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/subnode.hpp"
51 #include "prims/nativeLookup.hpp"
52 #include "prims/unsafe.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #ifdef TRACE_HAVE_INTRINSICS
55 #include "trace/traceMacros.hpp"
56 #endif
57
58 class LibraryIntrinsic : public InlineCallGenerator {
59 // Extend the set of intrinsics known to the runtime:
60 public:
61 private:
62 bool _is_virtual;
63 bool _does_virtual_dispatch;
64 int8_t _predicates_count; // Intrinsic is predicated by several conditions
65 int8_t _last_predicate; // Last generated predicate
66 vmIntrinsics::ID _intrinsic_id;
67
68 public:
69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
70 : InlineCallGenerator(m),
71 _is_virtual(is_virtual),
72 _does_virtual_dispatch(does_virtual_dispatch),
73 _predicates_count((int8_t)predicates_count),
74 _last_predicate((int8_t)-1),
75 _intrinsic_id(id)
76 {
77 }
78 virtual bool is_intrinsic() const { return true; }
79 virtual bool is_virtual() const { return _is_virtual; }
80 virtual bool is_predicated() const { return _predicates_count > 0; }
81 virtual int predicates_count() const { return _predicates_count; }
82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
83 virtual JVMState* generate(JVMState* jvms);
84 virtual Node* generate_predicate(JVMState* jvms, int predicate);
85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
86 };
87
88
89 // Local helper class for LibraryIntrinsic:
90 class LibraryCallKit : public GraphKit {
91 private:
92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
93 Node* _result; // the result node, if any
94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
95
96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
97
98 public:
99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
100 : GraphKit(jvms),
101 _intrinsic(intrinsic),
102 _result(NULL)
103 {
104 // Check if this is a root compile. In that case we don't have a caller.
105 if (!jvms->has_method()) {
106 _reexecute_sp = sp();
107 } else {
108 // Find out how many arguments the interpreter needs when deoptimizing
109 // and save the stack pointer value so it can used by uncommon_trap.
110 // We find the argument count by looking at the declared signature.
111 bool ignored_will_link;
112 ciSignature* declared_signature = NULL;
113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
116 }
117 }
118
119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
120
121 ciMethod* caller() const { return jvms()->method(); }
122 int bci() const { return jvms()->bci(); }
123 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
125 ciMethod* callee() const { return _intrinsic->method(); }
126
127 bool try_to_inline(int predicate);
128 Node* try_to_predicate(int predicate);
129
130 void push_result() {
131 // Push the result onto the stack.
132 if (!stopped() && result() != NULL) {
133 BasicType bt = result()->bottom_type()->basic_type();
134 push_node(bt, result());
135 }
136 }
137
138 private:
139 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
141 }
142
143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
144 void set_result(RegionNode* region, PhiNode* value);
145 Node* result() { return _result; }
146
147 virtual int reexecute_sp() { return _reexecute_sp; }
148
149 // Helper functions to inline natives
150 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
151 Node* generate_slow_guard(Node* test, RegionNode* region);
152 Node* generate_fair_guard(Node* test, RegionNode* region);
153 Node* generate_negative_guard(Node* index, RegionNode* region,
154 // resulting CastII of index:
155 Node* *pos_index = NULL);
156 Node* generate_limit_guard(Node* offset, Node* subseq_length,
157 Node* array_length,
158 RegionNode* region);
159 void generate_string_range_check(Node* array, Node* offset,
160 Node* length, bool char_count);
161 Node* generate_current_thread(Node* &tls_output);
162 Node* load_mirror_from_klass(Node* klass);
163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path,
165 int offset);
166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
174 RegionNode* region, int null_path) {
175 int offset = java_lang_Class::array_klass_offset_in_bytes();
176 return load_klass_from_mirror_common(mirror, never_see_null,
177 region, null_path,
178 offset);
179 }
180 Node* generate_access_flags_guard(Node* kls,
181 int modifier_mask, int modifier_bits,
182 RegionNode* region);
183 Node* generate_interface_guard(Node* kls, RegionNode* region);
184 Node* generate_array_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, false, false);
186 }
187 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, false, true);
189 }
190 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, true, false);
192 }
193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, true, true);
195 }
196 Node* generate_array_guard_common(Node* kls, RegionNode* region,
197 bool obj_array, bool not_array);
198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
200 bool is_virtual = false, bool is_static = false);
201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
202 return generate_method_call(method_id, false, true);
203 }
204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
205 return generate_method_call(method_id, true, false);
206 }
207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209
210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_indexOfChar();
217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_toBytesU();
219 bool inline_string_getCharsU();
220 bool inline_string_copy(bool compress);
221 bool inline_string_char_access(bool is_store);
222 Node* round_double_node(Node* n);
223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_math_native(vmIntrinsics::ID id);
225 bool inline_math(vmIntrinsics::ID id);
226 template <typename OverflowOp>
227 bool inline_math_overflow(Node* arg1, Node* arg2);
228 void inline_math_mathExact(Node* math, Node* test);
229 bool inline_math_addExactI(bool is_increment);
230 bool inline_math_addExactL(bool is_increment);
231 bool inline_math_multiplyExactI();
232 bool inline_math_multiplyExactL();
233 bool inline_math_negateExactI();
234 bool inline_math_negateExactL();
235 bool inline_math_subtractExactI(bool is_decrement);
236 bool inline_math_subtractExactL(bool is_decrement);
237 bool inline_min_max(vmIntrinsics::ID id);
238 bool inline_notify(vmIntrinsics::ID id);
239 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
240 // This returns Type::AnyPtr, RawPtr, or OopPtr.
241 int classify_unsafe_addr(Node* &base, Node* &offset);
242 Node* make_unsafe_address(Node* base, Node* offset);
243 // Helper for inline_unsafe_access.
244 // Generates the guards that check whether the result of
245 // Unsafe.getObject should be recorded in an SATB log buffer.
246 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
247
248 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
249 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
250 static bool klass_needs_init_guard(Node* kls);
251 bool inline_unsafe_allocate();
252 bool inline_unsafe_newArray(bool uninitialized);
253 bool inline_unsafe_copyMemory();
254 bool inline_native_currentThread();
255
256 bool inline_native_time_funcs(address method, const char* funcName);
257 #ifdef TRACE_HAVE_INTRINSICS
258 bool inline_native_classID();
259 bool inline_native_getBufferWriter();
260 #endif
261 bool inline_native_isInterrupted();
262 bool inline_native_Class_query(vmIntrinsics::ID id);
263 bool inline_native_subtype_check();
264 bool inline_native_getLength();
265 bool inline_array_copyOf(bool is_copyOfRange);
266 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
267 bool inline_preconditions_checkIndex();
268 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
269 bool inline_native_clone(bool is_virtual);
270 bool inline_native_Reflection_getCallerClass();
271 // Helper function for inlining native object hash method
272 bool inline_native_hashcode(bool is_virtual, bool is_static);
273 bool inline_native_getClass();
274
275 // Helper functions for inlining arraycopy
276 bool inline_arraycopy();
277 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
278 RegionNode* slow_region);
279 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
280 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
281 uint new_idx);
282
283 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
284 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
285 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
286 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
287 bool inline_unsafe_fence(vmIntrinsics::ID id);
288 bool inline_onspinwait();
289 bool inline_fp_conversions(vmIntrinsics::ID id);
290 bool inline_number_methods(vmIntrinsics::ID id);
291 bool inline_reference_get();
292 bool inline_Class_cast();
293 bool inline_aescrypt_Block(vmIntrinsics::ID id);
294 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
295 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
296 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
297 Node* inline_counterMode_AESCrypt_predicate();
298 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
299 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
300 bool inline_ghash_processBlocks();
301 bool inline_sha_implCompress(vmIntrinsics::ID id);
302 bool inline_digestBase_implCompressMB(int predicate);
303 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
304 bool long_state, address stubAddr, const char *stubName,
305 Node* src_start, Node* ofs, Node* limit);
306 Node* get_state_from_sha_object(Node *sha_object);
307 Node* get_state_from_sha5_object(Node *sha_object);
308 Node* inline_digestBase_implCompressMB_predicate(int predicate);
309 bool inline_encodeISOArray();
310 bool inline_updateCRC32();
311 bool inline_updateBytesCRC32();
312 bool inline_updateByteBufferCRC32();
313 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
314 bool inline_updateBytesCRC32C();
315 bool inline_updateDirectByteBufferCRC32C();
316 bool inline_updateBytesAdler32();
317 bool inline_updateByteBufferAdler32();
318 bool inline_multiplyToLen();
319 bool inline_hasNegatives();
320 bool inline_squareToLen();
321 bool inline_mulAdd();
322 bool inline_montgomeryMultiply();
323 bool inline_montgomerySquare();
324 bool inline_vectorizedMismatch();
325 bool inline_fma(vmIntrinsics::ID id);
326
327 bool inline_profileBoolean();
328 bool inline_isCompileConstant();
329 };
330
331 //---------------------------make_vm_intrinsic----------------------------
332 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
333 vmIntrinsics::ID id = m->intrinsic_id();
334 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
335
336 if (!m->is_loaded()) {
337 // Do not attempt to inline unloaded methods.
338 return NULL;
339 }
340
341 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
342 bool is_available = false;
343
344 {
345 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
346 // the compiler must transition to '_thread_in_vm' state because both
347 // methods access VM-internal data.
348 VM_ENTRY_MARK;
349 methodHandle mh(THREAD, m->get_Method());
350 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
351 !C->directive()->is_intrinsic_disabled(mh) &&
352 !vmIntrinsics::is_disabled_by_flags(mh);
353
354 }
355
356 if (is_available) {
357 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
358 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
359 return new LibraryIntrinsic(m, is_virtual,
360 vmIntrinsics::predicates_needed(id),
361 vmIntrinsics::does_virtual_dispatch(id),
362 (vmIntrinsics::ID) id);
363 } else {
364 return NULL;
365 }
366 }
367
368 //----------------------register_library_intrinsics-----------------------
369 // Initialize this file's data structures, for each Compile instance.
370 void Compile::register_library_intrinsics() {
371 // Nothing to do here.
372 }
373
374 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
375 LibraryCallKit kit(jvms, this);
376 Compile* C = kit.C;
377 int nodes = C->unique();
378 #ifndef PRODUCT
379 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
380 char buf[1000];
381 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
382 tty->print_cr("Intrinsic %s", str);
383 }
384 #endif
385 ciMethod* callee = kit.callee();
386 const int bci = kit.bci();
387
388 // Try to inline the intrinsic.
389 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
390 kit.try_to_inline(_last_predicate)) {
391 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
392 : "(intrinsic)";
393 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
394 if (C->print_intrinsics() || C->print_inlining()) {
395 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
396 }
397 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
398 if (C->log()) {
399 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
400 vmIntrinsics::name_at(intrinsic_id()),
401 (is_virtual() ? " virtual='1'" : ""),
402 C->unique() - nodes);
403 }
404 // Push the result from the inlined method onto the stack.
405 kit.push_result();
406 C->print_inlining_update(this);
407 return kit.transfer_exceptions_into_jvms();
408 }
409
410 // The intrinsic bailed out
411 if (jvms->has_method()) {
412 // Not a root compile.
413 const char* msg;
414 if (callee->intrinsic_candidate()) {
415 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
416 } else {
417 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
418 : "failed to inline (intrinsic), method not annotated";
419 }
420 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
421 if (C->print_intrinsics() || C->print_inlining()) {
422 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
423 }
424 } else {
425 // Root compile
426 ResourceMark rm;
427 stringStream msg_stream;
428 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
429 vmIntrinsics::name_at(intrinsic_id()),
430 is_virtual() ? " (virtual)" : "", bci);
431 const char *msg = msg_stream.as_string();
432 log_debug(jit, inlining)("%s", msg);
433 if (C->print_intrinsics() || C->print_inlining()) {
434 tty->print("%s", msg);
435 }
436 }
437 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
438 C->print_inlining_update(this);
439 return NULL;
440 }
441
442 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
443 LibraryCallKit kit(jvms, this);
444 Compile* C = kit.C;
445 int nodes = C->unique();
446 _last_predicate = predicate;
447 #ifndef PRODUCT
448 assert(is_predicated() && predicate < predicates_count(), "sanity");
449 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
450 char buf[1000];
451 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
452 tty->print_cr("Predicate for intrinsic %s", str);
453 }
454 #endif
455 ciMethod* callee = kit.callee();
456 const int bci = kit.bci();
457
458 Node* slow_ctl = kit.try_to_predicate(predicate);
459 if (!kit.failing()) {
460 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
461 : "(intrinsic, predicate)";
462 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
463 if (C->print_intrinsics() || C->print_inlining()) {
464 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
465 }
466 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
467 if (C->log()) {
468 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
469 vmIntrinsics::name_at(intrinsic_id()),
470 (is_virtual() ? " virtual='1'" : ""),
471 C->unique() - nodes);
472 }
473 return slow_ctl; // Could be NULL if the check folds.
474 }
475
476 // The intrinsic bailed out
477 if (jvms->has_method()) {
478 // Not a root compile.
479 const char* msg = "failed to generate predicate for intrinsic";
480 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
481 if (C->print_intrinsics() || C->print_inlining()) {
482 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
483 }
484 } else {
485 // Root compile
486 ResourceMark rm;
487 stringStream msg_stream;
488 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
489 vmIntrinsics::name_at(intrinsic_id()),
490 is_virtual() ? " (virtual)" : "", bci);
491 const char *msg = msg_stream.as_string();
492 log_debug(jit, inlining)("%s", msg);
493 if (C->print_intrinsics() || C->print_inlining()) {
494 C->print_inlining_stream()->print("%s", msg);
495 }
496 }
497 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
498 return NULL;
499 }
500
501 bool LibraryCallKit::try_to_inline(int predicate) {
502 // Handle symbolic names for otherwise undistinguished boolean switches:
503 const bool is_store = true;
504 const bool is_compress = true;
505 const bool is_static = true;
506 const bool is_volatile = true;
507
508 if (!jvms()->has_method()) {
509 // Root JVMState has a null method.
510 assert(map()->memory()->Opcode() == Op_Parm, "");
511 // Insert the memory aliasing node
512 set_all_memory(reset_memory());
513 }
514 assert(merged_memory(), "");
515
516
517 switch (intrinsic_id()) {
518 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
519 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
520 case vmIntrinsics::_getClass: return inline_native_getClass();
521
522 case vmIntrinsics::_dsin:
523 case vmIntrinsics::_dcos:
524 case vmIntrinsics::_dtan:
525 case vmIntrinsics::_dabs:
526 case vmIntrinsics::_datan2:
527 case vmIntrinsics::_dsqrt:
528 case vmIntrinsics::_dexp:
529 case vmIntrinsics::_dlog:
530 case vmIntrinsics::_dlog10:
531 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
532
533 case vmIntrinsics::_min:
534 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
535
536 case vmIntrinsics::_notify:
537 case vmIntrinsics::_notifyAll:
538 if (InlineNotify) {
539 return inline_notify(intrinsic_id());
540 }
541 return false;
542
543 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
544 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
545 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
546 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
547 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
548 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
549 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
550 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
551 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
552 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
553 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
554 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
555
556 case vmIntrinsics::_arraycopy: return inline_arraycopy();
557
558 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
559 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
560 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
561 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
562
563 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
564 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
565 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
566 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
567 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
568 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
569 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
570
571 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
572 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
573
574 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
575 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
576 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
577 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
578
579 case vmIntrinsics::_compressStringC:
580 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
581 case vmIntrinsics::_inflateStringC:
582 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
583
584 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
585 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
586 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
587 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
588 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
589 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
590 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
591 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
592 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
593
594 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
595 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
596 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
597 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
598 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
599 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
600 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
601 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
602 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
603
604 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
605 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
606 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
607 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
608 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
609 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
610 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
611 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
612 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
613
614 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
615 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
616 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
617 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
618 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
619 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
620 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
621 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
622 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
623
624 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
625 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
626 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
627 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
628
629 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
630 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
631 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
632 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
633
634 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
635 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
636 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
637 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
638 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
639 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
640 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
641 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
642 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
643
644 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
645 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
646 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
647 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
648 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
649 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
650 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
651 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
652 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
653
654 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
655 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
656 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
657 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
658 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
659 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
660 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
661 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
662 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
663
664 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
665 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
666 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
667 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
668 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
669 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
670 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
671 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
672 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
673
674 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
675 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
676 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
677 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
678 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
679
680 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
681 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
682 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
683 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
684 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
685 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
686 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
687 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
688 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
689 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
690 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
691 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
692 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
693 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
694 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
695 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
696 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
697 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
698 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
699 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
700
701 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
702 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
703 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
704 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
705 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
706 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
707 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
708 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
709 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
710 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
711 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
712 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
713 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
714 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
715 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
716
717 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
718 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
719 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
720 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
721
722 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
723 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
724 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
725 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
726 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
727
728 case vmIntrinsics::_loadFence:
729 case vmIntrinsics::_storeFence:
730 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
731
732 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
733
734 case vmIntrinsics::_currentThread: return inline_native_currentThread();
735 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
736
737 #ifdef TRACE_HAVE_INTRINSICS
738 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
739 case vmIntrinsics::_getClassId: return inline_native_classID();
740 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
741 #endif
742 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
743 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
744 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
745 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
746 case vmIntrinsics::_getLength: return inline_native_getLength();
747 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
748 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
749 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
750 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
751 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
752 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
753
754 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
755 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
756
757 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
758
759 case vmIntrinsics::_isInstance:
760 case vmIntrinsics::_getModifiers:
761 case vmIntrinsics::_isInterface:
762 case vmIntrinsics::_isArray:
763 case vmIntrinsics::_isPrimitive:
764 case vmIntrinsics::_getSuperclass:
765 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
766
767 case vmIntrinsics::_floatToRawIntBits:
768 case vmIntrinsics::_floatToIntBits:
769 case vmIntrinsics::_intBitsToFloat:
770 case vmIntrinsics::_doubleToRawLongBits:
771 case vmIntrinsics::_doubleToLongBits:
772 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
773
774 case vmIntrinsics::_numberOfLeadingZeros_i:
775 case vmIntrinsics::_numberOfLeadingZeros_l:
776 case vmIntrinsics::_numberOfTrailingZeros_i:
777 case vmIntrinsics::_numberOfTrailingZeros_l:
778 case vmIntrinsics::_bitCount_i:
779 case vmIntrinsics::_bitCount_l:
780 case vmIntrinsics::_reverseBytes_i:
781 case vmIntrinsics::_reverseBytes_l:
782 case vmIntrinsics::_reverseBytes_s:
783 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
784
785 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
786
787 case vmIntrinsics::_Reference_get: return inline_reference_get();
788
789 case vmIntrinsics::_Class_cast: return inline_Class_cast();
790
791 case vmIntrinsics::_aescrypt_encryptBlock:
792 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
793
794 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
795 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
796 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
797
798 case vmIntrinsics::_counterMode_AESCrypt:
799 return inline_counterMode_AESCrypt(intrinsic_id());
800
801 case vmIntrinsics::_sha_implCompress:
802 case vmIntrinsics::_sha2_implCompress:
803 case vmIntrinsics::_sha5_implCompress:
804 return inline_sha_implCompress(intrinsic_id());
805
806 case vmIntrinsics::_digestBase_implCompressMB:
807 return inline_digestBase_implCompressMB(predicate);
808
809 case vmIntrinsics::_multiplyToLen:
810 return inline_multiplyToLen();
811
812 case vmIntrinsics::_squareToLen:
813 return inline_squareToLen();
814
815 case vmIntrinsics::_mulAdd:
816 return inline_mulAdd();
817
818 case vmIntrinsics::_montgomeryMultiply:
819 return inline_montgomeryMultiply();
820 case vmIntrinsics::_montgomerySquare:
821 return inline_montgomerySquare();
822
823 case vmIntrinsics::_vectorizedMismatch:
824 return inline_vectorizedMismatch();
825
826 case vmIntrinsics::_ghash_processBlocks:
827 return inline_ghash_processBlocks();
828
829 case vmIntrinsics::_encodeISOArray:
830 case vmIntrinsics::_encodeByteISOArray:
831 return inline_encodeISOArray();
832
833 case vmIntrinsics::_updateCRC32:
834 return inline_updateCRC32();
835 case vmIntrinsics::_updateBytesCRC32:
836 return inline_updateBytesCRC32();
837 case vmIntrinsics::_updateByteBufferCRC32:
838 return inline_updateByteBufferCRC32();
839
840 case vmIntrinsics::_updateBytesCRC32C:
841 return inline_updateBytesCRC32C();
842 case vmIntrinsics::_updateDirectByteBufferCRC32C:
843 return inline_updateDirectByteBufferCRC32C();
844
845 case vmIntrinsics::_updateBytesAdler32:
846 return inline_updateBytesAdler32();
847 case vmIntrinsics::_updateByteBufferAdler32:
848 return inline_updateByteBufferAdler32();
849
850 case vmIntrinsics::_profileBoolean:
851 return inline_profileBoolean();
852 case vmIntrinsics::_isCompileConstant:
853 return inline_isCompileConstant();
854
855 case vmIntrinsics::_hasNegatives:
856 return inline_hasNegatives();
857
858 case vmIntrinsics::_fmaD:
859 case vmIntrinsics::_fmaF:
860 return inline_fma(intrinsic_id());
861
862 default:
863 // If you get here, it may be that someone has added a new intrinsic
864 // to the list in vmSymbols.hpp without implementing it here.
865 #ifndef PRODUCT
866 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
867 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
868 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
869 }
870 #endif
871 return false;
872 }
873 }
874
875 Node* LibraryCallKit::try_to_predicate(int predicate) {
876 if (!jvms()->has_method()) {
877 // Root JVMState has a null method.
878 assert(map()->memory()->Opcode() == Op_Parm, "");
879 // Insert the memory aliasing node
880 set_all_memory(reset_memory());
881 }
882 assert(merged_memory(), "");
883
884 switch (intrinsic_id()) {
885 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
886 return inline_cipherBlockChaining_AESCrypt_predicate(false);
887 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
888 return inline_cipherBlockChaining_AESCrypt_predicate(true);
889 case vmIntrinsics::_counterMode_AESCrypt:
890 return inline_counterMode_AESCrypt_predicate();
891 case vmIntrinsics::_digestBase_implCompressMB:
892 return inline_digestBase_implCompressMB_predicate(predicate);
893
894 default:
895 // If you get here, it may be that someone has added a new intrinsic
896 // to the list in vmSymbols.hpp without implementing it here.
897 #ifndef PRODUCT
898 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
899 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
900 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
901 }
902 #endif
903 Node* slow_ctl = control();
904 set_control(top()); // No fast path instrinsic
905 return slow_ctl;
906 }
907 }
908
909 //------------------------------set_result-------------------------------
910 // Helper function for finishing intrinsics.
911 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
912 record_for_igvn(region);
913 set_control(_gvn.transform(region));
914 set_result( _gvn.transform(value));
915 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
916 }
917
918 //------------------------------generate_guard---------------------------
919 // Helper function for generating guarded fast-slow graph structures.
920 // The given 'test', if true, guards a slow path. If the test fails
921 // then a fast path can be taken. (We generally hope it fails.)
922 // In all cases, GraphKit::control() is updated to the fast path.
923 // The returned value represents the control for the slow path.
924 // The return value is never 'top'; it is either a valid control
925 // or NULL if it is obvious that the slow path can never be taken.
926 // Also, if region and the slow control are not NULL, the slow edge
927 // is appended to the region.
928 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
929 if (stopped()) {
930 // Already short circuited.
931 return NULL;
932 }
933
934 // Build an if node and its projections.
935 // If test is true we take the slow path, which we assume is uncommon.
936 if (_gvn.type(test) == TypeInt::ZERO) {
937 // The slow branch is never taken. No need to build this guard.
938 return NULL;
939 }
940
941 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
942
943 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
944 if (if_slow == top()) {
945 // The slow branch is never taken. No need to build this guard.
946 return NULL;
947 }
948
949 if (region != NULL)
950 region->add_req(if_slow);
951
952 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
953 set_control(if_fast);
954
955 return if_slow;
956 }
957
958 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
959 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
960 }
961 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
962 return generate_guard(test, region, PROB_FAIR);
963 }
964
965 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
966 Node* *pos_index) {
967 if (stopped())
968 return NULL; // already stopped
969 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
970 return NULL; // index is already adequately typed
971 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
972 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
973 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
974 if (is_neg != NULL && pos_index != NULL) {
975 // Emulate effect of Parse::adjust_map_after_if.
976 Node* ccast = new CastIINode(index, TypeInt::POS);
977 ccast->set_req(0, control());
978 (*pos_index) = _gvn.transform(ccast);
979 }
980 return is_neg;
981 }
982
983 // Make sure that 'position' is a valid limit index, in [0..length].
984 // There are two equivalent plans for checking this:
985 // A. (offset + copyLength) unsigned<= arrayLength
986 // B. offset <= (arrayLength - copyLength)
987 // We require that all of the values above, except for the sum and
988 // difference, are already known to be non-negative.
989 // Plan A is robust in the face of overflow, if offset and copyLength
990 // are both hugely positive.
991 //
992 // Plan B is less direct and intuitive, but it does not overflow at
993 // all, since the difference of two non-negatives is always
994 // representable. Whenever Java methods must perform the equivalent
995 // check they generally use Plan B instead of Plan A.
996 // For the moment we use Plan A.
997 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
998 Node* subseq_length,
999 Node* array_length,
1000 RegionNode* region) {
1001 if (stopped())
1002 return NULL; // already stopped
1003 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1004 if (zero_offset && subseq_length->eqv_uncast(array_length))
1005 return NULL; // common case of whole-array copy
1006 Node* last = subseq_length;
1007 if (!zero_offset) // last += offset
1008 last = _gvn.transform(new AddINode(last, offset));
1009 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1010 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1011 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1012 return is_over;
1013 }
1014
1015 // Emit range checks for the given String.value byte array
1016 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1017 if (stopped()) {
1018 return; // already stopped
1019 }
1020 RegionNode* bailout = new RegionNode(1);
1021 record_for_igvn(bailout);
1022 if (char_count) {
1023 // Convert char count to byte count
1024 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1025 }
1026
1027 // Offset and count must not be negative
1028 generate_negative_guard(offset, bailout);
1029 generate_negative_guard(count, bailout);
1030 // Offset + count must not exceed length of array
1031 generate_limit_guard(offset, count, load_array_length(array), bailout);
1032
1033 if (bailout->req() > 1) {
1034 PreserveJVMState pjvms(this);
1035 set_control(_gvn.transform(bailout));
1036 uncommon_trap(Deoptimization::Reason_intrinsic,
1037 Deoptimization::Action_maybe_recompile);
1038 }
1039 }
1040
1041 //--------------------------generate_current_thread--------------------
1042 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1043 ciKlass* thread_klass = env()->Thread_klass();
1044 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1045 Node* thread = _gvn.transform(new ThreadLocalNode());
1046 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1047 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1048 tls_output = thread;
1049 return threadObj;
1050 }
1051
1052
1053 //------------------------------make_string_method_node------------------------
1054 // Helper method for String intrinsic functions. This version is called with
1055 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1056 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1057 // containing the lengths of str1 and str2.
1058 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1059 Node* result = NULL;
1060 switch (opcode) {
1061 case Op_StrIndexOf:
1062 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1063 str1_start, cnt1, str2_start, cnt2, ae);
1064 break;
1065 case Op_StrComp:
1066 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1067 str1_start, cnt1, str2_start, cnt2, ae);
1068 break;
1069 case Op_StrEquals:
1070 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1071 // Use the constant length if there is one because optimized match rule may exist.
1072 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1073 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1074 break;
1075 default:
1076 ShouldNotReachHere();
1077 return NULL;
1078 }
1079
1080 // All these intrinsics have checks.
1081 C->set_has_split_ifs(true); // Has chance for split-if optimization
1082
1083 return _gvn.transform(result);
1084 }
1085
1086 //------------------------------inline_string_compareTo------------------------
1087 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1088 Node* arg1 = argument(0);
1089 Node* arg2 = argument(1);
1090
1091 // Get start addr and length of first argument
1092 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1093 Node* arg1_cnt = load_array_length(arg1);
1094
1095 // Get start addr and length of second argument
1096 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1097 Node* arg2_cnt = load_array_length(arg2);
1098
1099 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1100 set_result(result);
1101 return true;
1102 }
1103
1104 //------------------------------inline_string_equals------------------------
1105 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1106 Node* arg1 = argument(0);
1107 Node* arg2 = argument(1);
1108
1109 // paths (plus control) merge
1110 RegionNode* region = new RegionNode(3);
1111 Node* phi = new PhiNode(region, TypeInt::BOOL);
1112
1113 if (!stopped()) {
1114 // Get start addr and length of first argument
1115 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1116 Node* arg1_cnt = load_array_length(arg1);
1117
1118 // Get start addr and length of second argument
1119 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1120 Node* arg2_cnt = load_array_length(arg2);
1121
1122 // Check for arg1_cnt != arg2_cnt
1123 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1124 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1125 Node* if_ne = generate_slow_guard(bol, NULL);
1126 if (if_ne != NULL) {
1127 phi->init_req(2, intcon(0));
1128 region->init_req(2, if_ne);
1129 }
1130
1131 // Check for count == 0 is done by assembler code for StrEquals.
1132
1133 if (!stopped()) {
1134 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1135 phi->init_req(1, equals);
1136 region->init_req(1, control());
1137 }
1138 }
1139
1140 // post merge
1141 set_control(_gvn.transform(region));
1142 record_for_igvn(region);
1143
1144 set_result(_gvn.transform(phi));
1145 return true;
1146 }
1147
1148 //------------------------------inline_array_equals----------------------------
1149 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1150 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1151 Node* arg1 = argument(0);
1152 Node* arg2 = argument(1);
1153
1154 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1155 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1156 return true;
1157 }
1158
1159 //------------------------------inline_hasNegatives------------------------------
1160 bool LibraryCallKit::inline_hasNegatives() {
1161 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1162 return false;
1163 }
1164
1165 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1166 // no receiver since it is static method
1167 Node* ba = argument(0);
1168 Node* offset = argument(1);
1169 Node* len = argument(2);
1170
1171 // Range checks
1172 generate_string_range_check(ba, offset, len, false);
1173 if (stopped()) {
1174 return true;
1175 }
1176 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1177 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1178 set_result(_gvn.transform(result));
1179 return true;
1180 }
1181
1182 bool LibraryCallKit::inline_preconditions_checkIndex() {
1183 Node* index = argument(0);
1184 Node* length = argument(1);
1185 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1186 return false;
1187 }
1188
1189 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1190 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1191
1192 {
1193 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1194 uncommon_trap(Deoptimization::Reason_intrinsic,
1195 Deoptimization::Action_make_not_entrant);
1196 }
1197
1198 if (stopped()) {
1199 return false;
1200 }
1201
1202 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1203 BoolTest::mask btest = BoolTest::lt;
1204 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1205 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1206 _gvn.set_type(rc, rc->Value(&_gvn));
1207 if (!rc_bool->is_Con()) {
1208 record_for_igvn(rc);
1209 }
1210 set_control(_gvn.transform(new IfTrueNode(rc)));
1211 {
1212 PreserveJVMState pjvms(this);
1213 set_control(_gvn.transform(new IfFalseNode(rc)));
1214 uncommon_trap(Deoptimization::Reason_range_check,
1215 Deoptimization::Action_make_not_entrant);
1216 }
1217
1218 if (stopped()) {
1219 return false;
1220 }
1221
1222 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1223 result->set_req(0, control());
1224 result = _gvn.transform(result);
1225 set_result(result);
1226 replace_in_map(index, result);
1227 return true;
1228 }
1229
1230 //------------------------------inline_string_indexOf------------------------
1231 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1232 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1233 return false;
1234 }
1235 Node* src = argument(0);
1236 Node* tgt = argument(1);
1237
1238 // Make the merge point
1239 RegionNode* result_rgn = new RegionNode(4);
1240 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1241
1242 // Get start addr and length of source string
1243 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1244 Node* src_count = load_array_length(src);
1245
1246 // Get start addr and length of substring
1247 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1248 Node* tgt_count = load_array_length(tgt);
1249
1250 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1251 // Divide src size by 2 if String is UTF16 encoded
1252 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1253 }
1254 if (ae == StrIntrinsicNode::UU) {
1255 // Divide substring size by 2 if String is UTF16 encoded
1256 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1257 }
1258
1259 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1260 if (result != NULL) {
1261 result_phi->init_req(3, result);
1262 result_rgn->init_req(3, control());
1263 }
1264 set_control(_gvn.transform(result_rgn));
1265 record_for_igvn(result_rgn);
1266 set_result(_gvn.transform(result_phi));
1267
1268 return true;
1269 }
1270
1271 //-----------------------------inline_string_indexOf-----------------------
1272 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1273 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1274 return false;
1275 }
1276 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1277 return false;
1278 }
1279 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1280 Node* src = argument(0); // byte[]
1281 Node* src_count = argument(1); // char count
1282 Node* tgt = argument(2); // byte[]
1283 Node* tgt_count = argument(3); // char count
1284 Node* from_index = argument(4); // char index
1285
1286 // Multiply byte array index by 2 if String is UTF16 encoded
1287 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1288 src_count = _gvn.transform(new SubINode(src_count, from_index));
1289 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1290 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1291
1292 // Range checks
1293 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1294 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1295 if (stopped()) {
1296 return true;
1297 }
1298
1299 RegionNode* region = new RegionNode(5);
1300 Node* phi = new PhiNode(region, TypeInt::INT);
1301
1302 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1303 if (result != NULL) {
1304 // The result is index relative to from_index if substring was found, -1 otherwise.
1305 // Generate code which will fold into cmove.
1306 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1307 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1308
1309 Node* if_lt = generate_slow_guard(bol, NULL);
1310 if (if_lt != NULL) {
1311 // result == -1
1312 phi->init_req(3, result);
1313 region->init_req(3, if_lt);
1314 }
1315 if (!stopped()) {
1316 result = _gvn.transform(new AddINode(result, from_index));
1317 phi->init_req(4, result);
1318 region->init_req(4, control());
1319 }
1320 }
1321
1322 set_control(_gvn.transform(region));
1323 record_for_igvn(region);
1324 set_result(_gvn.transform(phi));
1325
1326 return true;
1327 }
1328
1329 // Create StrIndexOfNode with fast path checks
1330 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1331 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1332 // Check for substr count > string count
1333 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1334 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1335 Node* if_gt = generate_slow_guard(bol, NULL);
1336 if (if_gt != NULL) {
1337 phi->init_req(1, intcon(-1));
1338 region->init_req(1, if_gt);
1339 }
1340 if (!stopped()) {
1341 // Check for substr count == 0
1342 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1343 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1344 Node* if_zero = generate_slow_guard(bol, NULL);
1345 if (if_zero != NULL) {
1346 phi->init_req(2, intcon(0));
1347 region->init_req(2, if_zero);
1348 }
1349 }
1350 if (!stopped()) {
1351 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1352 }
1353 return NULL;
1354 }
1355
1356 //-----------------------------inline_string_indexOfChar-----------------------
1357 bool LibraryCallKit::inline_string_indexOfChar() {
1358 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1359 return false;
1360 }
1361 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1362 return false;
1363 }
1364 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1365 Node* src = argument(0); // byte[]
1366 Node* tgt = argument(1); // tgt is int ch
1367 Node* from_index = argument(2);
1368 Node* max = argument(3);
1369
1370 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1371 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1372 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1373
1374 // Range checks
1375 generate_string_range_check(src, src_offset, src_count, true);
1376 if (stopped()) {
1377 return true;
1378 }
1379
1380 RegionNode* region = new RegionNode(3);
1381 Node* phi = new PhiNode(region, TypeInt::INT);
1382
1383 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1384 C->set_has_split_ifs(true); // Has chance for split-if optimization
1385 _gvn.transform(result);
1386
1387 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1388 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1389
1390 Node* if_lt = generate_slow_guard(bol, NULL);
1391 if (if_lt != NULL) {
1392 // result == -1
1393 phi->init_req(2, result);
1394 region->init_req(2, if_lt);
1395 }
1396 if (!stopped()) {
1397 result = _gvn.transform(new AddINode(result, from_index));
1398 phi->init_req(1, result);
1399 region->init_req(1, control());
1400 }
1401 set_control(_gvn.transform(region));
1402 record_for_igvn(region);
1403 set_result(_gvn.transform(phi));
1404
1405 return true;
1406 }
1407 //---------------------------inline_string_copy---------------------
1408 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1409 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1410 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1411 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1412 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1413 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1414 bool LibraryCallKit::inline_string_copy(bool compress) {
1415 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1416 return false;
1417 }
1418 int nargs = 5; // 2 oops, 3 ints
1419 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1420
1421 Node* src = argument(0);
1422 Node* src_offset = argument(1);
1423 Node* dst = argument(2);
1424 Node* dst_offset = argument(3);
1425 Node* length = argument(4);
1426
1427 // Check for allocation before we add nodes that would confuse
1428 // tightly_coupled_allocation()
1429 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1430
1431 // Figure out the size and type of the elements we will be copying.
1432 const Type* src_type = src->Value(&_gvn);
1433 const Type* dst_type = dst->Value(&_gvn);
1434 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1435 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1436 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1437 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1438 "Unsupported array types for inline_string_copy");
1439
1440 // Convert char[] offsets to byte[] offsets
1441 bool convert_src = (compress && src_elem == T_BYTE);
1442 bool convert_dst = (!compress && dst_elem == T_BYTE);
1443 if (convert_src) {
1444 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1445 } else if (convert_dst) {
1446 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1447 }
1448
1449 // Range checks
1450 generate_string_range_check(src, src_offset, length, convert_src);
1451 generate_string_range_check(dst, dst_offset, length, convert_dst);
1452 if (stopped()) {
1453 return true;
1454 }
1455
1456 Node* src_start = array_element_address(src, src_offset, src_elem);
1457 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1458 // 'src_start' points to src array + scaled offset
1459 // 'dst_start' points to dst array + scaled offset
1460 Node* count = NULL;
1461 if (compress) {
1462 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1463 } else {
1464 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1465 }
1466
1467 if (alloc != NULL) {
1468 if (alloc->maybe_set_complete(&_gvn)) {
1469 // "You break it, you buy it."
1470 InitializeNode* init = alloc->initialization();
1471 assert(init->is_complete(), "we just did this");
1472 init->set_complete_with_arraycopy();
1473 assert(dst->is_CheckCastPP(), "sanity");
1474 assert(dst->in(0)->in(0) == init, "dest pinned");
1475 }
1476 // Do not let stores that initialize this object be reordered with
1477 // a subsequent store that would make this object accessible by
1478 // other threads.
1479 // Record what AllocateNode this StoreStore protects so that
1480 // escape analysis can go from the MemBarStoreStoreNode to the
1481 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1482 // based on the escape status of the AllocateNode.
1483 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1484 }
1485 if (compress) {
1486 set_result(_gvn.transform(count));
1487 }
1488 return true;
1489 }
1490
1491 #ifdef _LP64
1492 #define XTOP ,top() /*additional argument*/
1493 #else //_LP64
1494 #define XTOP /*no additional argument*/
1495 #endif //_LP64
1496
1497 //------------------------inline_string_toBytesU--------------------------
1498 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1499 bool LibraryCallKit::inline_string_toBytesU() {
1500 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1501 return false;
1502 }
1503 // Get the arguments.
1504 Node* value = argument(0);
1505 Node* offset = argument(1);
1506 Node* length = argument(2);
1507
1508 Node* newcopy = NULL;
1509
1510 // Set the original stack and the reexecute bit for the interpreter to reexecute
1511 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1512 { PreserveReexecuteState preexecs(this);
1513 jvms()->set_should_reexecute(true);
1514
1515 // Check if a null path was taken unconditionally.
1516 value = null_check(value);
1517
1518 RegionNode* bailout = new RegionNode(1);
1519 record_for_igvn(bailout);
1520
1521 // Range checks
1522 generate_negative_guard(offset, bailout);
1523 generate_negative_guard(length, bailout);
1524 generate_limit_guard(offset, length, load_array_length(value), bailout);
1525 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1526 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1527
1528 if (bailout->req() > 1) {
1529 PreserveJVMState pjvms(this);
1530 set_control(_gvn.transform(bailout));
1531 uncommon_trap(Deoptimization::Reason_intrinsic,
1532 Deoptimization::Action_maybe_recompile);
1533 }
1534 if (stopped()) {
1535 return true;
1536 }
1537
1538 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1539 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1540 newcopy = new_array(klass_node, size, 0); // no arguments to push
1541 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1542
1543 // Calculate starting addresses.
1544 Node* src_start = array_element_address(value, offset, T_CHAR);
1545 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1546
1547 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1548 const TypeInt* toffset = gvn().type(offset)->is_int();
1549 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1550
1551 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1552 const char* copyfunc_name = "arraycopy";
1553 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1554 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1555 OptoRuntime::fast_arraycopy_Type(),
1556 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1557 src_start, dst_start, ConvI2X(length) XTOP);
1558 // Do not let reads from the cloned object float above the arraycopy.
1559 if (alloc != NULL) {
1560 if (alloc->maybe_set_complete(&_gvn)) {
1561 // "You break it, you buy it."
1562 InitializeNode* init = alloc->initialization();
1563 assert(init->is_complete(), "we just did this");
1564 init->set_complete_with_arraycopy();
1565 assert(newcopy->is_CheckCastPP(), "sanity");
1566 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1567 }
1568 // Do not let stores that initialize this object be reordered with
1569 // a subsequent store that would make this object accessible by
1570 // other threads.
1571 // Record what AllocateNode this StoreStore protects so that
1572 // escape analysis can go from the MemBarStoreStoreNode to the
1573 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1574 // based on the escape status of the AllocateNode.
1575 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1576 } else {
1577 insert_mem_bar(Op_MemBarCPUOrder);
1578 }
1579 } // original reexecute is set back here
1580
1581 C->set_has_split_ifs(true); // Has chance for split-if optimization
1582 if (!stopped()) {
1583 set_result(newcopy);
1584 }
1585 return true;
1586 }
1587
1588 //------------------------inline_string_getCharsU--------------------------
1589 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1590 bool LibraryCallKit::inline_string_getCharsU() {
1591 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1592 return false;
1593 }
1594
1595 // Get the arguments.
1596 Node* src = argument(0);
1597 Node* src_begin = argument(1);
1598 Node* src_end = argument(2); // exclusive offset (i < src_end)
1599 Node* dst = argument(3);
1600 Node* dst_begin = argument(4);
1601
1602 // Check for allocation before we add nodes that would confuse
1603 // tightly_coupled_allocation()
1604 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1605
1606 // Check if a null path was taken unconditionally.
1607 src = null_check(src);
1608 dst = null_check(dst);
1609 if (stopped()) {
1610 return true;
1611 }
1612
1613 // Get length and convert char[] offset to byte[] offset
1614 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1615 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1616
1617 // Range checks
1618 generate_string_range_check(src, src_begin, length, true);
1619 generate_string_range_check(dst, dst_begin, length, false);
1620 if (stopped()) {
1621 return true;
1622 }
1623
1624 if (!stopped()) {
1625 // Calculate starting addresses.
1626 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1627 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1628
1629 // Check if array addresses are aligned to HeapWordSize
1630 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1631 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1632 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1633 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1634
1635 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1636 const char* copyfunc_name = "arraycopy";
1637 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1638 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1639 OptoRuntime::fast_arraycopy_Type(),
1640 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1641 src_start, dst_start, ConvI2X(length) XTOP);
1642 // Do not let reads from the cloned object float above the arraycopy.
1643 if (alloc != NULL) {
1644 if (alloc->maybe_set_complete(&_gvn)) {
1645 // "You break it, you buy it."
1646 InitializeNode* init = alloc->initialization();
1647 assert(init->is_complete(), "we just did this");
1648 init->set_complete_with_arraycopy();
1649 assert(dst->is_CheckCastPP(), "sanity");
1650 assert(dst->in(0)->in(0) == init, "dest pinned");
1651 }
1652 // Do not let stores that initialize this object be reordered with
1653 // a subsequent store that would make this object accessible by
1654 // other threads.
1655 // Record what AllocateNode this StoreStore protects so that
1656 // escape analysis can go from the MemBarStoreStoreNode to the
1657 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1658 // based on the escape status of the AllocateNode.
1659 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1660 } else {
1661 insert_mem_bar(Op_MemBarCPUOrder);
1662 }
1663 }
1664
1665 C->set_has_split_ifs(true); // Has chance for split-if optimization
1666 return true;
1667 }
1668
1669 //----------------------inline_string_char_access----------------------------
1670 // Store/Load char to/from byte[] array.
1671 // static void StringUTF16.putChar(byte[] val, int index, int c)
1672 // static char StringUTF16.getChar(byte[] val, int index)
1673 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1674 Node* value = argument(0);
1675 Node* index = argument(1);
1676 Node* ch = is_store ? argument(2) : NULL;
1677
1678 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1679 // correctly requires matched array shapes.
1680 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1681 "sanity: byte[] and char[] bases agree");
1682 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1683 "sanity: byte[] and char[] scales agree");
1684
1685 // Bail when getChar over constants is requested: constant folding would
1686 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1687 // Java method would constant fold nicely instead.
1688 if (!is_store && value->is_Con() && index->is_Con()) {
1689 return false;
1690 }
1691
1692 Node* adr = array_element_address(value, index, T_CHAR);
1693 if (adr->is_top()) {
1694 return false;
1695 }
1696 if (is_store) {
1697 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1698 false, false, true /* mismatched */);
1699 } else {
1700 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1701 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1702 set_result(ch);
1703 }
1704 return true;
1705 }
1706
1707 //--------------------------round_double_node--------------------------------
1708 // Round a double node if necessary.
1709 Node* LibraryCallKit::round_double_node(Node* n) {
1710 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1711 n = _gvn.transform(new RoundDoubleNode(0, n));
1712 return n;
1713 }
1714
1715 //------------------------------inline_math-----------------------------------
1716 // public static double Math.abs(double)
1717 // public static double Math.sqrt(double)
1718 // public static double Math.log(double)
1719 // public static double Math.log10(double)
1720 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1721 Node* arg = round_double_node(argument(0));
1722 Node* n = NULL;
1723 switch (id) {
1724 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1725 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1726 default: fatal_unexpected_iid(id); break;
1727 }
1728 set_result(_gvn.transform(n));
1729 return true;
1730 }
1731
1732 //------------------------------runtime_math-----------------------------
1733 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1734 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1735 "must be (DD)D or (D)D type");
1736
1737 // Inputs
1738 Node* a = round_double_node(argument(0));
1739 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1740
1741 const TypePtr* no_memory_effects = NULL;
1742 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1743 no_memory_effects,
1744 a, top(), b, b ? top() : NULL);
1745 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1746 #ifdef ASSERT
1747 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1748 assert(value_top == top(), "second value must be top");
1749 #endif
1750
1751 set_result(value);
1752 return true;
1753 }
1754
1755 //------------------------------inline_math_native-----------------------------
1756 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1757 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1758 switch (id) {
1759 // These intrinsics are not properly supported on all hardware
1760 case vmIntrinsics::_dsin:
1761 return StubRoutines::dsin() != NULL ?
1762 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1763 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1764 case vmIntrinsics::_dcos:
1765 return StubRoutines::dcos() != NULL ?
1766 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1767 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1768 case vmIntrinsics::_dtan:
1769 return StubRoutines::dtan() != NULL ?
1770 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1771 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1772 case vmIntrinsics::_dlog:
1773 return StubRoutines::dlog() != NULL ?
1774 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1775 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1776 case vmIntrinsics::_dlog10:
1777 return StubRoutines::dlog10() != NULL ?
1778 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1779 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1780
1781 // These intrinsics are supported on all hardware
1782 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1783 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1784
1785 case vmIntrinsics::_dexp:
1786 return StubRoutines::dexp() != NULL ?
1787 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1788 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1789 case vmIntrinsics::_dpow:
1790 return StubRoutines::dpow() != NULL ?
1791 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1792 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1793 #undef FN_PTR
1794
1795 // These intrinsics are not yet correctly implemented
1796 case vmIntrinsics::_datan2:
1797 return false;
1798
1799 default:
1800 fatal_unexpected_iid(id);
1801 return false;
1802 }
1803 }
1804
1805 static bool is_simple_name(Node* n) {
1806 return (n->req() == 1 // constant
1807 || (n->is_Type() && n->as_Type()->type()->singleton())
1808 || n->is_Proj() // parameter or return value
1809 || n->is_Phi() // local of some sort
1810 );
1811 }
1812
1813 //----------------------------inline_notify-----------------------------------*
1814 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1815 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1816 address func;
1817 if (id == vmIntrinsics::_notify) {
1818 func = OptoRuntime::monitor_notify_Java();
1819 } else {
1820 func = OptoRuntime::monitor_notifyAll_Java();
1821 }
1822 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1823 make_slow_call_ex(call, env()->Throwable_klass(), false);
1824 return true;
1825 }
1826
1827
1828 //----------------------------inline_min_max-----------------------------------
1829 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1830 set_result(generate_min_max(id, argument(0), argument(1)));
1831 return true;
1832 }
1833
1834 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1835 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1836 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1837 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1838 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1839
1840 {
1841 PreserveJVMState pjvms(this);
1842 PreserveReexecuteState preexecs(this);
1843 jvms()->set_should_reexecute(true);
1844
1845 set_control(slow_path);
1846 set_i_o(i_o());
1847
1848 uncommon_trap(Deoptimization::Reason_intrinsic,
1849 Deoptimization::Action_none);
1850 }
1851
1852 set_control(fast_path);
1853 set_result(math);
1854 }
1855
1856 template <typename OverflowOp>
1857 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1858 typedef typename OverflowOp::MathOp MathOp;
1859
1860 MathOp* mathOp = new MathOp(arg1, arg2);
1861 Node* operation = _gvn.transform( mathOp );
1862 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1863 inline_math_mathExact(operation, ofcheck);
1864 return true;
1865 }
1866
1867 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1868 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1869 }
1870
1871 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1872 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1873 }
1874
1875 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1876 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1877 }
1878
1879 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1880 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1881 }
1882
1883 bool LibraryCallKit::inline_math_negateExactI() {
1884 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1885 }
1886
1887 bool LibraryCallKit::inline_math_negateExactL() {
1888 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1889 }
1890
1891 bool LibraryCallKit::inline_math_multiplyExactI() {
1892 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1893 }
1894
1895 bool LibraryCallKit::inline_math_multiplyExactL() {
1896 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1897 }
1898
1899 Node*
1900 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1901 // These are the candidate return value:
1902 Node* xvalue = x0;
1903 Node* yvalue = y0;
1904
1905 if (xvalue == yvalue) {
1906 return xvalue;
1907 }
1908
1909 bool want_max = (id == vmIntrinsics::_max);
1910
1911 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1912 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1913 if (txvalue == NULL || tyvalue == NULL) return top();
1914 // This is not really necessary, but it is consistent with a
1915 // hypothetical MaxINode::Value method:
1916 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1917
1918 // %%% This folding logic should (ideally) be in a different place.
1919 // Some should be inside IfNode, and there to be a more reliable
1920 // transformation of ?: style patterns into cmoves. We also want
1921 // more powerful optimizations around cmove and min/max.
1922
1923 // Try to find a dominating comparison of these guys.
1924 // It can simplify the index computation for Arrays.copyOf
1925 // and similar uses of System.arraycopy.
1926 // First, compute the normalized version of CmpI(x, y).
1927 int cmp_op = Op_CmpI;
1928 Node* xkey = xvalue;
1929 Node* ykey = yvalue;
1930 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1931 if (ideal_cmpxy->is_Cmp()) {
1932 // E.g., if we have CmpI(length - offset, count),
1933 // it might idealize to CmpI(length, count + offset)
1934 cmp_op = ideal_cmpxy->Opcode();
1935 xkey = ideal_cmpxy->in(1);
1936 ykey = ideal_cmpxy->in(2);
1937 }
1938
1939 // Start by locating any relevant comparisons.
1940 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1941 Node* cmpxy = NULL;
1942 Node* cmpyx = NULL;
1943 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1944 Node* cmp = start_from->fast_out(k);
1945 if (cmp->outcnt() > 0 && // must have prior uses
1946 cmp->in(0) == NULL && // must be context-independent
1947 cmp->Opcode() == cmp_op) { // right kind of compare
1948 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1949 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1950 }
1951 }
1952
1953 const int NCMPS = 2;
1954 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1955 int cmpn;
1956 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1957 if (cmps[cmpn] != NULL) break; // find a result
1958 }
1959 if (cmpn < NCMPS) {
1960 // Look for a dominating test that tells us the min and max.
1961 int depth = 0; // Limit search depth for speed
1962 Node* dom = control();
1963 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1964 if (++depth >= 100) break;
1965 Node* ifproj = dom;
1966 if (!ifproj->is_Proj()) continue;
1967 Node* iff = ifproj->in(0);
1968 if (!iff->is_If()) continue;
1969 Node* bol = iff->in(1);
1970 if (!bol->is_Bool()) continue;
1971 Node* cmp = bol->in(1);
1972 if (cmp == NULL) continue;
1973 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1974 if (cmps[cmpn] == cmp) break;
1975 if (cmpn == NCMPS) continue;
1976 BoolTest::mask btest = bol->as_Bool()->_test._test;
1977 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1978 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1979 // At this point, we know that 'x btest y' is true.
1980 switch (btest) {
1981 case BoolTest::eq:
1982 // They are proven equal, so we can collapse the min/max.
1983 // Either value is the answer. Choose the simpler.
1984 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1985 return yvalue;
1986 return xvalue;
1987 case BoolTest::lt: // x < y
1988 case BoolTest::le: // x <= y
1989 return (want_max ? yvalue : xvalue);
1990 case BoolTest::gt: // x > y
1991 case BoolTest::ge: // x >= y
1992 return (want_max ? xvalue : yvalue);
1993 }
1994 }
1995 }
1996
1997 // We failed to find a dominating test.
1998 // Let's pick a test that might GVN with prior tests.
1999 Node* best_bol = NULL;
2000 BoolTest::mask best_btest = BoolTest::illegal;
2001 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2002 Node* cmp = cmps[cmpn];
2003 if (cmp == NULL) continue;
2004 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2005 Node* bol = cmp->fast_out(j);
2006 if (!bol->is_Bool()) continue;
2007 BoolTest::mask btest = bol->as_Bool()->_test._test;
2008 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2009 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2010 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2011 best_bol = bol->as_Bool();
2012 best_btest = btest;
2013 }
2014 }
2015 }
2016
2017 Node* answer_if_true = NULL;
2018 Node* answer_if_false = NULL;
2019 switch (best_btest) {
2020 default:
2021 if (cmpxy == NULL)
2022 cmpxy = ideal_cmpxy;
2023 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2024 // and fall through:
2025 case BoolTest::lt: // x < y
2026 case BoolTest::le: // x <= y
2027 answer_if_true = (want_max ? yvalue : xvalue);
2028 answer_if_false = (want_max ? xvalue : yvalue);
2029 break;
2030 case BoolTest::gt: // x > y
2031 case BoolTest::ge: // x >= y
2032 answer_if_true = (want_max ? xvalue : yvalue);
2033 answer_if_false = (want_max ? yvalue : xvalue);
2034 break;
2035 }
2036
2037 jint hi, lo;
2038 if (want_max) {
2039 // We can sharpen the minimum.
2040 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2041 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2042 } else {
2043 // We can sharpen the maximum.
2044 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2045 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2046 }
2047
2048 // Use a flow-free graph structure, to avoid creating excess control edges
2049 // which could hinder other optimizations.
2050 // Since Math.min/max is often used with arraycopy, we want
2051 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2052 Node* cmov = CMoveNode::make(NULL, best_bol,
2053 answer_if_false, answer_if_true,
2054 TypeInt::make(lo, hi, widen));
2055
2056 return _gvn.transform(cmov);
2057
2058 /*
2059 // This is not as desirable as it may seem, since Min and Max
2060 // nodes do not have a full set of optimizations.
2061 // And they would interfere, anyway, with 'if' optimizations
2062 // and with CMoveI canonical forms.
2063 switch (id) {
2064 case vmIntrinsics::_min:
2065 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2066 case vmIntrinsics::_max:
2067 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2068 default:
2069 ShouldNotReachHere();
2070 }
2071 */
2072 }
2073
2074 inline int
2075 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2076 const TypePtr* base_type = TypePtr::NULL_PTR;
2077 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2078 if (base_type == NULL) {
2079 // Unknown type.
2080 return Type::AnyPtr;
2081 } else if (base_type == TypePtr::NULL_PTR) {
2082 // Since this is a NULL+long form, we have to switch to a rawptr.
2083 base = _gvn.transform(new CastX2PNode(offset));
2084 offset = MakeConX(0);
2085 return Type::RawPtr;
2086 } else if (base_type->base() == Type::RawPtr) {
2087 return Type::RawPtr;
2088 } else if (base_type->isa_oopptr()) {
2089 // Base is never null => always a heap address.
2090 if (base_type->ptr() == TypePtr::NotNull) {
2091 return Type::OopPtr;
2092 }
2093 // Offset is small => always a heap address.
2094 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2095 if (offset_type != NULL &&
2096 base_type->offset() == 0 && // (should always be?)
2097 offset_type->_lo >= 0 &&
2098 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2099 return Type::OopPtr;
2100 }
2101 // Otherwise, it might either be oop+off or NULL+addr.
2102 return Type::AnyPtr;
2103 } else {
2104 // No information:
2105 return Type::AnyPtr;
2106 }
2107 }
2108
2109 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2110 int kind = classify_unsafe_addr(base, offset);
2111 if (kind == Type::RawPtr) {
2112 return basic_plus_adr(top(), base, offset);
2113 } else {
2114 return basic_plus_adr(base, offset);
2115 }
2116 }
2117
2118 //--------------------------inline_number_methods-----------------------------
2119 // inline int Integer.numberOfLeadingZeros(int)
2120 // inline int Long.numberOfLeadingZeros(long)
2121 //
2122 // inline int Integer.numberOfTrailingZeros(int)
2123 // inline int Long.numberOfTrailingZeros(long)
2124 //
2125 // inline int Integer.bitCount(int)
2126 // inline int Long.bitCount(long)
2127 //
2128 // inline char Character.reverseBytes(char)
2129 // inline short Short.reverseBytes(short)
2130 // inline int Integer.reverseBytes(int)
2131 // inline long Long.reverseBytes(long)
2132 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2133 Node* arg = argument(0);
2134 Node* n = NULL;
2135 switch (id) {
2136 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2137 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2138 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2139 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2140 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2141 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2142 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2143 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2144 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2145 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2146 default: fatal_unexpected_iid(id); break;
2147 }
2148 set_result(_gvn.transform(n));
2149 return true;
2150 }
2151
2152 //----------------------------inline_unsafe_access----------------------------
2153
2154 // Helper that guards and inserts a pre-barrier.
2155 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2156 Node* pre_val, bool need_mem_bar) {
2157 // We could be accessing the referent field of a reference object. If so, when G1
2158 // is enabled, we need to log the value in the referent field in an SATB buffer.
2159 // This routine performs some compile time filters and generates suitable
2160 // runtime filters that guard the pre-barrier code.
2161 // Also add memory barrier for non volatile load from the referent field
2162 // to prevent commoning of loads across safepoint.
2163 if (!UseG1GC && !need_mem_bar)
2164 return;
2165
2166 // Some compile time checks.
2167
2168 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2169 const TypeX* otype = offset->find_intptr_t_type();
2170 if (otype != NULL && otype->is_con() &&
2171 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2172 // Constant offset but not the reference_offset so just return
2173 return;
2174 }
2175
2176 // We only need to generate the runtime guards for instances.
2177 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2178 if (btype != NULL) {
2179 if (btype->isa_aryptr()) {
2180 // Array type so nothing to do
2181 return;
2182 }
2183
2184 const TypeInstPtr* itype = btype->isa_instptr();
2185 if (itype != NULL) {
2186 // Can the klass of base_oop be statically determined to be
2187 // _not_ a sub-class of Reference and _not_ Object?
2188 ciKlass* klass = itype->klass();
2189 if ( klass->is_loaded() &&
2190 !klass->is_subtype_of(env()->Reference_klass()) &&
2191 !env()->Object_klass()->is_subtype_of(klass)) {
2192 return;
2193 }
2194 }
2195 }
2196
2197 // The compile time filters did not reject base_oop/offset so
2198 // we need to generate the following runtime filters
2199 //
2200 // if (offset == java_lang_ref_Reference::_reference_offset) {
2201 // if (instance_of(base, java.lang.ref.Reference)) {
2202 // pre_barrier(_, pre_val, ...);
2203 // }
2204 // }
2205
2206 float likely = PROB_LIKELY( 0.999);
2207 float unlikely = PROB_UNLIKELY(0.999);
2208
2209 IdealKit ideal(this);
2210 #define __ ideal.
2211
2212 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2213
2214 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2215 // Update graphKit memory and control from IdealKit.
2216 sync_kit(ideal);
2217
2218 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2219 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2220
2221 // Update IdealKit memory and control from graphKit.
2222 __ sync_kit(this);
2223
2224 Node* one = __ ConI(1);
2225 // is_instof == 0 if base_oop == NULL
2226 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2227
2228 // Update graphKit from IdeakKit.
2229 sync_kit(ideal);
2230
2231 // Use the pre-barrier to record the value in the referent field
2232 pre_barrier(false /* do_load */,
2233 __ ctrl(),
2234 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2235 pre_val /* pre_val */,
2236 T_OBJECT);
2237 if (need_mem_bar) {
2238 // Add memory barrier to prevent commoning reads from this field
2239 // across safepoint since GC can change its value.
2240 insert_mem_bar(Op_MemBarCPUOrder);
2241 }
2242 // Update IdealKit from graphKit.
2243 __ sync_kit(this);
2244
2245 } __ end_if(); // _ref_type != ref_none
2246 } __ end_if(); // offset == referent_offset
2247
2248 // Final sync IdealKit and GraphKit.
2249 final_sync(ideal);
2250 #undef __
2251 }
2252
2253
2254 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2255 // Attempt to infer a sharper value type from the offset and base type.
2256 ciKlass* sharpened_klass = NULL;
2257
2258 // See if it is an instance field, with an object type.
2259 if (alias_type->field() != NULL) {
2260 if (alias_type->field()->type()->is_klass()) {
2261 sharpened_klass = alias_type->field()->type()->as_klass();
2262 }
2263 }
2264
2265 // See if it is a narrow oop array.
2266 if (adr_type->isa_aryptr()) {
2267 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2268 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2269 if (elem_type != NULL) {
2270 sharpened_klass = elem_type->klass();
2271 }
2272 }
2273 }
2274
2275 // The sharpened class might be unloaded if there is no class loader
2276 // contraint in place.
2277 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2278 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2279
2280 #ifndef PRODUCT
2281 if (C->print_intrinsics() || C->print_inlining()) {
2282 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2283 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2284 }
2285 #endif
2286 // Sharpen the value type.
2287 return tjp;
2288 }
2289 return NULL;
2290 }
2291
2292 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2293 if (callee()->is_static()) return false; // caller must have the capability!
2294 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2295 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2296 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2297
2298 #ifndef PRODUCT
2299 {
2300 ResourceMark rm;
2301 // Check the signatures.
2302 ciSignature* sig = callee()->signature();
2303 #ifdef ASSERT
2304 if (!is_store) {
2305 // Object getObject(Object base, int/long offset), etc.
2306 BasicType rtype = sig->return_type()->basic_type();
2307 assert(rtype == type, "getter must return the expected value");
2308 assert(sig->count() == 2, "oop getter has 2 arguments");
2309 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2310 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2311 } else {
2312 // void putObject(Object base, int/long offset, Object x), etc.
2313 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2314 assert(sig->count() == 3, "oop putter has 3 arguments");
2315 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2316 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2317 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2318 assert(vtype == type, "putter must accept the expected value");
2319 }
2320 #endif // ASSERT
2321 }
2322 #endif //PRODUCT
2323
2324 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2325
2326 Node* receiver = argument(0); // type: oop
2327
2328 // Build address expression.
2329 Node* adr;
2330 Node* heap_base_oop = top();
2331 Node* offset = top();
2332 Node* val;
2333
2334 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2335 Node* base = argument(1); // type: oop
2336 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2337 offset = argument(2); // type: long
2338 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2339 // to be plain byte offsets, which are also the same as those accepted
2340 // by oopDesc::field_base.
2341 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2342 "fieldOffset must be byte-scaled");
2343 // 32-bit machines ignore the high half!
2344 offset = ConvL2X(offset);
2345 adr = make_unsafe_address(base, offset);
2346 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2347 heap_base_oop = base;
2348 } else if (type == T_OBJECT) {
2349 return false; // off-heap oop accesses are not supported
2350 }
2351
2352 // Can base be NULL? Otherwise, always on-heap access.
2353 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2354
2355 val = is_store ? argument(4) : NULL;
2356
2357 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2358
2359 // Try to categorize the address.
2360 Compile::AliasType* alias_type = C->alias_type(adr_type);
2361 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2362
2363 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2364 alias_type->adr_type() == TypeAryPtr::RANGE) {
2365 return false; // not supported
2366 }
2367
2368 bool mismatched = false;
2369 BasicType bt = alias_type->basic_type();
2370 if (bt != T_ILLEGAL) {
2371 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2372 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2373 // Alias type doesn't differentiate between byte[] and boolean[]).
2374 // Use address type to get the element type.
2375 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2376 }
2377 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2378 // accessing an array field with getObject is not a mismatch
2379 bt = T_OBJECT;
2380 }
2381 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2382 // Don't intrinsify mismatched object accesses
2383 return false;
2384 }
2385 mismatched = (bt != type);
2386 } else if (alias_type->adr_type()->isa_oopptr()) {
2387 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2388 }
2389
2390 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2391
2392 // First guess at the value type.
2393 const Type *value_type = Type::get_const_basic_type(type);
2394
2395 // We will need memory barriers unless we can determine a unique
2396 // alias category for this reference. (Note: If for some reason
2397 // the barriers get omitted and the unsafe reference begins to "pollute"
2398 // the alias analysis of the rest of the graph, either Compile::can_alias
2399 // or Compile::must_alias will throw a diagnostic assert.)
2400 bool need_mem_bar = false;
2401 switch (kind) {
2402 case Relaxed:
2403 need_mem_bar = mismatched && !adr_type->isa_aryptr();
2404 break;
2405 case Opaque:
2406 // Opaque uses CPUOrder membars for protection against code movement.
2407 case Acquire:
2408 case Release:
2409 case Volatile:
2410 need_mem_bar = true;
2411 break;
2412 default:
2413 ShouldNotReachHere();
2414 }
2415
2416 // Some accesses require access atomicity for all types, notably longs and doubles.
2417 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2418 bool requires_atomic_access = false;
2419 switch (kind) {
2420 case Relaxed:
2421 requires_atomic_access = AlwaysAtomicAccesses;
2422 break;
2423 case Opaque:
2424 // Opaque accesses are atomic.
2425 case Acquire:
2426 case Release:
2427 case Volatile:
2428 requires_atomic_access = true;
2429 break;
2430 default:
2431 ShouldNotReachHere();
2432 }
2433
2434 // Figure out the memory ordering.
2435 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2436 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2437
2438 // If we are reading the value of the referent field of a Reference
2439 // object (either by using Unsafe directly or through reflection)
2440 // then, if G1 is enabled, we need to record the referent in an
2441 // SATB log buffer using the pre-barrier mechanism.
2442 // Also we need to add memory barrier to prevent commoning reads
2443 // from this field across safepoint since GC can change its value.
2444 bool need_read_barrier = !is_store &&
2445 offset != top() && heap_base_oop != top();
2446
2447 if (!is_store && type == T_OBJECT) {
2448 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2449 if (tjp != NULL) {
2450 value_type = tjp;
2451 }
2452 }
2453
2454 receiver = null_check(receiver);
2455 if (stopped()) {
2456 return true;
2457 }
2458 // Heap pointers get a null-check from the interpreter,
2459 // as a courtesy. However, this is not guaranteed by Unsafe,
2460 // and it is not possible to fully distinguish unintended nulls
2461 // from intended ones in this API.
2462
2463 // We need to emit leading and trailing CPU membars (see below) in
2464 // addition to memory membars for special access modes. This is a little
2465 // too strong, but avoids the need to insert per-alias-type
2466 // volatile membars (for stores; compare Parse::do_put_xxx), which
2467 // we cannot do effectively here because we probably only have a
2468 // rough approximation of type.
2469
2470 switch(kind) {
2471 case Relaxed:
2472 case Opaque:
2473 case Acquire:
2474 break;
2475 case Release:
2476 case Volatile:
2477 if (is_store) {
2478 insert_mem_bar(Op_MemBarRelease);
2479 } else {
2480 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2481 insert_mem_bar(Op_MemBarVolatile);
2482 }
2483 }
2484 break;
2485 default:
2486 ShouldNotReachHere();
2487 }
2488
2489 // Memory barrier to prevent normal and 'unsafe' accesses from
2490 // bypassing each other. Happens after null checks, so the
2491 // exception paths do not take memory state from the memory barrier,
2492 // so there's no problems making a strong assert about mixing users
2493 // of safe & unsafe memory.
2494 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2495
2496 if (!is_store) {
2497 Node* p = NULL;
2498 // Try to constant fold a load from a constant field
2499 ciField* field = alias_type->field();
2500 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2501 // final or stable field
2502 p = make_constant_from_field(field, heap_base_oop);
2503 }
2504 if (p == NULL) {
2505 // To be valid, unsafe loads may depend on other conditions than
2506 // the one that guards them: pin the Load node
2507 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched);
2508 // load value
2509 switch (type) {
2510 case T_BOOLEAN:
2511 {
2512 // Normalize the value returned by getBoolean in the following cases
2513 if (mismatched ||
2514 heap_base_oop == top() || // - heap_base_oop is NULL or
2515 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2516 // and the unsafe access is made to large offset
2517 // (i.e., larger than the maximum offset necessary for any
2518 // field access)
2519 ) {
2520 IdealKit ideal = IdealKit(this);
2521 #define __ ideal.
2522 IdealVariable normalized_result(ideal);
2523 __ declarations_done();
2524 __ set(normalized_result, p);
2525 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2526 __ set(normalized_result, ideal.ConI(1));
2527 ideal.end_if();
2528 final_sync(ideal);
2529 p = __ value(normalized_result);
2530 #undef __
2531 }
2532 }
2533 case T_CHAR:
2534 case T_BYTE:
2535 case T_SHORT:
2536 case T_INT:
2537 case T_LONG:
2538 case T_FLOAT:
2539 case T_DOUBLE:
2540 break;
2541 case T_OBJECT:
2542 if (need_read_barrier) {
2543 // We do not require a mem bar inside pre_barrier if need_mem_bar
2544 // is set: the barriers would be emitted by us.
2545 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2546 }
2547 break;
2548 case T_ADDRESS:
2549 // Cast to an int type.
2550 p = _gvn.transform(new CastP2XNode(NULL, p));
2551 p = ConvX2UL(p);
2552 break;
2553 default:
2554 fatal("unexpected type %d: %s", type, type2name(type));
2555 break;
2556 }
2557 }
2558 // The load node has the control of the preceding MemBarCPUOrder. All
2559 // following nodes will have the control of the MemBarCPUOrder inserted at
2560 // the end of this method. So, pushing the load onto the stack at a later
2561 // point is fine.
2562 set_result(p);
2563 } else {
2564 // place effect of store into memory
2565 switch (type) {
2566 case T_DOUBLE:
2567 val = dstore_rounding(val);
2568 break;
2569 case T_ADDRESS:
2570 // Repackage the long as a pointer.
2571 val = ConvL2X(val);
2572 val = _gvn.transform(new CastX2PNode(val));
2573 break;
2574 }
2575
2576 if (type == T_OBJECT) {
2577 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2578 } else {
2579 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2580 }
2581 }
2582
2583 switch(kind) {
2584 case Relaxed:
2585 case Opaque:
2586 case Release:
2587 break;
2588 case Acquire:
2589 case Volatile:
2590 if (!is_store) {
2591 insert_mem_bar(Op_MemBarAcquire);
2592 } else {
2593 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2594 insert_mem_bar(Op_MemBarVolatile);
2595 }
2596 }
2597 break;
2598 default:
2599 ShouldNotReachHere();
2600 }
2601
2602 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2603
2604 return true;
2605 }
2606
2607 //----------------------------inline_unsafe_load_store----------------------------
2608 // This method serves a couple of different customers (depending on LoadStoreKind):
2609 //
2610 // LS_cmp_swap:
2611 //
2612 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2613 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2614 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2615 //
2616 // LS_cmp_swap_weak:
2617 //
2618 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2619 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2620 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2621 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2622 //
2623 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2624 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2625 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2626 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2627 //
2628 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2629 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2630 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2631 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2632 //
2633 // LS_cmp_exchange:
2634 //
2635 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2636 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2637 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2638 //
2639 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2640 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2641 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2642 //
2643 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2644 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2645 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2646 //
2647 // LS_get_add:
2648 //
2649 // int getAndAddInt( Object o, long offset, int delta)
2650 // long getAndAddLong(Object o, long offset, long delta)
2651 //
2652 // LS_get_set:
2653 //
2654 // int getAndSet(Object o, long offset, int newValue)
2655 // long getAndSet(Object o, long offset, long newValue)
2656 // Object getAndSet(Object o, long offset, Object newValue)
2657 //
2658 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2659 // This basic scheme here is the same as inline_unsafe_access, but
2660 // differs in enough details that combining them would make the code
2661 // overly confusing. (This is a true fact! I originally combined
2662 // them, but even I was confused by it!) As much code/comments as
2663 // possible are retained from inline_unsafe_access though to make
2664 // the correspondences clearer. - dl
2665
2666 if (callee()->is_static()) return false; // caller must have the capability!
2667
2668 #ifndef PRODUCT
2669 BasicType rtype;
2670 {
2671 ResourceMark rm;
2672 // Check the signatures.
2673 ciSignature* sig = callee()->signature();
2674 rtype = sig->return_type()->basic_type();
2675 switch(kind) {
2676 case LS_get_add:
2677 case LS_get_set: {
2678 // Check the signatures.
2679 #ifdef ASSERT
2680 assert(rtype == type, "get and set must return the expected type");
2681 assert(sig->count() == 3, "get and set has 3 arguments");
2682 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2683 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2684 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2685 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2686 #endif // ASSERT
2687 break;
2688 }
2689 case LS_cmp_swap:
2690 case LS_cmp_swap_weak: {
2691 // Check the signatures.
2692 #ifdef ASSERT
2693 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2694 assert(sig->count() == 4, "CAS has 4 arguments");
2695 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2696 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2697 #endif // ASSERT
2698 break;
2699 }
2700 case LS_cmp_exchange: {
2701 // Check the signatures.
2702 #ifdef ASSERT
2703 assert(rtype == type, "CAS must return the expected type");
2704 assert(sig->count() == 4, "CAS has 4 arguments");
2705 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2706 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2707 #endif // ASSERT
2708 break;
2709 }
2710 default:
2711 ShouldNotReachHere();
2712 }
2713 }
2714 #endif //PRODUCT
2715
2716 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2717
2718 // Get arguments:
2719 Node* receiver = NULL;
2720 Node* base = NULL;
2721 Node* offset = NULL;
2722 Node* oldval = NULL;
2723 Node* newval = NULL;
2724 switch(kind) {
2725 case LS_cmp_swap:
2726 case LS_cmp_swap_weak:
2727 case LS_cmp_exchange: {
2728 const bool two_slot_type = type2size[type] == 2;
2729 receiver = argument(0); // type: oop
2730 base = argument(1); // type: oop
2731 offset = argument(2); // type: long
2732 oldval = argument(4); // type: oop, int, or long
2733 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2734 break;
2735 }
2736 case LS_get_add:
2737 case LS_get_set: {
2738 receiver = argument(0); // type: oop
2739 base = argument(1); // type: oop
2740 offset = argument(2); // type: long
2741 oldval = NULL;
2742 newval = argument(4); // type: oop, int, or long
2743 break;
2744 }
2745 default:
2746 ShouldNotReachHere();
2747 }
2748
2749 // Build field offset expression.
2750 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2751 // to be plain byte offsets, which are also the same as those accepted
2752 // by oopDesc::field_base.
2753 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2754 // 32-bit machines ignore the high half of long offsets
2755 offset = ConvL2X(offset);
2756 Node* adr = make_unsafe_address(base, offset);
2757 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2758
2759 Compile::AliasType* alias_type = C->alias_type(adr_type);
2760 BasicType bt = alias_type->basic_type();
2761 if (bt != T_ILLEGAL &&
2762 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2763 // Don't intrinsify mismatched object accesses.
2764 return false;
2765 }
2766
2767 // For CAS, unlike inline_unsafe_access, there seems no point in
2768 // trying to refine types. Just use the coarse types here.
2769 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2770 const Type *value_type = Type::get_const_basic_type(type);
2771
2772 switch (kind) {
2773 case LS_get_set:
2774 case LS_cmp_exchange: {
2775 if (type == T_OBJECT) {
2776 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2777 if (tjp != NULL) {
2778 value_type = tjp;
2779 }
2780 }
2781 break;
2782 }
2783 case LS_cmp_swap:
2784 case LS_cmp_swap_weak:
2785 case LS_get_add:
2786 break;
2787 default:
2788 ShouldNotReachHere();
2789 }
2790
2791 // Null check receiver.
2792 receiver = null_check(receiver);
2793 if (stopped()) {
2794 return true;
2795 }
2796
2797 int alias_idx = C->get_alias_index(adr_type);
2798
2799 // Memory-model-wise, a LoadStore acts like a little synchronized
2800 // block, so needs barriers on each side. These don't translate
2801 // into actual barriers on most machines, but we still need rest of
2802 // compiler to respect ordering.
2803
2804 switch (access_kind) {
2805 case Relaxed:
2806 case Acquire:
2807 break;
2808 case Release:
2809 insert_mem_bar(Op_MemBarRelease);
2810 break;
2811 case Volatile:
2812 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2813 insert_mem_bar(Op_MemBarVolatile);
2814 } else {
2815 insert_mem_bar(Op_MemBarRelease);
2816 }
2817 break;
2818 default:
2819 ShouldNotReachHere();
2820 }
2821 insert_mem_bar(Op_MemBarCPUOrder);
2822
2823 // Figure out the memory ordering.
2824 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2825
2826 // 4984716: MemBars must be inserted before this
2827 // memory node in order to avoid a false
2828 // dependency which will confuse the scheduler.
2829 Node *mem = memory(alias_idx);
2830
2831 // For now, we handle only those cases that actually exist: ints,
2832 // longs, and Object. Adding others should be straightforward.
2833 Node* load_store = NULL;
2834 switch(type) {
2835 case T_BYTE:
2836 switch(kind) {
2837 case LS_get_add:
2838 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2839 break;
2840 case LS_get_set:
2841 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2842 break;
2843 case LS_cmp_swap_weak:
2844 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2845 break;
2846 case LS_cmp_swap:
2847 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2848 break;
2849 case LS_cmp_exchange:
2850 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2851 break;
2852 default:
2853 ShouldNotReachHere();
2854 }
2855 break;
2856 case T_SHORT:
2857 switch(kind) {
2858 case LS_get_add:
2859 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2860 break;
2861 case LS_get_set:
2862 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2863 break;
2864 case LS_cmp_swap_weak:
2865 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2866 break;
2867 case LS_cmp_swap:
2868 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2869 break;
2870 case LS_cmp_exchange:
2871 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2872 break;
2873 default:
2874 ShouldNotReachHere();
2875 }
2876 break;
2877 case T_INT:
2878 switch(kind) {
2879 case LS_get_add:
2880 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2881 break;
2882 case LS_get_set:
2883 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2884 break;
2885 case LS_cmp_swap_weak:
2886 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2887 break;
2888 case LS_cmp_swap:
2889 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2890 break;
2891 case LS_cmp_exchange:
2892 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2893 break;
2894 default:
2895 ShouldNotReachHere();
2896 }
2897 break;
2898 case T_LONG:
2899 switch(kind) {
2900 case LS_get_add:
2901 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2902 break;
2903 case LS_get_set:
2904 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2905 break;
2906 case LS_cmp_swap_weak:
2907 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2908 break;
2909 case LS_cmp_swap:
2910 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2911 break;
2912 case LS_cmp_exchange:
2913 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2914 break;
2915 default:
2916 ShouldNotReachHere();
2917 }
2918 break;
2919 case T_OBJECT:
2920 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2921 // could be delayed during Parse (for example, in adjust_map_after_if()).
2922 // Execute transformation here to avoid barrier generation in such case.
2923 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2924 newval = _gvn.makecon(TypePtr::NULL_PTR);
2925
2926 // Reference stores need a store barrier.
2927 switch(kind) {
2928 case LS_get_set: {
2929 // If pre-barrier must execute before the oop store, old value will require do_load here.
2930 if (!can_move_pre_barrier()) {
2931 pre_barrier(true /* do_load*/,
2932 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2933 NULL /* pre_val*/,
2934 T_OBJECT);
2935 } // Else move pre_barrier to use load_store value, see below.
2936 break;
2937 }
2938 case LS_cmp_swap_weak:
2939 case LS_cmp_swap:
2940 case LS_cmp_exchange: {
2941 // Same as for newval above:
2942 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2943 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2944 }
2945 // The only known value which might get overwritten is oldval.
2946 pre_barrier(false /* do_load */,
2947 control(), NULL, NULL, max_juint, NULL, NULL,
2948 oldval /* pre_val */,
2949 T_OBJECT);
2950 break;
2951 }
2952 default:
2953 ShouldNotReachHere();
2954 }
2955
2956 #ifdef _LP64
2957 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2958 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2959
2960 switch(kind) {
2961 case LS_get_set:
2962 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
2963 break;
2964 case LS_cmp_swap_weak: {
2965 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2966 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2967 break;
2968 }
2969 case LS_cmp_swap: {
2970 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2971 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2972 break;
2973 }
2974 case LS_cmp_exchange: {
2975 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2976 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
2977 break;
2978 }
2979 default:
2980 ShouldNotReachHere();
2981 }
2982 } else
2983 #endif
2984 switch (kind) {
2985 case LS_get_set:
2986 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2987 break;
2988 case LS_cmp_swap_weak:
2989 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2990 break;
2991 case LS_cmp_swap:
2992 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2993 break;
2994 case LS_cmp_exchange:
2995 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
2996 break;
2997 default:
2998 ShouldNotReachHere();
2999 }
3000
3001 // Emit the post barrier only when the actual store happened. This makes sense
3002 // to check only for LS_cmp_* that can fail to set the value.
3003 // LS_cmp_exchange does not produce any branches by default, so there is no
3004 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3005 // CompareAndExchange and move branches here, it would make sense to conditionalize
3006 // post_barriers for LS_cmp_exchange as well.
3007 //
3008 // CAS success path is marked more likely since we anticipate this is a performance
3009 // critical path, while CAS failure path can use the penalty for going through unlikely
3010 // path as backoff. Which is still better than doing a store barrier there.
3011 switch (kind) {
3012 case LS_get_set:
3013 case LS_cmp_exchange: {
3014 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3015 break;
3016 }
3017 case LS_cmp_swap_weak:
3018 case LS_cmp_swap: {
3019 IdealKit ideal(this);
3020 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3021 sync_kit(ideal);
3022 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3023 ideal.sync_kit(this);
3024 } ideal.end_if();
3025 final_sync(ideal);
3026 break;
3027 }
3028 default:
3029 ShouldNotReachHere();
3030 }
3031 break;
3032 default:
3033 fatal("unexpected type %d: %s", type, type2name(type));
3034 break;
3035 }
3036
3037 // SCMemProjNodes represent the memory state of a LoadStore. Their
3038 // main role is to prevent LoadStore nodes from being optimized away
3039 // when their results aren't used.
3040 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3041 set_memory(proj, alias_idx);
3042
3043 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3044 #ifdef _LP64
3045 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3046 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3047 }
3048 #endif
3049 if (can_move_pre_barrier() && kind == LS_get_set) {
3050 // Don't need to load pre_val. The old value is returned by load_store.
3051 // The pre_barrier can execute after the xchg as long as no safepoint
3052 // gets inserted between them.
3053 pre_barrier(false /* do_load */,
3054 control(), NULL, NULL, max_juint, NULL, NULL,
3055 load_store /* pre_val */,
3056 T_OBJECT);
3057 }
3058 }
3059
3060 // Add the trailing membar surrounding the access
3061 insert_mem_bar(Op_MemBarCPUOrder);
3062
3063 switch (access_kind) {
3064 case Relaxed:
3065 case Release:
3066 break; // do nothing
3067 case Acquire:
3068 case Volatile:
3069 insert_mem_bar(Op_MemBarAcquire);
3070 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3071 break;
3072 default:
3073 ShouldNotReachHere();
3074 }
3075
3076 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3077 set_result(load_store);
3078 return true;
3079 }
3080
3081 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3082 MemNode::MemOrd mo = MemNode::unset;
3083 switch(kind) {
3084 case Opaque:
3085 case Relaxed: mo = MemNode::unordered; break;
3086 case Acquire: mo = MemNode::acquire; break;
3087 case Release: mo = MemNode::release; break;
3088 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3089 default:
3090 ShouldNotReachHere();
3091 }
3092 guarantee(mo != MemNode::unset, "Should select memory ordering");
3093 return mo;
3094 }
3095
3096 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3097 MemNode::MemOrd mo = MemNode::unset;
3098 switch(kind) {
3099 case Opaque:
3100 case Relaxed: mo = MemNode::unordered; break;
3101 case Acquire: mo = MemNode::acquire; break;
3102 case Release: mo = MemNode::release; break;
3103 case Volatile: mo = MemNode::seqcst; break;
3104 default:
3105 ShouldNotReachHere();
3106 }
3107 guarantee(mo != MemNode::unset, "Should select memory ordering");
3108 return mo;
3109 }
3110
3111 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3112 // Regardless of form, don't allow previous ld/st to move down,
3113 // then issue acquire, release, or volatile mem_bar.
3114 insert_mem_bar(Op_MemBarCPUOrder);
3115 switch(id) {
3116 case vmIntrinsics::_loadFence:
3117 insert_mem_bar(Op_LoadFence);
3118 return true;
3119 case vmIntrinsics::_storeFence:
3120 insert_mem_bar(Op_StoreFence);
3121 return true;
3122 case vmIntrinsics::_fullFence:
3123 insert_mem_bar(Op_MemBarVolatile);
3124 return true;
3125 default:
3126 fatal_unexpected_iid(id);
3127 return false;
3128 }
3129 }
3130
3131 bool LibraryCallKit::inline_onspinwait() {
3132 insert_mem_bar(Op_OnSpinWait);
3133 return true;
3134 }
3135
3136 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3137 if (!kls->is_Con()) {
3138 return true;
3139 }
3140 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3141 if (klsptr == NULL) {
3142 return true;
3143 }
3144 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3145 // don't need a guard for a klass that is already initialized
3146 return !ik->is_initialized();
3147 }
3148
3149 //----------------------------inline_unsafe_allocate---------------------------
3150 // public native Object Unsafe.allocateInstance(Class<?> cls);
3151 bool LibraryCallKit::inline_unsafe_allocate() {
3152 if (callee()->is_static()) return false; // caller must have the capability!
3153
3154 null_check_receiver(); // null-check, then ignore
3155 Node* cls = null_check(argument(1));
3156 if (stopped()) return true;
3157
3158 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3159 kls = null_check(kls);
3160 if (stopped()) return true; // argument was like int.class
3161
3162 Node* test = NULL;
3163 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3164 // Note: The argument might still be an illegal value like
3165 // Serializable.class or Object[].class. The runtime will handle it.
3166 // But we must make an explicit check for initialization.
3167 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3168 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3169 // can generate code to load it as unsigned byte.
3170 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3171 Node* bits = intcon(InstanceKlass::fully_initialized);
3172 test = _gvn.transform(new SubINode(inst, bits));
3173 // The 'test' is non-zero if we need to take a slow path.
3174 }
3175
3176 Node* obj = new_instance(kls, test);
3177 set_result(obj);
3178 return true;
3179 }
3180
3181 //------------------------inline_native_time_funcs--------------
3182 // inline code for System.currentTimeMillis() and System.nanoTime()
3183 // these have the same type and signature
3184 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3185 const TypeFunc* tf = OptoRuntime::void_long_Type();
3186 const TypePtr* no_memory_effects = NULL;
3187 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3188 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3189 #ifdef ASSERT
3190 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3191 assert(value_top == top(), "second value must be top");
3192 #endif
3193 set_result(value);
3194 return true;
3195 }
3196
3197 #ifdef TRACE_HAVE_INTRINSICS
3198
3199 /*
3200 * oop -> myklass
3201 * myklass->trace_id |= USED
3202 * return myklass->trace_id & ~0x3
3203 */
3204 bool LibraryCallKit::inline_native_classID() {
3205 Node* cls = null_check(argument(0), T_OBJECT);
3206 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3207 kls = null_check(kls, T_OBJECT);
3208
3209 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3210 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3211 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3212
3213 Node* clsused = longcon(0x01l); // set the class bit
3214 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3215 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3216 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3217
3218 #ifdef TRACE_ID_META_BITS
3219 Node* mbits = longcon(~TRACE_ID_META_BITS);
3220 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3221 #endif
3222 #ifdef TRACE_ID_CLASS_SHIFT
3223 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3224 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3225 #endif
3226
3227 set_result(tvalue);
3228 return true;
3229
3230 }
3231
3232 bool LibraryCallKit::inline_native_getBufferWriter() {
3233 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3234
3235 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3236 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3237 );
3238
3239 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3240
3241 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3242 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3243
3244 IfNode* iff_jobj_null =
3245 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3246
3247 enum { _normal_path = 1,
3248 _null_path = 2,
3249 PATH_LIMIT };
3250
3251 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3252 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3253
3254 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3255 result_rgn->init_req(_null_path, jobj_is_null);
3256 result_val->init_req(_null_path, null());
3257
3258 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3259 result_rgn->init_req(_normal_path, jobj_is_not_null);
3260
3261 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3262 result_val->init_req(_normal_path, res);
3263
3264 set_result(result_rgn, result_val);
3265
3266 return true;
3267 }
3268
3269 #endif
3270
3271 //------------------------inline_native_currentThread------------------
3272 bool LibraryCallKit::inline_native_currentThread() {
3273 Node* junk = NULL;
3274 set_result(generate_current_thread(junk));
3275 return true;
3276 }
3277
3278 //------------------------inline_native_isInterrupted------------------
3279 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3280 bool LibraryCallKit::inline_native_isInterrupted() {
3281 // Add a fast path to t.isInterrupted(clear_int):
3282 // (t == Thread.current() &&
3283 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3284 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3285 // So, in the common case that the interrupt bit is false,
3286 // we avoid making a call into the VM. Even if the interrupt bit
3287 // is true, if the clear_int argument is false, we avoid the VM call.
3288 // However, if the receiver is not currentThread, we must call the VM,
3289 // because there must be some locking done around the operation.
3290
3291 // We only go to the fast case code if we pass two guards.
3292 // Paths which do not pass are accumulated in the slow_region.
3293
3294 enum {
3295 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3296 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3297 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3298 PATH_LIMIT
3299 };
3300
3301 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3302 // out of the function.
3303 insert_mem_bar(Op_MemBarCPUOrder);
3304
3305 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3306 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3307
3308 RegionNode* slow_region = new RegionNode(1);
3309 record_for_igvn(slow_region);
3310
3311 // (a) Receiving thread must be the current thread.
3312 Node* rec_thr = argument(0);
3313 Node* tls_ptr = NULL;
3314 Node* cur_thr = generate_current_thread(tls_ptr);
3315 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3316 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3317
3318 generate_slow_guard(bol_thr, slow_region);
3319
3320 // (b) Interrupt bit on TLS must be false.
3321 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3322 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3323 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3324
3325 // Set the control input on the field _interrupted read to prevent it floating up.
3326 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3327 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3328 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3329
3330 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3331
3332 // First fast path: if (!TLS._interrupted) return false;
3333 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3334 result_rgn->init_req(no_int_result_path, false_bit);
3335 result_val->init_req(no_int_result_path, intcon(0));
3336
3337 // drop through to next case
3338 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3339
3340 #ifndef _WINDOWS
3341 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3342 Node* clr_arg = argument(1);
3343 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3344 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3345 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3346
3347 // Second fast path: ... else if (!clear_int) return true;
3348 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3349 result_rgn->init_req(no_clear_result_path, false_arg);
3350 result_val->init_req(no_clear_result_path, intcon(1));
3351
3352 // drop through to next case
3353 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3354 #else
3355 // To return true on Windows you must read the _interrupted field
3356 // and check the event state i.e. take the slow path.
3357 #endif // _WINDOWS
3358
3359 // (d) Otherwise, go to the slow path.
3360 slow_region->add_req(control());
3361 set_control( _gvn.transform(slow_region));
3362
3363 if (stopped()) {
3364 // There is no slow path.
3365 result_rgn->init_req(slow_result_path, top());
3366 result_val->init_req(slow_result_path, top());
3367 } else {
3368 // non-virtual because it is a private non-static
3369 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3370
3371 Node* slow_val = set_results_for_java_call(slow_call);
3372 // this->control() comes from set_results_for_java_call
3373
3374 Node* fast_io = slow_call->in(TypeFunc::I_O);
3375 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3376
3377 // These two phis are pre-filled with copies of of the fast IO and Memory
3378 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3379 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3380
3381 result_rgn->init_req(slow_result_path, control());
3382 result_io ->init_req(slow_result_path, i_o());
3383 result_mem->init_req(slow_result_path, reset_memory());
3384 result_val->init_req(slow_result_path, slow_val);
3385
3386 set_all_memory(_gvn.transform(result_mem));
3387 set_i_o( _gvn.transform(result_io));
3388 }
3389
3390 C->set_has_split_ifs(true); // Has chance for split-if optimization
3391 set_result(result_rgn, result_val);
3392 return true;
3393 }
3394
3395 //---------------------------load_mirror_from_klass----------------------------
3396 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3397 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3398 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3399 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3400 }
3401
3402 //-----------------------load_klass_from_mirror_common-------------------------
3403 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3404 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3405 // and branch to the given path on the region.
3406 // If never_see_null, take an uncommon trap on null, so we can optimistically
3407 // compile for the non-null case.
3408 // If the region is NULL, force never_see_null = true.
3409 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3410 bool never_see_null,
3411 RegionNode* region,
3412 int null_path,
3413 int offset) {
3414 if (region == NULL) never_see_null = true;
3415 Node* p = basic_plus_adr(mirror, offset);
3416 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3417 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3418 Node* null_ctl = top();
3419 kls = null_check_oop(kls, &null_ctl, never_see_null);
3420 if (region != NULL) {
3421 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3422 region->init_req(null_path, null_ctl);
3423 } else {
3424 assert(null_ctl == top(), "no loose ends");
3425 }
3426 return kls;
3427 }
3428
3429 //--------------------(inline_native_Class_query helpers)---------------------
3430 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3431 // Fall through if (mods & mask) == bits, take the guard otherwise.
3432 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3433 // Branch around if the given klass has the given modifier bit set.
3434 // Like generate_guard, adds a new path onto the region.
3435 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3436 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3437 Node* mask = intcon(modifier_mask);
3438 Node* bits = intcon(modifier_bits);
3439 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3440 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3441 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3442 return generate_fair_guard(bol, region);
3443 }
3444 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3445 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3446 }
3447
3448 //-------------------------inline_native_Class_query-------------------
3449 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3450 const Type* return_type = TypeInt::BOOL;
3451 Node* prim_return_value = top(); // what happens if it's a primitive class?
3452 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3453 bool expect_prim = false; // most of these guys expect to work on refs
3454
3455 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3456
3457 Node* mirror = argument(0);
3458 Node* obj = top();
3459
3460 switch (id) {
3461 case vmIntrinsics::_isInstance:
3462 // nothing is an instance of a primitive type
3463 prim_return_value = intcon(0);
3464 obj = argument(1);
3465 break;
3466 case vmIntrinsics::_getModifiers:
3467 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3468 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3469 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3470 break;
3471 case vmIntrinsics::_isInterface:
3472 prim_return_value = intcon(0);
3473 break;
3474 case vmIntrinsics::_isArray:
3475 prim_return_value = intcon(0);
3476 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3477 break;
3478 case vmIntrinsics::_isPrimitive:
3479 prim_return_value = intcon(1);
3480 expect_prim = true; // obviously
3481 break;
3482 case vmIntrinsics::_getSuperclass:
3483 prim_return_value = null();
3484 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3485 break;
3486 case vmIntrinsics::_getClassAccessFlags:
3487 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3488 return_type = TypeInt::INT; // not bool! 6297094
3489 break;
3490 default:
3491 fatal_unexpected_iid(id);
3492 break;
3493 }
3494
3495 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3496 if (mirror_con == NULL) return false; // cannot happen?
3497
3498 #ifndef PRODUCT
3499 if (C->print_intrinsics() || C->print_inlining()) {
3500 ciType* k = mirror_con->java_mirror_type();
3501 if (k) {
3502 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3503 k->print_name();
3504 tty->cr();
3505 }
3506 }
3507 #endif
3508
3509 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3510 RegionNode* region = new RegionNode(PATH_LIMIT);
3511 record_for_igvn(region);
3512 PhiNode* phi = new PhiNode(region, return_type);
3513
3514 // The mirror will never be null of Reflection.getClassAccessFlags, however
3515 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3516 // if it is. See bug 4774291.
3517
3518 // For Reflection.getClassAccessFlags(), the null check occurs in
3519 // the wrong place; see inline_unsafe_access(), above, for a similar
3520 // situation.
3521 mirror = null_check(mirror);
3522 // If mirror or obj is dead, only null-path is taken.
3523 if (stopped()) return true;
3524
3525 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3526
3527 // Now load the mirror's klass metaobject, and null-check it.
3528 // Side-effects region with the control path if the klass is null.
3529 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3530 // If kls is null, we have a primitive mirror.
3531 phi->init_req(_prim_path, prim_return_value);
3532 if (stopped()) { set_result(region, phi); return true; }
3533 bool safe_for_replace = (region->in(_prim_path) == top());
3534
3535 Node* p; // handy temp
3536 Node* null_ctl;
3537
3538 // Now that we have the non-null klass, we can perform the real query.
3539 // For constant classes, the query will constant-fold in LoadNode::Value.
3540 Node* query_value = top();
3541 switch (id) {
3542 case vmIntrinsics::_isInstance:
3543 // nothing is an instance of a primitive type
3544 query_value = gen_instanceof(obj, kls, safe_for_replace);
3545 break;
3546
3547 case vmIntrinsics::_getModifiers:
3548 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3549 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3550 break;
3551
3552 case vmIntrinsics::_isInterface:
3553 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3554 if (generate_interface_guard(kls, region) != NULL)
3555 // A guard was added. If the guard is taken, it was an interface.
3556 phi->add_req(intcon(1));
3557 // If we fall through, it's a plain class.
3558 query_value = intcon(0);
3559 break;
3560
3561 case vmIntrinsics::_isArray:
3562 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3563 if (generate_array_guard(kls, region) != NULL)
3564 // A guard was added. If the guard is taken, it was an array.
3565 phi->add_req(intcon(1));
3566 // If we fall through, it's a plain class.
3567 query_value = intcon(0);
3568 break;
3569
3570 case vmIntrinsics::_isPrimitive:
3571 query_value = intcon(0); // "normal" path produces false
3572 break;
3573
3574 case vmIntrinsics::_getSuperclass:
3575 // The rules here are somewhat unfortunate, but we can still do better
3576 // with random logic than with a JNI call.
3577 // Interfaces store null or Object as _super, but must report null.
3578 // Arrays store an intermediate super as _super, but must report Object.
3579 // Other types can report the actual _super.
3580 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3581 if (generate_interface_guard(kls, region) != NULL)
3582 // A guard was added. If the guard is taken, it was an interface.
3583 phi->add_req(null());
3584 if (generate_array_guard(kls, region) != NULL)
3585 // A guard was added. If the guard is taken, it was an array.
3586 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3587 // If we fall through, it's a plain class. Get its _super.
3588 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3589 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3590 null_ctl = top();
3591 kls = null_check_oop(kls, &null_ctl);
3592 if (null_ctl != top()) {
3593 // If the guard is taken, Object.superClass is null (both klass and mirror).
3594 region->add_req(null_ctl);
3595 phi ->add_req(null());
3596 }
3597 if (!stopped()) {
3598 query_value = load_mirror_from_klass(kls);
3599 }
3600 break;
3601
3602 case vmIntrinsics::_getClassAccessFlags:
3603 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3604 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3605 break;
3606
3607 default:
3608 fatal_unexpected_iid(id);
3609 break;
3610 }
3611
3612 // Fall-through is the normal case of a query to a real class.
3613 phi->init_req(1, query_value);
3614 region->init_req(1, control());
3615
3616 C->set_has_split_ifs(true); // Has chance for split-if optimization
3617 set_result(region, phi);
3618 return true;
3619 }
3620
3621 //-------------------------inline_Class_cast-------------------
3622 bool LibraryCallKit::inline_Class_cast() {
3623 Node* mirror = argument(0); // Class
3624 Node* obj = argument(1);
3625 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3626 if (mirror_con == NULL) {
3627 return false; // dead path (mirror->is_top()).
3628 }
3629 if (obj == NULL || obj->is_top()) {
3630 return false; // dead path
3631 }
3632 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3633
3634 // First, see if Class.cast() can be folded statically.
3635 // java_mirror_type() returns non-null for compile-time Class constants.
3636 ciType* tm = mirror_con->java_mirror_type();
3637 if (tm != NULL && tm->is_klass() &&
3638 tp != NULL && tp->klass() != NULL) {
3639 if (!tp->klass()->is_loaded()) {
3640 // Don't use intrinsic when class is not loaded.
3641 return false;
3642 } else {
3643 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3644 if (static_res == Compile::SSC_always_true) {
3645 // isInstance() is true - fold the code.
3646 set_result(obj);
3647 return true;
3648 } else if (static_res == Compile::SSC_always_false) {
3649 // Don't use intrinsic, have to throw ClassCastException.
3650 // If the reference is null, the non-intrinsic bytecode will
3651 // be optimized appropriately.
3652 return false;
3653 }
3654 }
3655 }
3656
3657 // Bailout intrinsic and do normal inlining if exception path is frequent.
3658 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3659 return false;
3660 }
3661
3662 // Generate dynamic checks.
3663 // Class.cast() is java implementation of _checkcast bytecode.
3664 // Do checkcast (Parse::do_checkcast()) optimizations here.
3665
3666 mirror = null_check(mirror);
3667 // If mirror is dead, only null-path is taken.
3668 if (stopped()) {
3669 return true;
3670 }
3671
3672 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3673 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3674 RegionNode* region = new RegionNode(PATH_LIMIT);
3675 record_for_igvn(region);
3676
3677 // Now load the mirror's klass metaobject, and null-check it.
3678 // If kls is null, we have a primitive mirror and
3679 // nothing is an instance of a primitive type.
3680 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3681
3682 Node* res = top();
3683 if (!stopped()) {
3684 Node* bad_type_ctrl = top();
3685 // Do checkcast optimizations.
3686 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3687 region->init_req(_bad_type_path, bad_type_ctrl);
3688 }
3689 if (region->in(_prim_path) != top() ||
3690 region->in(_bad_type_path) != top()) {
3691 // Let Interpreter throw ClassCastException.
3692 PreserveJVMState pjvms(this);
3693 set_control(_gvn.transform(region));
3694 uncommon_trap(Deoptimization::Reason_intrinsic,
3695 Deoptimization::Action_maybe_recompile);
3696 }
3697 if (!stopped()) {
3698 set_result(res);
3699 }
3700 return true;
3701 }
3702
3703
3704 //--------------------------inline_native_subtype_check------------------------
3705 // This intrinsic takes the JNI calls out of the heart of
3706 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3707 bool LibraryCallKit::inline_native_subtype_check() {
3708 // Pull both arguments off the stack.
3709 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3710 args[0] = argument(0);
3711 args[1] = argument(1);
3712 Node* klasses[2]; // corresponding Klasses: superk, subk
3713 klasses[0] = klasses[1] = top();
3714
3715 enum {
3716 // A full decision tree on {superc is prim, subc is prim}:
3717 _prim_0_path = 1, // {P,N} => false
3718 // {P,P} & superc!=subc => false
3719 _prim_same_path, // {P,P} & superc==subc => true
3720 _prim_1_path, // {N,P} => false
3721 _ref_subtype_path, // {N,N} & subtype check wins => true
3722 _both_ref_path, // {N,N} & subtype check loses => false
3723 PATH_LIMIT
3724 };
3725
3726 RegionNode* region = new RegionNode(PATH_LIMIT);
3727 Node* phi = new PhiNode(region, TypeInt::BOOL);
3728 record_for_igvn(region);
3729
3730 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3731 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3732 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3733
3734 // First null-check both mirrors and load each mirror's klass metaobject.
3735 int which_arg;
3736 for (which_arg = 0; which_arg <= 1; which_arg++) {
3737 Node* arg = args[which_arg];
3738 arg = null_check(arg);
3739 if (stopped()) break;
3740 args[which_arg] = arg;
3741
3742 Node* p = basic_plus_adr(arg, class_klass_offset);
3743 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3744 klasses[which_arg] = _gvn.transform(kls);
3745 }
3746
3747 // Having loaded both klasses, test each for null.
3748 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3749 for (which_arg = 0; which_arg <= 1; which_arg++) {
3750 Node* kls = klasses[which_arg];
3751 Node* null_ctl = top();
3752 kls = null_check_oop(kls, &null_ctl, never_see_null);
3753 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3754 region->init_req(prim_path, null_ctl);
3755 if (stopped()) break;
3756 klasses[which_arg] = kls;
3757 }
3758
3759 if (!stopped()) {
3760 // now we have two reference types, in klasses[0..1]
3761 Node* subk = klasses[1]; // the argument to isAssignableFrom
3762 Node* superk = klasses[0]; // the receiver
3763 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3764 // now we have a successful reference subtype check
3765 region->set_req(_ref_subtype_path, control());
3766 }
3767
3768 // If both operands are primitive (both klasses null), then
3769 // we must return true when they are identical primitives.
3770 // It is convenient to test this after the first null klass check.
3771 set_control(region->in(_prim_0_path)); // go back to first null check
3772 if (!stopped()) {
3773 // Since superc is primitive, make a guard for the superc==subc case.
3774 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3775 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3776 generate_guard(bol_eq, region, PROB_FAIR);
3777 if (region->req() == PATH_LIMIT+1) {
3778 // A guard was added. If the added guard is taken, superc==subc.
3779 region->swap_edges(PATH_LIMIT, _prim_same_path);
3780 region->del_req(PATH_LIMIT);
3781 }
3782 region->set_req(_prim_0_path, control()); // Not equal after all.
3783 }
3784
3785 // these are the only paths that produce 'true':
3786 phi->set_req(_prim_same_path, intcon(1));
3787 phi->set_req(_ref_subtype_path, intcon(1));
3788
3789 // pull together the cases:
3790 assert(region->req() == PATH_LIMIT, "sane region");
3791 for (uint i = 1; i < region->req(); i++) {
3792 Node* ctl = region->in(i);
3793 if (ctl == NULL || ctl == top()) {
3794 region->set_req(i, top());
3795 phi ->set_req(i, top());
3796 } else if (phi->in(i) == NULL) {
3797 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3798 }
3799 }
3800
3801 set_control(_gvn.transform(region));
3802 set_result(_gvn.transform(phi));
3803 return true;
3804 }
3805
3806 //---------------------generate_array_guard_common------------------------
3807 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3808 bool obj_array, bool not_array) {
3809
3810 if (stopped()) {
3811 return NULL;
3812 }
3813
3814 // If obj_array/non_array==false/false:
3815 // Branch around if the given klass is in fact an array (either obj or prim).
3816 // If obj_array/non_array==false/true:
3817 // Branch around if the given klass is not an array klass of any kind.
3818 // If obj_array/non_array==true/true:
3819 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3820 // If obj_array/non_array==true/false:
3821 // Branch around if the kls is an oop array (Object[] or subtype)
3822 //
3823 // Like generate_guard, adds a new path onto the region.
3824 jint layout_con = 0;
3825 Node* layout_val = get_layout_helper(kls, layout_con);
3826 if (layout_val == NULL) {
3827 bool query = (obj_array
3828 ? Klass::layout_helper_is_objArray(layout_con)
3829 : Klass::layout_helper_is_array(layout_con));
3830 if (query == not_array) {
3831 return NULL; // never a branch
3832 } else { // always a branch
3833 Node* always_branch = control();
3834 if (region != NULL)
3835 region->add_req(always_branch);
3836 set_control(top());
3837 return always_branch;
3838 }
3839 }
3840 // Now test the correct condition.
3841 jint nval = (obj_array
3842 ? (jint)(Klass::_lh_array_tag_type_value
3843 << Klass::_lh_array_tag_shift)
3844 : Klass::_lh_neutral_value);
3845 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3846 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3847 // invert the test if we are looking for a non-array
3848 if (not_array) btest = BoolTest(btest).negate();
3849 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3850 return generate_fair_guard(bol, region);
3851 }
3852
3853
3854 //-----------------------inline_native_newArray--------------------------
3855 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3856 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3857 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3858 Node* mirror;
3859 Node* count_val;
3860 if (uninitialized) {
3861 mirror = argument(1);
3862 count_val = argument(2);
3863 } else {
3864 mirror = argument(0);
3865 count_val = argument(1);
3866 }
3867
3868 mirror = null_check(mirror);
3869 // If mirror or obj is dead, only null-path is taken.
3870 if (stopped()) return true;
3871
3872 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3873 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3874 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3875 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3876 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3877
3878 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3879 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3880 result_reg, _slow_path);
3881 Node* normal_ctl = control();
3882 Node* no_array_ctl = result_reg->in(_slow_path);
3883
3884 // Generate code for the slow case. We make a call to newArray().
3885 set_control(no_array_ctl);
3886 if (!stopped()) {
3887 // Either the input type is void.class, or else the
3888 // array klass has not yet been cached. Either the
3889 // ensuing call will throw an exception, or else it
3890 // will cache the array klass for next time.
3891 PreserveJVMState pjvms(this);
3892 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3893 Node* slow_result = set_results_for_java_call(slow_call);
3894 // this->control() comes from set_results_for_java_call
3895 result_reg->set_req(_slow_path, control());
3896 result_val->set_req(_slow_path, slow_result);
3897 result_io ->set_req(_slow_path, i_o());
3898 result_mem->set_req(_slow_path, reset_memory());
3899 }
3900
3901 set_control(normal_ctl);
3902 if (!stopped()) {
3903 // Normal case: The array type has been cached in the java.lang.Class.
3904 // The following call works fine even if the array type is polymorphic.
3905 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3906 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3907 result_reg->init_req(_normal_path, control());
3908 result_val->init_req(_normal_path, obj);
3909 result_io ->init_req(_normal_path, i_o());
3910 result_mem->init_req(_normal_path, reset_memory());
3911
3912 if (uninitialized) {
3913 // Mark the allocation so that zeroing is skipped
3914 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3915 alloc->maybe_set_complete(&_gvn);
3916 }
3917 }
3918
3919 // Return the combined state.
3920 set_i_o( _gvn.transform(result_io) );
3921 set_all_memory( _gvn.transform(result_mem));
3922
3923 C->set_has_split_ifs(true); // Has chance for split-if optimization
3924 set_result(result_reg, result_val);
3925 return true;
3926 }
3927
3928 //----------------------inline_native_getLength--------------------------
3929 // public static native int java.lang.reflect.Array.getLength(Object array);
3930 bool LibraryCallKit::inline_native_getLength() {
3931 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3932
3933 Node* array = null_check(argument(0));
3934 // If array is dead, only null-path is taken.
3935 if (stopped()) return true;
3936
3937 // Deoptimize if it is a non-array.
3938 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3939
3940 if (non_array != NULL) {
3941 PreserveJVMState pjvms(this);
3942 set_control(non_array);
3943 uncommon_trap(Deoptimization::Reason_intrinsic,
3944 Deoptimization::Action_maybe_recompile);
3945 }
3946
3947 // If control is dead, only non-array-path is taken.
3948 if (stopped()) return true;
3949
3950 // The works fine even if the array type is polymorphic.
3951 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3952 Node* result = load_array_length(array);
3953
3954 C->set_has_split_ifs(true); // Has chance for split-if optimization
3955 set_result(result);
3956 return true;
3957 }
3958
3959 //------------------------inline_array_copyOf----------------------------
3960 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3961 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3962 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3963 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3964
3965 // Get the arguments.
3966 Node* original = argument(0);
3967 Node* start = is_copyOfRange? argument(1): intcon(0);
3968 Node* end = is_copyOfRange? argument(2): argument(1);
3969 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3970
3971 Node* newcopy = NULL;
3972
3973 // Set the original stack and the reexecute bit for the interpreter to reexecute
3974 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3975 { PreserveReexecuteState preexecs(this);
3976 jvms()->set_should_reexecute(true);
3977
3978 array_type_mirror = null_check(array_type_mirror);
3979 original = null_check(original);
3980
3981 // Check if a null path was taken unconditionally.
3982 if (stopped()) return true;
3983
3984 Node* orig_length = load_array_length(original);
3985
3986 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3987 klass_node = null_check(klass_node);
3988
3989 RegionNode* bailout = new RegionNode(1);
3990 record_for_igvn(bailout);
3991
3992 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3993 // Bail out if that is so.
3994 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3995 if (not_objArray != NULL) {
3996 // Improve the klass node's type from the new optimistic assumption:
3997 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3998 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
3999 Node* cast = new CastPPNode(klass_node, akls);
4000 cast->init_req(0, control());
4001 klass_node = _gvn.transform(cast);
4002 }
4003
4004 // Bail out if either start or end is negative.
4005 generate_negative_guard(start, bailout, &start);
4006 generate_negative_guard(end, bailout, &end);
4007
4008 Node* length = end;
4009 if (_gvn.type(start) != TypeInt::ZERO) {
4010 length = _gvn.transform(new SubINode(end, start));
4011 }
4012
4013 // Bail out if length is negative.
4014 // Without this the new_array would throw
4015 // NegativeArraySizeException but IllegalArgumentException is what
4016 // should be thrown
4017 generate_negative_guard(length, bailout, &length);
4018
4019 if (bailout->req() > 1) {
4020 PreserveJVMState pjvms(this);
4021 set_control(_gvn.transform(bailout));
4022 uncommon_trap(Deoptimization::Reason_intrinsic,
4023 Deoptimization::Action_maybe_recompile);
4024 }
4025
4026 if (!stopped()) {
4027 // How many elements will we copy from the original?
4028 // The answer is MinI(orig_length - start, length).
4029 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4030 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4031
4032 // Generate a direct call to the right arraycopy function(s).
4033 // We know the copy is disjoint but we might not know if the
4034 // oop stores need checking.
4035 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4036 // This will fail a store-check if x contains any non-nulls.
4037
4038 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4039 // loads/stores but it is legal only if we're sure the
4040 // Arrays.copyOf would succeed. So we need all input arguments
4041 // to the copyOf to be validated, including that the copy to the
4042 // new array won't trigger an ArrayStoreException. That subtype
4043 // check can be optimized if we know something on the type of
4044 // the input array from type speculation.
4045 if (_gvn.type(klass_node)->singleton()) {
4046 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4047 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4048
4049 int test = C->static_subtype_check(superk, subk);
4050 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4051 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4052 if (t_original->speculative_type() != NULL) {
4053 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4054 }
4055 }
4056 }
4057
4058 bool validated = false;
4059 // Reason_class_check rather than Reason_intrinsic because we
4060 // want to intrinsify even if this traps.
4061 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4062 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4063 klass_node);
4064
4065 if (not_subtype_ctrl != top()) {
4066 PreserveJVMState pjvms(this);
4067 set_control(not_subtype_ctrl);
4068 uncommon_trap(Deoptimization::Reason_class_check,
4069 Deoptimization::Action_make_not_entrant);
4070 assert(stopped(), "Should be stopped");
4071 }
4072 validated = true;
4073 }
4074
4075 if (!stopped()) {
4076 newcopy = new_array(klass_node, length, 0); // no arguments to push
4077
4078 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4079 load_object_klass(original), klass_node);
4080 if (!is_copyOfRange) {
4081 ac->set_copyof(validated);
4082 } else {
4083 ac->set_copyofrange(validated);
4084 }
4085 Node* n = _gvn.transform(ac);
4086 if (n == ac) {
4087 ac->connect_outputs(this);
4088 } else {
4089 assert(validated, "shouldn't transform if all arguments not validated");
4090 set_all_memory(n);
4091 }
4092 }
4093 }
4094 } // original reexecute is set back here
4095
4096 C->set_has_split_ifs(true); // Has chance for split-if optimization
4097 if (!stopped()) {
4098 set_result(newcopy);
4099 }
4100 return true;
4101 }
4102
4103
4104 //----------------------generate_virtual_guard---------------------------
4105 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4106 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4107 RegionNode* slow_region) {
4108 ciMethod* method = callee();
4109 int vtable_index = method->vtable_index();
4110 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4111 "bad index %d", vtable_index);
4112 // Get the Method* out of the appropriate vtable entry.
4113 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4114 vtable_index*vtableEntry::size_in_bytes() +
4115 vtableEntry::method_offset_in_bytes();
4116 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4117 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4118
4119 // Compare the target method with the expected method (e.g., Object.hashCode).
4120 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4121
4122 Node* native_call = makecon(native_call_addr);
4123 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4124 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4125
4126 return generate_slow_guard(test_native, slow_region);
4127 }
4128
4129 //-----------------------generate_method_call----------------------------
4130 // Use generate_method_call to make a slow-call to the real
4131 // method if the fast path fails. An alternative would be to
4132 // use a stub like OptoRuntime::slow_arraycopy_Java.
4133 // This only works for expanding the current library call,
4134 // not another intrinsic. (E.g., don't use this for making an
4135 // arraycopy call inside of the copyOf intrinsic.)
4136 CallJavaNode*
4137 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4138 // When compiling the intrinsic method itself, do not use this technique.
4139 guarantee(callee() != C->method(), "cannot make slow-call to self");
4140
4141 ciMethod* method = callee();
4142 // ensure the JVMS we have will be correct for this call
4143 guarantee(method_id == method->intrinsic_id(), "must match");
4144
4145 const TypeFunc* tf = TypeFunc::make(method);
4146 CallJavaNode* slow_call;
4147 if (is_static) {
4148 assert(!is_virtual, "");
4149 slow_call = new CallStaticJavaNode(C, tf,
4150 SharedRuntime::get_resolve_static_call_stub(),
4151 method, bci());
4152 } else if (is_virtual) {
4153 null_check_receiver();
4154 int vtable_index = Method::invalid_vtable_index;
4155 if (UseInlineCaches) {
4156 // Suppress the vtable call
4157 } else {
4158 // hashCode and clone are not a miranda methods,
4159 // so the vtable index is fixed.
4160 // No need to use the linkResolver to get it.
4161 vtable_index = method->vtable_index();
4162 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4163 "bad index %d", vtable_index);
4164 }
4165 slow_call = new CallDynamicJavaNode(tf,
4166 SharedRuntime::get_resolve_virtual_call_stub(),
4167 method, vtable_index, bci());
4168 } else { // neither virtual nor static: opt_virtual
4169 null_check_receiver();
4170 slow_call = new CallStaticJavaNode(C, tf,
4171 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4172 method, bci());
4173 slow_call->set_optimized_virtual(true);
4174 }
4175 set_arguments_for_java_call(slow_call);
4176 set_edges_for_java_call(slow_call);
4177 return slow_call;
4178 }
4179
4180
4181 /**
4182 * Build special case code for calls to hashCode on an object. This call may
4183 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4184 * slightly different code.
4185 */
4186 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4187 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4188 assert(!(is_virtual && is_static), "either virtual, special, or static");
4189
4190 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4191
4192 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4193 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4194 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4195 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4196 Node* obj = NULL;
4197 if (!is_static) {
4198 // Check for hashing null object
4199 obj = null_check_receiver();
4200 if (stopped()) return true; // unconditionally null
4201 result_reg->init_req(_null_path, top());
4202 result_val->init_req(_null_path, top());
4203 } else {
4204 // Do a null check, and return zero if null.
4205 // System.identityHashCode(null) == 0
4206 obj = argument(0);
4207 Node* null_ctl = top();
4208 obj = null_check_oop(obj, &null_ctl);
4209 result_reg->init_req(_null_path, null_ctl);
4210 result_val->init_req(_null_path, _gvn.intcon(0));
4211 }
4212
4213 // Unconditionally null? Then return right away.
4214 if (stopped()) {
4215 set_control( result_reg->in(_null_path));
4216 if (!stopped())
4217 set_result(result_val->in(_null_path));
4218 return true;
4219 }
4220
4221 // We only go to the fast case code if we pass a number of guards. The
4222 // paths which do not pass are accumulated in the slow_region.
4223 RegionNode* slow_region = new RegionNode(1);
4224 record_for_igvn(slow_region);
4225
4226 // If this is a virtual call, we generate a funny guard. We pull out
4227 // the vtable entry corresponding to hashCode() from the target object.
4228 // If the target method which we are calling happens to be the native
4229 // Object hashCode() method, we pass the guard. We do not need this
4230 // guard for non-virtual calls -- the caller is known to be the native
4231 // Object hashCode().
4232 if (is_virtual) {
4233 // After null check, get the object's klass.
4234 Node* obj_klass = load_object_klass(obj);
4235 generate_virtual_guard(obj_klass, slow_region);
4236 }
4237
4238 // Get the header out of the object, use LoadMarkNode when available
4239 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4240 // The control of the load must be NULL. Otherwise, the load can move before
4241 // the null check after castPP removal.
4242 Node* no_ctrl = NULL;
4243 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4244
4245 // Test the header to see if it is unlocked.
4246 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4247 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4248 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4249 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4250 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4251
4252 generate_slow_guard(test_unlocked, slow_region);
4253
4254 // Get the hash value and check to see that it has been properly assigned.
4255 // We depend on hash_mask being at most 32 bits and avoid the use of
4256 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4257 // vm: see markOop.hpp.
4258 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4259 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4260 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4261 // This hack lets the hash bits live anywhere in the mark object now, as long
4262 // as the shift drops the relevant bits into the low 32 bits. Note that
4263 // Java spec says that HashCode is an int so there's no point in capturing
4264 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4265 hshifted_header = ConvX2I(hshifted_header);
4266 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4267
4268 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4269 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4270 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4271
4272 generate_slow_guard(test_assigned, slow_region);
4273
4274 Node* init_mem = reset_memory();
4275 // fill in the rest of the null path:
4276 result_io ->init_req(_null_path, i_o());
4277 result_mem->init_req(_null_path, init_mem);
4278
4279 result_val->init_req(_fast_path, hash_val);
4280 result_reg->init_req(_fast_path, control());
4281 result_io ->init_req(_fast_path, i_o());
4282 result_mem->init_req(_fast_path, init_mem);
4283
4284 // Generate code for the slow case. We make a call to hashCode().
4285 set_control(_gvn.transform(slow_region));
4286 if (!stopped()) {
4287 // No need for PreserveJVMState, because we're using up the present state.
4288 set_all_memory(init_mem);
4289 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4290 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4291 Node* slow_result = set_results_for_java_call(slow_call);
4292 // this->control() comes from set_results_for_java_call
4293 result_reg->init_req(_slow_path, control());
4294 result_val->init_req(_slow_path, slow_result);
4295 result_io ->set_req(_slow_path, i_o());
4296 result_mem ->set_req(_slow_path, reset_memory());
4297 }
4298
4299 // Return the combined state.
4300 set_i_o( _gvn.transform(result_io) );
4301 set_all_memory( _gvn.transform(result_mem));
4302
4303 set_result(result_reg, result_val);
4304 return true;
4305 }
4306
4307 //---------------------------inline_native_getClass----------------------------
4308 // public final native Class<?> java.lang.Object.getClass();
4309 //
4310 // Build special case code for calls to getClass on an object.
4311 bool LibraryCallKit::inline_native_getClass() {
4312 Node* obj = null_check_receiver();
4313 if (stopped()) return true;
4314 set_result(load_mirror_from_klass(load_object_klass(obj)));
4315 return true;
4316 }
4317
4318 //-----------------inline_native_Reflection_getCallerClass---------------------
4319 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4320 //
4321 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4322 //
4323 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4324 // in that it must skip particular security frames and checks for
4325 // caller sensitive methods.
4326 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4327 #ifndef PRODUCT
4328 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4329 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4330 }
4331 #endif
4332
4333 if (!jvms()->has_method()) {
4334 #ifndef PRODUCT
4335 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4336 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4337 }
4338 #endif
4339 return false;
4340 }
4341
4342 // Walk back up the JVM state to find the caller at the required
4343 // depth.
4344 JVMState* caller_jvms = jvms();
4345
4346 // Cf. JVM_GetCallerClass
4347 // NOTE: Start the loop at depth 1 because the current JVM state does
4348 // not include the Reflection.getCallerClass() frame.
4349 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4350 ciMethod* m = caller_jvms->method();
4351 switch (n) {
4352 case 0:
4353 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4354 break;
4355 case 1:
4356 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4357 if (!m->caller_sensitive()) {
4358 #ifndef PRODUCT
4359 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4360 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4361 }
4362 #endif
4363 return false; // bail-out; let JVM_GetCallerClass do the work
4364 }
4365 break;
4366 default:
4367 if (!m->is_ignored_by_security_stack_walk()) {
4368 // We have reached the desired frame; return the holder class.
4369 // Acquire method holder as java.lang.Class and push as constant.
4370 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4371 ciInstance* caller_mirror = caller_klass->java_mirror();
4372 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4373
4374 #ifndef PRODUCT
4375 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4376 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());
4377 tty->print_cr(" JVM state at this point:");
4378 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4379 ciMethod* m = jvms()->of_depth(i)->method();
4380 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4381 }
4382 }
4383 #endif
4384 return true;
4385 }
4386 break;
4387 }
4388 }
4389
4390 #ifndef PRODUCT
4391 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4392 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4393 tty->print_cr(" JVM state at this point:");
4394 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4395 ciMethod* m = jvms()->of_depth(i)->method();
4396 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4397 }
4398 }
4399 #endif
4400
4401 return false; // bail-out; let JVM_GetCallerClass do the work
4402 }
4403
4404 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4405 Node* arg = argument(0);
4406 Node* result = NULL;
4407
4408 switch (id) {
4409 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4410 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4411 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4412 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4413
4414 case vmIntrinsics::_doubleToLongBits: {
4415 // two paths (plus control) merge in a wood
4416 RegionNode *r = new RegionNode(3);
4417 Node *phi = new PhiNode(r, TypeLong::LONG);
4418
4419 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4420 // Build the boolean node
4421 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4422
4423 // Branch either way.
4424 // NaN case is less traveled, which makes all the difference.
4425 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4426 Node *opt_isnan = _gvn.transform(ifisnan);
4427 assert( opt_isnan->is_If(), "Expect an IfNode");
4428 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4429 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4430
4431 set_control(iftrue);
4432
4433 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4434 Node *slow_result = longcon(nan_bits); // return NaN
4435 phi->init_req(1, _gvn.transform( slow_result ));
4436 r->init_req(1, iftrue);
4437
4438 // Else fall through
4439 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4440 set_control(iffalse);
4441
4442 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4443 r->init_req(2, iffalse);
4444
4445 // Post merge
4446 set_control(_gvn.transform(r));
4447 record_for_igvn(r);
4448
4449 C->set_has_split_ifs(true); // Has chance for split-if optimization
4450 result = phi;
4451 assert(result->bottom_type()->isa_long(), "must be");
4452 break;
4453 }
4454
4455 case vmIntrinsics::_floatToIntBits: {
4456 // two paths (plus control) merge in a wood
4457 RegionNode *r = new RegionNode(3);
4458 Node *phi = new PhiNode(r, TypeInt::INT);
4459
4460 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4461 // Build the boolean node
4462 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4463
4464 // Branch either way.
4465 // NaN case is less traveled, which makes all the difference.
4466 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4467 Node *opt_isnan = _gvn.transform(ifisnan);
4468 assert( opt_isnan->is_If(), "Expect an IfNode");
4469 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4470 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4471
4472 set_control(iftrue);
4473
4474 static const jint nan_bits = 0x7fc00000;
4475 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4476 phi->init_req(1, _gvn.transform( slow_result ));
4477 r->init_req(1, iftrue);
4478
4479 // Else fall through
4480 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4481 set_control(iffalse);
4482
4483 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4484 r->init_req(2, iffalse);
4485
4486 // Post merge
4487 set_control(_gvn.transform(r));
4488 record_for_igvn(r);
4489
4490 C->set_has_split_ifs(true); // Has chance for split-if optimization
4491 result = phi;
4492 assert(result->bottom_type()->isa_int(), "must be");
4493 break;
4494 }
4495
4496 default:
4497 fatal_unexpected_iid(id);
4498 break;
4499 }
4500 set_result(_gvn.transform(result));
4501 return true;
4502 }
4503
4504 //----------------------inline_unsafe_copyMemory-------------------------
4505 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4506 bool LibraryCallKit::inline_unsafe_copyMemory() {
4507 if (callee()->is_static()) return false; // caller must have the capability!
4508 null_check_receiver(); // null-check receiver
4509 if (stopped()) return true;
4510
4511 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4512
4513 Node* src_ptr = argument(1); // type: oop
4514 Node* src_off = ConvL2X(argument(2)); // type: long
4515 Node* dst_ptr = argument(4); // type: oop
4516 Node* dst_off = ConvL2X(argument(5)); // type: long
4517 Node* size = ConvL2X(argument(7)); // type: long
4518
4519 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4520 "fieldOffset must be byte-scaled");
4521
4522 Node* src = make_unsafe_address(src_ptr, src_off);
4523 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4524
4525 // Conservatively insert a memory barrier on all memory slices.
4526 // Do not let writes of the copy source or destination float below the copy.
4527 insert_mem_bar(Op_MemBarCPUOrder);
4528
4529 // Call it. Note that the length argument is not scaled.
4530 make_runtime_call(RC_LEAF|RC_NO_FP,
4531 OptoRuntime::fast_arraycopy_Type(),
4532 StubRoutines::unsafe_arraycopy(),
4533 "unsafe_arraycopy",
4534 TypeRawPtr::BOTTOM,
4535 src, dst, size XTOP);
4536
4537 // Do not let reads of the copy destination float above the copy.
4538 insert_mem_bar(Op_MemBarCPUOrder);
4539
4540 return true;
4541 }
4542
4543 //------------------------clone_coping-----------------------------------
4544 // Helper function for inline_native_clone.
4545 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4546 assert(obj_size != NULL, "");
4547 Node* raw_obj = alloc_obj->in(1);
4548 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4549
4550 AllocateNode* alloc = NULL;
4551 if (ReduceBulkZeroing) {
4552 // We will be completely responsible for initializing this object -
4553 // mark Initialize node as complete.
4554 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4555 // The object was just allocated - there should be no any stores!
4556 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4557 // Mark as complete_with_arraycopy so that on AllocateNode
4558 // expansion, we know this AllocateNode is initialized by an array
4559 // copy and a StoreStore barrier exists after the array copy.
4560 alloc->initialization()->set_complete_with_arraycopy();
4561 }
4562
4563 // Copy the fastest available way.
4564 // TODO: generate fields copies for small objects instead.
4565 Node* src = obj;
4566 Node* dest = alloc_obj;
4567 Node* size = _gvn.transform(obj_size);
4568
4569 // Exclude the header but include array length to copy by 8 bytes words.
4570 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4571 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4572 instanceOopDesc::base_offset_in_bytes();
4573 // base_off:
4574 // 8 - 32-bit VM
4575 // 12 - 64-bit VM, compressed klass
4576 // 16 - 64-bit VM, normal klass
4577 if (base_off % BytesPerLong != 0) {
4578 assert(UseCompressedClassPointers, "");
4579 if (is_array) {
4580 // Exclude length to copy by 8 bytes words.
4581 base_off += sizeof(int);
4582 } else {
4583 // Include klass to copy by 8 bytes words.
4584 base_off = instanceOopDesc::klass_offset_in_bytes();
4585 }
4586 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4587 }
4588 src = basic_plus_adr(src, base_off);
4589 dest = basic_plus_adr(dest, base_off);
4590
4591 // Compute the length also, if needed:
4592 Node* countx = size;
4593 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4594 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4595
4596 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4597
4598 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4599 ac->set_clonebasic();
4600 Node* n = _gvn.transform(ac);
4601 if (n == ac) {
4602 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4603 } else {
4604 set_all_memory(n);
4605 }
4606
4607 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4608 if (card_mark) {
4609 assert(!is_array, "");
4610 // Put in store barrier for any and all oops we are sticking
4611 // into this object. (We could avoid this if we could prove
4612 // that the object type contains no oop fields at all.)
4613 Node* no_particular_value = NULL;
4614 Node* no_particular_field = NULL;
4615 int raw_adr_idx = Compile::AliasIdxRaw;
4616 post_barrier(control(),
4617 memory(raw_adr_type),
4618 alloc_obj,
4619 no_particular_field,
4620 raw_adr_idx,
4621 no_particular_value,
4622 T_OBJECT,
4623 false);
4624 }
4625
4626 // Do not let reads from the cloned object float above the arraycopy.
4627 if (alloc != NULL) {
4628 // Do not let stores that initialize this object be reordered with
4629 // a subsequent store that would make this object accessible by
4630 // other threads.
4631 // Record what AllocateNode this StoreStore protects so that
4632 // escape analysis can go from the MemBarStoreStoreNode to the
4633 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4634 // based on the escape status of the AllocateNode.
4635 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4636 } else {
4637 insert_mem_bar(Op_MemBarCPUOrder);
4638 }
4639 }
4640
4641 //------------------------inline_native_clone----------------------------
4642 // protected native Object java.lang.Object.clone();
4643 //
4644 // Here are the simple edge cases:
4645 // null receiver => normal trap
4646 // virtual and clone was overridden => slow path to out-of-line clone
4647 // not cloneable or finalizer => slow path to out-of-line Object.clone
4648 //
4649 // The general case has two steps, allocation and copying.
4650 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4651 //
4652 // Copying also has two cases, oop arrays and everything else.
4653 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4654 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4655 //
4656 // These steps fold up nicely if and when the cloned object's klass
4657 // can be sharply typed as an object array, a type array, or an instance.
4658 //
4659 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4660 PhiNode* result_val;
4661
4662 // Set the reexecute bit for the interpreter to reexecute
4663 // the bytecode that invokes Object.clone if deoptimization happens.
4664 { PreserveReexecuteState preexecs(this);
4665 jvms()->set_should_reexecute(true);
4666
4667 Node* obj = null_check_receiver();
4668 if (stopped()) return true;
4669
4670 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4671
4672 // If we are going to clone an instance, we need its exact type to
4673 // know the number and types of fields to convert the clone to
4674 // loads/stores. Maybe a speculative type can help us.
4675 if (!obj_type->klass_is_exact() &&
4676 obj_type->speculative_type() != NULL &&
4677 obj_type->speculative_type()->is_instance_klass()) {
4678 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4679 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4680 !spec_ik->has_injected_fields()) {
4681 ciKlass* k = obj_type->klass();
4682 if (!k->is_instance_klass() ||
4683 k->as_instance_klass()->is_interface() ||
4684 k->as_instance_klass()->has_subklass()) {
4685 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4686 }
4687 }
4688 }
4689
4690 Node* obj_klass = load_object_klass(obj);
4691 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4692 const TypeOopPtr* toop = ((tklass != NULL)
4693 ? tklass->as_instance_type()
4694 : TypeInstPtr::NOTNULL);
4695
4696 // Conservatively insert a memory barrier on all memory slices.
4697 // Do not let writes into the original float below the clone.
4698 insert_mem_bar(Op_MemBarCPUOrder);
4699
4700 // paths into result_reg:
4701 enum {
4702 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4703 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4704 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4705 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4706 PATH_LIMIT
4707 };
4708 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4709 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4710 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4711 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4712 record_for_igvn(result_reg);
4713
4714 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4715 int raw_adr_idx = Compile::AliasIdxRaw;
4716
4717 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4718 if (array_ctl != NULL) {
4719 // It's an array.
4720 PreserveJVMState pjvms(this);
4721 set_control(array_ctl);
4722 Node* obj_length = load_array_length(obj);
4723 Node* obj_size = NULL;
4724 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4725
4726 if (!use_ReduceInitialCardMarks()) {
4727 // If it is an oop array, it requires very special treatment,
4728 // because card marking is required on each card of the array.
4729 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4730 if (is_obja != NULL) {
4731 PreserveJVMState pjvms2(this);
4732 set_control(is_obja);
4733 // Generate a direct call to the right arraycopy function(s).
4734 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4735 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4736 ac->set_cloneoop();
4737 Node* n = _gvn.transform(ac);
4738 assert(n == ac, "cannot disappear");
4739 ac->connect_outputs(this);
4740
4741 result_reg->init_req(_objArray_path, control());
4742 result_val->init_req(_objArray_path, alloc_obj);
4743 result_i_o ->set_req(_objArray_path, i_o());
4744 result_mem ->set_req(_objArray_path, reset_memory());
4745 }
4746 }
4747 // Otherwise, there are no card marks to worry about.
4748 // (We can dispense with card marks if we know the allocation
4749 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4750 // causes the non-eden paths to take compensating steps to
4751 // simulate a fresh allocation, so that no further
4752 // card marks are required in compiled code to initialize
4753 // the object.)
4754
4755 if (!stopped()) {
4756 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4757
4758 // Present the results of the copy.
4759 result_reg->init_req(_array_path, control());
4760 result_val->init_req(_array_path, alloc_obj);
4761 result_i_o ->set_req(_array_path, i_o());
4762 result_mem ->set_req(_array_path, reset_memory());
4763 }
4764 }
4765
4766 // We only go to the instance fast case code if we pass a number of guards.
4767 // The paths which do not pass are accumulated in the slow_region.
4768 RegionNode* slow_region = new RegionNode(1);
4769 record_for_igvn(slow_region);
4770 if (!stopped()) {
4771 // It's an instance (we did array above). Make the slow-path tests.
4772 // If this is a virtual call, we generate a funny guard. We grab
4773 // the vtable entry corresponding to clone() from the target object.
4774 // If the target method which we are calling happens to be the
4775 // Object clone() method, we pass the guard. We do not need this
4776 // guard for non-virtual calls; the caller is known to be the native
4777 // Object clone().
4778 if (is_virtual) {
4779 generate_virtual_guard(obj_klass, slow_region);
4780 }
4781
4782 // The object must be easily cloneable and must not have a finalizer.
4783 // Both of these conditions may be checked in a single test.
4784 // We could optimize the test further, but we don't care.
4785 generate_access_flags_guard(obj_klass,
4786 // Test both conditions:
4787 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4788 // Must be cloneable but not finalizer:
4789 JVM_ACC_IS_CLONEABLE_FAST,
4790 slow_region);
4791 }
4792
4793 if (!stopped()) {
4794 // It's an instance, and it passed the slow-path tests.
4795 PreserveJVMState pjvms(this);
4796 Node* obj_size = NULL;
4797 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4798 // is reexecuted if deoptimization occurs and there could be problems when merging
4799 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4800 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4801
4802 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4803
4804 // Present the results of the slow call.
4805 result_reg->init_req(_instance_path, control());
4806 result_val->init_req(_instance_path, alloc_obj);
4807 result_i_o ->set_req(_instance_path, i_o());
4808 result_mem ->set_req(_instance_path, reset_memory());
4809 }
4810
4811 // Generate code for the slow case. We make a call to clone().
4812 set_control(_gvn.transform(slow_region));
4813 if (!stopped()) {
4814 PreserveJVMState pjvms(this);
4815 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4816 Node* slow_result = set_results_for_java_call(slow_call);
4817 // this->control() comes from set_results_for_java_call
4818 result_reg->init_req(_slow_path, control());
4819 result_val->init_req(_slow_path, slow_result);
4820 result_i_o ->set_req(_slow_path, i_o());
4821 result_mem ->set_req(_slow_path, reset_memory());
4822 }
4823
4824 // Return the combined state.
4825 set_control( _gvn.transform(result_reg));
4826 set_i_o( _gvn.transform(result_i_o));
4827 set_all_memory( _gvn.transform(result_mem));
4828 } // original reexecute is set back here
4829
4830 set_result(_gvn.transform(result_val));
4831 return true;
4832 }
4833
4834 // If we have a tighly coupled allocation, the arraycopy may take care
4835 // of the array initialization. If one of the guards we insert between
4836 // the allocation and the arraycopy causes a deoptimization, an
4837 // unitialized array will escape the compiled method. To prevent that
4838 // we set the JVM state for uncommon traps between the allocation and
4839 // the arraycopy to the state before the allocation so, in case of
4840 // deoptimization, we'll reexecute the allocation and the
4841 // initialization.
4842 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4843 if (alloc != NULL) {
4844 ciMethod* trap_method = alloc->jvms()->method();
4845 int trap_bci = alloc->jvms()->bci();
4846
4847 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4848 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4849 // Make sure there's no store between the allocation and the
4850 // arraycopy otherwise visible side effects could be rexecuted
4851 // in case of deoptimization and cause incorrect execution.
4852 bool no_interfering_store = true;
4853 Node* mem = alloc->in(TypeFunc::Memory);
4854 if (mem->is_MergeMem()) {
4855 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4856 Node* n = mms.memory();
4857 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4858 assert(n->is_Store(), "what else?");
4859 no_interfering_store = false;
4860 break;
4861 }
4862 }
4863 } else {
4864 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4865 Node* n = mms.memory();
4866 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4867 assert(n->is_Store(), "what else?");
4868 no_interfering_store = false;
4869 break;
4870 }
4871 }
4872 }
4873
4874 if (no_interfering_store) {
4875 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4876 uint size = alloc->req();
4877 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4878 old_jvms->set_map(sfpt);
4879 for (uint i = 0; i < size; i++) {
4880 sfpt->init_req(i, alloc->in(i));
4881 }
4882 // re-push array length for deoptimization
4883 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4884 old_jvms->set_sp(old_jvms->sp()+1);
4885 old_jvms->set_monoff(old_jvms->monoff()+1);
4886 old_jvms->set_scloff(old_jvms->scloff()+1);
4887 old_jvms->set_endoff(old_jvms->endoff()+1);
4888 old_jvms->set_should_reexecute(true);
4889
4890 sfpt->set_i_o(map()->i_o());
4891 sfpt->set_memory(map()->memory());
4892 sfpt->set_control(map()->control());
4893
4894 JVMState* saved_jvms = jvms();
4895 saved_reexecute_sp = _reexecute_sp;
4896
4897 set_jvms(sfpt->jvms());
4898 _reexecute_sp = jvms()->sp();
4899
4900 return saved_jvms;
4901 }
4902 }
4903 }
4904 return NULL;
4905 }
4906
4907 // In case of a deoptimization, we restart execution at the
4908 // allocation, allocating a new array. We would leave an uninitialized
4909 // array in the heap that GCs wouldn't expect. Move the allocation
4910 // after the traps so we don't allocate the array if we
4911 // deoptimize. This is possible because tightly_coupled_allocation()
4912 // guarantees there's no observer of the allocated array at this point
4913 // and the control flow is simple enough.
4914 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4915 int saved_reexecute_sp, uint new_idx) {
4916 if (saved_jvms != NULL && !stopped()) {
4917 assert(alloc != NULL, "only with a tightly coupled allocation");
4918 // restore JVM state to the state at the arraycopy
4919 saved_jvms->map()->set_control(map()->control());
4920 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4921 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4922 // If we've improved the types of some nodes (null check) while
4923 // emitting the guards, propagate them to the current state
4924 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4925 set_jvms(saved_jvms);
4926 _reexecute_sp = saved_reexecute_sp;
4927
4928 // Remove the allocation from above the guards
4929 CallProjections callprojs;
4930 alloc->extract_projections(&callprojs, true);
4931 InitializeNode* init = alloc->initialization();
4932 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4933 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4934 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4935 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4936
4937 // move the allocation here (after the guards)
4938 _gvn.hash_delete(alloc);
4939 alloc->set_req(TypeFunc::Control, control());
4940 alloc->set_req(TypeFunc::I_O, i_o());
4941 Node *mem = reset_memory();
4942 set_all_memory(mem);
4943 alloc->set_req(TypeFunc::Memory, mem);
4944 set_control(init->proj_out(TypeFunc::Control));
4945 set_i_o(callprojs.fallthrough_ioproj);
4946
4947 // Update memory as done in GraphKit::set_output_for_allocation()
4948 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4949 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4950 if (ary_type->isa_aryptr() && length_type != NULL) {
4951 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4952 }
4953 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4954 int elemidx = C->get_alias_index(telemref);
4955 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4956 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4957
4958 Node* allocx = _gvn.transform(alloc);
4959 assert(allocx == alloc, "where has the allocation gone?");
4960 assert(dest->is_CheckCastPP(), "not an allocation result?");
4961
4962 _gvn.hash_delete(dest);
4963 dest->set_req(0, control());
4964 Node* destx = _gvn.transform(dest);
4965 assert(destx == dest, "where has the allocation result gone?");
4966 }
4967 }
4968
4969
4970 //------------------------------inline_arraycopy-----------------------
4971 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4972 // Object dest, int destPos,
4973 // int length);
4974 bool LibraryCallKit::inline_arraycopy() {
4975 // Get the arguments.
4976 Node* src = argument(0); // type: oop
4977 Node* src_offset = argument(1); // type: int
4978 Node* dest = argument(2); // type: oop
4979 Node* dest_offset = argument(3); // type: int
4980 Node* length = argument(4); // type: int
4981
4982 uint new_idx = C->unique();
4983
4984 // Check for allocation before we add nodes that would confuse
4985 // tightly_coupled_allocation()
4986 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4987
4988 int saved_reexecute_sp = -1;
4989 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4990 // See arraycopy_restore_alloc_state() comment
4991 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4992 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4993 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4994 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4995
4996 // The following tests must be performed
4997 // (1) src and dest are arrays.
4998 // (2) src and dest arrays must have elements of the same BasicType
4999 // (3) src and dest must not be null.
5000 // (4) src_offset must not be negative.
5001 // (5) dest_offset must not be negative.
5002 // (6) length must not be negative.
5003 // (7) src_offset + length must not exceed length of src.
5004 // (8) dest_offset + length must not exceed length of dest.
5005 // (9) each element of an oop array must be assignable
5006
5007 // (3) src and dest must not be null.
5008 // always do this here because we need the JVM state for uncommon traps
5009 Node* null_ctl = top();
5010 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5011 assert(null_ctl->is_top(), "no null control here");
5012 dest = null_check(dest, T_ARRAY);
5013
5014 if (!can_emit_guards) {
5015 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5016 // guards but the arraycopy node could still take advantage of a
5017 // tightly allocated allocation. tightly_coupled_allocation() is
5018 // called again to make sure it takes the null check above into
5019 // account: the null check is mandatory and if it caused an
5020 // uncommon trap to be emitted then the allocation can't be
5021 // considered tightly coupled in this context.
5022 alloc = tightly_coupled_allocation(dest, NULL);
5023 }
5024
5025 bool validated = false;
5026
5027 const Type* src_type = _gvn.type(src);
5028 const Type* dest_type = _gvn.type(dest);
5029 const TypeAryPtr* top_src = src_type->isa_aryptr();
5030 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5031
5032 // Do we have the type of src?
5033 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5034 // Do we have the type of dest?
5035 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5036 // Is the type for src from speculation?
5037 bool src_spec = false;
5038 // Is the type for dest from speculation?
5039 bool dest_spec = false;
5040
5041 if ((!has_src || !has_dest) && can_emit_guards) {
5042 // We don't have sufficient type information, let's see if
5043 // speculative types can help. We need to have types for both src
5044 // and dest so that it pays off.
5045
5046 // Do we already have or could we have type information for src
5047 bool could_have_src = has_src;
5048 // Do we already have or could we have type information for dest
5049 bool could_have_dest = has_dest;
5050
5051 ciKlass* src_k = NULL;
5052 if (!has_src) {
5053 src_k = src_type->speculative_type_not_null();
5054 if (src_k != NULL && src_k->is_array_klass()) {
5055 could_have_src = true;
5056 }
5057 }
5058
5059 ciKlass* dest_k = NULL;
5060 if (!has_dest) {
5061 dest_k = dest_type->speculative_type_not_null();
5062 if (dest_k != NULL && dest_k->is_array_klass()) {
5063 could_have_dest = true;
5064 }
5065 }
5066
5067 if (could_have_src && could_have_dest) {
5068 // This is going to pay off so emit the required guards
5069 if (!has_src) {
5070 src = maybe_cast_profiled_obj(src, src_k, true);
5071 src_type = _gvn.type(src);
5072 top_src = src_type->isa_aryptr();
5073 has_src = (top_src != NULL && top_src->klass() != NULL);
5074 src_spec = true;
5075 }
5076 if (!has_dest) {
5077 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5078 dest_type = _gvn.type(dest);
5079 top_dest = dest_type->isa_aryptr();
5080 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5081 dest_spec = true;
5082 }
5083 }
5084 }
5085
5086 if (has_src && has_dest && can_emit_guards) {
5087 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5088 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5089 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5090 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5091
5092 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5093 // If both arrays are object arrays then having the exact types
5094 // for both will remove the need for a subtype check at runtime
5095 // before the call and may make it possible to pick a faster copy
5096 // routine (without a subtype check on every element)
5097 // Do we have the exact type of src?
5098 bool could_have_src = src_spec;
5099 // Do we have the exact type of dest?
5100 bool could_have_dest = dest_spec;
5101 ciKlass* src_k = top_src->klass();
5102 ciKlass* dest_k = top_dest->klass();
5103 if (!src_spec) {
5104 src_k = src_type->speculative_type_not_null();
5105 if (src_k != NULL && src_k->is_array_klass()) {
5106 could_have_src = true;
5107 }
5108 }
5109 if (!dest_spec) {
5110 dest_k = dest_type->speculative_type_not_null();
5111 if (dest_k != NULL && dest_k->is_array_klass()) {
5112 could_have_dest = true;
5113 }
5114 }
5115 if (could_have_src && could_have_dest) {
5116 // If we can have both exact types, emit the missing guards
5117 if (could_have_src && !src_spec) {
5118 src = maybe_cast_profiled_obj(src, src_k, true);
5119 }
5120 if (could_have_dest && !dest_spec) {
5121 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5122 }
5123 }
5124 }
5125 }
5126
5127 ciMethod* trap_method = method();
5128 int trap_bci = bci();
5129 if (saved_jvms != NULL) {
5130 trap_method = alloc->jvms()->method();
5131 trap_bci = alloc->jvms()->bci();
5132 }
5133
5134 bool negative_length_guard_generated = false;
5135
5136 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5137 can_emit_guards &&
5138 !src->is_top() && !dest->is_top()) {
5139 // validate arguments: enables transformation the ArrayCopyNode
5140 validated = true;
5141
5142 RegionNode* slow_region = new RegionNode(1);
5143 record_for_igvn(slow_region);
5144
5145 // (1) src and dest are arrays.
5146 generate_non_array_guard(load_object_klass(src), slow_region);
5147 generate_non_array_guard(load_object_klass(dest), slow_region);
5148
5149 // (2) src and dest arrays must have elements of the same BasicType
5150 // done at macro expansion or at Ideal transformation time
5151
5152 // (4) src_offset must not be negative.
5153 generate_negative_guard(src_offset, slow_region);
5154
5155 // (5) dest_offset must not be negative.
5156 generate_negative_guard(dest_offset, slow_region);
5157
5158 // (7) src_offset + length must not exceed length of src.
5159 generate_limit_guard(src_offset, length,
5160 load_array_length(src),
5161 slow_region);
5162
5163 // (8) dest_offset + length must not exceed length of dest.
5164 generate_limit_guard(dest_offset, length,
5165 load_array_length(dest),
5166 slow_region);
5167
5168 // (6) length must not be negative.
5169 // This is also checked in generate_arraycopy() during macro expansion, but
5170 // we also have to check it here for the case where the ArrayCopyNode will
5171 // be eliminated by Escape Analysis.
5172 if (EliminateAllocations) {
5173 generate_negative_guard(length, slow_region);
5174 negative_length_guard_generated = true;
5175 }
5176
5177 // (9) each element of an oop array must be assignable
5178 Node* src_klass = load_object_klass(src);
5179 Node* dest_klass = load_object_klass(dest);
5180 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5181
5182 if (not_subtype_ctrl != top()) {
5183 PreserveJVMState pjvms(this);
5184 set_control(not_subtype_ctrl);
5185 uncommon_trap(Deoptimization::Reason_intrinsic,
5186 Deoptimization::Action_make_not_entrant);
5187 assert(stopped(), "Should be stopped");
5188 }
5189 {
5190 PreserveJVMState pjvms(this);
5191 set_control(_gvn.transform(slow_region));
5192 uncommon_trap(Deoptimization::Reason_intrinsic,
5193 Deoptimization::Action_make_not_entrant);
5194 assert(stopped(), "Should be stopped");
5195 }
5196 }
5197
5198 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5199
5200 if (stopped()) {
5201 return true;
5202 }
5203
5204 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5205 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5206 // so the compiler has a chance to eliminate them: during macro expansion,
5207 // we have to set their control (CastPP nodes are eliminated).
5208 load_object_klass(src), load_object_klass(dest),
5209 load_array_length(src), load_array_length(dest));
5210
5211 ac->set_arraycopy(validated);
5212
5213 Node* n = _gvn.transform(ac);
5214 if (n == ac) {
5215 ac->connect_outputs(this);
5216 } else {
5217 assert(validated, "shouldn't transform if all arguments not validated");
5218 set_all_memory(n);
5219 }
5220
5221 return true;
5222 }
5223
5224
5225 // Helper function which determines if an arraycopy immediately follows
5226 // an allocation, with no intervening tests or other escapes for the object.
5227 AllocateArrayNode*
5228 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5229 RegionNode* slow_region) {
5230 if (stopped()) return NULL; // no fast path
5231 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5232
5233 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5234 if (alloc == NULL) return NULL;
5235
5236 Node* rawmem = memory(Compile::AliasIdxRaw);
5237 // Is the allocation's memory state untouched?
5238 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5239 // Bail out if there have been raw-memory effects since the allocation.
5240 // (Example: There might have been a call or safepoint.)
5241 return NULL;
5242 }
5243 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5244 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5245 return NULL;
5246 }
5247
5248 // There must be no unexpected observers of this allocation.
5249 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5250 Node* obs = ptr->fast_out(i);
5251 if (obs != this->map()) {
5252 return NULL;
5253 }
5254 }
5255
5256 // This arraycopy must unconditionally follow the allocation of the ptr.
5257 Node* alloc_ctl = ptr->in(0);
5258 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5259
5260 Node* ctl = control();
5261 while (ctl != alloc_ctl) {
5262 // There may be guards which feed into the slow_region.
5263 // Any other control flow means that we might not get a chance
5264 // to finish initializing the allocated object.
5265 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5266 IfNode* iff = ctl->in(0)->as_If();
5267 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5268 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5269 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5270 ctl = iff->in(0); // This test feeds the known slow_region.
5271 continue;
5272 }
5273 // One more try: Various low-level checks bottom out in
5274 // uncommon traps. If the debug-info of the trap omits
5275 // any reference to the allocation, as we've already
5276 // observed, then there can be no objection to the trap.
5277 bool found_trap = false;
5278 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5279 Node* obs = not_ctl->fast_out(j);
5280 if (obs->in(0) == not_ctl && obs->is_Call() &&
5281 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5282 found_trap = true; break;
5283 }
5284 }
5285 if (found_trap) {
5286 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5287 continue;
5288 }
5289 }
5290 return NULL;
5291 }
5292
5293 // If we get this far, we have an allocation which immediately
5294 // precedes the arraycopy, and we can take over zeroing the new object.
5295 // The arraycopy will finish the initialization, and provide
5296 // a new control state to which we will anchor the destination pointer.
5297
5298 return alloc;
5299 }
5300
5301 //-------------inline_encodeISOArray-----------------------------------
5302 // encode char[] to byte[] in ISO_8859_1
5303 bool LibraryCallKit::inline_encodeISOArray() {
5304 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5305 // no receiver since it is static method
5306 Node *src = argument(0);
5307 Node *src_offset = argument(1);
5308 Node *dst = argument(2);
5309 Node *dst_offset = argument(3);
5310 Node *length = argument(4);
5311
5312 const Type* src_type = src->Value(&_gvn);
5313 const Type* dst_type = dst->Value(&_gvn);
5314 const TypeAryPtr* top_src = src_type->isa_aryptr();
5315 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5316 if (top_src == NULL || top_src->klass() == NULL ||
5317 top_dest == NULL || top_dest->klass() == NULL) {
5318 // failed array check
5319 return false;
5320 }
5321
5322 // Figure out the size and type of the elements we will be copying.
5323 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5324 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5325 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5326 return false;
5327 }
5328
5329 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5330 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5331 // 'src_start' points to src array + scaled offset
5332 // 'dst_start' points to dst array + scaled offset
5333
5334 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5335 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5336 enc = _gvn.transform(enc);
5337 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5338 set_memory(res_mem, mtype);
5339 set_result(enc);
5340 return true;
5341 }
5342
5343 //-------------inline_multiplyToLen-----------------------------------
5344 bool LibraryCallKit::inline_multiplyToLen() {
5345 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5346
5347 address stubAddr = StubRoutines::multiplyToLen();
5348 if (stubAddr == NULL) {
5349 return false; // Intrinsic's stub is not implemented on this platform
5350 }
5351 const char* stubName = "multiplyToLen";
5352
5353 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5354
5355 // no receiver because it is a static method
5356 Node* x = argument(0);
5357 Node* xlen = argument(1);
5358 Node* y = argument(2);
5359 Node* ylen = argument(3);
5360 Node* z = argument(4);
5361
5362 const Type* x_type = x->Value(&_gvn);
5363 const Type* y_type = y->Value(&_gvn);
5364 const TypeAryPtr* top_x = x_type->isa_aryptr();
5365 const TypeAryPtr* top_y = y_type->isa_aryptr();
5366 if (top_x == NULL || top_x->klass() == NULL ||
5367 top_y == NULL || top_y->klass() == NULL) {
5368 // failed array check
5369 return false;
5370 }
5371
5372 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5373 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5374 if (x_elem != T_INT || y_elem != T_INT) {
5375 return false;
5376 }
5377
5378 // Set the original stack and the reexecute bit for the interpreter to reexecute
5379 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5380 // on the return from z array allocation in runtime.
5381 { PreserveReexecuteState preexecs(this);
5382 jvms()->set_should_reexecute(true);
5383
5384 Node* x_start = array_element_address(x, intcon(0), x_elem);
5385 Node* y_start = array_element_address(y, intcon(0), y_elem);
5386 // 'x_start' points to x array + scaled xlen
5387 // 'y_start' points to y array + scaled ylen
5388
5389 // Allocate the result array
5390 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5391 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5392 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5393
5394 IdealKit ideal(this);
5395
5396 #define __ ideal.
5397 Node* one = __ ConI(1);
5398 Node* zero = __ ConI(0);
5399 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5400 __ set(need_alloc, zero);
5401 __ set(z_alloc, z);
5402 __ if_then(z, BoolTest::eq, null()); {
5403 __ increment (need_alloc, one);
5404 } __ else_(); {
5405 // Update graphKit memory and control from IdealKit.
5406 sync_kit(ideal);
5407 Node* zlen_arg = load_array_length(z);
5408 // Update IdealKit memory and control from graphKit.
5409 __ sync_kit(this);
5410 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5411 __ increment (need_alloc, one);
5412 } __ end_if();
5413 } __ end_if();
5414
5415 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5416 // Update graphKit memory and control from IdealKit.
5417 sync_kit(ideal);
5418 Node * narr = new_array(klass_node, zlen, 1);
5419 // Update IdealKit memory and control from graphKit.
5420 __ sync_kit(this);
5421 __ set(z_alloc, narr);
5422 } __ end_if();
5423
5424 sync_kit(ideal);
5425 z = __ value(z_alloc);
5426 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5427 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5428 // Final sync IdealKit and GraphKit.
5429 final_sync(ideal);
5430 #undef __
5431
5432 Node* z_start = array_element_address(z, intcon(0), T_INT);
5433
5434 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5435 OptoRuntime::multiplyToLen_Type(),
5436 stubAddr, stubName, TypePtr::BOTTOM,
5437 x_start, xlen, y_start, ylen, z_start, zlen);
5438 } // original reexecute is set back here
5439
5440 C->set_has_split_ifs(true); // Has chance for split-if optimization
5441 set_result(z);
5442 return true;
5443 }
5444
5445 //-------------inline_squareToLen------------------------------------
5446 bool LibraryCallKit::inline_squareToLen() {
5447 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5448
5449 address stubAddr = StubRoutines::squareToLen();
5450 if (stubAddr == NULL) {
5451 return false; // Intrinsic's stub is not implemented on this platform
5452 }
5453 const char* stubName = "squareToLen";
5454
5455 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5456
5457 Node* x = argument(0);
5458 Node* len = argument(1);
5459 Node* z = argument(2);
5460 Node* zlen = argument(3);
5461
5462 const Type* x_type = x->Value(&_gvn);
5463 const Type* z_type = z->Value(&_gvn);
5464 const TypeAryPtr* top_x = x_type->isa_aryptr();
5465 const TypeAryPtr* top_z = z_type->isa_aryptr();
5466 if (top_x == NULL || top_x->klass() == NULL ||
5467 top_z == NULL || top_z->klass() == NULL) {
5468 // failed array check
5469 return false;
5470 }
5471
5472 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5473 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5474 if (x_elem != T_INT || z_elem != T_INT) {
5475 return false;
5476 }
5477
5478
5479 Node* x_start = array_element_address(x, intcon(0), x_elem);
5480 Node* z_start = array_element_address(z, intcon(0), z_elem);
5481
5482 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5483 OptoRuntime::squareToLen_Type(),
5484 stubAddr, stubName, TypePtr::BOTTOM,
5485 x_start, len, z_start, zlen);
5486
5487 set_result(z);
5488 return true;
5489 }
5490
5491 //-------------inline_mulAdd------------------------------------------
5492 bool LibraryCallKit::inline_mulAdd() {
5493 assert(UseMulAddIntrinsic, "not implemented on this platform");
5494
5495 address stubAddr = StubRoutines::mulAdd();
5496 if (stubAddr == NULL) {
5497 return false; // Intrinsic's stub is not implemented on this platform
5498 }
5499 const char* stubName = "mulAdd";
5500
5501 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5502
5503 Node* out = argument(0);
5504 Node* in = argument(1);
5505 Node* offset = argument(2);
5506 Node* len = argument(3);
5507 Node* k = argument(4);
5508
5509 const Type* out_type = out->Value(&_gvn);
5510 const Type* in_type = in->Value(&_gvn);
5511 const TypeAryPtr* top_out = out_type->isa_aryptr();
5512 const TypeAryPtr* top_in = in_type->isa_aryptr();
5513 if (top_out == NULL || top_out->klass() == NULL ||
5514 top_in == NULL || top_in->klass() == NULL) {
5515 // failed array check
5516 return false;
5517 }
5518
5519 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5520 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5521 if (out_elem != T_INT || in_elem != T_INT) {
5522 return false;
5523 }
5524
5525 Node* outlen = load_array_length(out);
5526 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5527 Node* out_start = array_element_address(out, intcon(0), out_elem);
5528 Node* in_start = array_element_address(in, intcon(0), in_elem);
5529
5530 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5531 OptoRuntime::mulAdd_Type(),
5532 stubAddr, stubName, TypePtr::BOTTOM,
5533 out_start,in_start, new_offset, len, k);
5534 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5535 set_result(result);
5536 return true;
5537 }
5538
5539 //-------------inline_montgomeryMultiply-----------------------------------
5540 bool LibraryCallKit::inline_montgomeryMultiply() {
5541 address stubAddr = StubRoutines::montgomeryMultiply();
5542 if (stubAddr == NULL) {
5543 return false; // Intrinsic's stub is not implemented on this platform
5544 }
5545
5546 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5547 const char* stubName = "montgomery_multiply";
5548
5549 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5550
5551 Node* a = argument(0);
5552 Node* b = argument(1);
5553 Node* n = argument(2);
5554 Node* len = argument(3);
5555 Node* inv = argument(4);
5556 Node* m = argument(6);
5557
5558 const Type* a_type = a->Value(&_gvn);
5559 const TypeAryPtr* top_a = a_type->isa_aryptr();
5560 const Type* b_type = b->Value(&_gvn);
5561 const TypeAryPtr* top_b = b_type->isa_aryptr();
5562 const Type* n_type = a->Value(&_gvn);
5563 const TypeAryPtr* top_n = n_type->isa_aryptr();
5564 const Type* m_type = a->Value(&_gvn);
5565 const TypeAryPtr* top_m = m_type->isa_aryptr();
5566 if (top_a == NULL || top_a->klass() == NULL ||
5567 top_b == NULL || top_b->klass() == NULL ||
5568 top_n == NULL || top_n->klass() == NULL ||
5569 top_m == NULL || top_m->klass() == NULL) {
5570 // failed array check
5571 return false;
5572 }
5573
5574 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5575 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5576 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5577 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5578 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5579 return false;
5580 }
5581
5582 // Make the call
5583 {
5584 Node* a_start = array_element_address(a, intcon(0), a_elem);
5585 Node* b_start = array_element_address(b, intcon(0), b_elem);
5586 Node* n_start = array_element_address(n, intcon(0), n_elem);
5587 Node* m_start = array_element_address(m, intcon(0), m_elem);
5588
5589 Node* call = make_runtime_call(RC_LEAF,
5590 OptoRuntime::montgomeryMultiply_Type(),
5591 stubAddr, stubName, TypePtr::BOTTOM,
5592 a_start, b_start, n_start, len, inv, top(),
5593 m_start);
5594 set_result(m);
5595 }
5596
5597 return true;
5598 }
5599
5600 bool LibraryCallKit::inline_montgomerySquare() {
5601 address stubAddr = StubRoutines::montgomerySquare();
5602 if (stubAddr == NULL) {
5603 return false; // Intrinsic's stub is not implemented on this platform
5604 }
5605
5606 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5607 const char* stubName = "montgomery_square";
5608
5609 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5610
5611 Node* a = argument(0);
5612 Node* n = argument(1);
5613 Node* len = argument(2);
5614 Node* inv = argument(3);
5615 Node* m = argument(5);
5616
5617 const Type* a_type = a->Value(&_gvn);
5618 const TypeAryPtr* top_a = a_type->isa_aryptr();
5619 const Type* n_type = a->Value(&_gvn);
5620 const TypeAryPtr* top_n = n_type->isa_aryptr();
5621 const Type* m_type = a->Value(&_gvn);
5622 const TypeAryPtr* top_m = m_type->isa_aryptr();
5623 if (top_a == NULL || top_a->klass() == NULL ||
5624 top_n == NULL || top_n->klass() == NULL ||
5625 top_m == NULL || top_m->klass() == NULL) {
5626 // failed array check
5627 return false;
5628 }
5629
5630 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5631 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5632 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5633 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5634 return false;
5635 }
5636
5637 // Make the call
5638 {
5639 Node* a_start = array_element_address(a, intcon(0), a_elem);
5640 Node* n_start = array_element_address(n, intcon(0), n_elem);
5641 Node* m_start = array_element_address(m, intcon(0), m_elem);
5642
5643 Node* call = make_runtime_call(RC_LEAF,
5644 OptoRuntime::montgomerySquare_Type(),
5645 stubAddr, stubName, TypePtr::BOTTOM,
5646 a_start, n_start, len, inv, top(),
5647 m_start);
5648 set_result(m);
5649 }
5650
5651 return true;
5652 }
5653
5654 //-------------inline_vectorizedMismatch------------------------------
5655 bool LibraryCallKit::inline_vectorizedMismatch() {
5656 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5657
5658 address stubAddr = StubRoutines::vectorizedMismatch();
5659 if (stubAddr == NULL) {
5660 return false; // Intrinsic's stub is not implemented on this platform
5661 }
5662 const char* stubName = "vectorizedMismatch";
5663 int size_l = callee()->signature()->size();
5664 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5665
5666 Node* obja = argument(0);
5667 Node* aoffset = argument(1);
5668 Node* objb = argument(3);
5669 Node* boffset = argument(4);
5670 Node* length = argument(6);
5671 Node* scale = argument(7);
5672
5673 const Type* a_type = obja->Value(&_gvn);
5674 const Type* b_type = objb->Value(&_gvn);
5675 const TypeAryPtr* top_a = a_type->isa_aryptr();
5676 const TypeAryPtr* top_b = b_type->isa_aryptr();
5677 if (top_a == NULL || top_a->klass() == NULL ||
5678 top_b == NULL || top_b->klass() == NULL) {
5679 // failed array check
5680 return false;
5681 }
5682
5683 Node* call;
5684 jvms()->set_should_reexecute(true);
5685
5686 Node* obja_adr = make_unsafe_address(obja, aoffset);
5687 Node* objb_adr = make_unsafe_address(objb, boffset);
5688
5689 call = make_runtime_call(RC_LEAF,
5690 OptoRuntime::vectorizedMismatch_Type(),
5691 stubAddr, stubName, TypePtr::BOTTOM,
5692 obja_adr, objb_adr, length, scale);
5693
5694 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5695 set_result(result);
5696 return true;
5697 }
5698
5699 /**
5700 * Calculate CRC32 for byte.
5701 * int java.util.zip.CRC32.update(int crc, int b)
5702 */
5703 bool LibraryCallKit::inline_updateCRC32() {
5704 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5705 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5706 // no receiver since it is static method
5707 Node* crc = argument(0); // type: int
5708 Node* b = argument(1); // type: int
5709
5710 /*
5711 * int c = ~ crc;
5712 * b = timesXtoThe32[(b ^ c) & 0xFF];
5713 * b = b ^ (c >>> 8);
5714 * crc = ~b;
5715 */
5716
5717 Node* M1 = intcon(-1);
5718 crc = _gvn.transform(new XorINode(crc, M1));
5719 Node* result = _gvn.transform(new XorINode(crc, b));
5720 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5721
5722 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5723 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5724 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5725 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5726
5727 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5728 result = _gvn.transform(new XorINode(crc, result));
5729 result = _gvn.transform(new XorINode(result, M1));
5730 set_result(result);
5731 return true;
5732 }
5733
5734 /**
5735 * Calculate CRC32 for byte[] array.
5736 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5737 */
5738 bool LibraryCallKit::inline_updateBytesCRC32() {
5739 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5740 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5741 // no receiver since it is static method
5742 Node* crc = argument(0); // type: int
5743 Node* src = argument(1); // type: oop
5744 Node* offset = argument(2); // type: int
5745 Node* length = argument(3); // type: int
5746
5747 const Type* src_type = src->Value(&_gvn);
5748 const TypeAryPtr* top_src = src_type->isa_aryptr();
5749 if (top_src == NULL || top_src->klass() == NULL) {
5750 // failed array check
5751 return false;
5752 }
5753
5754 // Figure out the size and type of the elements we will be copying.
5755 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5756 if (src_elem != T_BYTE) {
5757 return false;
5758 }
5759
5760 // 'src_start' points to src array + scaled offset
5761 Node* src_start = array_element_address(src, offset, src_elem);
5762
5763 // We assume that range check is done by caller.
5764 // TODO: generate range check (offset+length < src.length) in debug VM.
5765
5766 // Call the stub.
5767 address stubAddr = StubRoutines::updateBytesCRC32();
5768 const char *stubName = "updateBytesCRC32";
5769
5770 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5771 stubAddr, stubName, TypePtr::BOTTOM,
5772 crc, src_start, length);
5773 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5774 set_result(result);
5775 return true;
5776 }
5777
5778 /**
5779 * Calculate CRC32 for ByteBuffer.
5780 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5781 */
5782 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5783 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5784 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5785 // no receiver since it is static method
5786 Node* crc = argument(0); // type: int
5787 Node* src = argument(1); // type: long
5788 Node* offset = argument(3); // type: int
5789 Node* length = argument(4); // type: int
5790
5791 src = ConvL2X(src); // adjust Java long to machine word
5792 Node* base = _gvn.transform(new CastX2PNode(src));
5793 offset = ConvI2X(offset);
5794
5795 // 'src_start' points to src array + scaled offset
5796 Node* src_start = basic_plus_adr(top(), base, offset);
5797
5798 // Call the stub.
5799 address stubAddr = StubRoutines::updateBytesCRC32();
5800 const char *stubName = "updateBytesCRC32";
5801
5802 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5803 stubAddr, stubName, TypePtr::BOTTOM,
5804 crc, src_start, length);
5805 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5806 set_result(result);
5807 return true;
5808 }
5809
5810 //------------------------------get_table_from_crc32c_class-----------------------
5811 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5812 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5813 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5814
5815 return table;
5816 }
5817
5818 //------------------------------inline_updateBytesCRC32C-----------------------
5819 //
5820 // Calculate CRC32C for byte[] array.
5821 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5822 //
5823 bool LibraryCallKit::inline_updateBytesCRC32C() {
5824 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5825 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5826 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5827 // no receiver since it is a static method
5828 Node* crc = argument(0); // type: int
5829 Node* src = argument(1); // type: oop
5830 Node* offset = argument(2); // type: int
5831 Node* end = argument(3); // type: int
5832
5833 Node* length = _gvn.transform(new SubINode(end, offset));
5834
5835 const Type* src_type = src->Value(&_gvn);
5836 const TypeAryPtr* top_src = src_type->isa_aryptr();
5837 if (top_src == NULL || top_src->klass() == NULL) {
5838 // failed array check
5839 return false;
5840 }
5841
5842 // Figure out the size and type of the elements we will be copying.
5843 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5844 if (src_elem != T_BYTE) {
5845 return false;
5846 }
5847
5848 // 'src_start' points to src array + scaled offset
5849 Node* src_start = array_element_address(src, offset, src_elem);
5850
5851 // static final int[] byteTable in class CRC32C
5852 Node* table = get_table_from_crc32c_class(callee()->holder());
5853 Node* table_start = array_element_address(table, intcon(0), T_INT);
5854
5855 // We assume that range check is done by caller.
5856 // TODO: generate range check (offset+length < src.length) in debug VM.
5857
5858 // Call the stub.
5859 address stubAddr = StubRoutines::updateBytesCRC32C();
5860 const char *stubName = "updateBytesCRC32C";
5861
5862 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5863 stubAddr, stubName, TypePtr::BOTTOM,
5864 crc, src_start, length, table_start);
5865 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5866 set_result(result);
5867 return true;
5868 }
5869
5870 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5871 //
5872 // Calculate CRC32C for DirectByteBuffer.
5873 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5874 //
5875 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5876 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5877 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5878 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5879 // no receiver since it is a static method
5880 Node* crc = argument(0); // type: int
5881 Node* src = argument(1); // type: long
5882 Node* offset = argument(3); // type: int
5883 Node* end = argument(4); // type: int
5884
5885 Node* length = _gvn.transform(new SubINode(end, offset));
5886
5887 src = ConvL2X(src); // adjust Java long to machine word
5888 Node* base = _gvn.transform(new CastX2PNode(src));
5889 offset = ConvI2X(offset);
5890
5891 // 'src_start' points to src array + scaled offset
5892 Node* src_start = basic_plus_adr(top(), base, offset);
5893
5894 // static final int[] byteTable in class CRC32C
5895 Node* table = get_table_from_crc32c_class(callee()->holder());
5896 Node* table_start = array_element_address(table, intcon(0), T_INT);
5897
5898 // Call the stub.
5899 address stubAddr = StubRoutines::updateBytesCRC32C();
5900 const char *stubName = "updateBytesCRC32C";
5901
5902 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5903 stubAddr, stubName, TypePtr::BOTTOM,
5904 crc, src_start, length, table_start);
5905 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5906 set_result(result);
5907 return true;
5908 }
5909
5910 //------------------------------inline_updateBytesAdler32----------------------
5911 //
5912 // Calculate Adler32 checksum for byte[] array.
5913 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5914 //
5915 bool LibraryCallKit::inline_updateBytesAdler32() {
5916 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5917 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5918 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5919 // no receiver since it is static method
5920 Node* crc = argument(0); // type: int
5921 Node* src = argument(1); // type: oop
5922 Node* offset = argument(2); // type: int
5923 Node* length = argument(3); // type: int
5924
5925 const Type* src_type = src->Value(&_gvn);
5926 const TypeAryPtr* top_src = src_type->isa_aryptr();
5927 if (top_src == NULL || top_src->klass() == NULL) {
5928 // failed array check
5929 return false;
5930 }
5931
5932 // Figure out the size and type of the elements we will be copying.
5933 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5934 if (src_elem != T_BYTE) {
5935 return false;
5936 }
5937
5938 // 'src_start' points to src array + scaled offset
5939 Node* src_start = array_element_address(src, offset, src_elem);
5940
5941 // We assume that range check is done by caller.
5942 // TODO: generate range check (offset+length < src.length) in debug VM.
5943
5944 // Call the stub.
5945 address stubAddr = StubRoutines::updateBytesAdler32();
5946 const char *stubName = "updateBytesAdler32";
5947
5948 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5949 stubAddr, stubName, TypePtr::BOTTOM,
5950 crc, src_start, length);
5951 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5952 set_result(result);
5953 return true;
5954 }
5955
5956 //------------------------------inline_updateByteBufferAdler32---------------
5957 //
5958 // Calculate Adler32 checksum for DirectByteBuffer.
5959 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5960 //
5961 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5962 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5963 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5964 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5965 // no receiver since it is static method
5966 Node* crc = argument(0); // type: int
5967 Node* src = argument(1); // type: long
5968 Node* offset = argument(3); // type: int
5969 Node* length = argument(4); // type: int
5970
5971 src = ConvL2X(src); // adjust Java long to machine word
5972 Node* base = _gvn.transform(new CastX2PNode(src));
5973 offset = ConvI2X(offset);
5974
5975 // 'src_start' points to src array + scaled offset
5976 Node* src_start = basic_plus_adr(top(), base, offset);
5977
5978 // Call the stub.
5979 address stubAddr = StubRoutines::updateBytesAdler32();
5980 const char *stubName = "updateBytesAdler32";
5981
5982 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5983 stubAddr, stubName, TypePtr::BOTTOM,
5984 crc, src_start, length);
5985
5986 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5987 set_result(result);
5988 return true;
5989 }
5990
5991 //----------------------------inline_reference_get----------------------------
5992 // public T java.lang.ref.Reference.get();
5993 bool LibraryCallKit::inline_reference_get() {
5994 const int referent_offset = java_lang_ref_Reference::referent_offset;
5995 guarantee(referent_offset > 0, "should have already been set");
5996
5997 // Get the argument:
5998 Node* reference_obj = null_check_receiver();
5999 if (stopped()) return true;
6000
6001 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6002
6003 ciInstanceKlass* klass = env()->Object_klass();
6004 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6005
6006 Node* no_ctrl = NULL;
6007 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6008
6009 // Use the pre-barrier to record the value in the referent field
6010 pre_barrier(false /* do_load */,
6011 control(),
6012 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6013 result /* pre_val */,
6014 T_OBJECT);
6015
6016 // Add memory barrier to prevent commoning reads from this field
6017 // across safepoint since GC can change its value.
6018 insert_mem_bar(Op_MemBarCPUOrder);
6019
6020 set_result(result);
6021 return true;
6022 }
6023
6024
6025 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6026 bool is_exact=true, bool is_static=false,
6027 ciInstanceKlass * fromKls=NULL) {
6028 if (fromKls == NULL) {
6029 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6030 assert(tinst != NULL, "obj is null");
6031 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6032 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6033 fromKls = tinst->klass()->as_instance_klass();
6034 } else {
6035 assert(is_static, "only for static field access");
6036 }
6037 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6038 ciSymbol::make(fieldTypeString),
6039 is_static);
6040
6041 assert (field != NULL, "undefined field");
6042 if (field == NULL) return (Node *) NULL;
6043
6044 if (is_static) {
6045 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6046 fromObj = makecon(tip);
6047 }
6048
6049 // Next code copied from Parse::do_get_xxx():
6050
6051 // Compute address and memory type.
6052 int offset = field->offset_in_bytes();
6053 bool is_vol = field->is_volatile();
6054 ciType* field_klass = field->type();
6055 assert(field_klass->is_loaded(), "should be loaded");
6056 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6057 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6058 BasicType bt = field->layout_type();
6059
6060 // Build the resultant type of the load
6061 const Type *type;
6062 if (bt == T_OBJECT) {
6063 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6064 } else {
6065 type = Type::get_const_basic_type(bt);
6066 }
6067
6068 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6069 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6070 }
6071 // Build the load.
6072 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6073 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6074 // If reference is volatile, prevent following memory ops from
6075 // floating up past the volatile read. Also prevents commoning
6076 // another volatile read.
6077 if (is_vol) {
6078 // Memory barrier includes bogus read of value to force load BEFORE membar
6079 insert_mem_bar(Op_MemBarAcquire, loadedField);
6080 }
6081 return loadedField;
6082 }
6083
6084 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6085 bool is_exact = true, bool is_static = false,
6086 ciInstanceKlass * fromKls = NULL) {
6087 if (fromKls == NULL) {
6088 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6089 assert(tinst != NULL, "obj is null");
6090 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6091 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6092 fromKls = tinst->klass()->as_instance_klass();
6093 }
6094 else {
6095 assert(is_static, "only for static field access");
6096 }
6097 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6098 ciSymbol::make(fieldTypeString),
6099 is_static);
6100
6101 assert(field != NULL, "undefined field");
6102 assert(!field->is_volatile(), "not defined for volatile fields");
6103
6104 if (is_static) {
6105 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6106 fromObj = makecon(tip);
6107 }
6108
6109 // Next code copied from Parse::do_get_xxx():
6110
6111 // Compute address and memory type.
6112 int offset = field->offset_in_bytes();
6113 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6114
6115 return adr;
6116 }
6117
6118 //------------------------------inline_aescrypt_Block-----------------------
6119 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6120 address stubAddr = NULL;
6121 const char *stubName;
6122 assert(UseAES, "need AES instruction support");
6123
6124 switch(id) {
6125 case vmIntrinsics::_aescrypt_encryptBlock:
6126 stubAddr = StubRoutines::aescrypt_encryptBlock();
6127 stubName = "aescrypt_encryptBlock";
6128 break;
6129 case vmIntrinsics::_aescrypt_decryptBlock:
6130 stubAddr = StubRoutines::aescrypt_decryptBlock();
6131 stubName = "aescrypt_decryptBlock";
6132 break;
6133 }
6134 if (stubAddr == NULL) return false;
6135
6136 Node* aescrypt_object = argument(0);
6137 Node* src = argument(1);
6138 Node* src_offset = argument(2);
6139 Node* dest = argument(3);
6140 Node* dest_offset = argument(4);
6141
6142 // (1) src and dest are arrays.
6143 const Type* src_type = src->Value(&_gvn);
6144 const Type* dest_type = dest->Value(&_gvn);
6145 const TypeAryPtr* top_src = src_type->isa_aryptr();
6146 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6147 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6148
6149 // for the quick and dirty code we will skip all the checks.
6150 // we are just trying to get the call to be generated.
6151 Node* src_start = src;
6152 Node* dest_start = dest;
6153 if (src_offset != NULL || dest_offset != NULL) {
6154 assert(src_offset != NULL && dest_offset != NULL, "");
6155 src_start = array_element_address(src, src_offset, T_BYTE);
6156 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6157 }
6158
6159 // now need to get the start of its expanded key array
6160 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6161 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6162 if (k_start == NULL) return false;
6163
6164 if (Matcher::pass_original_key_for_aes()) {
6165 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6166 // compatibility issues between Java key expansion and SPARC crypto instructions
6167 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6168 if (original_k_start == NULL) return false;
6169
6170 // Call the stub.
6171 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6172 stubAddr, stubName, TypePtr::BOTTOM,
6173 src_start, dest_start, k_start, original_k_start);
6174 } else {
6175 // Call the stub.
6176 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6177 stubAddr, stubName, TypePtr::BOTTOM,
6178 src_start, dest_start, k_start);
6179 }
6180
6181 return true;
6182 }
6183
6184 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6185 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6186 address stubAddr = NULL;
6187 const char *stubName = NULL;
6188
6189 assert(UseAES, "need AES instruction support");
6190
6191 switch(id) {
6192 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6193 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6194 stubName = "cipherBlockChaining_encryptAESCrypt";
6195 break;
6196 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6197 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6198 stubName = "cipherBlockChaining_decryptAESCrypt";
6199 break;
6200 }
6201 if (stubAddr == NULL) return false;
6202
6203 Node* cipherBlockChaining_object = argument(0);
6204 Node* src = argument(1);
6205 Node* src_offset = argument(2);
6206 Node* len = argument(3);
6207 Node* dest = argument(4);
6208 Node* dest_offset = argument(5);
6209
6210 // (1) src and dest are arrays.
6211 const Type* src_type = src->Value(&_gvn);
6212 const Type* dest_type = dest->Value(&_gvn);
6213 const TypeAryPtr* top_src = src_type->isa_aryptr();
6214 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6215 assert (top_src != NULL && top_src->klass() != NULL
6216 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6217
6218 // checks are the responsibility of the caller
6219 Node* src_start = src;
6220 Node* dest_start = dest;
6221 if (src_offset != NULL || dest_offset != NULL) {
6222 assert(src_offset != NULL && dest_offset != NULL, "");
6223 src_start = array_element_address(src, src_offset, T_BYTE);
6224 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6225 }
6226
6227 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6228 // (because of the predicated logic executed earlier).
6229 // so we cast it here safely.
6230 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6231
6232 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6233 if (embeddedCipherObj == NULL) return false;
6234
6235 // cast it to what we know it will be at runtime
6236 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6237 assert(tinst != NULL, "CBC obj is null");
6238 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6239 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6240 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6241
6242 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6243 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6244 const TypeOopPtr* xtype = aklass->as_instance_type();
6245 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6246 aescrypt_object = _gvn.transform(aescrypt_object);
6247
6248 // we need to get the start of the aescrypt_object's expanded key array
6249 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6250 if (k_start == NULL) return false;
6251
6252 // similarly, get the start address of the r vector
6253 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6254 if (objRvec == NULL) return false;
6255 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6256
6257 Node* cbcCrypt;
6258 if (Matcher::pass_original_key_for_aes()) {
6259 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6260 // compatibility issues between Java key expansion and SPARC crypto instructions
6261 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6262 if (original_k_start == NULL) return false;
6263
6264 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6265 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6266 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6267 stubAddr, stubName, TypePtr::BOTTOM,
6268 src_start, dest_start, k_start, r_start, len, original_k_start);
6269 } else {
6270 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6271 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6272 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6273 stubAddr, stubName, TypePtr::BOTTOM,
6274 src_start, dest_start, k_start, r_start, len);
6275 }
6276
6277 // return cipher length (int)
6278 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6279 set_result(retvalue);
6280 return true;
6281 }
6282
6283 //------------------------------inline_counterMode_AESCrypt-----------------------
6284 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6285 assert(UseAES, "need AES instruction support");
6286 if (!UseAESCTRIntrinsics) return false;
6287
6288 address stubAddr = NULL;
6289 const char *stubName = NULL;
6290 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6291 stubAddr = StubRoutines::counterMode_AESCrypt();
6292 stubName = "counterMode_AESCrypt";
6293 }
6294 if (stubAddr == NULL) return false;
6295
6296 Node* counterMode_object = argument(0);
6297 Node* src = argument(1);
6298 Node* src_offset = argument(2);
6299 Node* len = argument(3);
6300 Node* dest = argument(4);
6301 Node* dest_offset = argument(5);
6302
6303 // (1) src and dest are arrays.
6304 const Type* src_type = src->Value(&_gvn);
6305 const Type* dest_type = dest->Value(&_gvn);
6306 const TypeAryPtr* top_src = src_type->isa_aryptr();
6307 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6308 assert(top_src != NULL && top_src->klass() != NULL &&
6309 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6310
6311 // checks are the responsibility of the caller
6312 Node* src_start = src;
6313 Node* dest_start = dest;
6314 if (src_offset != NULL || dest_offset != NULL) {
6315 assert(src_offset != NULL && dest_offset != NULL, "");
6316 src_start = array_element_address(src, src_offset, T_BYTE);
6317 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6318 }
6319
6320 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6321 // (because of the predicated logic executed earlier).
6322 // so we cast it here safely.
6323 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6324 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6325 if (embeddedCipherObj == NULL) return false;
6326 // cast it to what we know it will be at runtime
6327 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6328 assert(tinst != NULL, "CTR obj is null");
6329 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6330 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6331 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6332 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6333 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6334 const TypeOopPtr* xtype = aklass->as_instance_type();
6335 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6336 aescrypt_object = _gvn.transform(aescrypt_object);
6337 // we need to get the start of the aescrypt_object's expanded key array
6338 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6339 if (k_start == NULL) return false;
6340 // similarly, get the start address of the r vector
6341 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6342 if (obj_counter == NULL) return false;
6343 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6344
6345 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6346 if (saved_encCounter == NULL) return false;
6347 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6348 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6349
6350 Node* ctrCrypt;
6351 if (Matcher::pass_original_key_for_aes()) {
6352 // no SPARC version for AES/CTR intrinsics now.
6353 return false;
6354 }
6355 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6356 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6357 OptoRuntime::counterMode_aescrypt_Type(),
6358 stubAddr, stubName, TypePtr::BOTTOM,
6359 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6360
6361 // return cipher length (int)
6362 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6363 set_result(retvalue);
6364 return true;
6365 }
6366
6367 //------------------------------get_key_start_from_aescrypt_object-----------------------
6368 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6369 #if defined(PPC64) || defined(S390)
6370 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6371 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6372 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6373 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6374 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6375 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6376 if (objSessionK == NULL) {
6377 return (Node *) NULL;
6378 }
6379 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6380 #else
6381 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6382 #endif // PPC64
6383 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6384 if (objAESCryptKey == NULL) return (Node *) NULL;
6385
6386 // now have the array, need to get the start address of the K array
6387 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6388 return k_start;
6389 }
6390
6391 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6392 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6393 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6394 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6395 if (objAESCryptKey == NULL) return (Node *) NULL;
6396
6397 // now have the array, need to get the start address of the lastKey array
6398 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6399 return original_k_start;
6400 }
6401
6402 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6403 // Return node representing slow path of predicate check.
6404 // the pseudo code we want to emulate with this predicate is:
6405 // for encryption:
6406 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6407 // for decryption:
6408 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6409 // note cipher==plain is more conservative than the original java code but that's OK
6410 //
6411 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6412 // The receiver was checked for NULL already.
6413 Node* objCBC = argument(0);
6414
6415 // Load embeddedCipher field of CipherBlockChaining object.
6416 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6417
6418 // get AESCrypt klass for instanceOf check
6419 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6420 // will have same classloader as CipherBlockChaining object
6421 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6422 assert(tinst != NULL, "CBCobj is null");
6423 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6424
6425 // we want to do an instanceof comparison against the AESCrypt class
6426 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6427 if (!klass_AESCrypt->is_loaded()) {
6428 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6429 Node* ctrl = control();
6430 set_control(top()); // no regular fast path
6431 return ctrl;
6432 }
6433 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6434
6435 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6436 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6437 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6438
6439 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6440
6441 // for encryption, we are done
6442 if (!decrypting)
6443 return instof_false; // even if it is NULL
6444
6445 // for decryption, we need to add a further check to avoid
6446 // taking the intrinsic path when cipher and plain are the same
6447 // see the original java code for why.
6448 RegionNode* region = new RegionNode(3);
6449 region->init_req(1, instof_false);
6450 Node* src = argument(1);
6451 Node* dest = argument(4);
6452 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6453 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6454 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6455 region->init_req(2, src_dest_conjoint);
6456
6457 record_for_igvn(region);
6458 return _gvn.transform(region);
6459 }
6460
6461 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6462 // Return node representing slow path of predicate check.
6463 // the pseudo code we want to emulate with this predicate is:
6464 // for encryption:
6465 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6466 // for decryption:
6467 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6468 // note cipher==plain is more conservative than the original java code but that's OK
6469 //
6470
6471 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6472 // The receiver was checked for NULL already.
6473 Node* objCTR = argument(0);
6474
6475 // Load embeddedCipher field of CipherBlockChaining object.
6476 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6477
6478 // get AESCrypt klass for instanceOf check
6479 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6480 // will have same classloader as CipherBlockChaining object
6481 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6482 assert(tinst != NULL, "CTRobj is null");
6483 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6484
6485 // we want to do an instanceof comparison against the AESCrypt class
6486 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6487 if (!klass_AESCrypt->is_loaded()) {
6488 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6489 Node* ctrl = control();
6490 set_control(top()); // no regular fast path
6491 return ctrl;
6492 }
6493
6494 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6495 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6496 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6497 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6498 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6499
6500 return instof_false; // even if it is NULL
6501 }
6502
6503 //------------------------------inline_ghash_processBlocks
6504 bool LibraryCallKit::inline_ghash_processBlocks() {
6505 address stubAddr;
6506 const char *stubName;
6507 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6508
6509 stubAddr = StubRoutines::ghash_processBlocks();
6510 stubName = "ghash_processBlocks";
6511
6512 Node* data = argument(0);
6513 Node* offset = argument(1);
6514 Node* len = argument(2);
6515 Node* state = argument(3);
6516 Node* subkeyH = argument(4);
6517
6518 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6519 assert(state_start, "state is NULL");
6520 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6521 assert(subkeyH_start, "subkeyH is NULL");
6522 Node* data_start = array_element_address(data, offset, T_BYTE);
6523 assert(data_start, "data is NULL");
6524
6525 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6526 OptoRuntime::ghash_processBlocks_Type(),
6527 stubAddr, stubName, TypePtr::BOTTOM,
6528 state_start, subkeyH_start, data_start, len);
6529 return true;
6530 }
6531
6532 //------------------------------inline_sha_implCompress-----------------------
6533 //
6534 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6535 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6536 //
6537 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6538 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6539 //
6540 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6541 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6542 //
6543 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6544 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6545
6546 Node* sha_obj = argument(0);
6547 Node* src = argument(1); // type oop
6548 Node* ofs = argument(2); // type int
6549
6550 const Type* src_type = src->Value(&_gvn);
6551 const TypeAryPtr* top_src = src_type->isa_aryptr();
6552 if (top_src == NULL || top_src->klass() == NULL) {
6553 // failed array check
6554 return false;
6555 }
6556 // Figure out the size and type of the elements we will be copying.
6557 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6558 if (src_elem != T_BYTE) {
6559 return false;
6560 }
6561 // 'src_start' points to src array + offset
6562 Node* src_start = array_element_address(src, ofs, src_elem);
6563 Node* state = NULL;
6564 address stubAddr;
6565 const char *stubName;
6566
6567 switch(id) {
6568 case vmIntrinsics::_sha_implCompress:
6569 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6570 state = get_state_from_sha_object(sha_obj);
6571 stubAddr = StubRoutines::sha1_implCompress();
6572 stubName = "sha1_implCompress";
6573 break;
6574 case vmIntrinsics::_sha2_implCompress:
6575 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6576 state = get_state_from_sha_object(sha_obj);
6577 stubAddr = StubRoutines::sha256_implCompress();
6578 stubName = "sha256_implCompress";
6579 break;
6580 case vmIntrinsics::_sha5_implCompress:
6581 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6582 state = get_state_from_sha5_object(sha_obj);
6583 stubAddr = StubRoutines::sha512_implCompress();
6584 stubName = "sha512_implCompress";
6585 break;
6586 default:
6587 fatal_unexpected_iid(id);
6588 return false;
6589 }
6590 if (state == NULL) return false;
6591
6592 // Call the stub.
6593 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6594 stubAddr, stubName, TypePtr::BOTTOM,
6595 src_start, state);
6596
6597 return true;
6598 }
6599
6600 //------------------------------inline_digestBase_implCompressMB-----------------------
6601 //
6602 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6603 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6604 //
6605 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6606 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6607 "need SHA1/SHA256/SHA512 instruction support");
6608 assert((uint)predicate < 3, "sanity");
6609 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6610
6611 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6612 Node* src = argument(1); // byte[] array
6613 Node* ofs = argument(2); // type int
6614 Node* limit = argument(3); // type int
6615
6616 const Type* src_type = src->Value(&_gvn);
6617 const TypeAryPtr* top_src = src_type->isa_aryptr();
6618 if (top_src == NULL || top_src->klass() == NULL) {
6619 // failed array check
6620 return false;
6621 }
6622 // Figure out the size and type of the elements we will be copying.
6623 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6624 if (src_elem != T_BYTE) {
6625 return false;
6626 }
6627 // 'src_start' points to src array + offset
6628 Node* src_start = array_element_address(src, ofs, src_elem);
6629
6630 const char* klass_SHA_name = NULL;
6631 const char* stub_name = NULL;
6632 address stub_addr = NULL;
6633 bool long_state = false;
6634
6635 switch (predicate) {
6636 case 0:
6637 if (UseSHA1Intrinsics) {
6638 klass_SHA_name = "sun/security/provider/SHA";
6639 stub_name = "sha1_implCompressMB";
6640 stub_addr = StubRoutines::sha1_implCompressMB();
6641 }
6642 break;
6643 case 1:
6644 if (UseSHA256Intrinsics) {
6645 klass_SHA_name = "sun/security/provider/SHA2";
6646 stub_name = "sha256_implCompressMB";
6647 stub_addr = StubRoutines::sha256_implCompressMB();
6648 }
6649 break;
6650 case 2:
6651 if (UseSHA512Intrinsics) {
6652 klass_SHA_name = "sun/security/provider/SHA5";
6653 stub_name = "sha512_implCompressMB";
6654 stub_addr = StubRoutines::sha512_implCompressMB();
6655 long_state = true;
6656 }
6657 break;
6658 default:
6659 fatal("unknown SHA intrinsic predicate: %d", predicate);
6660 }
6661 if (klass_SHA_name != NULL) {
6662 // get DigestBase klass to lookup for SHA klass
6663 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6664 assert(tinst != NULL, "digestBase_obj is not instance???");
6665 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6666
6667 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6668 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6669 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6670 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6671 }
6672 return false;
6673 }
6674 //------------------------------inline_sha_implCompressMB-----------------------
6675 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6676 bool long_state, address stubAddr, const char *stubName,
6677 Node* src_start, Node* ofs, Node* limit) {
6678 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6679 const TypeOopPtr* xtype = aklass->as_instance_type();
6680 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6681 sha_obj = _gvn.transform(sha_obj);
6682
6683 Node* state;
6684 if (long_state) {
6685 state = get_state_from_sha5_object(sha_obj);
6686 } else {
6687 state = get_state_from_sha_object(sha_obj);
6688 }
6689 if (state == NULL) return false;
6690
6691 // Call the stub.
6692 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6693 OptoRuntime::digestBase_implCompressMB_Type(),
6694 stubAddr, stubName, TypePtr::BOTTOM,
6695 src_start, state, ofs, limit);
6696 // return ofs (int)
6697 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6698 set_result(result);
6699
6700 return true;
6701 }
6702
6703 //------------------------------get_state_from_sha_object-----------------------
6704 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6705 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6706 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6707 if (sha_state == NULL) return (Node *) NULL;
6708
6709 // now have the array, need to get the start address of the state array
6710 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6711 return state;
6712 }
6713
6714 //------------------------------get_state_from_sha5_object-----------------------
6715 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6716 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6717 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6718 if (sha_state == NULL) return (Node *) NULL;
6719
6720 // now have the array, need to get the start address of the state array
6721 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6722 return state;
6723 }
6724
6725 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6726 // Return node representing slow path of predicate check.
6727 // the pseudo code we want to emulate with this predicate is:
6728 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6729 //
6730 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6731 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6732 "need SHA1/SHA256/SHA512 instruction support");
6733 assert((uint)predicate < 3, "sanity");
6734
6735 // The receiver was checked for NULL already.
6736 Node* digestBaseObj = argument(0);
6737
6738 // get DigestBase klass for instanceOf check
6739 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6740 assert(tinst != NULL, "digestBaseObj is null");
6741 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6742
6743 const char* klass_SHA_name = NULL;
6744 switch (predicate) {
6745 case 0:
6746 if (UseSHA1Intrinsics) {
6747 // we want to do an instanceof comparison against the SHA class
6748 klass_SHA_name = "sun/security/provider/SHA";
6749 }
6750 break;
6751 case 1:
6752 if (UseSHA256Intrinsics) {
6753 // we want to do an instanceof comparison against the SHA2 class
6754 klass_SHA_name = "sun/security/provider/SHA2";
6755 }
6756 break;
6757 case 2:
6758 if (UseSHA512Intrinsics) {
6759 // we want to do an instanceof comparison against the SHA5 class
6760 klass_SHA_name = "sun/security/provider/SHA5";
6761 }
6762 break;
6763 default:
6764 fatal("unknown SHA intrinsic predicate: %d", predicate);
6765 }
6766
6767 ciKlass* klass_SHA = NULL;
6768 if (klass_SHA_name != NULL) {
6769 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6770 }
6771 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6772 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6773 Node* ctrl = control();
6774 set_control(top()); // no intrinsic path
6775 return ctrl;
6776 }
6777 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6778
6779 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6780 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6781 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6782 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6783
6784 return instof_false; // even if it is NULL
6785 }
6786
6787 //-------------inline_fma-----------------------------------
6788 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6789 Node *a = NULL;
6790 Node *b = NULL;
6791 Node *c = NULL;
6792 Node* result = NULL;
6793 switch (id) {
6794 case vmIntrinsics::_fmaD:
6795 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6796 // no receiver since it is static method
6797 a = round_double_node(argument(0));
6798 b = round_double_node(argument(2));
6799 c = round_double_node(argument(4));
6800 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6801 break;
6802 case vmIntrinsics::_fmaF:
6803 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6804 a = argument(0);
6805 b = argument(1);
6806 c = argument(2);
6807 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6808 break;
6809 default:
6810 fatal_unexpected_iid(id); break;
6811 }
6812 set_result(result);
6813 return true;
6814 }
6815
6816 bool LibraryCallKit::inline_profileBoolean() {
6817 Node* counts = argument(1);
6818 const TypeAryPtr* ary = NULL;
6819 ciArray* aobj = NULL;
6820 if (counts->is_Con()
6821 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6822 && (aobj = ary->const_oop()->as_array()) != NULL
6823 && (aobj->length() == 2)) {
6824 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6825 jint false_cnt = aobj->element_value(0).as_int();
6826 jint true_cnt = aobj->element_value(1).as_int();
6827
6828 if (C->log() != NULL) {
6829 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6830 false_cnt, true_cnt);
6831 }
6832
6833 if (false_cnt + true_cnt == 0) {
6834 // According to profile, never executed.
6835 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6836 Deoptimization::Action_reinterpret);
6837 return true;
6838 }
6839
6840 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6841 // is a number of each value occurrences.
6842 Node* result = argument(0);
6843 if (false_cnt == 0 || true_cnt == 0) {
6844 // According to profile, one value has been never seen.
6845 int expected_val = (false_cnt == 0) ? 1 : 0;
6846
6847 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6848 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6849
6850 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6851 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6852 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6853
6854 { // Slow path: uncommon trap for never seen value and then reexecute
6855 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6856 // the value has been seen at least once.
6857 PreserveJVMState pjvms(this);
6858 PreserveReexecuteState preexecs(this);
6859 jvms()->set_should_reexecute(true);
6860
6861 set_control(slow_path);
6862 set_i_o(i_o());
6863
6864 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6865 Deoptimization::Action_reinterpret);
6866 }
6867 // The guard for never seen value enables sharpening of the result and
6868 // returning a constant. It allows to eliminate branches on the same value
6869 // later on.
6870 set_control(fast_path);
6871 result = intcon(expected_val);
6872 }
6873 // Stop profiling.
6874 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6875 // By replacing method body with profile data (represented as ProfileBooleanNode
6876 // on IR level) we effectively disable profiling.
6877 // It enables full speed execution once optimized code is generated.
6878 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6879 C->record_for_igvn(profile);
6880 set_result(profile);
6881 return true;
6882 } else {
6883 // Continue profiling.
6884 // Profile data isn't available at the moment. So, execute method's bytecode version.
6885 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6886 // is compiled and counters aren't available since corresponding MethodHandle
6887 // isn't a compile-time constant.
6888 return false;
6889 }
6890 }
6891
6892 bool LibraryCallKit::inline_isCompileConstant() {
6893 Node* n = argument(0);
6894 set_result(n->is_Con() ? intcon(1) : intcon(0));
6895 return true;
6896 }