here
5492
5493 C->set_has_split_ifs(true); // Has chance for split-if optimization
5494 set_result(z);
5495 return true;
5496 }
5497
5498 //-------------inline_squareToLen------------------------------------
5499 bool LibraryCallKit::inline_squareToLen() {
5500 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5501
5502 address stubAddr = StubRoutines::squareToLen();
5503 if (stubAddr == NULL) {
5504 return false; // Intrinsic's stub is not implemented on this platform
5505 }
5506 const char* stubName = "squareToLen";
5507
5508 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5509
5510 Node* x = argument(0);
5511 Node* len = argument(1);
5512 Node* z = argument(2);
5513 Node* zlen = argument(3);
5514
5515 const Type* x_type = x->Value(&_gvn);
5516 const Type* z_type = z->Value(&_gvn);
5517 const TypeAryPtr* top_x = x_type->isa_aryptr();
5518 const TypeAryPtr* top_z = z_type->isa_aryptr();
5519 if (top_x == NULL || top_x->klass() == NULL ||
5520 top_z == NULL || top_z->klass() == NULL) {
5521 // failed array check
5522 return false;
5523 }
5524
5525 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5526 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5527 if (x_elem != T_INT || z_elem != T_INT) {
5528 return false;
5529 }
5530
5531
5532 Node* x_start = array_element_address(x, intcon(0), x_elem);
5533 Node* z_start = array_element_address(z, intcon(0), z_elem);
5534
5535 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5536 OptoRuntime::squareToLen_Type(),
5537 stubAddr, stubName, TypePtr::BOTTOM,
5538 x_start, len, z_start, zlen);
5539
5540 set_result(z);
5541 return true;
5542 }
5543
5544 //-------------inline_mulAdd------------------------------------------
5545 bool LibraryCallKit::inline_mulAdd() {
5546 assert(UseMulAddIntrinsic, "not implemented on this platform");
5547
5548 address stubAddr = StubRoutines::mulAdd();
5549 if (stubAddr == NULL) {
5550 return false; // Intrinsic's stub is not implemented on this platform
5551 }
5552 const char* stubName = "mulAdd";
5553
5554 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5555
5556 Node* out = argument(0);
5557 Node* in = argument(1);
5558 Node* offset = argument(2);
5559 Node* len = argument(3);
5560 Node* k = argument(4);
5561
5562 const Type* out_type = out->Value(&_gvn);
5563 const Type* in_type = in->Value(&_gvn);
5564 const TypeAryPtr* top_out = out_type->isa_aryptr();
5565 const TypeAryPtr* top_in = in_type->isa_aryptr();
5566 if (top_out == NULL || top_out->klass() == NULL ||
5567 top_in == NULL || top_in->klass() == NULL) {
5568 // failed array check
5569 return false;
5570 }
5571
5572 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5573 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5574 if (out_elem != T_INT || in_elem != T_INT) {
5575 return false;
5576 }
5577
5578 Node* outlen = load_array_length(out);
5579 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5580 Node* out_start = array_element_address(out, intcon(0), out_elem);
5581 Node* in_start = array_element_address(in, intcon(0), in_elem);
5582
5583 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5584 OptoRuntime::mulAdd_Type(),
5585 stubAddr, stubName, TypePtr::BOTTOM,
5586 out_start,in_start, new_offset, len, k);
5587 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5588 set_result(result);
5589 return true;
5590 }
5591
5592 //-------------inline_montgomeryMultiply-----------------------------------
5593 bool LibraryCallKit::inline_montgomeryMultiply() {
5594 address stubAddr = StubRoutines::montgomeryMultiply();
5595 if (stubAddr == NULL) {
5596 return false; // Intrinsic's stub is not implemented on this platform
5597 }
5598
5599 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5600 const char* stubName = "montgomery_multiply";
5601
5602 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5603
5604 Node* a = argument(0);
5605 Node* b = argument(1);
5606 Node* n = argument(2);
5607 Node* len = argument(3);
5608 Node* inv = argument(4);
5609 Node* m = argument(6);
5610
5611 const Type* a_type = a->Value(&_gvn);
5612 const TypeAryPtr* top_a = a_type->isa_aryptr();
5613 const Type* b_type = b->Value(&_gvn);
5614 const TypeAryPtr* top_b = b_type->isa_aryptr();
5615 const Type* n_type = a->Value(&_gvn);
5616 const TypeAryPtr* top_n = n_type->isa_aryptr();
5617 const Type* m_type = a->Value(&_gvn);
5618 const TypeAryPtr* top_m = m_type->isa_aryptr();
5619 if (top_a == NULL || top_a->klass() == NULL ||
5620 top_b == NULL || top_b->klass() == NULL ||
5621 top_n == NULL || top_n->klass() == NULL ||
5622 top_m == NULL || top_m->klass() == NULL) {
5623 // failed array check
5624 return false;
5625 }
5626
5627 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5628 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5629 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5630 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5631 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5632 return false;
5633 }
5634
5635 // Make the call
5636 {
5637 Node* a_start = array_element_address(a, intcon(0), a_elem);
5638 Node* b_start = array_element_address(b, intcon(0), b_elem);
5639 Node* n_start = array_element_address(n, intcon(0), n_elem);
5640 Node* m_start = array_element_address(m, intcon(0), m_elem);
5641
5642 Node* call = make_runtime_call(RC_LEAF,
5643 OptoRuntime::montgomeryMultiply_Type(),
5644 stubAddr, stubName, TypePtr::BOTTOM,
5645 a_start, b_start, n_start, len, inv, top(),
5646 m_start);
5647 set_result(m);
5648 }
5649
5650 return true;
5651 }
5652
5653 bool LibraryCallKit::inline_montgomerySquare() {
5654 address stubAddr = StubRoutines::montgomerySquare();
5655 if (stubAddr == NULL) {
5656 return false; // Intrinsic's stub is not implemented on this platform
5657 }
5658
5659 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5660 const char* stubName = "montgomery_square";
5661
5662 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5663
5664 Node* a = argument(0);
5665 Node* n = argument(1);
5666 Node* len = argument(2);
5667 Node* inv = argument(3);
5668 Node* m = argument(5);
5669
5670 const Type* a_type = a->Value(&_gvn);
5671 const TypeAryPtr* top_a = a_type->isa_aryptr();
5672 const Type* n_type = a->Value(&_gvn);
5673 const TypeAryPtr* top_n = n_type->isa_aryptr();
5674 const Type* m_type = a->Value(&_gvn);
5675 const TypeAryPtr* top_m = m_type->isa_aryptr();
5676 if (top_a == NULL || top_a->klass() == NULL ||
5677 top_n == NULL || top_n->klass() == NULL ||
5678 top_m == NULL || top_m->klass() == NULL) {
5679 // failed array check
5680 return false;
5681 }
5682
5683 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5684 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5685 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5686 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5687 return false;
5688 }
5689
5690 // Make the call
5691 {
5692 Node* a_start = array_element_address(a, intcon(0), a_elem);
5693 Node* n_start = array_element_address(n, intcon(0), n_elem);
5694 Node* m_start = array_element_address(m, intcon(0), m_elem);
5695
5696 Node* call = make_runtime_call(RC_LEAF,
5697 OptoRuntime::montgomerySquare_Type(),
5698 stubAddr, stubName, TypePtr::BOTTOM,
5699 a_start, n_start, len, inv, top(),
5700 m_start);
5701 set_result(m);
5702 }
5703
5704 return true;
5705 }
5706
5707 //-------------inline_vectorizedMismatch------------------------------
5708 bool LibraryCallKit::inline_vectorizedMismatch() {
5709 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5710
5711 address stubAddr = StubRoutines::vectorizedMismatch();
5712 if (stubAddr == NULL) {
5713 return false; // Intrinsic's stub is not implemented on this platform
5714 }
5715 const char* stubName = "vectorizedMismatch";
5716 int size_l = callee()->signature()->size();
5717 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5718
5719 Node* obja = argument(0);
5720 Node* aoffset = argument(1);
5721 Node* objb = argument(3);
5722 Node* boffset = argument(4);
5723 Node* length = argument(6);
5724 Node* scale = argument(7);
5725
5726 const Type* a_type = obja->Value(&_gvn);
5727 const Type* b_type = objb->Value(&_gvn);
5728 const TypeAryPtr* top_a = a_type->isa_aryptr();
5729 const TypeAryPtr* top_b = b_type->isa_aryptr();
5730 if (top_a == NULL || top_a->klass() == NULL ||
5731 top_b == NULL || top_b->klass() == NULL) {
5732 // failed array check
5733 return false;
5734 }
5735
5736 Node* call;
5737 jvms()->set_should_reexecute(true);
5738
5739 Node* obja_adr = make_unsafe_address(obja, aoffset);
5740 Node* objb_adr = make_unsafe_address(objb, boffset);
5741
5742 call = make_runtime_call(RC_LEAF,
5743 OptoRuntime::vectorizedMismatch_Type(),
5744 stubAddr, stubName, TypePtr::BOTTOM,
5745 obja_adr, objb_adr, length, scale);
5746
5747 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5748 set_result(result);
5749 return true;
5750 }
5751
5752 /**
5753 * Calculate CRC32 for byte.
5754 * int java.util.zip.CRC32.update(int crc, int b)
5755 */
5756 bool LibraryCallKit::inline_updateCRC32() {
5757 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5758 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5759 // no receiver since it is static method
5760 Node* crc = argument(0); // type: int
5761 Node* b = argument(1); // type: int
5762
5763 /*
5764 * int c = ~ crc;
5765 * b = timesXtoThe32[(b ^ c) & 0xFF];
5766 * b = b ^ (c >>> 8);
5767 * crc = ~b;
5768 */
5769
5770 Node* M1 = intcon(-1);
5771 crc = _gvn.transform(new XorINode(crc, M1));
5772 Node* result = _gvn.transform(new XorINode(crc, b));
5773 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5774
5775 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5776 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5777 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5778 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5779
5780 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5781 result = _gvn.transform(new XorINode(crc, result));
5782 result = _gvn.transform(new XorINode(result, M1));
5783 set_result(result);
5784 return true;
5785 }
5786
5787 /**
5788 * Calculate CRC32 for byte[] array.
5789 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5790 */
5791 bool LibraryCallKit::inline_updateBytesCRC32() {
5792 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5793 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5794 // no receiver since it is static method
5795 Node* crc = argument(0); // type: int
5796 Node* src = argument(1); // type: oop
5797 Node* offset = argument(2); // type: int
5798 Node* length = argument(3); // type: int
5799
5800 const Type* src_type = src->Value(&_gvn);
5801 const TypeAryPtr* top_src = src_type->isa_aryptr();
5802 if (top_src == NULL || top_src->klass() == NULL) {
5803 // failed array check
5804 return false;
5805 }
5806
5807 // Figure out the size and type of the elements we will be copying.
5808 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5809 if (src_elem != T_BYTE) {
5810 return false;
5811 }
5812
5813 // 'src_start' points to src array + scaled offset
5814 Node* src_start = array_element_address(src, offset, src_elem);
5815
5816 // We assume that range check is done by caller.
5817 // TODO: generate range check (offset+length < src.length) in debug VM.
5818
5819 // Call the stub.
5820 address stubAddr = StubRoutines::updateBytesCRC32();
5821 const char *stubName = "updateBytesCRC32";
5822
5823 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5824 stubAddr, stubName, TypePtr::BOTTOM,
5825 crc, src_start, length);
5826 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5827 set_result(result);
5828 return true;
5829 }
5830
5831 /**
5832 * Calculate CRC32 for ByteBuffer.
5833 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5834 */
5835 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5836 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5837 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5838 // no receiver since it is static method
5839 Node* crc = argument(0); // type: int
5840 Node* src = argument(1); // type: long
5841 Node* offset = argument(3); // type: int
5842 Node* length = argument(4); // type: int
5843
5844 src = ConvL2X(src); // adjust Java long to machine word
5845 Node* base = _gvn.transform(new CastX2PNode(src));
5846 offset = ConvI2X(offset);
5847
5848 // 'src_start' points to src array + scaled offset
5849 Node* src_start = basic_plus_adr(top(), base, offset);
5850
5851 // Call the stub.
5852 address stubAddr = StubRoutines::updateBytesCRC32();
5853 const char *stubName = "updateBytesCRC32";
5854
5855 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5856 stubAddr, stubName, TypePtr::BOTTOM,
5857 crc, src_start, length);
5858 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5859 set_result(result);
5860 return true;
5861 }
5862
5863 //------------------------------get_table_from_crc32c_class-----------------------
5864 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5865 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5866 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5867
5868 return table;
5869 }
5870
5871 //------------------------------inline_updateBytesCRC32C-----------------------
5872 //
5873 // Calculate CRC32C for byte[] array.
5874 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5875 //
5876 bool LibraryCallKit::inline_updateBytesCRC32C() {
5877 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5878 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5879 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5880 // no receiver since it is a static method
5881 Node* crc = argument(0); // type: int
5882 Node* src = argument(1); // type: oop
5883 Node* offset = argument(2); // type: int
5884 Node* end = argument(3); // type: int
5885
5886 Node* length = _gvn.transform(new SubINode(end, offset));
5887
5888 const Type* src_type = src->Value(&_gvn);
5889 const TypeAryPtr* top_src = src_type->isa_aryptr();
5890 if (top_src == NULL || top_src->klass() == NULL) {
5891 // failed array check
5892 return false;
5893 }
5894
5895 // Figure out the size and type of the elements we will be copying.
5896 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5897 if (src_elem != T_BYTE) {
5898 return false;
5899 }
5900
5901 // 'src_start' points to src array + scaled offset
5902 Node* src_start = array_element_address(src, offset, src_elem);
5903
5904 // static final int[] byteTable in class CRC32C
5905 Node* table = get_table_from_crc32c_class(callee()->holder());
5906 Node* table_start = array_element_address(table, intcon(0), T_INT);
5907
5908 // We assume that range check is done by caller.
5909 // TODO: generate range check (offset+length < src.length) in debug VM.
5910
5911 // Call the stub.
5912 address stubAddr = StubRoutines::updateBytesCRC32C();
5913 const char *stubName = "updateBytesCRC32C";
5914
5915 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5916 stubAddr, stubName, TypePtr::BOTTOM,
5917 crc, src_start, length, table_start);
5918 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5919 set_result(result);
5920 return true;
5921 }
5922
5923 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5924 //
5925 // Calculate CRC32C for DirectByteBuffer.
5926 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5927 //
5928 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5929 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5930 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5931 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5932 // no receiver since it is a static method
5933 Node* crc = argument(0); // type: int
5934 Node* src = argument(1); // type: long
5935 Node* offset = argument(3); // type: int
5936 Node* end = argument(4); // type: int
5937
5938 Node* length = _gvn.transform(new SubINode(end, offset));
5939
5940 src = ConvL2X(src); // adjust Java long to machine word
5941 Node* base = _gvn.transform(new CastX2PNode(src));
5942 offset = ConvI2X(offset);
5943
5944 // 'src_start' points to src array + scaled offset
5945 Node* src_start = basic_plus_adr(top(), base, offset);
5946
5947 // static final int[] byteTable in class CRC32C
5948 Node* table = get_table_from_crc32c_class(callee()->holder());
5949 Node* table_start = array_element_address(table, intcon(0), T_INT);
5950
5951 // Call the stub.
5952 address stubAddr = StubRoutines::updateBytesCRC32C();
5953 const char *stubName = "updateBytesCRC32C";
5954
5955 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5956 stubAddr, stubName, TypePtr::BOTTOM,
5957 crc, src_start, length, table_start);
5958 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5959 set_result(result);
5960 return true;
5961 }
5962
5963 //------------------------------inline_updateBytesAdler32----------------------
5964 //
5965 // Calculate Adler32 checksum for byte[] array.
5966 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5967 //
5968 bool LibraryCallKit::inline_updateBytesAdler32() {
5969 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5970 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5971 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5972 // no receiver since it is static method
5973 Node* crc = argument(0); // type: int
5974 Node* src = argument(1); // type: oop
5975 Node* offset = argument(2); // type: int
5976 Node* length = argument(3); // type: int
5977
5978 const Type* src_type = src->Value(&_gvn);
5979 const TypeAryPtr* top_src = src_type->isa_aryptr();
5980 if (top_src == NULL || top_src->klass() == NULL) {
5981 // failed array check
5982 return false;
5983 }
5984
5985 // Figure out the size and type of the elements we will be copying.
5986 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5987 if (src_elem != T_BYTE) {
5988 return false;
5989 }
5990
5991 // 'src_start' points to src array + scaled offset
5992 Node* src_start = array_element_address(src, offset, src_elem);
5993
5994 // We assume that range check is done by caller.
5995 // TODO: generate range check (offset+length < src.length) in debug VM.
5996
5997 // Call the stub.
5998 address stubAddr = StubRoutines::updateBytesAdler32();
5999 const char *stubName = "updateBytesAdler32";
6000
6001 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6002 stubAddr, stubName, TypePtr::BOTTOM,
6003 crc, src_start, length);
6004 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6005 set_result(result);
6006 return true;
6007 }
6008
6009 //------------------------------inline_updateByteBufferAdler32---------------
6010 //
6011 // Calculate Adler32 checksum for DirectByteBuffer.
6012 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6013 //
6014 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6015 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6016 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6017 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6018 // no receiver since it is static method
6019 Node* crc = argument(0); // type: int
6020 Node* src = argument(1); // type: long
6021 Node* offset = argument(3); // type: int
6022 Node* length = argument(4); // type: int
6023
6024 src = ConvL2X(src); // adjust Java long to machine word
6025 Node* base = _gvn.transform(new CastX2PNode(src));
6026 offset = ConvI2X(offset);
6027
6028 // 'src_start' points to src array + scaled offset
6029 Node* src_start = basic_plus_adr(top(), base, offset);
6030
6031 // Call the stub.
6032 address stubAddr = StubRoutines::updateBytesAdler32();
6033 const char *stubName = "updateBytesAdler32";
6034
6035 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6036 stubAddr, stubName, TypePtr::BOTTOM,
6037 crc, src_start, length);
6038
6039 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6040 set_result(result);
6041 return true;
6042 }
6043
6044 //----------------------------inline_reference_get----------------------------
6045 // public T java.lang.ref.Reference.get();
6046 bool LibraryCallKit::inline_reference_get() {
6047 const int referent_offset = java_lang_ref_Reference::referent_offset;
6048 guarantee(referent_offset > 0, "should have already been set");
6049
6050 // Get the argument:
6051 Node* reference_obj = null_check_receiver();
6052 if (stopped()) return true;
6053
6054 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6055
6056 ciInstanceKlass* klass = env()->Object_klass();
6057 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6058
6059 Node* no_ctrl = NULL;
6060 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6061
6062 // Use the pre-barrier to record the value in the referent field
6063 pre_barrier(false /* do_load */,
6064 control(),
6065 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6066 result /* pre_val */,
6067 T_OBJECT);
6068
6069 // Add memory barrier to prevent commoning reads from this field
6070 // across safepoint since GC can change its value.
6071 insert_mem_bar(Op_MemBarCPUOrder);
6072
6073 set_result(result);
6074 return true;
6075 }
6076
6077
6078 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6079 bool is_exact=true, bool is_static=false,
6080 ciInstanceKlass * fromKls=NULL) {
6081 if (fromKls == NULL) {
6082 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6083 assert(tinst != NULL, "obj is null");
6084 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6085 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6086 fromKls = tinst->klass()->as_instance_klass();
6087 } else {
6088 assert(is_static, "only for static field access");
6089 }
6090 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6091 ciSymbol::make(fieldTypeString),
6092 is_static);
6093
6094 assert (field != NULL, "undefined field");
6095 if (field == NULL) return (Node *) NULL;
6096
6097 if (is_static) {
6098 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6099 fromObj = makecon(tip);
6100 }
6101
6102 // Next code copied from Parse::do_get_xxx():
6103
6104 // Compute address and memory type.
6105 int offset = field->offset_in_bytes();
6106 bool is_vol = field->is_volatile();
6107 ciType* field_klass = field->type();
6108 assert(field_klass->is_loaded(), "should be loaded");
6109 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6110 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6111 BasicType bt = field->layout_type();
6112
6113 // Build the resultant type of the load
6114 const Type *type;
6115 if (bt == T_OBJECT) {
6116 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6117 } else {
6118 type = Type::get_const_basic_type(bt);
6119 }
6120
6121 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6122 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6123 }
6124 // Build the load.
6125 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6126 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6127 // If reference is volatile, prevent following memory ops from
6128 // floating up past the volatile read. Also prevents commoning
6129 // another volatile read.
6130 if (is_vol) {
6131 // Memory barrier includes bogus read of value to force load BEFORE membar
6132 insert_mem_bar(Op_MemBarAcquire, loadedField);
6133 }
6134 return loadedField;
6135 }
6136
6137 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6138 bool is_exact = true, bool is_static = false,
6139 ciInstanceKlass * fromKls = NULL) {
6140 if (fromKls == NULL) {
6141 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6142 assert(tinst != NULL, "obj is null");
6143 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6144 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6145 fromKls = tinst->klass()->as_instance_klass();
6146 }
6147 else {
6148 assert(is_static, "only for static field access");
6149 }
6150 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6151 ciSymbol::make(fieldTypeString),
6152 is_static);
6153
6154 assert(field != NULL, "undefined field");
6155 assert(!field->is_volatile(), "not defined for volatile fields");
6156
6157 if (is_static) {
6158 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6159 fromObj = makecon(tip);
6160 }
6161
6162 // Next code copied from Parse::do_get_xxx():
6163
6164 // Compute address and memory type.
6165 int offset = field->offset_in_bytes();
6166 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6167
6168 return adr;
6169 }
6170
6171 //------------------------------inline_aescrypt_Block-----------------------
6172 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6173 address stubAddr = NULL;
6174 const char *stubName;
6175 assert(UseAES, "need AES instruction support");
6176
6177 switch(id) {
6178 case vmIntrinsics::_aescrypt_encryptBlock:
6179 stubAddr = StubRoutines::aescrypt_encryptBlock();
6180 stubName = "aescrypt_encryptBlock";
6181 break;
6182 case vmIntrinsics::_aescrypt_decryptBlock:
6183 stubAddr = StubRoutines::aescrypt_decryptBlock();
6184 stubName = "aescrypt_decryptBlock";
6185 break;
6186 }
6187 if (stubAddr == NULL) return false;
6188
6189 Node* aescrypt_object = argument(0);
6190 Node* src = argument(1);
6191 Node* src_offset = argument(2);
6192 Node* dest = argument(3);
6193 Node* dest_offset = argument(4);
6194
6195 // (1) src and dest are arrays.
6196 const Type* src_type = src->Value(&_gvn);
6197 const Type* dest_type = dest->Value(&_gvn);
6198 const TypeAryPtr* top_src = src_type->isa_aryptr();
6199 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6200 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6201
6202 // for the quick and dirty code we will skip all the checks.
6203 // we are just trying to get the call to be generated.
6204 Node* src_start = src;
6205 Node* dest_start = dest;
6206 if (src_offset != NULL || dest_offset != NULL) {
6207 assert(src_offset != NULL && dest_offset != NULL, "");
6208 src_start = array_element_address(src, src_offset, T_BYTE);
6209 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6210 }
6211
6212 // now need to get the start of its expanded key array
6213 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6214 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6215 if (k_start == NULL) return false;
6216
6217 if (Matcher::pass_original_key_for_aes()) {
6218 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6219 // compatibility issues between Java key expansion and SPARC crypto instructions
6220 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6221 if (original_k_start == NULL) return false;
6222
6223 // Call the stub.
6224 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6225 stubAddr, stubName, TypePtr::BOTTOM,
6226 src_start, dest_start, k_start, original_k_start);
6227 } else {
6228 // Call the stub.
6229 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6230 stubAddr, stubName, TypePtr::BOTTOM,
6231 src_start, dest_start, k_start);
6232 }
6233
6234 return true;
6235 }
6236
6237 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6238 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6239 address stubAddr = NULL;
6240 const char *stubName = NULL;
6241
6242 assert(UseAES, "need AES instruction support");
6243
6244 switch(id) {
6245 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6246 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6247 stubName = "cipherBlockChaining_encryptAESCrypt";
6248 break;
6249 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6250 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6251 stubName = "cipherBlockChaining_decryptAESCrypt";
6252 break;
6253 }
6254 if (stubAddr == NULL) return false;
6255
6256 Node* cipherBlockChaining_object = argument(0);
6257 Node* src = argument(1);
6258 Node* src_offset = argument(2);
6259 Node* len = argument(3);
6260 Node* dest = argument(4);
6261 Node* dest_offset = argument(5);
6262
6263 // (1) src and dest are arrays.
6264 const Type* src_type = src->Value(&_gvn);
6265 const Type* dest_type = dest->Value(&_gvn);
6266 const TypeAryPtr* top_src = src_type->isa_aryptr();
6267 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6268 assert (top_src != NULL && top_src->klass() != NULL
6269 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6270
6271 // checks are the responsibility of the caller
6272 Node* src_start = src;
6273 Node* dest_start = dest;
6274 if (src_offset != NULL || dest_offset != NULL) {
6275 assert(src_offset != NULL && dest_offset != NULL, "");
6276 src_start = array_element_address(src, src_offset, T_BYTE);
6277 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6278 }
6279
6280 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6281 // (because of the predicated logic executed earlier).
6282 // so we cast it here safely.
6283 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6284
6285 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6286 if (embeddedCipherObj == NULL) return false;
6287
6288 // cast it to what we know it will be at runtime
6289 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6290 assert(tinst != NULL, "CBC obj is null");
6291 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6292 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6293 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6294
6295 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6296 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6297 const TypeOopPtr* xtype = aklass->as_instance_type();
6298 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6299 aescrypt_object = _gvn.transform(aescrypt_object);
6300
6301 // we need to get the start of the aescrypt_object's expanded key array
6302 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6303 if (k_start == NULL) return false;
6304
6305 // similarly, get the start address of the r vector
6306 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6307 if (objRvec == NULL) return false;
6308 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6309
6310 Node* cbcCrypt;
6311 if (Matcher::pass_original_key_for_aes()) {
6312 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6313 // compatibility issues between Java key expansion and SPARC crypto instructions
6314 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6315 if (original_k_start == NULL) return false;
6316
6317 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6318 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6319 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6320 stubAddr, stubName, TypePtr::BOTTOM,
6321 src_start, dest_start, k_start, r_start, len, original_k_start);
6322 } else {
6323 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6324 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6325 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6326 stubAddr, stubName, TypePtr::BOTTOM,
6327 src_start, dest_start, k_start, r_start, len);
6328 }
6329
6330 // return cipher length (int)
6331 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6332 set_result(retvalue);
6333 return true;
6334 }
6335
6336 //------------------------------inline_counterMode_AESCrypt-----------------------
6337 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6338 assert(UseAES, "need AES instruction support");
6339 if (!UseAESCTRIntrinsics) return false;
6340
6341 address stubAddr = NULL;
6342 const char *stubName = NULL;
6343 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6344 stubAddr = StubRoutines::counterMode_AESCrypt();
6345 stubName = "counterMode_AESCrypt";
6346 }
6347 if (stubAddr == NULL) return false;
6348
6349 Node* counterMode_object = argument(0);
6350 Node* src = argument(1);
6351 Node* src_offset = argument(2);
6352 Node* len = argument(3);
6353 Node* dest = argument(4);
6354 Node* dest_offset = argument(5);
6355
6356 // (1) src and dest are arrays.
6357 const Type* src_type = src->Value(&_gvn);
6358 const Type* dest_type = dest->Value(&_gvn);
6359 const TypeAryPtr* top_src = src_type->isa_aryptr();
6360 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6361 assert(top_src != NULL && top_src->klass() != NULL &&
6362 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6363
6364 // checks are the responsibility of the caller
6365 Node* src_start = src;
6366 Node* dest_start = dest;
6367 if (src_offset != NULL || dest_offset != NULL) {
6368 assert(src_offset != NULL && dest_offset != NULL, "");
6369 src_start = array_element_address(src, src_offset, T_BYTE);
6370 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6371 }
6372
6373 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6374 // (because of the predicated logic executed earlier).
6375 // so we cast it here safely.
6376 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6377 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6378 if (embeddedCipherObj == NULL) return false;
6379 // cast it to what we know it will be at runtime
6380 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6381 assert(tinst != NULL, "CTR obj is null");
6382 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6383 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6384 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6385 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6386 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6387 const TypeOopPtr* xtype = aklass->as_instance_type();
6388 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6389 aescrypt_object = _gvn.transform(aescrypt_object);
6390 // we need to get the start of the aescrypt_object's expanded key array
6391 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6392 if (k_start == NULL) return false;
6393 // similarly, get the start address of the r vector
6394 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6395 if (obj_counter == NULL) return false;
6396 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6397
6398 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6399 if (saved_encCounter == NULL) return false;
6400 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6401 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6402
6403 Node* ctrCrypt;
6404 if (Matcher::pass_original_key_for_aes()) {
6405 // no SPARC version for AES/CTR intrinsics now.
6406 return false;
6407 }
6408 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6409 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6410 OptoRuntime::counterMode_aescrypt_Type(),
6411 stubAddr, stubName, TypePtr::BOTTOM,
6412 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6413
6414 // return cipher length (int)
6415 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6416 set_result(retvalue);
6417 return true;
6418 }
6419
6420 //------------------------------get_key_start_from_aescrypt_object-----------------------
6421 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6422 #if defined(PPC64) || defined(S390)
6423 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6424 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6425 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6426 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6427 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6428 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6429 if (objSessionK == NULL) {
6430 return (Node *) NULL;
6431 }
6432 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6433 #else
6434 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6435 #endif // PPC64
6436 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6437 if (objAESCryptKey == NULL) return (Node *) NULL;
6438
6439 // now have the array, need to get the start address of the K array
6440 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6441 return k_start;
6442 }
6443
6444 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6445 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6446 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6447 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6448 if (objAESCryptKey == NULL) return (Node *) NULL;
6449
6450 // now have the array, need to get the start address of the lastKey array
6451 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6452 return original_k_start;
6453 }
6454
6455 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6456 // Return node representing slow path of predicate check.
6457 // the pseudo code we want to emulate with this predicate is:
6458 // for encryption:
6459 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6460 // for decryption:
6461 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6462 // note cipher==plain is more conservative than the original java code but that's OK
6463 //
6464 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6465 // The receiver was checked for NULL already.
6466 Node* objCBC = argument(0);
6467
6468 // Load embeddedCipher field of CipherBlockChaining object.
6469 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6470
6471 // get AESCrypt klass for instanceOf check
6472 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6473 // will have same classloader as CipherBlockChaining object
6474 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6475 assert(tinst != NULL, "CBCobj is null");
6476 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6477
6478 // we want to do an instanceof comparison against the AESCrypt class
6479 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6480 if (!klass_AESCrypt->is_loaded()) {
6481 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6482 Node* ctrl = control();
6483 set_control(top()); // no regular fast path
6484 return ctrl;
6485 }
6486 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6487
6488 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6489 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6490 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6491
6492 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6493
6494 // for encryption, we are done
6495 if (!decrypting)
6496 return instof_false; // even if it is NULL
6497
6498 // for decryption, we need to add a further check to avoid
6499 // taking the intrinsic path when cipher and plain are the same
6500 // see the original java code for why.
6501 RegionNode* region = new RegionNode(3);
6502 region->init_req(1, instof_false);
6503 Node* src = argument(1);
6504 Node* dest = argument(4);
6505 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6506 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6507 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6508 region->init_req(2, src_dest_conjoint);
6509
6510 record_for_igvn(region);
6511 return _gvn.transform(region);
6512 }
6513
6514 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6515 // Return node representing slow path of predicate check.
6516 // the pseudo code we want to emulate with this predicate is:
6517 // for encryption:
6518 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6519 // for decryption:
6520 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6521 // note cipher==plain is more conservative than the original java code but that's OK
6522 //
6523
6524 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6525 // The receiver was checked for NULL already.
6526 Node* objCTR = argument(0);
6527
6528 // Load embeddedCipher field of CipherBlockChaining object.
6529 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6530
6531 // get AESCrypt klass for instanceOf check
6532 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6533 // will have same classloader as CipherBlockChaining object
6534 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6535 assert(tinst != NULL, "CTRobj is null");
6536 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6537
6538 // we want to do an instanceof comparison against the AESCrypt class
6539 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6540 if (!klass_AESCrypt->is_loaded()) {
6541 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6542 Node* ctrl = control();
6543 set_control(top()); // no regular fast path
6544 return ctrl;
6545 }
6546
6547 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6548 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6549 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6550 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6551 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6552
6553 return instof_false; // even if it is NULL
6554 }
6555
6556 //------------------------------inline_ghash_processBlocks
6557 bool LibraryCallKit::inline_ghash_processBlocks() {
6558 address stubAddr;
6559 const char *stubName;
6560 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6561
6562 stubAddr = StubRoutines::ghash_processBlocks();
6563 stubName = "ghash_processBlocks";
6564
6565 Node* data = argument(0);
6566 Node* offset = argument(1);
6567 Node* len = argument(2);
6568 Node* state = argument(3);
6569 Node* subkeyH = argument(4);
6570
6571 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6572 assert(state_start, "state is NULL");
6573 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6574 assert(subkeyH_start, "subkeyH is NULL");
6575 Node* data_start = array_element_address(data, offset, T_BYTE);
6576 assert(data_start, "data is NULL");
6577
6578 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6579 OptoRuntime::ghash_processBlocks_Type(),
6580 stubAddr, stubName, TypePtr::BOTTOM,
6581 state_start, subkeyH_start, data_start, len);
6582 return true;
6583 }
6584
6585 //------------------------------inline_sha_implCompress-----------------------
6586 //
6587 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6588 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6589 //
6590 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6591 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6592 //
6593 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6594 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6595 //
6596 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6597 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6598
6599 Node* sha_obj = argument(0);
6600 Node* src = argument(1); // type oop
6601 Node* ofs = argument(2); // type int
6602
6603 const Type* src_type = src->Value(&_gvn);
6604 const TypeAryPtr* top_src = src_type->isa_aryptr();
6605 if (top_src == NULL || top_src->klass() == NULL) {
6606 // failed array check
6607 return false;
6608 }
6609 // Figure out the size and type of the elements we will be copying.
6610 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6611 if (src_elem != T_BYTE) {
6612 return false;
6613 }
6614 // 'src_start' points to src array + offset
6615 Node* src_start = array_element_address(src, ofs, src_elem);
6616 Node* state = NULL;
6617 address stubAddr;
6618 const char *stubName;
6619
6620 switch(id) {
6621 case vmIntrinsics::_sha_implCompress:
6622 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6623 state = get_state_from_sha_object(sha_obj);
6624 stubAddr = StubRoutines::sha1_implCompress();
6625 stubName = "sha1_implCompress";
6626 break;
6627 case vmIntrinsics::_sha2_implCompress:
6628 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6629 state = get_state_from_sha_object(sha_obj);
6630 stubAddr = StubRoutines::sha256_implCompress();
6631 stubName = "sha256_implCompress";
6632 break;
6633 case vmIntrinsics::_sha5_implCompress:
6634 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6635 state = get_state_from_sha5_object(sha_obj);
6636 stubAddr = StubRoutines::sha512_implCompress();
6637 stubName = "sha512_implCompress";
6638 break;
6639 default:
6640 fatal_unexpected_iid(id);
6641 return false;
6642 }
6643 if (state == NULL) return false;
6644
6645 // Call the stub.
6646 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6647 stubAddr, stubName, TypePtr::BOTTOM,
6648 src_start, state);
6649
6650 return true;
6651 }
6652
6653 //------------------------------inline_digestBase_implCompressMB-----------------------
6654 //
6655 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6656 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6657 //
6658 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6659 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6660 "need SHA1/SHA256/SHA512 instruction support");
6661 assert((uint)predicate < 3, "sanity");
6662 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6663
6664 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6665 Node* src = argument(1); // byte[] array
6666 Node* ofs = argument(2); // type int
6667 Node* limit = argument(3); // type int
6668
6669 const Type* src_type = src->Value(&_gvn);
6670 const TypeAryPtr* top_src = src_type->isa_aryptr();
6671 if (top_src == NULL || top_src->klass() == NULL) {
6672 // failed array check
6673 return false;
6674 }
6675 // Figure out the size and type of the elements we will be copying.
6676 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6677 if (src_elem != T_BYTE) {
6678 return false;
6679 }
6680 // 'src_start' points to src array + offset
6681 Node* src_start = array_element_address(src, ofs, src_elem);
6682
6683 const char* klass_SHA_name = NULL;
6684 const char* stub_name = NULL;
6685 address stub_addr = NULL;
6686 bool long_state = false;
6687
6688 switch (predicate) {
6689 case 0:
6690 if (UseSHA1Intrinsics) {
6691 klass_SHA_name = "sun/security/provider/SHA";
6692 stub_name = "sha1_implCompressMB";
6693 stub_addr = StubRoutines::sha1_implCompressMB();
6694 }
6695 break;
6696 case 1:
6697 if (UseSHA256Intrinsics) {
6698 klass_SHA_name = "sun/security/provider/SHA2";
6699 stub_name = "sha256_implCompressMB";
6700 stub_addr = StubRoutines::sha256_implCompressMB();
6701 }
6702 break;
6703 case 2:
6704 if (UseSHA512Intrinsics) {
6705 klass_SHA_name = "sun/security/provider/SHA5";
6706 stub_name = "sha512_implCompressMB";
6707 stub_addr = StubRoutines::sha512_implCompressMB();
6708 long_state = true;
6709 }
6710 break;
6711 default:
6712 fatal("unknown SHA intrinsic predicate: %d", predicate);
6713 }
6714 if (klass_SHA_name != NULL) {
6715 // get DigestBase klass to lookup for SHA klass
6716 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6717 assert(tinst != NULL, "digestBase_obj is not instance???");
6718 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6719
6720 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6721 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6722 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6723 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6724 }
6725 return false;
6726 }
6727 //------------------------------inline_sha_implCompressMB-----------------------
6728 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6729 bool long_state, address stubAddr, const char *stubName,
6730 Node* src_start, Node* ofs, Node* limit) {
6731 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6732 const TypeOopPtr* xtype = aklass->as_instance_type();
6733 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6734 sha_obj = _gvn.transform(sha_obj);
6735
6736 Node* state;
6737 if (long_state) {
6738 state = get_state_from_sha5_object(sha_obj);
6739 } else {
6740 state = get_state_from_sha_object(sha_obj);
6741 }
6742 if (state == NULL) return false;
6743
6744 // Call the stub.
6745 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6746 OptoRuntime::digestBase_implCompressMB_Type(),
6747 stubAddr, stubName, TypePtr::BOTTOM,
6748 src_start, state, ofs, limit);
6749 // return ofs (int)
6750 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6751 set_result(result);
6752
6753 return true;
6754 }
6755
6756 //------------------------------get_state_from_sha_object-----------------------
6757 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6758 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6759 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6760 if (sha_state == NULL) return (Node *) NULL;
6761
6762 // now have the array, need to get the start address of the state array
6763 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6764 return state;
6765 }
6766
6767 //------------------------------get_state_from_sha5_object-----------------------
6768 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6769 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6770 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6771 if (sha_state == NULL) return (Node *) NULL;
6772
6773 // now have the array, need to get the start address of the state array
6774 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6775 return state;
6776 }
6777
6778 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6779 // Return node representing slow path of predicate check.
6780 // the pseudo code we want to emulate with this predicate is:
6781 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6782 //
6783 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6784 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6785 "need SHA1/SHA256/SHA512 instruction support");
6786 assert((uint)predicate < 3, "sanity");
6787
6788 // The receiver was checked for NULL already.
6789 Node* digestBaseObj = argument(0);
6790
6791 // get DigestBase klass for instanceOf check
6792 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6793 assert(tinst != NULL, "digestBaseObj is null");
6794 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6795
6796 const char* klass_SHA_name = NULL;
6797 switch (predicate) {
6798 case 0:
6799 if (UseSHA1Intrinsics) {
6800 // we want to do an instanceof comparison against the SHA class
6801 klass_SHA_name = "sun/security/provider/SHA";
6802 }
6803 break;
6804 case 1:
6805 if (UseSHA256Intrinsics) {
6806 // we want to do an instanceof comparison against the SHA2 class
6807 klass_SHA_name = "sun/security/provider/SHA2";
6808 }
6809 break;
6810 case 2:
6811 if (UseSHA512Intrinsics) {
6812 // we want to do an instanceof comparison against the SHA5 class
6813 klass_SHA_name = "sun/security/provider/SHA5";
6814 }
6815 break;
6816 default:
6817 fatal("unknown SHA intrinsic predicate: %d", predicate);
6818 }
6819
6820 ciKlass* klass_SHA = NULL;
6821 if (klass_SHA_name != NULL) {
6822 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6823 }
6824 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6825 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6826 Node* ctrl = control();
6827 set_control(top()); // no intrinsic path
6828 return ctrl;
6829 }
6830 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6831
6832 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6833 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6834 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6835 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6836
6837 return instof_false; // even if it is NULL
6838 }
6839
6840 //-------------inline_fma-----------------------------------
6841 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6842 Node *a = NULL;
6843 Node *b = NULL;
6844 Node *c = NULL;
6845 Node* result = NULL;
6846 switch (id) {
6847 case vmIntrinsics::_fmaD:
6848 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6849 // no receiver since it is static method
6850 a = round_double_node(argument(0));
6851 b = round_double_node(argument(2));
6852 c = round_double_node(argument(4));
6853 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6854 break;
6855 case vmIntrinsics::_fmaF:
6856 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6857 a = argument(0);
6858 b = argument(1);
6859 c = argument(2);
6860 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6861 break;
6862 default:
6863 fatal_unexpected_iid(id); break;
6864 }
6865 set_result(result);
6866 return true;
6867 }
6868
6869 bool LibraryCallKit::inline_profileBoolean() {
6870 Node* counts = argument(1);
6871 const TypeAryPtr* ary = NULL;
6872 ciArray* aobj = NULL;
6873 if (counts->is_Con()
6874 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6875 && (aobj = ary->const_oop()->as_array()) != NULL
6876 && (aobj->length() == 2)) {
6877 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6878 jint false_cnt = aobj->element_value(0).as_int();
6879 jint true_cnt = aobj->element_value(1).as_int();
6880
6881 if (C->log() != NULL) {
6882 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6883 false_cnt, true_cnt);
6884 }
6885
6886 if (false_cnt + true_cnt == 0) {
6887 // According to profile, never executed.
6888 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6889 Deoptimization::Action_reinterpret);
6890 return true;
6891 }
6892
6893 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6894 // is a number of each value occurrences.
6895 Node* result = argument(0);
6896 if (false_cnt == 0 || true_cnt == 0) {
6897 // According to profile, one value has been never seen.
6898 int expected_val = (false_cnt == 0) ? 1 : 0;
6899
6900 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6901 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6902
6903 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6904 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6905 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6906
6907 { // Slow path: uncommon trap for never seen value and then reexecute
6908 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6909 // the value has been seen at least once.
6910 PreserveJVMState pjvms(this);
6911 PreserveReexecuteState preexecs(this);
6912 jvms()->set_should_reexecute(true);
6913
6914 set_control(slow_path);
6915 set_i_o(i_o());
6916
6917 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6918 Deoptimization::Action_reinterpret);
6919 }
6920 // The guard for never seen value enables sharpening of the result and
6921 // returning a constant. It allows to eliminate branches on the same value
6922 // later on.
6923 set_control(fast_path);
6924 result = intcon(expected_val);
6925 }
6926 // Stop profiling.
6927 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6928 // By replacing method body with profile data (represented as ProfileBooleanNode
6929 // on IR level) we effectively disable profiling.
6930 // It enables full speed execution once optimized code is generated.
6931 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6932 C->record_for_igvn(profile);
6933 set_result(profile);
6934 return true;
6935 } else {
6936 // Continue profiling.
6937 // Profile data isn't available at the moment. So, execute method's bytecode version.
6938 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6939 // is compiled and counters aren't available since corresponding MethodHandle
6940 // isn't a compile-time constant.
6941 return false;
6942 }
6943 }
6944
6945 bool LibraryCallKit::inline_isCompileConstant() {
6946 Node* n = argument(0);
6947 set_result(n->is_Con() ? intcon(1) : intcon(0));
6948 return true;
6949 }