1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "prims/unsafe.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #ifdef TRACE_HAVE_INTRINSICS
56 #include "trace/traceMacros.hpp"
57 #endif
58
59 class LibraryIntrinsic : public InlineCallGenerator {
60 // Extend the set of intrinsics known to the runtime:
61 public:
62 private:
63 bool _is_virtual;
64 bool _does_virtual_dispatch;
65 int8_t _predicates_count; // Intrinsic is predicated by several conditions
66 int8_t _last_predicate; // Last generated predicate
67 vmIntrinsics::ID _intrinsic_id;
68
69 public:
70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
71 : InlineCallGenerator(m),
72 _is_virtual(is_virtual),
73 _does_virtual_dispatch(does_virtual_dispatch),
74 _predicates_count((int8_t)predicates_count),
75 _last_predicate((int8_t)-1),
76 _intrinsic_id(id)
77 {
78 }
79 virtual bool is_intrinsic() const { return true; }
80 virtual bool is_virtual() const { return _is_virtual; }
81 virtual bool is_predicated() const { return _predicates_count > 0; }
82 virtual int predicates_count() const { return _predicates_count; }
83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
84 virtual JVMState* generate(JVMState* jvms);
85 virtual Node* generate_predicate(JVMState* jvms, int predicate);
86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
87 };
88
89
90 // Local helper class for LibraryIntrinsic:
91 class LibraryCallKit : public GraphKit {
92 private:
93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
94 Node* _result; // the result node, if any
95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
96
97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
98
99 public:
100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
101 : GraphKit(jvms),
102 _intrinsic(intrinsic),
103 _result(NULL)
104 {
105 // Check if this is a root compile. In that case we don't have a caller.
106 if (!jvms->has_method()) {
107 _reexecute_sp = sp();
108 } else {
109 // Find out how many arguments the interpreter needs when deoptimizing
110 // and save the stack pointer value so it can used by uncommon_trap.
111 // We find the argument count by looking at the declared signature.
112 bool ignored_will_link;
113 ciSignature* declared_signature = NULL;
114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
117 }
118 }
119
120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
121
122 ciMethod* caller() const { return jvms()->method(); }
123 int bci() const { return jvms()->bci(); }
124 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
126 ciMethod* callee() const { return _intrinsic->method(); }
127
128 bool try_to_inline(int predicate);
129 Node* try_to_predicate(int predicate);
130
131 void push_result() {
132 // Push the result onto the stack.
133 if (!stopped() && result() != NULL) {
134 BasicType bt = result()->bottom_type()->basic_type();
135 push_node(bt, result());
136 }
137 }
138
139 private:
140 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
142 }
143
144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
145 void set_result(RegionNode* region, PhiNode* value);
146 Node* result() { return _result; }
147
148 virtual int reexecute_sp() { return _reexecute_sp; }
149
150 // Helper functions to inline natives
151 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
152 Node* generate_slow_guard(Node* test, RegionNode* region);
153 Node* generate_fair_guard(Node* test, RegionNode* region);
154 Node* generate_negative_guard(Node* index, RegionNode* region,
155 // resulting CastII of index:
156 Node* *pos_index = NULL);
157 Node* generate_limit_guard(Node* offset, Node* subseq_length,
158 Node* array_length,
159 RegionNode* region);
160 void generate_string_range_check(Node* array, Node* offset,
161 Node* length, bool char_count);
162 Node* generate_current_thread(Node* &tls_output);
163 Node* load_mirror_from_klass(Node* klass);
164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path,
166 int offset);
167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
175 RegionNode* region, int null_path) {
176 int offset = java_lang_Class::array_klass_offset_in_bytes();
177 return load_klass_from_mirror_common(mirror, never_see_null,
178 region, null_path,
179 offset);
180 }
181 Node* generate_access_flags_guard(Node* kls,
182 int modifier_mask, int modifier_bits,
183 RegionNode* region);
184 Node* generate_interface_guard(Node* kls, RegionNode* region);
185 Node* generate_array_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, false, false);
187 }
188 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
189 return generate_array_guard_common(kls, region, false, true);
190 }
191 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
192 return generate_array_guard_common(kls, region, true, false);
193 }
194 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
195 return generate_array_guard_common(kls, region, true, true);
196 }
197 Node* generate_array_guard_common(Node* kls, RegionNode* region,
198 bool obj_array, bool not_array);
199 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
200 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
201 bool is_virtual = false, bool is_static = false);
202 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
203 return generate_method_call(method_id, false, true);
204 }
205 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
206 return generate_method_call(method_id, true, false);
207 }
208 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210
211 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
215 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
216 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_indexOfChar();
218 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_toBytesU();
220 bool inline_string_getCharsU();
221 bool inline_string_copy(bool compress);
222 bool inline_string_char_access(bool is_store);
223 Node* round_double_node(Node* n);
224 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_math_native(vmIntrinsics::ID id);
226 bool inline_math(vmIntrinsics::ID id);
227 template <typename OverflowOp>
228 bool inline_math_overflow(Node* arg1, Node* arg2);
229 void inline_math_mathExact(Node* math, Node* test);
230 bool inline_math_addExactI(bool is_increment);
231 bool inline_math_addExactL(bool is_increment);
232 bool inline_math_multiplyExactI();
233 bool inline_math_multiplyExactL();
234 bool inline_math_negateExactI();
235 bool inline_math_negateExactL();
236 bool inline_math_subtractExactI(bool is_decrement);
237 bool inline_math_subtractExactL(bool is_decrement);
238 bool inline_min_max(vmIntrinsics::ID id);
239 bool inline_notify(vmIntrinsics::ID id);
240 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
241 // This returns Type::AnyPtr, RawPtr, or OopPtr.
242 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
243 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
244 // Helper for inline_unsafe_access.
245 // Generates the guards that check whether the result of
246 // Unsafe.getObject should be recorded in an SATB log buffer.
247 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
248
249 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
250 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
251 static bool klass_needs_init_guard(Node* kls);
252 bool inline_unsafe_allocate();
253 bool inline_unsafe_newArray(bool uninitialized);
254 bool inline_unsafe_copyMemory();
255 bool inline_native_currentThread();
256
257 bool inline_native_time_funcs(address method, const char* funcName);
258 #ifdef TRACE_HAVE_INTRINSICS
259 bool inline_native_classID();
260 bool inline_native_getBufferWriter();
261 #endif
262 bool inline_native_isInterrupted();
263 bool inline_native_Class_query(vmIntrinsics::ID id);
264 bool inline_native_subtype_check();
265 bool inline_native_getLength();
266 bool inline_array_copyOf(bool is_copyOfRange);
267 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
268 bool inline_preconditions_checkIndex();
269 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
270 bool inline_native_clone(bool is_virtual);
271 bool inline_native_Reflection_getCallerClass();
272 // Helper function for inlining native object hash method
273 bool inline_native_hashcode(bool is_virtual, bool is_static);
274 bool inline_native_getClass();
275
276 // Helper functions for inlining arraycopy
277 bool inline_arraycopy();
278 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
279 RegionNode* slow_region);
280 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
281 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
282 uint new_idx);
283
284 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
285 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
286 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
288 bool inline_unsafe_fence(vmIntrinsics::ID id);
289 bool inline_onspinwait();
290 bool inline_fp_conversions(vmIntrinsics::ID id);
291 bool inline_number_methods(vmIntrinsics::ID id);
292 bool inline_reference_get();
293 bool inline_Class_cast();
294 bool inline_aescrypt_Block(vmIntrinsics::ID id);
295 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
296 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
297 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
298 Node* inline_counterMode_AESCrypt_predicate();
299 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
300 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
301 bool inline_ghash_processBlocks();
302 bool inline_sha_implCompress(vmIntrinsics::ID id);
303 bool inline_digestBase_implCompressMB(int predicate);
304 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
305 bool long_state, address stubAddr, const char *stubName,
306 Node* src_start, Node* ofs, Node* limit);
307 Node* get_state_from_sha_object(Node *sha_object);
308 Node* get_state_from_sha5_object(Node *sha_object);
309 Node* inline_digestBase_implCompressMB_predicate(int predicate);
310 bool inline_encodeISOArray();
311 bool inline_updateCRC32();
312 bool inline_updateBytesCRC32();
313 bool inline_updateByteBufferCRC32();
314 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
315 bool inline_updateBytesCRC32C();
316 bool inline_updateDirectByteBufferCRC32C();
317 bool inline_updateBytesAdler32();
318 bool inline_updateByteBufferAdler32();
319 bool inline_multiplyToLen();
320 bool inline_hasNegatives();
321 bool inline_squareToLen();
322 bool inline_mulAdd();
323 bool inline_montgomeryMultiply();
324 bool inline_montgomerySquare();
325 bool inline_vectorizedMismatch();
326 bool inline_fma(vmIntrinsics::ID id);
327
328 bool inline_profileBoolean();
329 bool inline_isCompileConstant();
330 };
331
332 //---------------------------make_vm_intrinsic----------------------------
333 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
334 vmIntrinsics::ID id = m->intrinsic_id();
335 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
336
337 if (!m->is_loaded()) {
338 // Do not attempt to inline unloaded methods.
339 return NULL;
340 }
341
342 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
343 bool is_available = false;
344
345 {
346 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
347 // the compiler must transition to '_thread_in_vm' state because both
348 // methods access VM-internal data.
349 VM_ENTRY_MARK;
350 methodHandle mh(THREAD, m->get_Method());
351 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
352 !C->directive()->is_intrinsic_disabled(mh) &&
353 !vmIntrinsics::is_disabled_by_flags(mh);
354
355 }
356
357 if (is_available) {
358 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
359 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
360 return new LibraryIntrinsic(m, is_virtual,
361 vmIntrinsics::predicates_needed(id),
362 vmIntrinsics::does_virtual_dispatch(id),
363 (vmIntrinsics::ID) id);
364 } else {
365 return NULL;
366 }
367 }
368
369 //----------------------register_library_intrinsics-----------------------
370 // Initialize this file's data structures, for each Compile instance.
371 void Compile::register_library_intrinsics() {
372 // Nothing to do here.
373 }
374
375 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
376 LibraryCallKit kit(jvms, this);
377 Compile* C = kit.C;
378 int nodes = C->unique();
379 #ifndef PRODUCT
380 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
381 char buf[1000];
382 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
383 tty->print_cr("Intrinsic %s", str);
384 }
385 #endif
386 ciMethod* callee = kit.callee();
387 const int bci = kit.bci();
388
389 // Try to inline the intrinsic.
390 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
391 kit.try_to_inline(_last_predicate)) {
392 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
393 : "(intrinsic)";
394 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
395 if (C->print_intrinsics() || C->print_inlining()) {
396 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
397 }
398 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
399 if (C->log()) {
400 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
401 vmIntrinsics::name_at(intrinsic_id()),
402 (is_virtual() ? " virtual='1'" : ""),
403 C->unique() - nodes);
404 }
405 // Push the result from the inlined method onto the stack.
406 kit.push_result();
407 C->print_inlining_update(this);
408 return kit.transfer_exceptions_into_jvms();
409 }
410
411 // The intrinsic bailed out
412 if (jvms->has_method()) {
413 // Not a root compile.
414 const char* msg;
415 if (callee->intrinsic_candidate()) {
416 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
417 } else {
418 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
419 : "failed to inline (intrinsic), method not annotated";
420 }
421 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
422 if (C->print_intrinsics() || C->print_inlining()) {
423 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
424 }
425 } else {
426 // Root compile
427 ResourceMark rm;
428 stringStream msg_stream;
429 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
430 vmIntrinsics::name_at(intrinsic_id()),
431 is_virtual() ? " (virtual)" : "", bci);
432 const char *msg = msg_stream.as_string();
433 log_debug(jit, inlining)("%s", msg);
434 if (C->print_intrinsics() || C->print_inlining()) {
435 tty->print("%s", msg);
436 }
437 }
438 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
439 C->print_inlining_update(this);
440 return NULL;
441 }
442
443 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
444 LibraryCallKit kit(jvms, this);
445 Compile* C = kit.C;
446 int nodes = C->unique();
447 _last_predicate = predicate;
448 #ifndef PRODUCT
449 assert(is_predicated() && predicate < predicates_count(), "sanity");
450 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
451 char buf[1000];
452 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
453 tty->print_cr("Predicate for intrinsic %s", str);
454 }
455 #endif
456 ciMethod* callee = kit.callee();
457 const int bci = kit.bci();
458
459 Node* slow_ctl = kit.try_to_predicate(predicate);
460 if (!kit.failing()) {
461 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
462 : "(intrinsic, predicate)";
463 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
464 if (C->print_intrinsics() || C->print_inlining()) {
465 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
466 }
467 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
468 if (C->log()) {
469 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
470 vmIntrinsics::name_at(intrinsic_id()),
471 (is_virtual() ? " virtual='1'" : ""),
472 C->unique() - nodes);
473 }
474 return slow_ctl; // Could be NULL if the check folds.
475 }
476
477 // The intrinsic bailed out
478 if (jvms->has_method()) {
479 // Not a root compile.
480 const char* msg = "failed to generate predicate for intrinsic";
481 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
482 if (C->print_intrinsics() || C->print_inlining()) {
483 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
484 }
485 } else {
486 // Root compile
487 ResourceMark rm;
488 stringStream msg_stream;
489 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
490 vmIntrinsics::name_at(intrinsic_id()),
491 is_virtual() ? " (virtual)" : "", bci);
492 const char *msg = msg_stream.as_string();
493 log_debug(jit, inlining)("%s", msg);
494 if (C->print_intrinsics() || C->print_inlining()) {
495 C->print_inlining_stream()->print("%s", msg);
496 }
497 }
498 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
499 return NULL;
500 }
501
502 bool LibraryCallKit::try_to_inline(int predicate) {
503 // Handle symbolic names for otherwise undistinguished boolean switches:
504 const bool is_store = true;
505 const bool is_compress = true;
506 const bool is_static = true;
507 const bool is_volatile = true;
508
509 if (!jvms()->has_method()) {
510 // Root JVMState has a null method.
511 assert(map()->memory()->Opcode() == Op_Parm, "");
512 // Insert the memory aliasing node
513 set_all_memory(reset_memory());
514 }
515 assert(merged_memory(), "");
516
517
518 switch (intrinsic_id()) {
519 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
520 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
521 case vmIntrinsics::_getClass: return inline_native_getClass();
522
523 case vmIntrinsics::_dsin:
524 case vmIntrinsics::_dcos:
525 case vmIntrinsics::_dtan:
526 case vmIntrinsics::_dabs:
527 case vmIntrinsics::_datan2:
528 case vmIntrinsics::_dsqrt:
529 case vmIntrinsics::_dexp:
530 case vmIntrinsics::_dlog:
531 case vmIntrinsics::_dlog10:
532 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
533
534 case vmIntrinsics::_min:
535 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
536
537 case vmIntrinsics::_notify:
538 case vmIntrinsics::_notifyAll:
539 if (InlineNotify) {
540 return inline_notify(intrinsic_id());
541 }
542 return false;
543
544 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
545 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
546 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
547 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
548 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
549 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
550 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
551 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
552 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
553 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
554 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
555 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
556
557 case vmIntrinsics::_arraycopy: return inline_arraycopy();
558
559 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
560 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
561 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
562 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
563
564 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
565 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
566 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
567 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
568 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
569 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
570 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
571
572 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
573 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
574
575 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
576 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
577 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
578 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
579
580 case vmIntrinsics::_compressStringC:
581 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
582 case vmIntrinsics::_inflateStringC:
583 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
584
585 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
586 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
587 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
588 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
589 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
590 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
591 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
592 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
593 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
594
595 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
596 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
597 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
598 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
599 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
600 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
601 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
602 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
603 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
604
605 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
606 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
607 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
608 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
609 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
610 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
611 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
612 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
613 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
614
615 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
616 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
617 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
618 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
619 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
620 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
621 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
622 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
623 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
624
625 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
626 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
627 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
628 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
629
630 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
631 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
632 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
633 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
634
635 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
636 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
637 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
638 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
639 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
640 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
641 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
642 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
643 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
644
645 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
646 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
647 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
648 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
649 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
650 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
651 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
652 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
653 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
654
655 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
656 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
657 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
658 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
659 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
660 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
661 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
662 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
663 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
664
665 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
666 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
667 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
668 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
669 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
670 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
671 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
672 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
673 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
674
675 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
676 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
677 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
678 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
679 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
680
681 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
682 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
683 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
684 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
685 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
686 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
687 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
688 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
689 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
690 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
691 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
692 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
693 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
694 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
695 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
696 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
697 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
698 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
699 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
700 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
701
702 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
703 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
704 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
705 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
706 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
707 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
708 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
709 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
710 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
711 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
712 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
713 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
714 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
715 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
716 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
717
718 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
719 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
720 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
721 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
722
723 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
724 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
725 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
726 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
727 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
728
729 case vmIntrinsics::_loadFence:
730 case vmIntrinsics::_storeFence:
731 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
732
733 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
734
735 case vmIntrinsics::_currentThread: return inline_native_currentThread();
736 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
737
738 #ifdef TRACE_HAVE_INTRINSICS
739 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
740 case vmIntrinsics::_getClassId: return inline_native_classID();
741 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
742 #endif
743 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
744 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
745 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
746 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
747 case vmIntrinsics::_getLength: return inline_native_getLength();
748 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
749 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
750 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
751 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
752 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
753 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
754
755 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
756 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
757
758 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
759
760 case vmIntrinsics::_isInstance:
761 case vmIntrinsics::_getModifiers:
762 case vmIntrinsics::_isInterface:
763 case vmIntrinsics::_isArray:
764 case vmIntrinsics::_isPrimitive:
765 case vmIntrinsics::_getSuperclass:
766 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
767
768 case vmIntrinsics::_floatToRawIntBits:
769 case vmIntrinsics::_floatToIntBits:
770 case vmIntrinsics::_intBitsToFloat:
771 case vmIntrinsics::_doubleToRawLongBits:
772 case vmIntrinsics::_doubleToLongBits:
773 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
774
775 case vmIntrinsics::_numberOfLeadingZeros_i:
776 case vmIntrinsics::_numberOfLeadingZeros_l:
777 case vmIntrinsics::_numberOfTrailingZeros_i:
778 case vmIntrinsics::_numberOfTrailingZeros_l:
779 case vmIntrinsics::_bitCount_i:
780 case vmIntrinsics::_bitCount_l:
781 case vmIntrinsics::_reverseBytes_i:
782 case vmIntrinsics::_reverseBytes_l:
783 case vmIntrinsics::_reverseBytes_s:
784 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
785
786 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
787
788 case vmIntrinsics::_Reference_get: return inline_reference_get();
789
790 case vmIntrinsics::_Class_cast: return inline_Class_cast();
791
792 case vmIntrinsics::_aescrypt_encryptBlock:
793 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
794
795 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
796 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
797 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
798
799 case vmIntrinsics::_counterMode_AESCrypt:
800 return inline_counterMode_AESCrypt(intrinsic_id());
801
802 case vmIntrinsics::_sha_implCompress:
803 case vmIntrinsics::_sha2_implCompress:
804 case vmIntrinsics::_sha5_implCompress:
805 return inline_sha_implCompress(intrinsic_id());
806
807 case vmIntrinsics::_digestBase_implCompressMB:
808 return inline_digestBase_implCompressMB(predicate);
809
810 case vmIntrinsics::_multiplyToLen:
811 return inline_multiplyToLen();
812
813 case vmIntrinsics::_squareToLen:
814 return inline_squareToLen();
815
816 case vmIntrinsics::_mulAdd:
817 return inline_mulAdd();
818
819 case vmIntrinsics::_montgomeryMultiply:
820 return inline_montgomeryMultiply();
821 case vmIntrinsics::_montgomerySquare:
822 return inline_montgomerySquare();
823
824 case vmIntrinsics::_vectorizedMismatch:
825 return inline_vectorizedMismatch();
826
827 case vmIntrinsics::_ghash_processBlocks:
828 return inline_ghash_processBlocks();
829
830 case vmIntrinsics::_encodeISOArray:
831 case vmIntrinsics::_encodeByteISOArray:
832 return inline_encodeISOArray();
833
834 case vmIntrinsics::_updateCRC32:
835 return inline_updateCRC32();
836 case vmIntrinsics::_updateBytesCRC32:
837 return inline_updateBytesCRC32();
838 case vmIntrinsics::_updateByteBufferCRC32:
839 return inline_updateByteBufferCRC32();
840
841 case vmIntrinsics::_updateBytesCRC32C:
842 return inline_updateBytesCRC32C();
843 case vmIntrinsics::_updateDirectByteBufferCRC32C:
844 return inline_updateDirectByteBufferCRC32C();
845
846 case vmIntrinsics::_updateBytesAdler32:
847 return inline_updateBytesAdler32();
848 case vmIntrinsics::_updateByteBufferAdler32:
849 return inline_updateByteBufferAdler32();
850
851 case vmIntrinsics::_profileBoolean:
852 return inline_profileBoolean();
853 case vmIntrinsics::_isCompileConstant:
854 return inline_isCompileConstant();
855
856 case vmIntrinsics::_hasNegatives:
857 return inline_hasNegatives();
858
859 case vmIntrinsics::_fmaD:
860 case vmIntrinsics::_fmaF:
861 return inline_fma(intrinsic_id());
862
863 default:
864 // If you get here, it may be that someone has added a new intrinsic
865 // to the list in vmSymbols.hpp without implementing it here.
866 #ifndef PRODUCT
867 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
868 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
869 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
870 }
871 #endif
872 return false;
873 }
874 }
875
876 Node* LibraryCallKit::try_to_predicate(int predicate) {
877 if (!jvms()->has_method()) {
878 // Root JVMState has a null method.
879 assert(map()->memory()->Opcode() == Op_Parm, "");
880 // Insert the memory aliasing node
881 set_all_memory(reset_memory());
882 }
883 assert(merged_memory(), "");
884
885 switch (intrinsic_id()) {
886 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
887 return inline_cipherBlockChaining_AESCrypt_predicate(false);
888 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
889 return inline_cipherBlockChaining_AESCrypt_predicate(true);
890 case vmIntrinsics::_counterMode_AESCrypt:
891 return inline_counterMode_AESCrypt_predicate();
892 case vmIntrinsics::_digestBase_implCompressMB:
893 return inline_digestBase_implCompressMB_predicate(predicate);
894
895 default:
896 // If you get here, it may be that someone has added a new intrinsic
897 // to the list in vmSymbols.hpp without implementing it here.
898 #ifndef PRODUCT
899 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
900 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
901 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
902 }
903 #endif
904 Node* slow_ctl = control();
905 set_control(top()); // No fast path instrinsic
906 return slow_ctl;
907 }
908 }
909
910 //------------------------------set_result-------------------------------
911 // Helper function for finishing intrinsics.
912 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
913 record_for_igvn(region);
914 set_control(_gvn.transform(region));
915 set_result( _gvn.transform(value));
916 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
917 }
918
919 //------------------------------generate_guard---------------------------
920 // Helper function for generating guarded fast-slow graph structures.
921 // The given 'test', if true, guards a slow path. If the test fails
922 // then a fast path can be taken. (We generally hope it fails.)
923 // In all cases, GraphKit::control() is updated to the fast path.
924 // The returned value represents the control for the slow path.
925 // The return value is never 'top'; it is either a valid control
926 // or NULL if it is obvious that the slow path can never be taken.
927 // Also, if region and the slow control are not NULL, the slow edge
928 // is appended to the region.
929 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
930 if (stopped()) {
931 // Already short circuited.
932 return NULL;
933 }
934
935 // Build an if node and its projections.
936 // If test is true we take the slow path, which we assume is uncommon.
937 if (_gvn.type(test) == TypeInt::ZERO) {
938 // The slow branch is never taken. No need to build this guard.
939 return NULL;
940 }
941
942 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
943
944 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
945 if (if_slow == top()) {
946 // The slow branch is never taken. No need to build this guard.
947 return NULL;
948 }
949
950 if (region != NULL)
951 region->add_req(if_slow);
952
953 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
954 set_control(if_fast);
955
956 return if_slow;
957 }
958
959 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
960 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
961 }
962 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
963 return generate_guard(test, region, PROB_FAIR);
964 }
965
966 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
967 Node* *pos_index) {
968 if (stopped())
969 return NULL; // already stopped
970 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
971 return NULL; // index is already adequately typed
972 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
973 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
974 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
975 if (is_neg != NULL && pos_index != NULL) {
976 // Emulate effect of Parse::adjust_map_after_if.
977 Node* ccast = new CastIINode(index, TypeInt::POS);
978 ccast->set_req(0, control());
979 (*pos_index) = _gvn.transform(ccast);
980 }
981 return is_neg;
982 }
983
984 // Make sure that 'position' is a valid limit index, in [0..length].
985 // There are two equivalent plans for checking this:
986 // A. (offset + copyLength) unsigned<= arrayLength
987 // B. offset <= (arrayLength - copyLength)
988 // We require that all of the values above, except for the sum and
989 // difference, are already known to be non-negative.
990 // Plan A is robust in the face of overflow, if offset and copyLength
991 // are both hugely positive.
992 //
993 // Plan B is less direct and intuitive, but it does not overflow at
994 // all, since the difference of two non-negatives is always
995 // representable. Whenever Java methods must perform the equivalent
996 // check they generally use Plan B instead of Plan A.
997 // For the moment we use Plan A.
998 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
999 Node* subseq_length,
1000 Node* array_length,
1001 RegionNode* region) {
1002 if (stopped())
1003 return NULL; // already stopped
1004 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1005 if (zero_offset && subseq_length->eqv_uncast(array_length))
1006 return NULL; // common case of whole-array copy
1007 Node* last = subseq_length;
1008 if (!zero_offset) // last += offset
1009 last = _gvn.transform(new AddINode(last, offset));
1010 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1011 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1012 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1013 return is_over;
1014 }
1015
1016 // Emit range checks for the given String.value byte array
1017 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1018 if (stopped()) {
1019 return; // already stopped
1020 }
1021 RegionNode* bailout = new RegionNode(1);
1022 record_for_igvn(bailout);
1023 if (char_count) {
1024 // Convert char count to byte count
1025 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1026 }
1027
1028 // Offset and count must not be negative
1029 generate_negative_guard(offset, bailout);
1030 generate_negative_guard(count, bailout);
1031 // Offset + count must not exceed length of array
1032 generate_limit_guard(offset, count, load_array_length(array), bailout);
1033
1034 if (bailout->req() > 1) {
1035 PreserveJVMState pjvms(this);
1036 set_control(_gvn.transform(bailout));
1037 uncommon_trap(Deoptimization::Reason_intrinsic,
1038 Deoptimization::Action_maybe_recompile);
1039 }
1040 }
1041
1042 //--------------------------generate_current_thread--------------------
1043 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1044 ciKlass* thread_klass = env()->Thread_klass();
1045 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1046 Node* thread = _gvn.transform(new ThreadLocalNode());
1047 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1048 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1049 tls_output = thread;
1050 return threadObj;
1051 }
1052
1053
1054 //------------------------------make_string_method_node------------------------
1055 // Helper method for String intrinsic functions. This version is called with
1056 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1057 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1058 // containing the lengths of str1 and str2.
1059 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1060 Node* result = NULL;
1061 switch (opcode) {
1062 case Op_StrIndexOf:
1063 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1064 str1_start, cnt1, str2_start, cnt2, ae);
1065 break;
1066 case Op_StrComp:
1067 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1068 str1_start, cnt1, str2_start, cnt2, ae);
1069 break;
1070 case Op_StrEquals:
1071 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1072 // Use the constant length if there is one because optimized match rule may exist.
1073 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1074 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1075 break;
1076 default:
1077 ShouldNotReachHere();
1078 return NULL;
1079 }
1080
1081 // All these intrinsics have checks.
1082 C->set_has_split_ifs(true); // Has chance for split-if optimization
1083
1084 return _gvn.transform(result);
1085 }
1086
1087 //------------------------------inline_string_compareTo------------------------
1088 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1089 Node* arg1 = argument(0);
1090 Node* arg2 = argument(1);
1091
1092 // Get start addr and length of first argument
1093 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1094 Node* arg1_cnt = load_array_length(arg1);
1095
1096 // Get start addr and length of second argument
1097 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1098 Node* arg2_cnt = load_array_length(arg2);
1099
1100 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1101 set_result(result);
1102 return true;
1103 }
1104
1105 //------------------------------inline_string_equals------------------------
1106 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1107 Node* arg1 = argument(0);
1108 Node* arg2 = argument(1);
1109
1110 // paths (plus control) merge
1111 RegionNode* region = new RegionNode(3);
1112 Node* phi = new PhiNode(region, TypeInt::BOOL);
1113
1114 if (!stopped()) {
1115 // Get start addr and length of first argument
1116 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1117 Node* arg1_cnt = load_array_length(arg1);
1118
1119 // Get start addr and length of second argument
1120 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1121 Node* arg2_cnt = load_array_length(arg2);
1122
1123 // Check for arg1_cnt != arg2_cnt
1124 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1125 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1126 Node* if_ne = generate_slow_guard(bol, NULL);
1127 if (if_ne != NULL) {
1128 phi->init_req(2, intcon(0));
1129 region->init_req(2, if_ne);
1130 }
1131
1132 // Check for count == 0 is done by assembler code for StrEquals.
1133
1134 if (!stopped()) {
1135 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1136 phi->init_req(1, equals);
1137 region->init_req(1, control());
1138 }
1139 }
1140
1141 // post merge
1142 set_control(_gvn.transform(region));
1143 record_for_igvn(region);
1144
1145 set_result(_gvn.transform(phi));
1146 return true;
1147 }
1148
1149 //------------------------------inline_array_equals----------------------------
1150 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1151 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1152 Node* arg1 = argument(0);
1153 Node* arg2 = argument(1);
1154
1155 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1156 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1157 return true;
1158 }
1159
1160 //------------------------------inline_hasNegatives------------------------------
1161 bool LibraryCallKit::inline_hasNegatives() {
1162 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1163 return false;
1164 }
1165
1166 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1167 // no receiver since it is static method
1168 Node* ba = argument(0);
1169 Node* offset = argument(1);
1170 Node* len = argument(2);
1171
1172 // Range checks
1173 generate_string_range_check(ba, offset, len, false);
1174 if (stopped()) {
1175 return true;
1176 }
1177 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1178 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1179 set_result(_gvn.transform(result));
1180 return true;
1181 }
1182
1183 bool LibraryCallKit::inline_preconditions_checkIndex() {
1184 Node* index = argument(0);
1185 Node* length = argument(1);
1186 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1187 return false;
1188 }
1189
1190 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1191 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1192
1193 {
1194 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1195 uncommon_trap(Deoptimization::Reason_intrinsic,
1196 Deoptimization::Action_make_not_entrant);
1197 }
1198
1199 if (stopped()) {
1200 return false;
1201 }
1202
1203 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1204 BoolTest::mask btest = BoolTest::lt;
1205 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1206 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1207 _gvn.set_type(rc, rc->Value(&_gvn));
1208 if (!rc_bool->is_Con()) {
1209 record_for_igvn(rc);
1210 }
1211 set_control(_gvn.transform(new IfTrueNode(rc)));
1212 {
1213 PreserveJVMState pjvms(this);
1214 set_control(_gvn.transform(new IfFalseNode(rc)));
1215 uncommon_trap(Deoptimization::Reason_range_check,
1216 Deoptimization::Action_make_not_entrant);
1217 }
1218
1219 if (stopped()) {
1220 return false;
1221 }
1222
1223 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1224 result->set_req(0, control());
1225 result = _gvn.transform(result);
1226 set_result(result);
1227 replace_in_map(index, result);
1228 return true;
1229 }
1230
1231 //------------------------------inline_string_indexOf------------------------
1232 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1233 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1234 return false;
1235 }
1236 Node* src = argument(0);
1237 Node* tgt = argument(1);
1238
1239 // Make the merge point
1240 RegionNode* result_rgn = new RegionNode(4);
1241 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1242
1243 // Get start addr and length of source string
1244 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1245 Node* src_count = load_array_length(src);
1246
1247 // Get start addr and length of substring
1248 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1249 Node* tgt_count = load_array_length(tgt);
1250
1251 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1252 // Divide src size by 2 if String is UTF16 encoded
1253 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1254 }
1255 if (ae == StrIntrinsicNode::UU) {
1256 // Divide substring size by 2 if String is UTF16 encoded
1257 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1258 }
1259
1260 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1261 if (result != NULL) {
1262 result_phi->init_req(3, result);
1263 result_rgn->init_req(3, control());
1264 }
1265 set_control(_gvn.transform(result_rgn));
1266 record_for_igvn(result_rgn);
1267 set_result(_gvn.transform(result_phi));
1268
1269 return true;
1270 }
1271
1272 //-----------------------------inline_string_indexOf-----------------------
1273 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1274 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1275 return false;
1276 }
1277 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1278 return false;
1279 }
1280 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1281 Node* src = argument(0); // byte[]
1282 Node* src_count = argument(1); // char count
1283 Node* tgt = argument(2); // byte[]
1284 Node* tgt_count = argument(3); // char count
1285 Node* from_index = argument(4); // char index
1286
1287 // Multiply byte array index by 2 if String is UTF16 encoded
1288 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1289 src_count = _gvn.transform(new SubINode(src_count, from_index));
1290 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1291 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1292
1293 // Range checks
1294 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1295 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1296 if (stopped()) {
1297 return true;
1298 }
1299
1300 RegionNode* region = new RegionNode(5);
1301 Node* phi = new PhiNode(region, TypeInt::INT);
1302
1303 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1304 if (result != NULL) {
1305 // The result is index relative to from_index if substring was found, -1 otherwise.
1306 // Generate code which will fold into cmove.
1307 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1308 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1309
1310 Node* if_lt = generate_slow_guard(bol, NULL);
1311 if (if_lt != NULL) {
1312 // result == -1
1313 phi->init_req(3, result);
1314 region->init_req(3, if_lt);
1315 }
1316 if (!stopped()) {
1317 result = _gvn.transform(new AddINode(result, from_index));
1318 phi->init_req(4, result);
1319 region->init_req(4, control());
1320 }
1321 }
1322
1323 set_control(_gvn.transform(region));
1324 record_for_igvn(region);
1325 set_result(_gvn.transform(phi));
1326
1327 return true;
1328 }
1329
1330 // Create StrIndexOfNode with fast path checks
1331 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1332 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1333 // Check for substr count > string count
1334 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1335 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1336 Node* if_gt = generate_slow_guard(bol, NULL);
1337 if (if_gt != NULL) {
1338 phi->init_req(1, intcon(-1));
1339 region->init_req(1, if_gt);
1340 }
1341 if (!stopped()) {
1342 // Check for substr count == 0
1343 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1344 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1345 Node* if_zero = generate_slow_guard(bol, NULL);
1346 if (if_zero != NULL) {
1347 phi->init_req(2, intcon(0));
1348 region->init_req(2, if_zero);
1349 }
1350 }
1351 if (!stopped()) {
1352 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1353 }
1354 return NULL;
1355 }
1356
1357 //-----------------------------inline_string_indexOfChar-----------------------
1358 bool LibraryCallKit::inline_string_indexOfChar() {
1359 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1360 return false;
1361 }
1362 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1363 return false;
1364 }
1365 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1366 Node* src = argument(0); // byte[]
1367 Node* tgt = argument(1); // tgt is int ch
1368 Node* from_index = argument(2);
1369 Node* max = argument(3);
1370
1371 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1372 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1373 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1374
1375 // Range checks
1376 generate_string_range_check(src, src_offset, src_count, true);
1377 if (stopped()) {
1378 return true;
1379 }
1380
1381 RegionNode* region = new RegionNode(3);
1382 Node* phi = new PhiNode(region, TypeInt::INT);
1383
1384 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1385 C->set_has_split_ifs(true); // Has chance for split-if optimization
1386 _gvn.transform(result);
1387
1388 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1389 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1390
1391 Node* if_lt = generate_slow_guard(bol, NULL);
1392 if (if_lt != NULL) {
1393 // result == -1
1394 phi->init_req(2, result);
1395 region->init_req(2, if_lt);
1396 }
1397 if (!stopped()) {
1398 result = _gvn.transform(new AddINode(result, from_index));
1399 phi->init_req(1, result);
1400 region->init_req(1, control());
1401 }
1402 set_control(_gvn.transform(region));
1403 record_for_igvn(region);
1404 set_result(_gvn.transform(phi));
1405
1406 return true;
1407 }
1408 //---------------------------inline_string_copy---------------------
1409 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1410 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1411 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1412 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1413 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1414 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1415 bool LibraryCallKit::inline_string_copy(bool compress) {
1416 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1417 return false;
1418 }
1419 int nargs = 5; // 2 oops, 3 ints
1420 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1421
1422 Node* src = argument(0);
1423 Node* src_offset = argument(1);
1424 Node* dst = argument(2);
1425 Node* dst_offset = argument(3);
1426 Node* length = argument(4);
1427
1428 // Check for allocation before we add nodes that would confuse
1429 // tightly_coupled_allocation()
1430 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1431
1432 // Figure out the size and type of the elements we will be copying.
1433 const Type* src_type = src->Value(&_gvn);
1434 const Type* dst_type = dst->Value(&_gvn);
1435 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1436 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1437 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1438 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1439 "Unsupported array types for inline_string_copy");
1440
1441 // Convert char[] offsets to byte[] offsets
1442 bool convert_src = (compress && src_elem == T_BYTE);
1443 bool convert_dst = (!compress && dst_elem == T_BYTE);
1444 if (convert_src) {
1445 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1446 } else if (convert_dst) {
1447 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1448 }
1449
1450 // Range checks
1451 generate_string_range_check(src, src_offset, length, convert_src);
1452 generate_string_range_check(dst, dst_offset, length, convert_dst);
1453 if (stopped()) {
1454 return true;
1455 }
1456
1457 Node* src_start = array_element_address(src, src_offset, src_elem);
1458 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1459 // 'src_start' points to src array + scaled offset
1460 // 'dst_start' points to dst array + scaled offset
1461 Node* count = NULL;
1462 if (compress) {
1463 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1464 } else {
1465 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1466 }
1467
1468 if (alloc != NULL) {
1469 if (alloc->maybe_set_complete(&_gvn)) {
1470 // "You break it, you buy it."
1471 InitializeNode* init = alloc->initialization();
1472 assert(init->is_complete(), "we just did this");
1473 init->set_complete_with_arraycopy();
1474 assert(dst->is_CheckCastPP(), "sanity");
1475 assert(dst->in(0)->in(0) == init, "dest pinned");
1476 }
1477 // Do not let stores that initialize this object be reordered with
1478 // a subsequent store that would make this object accessible by
1479 // other threads.
1480 // Record what AllocateNode this StoreStore protects so that
1481 // escape analysis can go from the MemBarStoreStoreNode to the
1482 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1483 // based on the escape status of the AllocateNode.
1484 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1485 }
1486 if (compress) {
1487 set_result(_gvn.transform(count));
1488 }
1489 return true;
1490 }
1491
1492 #ifdef _LP64
1493 #define XTOP ,top() /*additional argument*/
1494 #else //_LP64
1495 #define XTOP /*no additional argument*/
1496 #endif //_LP64
1497
1498 //------------------------inline_string_toBytesU--------------------------
1499 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1500 bool LibraryCallKit::inline_string_toBytesU() {
1501 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1502 return false;
1503 }
1504 // Get the arguments.
1505 Node* value = argument(0);
1506 Node* offset = argument(1);
1507 Node* length = argument(2);
1508
1509 Node* newcopy = NULL;
1510
1511 // Set the original stack and the reexecute bit for the interpreter to reexecute
1512 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1513 { PreserveReexecuteState preexecs(this);
1514 jvms()->set_should_reexecute(true);
1515
1516 // Check if a null path was taken unconditionally.
1517 value = null_check(value);
1518
1519 RegionNode* bailout = new RegionNode(1);
1520 record_for_igvn(bailout);
1521
1522 // Range checks
1523 generate_negative_guard(offset, bailout);
1524 generate_negative_guard(length, bailout);
1525 generate_limit_guard(offset, length, load_array_length(value), bailout);
1526 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1527 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1528
1529 if (bailout->req() > 1) {
1530 PreserveJVMState pjvms(this);
1531 set_control(_gvn.transform(bailout));
1532 uncommon_trap(Deoptimization::Reason_intrinsic,
1533 Deoptimization::Action_maybe_recompile);
1534 }
1535 if (stopped()) {
1536 return true;
1537 }
1538
1539 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1540 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1541 newcopy = new_array(klass_node, size, 0); // no arguments to push
1542 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1543
1544 // Calculate starting addresses.
1545 Node* src_start = array_element_address(value, offset, T_CHAR);
1546 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1547
1548 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1549 const TypeInt* toffset = gvn().type(offset)->is_int();
1550 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1551
1552 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1553 const char* copyfunc_name = "arraycopy";
1554 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1555 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1556 OptoRuntime::fast_arraycopy_Type(),
1557 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1558 src_start, dst_start, ConvI2X(length) XTOP);
1559 // Do not let reads from the cloned object float above the arraycopy.
1560 if (alloc != NULL) {
1561 if (alloc->maybe_set_complete(&_gvn)) {
1562 // "You break it, you buy it."
1563 InitializeNode* init = alloc->initialization();
1564 assert(init->is_complete(), "we just did this");
1565 init->set_complete_with_arraycopy();
1566 assert(newcopy->is_CheckCastPP(), "sanity");
1567 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1568 }
1569 // Do not let stores that initialize this object be reordered with
1570 // a subsequent store that would make this object accessible by
1571 // other threads.
1572 // Record what AllocateNode this StoreStore protects so that
1573 // escape analysis can go from the MemBarStoreStoreNode to the
1574 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1575 // based on the escape status of the AllocateNode.
1576 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1577 } else {
1578 insert_mem_bar(Op_MemBarCPUOrder);
1579 }
1580 } // original reexecute is set back here
1581
1582 C->set_has_split_ifs(true); // Has chance for split-if optimization
1583 if (!stopped()) {
1584 set_result(newcopy);
1585 }
1586 return true;
1587 }
1588
1589 //------------------------inline_string_getCharsU--------------------------
1590 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1591 bool LibraryCallKit::inline_string_getCharsU() {
1592 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1593 return false;
1594 }
1595
1596 // Get the arguments.
1597 Node* src = argument(0);
1598 Node* src_begin = argument(1);
1599 Node* src_end = argument(2); // exclusive offset (i < src_end)
1600 Node* dst = argument(3);
1601 Node* dst_begin = argument(4);
1602
1603 // Check for allocation before we add nodes that would confuse
1604 // tightly_coupled_allocation()
1605 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1606
1607 // Check if a null path was taken unconditionally.
1608 src = null_check(src);
1609 dst = null_check(dst);
1610 if (stopped()) {
1611 return true;
1612 }
1613
1614 // Get length and convert char[] offset to byte[] offset
1615 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1616 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1617
1618 // Range checks
1619 generate_string_range_check(src, src_begin, length, true);
1620 generate_string_range_check(dst, dst_begin, length, false);
1621 if (stopped()) {
1622 return true;
1623 }
1624
1625 if (!stopped()) {
1626 // Calculate starting addresses.
1627 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1628 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1629
1630 // Check if array addresses are aligned to HeapWordSize
1631 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1632 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1633 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1634 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1635
1636 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1637 const char* copyfunc_name = "arraycopy";
1638 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1639 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1640 OptoRuntime::fast_arraycopy_Type(),
1641 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1642 src_start, dst_start, ConvI2X(length) XTOP);
1643 // Do not let reads from the cloned object float above the arraycopy.
1644 if (alloc != NULL) {
1645 if (alloc->maybe_set_complete(&_gvn)) {
1646 // "You break it, you buy it."
1647 InitializeNode* init = alloc->initialization();
1648 assert(init->is_complete(), "we just did this");
1649 init->set_complete_with_arraycopy();
1650 assert(dst->is_CheckCastPP(), "sanity");
1651 assert(dst->in(0)->in(0) == init, "dest pinned");
1652 }
1653 // Do not let stores that initialize this object be reordered with
1654 // a subsequent store that would make this object accessible by
1655 // other threads.
1656 // Record what AllocateNode this StoreStore protects so that
1657 // escape analysis can go from the MemBarStoreStoreNode to the
1658 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1659 // based on the escape status of the AllocateNode.
1660 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1661 } else {
1662 insert_mem_bar(Op_MemBarCPUOrder);
1663 }
1664 }
1665
1666 C->set_has_split_ifs(true); // Has chance for split-if optimization
1667 return true;
1668 }
1669
1670 //----------------------inline_string_char_access----------------------------
1671 // Store/Load char to/from byte[] array.
1672 // static void StringUTF16.putChar(byte[] val, int index, int c)
1673 // static char StringUTF16.getChar(byte[] val, int index)
1674 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1675 Node* value = argument(0);
1676 Node* index = argument(1);
1677 Node* ch = is_store ? argument(2) : NULL;
1678
1679 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1680 // correctly requires matched array shapes.
1681 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1682 "sanity: byte[] and char[] bases agree");
1683 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1684 "sanity: byte[] and char[] scales agree");
1685
1686 // Bail when getChar over constants is requested: constant folding would
1687 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1688 // Java method would constant fold nicely instead.
1689 if (!is_store && value->is_Con() && index->is_Con()) {
1690 return false;
1691 }
1692
1693 Node* adr = array_element_address(value, index, T_CHAR);
1694 if (adr->is_top()) {
1695 return false;
1696 }
1697 if (is_store) {
1698 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1699 false, false, true /* mismatched */);
1700 } else {
1701 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1702 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1703 set_result(ch);
1704 }
1705 return true;
1706 }
1707
1708 //--------------------------round_double_node--------------------------------
1709 // Round a double node if necessary.
1710 Node* LibraryCallKit::round_double_node(Node* n) {
1711 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1712 n = _gvn.transform(new RoundDoubleNode(0, n));
1713 return n;
1714 }
1715
1716 //------------------------------inline_math-----------------------------------
1717 // public static double Math.abs(double)
1718 // public static double Math.sqrt(double)
1719 // public static double Math.log(double)
1720 // public static double Math.log10(double)
1721 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1722 Node* arg = round_double_node(argument(0));
1723 Node* n = NULL;
1724 switch (id) {
1725 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1726 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1727 default: fatal_unexpected_iid(id); break;
1728 }
1729 set_result(_gvn.transform(n));
1730 return true;
1731 }
1732
1733 //------------------------------runtime_math-----------------------------
1734 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1735 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1736 "must be (DD)D or (D)D type");
1737
1738 // Inputs
1739 Node* a = round_double_node(argument(0));
1740 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1741
1742 const TypePtr* no_memory_effects = NULL;
1743 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1744 no_memory_effects,
1745 a, top(), b, b ? top() : NULL);
1746 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1747 #ifdef ASSERT
1748 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1749 assert(value_top == top(), "second value must be top");
1750 #endif
1751
1752 set_result(value);
1753 return true;
1754 }
1755
1756 //------------------------------inline_math_native-----------------------------
1757 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1758 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1759 switch (id) {
1760 // These intrinsics are not properly supported on all hardware
1761 case vmIntrinsics::_dsin:
1762 return StubRoutines::dsin() != NULL ?
1763 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1764 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1765 case vmIntrinsics::_dcos:
1766 return StubRoutines::dcos() != NULL ?
1767 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1768 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1769 case vmIntrinsics::_dtan:
1770 return StubRoutines::dtan() != NULL ?
1771 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1772 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1773 case vmIntrinsics::_dlog:
1774 return StubRoutines::dlog() != NULL ?
1775 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1776 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1777 case vmIntrinsics::_dlog10:
1778 return StubRoutines::dlog10() != NULL ?
1779 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1780 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1781
1782 // These intrinsics are supported on all hardware
1783 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1784 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1785
1786 case vmIntrinsics::_dexp:
1787 return StubRoutines::dexp() != NULL ?
1788 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1789 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1790 case vmIntrinsics::_dpow:
1791 return StubRoutines::dpow() != NULL ?
1792 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1793 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1794 #undef FN_PTR
1795
1796 // These intrinsics are not yet correctly implemented
1797 case vmIntrinsics::_datan2:
1798 return false;
1799
1800 default:
1801 fatal_unexpected_iid(id);
1802 return false;
1803 }
1804 }
1805
1806 static bool is_simple_name(Node* n) {
1807 return (n->req() == 1 // constant
1808 || (n->is_Type() && n->as_Type()->type()->singleton())
1809 || n->is_Proj() // parameter or return value
1810 || n->is_Phi() // local of some sort
1811 );
1812 }
1813
1814 //----------------------------inline_notify-----------------------------------*
1815 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1816 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1817 address func;
1818 if (id == vmIntrinsics::_notify) {
1819 func = OptoRuntime::monitor_notify_Java();
1820 } else {
1821 func = OptoRuntime::monitor_notifyAll_Java();
1822 }
1823 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1824 make_slow_call_ex(call, env()->Throwable_klass(), false);
1825 return true;
1826 }
1827
1828
1829 //----------------------------inline_min_max-----------------------------------
1830 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1831 set_result(generate_min_max(id, argument(0), argument(1)));
1832 return true;
1833 }
1834
1835 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1836 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1837 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1838 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1839 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1840
1841 {
1842 PreserveJVMState pjvms(this);
1843 PreserveReexecuteState preexecs(this);
1844 jvms()->set_should_reexecute(true);
1845
1846 set_control(slow_path);
1847 set_i_o(i_o());
1848
1849 uncommon_trap(Deoptimization::Reason_intrinsic,
1850 Deoptimization::Action_none);
1851 }
1852
1853 set_control(fast_path);
1854 set_result(math);
1855 }
1856
1857 template <typename OverflowOp>
1858 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1859 typedef typename OverflowOp::MathOp MathOp;
1860
1861 MathOp* mathOp = new MathOp(arg1, arg2);
1862 Node* operation = _gvn.transform( mathOp );
1863 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1864 inline_math_mathExact(operation, ofcheck);
1865 return true;
1866 }
1867
1868 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1869 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1870 }
1871
1872 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1873 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1874 }
1875
1876 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1877 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1878 }
1879
1880 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1881 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1882 }
1883
1884 bool LibraryCallKit::inline_math_negateExactI() {
1885 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1886 }
1887
1888 bool LibraryCallKit::inline_math_negateExactL() {
1889 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1890 }
1891
1892 bool LibraryCallKit::inline_math_multiplyExactI() {
1893 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1894 }
1895
1896 bool LibraryCallKit::inline_math_multiplyExactL() {
1897 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1898 }
1899
1900 Node*
1901 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1902 // These are the candidate return value:
1903 Node* xvalue = x0;
1904 Node* yvalue = y0;
1905
1906 if (xvalue == yvalue) {
1907 return xvalue;
1908 }
1909
1910 bool want_max = (id == vmIntrinsics::_max);
1911
1912 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1913 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1914 if (txvalue == NULL || tyvalue == NULL) return top();
1915 // This is not really necessary, but it is consistent with a
1916 // hypothetical MaxINode::Value method:
1917 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1918
1919 // %%% This folding logic should (ideally) be in a different place.
1920 // Some should be inside IfNode, and there to be a more reliable
1921 // transformation of ?: style patterns into cmoves. We also want
1922 // more powerful optimizations around cmove and min/max.
1923
1924 // Try to find a dominating comparison of these guys.
1925 // It can simplify the index computation for Arrays.copyOf
1926 // and similar uses of System.arraycopy.
1927 // First, compute the normalized version of CmpI(x, y).
1928 int cmp_op = Op_CmpI;
1929 Node* xkey = xvalue;
1930 Node* ykey = yvalue;
1931 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1932 if (ideal_cmpxy->is_Cmp()) {
1933 // E.g., if we have CmpI(length - offset, count),
1934 // it might idealize to CmpI(length, count + offset)
1935 cmp_op = ideal_cmpxy->Opcode();
1936 xkey = ideal_cmpxy->in(1);
1937 ykey = ideal_cmpxy->in(2);
1938 }
1939
1940 // Start by locating any relevant comparisons.
1941 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1942 Node* cmpxy = NULL;
1943 Node* cmpyx = NULL;
1944 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1945 Node* cmp = start_from->fast_out(k);
1946 if (cmp->outcnt() > 0 && // must have prior uses
1947 cmp->in(0) == NULL && // must be context-independent
1948 cmp->Opcode() == cmp_op) { // right kind of compare
1949 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1950 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1951 }
1952 }
1953
1954 const int NCMPS = 2;
1955 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1956 int cmpn;
1957 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1958 if (cmps[cmpn] != NULL) break; // find a result
1959 }
1960 if (cmpn < NCMPS) {
1961 // Look for a dominating test that tells us the min and max.
1962 int depth = 0; // Limit search depth for speed
1963 Node* dom = control();
1964 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1965 if (++depth >= 100) break;
1966 Node* ifproj = dom;
1967 if (!ifproj->is_Proj()) continue;
1968 Node* iff = ifproj->in(0);
1969 if (!iff->is_If()) continue;
1970 Node* bol = iff->in(1);
1971 if (!bol->is_Bool()) continue;
1972 Node* cmp = bol->in(1);
1973 if (cmp == NULL) continue;
1974 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1975 if (cmps[cmpn] == cmp) break;
1976 if (cmpn == NCMPS) continue;
1977 BoolTest::mask btest = bol->as_Bool()->_test._test;
1978 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1979 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1980 // At this point, we know that 'x btest y' is true.
1981 switch (btest) {
1982 case BoolTest::eq:
1983 // They are proven equal, so we can collapse the min/max.
1984 // Either value is the answer. Choose the simpler.
1985 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1986 return yvalue;
1987 return xvalue;
1988 case BoolTest::lt: // x < y
1989 case BoolTest::le: // x <= y
1990 return (want_max ? yvalue : xvalue);
1991 case BoolTest::gt: // x > y
1992 case BoolTest::ge: // x >= y
1993 return (want_max ? xvalue : yvalue);
1994 default:
1995 break;
1996 }
1997 }
1998 }
1999
2000 // We failed to find a dominating test.
2001 // Let's pick a test that might GVN with prior tests.
2002 Node* best_bol = NULL;
2003 BoolTest::mask best_btest = BoolTest::illegal;
2004 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2005 Node* cmp = cmps[cmpn];
2006 if (cmp == NULL) continue;
2007 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2008 Node* bol = cmp->fast_out(j);
2009 if (!bol->is_Bool()) continue;
2010 BoolTest::mask btest = bol->as_Bool()->_test._test;
2011 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2012 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2013 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2014 best_bol = bol->as_Bool();
2015 best_btest = btest;
2016 }
2017 }
2018 }
2019
2020 Node* answer_if_true = NULL;
2021 Node* answer_if_false = NULL;
2022 switch (best_btest) {
2023 default:
2024 if (cmpxy == NULL)
2025 cmpxy = ideal_cmpxy;
2026 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2027 // and fall through:
2028 case BoolTest::lt: // x < y
2029 case BoolTest::le: // x <= y
2030 answer_if_true = (want_max ? yvalue : xvalue);
2031 answer_if_false = (want_max ? xvalue : yvalue);
2032 break;
2033 case BoolTest::gt: // x > y
2034 case BoolTest::ge: // x >= y
2035 answer_if_true = (want_max ? xvalue : yvalue);
2036 answer_if_false = (want_max ? yvalue : xvalue);
2037 break;
2038 }
2039
2040 jint hi, lo;
2041 if (want_max) {
2042 // We can sharpen the minimum.
2043 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2044 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2045 } else {
2046 // We can sharpen the maximum.
2047 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2048 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2049 }
2050
2051 // Use a flow-free graph structure, to avoid creating excess control edges
2052 // which could hinder other optimizations.
2053 // Since Math.min/max is often used with arraycopy, we want
2054 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2055 Node* cmov = CMoveNode::make(NULL, best_bol,
2056 answer_if_false, answer_if_true,
2057 TypeInt::make(lo, hi, widen));
2058
2059 return _gvn.transform(cmov);
2060
2061 /*
2062 // This is not as desirable as it may seem, since Min and Max
2063 // nodes do not have a full set of optimizations.
2064 // And they would interfere, anyway, with 'if' optimizations
2065 // and with CMoveI canonical forms.
2066 switch (id) {
2067 case vmIntrinsics::_min:
2068 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2069 case vmIntrinsics::_max:
2070 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2071 default:
2072 ShouldNotReachHere();
2073 }
2074 */
2075 }
2076
2077 inline int
2078 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2079 const TypePtr* base_type = TypePtr::NULL_PTR;
2080 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2081 if (base_type == NULL) {
2082 // Unknown type.
2083 return Type::AnyPtr;
2084 } else if (base_type == TypePtr::NULL_PTR) {
2085 // Since this is a NULL+long form, we have to switch to a rawptr.
2086 base = _gvn.transform(new CastX2PNode(offset));
2087 offset = MakeConX(0);
2088 return Type::RawPtr;
2089 } else if (base_type->base() == Type::RawPtr) {
2090 return Type::RawPtr;
2091 } else if (base_type->isa_oopptr()) {
2092 // Base is never null => always a heap address.
2093 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2094 return Type::OopPtr;
2095 }
2096 // Offset is small => always a heap address.
2097 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2098 if (offset_type != NULL &&
2099 base_type->offset() == 0 && // (should always be?)
2100 offset_type->_lo >= 0 &&
2101 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2102 return Type::OopPtr;
2103 } else if (type == T_OBJECT) {
2104 // off heap access to an oop doesn't make any sense. Has to be on
2105 // heap.
2106 return Type::OopPtr;
2107 }
2108 // Otherwise, it might either be oop+off or NULL+addr.
2109 return Type::AnyPtr;
2110 } else {
2111 // No information:
2112 return Type::AnyPtr;
2113 }
2114 }
2115
2116 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2117 Node* uncasted_base = base;
2118 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2119 if (kind == Type::RawPtr) {
2120 return basic_plus_adr(top(), uncasted_base, offset);
2121 } else if (kind == Type::AnyPtr) {
2122 assert(base == uncasted_base, "unexpected base change");
2123 if (can_cast) {
2124 if (!_gvn.type(base)->speculative_maybe_null() &&
2125 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2126 // According to profiling, this access is always on
2127 // heap. Casting the base to not null and thus avoiding membars
2128 // around the access should allow better optimizations
2129 Node* null_ctl = top();
2130 base = null_check_oop(base, &null_ctl, true, true, true);
2131 assert(null_ctl->is_top(), "no null control here");
2132 return basic_plus_adr(base, offset);
2133 } else if (_gvn.type(base)->speculative_always_null() &&
2134 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2135 // According to profiling, this access is always off
2136 // heap.
2137 base = null_assert(base);
2138 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2139 offset = MakeConX(0);
2140 return basic_plus_adr(top(), raw_base, offset);
2141 }
2142 }
2143 // We don't know if it's an on heap or off heap access. Fall back
2144 // to raw memory access.
2145 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2146 return basic_plus_adr(top(), raw, offset);
2147 } else {
2148 assert(base == uncasted_base, "unexpected base change");
2149 // We know it's an on heap access so base can't be null
2150 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2151 base = must_be_not_null(base, true);
2152 }
2153 return basic_plus_adr(base, offset);
2154 }
2155 }
2156
2157 //--------------------------inline_number_methods-----------------------------
2158 // inline int Integer.numberOfLeadingZeros(int)
2159 // inline int Long.numberOfLeadingZeros(long)
2160 //
2161 // inline int Integer.numberOfTrailingZeros(int)
2162 // inline int Long.numberOfTrailingZeros(long)
2163 //
2164 // inline int Integer.bitCount(int)
2165 // inline int Long.bitCount(long)
2166 //
2167 // inline char Character.reverseBytes(char)
2168 // inline short Short.reverseBytes(short)
2169 // inline int Integer.reverseBytes(int)
2170 // inline long Long.reverseBytes(long)
2171 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2172 Node* arg = argument(0);
2173 Node* n = NULL;
2174 switch (id) {
2175 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2176 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2177 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2178 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2179 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2180 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2181 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2182 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2183 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2184 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2185 default: fatal_unexpected_iid(id); break;
2186 }
2187 set_result(_gvn.transform(n));
2188 return true;
2189 }
2190
2191 //----------------------------inline_unsafe_access----------------------------
2192
2193 // Helper that guards and inserts a pre-barrier.
2194 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2195 Node* pre_val, bool need_mem_bar) {
2196 // We could be accessing the referent field of a reference object. If so, when G1
2197 // is enabled, we need to log the value in the referent field in an SATB buffer.
2198 // This routine performs some compile time filters and generates suitable
2199 // runtime filters that guard the pre-barrier code.
2200 // Also add memory barrier for non volatile load from the referent field
2201 // to prevent commoning of loads across safepoint.
2202 if (!UseG1GC && !need_mem_bar)
2203 return;
2204
2205 // Some compile time checks.
2206
2207 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2208 const TypeX* otype = offset->find_intptr_t_type();
2209 if (otype != NULL && otype->is_con() &&
2210 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2211 // Constant offset but not the reference_offset so just return
2212 return;
2213 }
2214
2215 // We only need to generate the runtime guards for instances.
2216 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2217 if (btype != NULL) {
2218 if (btype->isa_aryptr()) {
2219 // Array type so nothing to do
2220 return;
2221 }
2222
2223 const TypeInstPtr* itype = btype->isa_instptr();
2224 if (itype != NULL) {
2225 // Can the klass of base_oop be statically determined to be
2226 // _not_ a sub-class of Reference and _not_ Object?
2227 ciKlass* klass = itype->klass();
2228 if ( klass->is_loaded() &&
2229 !klass->is_subtype_of(env()->Reference_klass()) &&
2230 !env()->Object_klass()->is_subtype_of(klass)) {
2231 return;
2232 }
2233 }
2234 }
2235
2236 // The compile time filters did not reject base_oop/offset so
2237 // we need to generate the following runtime filters
2238 //
2239 // if (offset == java_lang_ref_Reference::_reference_offset) {
2240 // if (instance_of(base, java.lang.ref.Reference)) {
2241 // pre_barrier(_, pre_val, ...);
2242 // }
2243 // }
2244
2245 float likely = PROB_LIKELY( 0.999);
2246 float unlikely = PROB_UNLIKELY(0.999);
2247
2248 IdealKit ideal(this);
2249 #define __ ideal.
2250
2251 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2252
2253 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2254 // Update graphKit memory and control from IdealKit.
2255 sync_kit(ideal);
2256
2257 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2258 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2259
2260 // Update IdealKit memory and control from graphKit.
2261 __ sync_kit(this);
2262
2263 Node* one = __ ConI(1);
2264 // is_instof == 0 if base_oop == NULL
2265 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2266
2267 // Update graphKit from IdeakKit.
2268 sync_kit(ideal);
2269
2270 // Use the pre-barrier to record the value in the referent field
2271 pre_barrier(false /* do_load */,
2272 __ ctrl(),
2273 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2274 pre_val /* pre_val */,
2275 T_OBJECT);
2276 if (need_mem_bar) {
2277 // Add memory barrier to prevent commoning reads from this field
2278 // across safepoint since GC can change its value.
2279 insert_mem_bar(Op_MemBarCPUOrder);
2280 }
2281 // Update IdealKit from graphKit.
2282 __ sync_kit(this);
2283
2284 } __ end_if(); // _ref_type != ref_none
2285 } __ end_if(); // offset == referent_offset
2286
2287 // Final sync IdealKit and GraphKit.
2288 final_sync(ideal);
2289 #undef __
2290 }
2291
2292
2293 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2294 // Attempt to infer a sharper value type from the offset and base type.
2295 ciKlass* sharpened_klass = NULL;
2296
2297 // See if it is an instance field, with an object type.
2298 if (alias_type->field() != NULL) {
2299 if (alias_type->field()->type()->is_klass()) {
2300 sharpened_klass = alias_type->field()->type()->as_klass();
2301 }
2302 }
2303
2304 // See if it is a narrow oop array.
2305 if (adr_type->isa_aryptr()) {
2306 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2307 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2308 if (elem_type != NULL) {
2309 sharpened_klass = elem_type->klass();
2310 }
2311 }
2312 }
2313
2314 // The sharpened class might be unloaded if there is no class loader
2315 // contraint in place.
2316 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2317 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2318
2319 #ifndef PRODUCT
2320 if (C->print_intrinsics() || C->print_inlining()) {
2321 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2322 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2323 }
2324 #endif
2325 // Sharpen the value type.
2326 return tjp;
2327 }
2328 return NULL;
2329 }
2330
2331 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2332 if (callee()->is_static()) return false; // caller must have the capability!
2333 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2334 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2335 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2336
2337 #ifndef PRODUCT
2338 {
2339 ResourceMark rm;
2340 // Check the signatures.
2341 ciSignature* sig = callee()->signature();
2342 #ifdef ASSERT
2343 if (!is_store) {
2344 // Object getObject(Object base, int/long offset), etc.
2345 BasicType rtype = sig->return_type()->basic_type();
2346 assert(rtype == type, "getter must return the expected value");
2347 assert(sig->count() == 2, "oop getter has 2 arguments");
2348 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2349 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2350 } else {
2351 // void putObject(Object base, int/long offset, Object x), etc.
2352 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2353 assert(sig->count() == 3, "oop putter has 3 arguments");
2354 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2355 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2356 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2357 assert(vtype == type, "putter must accept the expected value");
2358 }
2359 #endif // ASSERT
2360 }
2361 #endif //PRODUCT
2362
2363 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2364
2365 Node* receiver = argument(0); // type: oop
2366
2367 // Build address expression.
2368 Node* adr;
2369 Node* heap_base_oop = top();
2370 Node* offset = top();
2371 Node* val;
2372
2373 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2374 Node* base = argument(1); // type: oop
2375 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2376 offset = argument(2); // type: long
2377 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2378 // to be plain byte offsets, which are also the same as those accepted
2379 // by oopDesc::field_base.
2380 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2381 "fieldOffset must be byte-scaled");
2382 // 32-bit machines ignore the high half!
2383 offset = ConvL2X(offset);
2384 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2385
2386 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2387 heap_base_oop = base;
2388 } else if (type == T_OBJECT) {
2389 return false; // off-heap oop accesses are not supported
2390 }
2391
2392 // Can base be NULL? Otherwise, always on-heap access.
2393 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2394
2395 val = is_store ? argument(4) : NULL;
2396
2397 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2398
2399 // Try to categorize the address.
2400 Compile::AliasType* alias_type = C->alias_type(adr_type);
2401 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2402
2403 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2404 alias_type->adr_type() == TypeAryPtr::RANGE) {
2405 return false; // not supported
2406 }
2407
2408 bool mismatched = false;
2409 BasicType bt = alias_type->basic_type();
2410 if (bt != T_ILLEGAL) {
2411 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2412 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2413 // Alias type doesn't differentiate between byte[] and boolean[]).
2414 // Use address type to get the element type.
2415 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2416 }
2417 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2418 // accessing an array field with getObject is not a mismatch
2419 bt = T_OBJECT;
2420 }
2421 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2422 // Don't intrinsify mismatched object accesses
2423 return false;
2424 }
2425 mismatched = (bt != type);
2426 } else if (alias_type->adr_type()->isa_oopptr()) {
2427 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2428 }
2429
2430 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2431
2432 // First guess at the value type.
2433 const Type *value_type = Type::get_const_basic_type(type);
2434
2435 // We will need memory barriers unless we can determine a unique
2436 // alias category for this reference. (Note: If for some reason
2437 // the barriers get omitted and the unsafe reference begins to "pollute"
2438 // the alias analysis of the rest of the graph, either Compile::can_alias
2439 // or Compile::must_alias will throw a diagnostic assert.)
2440 bool need_mem_bar = false;
2441 switch (kind) {
2442 case Relaxed:
2443 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2444 break;
2445 case Opaque:
2446 // Opaque uses CPUOrder membars for protection against code movement.
2447 case Acquire:
2448 case Release:
2449 case Volatile:
2450 need_mem_bar = true;
2451 break;
2452 default:
2453 ShouldNotReachHere();
2454 }
2455
2456 // Some accesses require access atomicity for all types, notably longs and doubles.
2457 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2458 bool requires_atomic_access = false;
2459 switch (kind) {
2460 case Relaxed:
2461 requires_atomic_access = AlwaysAtomicAccesses;
2462 break;
2463 case Opaque:
2464 // Opaque accesses are atomic.
2465 case Acquire:
2466 case Release:
2467 case Volatile:
2468 requires_atomic_access = true;
2469 break;
2470 default:
2471 ShouldNotReachHere();
2472 }
2473
2474 // Figure out the memory ordering.
2475 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2476 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2477
2478 // If we are reading the value of the referent field of a Reference
2479 // object (either by using Unsafe directly or through reflection)
2480 // then, if G1 is enabled, we need to record the referent in an
2481 // SATB log buffer using the pre-barrier mechanism.
2482 // Also we need to add memory barrier to prevent commoning reads
2483 // from this field across safepoint since GC can change its value.
2484 bool need_read_barrier = !is_store &&
2485 offset != top() && heap_base_oop != top();
2486
2487 if (!is_store && type == T_OBJECT) {
2488 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2489 if (tjp != NULL) {
2490 value_type = tjp;
2491 }
2492 }
2493
2494 receiver = null_check(receiver);
2495 if (stopped()) {
2496 return true;
2497 }
2498 // Heap pointers get a null-check from the interpreter,
2499 // as a courtesy. However, this is not guaranteed by Unsafe,
2500 // and it is not possible to fully distinguish unintended nulls
2501 // from intended ones in this API.
2502
2503 // We need to emit leading and trailing CPU membars (see below) in
2504 // addition to memory membars for special access modes. This is a little
2505 // too strong, but avoids the need to insert per-alias-type
2506 // volatile membars (for stores; compare Parse::do_put_xxx), which
2507 // we cannot do effectively here because we probably only have a
2508 // rough approximation of type.
2509
2510 switch(kind) {
2511 case Relaxed:
2512 case Opaque:
2513 case Acquire:
2514 break;
2515 case Release:
2516 case Volatile:
2517 if (is_store) {
2518 insert_mem_bar(Op_MemBarRelease);
2519 } else {
2520 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2521 insert_mem_bar(Op_MemBarVolatile);
2522 }
2523 }
2524 break;
2525 default:
2526 ShouldNotReachHere();
2527 }
2528
2529 // Memory barrier to prevent normal and 'unsafe' accesses from
2530 // bypassing each other. Happens after null checks, so the
2531 // exception paths do not take memory state from the memory barrier,
2532 // so there's no problems making a strong assert about mixing users
2533 // of safe & unsafe memory.
2534 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2535
2536 if (!is_store) {
2537 Node* p = NULL;
2538 // Try to constant fold a load from a constant field
2539 ciField* field = alias_type->field();
2540 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2541 // final or stable field
2542 p = make_constant_from_field(field, heap_base_oop);
2543 }
2544 if (p == NULL) {
2545 // To be valid, unsafe loads may depend on other conditions than
2546 // the one that guards them: pin the Load node
2547 LoadNode::ControlDependency dep = LoadNode::Pinned;
2548 Node* ctrl = control();
2549 if (adr_type->isa_instptr()) {
2550 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2551 intptr_t offset = Type::OffsetBot;
2552 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2553 if (offset >= 0) {
2554 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2555 if (offset < s) {
2556 // Guaranteed to be a valid access, no need to pin it
2557 dep = LoadNode::DependsOnlyOnTest;
2558 ctrl = NULL;
2559 }
2560 }
2561 }
2562 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2563 // load value
2564 switch (type) {
2565 case T_BOOLEAN:
2566 {
2567 // Normalize the value returned by getBoolean in the following cases
2568 if (mismatched ||
2569 heap_base_oop == top() || // - heap_base_oop is NULL or
2570 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2571 // and the unsafe access is made to large offset
2572 // (i.e., larger than the maximum offset necessary for any
2573 // field access)
2574 ) {
2575 IdealKit ideal = IdealKit(this);
2576 #define __ ideal.
2577 IdealVariable normalized_result(ideal);
2578 __ declarations_done();
2579 __ set(normalized_result, p);
2580 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2581 __ set(normalized_result, ideal.ConI(1));
2582 ideal.end_if();
2583 final_sync(ideal);
2584 p = __ value(normalized_result);
2585 #undef __
2586 }
2587 }
2588 case T_CHAR:
2589 case T_BYTE:
2590 case T_SHORT:
2591 case T_INT:
2592 case T_LONG:
2593 case T_FLOAT:
2594 case T_DOUBLE:
2595 break;
2596 case T_OBJECT:
2597 if (need_read_barrier) {
2598 // We do not require a mem bar inside pre_barrier if need_mem_bar
2599 // is set: the barriers would be emitted by us.
2600 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2601 }
2602 break;
2603 case T_ADDRESS:
2604 // Cast to an int type.
2605 p = _gvn.transform(new CastP2XNode(NULL, p));
2606 p = ConvX2UL(p);
2607 break;
2608 default:
2609 fatal("unexpected type %d: %s", type, type2name(type));
2610 break;
2611 }
2612 }
2613 // The load node has the control of the preceding MemBarCPUOrder. All
2614 // following nodes will have the control of the MemBarCPUOrder inserted at
2615 // the end of this method. So, pushing the load onto the stack at a later
2616 // point is fine.
2617 set_result(p);
2618 } else {
2619 // place effect of store into memory
2620 switch (type) {
2621 case T_DOUBLE:
2622 val = dstore_rounding(val);
2623 break;
2624 case T_ADDRESS:
2625 // Repackage the long as a pointer.
2626 val = ConvL2X(val);
2627 val = _gvn.transform(new CastX2PNode(val));
2628 break;
2629 default:
2630 break;
2631 }
2632
2633 if (type == T_OBJECT) {
2634 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2635 } else {
2636 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2637 }
2638 }
2639
2640 switch(kind) {
2641 case Relaxed:
2642 case Opaque:
2643 case Release:
2644 break;
2645 case Acquire:
2646 case Volatile:
2647 if (!is_store) {
2648 insert_mem_bar(Op_MemBarAcquire);
2649 } else {
2650 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2651 insert_mem_bar(Op_MemBarVolatile);
2652 }
2653 }
2654 break;
2655 default:
2656 ShouldNotReachHere();
2657 }
2658
2659 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2660
2661 return true;
2662 }
2663
2664 //----------------------------inline_unsafe_load_store----------------------------
2665 // This method serves a couple of different customers (depending on LoadStoreKind):
2666 //
2667 // LS_cmp_swap:
2668 //
2669 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2670 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2671 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2672 //
2673 // LS_cmp_swap_weak:
2674 //
2675 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2676 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2677 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2678 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2679 //
2680 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2681 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2682 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2683 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2684 //
2685 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2686 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2687 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2688 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2689 //
2690 // LS_cmp_exchange:
2691 //
2692 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2693 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2694 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2695 //
2696 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2697 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2698 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2699 //
2700 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2701 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2702 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2703 //
2704 // LS_get_add:
2705 //
2706 // int getAndAddInt( Object o, long offset, int delta)
2707 // long getAndAddLong(Object o, long offset, long delta)
2708 //
2709 // LS_get_set:
2710 //
2711 // int getAndSet(Object o, long offset, int newValue)
2712 // long getAndSet(Object o, long offset, long newValue)
2713 // Object getAndSet(Object o, long offset, Object newValue)
2714 //
2715 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2716 // This basic scheme here is the same as inline_unsafe_access, but
2717 // differs in enough details that combining them would make the code
2718 // overly confusing. (This is a true fact! I originally combined
2719 // them, but even I was confused by it!) As much code/comments as
2720 // possible are retained from inline_unsafe_access though to make
2721 // the correspondences clearer. - dl
2722
2723 if (callee()->is_static()) return false; // caller must have the capability!
2724
2725 #ifndef PRODUCT
2726 BasicType rtype;
2727 {
2728 ResourceMark rm;
2729 // Check the signatures.
2730 ciSignature* sig = callee()->signature();
2731 rtype = sig->return_type()->basic_type();
2732 switch(kind) {
2733 case LS_get_add:
2734 case LS_get_set: {
2735 // Check the signatures.
2736 #ifdef ASSERT
2737 assert(rtype == type, "get and set must return the expected type");
2738 assert(sig->count() == 3, "get and set has 3 arguments");
2739 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2740 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2741 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2742 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2743 #endif // ASSERT
2744 break;
2745 }
2746 case LS_cmp_swap:
2747 case LS_cmp_swap_weak: {
2748 // Check the signatures.
2749 #ifdef ASSERT
2750 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2751 assert(sig->count() == 4, "CAS has 4 arguments");
2752 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2753 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2754 #endif // ASSERT
2755 break;
2756 }
2757 case LS_cmp_exchange: {
2758 // Check the signatures.
2759 #ifdef ASSERT
2760 assert(rtype == type, "CAS must return the expected type");
2761 assert(sig->count() == 4, "CAS has 4 arguments");
2762 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2763 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2764 #endif // ASSERT
2765 break;
2766 }
2767 default:
2768 ShouldNotReachHere();
2769 }
2770 }
2771 #endif //PRODUCT
2772
2773 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2774
2775 // Get arguments:
2776 Node* receiver = NULL;
2777 Node* base = NULL;
2778 Node* offset = NULL;
2779 Node* oldval = NULL;
2780 Node* newval = NULL;
2781 switch(kind) {
2782 case LS_cmp_swap:
2783 case LS_cmp_swap_weak:
2784 case LS_cmp_exchange: {
2785 const bool two_slot_type = type2size[type] == 2;
2786 receiver = argument(0); // type: oop
2787 base = argument(1); // type: oop
2788 offset = argument(2); // type: long
2789 oldval = argument(4); // type: oop, int, or long
2790 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2791 break;
2792 }
2793 case LS_get_add:
2794 case LS_get_set: {
2795 receiver = argument(0); // type: oop
2796 base = argument(1); // type: oop
2797 offset = argument(2); // type: long
2798 oldval = NULL;
2799 newval = argument(4); // type: oop, int, or long
2800 break;
2801 }
2802 default:
2803 ShouldNotReachHere();
2804 }
2805
2806 // Build field offset expression.
2807 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2808 // to be plain byte offsets, which are also the same as those accepted
2809 // by oopDesc::field_base.
2810 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2811 // 32-bit machines ignore the high half of long offsets
2812 offset = ConvL2X(offset);
2813 Node* adr = make_unsafe_address(base, offset, type, false);
2814 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2815
2816 Compile::AliasType* alias_type = C->alias_type(adr_type);
2817 BasicType bt = alias_type->basic_type();
2818 if (bt != T_ILLEGAL &&
2819 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2820 // Don't intrinsify mismatched object accesses.
2821 return false;
2822 }
2823
2824 // For CAS, unlike inline_unsafe_access, there seems no point in
2825 // trying to refine types. Just use the coarse types here.
2826 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2827 const Type *value_type = Type::get_const_basic_type(type);
2828
2829 switch (kind) {
2830 case LS_get_set:
2831 case LS_cmp_exchange: {
2832 if (type == T_OBJECT) {
2833 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2834 if (tjp != NULL) {
2835 value_type = tjp;
2836 }
2837 }
2838 break;
2839 }
2840 case LS_cmp_swap:
2841 case LS_cmp_swap_weak:
2842 case LS_get_add:
2843 break;
2844 default:
2845 ShouldNotReachHere();
2846 }
2847
2848 // Null check receiver.
2849 receiver = null_check(receiver);
2850 if (stopped()) {
2851 return true;
2852 }
2853
2854 int alias_idx = C->get_alias_index(adr_type);
2855
2856 // Memory-model-wise, a LoadStore acts like a little synchronized
2857 // block, so needs barriers on each side. These don't translate
2858 // into actual barriers on most machines, but we still need rest of
2859 // compiler to respect ordering.
2860
2861 switch (access_kind) {
2862 case Relaxed:
2863 case Acquire:
2864 break;
2865 case Release:
2866 insert_mem_bar(Op_MemBarRelease);
2867 break;
2868 case Volatile:
2869 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2870 insert_mem_bar(Op_MemBarVolatile);
2871 } else {
2872 insert_mem_bar(Op_MemBarRelease);
2873 }
2874 break;
2875 default:
2876 ShouldNotReachHere();
2877 }
2878 insert_mem_bar(Op_MemBarCPUOrder);
2879
2880 // Figure out the memory ordering.
2881 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2882
2883 // 4984716: MemBars must be inserted before this
2884 // memory node in order to avoid a false
2885 // dependency which will confuse the scheduler.
2886 Node *mem = memory(alias_idx);
2887
2888 // For now, we handle only those cases that actually exist: ints,
2889 // longs, and Object. Adding others should be straightforward.
2890 Node* load_store = NULL;
2891 switch(type) {
2892 case T_BYTE:
2893 switch(kind) {
2894 case LS_get_add:
2895 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2896 break;
2897 case LS_get_set:
2898 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2899 break;
2900 case LS_cmp_swap_weak:
2901 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2902 break;
2903 case LS_cmp_swap:
2904 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2905 break;
2906 case LS_cmp_exchange:
2907 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2908 break;
2909 default:
2910 ShouldNotReachHere();
2911 }
2912 break;
2913 case T_SHORT:
2914 switch(kind) {
2915 case LS_get_add:
2916 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2917 break;
2918 case LS_get_set:
2919 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2920 break;
2921 case LS_cmp_swap_weak:
2922 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2923 break;
2924 case LS_cmp_swap:
2925 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2926 break;
2927 case LS_cmp_exchange:
2928 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2929 break;
2930 default:
2931 ShouldNotReachHere();
2932 }
2933 break;
2934 case T_INT:
2935 switch(kind) {
2936 case LS_get_add:
2937 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2938 break;
2939 case LS_get_set:
2940 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2941 break;
2942 case LS_cmp_swap_weak:
2943 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2944 break;
2945 case LS_cmp_swap:
2946 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2947 break;
2948 case LS_cmp_exchange:
2949 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2950 break;
2951 default:
2952 ShouldNotReachHere();
2953 }
2954 break;
2955 case T_LONG:
2956 switch(kind) {
2957 case LS_get_add:
2958 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2959 break;
2960 case LS_get_set:
2961 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2962 break;
2963 case LS_cmp_swap_weak:
2964 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2965 break;
2966 case LS_cmp_swap:
2967 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2968 break;
2969 case LS_cmp_exchange:
2970 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2971 break;
2972 default:
2973 ShouldNotReachHere();
2974 }
2975 break;
2976 case T_OBJECT:
2977 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2978 // could be delayed during Parse (for example, in adjust_map_after_if()).
2979 // Execute transformation here to avoid barrier generation in such case.
2980 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2981 newval = _gvn.makecon(TypePtr::NULL_PTR);
2982
2983 // Reference stores need a store barrier.
2984 switch(kind) {
2985 case LS_get_set: {
2986 // If pre-barrier must execute before the oop store, old value will require do_load here.
2987 if (!can_move_pre_barrier()) {
2988 pre_barrier(true /* do_load*/,
2989 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2990 NULL /* pre_val*/,
2991 T_OBJECT);
2992 } // Else move pre_barrier to use load_store value, see below.
2993 break;
2994 }
2995 case LS_cmp_swap_weak:
2996 case LS_cmp_swap:
2997 case LS_cmp_exchange: {
2998 // Same as for newval above:
2999 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3000 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3001 }
3002 // The only known value which might get overwritten is oldval.
3003 pre_barrier(false /* do_load */,
3004 control(), NULL, NULL, max_juint, NULL, NULL,
3005 oldval /* pre_val */,
3006 T_OBJECT);
3007 break;
3008 }
3009 default:
3010 ShouldNotReachHere();
3011 }
3012
3013 #ifdef _LP64
3014 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3015 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3016
3017 switch(kind) {
3018 case LS_get_set:
3019 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3020 break;
3021 case LS_cmp_swap_weak: {
3022 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3023 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3024 break;
3025 }
3026 case LS_cmp_swap: {
3027 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3028 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3029 break;
3030 }
3031 case LS_cmp_exchange: {
3032 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3033 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3034 break;
3035 }
3036 default:
3037 ShouldNotReachHere();
3038 }
3039 } else
3040 #endif
3041 switch (kind) {
3042 case LS_get_set:
3043 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3044 break;
3045 case LS_cmp_swap_weak:
3046 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3047 break;
3048 case LS_cmp_swap:
3049 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3050 break;
3051 case LS_cmp_exchange:
3052 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3053 break;
3054 default:
3055 ShouldNotReachHere();
3056 }
3057
3058 // Emit the post barrier only when the actual store happened. This makes sense
3059 // to check only for LS_cmp_* that can fail to set the value.
3060 // LS_cmp_exchange does not produce any branches by default, so there is no
3061 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3062 // CompareAndExchange and move branches here, it would make sense to conditionalize
3063 // post_barriers for LS_cmp_exchange as well.
3064 //
3065 // CAS success path is marked more likely since we anticipate this is a performance
3066 // critical path, while CAS failure path can use the penalty for going through unlikely
3067 // path as backoff. Which is still better than doing a store barrier there.
3068 switch (kind) {
3069 case LS_get_set:
3070 case LS_cmp_exchange: {
3071 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3072 break;
3073 }
3074 case LS_cmp_swap_weak:
3075 case LS_cmp_swap: {
3076 IdealKit ideal(this);
3077 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3078 sync_kit(ideal);
3079 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3080 ideal.sync_kit(this);
3081 } ideal.end_if();
3082 final_sync(ideal);
3083 break;
3084 }
3085 default:
3086 ShouldNotReachHere();
3087 }
3088 break;
3089 default:
3090 fatal("unexpected type %d: %s", type, type2name(type));
3091 break;
3092 }
3093
3094 // SCMemProjNodes represent the memory state of a LoadStore. Their
3095 // main role is to prevent LoadStore nodes from being optimized away
3096 // when their results aren't used.
3097 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3098 set_memory(proj, alias_idx);
3099
3100 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3101 #ifdef _LP64
3102 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3103 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3104 }
3105 #endif
3106 if (can_move_pre_barrier() && kind == LS_get_set) {
3107 // Don't need to load pre_val. The old value is returned by load_store.
3108 // The pre_barrier can execute after the xchg as long as no safepoint
3109 // gets inserted between them.
3110 pre_barrier(false /* do_load */,
3111 control(), NULL, NULL, max_juint, NULL, NULL,
3112 load_store /* pre_val */,
3113 T_OBJECT);
3114 }
3115 }
3116
3117 // Add the trailing membar surrounding the access
3118 insert_mem_bar(Op_MemBarCPUOrder);
3119
3120 switch (access_kind) {
3121 case Relaxed:
3122 case Release:
3123 break; // do nothing
3124 case Acquire:
3125 case Volatile:
3126 insert_mem_bar(Op_MemBarAcquire);
3127 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3128 break;
3129 default:
3130 ShouldNotReachHere();
3131 }
3132
3133 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3134 set_result(load_store);
3135 return true;
3136 }
3137
3138 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3139 MemNode::MemOrd mo = MemNode::unset;
3140 switch(kind) {
3141 case Opaque:
3142 case Relaxed: mo = MemNode::unordered; break;
3143 case Acquire: mo = MemNode::acquire; break;
3144 case Release: mo = MemNode::release; break;
3145 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3146 default:
3147 ShouldNotReachHere();
3148 }
3149 guarantee(mo != MemNode::unset, "Should select memory ordering");
3150 return mo;
3151 }
3152
3153 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3154 MemNode::MemOrd mo = MemNode::unset;
3155 switch(kind) {
3156 case Opaque:
3157 case Relaxed: mo = MemNode::unordered; break;
3158 case Acquire: mo = MemNode::acquire; break;
3159 case Release: mo = MemNode::release; break;
3160 case Volatile: mo = MemNode::seqcst; break;
3161 default:
3162 ShouldNotReachHere();
3163 }
3164 guarantee(mo != MemNode::unset, "Should select memory ordering");
3165 return mo;
3166 }
3167
3168 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3169 // Regardless of form, don't allow previous ld/st to move down,
3170 // then issue acquire, release, or volatile mem_bar.
3171 insert_mem_bar(Op_MemBarCPUOrder);
3172 switch(id) {
3173 case vmIntrinsics::_loadFence:
3174 insert_mem_bar(Op_LoadFence);
3175 return true;
3176 case vmIntrinsics::_storeFence:
3177 insert_mem_bar(Op_StoreFence);
3178 return true;
3179 case vmIntrinsics::_fullFence:
3180 insert_mem_bar(Op_MemBarVolatile);
3181 return true;
3182 default:
3183 fatal_unexpected_iid(id);
3184 return false;
3185 }
3186 }
3187
3188 bool LibraryCallKit::inline_onspinwait() {
3189 insert_mem_bar(Op_OnSpinWait);
3190 return true;
3191 }
3192
3193 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3194 if (!kls->is_Con()) {
3195 return true;
3196 }
3197 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3198 if (klsptr == NULL) {
3199 return true;
3200 }
3201 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3202 // don't need a guard for a klass that is already initialized
3203 return !ik->is_initialized();
3204 }
3205
3206 //----------------------------inline_unsafe_allocate---------------------------
3207 // public native Object Unsafe.allocateInstance(Class<?> cls);
3208 bool LibraryCallKit::inline_unsafe_allocate() {
3209 if (callee()->is_static()) return false; // caller must have the capability!
3210
3211 null_check_receiver(); // null-check, then ignore
3212 Node* cls = null_check(argument(1));
3213 if (stopped()) return true;
3214
3215 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3216 kls = null_check(kls);
3217 if (stopped()) return true; // argument was like int.class
3218
3219 Node* test = NULL;
3220 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3221 // Note: The argument might still be an illegal value like
3222 // Serializable.class or Object[].class. The runtime will handle it.
3223 // But we must make an explicit check for initialization.
3224 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3225 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3226 // can generate code to load it as unsigned byte.
3227 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3228 Node* bits = intcon(InstanceKlass::fully_initialized);
3229 test = _gvn.transform(new SubINode(inst, bits));
3230 // The 'test' is non-zero if we need to take a slow path.
3231 }
3232
3233 Node* obj = new_instance(kls, test);
3234 set_result(obj);
3235 return true;
3236 }
3237
3238 //------------------------inline_native_time_funcs--------------
3239 // inline code for System.currentTimeMillis() and System.nanoTime()
3240 // these have the same type and signature
3241 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3242 const TypeFunc* tf = OptoRuntime::void_long_Type();
3243 const TypePtr* no_memory_effects = NULL;
3244 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3245 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3246 #ifdef ASSERT
3247 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3248 assert(value_top == top(), "second value must be top");
3249 #endif
3250 set_result(value);
3251 return true;
3252 }
3253
3254 #ifdef TRACE_HAVE_INTRINSICS
3255
3256 /*
3257 * oop -> myklass
3258 * myklass->trace_id |= USED
3259 * return myklass->trace_id & ~0x3
3260 */
3261 bool LibraryCallKit::inline_native_classID() {
3262 Node* cls = null_check(argument(0), T_OBJECT);
3263 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3264 kls = null_check(kls, T_OBJECT);
3265
3266 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3267 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3268 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3269
3270 Node* clsused = longcon(0x01l); // set the class bit
3271 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3272 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3273 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3274
3275 #ifdef TRACE_ID_META_BITS
3276 Node* mbits = longcon(~TRACE_ID_META_BITS);
3277 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3278 #endif
3279 #ifdef TRACE_ID_CLASS_SHIFT
3280 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3281 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3282 #endif
3283
3284 set_result(tvalue);
3285 return true;
3286
3287 }
3288
3289 bool LibraryCallKit::inline_native_getBufferWriter() {
3290 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3291
3292 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3293 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3294 );
3295
3296 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3297
3298 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3299 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3300
3301 IfNode* iff_jobj_null =
3302 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3303
3304 enum { _normal_path = 1,
3305 _null_path = 2,
3306 PATH_LIMIT };
3307
3308 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3309 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3310
3311 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3312 result_rgn->init_req(_null_path, jobj_is_null);
3313 result_val->init_req(_null_path, null());
3314
3315 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3316 result_rgn->init_req(_normal_path, jobj_is_not_null);
3317
3318 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3319 result_val->init_req(_normal_path, res);
3320
3321 set_result(result_rgn, result_val);
3322
3323 return true;
3324 }
3325
3326 #endif
3327
3328 //------------------------inline_native_currentThread------------------
3329 bool LibraryCallKit::inline_native_currentThread() {
3330 Node* junk = NULL;
3331 set_result(generate_current_thread(junk));
3332 return true;
3333 }
3334
3335 //------------------------inline_native_isInterrupted------------------
3336 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3337 bool LibraryCallKit::inline_native_isInterrupted() {
3338 // Add a fast path to t.isInterrupted(clear_int):
3339 // (t == Thread.current() &&
3340 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3341 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3342 // So, in the common case that the interrupt bit is false,
3343 // we avoid making a call into the VM. Even if the interrupt bit
3344 // is true, if the clear_int argument is false, we avoid the VM call.
3345 // However, if the receiver is not currentThread, we must call the VM,
3346 // because there must be some locking done around the operation.
3347
3348 // We only go to the fast case code if we pass two guards.
3349 // Paths which do not pass are accumulated in the slow_region.
3350
3351 enum {
3352 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3353 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3354 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3355 PATH_LIMIT
3356 };
3357
3358 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3359 // out of the function.
3360 insert_mem_bar(Op_MemBarCPUOrder);
3361
3362 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3363 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3364
3365 RegionNode* slow_region = new RegionNode(1);
3366 record_for_igvn(slow_region);
3367
3368 // (a) Receiving thread must be the current thread.
3369 Node* rec_thr = argument(0);
3370 Node* tls_ptr = NULL;
3371 Node* cur_thr = generate_current_thread(tls_ptr);
3372 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3373 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3374
3375 generate_slow_guard(bol_thr, slow_region);
3376
3377 // (b) Interrupt bit on TLS must be false.
3378 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3379 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3380 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3381
3382 // Set the control input on the field _interrupted read to prevent it floating up.
3383 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3384 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3385 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3386
3387 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3388
3389 // First fast path: if (!TLS._interrupted) return false;
3390 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3391 result_rgn->init_req(no_int_result_path, false_bit);
3392 result_val->init_req(no_int_result_path, intcon(0));
3393
3394 // drop through to next case
3395 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3396
3397 #ifndef _WINDOWS
3398 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3399 Node* clr_arg = argument(1);
3400 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3401 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3402 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3403
3404 // Second fast path: ... else if (!clear_int) return true;
3405 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3406 result_rgn->init_req(no_clear_result_path, false_arg);
3407 result_val->init_req(no_clear_result_path, intcon(1));
3408
3409 // drop through to next case
3410 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3411 #else
3412 // To return true on Windows you must read the _interrupted field
3413 // and check the event state i.e. take the slow path.
3414 #endif // _WINDOWS
3415
3416 // (d) Otherwise, go to the slow path.
3417 slow_region->add_req(control());
3418 set_control( _gvn.transform(slow_region));
3419
3420 if (stopped()) {
3421 // There is no slow path.
3422 result_rgn->init_req(slow_result_path, top());
3423 result_val->init_req(slow_result_path, top());
3424 } else {
3425 // non-virtual because it is a private non-static
3426 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3427
3428 Node* slow_val = set_results_for_java_call(slow_call);
3429 // this->control() comes from set_results_for_java_call
3430
3431 Node* fast_io = slow_call->in(TypeFunc::I_O);
3432 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3433
3434 // These two phis are pre-filled with copies of of the fast IO and Memory
3435 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3436 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3437
3438 result_rgn->init_req(slow_result_path, control());
3439 result_io ->init_req(slow_result_path, i_o());
3440 result_mem->init_req(slow_result_path, reset_memory());
3441 result_val->init_req(slow_result_path, slow_val);
3442
3443 set_all_memory(_gvn.transform(result_mem));
3444 set_i_o( _gvn.transform(result_io));
3445 }
3446
3447 C->set_has_split_ifs(true); // Has chance for split-if optimization
3448 set_result(result_rgn, result_val);
3449 return true;
3450 }
3451
3452 //---------------------------load_mirror_from_klass----------------------------
3453 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3454 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3455 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3456 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3457 }
3458
3459 //-----------------------load_klass_from_mirror_common-------------------------
3460 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3461 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3462 // and branch to the given path on the region.
3463 // If never_see_null, take an uncommon trap on null, so we can optimistically
3464 // compile for the non-null case.
3465 // If the region is NULL, force never_see_null = true.
3466 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3467 bool never_see_null,
3468 RegionNode* region,
3469 int null_path,
3470 int offset) {
3471 if (region == NULL) never_see_null = true;
3472 Node* p = basic_plus_adr(mirror, offset);
3473 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3474 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3475 Node* null_ctl = top();
3476 kls = null_check_oop(kls, &null_ctl, never_see_null);
3477 if (region != NULL) {
3478 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3479 region->init_req(null_path, null_ctl);
3480 } else {
3481 assert(null_ctl == top(), "no loose ends");
3482 }
3483 return kls;
3484 }
3485
3486 //--------------------(inline_native_Class_query helpers)---------------------
3487 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3488 // Fall through if (mods & mask) == bits, take the guard otherwise.
3489 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3490 // Branch around if the given klass has the given modifier bit set.
3491 // Like generate_guard, adds a new path onto the region.
3492 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3493 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3494 Node* mask = intcon(modifier_mask);
3495 Node* bits = intcon(modifier_bits);
3496 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3497 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3498 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3499 return generate_fair_guard(bol, region);
3500 }
3501 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3502 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3503 }
3504
3505 //-------------------------inline_native_Class_query-------------------
3506 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3507 const Type* return_type = TypeInt::BOOL;
3508 Node* prim_return_value = top(); // what happens if it's a primitive class?
3509 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3510 bool expect_prim = false; // most of these guys expect to work on refs
3511
3512 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3513
3514 Node* mirror = argument(0);
3515 Node* obj = top();
3516
3517 switch (id) {
3518 case vmIntrinsics::_isInstance:
3519 // nothing is an instance of a primitive type
3520 prim_return_value = intcon(0);
3521 obj = argument(1);
3522 break;
3523 case vmIntrinsics::_getModifiers:
3524 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3525 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3526 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3527 break;
3528 case vmIntrinsics::_isInterface:
3529 prim_return_value = intcon(0);
3530 break;
3531 case vmIntrinsics::_isArray:
3532 prim_return_value = intcon(0);
3533 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3534 break;
3535 case vmIntrinsics::_isPrimitive:
3536 prim_return_value = intcon(1);
3537 expect_prim = true; // obviously
3538 break;
3539 case vmIntrinsics::_getSuperclass:
3540 prim_return_value = null();
3541 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3542 break;
3543 case vmIntrinsics::_getClassAccessFlags:
3544 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3545 return_type = TypeInt::INT; // not bool! 6297094
3546 break;
3547 default:
3548 fatal_unexpected_iid(id);
3549 break;
3550 }
3551
3552 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3553 if (mirror_con == NULL) return false; // cannot happen?
3554
3555 #ifndef PRODUCT
3556 if (C->print_intrinsics() || C->print_inlining()) {
3557 ciType* k = mirror_con->java_mirror_type();
3558 if (k) {
3559 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3560 k->print_name();
3561 tty->cr();
3562 }
3563 }
3564 #endif
3565
3566 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3567 RegionNode* region = new RegionNode(PATH_LIMIT);
3568 record_for_igvn(region);
3569 PhiNode* phi = new PhiNode(region, return_type);
3570
3571 // The mirror will never be null of Reflection.getClassAccessFlags, however
3572 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3573 // if it is. See bug 4774291.
3574
3575 // For Reflection.getClassAccessFlags(), the null check occurs in
3576 // the wrong place; see inline_unsafe_access(), above, for a similar
3577 // situation.
3578 mirror = null_check(mirror);
3579 // If mirror or obj is dead, only null-path is taken.
3580 if (stopped()) return true;
3581
3582 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3583
3584 // Now load the mirror's klass metaobject, and null-check it.
3585 // Side-effects region with the control path if the klass is null.
3586 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3587 // If kls is null, we have a primitive mirror.
3588 phi->init_req(_prim_path, prim_return_value);
3589 if (stopped()) { set_result(region, phi); return true; }
3590 bool safe_for_replace = (region->in(_prim_path) == top());
3591
3592 Node* p; // handy temp
3593 Node* null_ctl;
3594
3595 // Now that we have the non-null klass, we can perform the real query.
3596 // For constant classes, the query will constant-fold in LoadNode::Value.
3597 Node* query_value = top();
3598 switch (id) {
3599 case vmIntrinsics::_isInstance:
3600 // nothing is an instance of a primitive type
3601 query_value = gen_instanceof(obj, kls, safe_for_replace);
3602 break;
3603
3604 case vmIntrinsics::_getModifiers:
3605 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3606 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3607 break;
3608
3609 case vmIntrinsics::_isInterface:
3610 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3611 if (generate_interface_guard(kls, region) != NULL)
3612 // A guard was added. If the guard is taken, it was an interface.
3613 phi->add_req(intcon(1));
3614 // If we fall through, it's a plain class.
3615 query_value = intcon(0);
3616 break;
3617
3618 case vmIntrinsics::_isArray:
3619 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3620 if (generate_array_guard(kls, region) != NULL)
3621 // A guard was added. If the guard is taken, it was an array.
3622 phi->add_req(intcon(1));
3623 // If we fall through, it's a plain class.
3624 query_value = intcon(0);
3625 break;
3626
3627 case vmIntrinsics::_isPrimitive:
3628 query_value = intcon(0); // "normal" path produces false
3629 break;
3630
3631 case vmIntrinsics::_getSuperclass:
3632 // The rules here are somewhat unfortunate, but we can still do better
3633 // with random logic than with a JNI call.
3634 // Interfaces store null or Object as _super, but must report null.
3635 // Arrays store an intermediate super as _super, but must report Object.
3636 // Other types can report the actual _super.
3637 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3638 if (generate_interface_guard(kls, region) != NULL)
3639 // A guard was added. If the guard is taken, it was an interface.
3640 phi->add_req(null());
3641 if (generate_array_guard(kls, region) != NULL)
3642 // A guard was added. If the guard is taken, it was an array.
3643 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3644 // If we fall through, it's a plain class. Get its _super.
3645 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3646 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3647 null_ctl = top();
3648 kls = null_check_oop(kls, &null_ctl);
3649 if (null_ctl != top()) {
3650 // If the guard is taken, Object.superClass is null (both klass and mirror).
3651 region->add_req(null_ctl);
3652 phi ->add_req(null());
3653 }
3654 if (!stopped()) {
3655 query_value = load_mirror_from_klass(kls);
3656 }
3657 break;
3658
3659 case vmIntrinsics::_getClassAccessFlags:
3660 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3661 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3662 break;
3663
3664 default:
3665 fatal_unexpected_iid(id);
3666 break;
3667 }
3668
3669 // Fall-through is the normal case of a query to a real class.
3670 phi->init_req(1, query_value);
3671 region->init_req(1, control());
3672
3673 C->set_has_split_ifs(true); // Has chance for split-if optimization
3674 set_result(region, phi);
3675 return true;
3676 }
3677
3678 //-------------------------inline_Class_cast-------------------
3679 bool LibraryCallKit::inline_Class_cast() {
3680 Node* mirror = argument(0); // Class
3681 Node* obj = argument(1);
3682 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3683 if (mirror_con == NULL) {
3684 return false; // dead path (mirror->is_top()).
3685 }
3686 if (obj == NULL || obj->is_top()) {
3687 return false; // dead path
3688 }
3689 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3690
3691 // First, see if Class.cast() can be folded statically.
3692 // java_mirror_type() returns non-null for compile-time Class constants.
3693 ciType* tm = mirror_con->java_mirror_type();
3694 if (tm != NULL && tm->is_klass() &&
3695 tp != NULL && tp->klass() != NULL) {
3696 if (!tp->klass()->is_loaded()) {
3697 // Don't use intrinsic when class is not loaded.
3698 return false;
3699 } else {
3700 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3701 if (static_res == Compile::SSC_always_true) {
3702 // isInstance() is true - fold the code.
3703 set_result(obj);
3704 return true;
3705 } else if (static_res == Compile::SSC_always_false) {
3706 // Don't use intrinsic, have to throw ClassCastException.
3707 // If the reference is null, the non-intrinsic bytecode will
3708 // be optimized appropriately.
3709 return false;
3710 }
3711 }
3712 }
3713
3714 // Bailout intrinsic and do normal inlining if exception path is frequent.
3715 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3716 return false;
3717 }
3718
3719 // Generate dynamic checks.
3720 // Class.cast() is java implementation of _checkcast bytecode.
3721 // Do checkcast (Parse::do_checkcast()) optimizations here.
3722
3723 mirror = null_check(mirror);
3724 // If mirror is dead, only null-path is taken.
3725 if (stopped()) {
3726 return true;
3727 }
3728
3729 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3730 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3731 RegionNode* region = new RegionNode(PATH_LIMIT);
3732 record_for_igvn(region);
3733
3734 // Now load the mirror's klass metaobject, and null-check it.
3735 // If kls is null, we have a primitive mirror and
3736 // nothing is an instance of a primitive type.
3737 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3738
3739 Node* res = top();
3740 if (!stopped()) {
3741 Node* bad_type_ctrl = top();
3742 // Do checkcast optimizations.
3743 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3744 region->init_req(_bad_type_path, bad_type_ctrl);
3745 }
3746 if (region->in(_prim_path) != top() ||
3747 region->in(_bad_type_path) != top()) {
3748 // Let Interpreter throw ClassCastException.
3749 PreserveJVMState pjvms(this);
3750 set_control(_gvn.transform(region));
3751 uncommon_trap(Deoptimization::Reason_intrinsic,
3752 Deoptimization::Action_maybe_recompile);
3753 }
3754 if (!stopped()) {
3755 set_result(res);
3756 }
3757 return true;
3758 }
3759
3760
3761 //--------------------------inline_native_subtype_check------------------------
3762 // This intrinsic takes the JNI calls out of the heart of
3763 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3764 bool LibraryCallKit::inline_native_subtype_check() {
3765 // Pull both arguments off the stack.
3766 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3767 args[0] = argument(0);
3768 args[1] = argument(1);
3769 Node* klasses[2]; // corresponding Klasses: superk, subk
3770 klasses[0] = klasses[1] = top();
3771
3772 enum {
3773 // A full decision tree on {superc is prim, subc is prim}:
3774 _prim_0_path = 1, // {P,N} => false
3775 // {P,P} & superc!=subc => false
3776 _prim_same_path, // {P,P} & superc==subc => true
3777 _prim_1_path, // {N,P} => false
3778 _ref_subtype_path, // {N,N} & subtype check wins => true
3779 _both_ref_path, // {N,N} & subtype check loses => false
3780 PATH_LIMIT
3781 };
3782
3783 RegionNode* region = new RegionNode(PATH_LIMIT);
3784 Node* phi = new PhiNode(region, TypeInt::BOOL);
3785 record_for_igvn(region);
3786
3787 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3788 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3789 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3790
3791 // First null-check both mirrors and load each mirror's klass metaobject.
3792 int which_arg;
3793 for (which_arg = 0; which_arg <= 1; which_arg++) {
3794 Node* arg = args[which_arg];
3795 arg = null_check(arg);
3796 if (stopped()) break;
3797 args[which_arg] = arg;
3798
3799 Node* p = basic_plus_adr(arg, class_klass_offset);
3800 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3801 klasses[which_arg] = _gvn.transform(kls);
3802 }
3803
3804 // Having loaded both klasses, test each for null.
3805 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3806 for (which_arg = 0; which_arg <= 1; which_arg++) {
3807 Node* kls = klasses[which_arg];
3808 Node* null_ctl = top();
3809 kls = null_check_oop(kls, &null_ctl, never_see_null);
3810 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3811 region->init_req(prim_path, null_ctl);
3812 if (stopped()) break;
3813 klasses[which_arg] = kls;
3814 }
3815
3816 if (!stopped()) {
3817 // now we have two reference types, in klasses[0..1]
3818 Node* subk = klasses[1]; // the argument to isAssignableFrom
3819 Node* superk = klasses[0]; // the receiver
3820 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3821 // now we have a successful reference subtype check
3822 region->set_req(_ref_subtype_path, control());
3823 }
3824
3825 // If both operands are primitive (both klasses null), then
3826 // we must return true when they are identical primitives.
3827 // It is convenient to test this after the first null klass check.
3828 set_control(region->in(_prim_0_path)); // go back to first null check
3829 if (!stopped()) {
3830 // Since superc is primitive, make a guard for the superc==subc case.
3831 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3832 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3833 generate_guard(bol_eq, region, PROB_FAIR);
3834 if (region->req() == PATH_LIMIT+1) {
3835 // A guard was added. If the added guard is taken, superc==subc.
3836 region->swap_edges(PATH_LIMIT, _prim_same_path);
3837 region->del_req(PATH_LIMIT);
3838 }
3839 region->set_req(_prim_0_path, control()); // Not equal after all.
3840 }
3841
3842 // these are the only paths that produce 'true':
3843 phi->set_req(_prim_same_path, intcon(1));
3844 phi->set_req(_ref_subtype_path, intcon(1));
3845
3846 // pull together the cases:
3847 assert(region->req() == PATH_LIMIT, "sane region");
3848 for (uint i = 1; i < region->req(); i++) {
3849 Node* ctl = region->in(i);
3850 if (ctl == NULL || ctl == top()) {
3851 region->set_req(i, top());
3852 phi ->set_req(i, top());
3853 } else if (phi->in(i) == NULL) {
3854 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3855 }
3856 }
3857
3858 set_control(_gvn.transform(region));
3859 set_result(_gvn.transform(phi));
3860 return true;
3861 }
3862
3863 //---------------------generate_array_guard_common------------------------
3864 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3865 bool obj_array, bool not_array) {
3866
3867 if (stopped()) {
3868 return NULL;
3869 }
3870
3871 // If obj_array/non_array==false/false:
3872 // Branch around if the given klass is in fact an array (either obj or prim).
3873 // If obj_array/non_array==false/true:
3874 // Branch around if the given klass is not an array klass of any kind.
3875 // If obj_array/non_array==true/true:
3876 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3877 // If obj_array/non_array==true/false:
3878 // Branch around if the kls is an oop array (Object[] or subtype)
3879 //
3880 // Like generate_guard, adds a new path onto the region.
3881 jint layout_con = 0;
3882 Node* layout_val = get_layout_helper(kls, layout_con);
3883 if (layout_val == NULL) {
3884 bool query = (obj_array
3885 ? Klass::layout_helper_is_objArray(layout_con)
3886 : Klass::layout_helper_is_array(layout_con));
3887 if (query == not_array) {
3888 return NULL; // never a branch
3889 } else { // always a branch
3890 Node* always_branch = control();
3891 if (region != NULL)
3892 region->add_req(always_branch);
3893 set_control(top());
3894 return always_branch;
3895 }
3896 }
3897 // Now test the correct condition.
3898 jint nval = (obj_array
3899 ? (jint)(Klass::_lh_array_tag_type_value
3900 << Klass::_lh_array_tag_shift)
3901 : Klass::_lh_neutral_value);
3902 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3903 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3904 // invert the test if we are looking for a non-array
3905 if (not_array) btest = BoolTest(btest).negate();
3906 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3907 return generate_fair_guard(bol, region);
3908 }
3909
3910
3911 //-----------------------inline_native_newArray--------------------------
3912 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3913 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3914 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3915 Node* mirror;
3916 Node* count_val;
3917 if (uninitialized) {
3918 mirror = argument(1);
3919 count_val = argument(2);
3920 } else {
3921 mirror = argument(0);
3922 count_val = argument(1);
3923 }
3924
3925 mirror = null_check(mirror);
3926 // If mirror or obj is dead, only null-path is taken.
3927 if (stopped()) return true;
3928
3929 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3930 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3931 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3932 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3933 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3934
3935 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3936 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3937 result_reg, _slow_path);
3938 Node* normal_ctl = control();
3939 Node* no_array_ctl = result_reg->in(_slow_path);
3940
3941 // Generate code for the slow case. We make a call to newArray().
3942 set_control(no_array_ctl);
3943 if (!stopped()) {
3944 // Either the input type is void.class, or else the
3945 // array klass has not yet been cached. Either the
3946 // ensuing call will throw an exception, or else it
3947 // will cache the array klass for next time.
3948 PreserveJVMState pjvms(this);
3949 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3950 Node* slow_result = set_results_for_java_call(slow_call);
3951 // this->control() comes from set_results_for_java_call
3952 result_reg->set_req(_slow_path, control());
3953 result_val->set_req(_slow_path, slow_result);
3954 result_io ->set_req(_slow_path, i_o());
3955 result_mem->set_req(_slow_path, reset_memory());
3956 }
3957
3958 set_control(normal_ctl);
3959 if (!stopped()) {
3960 // Normal case: The array type has been cached in the java.lang.Class.
3961 // The following call works fine even if the array type is polymorphic.
3962 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3963 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3964 result_reg->init_req(_normal_path, control());
3965 result_val->init_req(_normal_path, obj);
3966 result_io ->init_req(_normal_path, i_o());
3967 result_mem->init_req(_normal_path, reset_memory());
3968
3969 if (uninitialized) {
3970 // Mark the allocation so that zeroing is skipped
3971 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3972 alloc->maybe_set_complete(&_gvn);
3973 }
3974 }
3975
3976 // Return the combined state.
3977 set_i_o( _gvn.transform(result_io) );
3978 set_all_memory( _gvn.transform(result_mem));
3979
3980 C->set_has_split_ifs(true); // Has chance for split-if optimization
3981 set_result(result_reg, result_val);
3982 return true;
3983 }
3984
3985 //----------------------inline_native_getLength--------------------------
3986 // public static native int java.lang.reflect.Array.getLength(Object array);
3987 bool LibraryCallKit::inline_native_getLength() {
3988 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3989
3990 Node* array = null_check(argument(0));
3991 // If array is dead, only null-path is taken.
3992 if (stopped()) return true;
3993
3994 // Deoptimize if it is a non-array.
3995 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3996
3997 if (non_array != NULL) {
3998 PreserveJVMState pjvms(this);
3999 set_control(non_array);
4000 uncommon_trap(Deoptimization::Reason_intrinsic,
4001 Deoptimization::Action_maybe_recompile);
4002 }
4003
4004 // If control is dead, only non-array-path is taken.
4005 if (stopped()) return true;
4006
4007 // The works fine even if the array type is polymorphic.
4008 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4009 Node* result = load_array_length(array);
4010
4011 C->set_has_split_ifs(true); // Has chance for split-if optimization
4012 set_result(result);
4013 return true;
4014 }
4015
4016 //------------------------inline_array_copyOf----------------------------
4017 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4018 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4019 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4020 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4021
4022 // Get the arguments.
4023 Node* original = argument(0);
4024 Node* start = is_copyOfRange? argument(1): intcon(0);
4025 Node* end = is_copyOfRange? argument(2): argument(1);
4026 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4027
4028 Node* newcopy = NULL;
4029
4030 // Set the original stack and the reexecute bit for the interpreter to reexecute
4031 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4032 { PreserveReexecuteState preexecs(this);
4033 jvms()->set_should_reexecute(true);
4034
4035 array_type_mirror = null_check(array_type_mirror);
4036 original = null_check(original);
4037
4038 // Check if a null path was taken unconditionally.
4039 if (stopped()) return true;
4040
4041 Node* orig_length = load_array_length(original);
4042
4043 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4044 klass_node = null_check(klass_node);
4045
4046 RegionNode* bailout = new RegionNode(1);
4047 record_for_igvn(bailout);
4048
4049 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4050 // Bail out if that is so.
4051 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4052 if (not_objArray != NULL) {
4053 // Improve the klass node's type from the new optimistic assumption:
4054 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4055 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4056 Node* cast = new CastPPNode(klass_node, akls);
4057 cast->init_req(0, control());
4058 klass_node = _gvn.transform(cast);
4059 }
4060
4061 // Bail out if either start or end is negative.
4062 generate_negative_guard(start, bailout, &start);
4063 generate_negative_guard(end, bailout, &end);
4064
4065 Node* length = end;
4066 if (_gvn.type(start) != TypeInt::ZERO) {
4067 length = _gvn.transform(new SubINode(end, start));
4068 }
4069
4070 // Bail out if length is negative.
4071 // Without this the new_array would throw
4072 // NegativeArraySizeException but IllegalArgumentException is what
4073 // should be thrown
4074 generate_negative_guard(length, bailout, &length);
4075
4076 if (bailout->req() > 1) {
4077 PreserveJVMState pjvms(this);
4078 set_control(_gvn.transform(bailout));
4079 uncommon_trap(Deoptimization::Reason_intrinsic,
4080 Deoptimization::Action_maybe_recompile);
4081 }
4082
4083 if (!stopped()) {
4084 // How many elements will we copy from the original?
4085 // The answer is MinI(orig_length - start, length).
4086 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4087 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4088
4089 // Generate a direct call to the right arraycopy function(s).
4090 // We know the copy is disjoint but we might not know if the
4091 // oop stores need checking.
4092 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4093 // This will fail a store-check if x contains any non-nulls.
4094
4095 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4096 // loads/stores but it is legal only if we're sure the
4097 // Arrays.copyOf would succeed. So we need all input arguments
4098 // to the copyOf to be validated, including that the copy to the
4099 // new array won't trigger an ArrayStoreException. That subtype
4100 // check can be optimized if we know something on the type of
4101 // the input array from type speculation.
4102 if (_gvn.type(klass_node)->singleton()) {
4103 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4104 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4105
4106 int test = C->static_subtype_check(superk, subk);
4107 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4108 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4109 if (t_original->speculative_type() != NULL) {
4110 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4111 }
4112 }
4113 }
4114
4115 bool validated = false;
4116 // Reason_class_check rather than Reason_intrinsic because we
4117 // want to intrinsify even if this traps.
4118 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4119 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4120 klass_node);
4121
4122 if (not_subtype_ctrl != top()) {
4123 PreserveJVMState pjvms(this);
4124 set_control(not_subtype_ctrl);
4125 uncommon_trap(Deoptimization::Reason_class_check,
4126 Deoptimization::Action_make_not_entrant);
4127 assert(stopped(), "Should be stopped");
4128 }
4129 validated = true;
4130 }
4131
4132 if (!stopped()) {
4133 newcopy = new_array(klass_node, length, 0); // no arguments to push
4134
4135 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4136 load_object_klass(original), klass_node);
4137 if (!is_copyOfRange) {
4138 ac->set_copyof(validated);
4139 } else {
4140 ac->set_copyofrange(validated);
4141 }
4142 Node* n = _gvn.transform(ac);
4143 if (n == ac) {
4144 ac->connect_outputs(this);
4145 } else {
4146 assert(validated, "shouldn't transform if all arguments not validated");
4147 set_all_memory(n);
4148 }
4149 }
4150 }
4151 } // original reexecute is set back here
4152
4153 C->set_has_split_ifs(true); // Has chance for split-if optimization
4154 if (!stopped()) {
4155 set_result(newcopy);
4156 }
4157 return true;
4158 }
4159
4160
4161 //----------------------generate_virtual_guard---------------------------
4162 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4163 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4164 RegionNode* slow_region) {
4165 ciMethod* method = callee();
4166 int vtable_index = method->vtable_index();
4167 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4168 "bad index %d", vtable_index);
4169 // Get the Method* out of the appropriate vtable entry.
4170 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4171 vtable_index*vtableEntry::size_in_bytes() +
4172 vtableEntry::method_offset_in_bytes();
4173 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4174 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4175
4176 // Compare the target method with the expected method (e.g., Object.hashCode).
4177 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4178
4179 Node* native_call = makecon(native_call_addr);
4180 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4181 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4182
4183 return generate_slow_guard(test_native, slow_region);
4184 }
4185
4186 //-----------------------generate_method_call----------------------------
4187 // Use generate_method_call to make a slow-call to the real
4188 // method if the fast path fails. An alternative would be to
4189 // use a stub like OptoRuntime::slow_arraycopy_Java.
4190 // This only works for expanding the current library call,
4191 // not another intrinsic. (E.g., don't use this for making an
4192 // arraycopy call inside of the copyOf intrinsic.)
4193 CallJavaNode*
4194 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4195 // When compiling the intrinsic method itself, do not use this technique.
4196 guarantee(callee() != C->method(), "cannot make slow-call to self");
4197
4198 ciMethod* method = callee();
4199 // ensure the JVMS we have will be correct for this call
4200 guarantee(method_id == method->intrinsic_id(), "must match");
4201
4202 const TypeFunc* tf = TypeFunc::make(method);
4203 CallJavaNode* slow_call;
4204 if (is_static) {
4205 assert(!is_virtual, "");
4206 slow_call = new CallStaticJavaNode(C, tf,
4207 SharedRuntime::get_resolve_static_call_stub(),
4208 method, bci());
4209 } else if (is_virtual) {
4210 null_check_receiver();
4211 int vtable_index = Method::invalid_vtable_index;
4212 if (UseInlineCaches) {
4213 // Suppress the vtable call
4214 } else {
4215 // hashCode and clone are not a miranda methods,
4216 // so the vtable index is fixed.
4217 // No need to use the linkResolver to get it.
4218 vtable_index = method->vtable_index();
4219 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4220 "bad index %d", vtable_index);
4221 }
4222 slow_call = new CallDynamicJavaNode(tf,
4223 SharedRuntime::get_resolve_virtual_call_stub(),
4224 method, vtable_index, bci());
4225 } else { // neither virtual nor static: opt_virtual
4226 null_check_receiver();
4227 slow_call = new CallStaticJavaNode(C, tf,
4228 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4229 method, bci());
4230 slow_call->set_optimized_virtual(true);
4231 }
4232 set_arguments_for_java_call(slow_call);
4233 set_edges_for_java_call(slow_call);
4234 return slow_call;
4235 }
4236
4237
4238 /**
4239 * Build special case code for calls to hashCode on an object. This call may
4240 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4241 * slightly different code.
4242 */
4243 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4244 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4245 assert(!(is_virtual && is_static), "either virtual, special, or static");
4246
4247 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4248
4249 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4250 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4251 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4252 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4253 Node* obj = NULL;
4254 if (!is_static) {
4255 // Check for hashing null object
4256 obj = null_check_receiver();
4257 if (stopped()) return true; // unconditionally null
4258 result_reg->init_req(_null_path, top());
4259 result_val->init_req(_null_path, top());
4260 } else {
4261 // Do a null check, and return zero if null.
4262 // System.identityHashCode(null) == 0
4263 obj = argument(0);
4264 Node* null_ctl = top();
4265 obj = null_check_oop(obj, &null_ctl);
4266 result_reg->init_req(_null_path, null_ctl);
4267 result_val->init_req(_null_path, _gvn.intcon(0));
4268 }
4269
4270 // Unconditionally null? Then return right away.
4271 if (stopped()) {
4272 set_control( result_reg->in(_null_path));
4273 if (!stopped())
4274 set_result(result_val->in(_null_path));
4275 return true;
4276 }
4277
4278 // We only go to the fast case code if we pass a number of guards. The
4279 // paths which do not pass are accumulated in the slow_region.
4280 RegionNode* slow_region = new RegionNode(1);
4281 record_for_igvn(slow_region);
4282
4283 // If this is a virtual call, we generate a funny guard. We pull out
4284 // the vtable entry corresponding to hashCode() from the target object.
4285 // If the target method which we are calling happens to be the native
4286 // Object hashCode() method, we pass the guard. We do not need this
4287 // guard for non-virtual calls -- the caller is known to be the native
4288 // Object hashCode().
4289 if (is_virtual) {
4290 // After null check, get the object's klass.
4291 Node* obj_klass = load_object_klass(obj);
4292 generate_virtual_guard(obj_klass, slow_region);
4293 }
4294
4295 // Get the header out of the object, use LoadMarkNode when available
4296 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4297 // The control of the load must be NULL. Otherwise, the load can move before
4298 // the null check after castPP removal.
4299 Node* no_ctrl = NULL;
4300 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4301
4302 // Test the header to see if it is unlocked.
4303 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4304 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4305 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4306 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4307 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4308
4309 generate_slow_guard(test_unlocked, slow_region);
4310
4311 // Get the hash value and check to see that it has been properly assigned.
4312 // We depend on hash_mask being at most 32 bits and avoid the use of
4313 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4314 // vm: see markOop.hpp.
4315 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4316 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4317 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4318 // This hack lets the hash bits live anywhere in the mark object now, as long
4319 // as the shift drops the relevant bits into the low 32 bits. Note that
4320 // Java spec says that HashCode is an int so there's no point in capturing
4321 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4322 hshifted_header = ConvX2I(hshifted_header);
4323 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4324
4325 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4326 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4327 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4328
4329 generate_slow_guard(test_assigned, slow_region);
4330
4331 Node* init_mem = reset_memory();
4332 // fill in the rest of the null path:
4333 result_io ->init_req(_null_path, i_o());
4334 result_mem->init_req(_null_path, init_mem);
4335
4336 result_val->init_req(_fast_path, hash_val);
4337 result_reg->init_req(_fast_path, control());
4338 result_io ->init_req(_fast_path, i_o());
4339 result_mem->init_req(_fast_path, init_mem);
4340
4341 // Generate code for the slow case. We make a call to hashCode().
4342 set_control(_gvn.transform(slow_region));
4343 if (!stopped()) {
4344 // No need for PreserveJVMState, because we're using up the present state.
4345 set_all_memory(init_mem);
4346 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4347 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4348 Node* slow_result = set_results_for_java_call(slow_call);
4349 // this->control() comes from set_results_for_java_call
4350 result_reg->init_req(_slow_path, control());
4351 result_val->init_req(_slow_path, slow_result);
4352 result_io ->set_req(_slow_path, i_o());
4353 result_mem ->set_req(_slow_path, reset_memory());
4354 }
4355
4356 // Return the combined state.
4357 set_i_o( _gvn.transform(result_io) );
4358 set_all_memory( _gvn.transform(result_mem));
4359
4360 set_result(result_reg, result_val);
4361 return true;
4362 }
4363
4364 //---------------------------inline_native_getClass----------------------------
4365 // public final native Class<?> java.lang.Object.getClass();
4366 //
4367 // Build special case code for calls to getClass on an object.
4368 bool LibraryCallKit::inline_native_getClass() {
4369 Node* obj = null_check_receiver();
4370 if (stopped()) return true;
4371 set_result(load_mirror_from_klass(load_object_klass(obj)));
4372 return true;
4373 }
4374
4375 //-----------------inline_native_Reflection_getCallerClass---------------------
4376 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4377 //
4378 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4379 //
4380 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4381 // in that it must skip particular security frames and checks for
4382 // caller sensitive methods.
4383 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4384 #ifndef PRODUCT
4385 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4386 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4387 }
4388 #endif
4389
4390 if (!jvms()->has_method()) {
4391 #ifndef PRODUCT
4392 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4393 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4394 }
4395 #endif
4396 return false;
4397 }
4398
4399 // Walk back up the JVM state to find the caller at the required
4400 // depth.
4401 JVMState* caller_jvms = jvms();
4402
4403 // Cf. JVM_GetCallerClass
4404 // NOTE: Start the loop at depth 1 because the current JVM state does
4405 // not include the Reflection.getCallerClass() frame.
4406 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4407 ciMethod* m = caller_jvms->method();
4408 switch (n) {
4409 case 0:
4410 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4411 break;
4412 case 1:
4413 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4414 if (!m->caller_sensitive()) {
4415 #ifndef PRODUCT
4416 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4417 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4418 }
4419 #endif
4420 return false; // bail-out; let JVM_GetCallerClass do the work
4421 }
4422 break;
4423 default:
4424 if (!m->is_ignored_by_security_stack_walk()) {
4425 // We have reached the desired frame; return the holder class.
4426 // Acquire method holder as java.lang.Class and push as constant.
4427 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4428 ciInstance* caller_mirror = caller_klass->java_mirror();
4429 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4430
4431 #ifndef PRODUCT
4432 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4433 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());
4434 tty->print_cr(" JVM state at this point:");
4435 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4436 ciMethod* m = jvms()->of_depth(i)->method();
4437 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4438 }
4439 }
4440 #endif
4441 return true;
4442 }
4443 break;
4444 }
4445 }
4446
4447 #ifndef PRODUCT
4448 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4449 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4450 tty->print_cr(" JVM state at this point:");
4451 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4452 ciMethod* m = jvms()->of_depth(i)->method();
4453 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4454 }
4455 }
4456 #endif
4457
4458 return false; // bail-out; let JVM_GetCallerClass do the work
4459 }
4460
4461 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4462 Node* arg = argument(0);
4463 Node* result = NULL;
4464
4465 switch (id) {
4466 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4467 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4468 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4469 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4470
4471 case vmIntrinsics::_doubleToLongBits: {
4472 // two paths (plus control) merge in a wood
4473 RegionNode *r = new RegionNode(3);
4474 Node *phi = new PhiNode(r, TypeLong::LONG);
4475
4476 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4477 // Build the boolean node
4478 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4479
4480 // Branch either way.
4481 // NaN case is less traveled, which makes all the difference.
4482 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4483 Node *opt_isnan = _gvn.transform(ifisnan);
4484 assert( opt_isnan->is_If(), "Expect an IfNode");
4485 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4486 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4487
4488 set_control(iftrue);
4489
4490 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4491 Node *slow_result = longcon(nan_bits); // return NaN
4492 phi->init_req(1, _gvn.transform( slow_result ));
4493 r->init_req(1, iftrue);
4494
4495 // Else fall through
4496 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4497 set_control(iffalse);
4498
4499 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4500 r->init_req(2, iffalse);
4501
4502 // Post merge
4503 set_control(_gvn.transform(r));
4504 record_for_igvn(r);
4505
4506 C->set_has_split_ifs(true); // Has chance for split-if optimization
4507 result = phi;
4508 assert(result->bottom_type()->isa_long(), "must be");
4509 break;
4510 }
4511
4512 case vmIntrinsics::_floatToIntBits: {
4513 // two paths (plus control) merge in a wood
4514 RegionNode *r = new RegionNode(3);
4515 Node *phi = new PhiNode(r, TypeInt::INT);
4516
4517 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4518 // Build the boolean node
4519 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4520
4521 // Branch either way.
4522 // NaN case is less traveled, which makes all the difference.
4523 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4524 Node *opt_isnan = _gvn.transform(ifisnan);
4525 assert( opt_isnan->is_If(), "Expect an IfNode");
4526 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4527 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4528
4529 set_control(iftrue);
4530
4531 static const jint nan_bits = 0x7fc00000;
4532 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4533 phi->init_req(1, _gvn.transform( slow_result ));
4534 r->init_req(1, iftrue);
4535
4536 // Else fall through
4537 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4538 set_control(iffalse);
4539
4540 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4541 r->init_req(2, iffalse);
4542
4543 // Post merge
4544 set_control(_gvn.transform(r));
4545 record_for_igvn(r);
4546
4547 C->set_has_split_ifs(true); // Has chance for split-if optimization
4548 result = phi;
4549 assert(result->bottom_type()->isa_int(), "must be");
4550 break;
4551 }
4552
4553 default:
4554 fatal_unexpected_iid(id);
4555 break;
4556 }
4557 set_result(_gvn.transform(result));
4558 return true;
4559 }
4560
4561 //----------------------inline_unsafe_copyMemory-------------------------
4562 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4563 bool LibraryCallKit::inline_unsafe_copyMemory() {
4564 if (callee()->is_static()) return false; // caller must have the capability!
4565 null_check_receiver(); // null-check receiver
4566 if (stopped()) return true;
4567
4568 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4569
4570 Node* src_ptr = argument(1); // type: oop
4571 Node* src_off = ConvL2X(argument(2)); // type: long
4572 Node* dst_ptr = argument(4); // type: oop
4573 Node* dst_off = ConvL2X(argument(5)); // type: long
4574 Node* size = ConvL2X(argument(7)); // type: long
4575
4576 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4577 "fieldOffset must be byte-scaled");
4578
4579 Node* src = make_unsafe_address(src_ptr, src_off);
4580 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4581
4582 // Conservatively insert a memory barrier on all memory slices.
4583 // Do not let writes of the copy source or destination float below the copy.
4584 insert_mem_bar(Op_MemBarCPUOrder);
4585
4586 // Call it. Note that the length argument is not scaled.
4587 make_runtime_call(RC_LEAF|RC_NO_FP,
4588 OptoRuntime::fast_arraycopy_Type(),
4589 StubRoutines::unsafe_arraycopy(),
4590 "unsafe_arraycopy",
4591 TypeRawPtr::BOTTOM,
4592 src, dst, size XTOP);
4593
4594 // Do not let reads of the copy destination float above the copy.
4595 insert_mem_bar(Op_MemBarCPUOrder);
4596
4597 return true;
4598 }
4599
4600 //------------------------clone_coping-----------------------------------
4601 // Helper function for inline_native_clone.
4602 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4603 assert(obj_size != NULL, "");
4604 Node* raw_obj = alloc_obj->in(1);
4605 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4606
4607 AllocateNode* alloc = NULL;
4608 if (ReduceBulkZeroing) {
4609 // We will be completely responsible for initializing this object -
4610 // mark Initialize node as complete.
4611 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4612 // The object was just allocated - there should be no any stores!
4613 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4614 // Mark as complete_with_arraycopy so that on AllocateNode
4615 // expansion, we know this AllocateNode is initialized by an array
4616 // copy and a StoreStore barrier exists after the array copy.
4617 alloc->initialization()->set_complete_with_arraycopy();
4618 }
4619
4620 // Copy the fastest available way.
4621 // TODO: generate fields copies for small objects instead.
4622 Node* src = obj;
4623 Node* dest = alloc_obj;
4624 Node* size = _gvn.transform(obj_size);
4625
4626 // Exclude the header but include array length to copy by 8 bytes words.
4627 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4628 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4629 instanceOopDesc::base_offset_in_bytes();
4630 // base_off:
4631 // 8 - 32-bit VM
4632 // 12 - 64-bit VM, compressed klass
4633 // 16 - 64-bit VM, normal klass
4634 if (base_off % BytesPerLong != 0) {
4635 assert(UseCompressedClassPointers, "");
4636 if (is_array) {
4637 // Exclude length to copy by 8 bytes words.
4638 base_off += sizeof(int);
4639 } else {
4640 // Include klass to copy by 8 bytes words.
4641 base_off = instanceOopDesc::klass_offset_in_bytes();
4642 }
4643 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4644 }
4645 src = basic_plus_adr(src, base_off);
4646 dest = basic_plus_adr(dest, base_off);
4647
4648 // Compute the length also, if needed:
4649 Node* countx = size;
4650 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4651 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4652
4653 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4654
4655 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4656 ac->set_clonebasic();
4657 Node* n = _gvn.transform(ac);
4658 if (n == ac) {
4659 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4660 } else {
4661 set_all_memory(n);
4662 }
4663
4664 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4665 if (card_mark) {
4666 assert(!is_array, "");
4667 // Put in store barrier for any and all oops we are sticking
4668 // into this object. (We could avoid this if we could prove
4669 // that the object type contains no oop fields at all.)
4670 Node* no_particular_value = NULL;
4671 Node* no_particular_field = NULL;
4672 int raw_adr_idx = Compile::AliasIdxRaw;
4673 post_barrier(control(),
4674 memory(raw_adr_type),
4675 alloc_obj,
4676 no_particular_field,
4677 raw_adr_idx,
4678 no_particular_value,
4679 T_OBJECT,
4680 false);
4681 }
4682
4683 // Do not let reads from the cloned object float above the arraycopy.
4684 if (alloc != NULL) {
4685 // Do not let stores that initialize this object be reordered with
4686 // a subsequent store that would make this object accessible by
4687 // other threads.
4688 // Record what AllocateNode this StoreStore protects so that
4689 // escape analysis can go from the MemBarStoreStoreNode to the
4690 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4691 // based on the escape status of the AllocateNode.
4692 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4693 } else {
4694 insert_mem_bar(Op_MemBarCPUOrder);
4695 }
4696 }
4697
4698 //------------------------inline_native_clone----------------------------
4699 // protected native Object java.lang.Object.clone();
4700 //
4701 // Here are the simple edge cases:
4702 // null receiver => normal trap
4703 // virtual and clone was overridden => slow path to out-of-line clone
4704 // not cloneable or finalizer => slow path to out-of-line Object.clone
4705 //
4706 // The general case has two steps, allocation and copying.
4707 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4708 //
4709 // Copying also has two cases, oop arrays and everything else.
4710 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4711 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4712 //
4713 // These steps fold up nicely if and when the cloned object's klass
4714 // can be sharply typed as an object array, a type array, or an instance.
4715 //
4716 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4717 PhiNode* result_val;
4718
4719 // Set the reexecute bit for the interpreter to reexecute
4720 // the bytecode that invokes Object.clone if deoptimization happens.
4721 { PreserveReexecuteState preexecs(this);
4722 jvms()->set_should_reexecute(true);
4723
4724 Node* obj = null_check_receiver();
4725 if (stopped()) return true;
4726
4727 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4728
4729 // If we are going to clone an instance, we need its exact type to
4730 // know the number and types of fields to convert the clone to
4731 // loads/stores. Maybe a speculative type can help us.
4732 if (!obj_type->klass_is_exact() &&
4733 obj_type->speculative_type() != NULL &&
4734 obj_type->speculative_type()->is_instance_klass()) {
4735 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4736 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4737 !spec_ik->has_injected_fields()) {
4738 ciKlass* k = obj_type->klass();
4739 if (!k->is_instance_klass() ||
4740 k->as_instance_klass()->is_interface() ||
4741 k->as_instance_klass()->has_subklass()) {
4742 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4743 }
4744 }
4745 }
4746
4747 Node* obj_klass = load_object_klass(obj);
4748 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4749 const TypeOopPtr* toop = ((tklass != NULL)
4750 ? tklass->as_instance_type()
4751 : TypeInstPtr::NOTNULL);
4752
4753 // Conservatively insert a memory barrier on all memory slices.
4754 // Do not let writes into the original float below the clone.
4755 insert_mem_bar(Op_MemBarCPUOrder);
4756
4757 // paths into result_reg:
4758 enum {
4759 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4760 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4761 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4762 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4763 PATH_LIMIT
4764 };
4765 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4766 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4767 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4768 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4769 record_for_igvn(result_reg);
4770
4771 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4772 int raw_adr_idx = Compile::AliasIdxRaw;
4773
4774 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4775 if (array_ctl != NULL) {
4776 // It's an array.
4777 PreserveJVMState pjvms(this);
4778 set_control(array_ctl);
4779 Node* obj_length = load_array_length(obj);
4780 Node* obj_size = NULL;
4781 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4782
4783 if (!use_ReduceInitialCardMarks()) {
4784 // If it is an oop array, it requires very special treatment,
4785 // because card marking is required on each card of the array.
4786 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4787 if (is_obja != NULL) {
4788 PreserveJVMState pjvms2(this);
4789 set_control(is_obja);
4790 // Generate a direct call to the right arraycopy function(s).
4791 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4792 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4793 ac->set_cloneoop();
4794 Node* n = _gvn.transform(ac);
4795 assert(n == ac, "cannot disappear");
4796 ac->connect_outputs(this);
4797
4798 result_reg->init_req(_objArray_path, control());
4799 result_val->init_req(_objArray_path, alloc_obj);
4800 result_i_o ->set_req(_objArray_path, i_o());
4801 result_mem ->set_req(_objArray_path, reset_memory());
4802 }
4803 }
4804 // Otherwise, there are no card marks to worry about.
4805 // (We can dispense with card marks if we know the allocation
4806 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4807 // causes the non-eden paths to take compensating steps to
4808 // simulate a fresh allocation, so that no further
4809 // card marks are required in compiled code to initialize
4810 // the object.)
4811
4812 if (!stopped()) {
4813 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4814
4815 // Present the results of the copy.
4816 result_reg->init_req(_array_path, control());
4817 result_val->init_req(_array_path, alloc_obj);
4818 result_i_o ->set_req(_array_path, i_o());
4819 result_mem ->set_req(_array_path, reset_memory());
4820 }
4821 }
4822
4823 // We only go to the instance fast case code if we pass a number of guards.
4824 // The paths which do not pass are accumulated in the slow_region.
4825 RegionNode* slow_region = new RegionNode(1);
4826 record_for_igvn(slow_region);
4827 if (!stopped()) {
4828 // It's an instance (we did array above). Make the slow-path tests.
4829 // If this is a virtual call, we generate a funny guard. We grab
4830 // the vtable entry corresponding to clone() from the target object.
4831 // If the target method which we are calling happens to be the
4832 // Object clone() method, we pass the guard. We do not need this
4833 // guard for non-virtual calls; the caller is known to be the native
4834 // Object clone().
4835 if (is_virtual) {
4836 generate_virtual_guard(obj_klass, slow_region);
4837 }
4838
4839 // The object must be easily cloneable and must not have a finalizer.
4840 // Both of these conditions may be checked in a single test.
4841 // We could optimize the test further, but we don't care.
4842 generate_access_flags_guard(obj_klass,
4843 // Test both conditions:
4844 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4845 // Must be cloneable but not finalizer:
4846 JVM_ACC_IS_CLONEABLE_FAST,
4847 slow_region);
4848 }
4849
4850 if (!stopped()) {
4851 // It's an instance, and it passed the slow-path tests.
4852 PreserveJVMState pjvms(this);
4853 Node* obj_size = NULL;
4854 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4855 // is reexecuted if deoptimization occurs and there could be problems when merging
4856 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4857 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4858
4859 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4860
4861 // Present the results of the slow call.
4862 result_reg->init_req(_instance_path, control());
4863 result_val->init_req(_instance_path, alloc_obj);
4864 result_i_o ->set_req(_instance_path, i_o());
4865 result_mem ->set_req(_instance_path, reset_memory());
4866 }
4867
4868 // Generate code for the slow case. We make a call to clone().
4869 set_control(_gvn.transform(slow_region));
4870 if (!stopped()) {
4871 PreserveJVMState pjvms(this);
4872 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4873 Node* slow_result = set_results_for_java_call(slow_call);
4874 // this->control() comes from set_results_for_java_call
4875 result_reg->init_req(_slow_path, control());
4876 result_val->init_req(_slow_path, slow_result);
4877 result_i_o ->set_req(_slow_path, i_o());
4878 result_mem ->set_req(_slow_path, reset_memory());
4879 }
4880
4881 // Return the combined state.
4882 set_control( _gvn.transform(result_reg));
4883 set_i_o( _gvn.transform(result_i_o));
4884 set_all_memory( _gvn.transform(result_mem));
4885 } // original reexecute is set back here
4886
4887 set_result(_gvn.transform(result_val));
4888 return true;
4889 }
4890
4891 // If we have a tighly coupled allocation, the arraycopy may take care
4892 // of the array initialization. If one of the guards we insert between
4893 // the allocation and the arraycopy causes a deoptimization, an
4894 // unitialized array will escape the compiled method. To prevent that
4895 // we set the JVM state for uncommon traps between the allocation and
4896 // the arraycopy to the state before the allocation so, in case of
4897 // deoptimization, we'll reexecute the allocation and the
4898 // initialization.
4899 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4900 if (alloc != NULL) {
4901 ciMethod* trap_method = alloc->jvms()->method();
4902 int trap_bci = alloc->jvms()->bci();
4903
4904 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4905 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4906 // Make sure there's no store between the allocation and the
4907 // arraycopy otherwise visible side effects could be rexecuted
4908 // in case of deoptimization and cause incorrect execution.
4909 bool no_interfering_store = true;
4910 Node* mem = alloc->in(TypeFunc::Memory);
4911 if (mem->is_MergeMem()) {
4912 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4913 Node* n = mms.memory();
4914 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4915 assert(n->is_Store(), "what else?");
4916 no_interfering_store = false;
4917 break;
4918 }
4919 }
4920 } else {
4921 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4922 Node* n = mms.memory();
4923 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4924 assert(n->is_Store(), "what else?");
4925 no_interfering_store = false;
4926 break;
4927 }
4928 }
4929 }
4930
4931 if (no_interfering_store) {
4932 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4933 uint size = alloc->req();
4934 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4935 old_jvms->set_map(sfpt);
4936 for (uint i = 0; i < size; i++) {
4937 sfpt->init_req(i, alloc->in(i));
4938 }
4939 // re-push array length for deoptimization
4940 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4941 old_jvms->set_sp(old_jvms->sp()+1);
4942 old_jvms->set_monoff(old_jvms->monoff()+1);
4943 old_jvms->set_scloff(old_jvms->scloff()+1);
4944 old_jvms->set_endoff(old_jvms->endoff()+1);
4945 old_jvms->set_should_reexecute(true);
4946
4947 sfpt->set_i_o(map()->i_o());
4948 sfpt->set_memory(map()->memory());
4949 sfpt->set_control(map()->control());
4950
4951 JVMState* saved_jvms = jvms();
4952 saved_reexecute_sp = _reexecute_sp;
4953
4954 set_jvms(sfpt->jvms());
4955 _reexecute_sp = jvms()->sp();
4956
4957 return saved_jvms;
4958 }
4959 }
4960 }
4961 return NULL;
4962 }
4963
4964 // In case of a deoptimization, we restart execution at the
4965 // allocation, allocating a new array. We would leave an uninitialized
4966 // array in the heap that GCs wouldn't expect. Move the allocation
4967 // after the traps so we don't allocate the array if we
4968 // deoptimize. This is possible because tightly_coupled_allocation()
4969 // guarantees there's no observer of the allocated array at this point
4970 // and the control flow is simple enough.
4971 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4972 int saved_reexecute_sp, uint new_idx) {
4973 if (saved_jvms != NULL && !stopped()) {
4974 assert(alloc != NULL, "only with a tightly coupled allocation");
4975 // restore JVM state to the state at the arraycopy
4976 saved_jvms->map()->set_control(map()->control());
4977 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4978 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4979 // If we've improved the types of some nodes (null check) while
4980 // emitting the guards, propagate them to the current state
4981 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4982 set_jvms(saved_jvms);
4983 _reexecute_sp = saved_reexecute_sp;
4984
4985 // Remove the allocation from above the guards
4986 CallProjections callprojs;
4987 alloc->extract_projections(&callprojs, true);
4988 InitializeNode* init = alloc->initialization();
4989 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4990 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4991 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4992 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4993
4994 // move the allocation here (after the guards)
4995 _gvn.hash_delete(alloc);
4996 alloc->set_req(TypeFunc::Control, control());
4997 alloc->set_req(TypeFunc::I_O, i_o());
4998 Node *mem = reset_memory();
4999 set_all_memory(mem);
5000 alloc->set_req(TypeFunc::Memory, mem);
5001 set_control(init->proj_out(TypeFunc::Control));
5002 set_i_o(callprojs.fallthrough_ioproj);
5003
5004 // Update memory as done in GraphKit::set_output_for_allocation()
5005 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5006 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5007 if (ary_type->isa_aryptr() && length_type != NULL) {
5008 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5009 }
5010 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5011 int elemidx = C->get_alias_index(telemref);
5012 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
5013 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
5014
5015 Node* allocx = _gvn.transform(alloc);
5016 assert(allocx == alloc, "where has the allocation gone?");
5017 assert(dest->is_CheckCastPP(), "not an allocation result?");
5018
5019 _gvn.hash_delete(dest);
5020 dest->set_req(0, control());
5021 Node* destx = _gvn.transform(dest);
5022 assert(destx == dest, "where has the allocation result gone?");
5023 }
5024 }
5025
5026
5027 //------------------------------inline_arraycopy-----------------------
5028 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5029 // Object dest, int destPos,
5030 // int length);
5031 bool LibraryCallKit::inline_arraycopy() {
5032 // Get the arguments.
5033 Node* src = argument(0); // type: oop
5034 Node* src_offset = argument(1); // type: int
5035 Node* dest = argument(2); // type: oop
5036 Node* dest_offset = argument(3); // type: int
5037 Node* length = argument(4); // type: int
5038
5039 uint new_idx = C->unique();
5040
5041 // Check for allocation before we add nodes that would confuse
5042 // tightly_coupled_allocation()
5043 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5044
5045 int saved_reexecute_sp = -1;
5046 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5047 // See arraycopy_restore_alloc_state() comment
5048 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5049 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5050 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5051 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5052
5053 // The following tests must be performed
5054 // (1) src and dest are arrays.
5055 // (2) src and dest arrays must have elements of the same BasicType
5056 // (3) src and dest must not be null.
5057 // (4) src_offset must not be negative.
5058 // (5) dest_offset must not be negative.
5059 // (6) length must not be negative.
5060 // (7) src_offset + length must not exceed length of src.
5061 // (8) dest_offset + length must not exceed length of dest.
5062 // (9) each element of an oop array must be assignable
5063
5064 // (3) src and dest must not be null.
5065 // always do this here because we need the JVM state for uncommon traps
5066 Node* null_ctl = top();
5067 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5068 assert(null_ctl->is_top(), "no null control here");
5069 dest = null_check(dest, T_ARRAY);
5070
5071 if (!can_emit_guards) {
5072 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5073 // guards but the arraycopy node could still take advantage of a
5074 // tightly allocated allocation. tightly_coupled_allocation() is
5075 // called again to make sure it takes the null check above into
5076 // account: the null check is mandatory and if it caused an
5077 // uncommon trap to be emitted then the allocation can't be
5078 // considered tightly coupled in this context.
5079 alloc = tightly_coupled_allocation(dest, NULL);
5080 }
5081
5082 bool validated = false;
5083
5084 const Type* src_type = _gvn.type(src);
5085 const Type* dest_type = _gvn.type(dest);
5086 const TypeAryPtr* top_src = src_type->isa_aryptr();
5087 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5088
5089 // Do we have the type of src?
5090 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5091 // Do we have the type of dest?
5092 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5093 // Is the type for src from speculation?
5094 bool src_spec = false;
5095 // Is the type for dest from speculation?
5096 bool dest_spec = false;
5097
5098 if ((!has_src || !has_dest) && can_emit_guards) {
5099 // We don't have sufficient type information, let's see if
5100 // speculative types can help. We need to have types for both src
5101 // and dest so that it pays off.
5102
5103 // Do we already have or could we have type information for src
5104 bool could_have_src = has_src;
5105 // Do we already have or could we have type information for dest
5106 bool could_have_dest = has_dest;
5107
5108 ciKlass* src_k = NULL;
5109 if (!has_src) {
5110 src_k = src_type->speculative_type_not_null();
5111 if (src_k != NULL && src_k->is_array_klass()) {
5112 could_have_src = true;
5113 }
5114 }
5115
5116 ciKlass* dest_k = NULL;
5117 if (!has_dest) {
5118 dest_k = dest_type->speculative_type_not_null();
5119 if (dest_k != NULL && dest_k->is_array_klass()) {
5120 could_have_dest = true;
5121 }
5122 }
5123
5124 if (could_have_src && could_have_dest) {
5125 // This is going to pay off so emit the required guards
5126 if (!has_src) {
5127 src = maybe_cast_profiled_obj(src, src_k, true);
5128 src_type = _gvn.type(src);
5129 top_src = src_type->isa_aryptr();
5130 has_src = (top_src != NULL && top_src->klass() != NULL);
5131 src_spec = true;
5132 }
5133 if (!has_dest) {
5134 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5135 dest_type = _gvn.type(dest);
5136 top_dest = dest_type->isa_aryptr();
5137 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5138 dest_spec = true;
5139 }
5140 }
5141 }
5142
5143 if (has_src && has_dest && can_emit_guards) {
5144 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5145 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5146 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5147 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5148
5149 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5150 // If both arrays are object arrays then having the exact types
5151 // for both will remove the need for a subtype check at runtime
5152 // before the call and may make it possible to pick a faster copy
5153 // routine (without a subtype check on every element)
5154 // Do we have the exact type of src?
5155 bool could_have_src = src_spec;
5156 // Do we have the exact type of dest?
5157 bool could_have_dest = dest_spec;
5158 ciKlass* src_k = top_src->klass();
5159 ciKlass* dest_k = top_dest->klass();
5160 if (!src_spec) {
5161 src_k = src_type->speculative_type_not_null();
5162 if (src_k != NULL && src_k->is_array_klass()) {
5163 could_have_src = true;
5164 }
5165 }
5166 if (!dest_spec) {
5167 dest_k = dest_type->speculative_type_not_null();
5168 if (dest_k != NULL && dest_k->is_array_klass()) {
5169 could_have_dest = true;
5170 }
5171 }
5172 if (could_have_src && could_have_dest) {
5173 // If we can have both exact types, emit the missing guards
5174 if (could_have_src && !src_spec) {
5175 src = maybe_cast_profiled_obj(src, src_k, true);
5176 }
5177 if (could_have_dest && !dest_spec) {
5178 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5179 }
5180 }
5181 }
5182 }
5183
5184 ciMethod* trap_method = method();
5185 int trap_bci = bci();
5186 if (saved_jvms != NULL) {
5187 trap_method = alloc->jvms()->method();
5188 trap_bci = alloc->jvms()->bci();
5189 }
5190
5191 bool negative_length_guard_generated = false;
5192
5193 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5194 can_emit_guards &&
5195 !src->is_top() && !dest->is_top()) {
5196 // validate arguments: enables transformation the ArrayCopyNode
5197 validated = true;
5198
5199 RegionNode* slow_region = new RegionNode(1);
5200 record_for_igvn(slow_region);
5201
5202 // (1) src and dest are arrays.
5203 generate_non_array_guard(load_object_klass(src), slow_region);
5204 generate_non_array_guard(load_object_klass(dest), slow_region);
5205
5206 // (2) src and dest arrays must have elements of the same BasicType
5207 // done at macro expansion or at Ideal transformation time
5208
5209 // (4) src_offset must not be negative.
5210 generate_negative_guard(src_offset, slow_region);
5211
5212 // (5) dest_offset must not be negative.
5213 generate_negative_guard(dest_offset, slow_region);
5214
5215 // (7) src_offset + length must not exceed length of src.
5216 generate_limit_guard(src_offset, length,
5217 load_array_length(src),
5218 slow_region);
5219
5220 // (8) dest_offset + length must not exceed length of dest.
5221 generate_limit_guard(dest_offset, length,
5222 load_array_length(dest),
5223 slow_region);
5224
5225 // (6) length must not be negative.
5226 // This is also checked in generate_arraycopy() during macro expansion, but
5227 // we also have to check it here for the case where the ArrayCopyNode will
5228 // be eliminated by Escape Analysis.
5229 if (EliminateAllocations) {
5230 generate_negative_guard(length, slow_region);
5231 negative_length_guard_generated = true;
5232 }
5233
5234 // (9) each element of an oop array must be assignable
5235 Node* src_klass = load_object_klass(src);
5236 Node* dest_klass = load_object_klass(dest);
5237 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5238
5239 if (not_subtype_ctrl != top()) {
5240 PreserveJVMState pjvms(this);
5241 set_control(not_subtype_ctrl);
5242 uncommon_trap(Deoptimization::Reason_intrinsic,
5243 Deoptimization::Action_make_not_entrant);
5244 assert(stopped(), "Should be stopped");
5245 }
5246 {
5247 PreserveJVMState pjvms(this);
5248 set_control(_gvn.transform(slow_region));
5249 uncommon_trap(Deoptimization::Reason_intrinsic,
5250 Deoptimization::Action_make_not_entrant);
5251 assert(stopped(), "Should be stopped");
5252 }
5253 }
5254
5255 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5256
5257 if (stopped()) {
5258 return true;
5259 }
5260
5261 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5262 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5263 // so the compiler has a chance to eliminate them: during macro expansion,
5264 // we have to set their control (CastPP nodes are eliminated).
5265 load_object_klass(src), load_object_klass(dest),
5266 load_array_length(src), load_array_length(dest));
5267
5268 ac->set_arraycopy(validated);
5269
5270 Node* n = _gvn.transform(ac);
5271 if (n == ac) {
5272 ac->connect_outputs(this);
5273 } else {
5274 assert(validated, "shouldn't transform if all arguments not validated");
5275 set_all_memory(n);
5276 }
5277
5278 return true;
5279 }
5280
5281
5282 // Helper function which determines if an arraycopy immediately follows
5283 // an allocation, with no intervening tests or other escapes for the object.
5284 AllocateArrayNode*
5285 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5286 RegionNode* slow_region) {
5287 if (stopped()) return NULL; // no fast path
5288 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5289
5290 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5291 if (alloc == NULL) return NULL;
5292
5293 Node* rawmem = memory(Compile::AliasIdxRaw);
5294 // Is the allocation's memory state untouched?
5295 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5296 // Bail out if there have been raw-memory effects since the allocation.
5297 // (Example: There might have been a call or safepoint.)
5298 return NULL;
5299 }
5300 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5301 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5302 return NULL;
5303 }
5304
5305 // There must be no unexpected observers of this allocation.
5306 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5307 Node* obs = ptr->fast_out(i);
5308 if (obs != this->map()) {
5309 return NULL;
5310 }
5311 }
5312
5313 // This arraycopy must unconditionally follow the allocation of the ptr.
5314 Node* alloc_ctl = ptr->in(0);
5315 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5316
5317 Node* ctl = control();
5318 while (ctl != alloc_ctl) {
5319 // There may be guards which feed into the slow_region.
5320 // Any other control flow means that we might not get a chance
5321 // to finish initializing the allocated object.
5322 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5323 IfNode* iff = ctl->in(0)->as_If();
5324 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5325 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5326 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5327 ctl = iff->in(0); // This test feeds the known slow_region.
5328 continue;
5329 }
5330 // One more try: Various low-level checks bottom out in
5331 // uncommon traps. If the debug-info of the trap omits
5332 // any reference to the allocation, as we've already
5333 // observed, then there can be no objection to the trap.
5334 bool found_trap = false;
5335 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5336 Node* obs = not_ctl->fast_out(j);
5337 if (obs->in(0) == not_ctl && obs->is_Call() &&
5338 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5339 found_trap = true; break;
5340 }
5341 }
5342 if (found_trap) {
5343 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5344 continue;
5345 }
5346 }
5347 return NULL;
5348 }
5349
5350 // If we get this far, we have an allocation which immediately
5351 // precedes the arraycopy, and we can take over zeroing the new object.
5352 // The arraycopy will finish the initialization, and provide
5353 // a new control state to which we will anchor the destination pointer.
5354
5355 return alloc;
5356 }
5357
5358 //-------------inline_encodeISOArray-----------------------------------
5359 // encode char[] to byte[] in ISO_8859_1
5360 bool LibraryCallKit::inline_encodeISOArray() {
5361 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5362 // no receiver since it is static method
5363 Node *src = argument(0);
5364 Node *src_offset = argument(1);
5365 Node *dst = argument(2);
5366 Node *dst_offset = argument(3);
5367 Node *length = argument(4);
5368
5369 const Type* src_type = src->Value(&_gvn);
5370 const Type* dst_type = dst->Value(&_gvn);
5371 const TypeAryPtr* top_src = src_type->isa_aryptr();
5372 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5373 if (top_src == NULL || top_src->klass() == NULL ||
5374 top_dest == NULL || top_dest->klass() == NULL) {
5375 // failed array check
5376 return false;
5377 }
5378
5379 // Figure out the size and type of the elements we will be copying.
5380 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5381 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5382 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5383 return false;
5384 }
5385
5386 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5387 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5388 // 'src_start' points to src array + scaled offset
5389 // 'dst_start' points to dst array + scaled offset
5390
5391 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5392 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5393 enc = _gvn.transform(enc);
5394 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5395 set_memory(res_mem, mtype);
5396 set_result(enc);
5397 return true;
5398 }
5399
5400 //-------------inline_multiplyToLen-----------------------------------
5401 bool LibraryCallKit::inline_multiplyToLen() {
5402 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5403
5404 address stubAddr = StubRoutines::multiplyToLen();
5405 if (stubAddr == NULL) {
5406 return false; // Intrinsic's stub is not implemented on this platform
5407 }
5408 const char* stubName = "multiplyToLen";
5409
5410 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5411
5412 // no receiver because it is a static method
5413 Node* x = argument(0);
5414 Node* xlen = argument(1);
5415 Node* y = argument(2);
5416 Node* ylen = argument(3);
5417 Node* z = argument(4);
5418
5419 const Type* x_type = x->Value(&_gvn);
5420 const Type* y_type = y->Value(&_gvn);
5421 const TypeAryPtr* top_x = x_type->isa_aryptr();
5422 const TypeAryPtr* top_y = y_type->isa_aryptr();
5423 if (top_x == NULL || top_x->klass() == NULL ||
5424 top_y == NULL || top_y->klass() == NULL) {
5425 // failed array check
5426 return false;
5427 }
5428
5429 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5430 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5431 if (x_elem != T_INT || y_elem != T_INT) {
5432 return false;
5433 }
5434
5435 // Set the original stack and the reexecute bit for the interpreter to reexecute
5436 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5437 // on the return from z array allocation in runtime.
5438 { PreserveReexecuteState preexecs(this);
5439 jvms()->set_should_reexecute(true);
5440
5441 Node* x_start = array_element_address(x, intcon(0), x_elem);
5442 Node* y_start = array_element_address(y, intcon(0), y_elem);
5443 // 'x_start' points to x array + scaled xlen
5444 // 'y_start' points to y array + scaled ylen
5445
5446 // Allocate the result array
5447 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5448 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5449 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5450
5451 IdealKit ideal(this);
5452
5453 #define __ ideal.
5454 Node* one = __ ConI(1);
5455 Node* zero = __ ConI(0);
5456 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5457 __ set(need_alloc, zero);
5458 __ set(z_alloc, z);
5459 __ if_then(z, BoolTest::eq, null()); {
5460 __ increment (need_alloc, one);
5461 } __ else_(); {
5462 // Update graphKit memory and control from IdealKit.
5463 sync_kit(ideal);
5464 Node* zlen_arg = load_array_length(z);
5465 // Update IdealKit memory and control from graphKit.
5466 __ sync_kit(this);
5467 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5468 __ increment (need_alloc, one);
5469 } __ end_if();
5470 } __ end_if();
5471
5472 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5473 // Update graphKit memory and control from IdealKit.
5474 sync_kit(ideal);
5475 Node * narr = new_array(klass_node, zlen, 1);
5476 // Update IdealKit memory and control from graphKit.
5477 __ sync_kit(this);
5478 __ set(z_alloc, narr);
5479 } __ end_if();
5480
5481 sync_kit(ideal);
5482 z = __ value(z_alloc);
5483 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5484 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5485 // Final sync IdealKit and GraphKit.
5486 final_sync(ideal);
5487 #undef __
5488
5489 Node* z_start = array_element_address(z, intcon(0), T_INT);
5490
5491 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5492 OptoRuntime::multiplyToLen_Type(),
5493 stubAddr, stubName, TypePtr::BOTTOM,
5494 x_start, xlen, y_start, ylen, z_start, zlen);
5495 } // original reexecute is set back here
5496
5497 C->set_has_split_ifs(true); // Has chance for split-if optimization
5498 set_result(z);
5499 return true;
5500 }
5501
5502 //-------------inline_squareToLen------------------------------------
5503 bool LibraryCallKit::inline_squareToLen() {
5504 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5505
5506 address stubAddr = StubRoutines::squareToLen();
5507 if (stubAddr == NULL) {
5508 return false; // Intrinsic's stub is not implemented on this platform
5509 }
5510 const char* stubName = "squareToLen";
5511
5512 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5513
5514 Node* x = argument(0);
5515 Node* len = argument(1);
5516 Node* z = argument(2);
5517 Node* zlen = argument(3);
5518
5519 const Type* x_type = x->Value(&_gvn);
5520 const Type* z_type = z->Value(&_gvn);
5521 const TypeAryPtr* top_x = x_type->isa_aryptr();
5522 const TypeAryPtr* top_z = z_type->isa_aryptr();
5523 if (top_x == NULL || top_x->klass() == NULL ||
5524 top_z == NULL || top_z->klass() == NULL) {
5525 // failed array check
5526 return false;
5527 }
5528
5529 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5530 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5531 if (x_elem != T_INT || z_elem != T_INT) {
5532 return false;
5533 }
5534
5535
5536 Node* x_start = array_element_address(x, intcon(0), x_elem);
5537 Node* z_start = array_element_address(z, intcon(0), z_elem);
5538
5539 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5540 OptoRuntime::squareToLen_Type(),
5541 stubAddr, stubName, TypePtr::BOTTOM,
5542 x_start, len, z_start, zlen);
5543
5544 set_result(z);
5545 return true;
5546 }
5547
5548 //-------------inline_mulAdd------------------------------------------
5549 bool LibraryCallKit::inline_mulAdd() {
5550 assert(UseMulAddIntrinsic, "not implemented on this platform");
5551
5552 address stubAddr = StubRoutines::mulAdd();
5553 if (stubAddr == NULL) {
5554 return false; // Intrinsic's stub is not implemented on this platform
5555 }
5556 const char* stubName = "mulAdd";
5557
5558 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5559
5560 Node* out = argument(0);
5561 Node* in = argument(1);
5562 Node* offset = argument(2);
5563 Node* len = argument(3);
5564 Node* k = argument(4);
5565
5566 const Type* out_type = out->Value(&_gvn);
5567 const Type* in_type = in->Value(&_gvn);
5568 const TypeAryPtr* top_out = out_type->isa_aryptr();
5569 const TypeAryPtr* top_in = in_type->isa_aryptr();
5570 if (top_out == NULL || top_out->klass() == NULL ||
5571 top_in == NULL || top_in->klass() == NULL) {
5572 // failed array check
5573 return false;
5574 }
5575
5576 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5577 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5578 if (out_elem != T_INT || in_elem != T_INT) {
5579 return false;
5580 }
5581
5582 Node* outlen = load_array_length(out);
5583 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5584 Node* out_start = array_element_address(out, intcon(0), out_elem);
5585 Node* in_start = array_element_address(in, intcon(0), in_elem);
5586
5587 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5588 OptoRuntime::mulAdd_Type(),
5589 stubAddr, stubName, TypePtr::BOTTOM,
5590 out_start,in_start, new_offset, len, k);
5591 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5592 set_result(result);
5593 return true;
5594 }
5595
5596 //-------------inline_montgomeryMultiply-----------------------------------
5597 bool LibraryCallKit::inline_montgomeryMultiply() {
5598 address stubAddr = StubRoutines::montgomeryMultiply();
5599 if (stubAddr == NULL) {
5600 return false; // Intrinsic's stub is not implemented on this platform
5601 }
5602
5603 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5604 const char* stubName = "montgomery_multiply";
5605
5606 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5607
5608 Node* a = argument(0);
5609 Node* b = argument(1);
5610 Node* n = argument(2);
5611 Node* len = argument(3);
5612 Node* inv = argument(4);
5613 Node* m = argument(6);
5614
5615 const Type* a_type = a->Value(&_gvn);
5616 const TypeAryPtr* top_a = a_type->isa_aryptr();
5617 const Type* b_type = b->Value(&_gvn);
5618 const TypeAryPtr* top_b = b_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_b == NULL || top_b->klass() == NULL ||
5625 top_n == NULL || top_n->klass() == NULL ||
5626 top_m == NULL || top_m->klass() == NULL) {
5627 // failed array check
5628 return false;
5629 }
5630
5631 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5632 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5633 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5634 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5635 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5636 return false;
5637 }
5638
5639 // Make the call
5640 {
5641 Node* a_start = array_element_address(a, intcon(0), a_elem);
5642 Node* b_start = array_element_address(b, intcon(0), b_elem);
5643 Node* n_start = array_element_address(n, intcon(0), n_elem);
5644 Node* m_start = array_element_address(m, intcon(0), m_elem);
5645
5646 Node* call = make_runtime_call(RC_LEAF,
5647 OptoRuntime::montgomeryMultiply_Type(),
5648 stubAddr, stubName, TypePtr::BOTTOM,
5649 a_start, b_start, n_start, len, inv, top(),
5650 m_start);
5651 set_result(m);
5652 }
5653
5654 return true;
5655 }
5656
5657 bool LibraryCallKit::inline_montgomerySquare() {
5658 address stubAddr = StubRoutines::montgomerySquare();
5659 if (stubAddr == NULL) {
5660 return false; // Intrinsic's stub is not implemented on this platform
5661 }
5662
5663 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5664 const char* stubName = "montgomery_square";
5665
5666 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5667
5668 Node* a = argument(0);
5669 Node* n = argument(1);
5670 Node* len = argument(2);
5671 Node* inv = argument(3);
5672 Node* m = argument(5);
5673
5674 const Type* a_type = a->Value(&_gvn);
5675 const TypeAryPtr* top_a = a_type->isa_aryptr();
5676 const Type* n_type = a->Value(&_gvn);
5677 const TypeAryPtr* top_n = n_type->isa_aryptr();
5678 const Type* m_type = a->Value(&_gvn);
5679 const TypeAryPtr* top_m = m_type->isa_aryptr();
5680 if (top_a == NULL || top_a->klass() == NULL ||
5681 top_n == NULL || top_n->klass() == NULL ||
5682 top_m == NULL || top_m->klass() == NULL) {
5683 // failed array check
5684 return false;
5685 }
5686
5687 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5688 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5689 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5690 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5691 return false;
5692 }
5693
5694 // Make the call
5695 {
5696 Node* a_start = array_element_address(a, intcon(0), a_elem);
5697 Node* n_start = array_element_address(n, intcon(0), n_elem);
5698 Node* m_start = array_element_address(m, intcon(0), m_elem);
5699
5700 Node* call = make_runtime_call(RC_LEAF,
5701 OptoRuntime::montgomerySquare_Type(),
5702 stubAddr, stubName, TypePtr::BOTTOM,
5703 a_start, n_start, len, inv, top(),
5704 m_start);
5705 set_result(m);
5706 }
5707
5708 return true;
5709 }
5710
5711 //-------------inline_vectorizedMismatch------------------------------
5712 bool LibraryCallKit::inline_vectorizedMismatch() {
5713 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5714
5715 address stubAddr = StubRoutines::vectorizedMismatch();
5716 if (stubAddr == NULL) {
5717 return false; // Intrinsic's stub is not implemented on this platform
5718 }
5719 const char* stubName = "vectorizedMismatch";
5720 int size_l = callee()->signature()->size();
5721 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5722
5723 Node* obja = argument(0);
5724 Node* aoffset = argument(1);
5725 Node* objb = argument(3);
5726 Node* boffset = argument(4);
5727 Node* length = argument(6);
5728 Node* scale = argument(7);
5729
5730 const Type* a_type = obja->Value(&_gvn);
5731 const Type* b_type = objb->Value(&_gvn);
5732 const TypeAryPtr* top_a = a_type->isa_aryptr();
5733 const TypeAryPtr* top_b = b_type->isa_aryptr();
5734 if (top_a == NULL || top_a->klass() == NULL ||
5735 top_b == NULL || top_b->klass() == NULL) {
5736 // failed array check
5737 return false;
5738 }
5739
5740 Node* call;
5741 jvms()->set_should_reexecute(true);
5742
5743 Node* obja_adr = make_unsafe_address(obja, aoffset);
5744 Node* objb_adr = make_unsafe_address(objb, boffset);
5745
5746 call = make_runtime_call(RC_LEAF,
5747 OptoRuntime::vectorizedMismatch_Type(),
5748 stubAddr, stubName, TypePtr::BOTTOM,
5749 obja_adr, objb_adr, length, scale);
5750
5751 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5752 set_result(result);
5753 return true;
5754 }
5755
5756 /**
5757 * Calculate CRC32 for byte.
5758 * int java.util.zip.CRC32.update(int crc, int b)
5759 */
5760 bool LibraryCallKit::inline_updateCRC32() {
5761 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5762 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5763 // no receiver since it is static method
5764 Node* crc = argument(0); // type: int
5765 Node* b = argument(1); // type: int
5766
5767 /*
5768 * int c = ~ crc;
5769 * b = timesXtoThe32[(b ^ c) & 0xFF];
5770 * b = b ^ (c >>> 8);
5771 * crc = ~b;
5772 */
5773
5774 Node* M1 = intcon(-1);
5775 crc = _gvn.transform(new XorINode(crc, M1));
5776 Node* result = _gvn.transform(new XorINode(crc, b));
5777 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5778
5779 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5780 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5781 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5782 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5783
5784 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5785 result = _gvn.transform(new XorINode(crc, result));
5786 result = _gvn.transform(new XorINode(result, M1));
5787 set_result(result);
5788 return true;
5789 }
5790
5791 /**
5792 * Calculate CRC32 for byte[] array.
5793 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5794 */
5795 bool LibraryCallKit::inline_updateBytesCRC32() {
5796 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5797 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5798 // no receiver since it is static method
5799 Node* crc = argument(0); // type: int
5800 Node* src = argument(1); // type: oop
5801 Node* offset = argument(2); // type: int
5802 Node* length = argument(3); // type: int
5803
5804 const Type* src_type = src->Value(&_gvn);
5805 const TypeAryPtr* top_src = src_type->isa_aryptr();
5806 if (top_src == NULL || top_src->klass() == NULL) {
5807 // failed array check
5808 return false;
5809 }
5810
5811 // Figure out the size and type of the elements we will be copying.
5812 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5813 if (src_elem != T_BYTE) {
5814 return false;
5815 }
5816
5817 // 'src_start' points to src array + scaled offset
5818 Node* src_start = array_element_address(src, offset, src_elem);
5819
5820 // We assume that range check is done by caller.
5821 // TODO: generate range check (offset+length < src.length) in debug VM.
5822
5823 // Call the stub.
5824 address stubAddr = StubRoutines::updateBytesCRC32();
5825 const char *stubName = "updateBytesCRC32";
5826
5827 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5828 stubAddr, stubName, TypePtr::BOTTOM,
5829 crc, src_start, length);
5830 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5831 set_result(result);
5832 return true;
5833 }
5834
5835 /**
5836 * Calculate CRC32 for ByteBuffer.
5837 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5838 */
5839 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5840 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5841 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5842 // no receiver since it is static method
5843 Node* crc = argument(0); // type: int
5844 Node* src = argument(1); // type: long
5845 Node* offset = argument(3); // type: int
5846 Node* length = argument(4); // type: int
5847
5848 src = ConvL2X(src); // adjust Java long to machine word
5849 Node* base = _gvn.transform(new CastX2PNode(src));
5850 offset = ConvI2X(offset);
5851
5852 // 'src_start' points to src array + scaled offset
5853 Node* src_start = basic_plus_adr(top(), base, offset);
5854
5855 // Call the stub.
5856 address stubAddr = StubRoutines::updateBytesCRC32();
5857 const char *stubName = "updateBytesCRC32";
5858
5859 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5860 stubAddr, stubName, TypePtr::BOTTOM,
5861 crc, src_start, length);
5862 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5863 set_result(result);
5864 return true;
5865 }
5866
5867 //------------------------------get_table_from_crc32c_class-----------------------
5868 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5869 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5870 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5871
5872 return table;
5873 }
5874
5875 //------------------------------inline_updateBytesCRC32C-----------------------
5876 //
5877 // Calculate CRC32C for byte[] array.
5878 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5879 //
5880 bool LibraryCallKit::inline_updateBytesCRC32C() {
5881 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5882 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5883 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5884 // no receiver since it is a static method
5885 Node* crc = argument(0); // type: int
5886 Node* src = argument(1); // type: oop
5887 Node* offset = argument(2); // type: int
5888 Node* end = argument(3); // type: int
5889
5890 Node* length = _gvn.transform(new SubINode(end, offset));
5891
5892 const Type* src_type = src->Value(&_gvn);
5893 const TypeAryPtr* top_src = src_type->isa_aryptr();
5894 if (top_src == NULL || top_src->klass() == NULL) {
5895 // failed array check
5896 return false;
5897 }
5898
5899 // Figure out the size and type of the elements we will be copying.
5900 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5901 if (src_elem != T_BYTE) {
5902 return false;
5903 }
5904
5905 // 'src_start' points to src array + scaled offset
5906 Node* src_start = array_element_address(src, offset, src_elem);
5907
5908 // static final int[] byteTable in class CRC32C
5909 Node* table = get_table_from_crc32c_class(callee()->holder());
5910 Node* table_start = array_element_address(table, intcon(0), T_INT);
5911
5912 // We assume that range check is done by caller.
5913 // TODO: generate range check (offset+length < src.length) in debug VM.
5914
5915 // Call the stub.
5916 address stubAddr = StubRoutines::updateBytesCRC32C();
5917 const char *stubName = "updateBytesCRC32C";
5918
5919 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5920 stubAddr, stubName, TypePtr::BOTTOM,
5921 crc, src_start, length, table_start);
5922 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5923 set_result(result);
5924 return true;
5925 }
5926
5927 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5928 //
5929 // Calculate CRC32C for DirectByteBuffer.
5930 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5931 //
5932 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5933 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5934 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5935 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5936 // no receiver since it is a static method
5937 Node* crc = argument(0); // type: int
5938 Node* src = argument(1); // type: long
5939 Node* offset = argument(3); // type: int
5940 Node* end = argument(4); // type: int
5941
5942 Node* length = _gvn.transform(new SubINode(end, offset));
5943
5944 src = ConvL2X(src); // adjust Java long to machine word
5945 Node* base = _gvn.transform(new CastX2PNode(src));
5946 offset = ConvI2X(offset);
5947
5948 // 'src_start' points to src array + scaled offset
5949 Node* src_start = basic_plus_adr(top(), base, offset);
5950
5951 // static final int[] byteTable in class CRC32C
5952 Node* table = get_table_from_crc32c_class(callee()->holder());
5953 Node* table_start = array_element_address(table, intcon(0), T_INT);
5954
5955 // Call the stub.
5956 address stubAddr = StubRoutines::updateBytesCRC32C();
5957 const char *stubName = "updateBytesCRC32C";
5958
5959 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5960 stubAddr, stubName, TypePtr::BOTTOM,
5961 crc, src_start, length, table_start);
5962 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5963 set_result(result);
5964 return true;
5965 }
5966
5967 //------------------------------inline_updateBytesAdler32----------------------
5968 //
5969 // Calculate Adler32 checksum for byte[] array.
5970 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5971 //
5972 bool LibraryCallKit::inline_updateBytesAdler32() {
5973 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5974 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5975 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5976 // no receiver since it is static method
5977 Node* crc = argument(0); // type: int
5978 Node* src = argument(1); // type: oop
5979 Node* offset = argument(2); // type: int
5980 Node* length = argument(3); // type: int
5981
5982 const Type* src_type = src->Value(&_gvn);
5983 const TypeAryPtr* top_src = src_type->isa_aryptr();
5984 if (top_src == NULL || top_src->klass() == NULL) {
5985 // failed array check
5986 return false;
5987 }
5988
5989 // Figure out the size and type of the elements we will be copying.
5990 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5991 if (src_elem != T_BYTE) {
5992 return false;
5993 }
5994
5995 // 'src_start' points to src array + scaled offset
5996 Node* src_start = array_element_address(src, offset, src_elem);
5997
5998 // We assume that range check is done by caller.
5999 // TODO: generate range check (offset+length < src.length) in debug VM.
6000
6001 // Call the stub.
6002 address stubAddr = StubRoutines::updateBytesAdler32();
6003 const char *stubName = "updateBytesAdler32";
6004
6005 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6006 stubAddr, stubName, TypePtr::BOTTOM,
6007 crc, src_start, length);
6008 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6009 set_result(result);
6010 return true;
6011 }
6012
6013 //------------------------------inline_updateByteBufferAdler32---------------
6014 //
6015 // Calculate Adler32 checksum for DirectByteBuffer.
6016 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6017 //
6018 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6019 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6020 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6021 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6022 // no receiver since it is static method
6023 Node* crc = argument(0); // type: int
6024 Node* src = argument(1); // type: long
6025 Node* offset = argument(3); // type: int
6026 Node* length = argument(4); // type: int
6027
6028 src = ConvL2X(src); // adjust Java long to machine word
6029 Node* base = _gvn.transform(new CastX2PNode(src));
6030 offset = ConvI2X(offset);
6031
6032 // 'src_start' points to src array + scaled offset
6033 Node* src_start = basic_plus_adr(top(), base, offset);
6034
6035 // Call the stub.
6036 address stubAddr = StubRoutines::updateBytesAdler32();
6037 const char *stubName = "updateBytesAdler32";
6038
6039 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6040 stubAddr, stubName, TypePtr::BOTTOM,
6041 crc, src_start, length);
6042
6043 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6044 set_result(result);
6045 return true;
6046 }
6047
6048 //----------------------------inline_reference_get----------------------------
6049 // public T java.lang.ref.Reference.get();
6050 bool LibraryCallKit::inline_reference_get() {
6051 const int referent_offset = java_lang_ref_Reference::referent_offset;
6052 guarantee(referent_offset > 0, "should have already been set");
6053
6054 // Get the argument:
6055 Node* reference_obj = null_check_receiver();
6056 if (stopped()) return true;
6057
6058 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6059
6060 ciInstanceKlass* klass = env()->Object_klass();
6061 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6062
6063 Node* no_ctrl = NULL;
6064 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6065
6066 // Use the pre-barrier to record the value in the referent field
6067 pre_barrier(false /* do_load */,
6068 control(),
6069 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6070 result /* pre_val */,
6071 T_OBJECT);
6072
6073 // Add memory barrier to prevent commoning reads from this field
6074 // across safepoint since GC can change its value.
6075 insert_mem_bar(Op_MemBarCPUOrder);
6076
6077 set_result(result);
6078 return true;
6079 }
6080
6081
6082 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6083 bool is_exact=true, bool is_static=false,
6084 ciInstanceKlass * fromKls=NULL) {
6085 if (fromKls == NULL) {
6086 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6087 assert(tinst != NULL, "obj is null");
6088 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6089 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6090 fromKls = tinst->klass()->as_instance_klass();
6091 } else {
6092 assert(is_static, "only for static field access");
6093 }
6094 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6095 ciSymbol::make(fieldTypeString),
6096 is_static);
6097
6098 assert (field != NULL, "undefined field");
6099 if (field == NULL) return (Node *) NULL;
6100
6101 if (is_static) {
6102 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6103 fromObj = makecon(tip);
6104 }
6105
6106 // Next code copied from Parse::do_get_xxx():
6107
6108 // Compute address and memory type.
6109 int offset = field->offset_in_bytes();
6110 bool is_vol = field->is_volatile();
6111 ciType* field_klass = field->type();
6112 assert(field_klass->is_loaded(), "should be loaded");
6113 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6114 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6115 BasicType bt = field->layout_type();
6116
6117 // Build the resultant type of the load
6118 const Type *type;
6119 if (bt == T_OBJECT) {
6120 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6121 } else {
6122 type = Type::get_const_basic_type(bt);
6123 }
6124
6125 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6126 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6127 }
6128 // Build the load.
6129 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6130 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6131 // If reference is volatile, prevent following memory ops from
6132 // floating up past the volatile read. Also prevents commoning
6133 // another volatile read.
6134 if (is_vol) {
6135 // Memory barrier includes bogus read of value to force load BEFORE membar
6136 insert_mem_bar(Op_MemBarAcquire, loadedField);
6137 }
6138 return loadedField;
6139 }
6140
6141 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6142 bool is_exact = true, bool is_static = false,
6143 ciInstanceKlass * fromKls = NULL) {
6144 if (fromKls == NULL) {
6145 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6146 assert(tinst != NULL, "obj is null");
6147 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6148 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6149 fromKls = tinst->klass()->as_instance_klass();
6150 }
6151 else {
6152 assert(is_static, "only for static field access");
6153 }
6154 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6155 ciSymbol::make(fieldTypeString),
6156 is_static);
6157
6158 assert(field != NULL, "undefined field");
6159 assert(!field->is_volatile(), "not defined for volatile fields");
6160
6161 if (is_static) {
6162 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6163 fromObj = makecon(tip);
6164 }
6165
6166 // Next code copied from Parse::do_get_xxx():
6167
6168 // Compute address and memory type.
6169 int offset = field->offset_in_bytes();
6170 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6171
6172 return adr;
6173 }
6174
6175 //------------------------------inline_aescrypt_Block-----------------------
6176 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6177 address stubAddr = NULL;
6178 const char *stubName;
6179 assert(UseAES, "need AES instruction support");
6180
6181 switch(id) {
6182 case vmIntrinsics::_aescrypt_encryptBlock:
6183 stubAddr = StubRoutines::aescrypt_encryptBlock();
6184 stubName = "aescrypt_encryptBlock";
6185 break;
6186 case vmIntrinsics::_aescrypt_decryptBlock:
6187 stubAddr = StubRoutines::aescrypt_decryptBlock();
6188 stubName = "aescrypt_decryptBlock";
6189 break;
6190 default:
6191 break;
6192 }
6193 if (stubAddr == NULL) return false;
6194
6195 Node* aescrypt_object = argument(0);
6196 Node* src = argument(1);
6197 Node* src_offset = argument(2);
6198 Node* dest = argument(3);
6199 Node* dest_offset = argument(4);
6200
6201 // (1) src and dest are arrays.
6202 const Type* src_type = src->Value(&_gvn);
6203 const Type* dest_type = dest->Value(&_gvn);
6204 const TypeAryPtr* top_src = src_type->isa_aryptr();
6205 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6206 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6207
6208 // for the quick and dirty code we will skip all the checks.
6209 // we are just trying to get the call to be generated.
6210 Node* src_start = src;
6211 Node* dest_start = dest;
6212 if (src_offset != NULL || dest_offset != NULL) {
6213 assert(src_offset != NULL && dest_offset != NULL, "");
6214 src_start = array_element_address(src, src_offset, T_BYTE);
6215 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6216 }
6217
6218 // now need to get the start of its expanded key array
6219 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6220 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6221 if (k_start == NULL) return false;
6222
6223 if (Matcher::pass_original_key_for_aes()) {
6224 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6225 // compatibility issues between Java key expansion and SPARC crypto instructions
6226 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6227 if (original_k_start == NULL) return false;
6228
6229 // Call the stub.
6230 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6231 stubAddr, stubName, TypePtr::BOTTOM,
6232 src_start, dest_start, k_start, original_k_start);
6233 } else {
6234 // Call the stub.
6235 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6236 stubAddr, stubName, TypePtr::BOTTOM,
6237 src_start, dest_start, k_start);
6238 }
6239
6240 return true;
6241 }
6242
6243 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6244 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6245 address stubAddr = NULL;
6246 const char *stubName = NULL;
6247
6248 assert(UseAES, "need AES instruction support");
6249
6250 switch(id) {
6251 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6252 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6253 stubName = "cipherBlockChaining_encryptAESCrypt";
6254 break;
6255 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6256 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6257 stubName = "cipherBlockChaining_decryptAESCrypt";
6258 break;
6259 default:
6260 break;
6261 }
6262 if (stubAddr == NULL) return false;
6263
6264 Node* cipherBlockChaining_object = argument(0);
6265 Node* src = argument(1);
6266 Node* src_offset = argument(2);
6267 Node* len = argument(3);
6268 Node* dest = argument(4);
6269 Node* dest_offset = argument(5);
6270
6271 // (1) src and dest are arrays.
6272 const Type* src_type = src->Value(&_gvn);
6273 const Type* dest_type = dest->Value(&_gvn);
6274 const TypeAryPtr* top_src = src_type->isa_aryptr();
6275 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6276 assert (top_src != NULL && top_src->klass() != NULL
6277 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6278
6279 // checks are the responsibility of the caller
6280 Node* src_start = src;
6281 Node* dest_start = dest;
6282 if (src_offset != NULL || dest_offset != NULL) {
6283 assert(src_offset != NULL && dest_offset != NULL, "");
6284 src_start = array_element_address(src, src_offset, T_BYTE);
6285 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6286 }
6287
6288 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6289 // (because of the predicated logic executed earlier).
6290 // so we cast it here safely.
6291 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6292
6293 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6294 if (embeddedCipherObj == NULL) return false;
6295
6296 // cast it to what we know it will be at runtime
6297 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6298 assert(tinst != NULL, "CBC obj is null");
6299 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6300 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6301 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6302
6303 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6304 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6305 const TypeOopPtr* xtype = aklass->as_instance_type();
6306 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6307 aescrypt_object = _gvn.transform(aescrypt_object);
6308
6309 // we need to get the start of the aescrypt_object's expanded key array
6310 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6311 if (k_start == NULL) return false;
6312
6313 // similarly, get the start address of the r vector
6314 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6315 if (objRvec == NULL) return false;
6316 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6317
6318 Node* cbcCrypt;
6319 if (Matcher::pass_original_key_for_aes()) {
6320 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6321 // compatibility issues between Java key expansion and SPARC crypto instructions
6322 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6323 if (original_k_start == NULL) return false;
6324
6325 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6326 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6327 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6328 stubAddr, stubName, TypePtr::BOTTOM,
6329 src_start, dest_start, k_start, r_start, len, original_k_start);
6330 } else {
6331 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6332 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6333 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6334 stubAddr, stubName, TypePtr::BOTTOM,
6335 src_start, dest_start, k_start, r_start, len);
6336 }
6337
6338 // return cipher length (int)
6339 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6340 set_result(retvalue);
6341 return true;
6342 }
6343
6344 //------------------------------inline_counterMode_AESCrypt-----------------------
6345 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6346 assert(UseAES, "need AES instruction support");
6347 if (!UseAESCTRIntrinsics) return false;
6348
6349 address stubAddr = NULL;
6350 const char *stubName = NULL;
6351 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6352 stubAddr = StubRoutines::counterMode_AESCrypt();
6353 stubName = "counterMode_AESCrypt";
6354 }
6355 if (stubAddr == NULL) return false;
6356
6357 Node* counterMode_object = argument(0);
6358 Node* src = argument(1);
6359 Node* src_offset = argument(2);
6360 Node* len = argument(3);
6361 Node* dest = argument(4);
6362 Node* dest_offset = argument(5);
6363
6364 // (1) src and dest are arrays.
6365 const Type* src_type = src->Value(&_gvn);
6366 const Type* dest_type = dest->Value(&_gvn);
6367 const TypeAryPtr* top_src = src_type->isa_aryptr();
6368 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6369 assert(top_src != NULL && top_src->klass() != NULL &&
6370 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6371
6372 // checks are the responsibility of the caller
6373 Node* src_start = src;
6374 Node* dest_start = dest;
6375 if (src_offset != NULL || dest_offset != NULL) {
6376 assert(src_offset != NULL && dest_offset != NULL, "");
6377 src_start = array_element_address(src, src_offset, T_BYTE);
6378 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6379 }
6380
6381 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6382 // (because of the predicated logic executed earlier).
6383 // so we cast it here safely.
6384 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6385 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6386 if (embeddedCipherObj == NULL) return false;
6387 // cast it to what we know it will be at runtime
6388 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6389 assert(tinst != NULL, "CTR obj is null");
6390 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6391 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6392 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6393 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6394 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6395 const TypeOopPtr* xtype = aklass->as_instance_type();
6396 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6397 aescrypt_object = _gvn.transform(aescrypt_object);
6398 // we need to get the start of the aescrypt_object's expanded key array
6399 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6400 if (k_start == NULL) return false;
6401 // similarly, get the start address of the r vector
6402 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6403 if (obj_counter == NULL) return false;
6404 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6405
6406 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6407 if (saved_encCounter == NULL) return false;
6408 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6409 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6410
6411 Node* ctrCrypt;
6412 if (Matcher::pass_original_key_for_aes()) {
6413 // no SPARC version for AES/CTR intrinsics now.
6414 return false;
6415 }
6416 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6417 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6418 OptoRuntime::counterMode_aescrypt_Type(),
6419 stubAddr, stubName, TypePtr::BOTTOM,
6420 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6421
6422 // return cipher length (int)
6423 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6424 set_result(retvalue);
6425 return true;
6426 }
6427
6428 //------------------------------get_key_start_from_aescrypt_object-----------------------
6429 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6430 #if defined(PPC64) || defined(S390)
6431 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6432 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6433 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6434 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6435 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6436 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6437 if (objSessionK == NULL) {
6438 return (Node *) NULL;
6439 }
6440 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6441 #else
6442 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6443 #endif // PPC64
6444 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6445 if (objAESCryptKey == NULL) return (Node *) NULL;
6446
6447 // now have the array, need to get the start address of the K array
6448 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6449 return k_start;
6450 }
6451
6452 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6453 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6454 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6455 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6456 if (objAESCryptKey == NULL) return (Node *) NULL;
6457
6458 // now have the array, need to get the start address of the lastKey array
6459 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6460 return original_k_start;
6461 }
6462
6463 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6464 // Return node representing slow path of predicate check.
6465 // the pseudo code we want to emulate with this predicate is:
6466 // for encryption:
6467 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6468 // for decryption:
6469 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6470 // note cipher==plain is more conservative than the original java code but that's OK
6471 //
6472 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6473 // The receiver was checked for NULL already.
6474 Node* objCBC = argument(0);
6475
6476 // Load embeddedCipher field of CipherBlockChaining object.
6477 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6478
6479 // get AESCrypt klass for instanceOf check
6480 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6481 // will have same classloader as CipherBlockChaining object
6482 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6483 assert(tinst != NULL, "CBCobj is null");
6484 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6485
6486 // we want to do an instanceof comparison against the AESCrypt class
6487 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6488 if (!klass_AESCrypt->is_loaded()) {
6489 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6490 Node* ctrl = control();
6491 set_control(top()); // no regular fast path
6492 return ctrl;
6493 }
6494 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6495
6496 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6497 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6498 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6499
6500 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6501
6502 // for encryption, we are done
6503 if (!decrypting)
6504 return instof_false; // even if it is NULL
6505
6506 // for decryption, we need to add a further check to avoid
6507 // taking the intrinsic path when cipher and plain are the same
6508 // see the original java code for why.
6509 RegionNode* region = new RegionNode(3);
6510 region->init_req(1, instof_false);
6511 Node* src = argument(1);
6512 Node* dest = argument(4);
6513 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6514 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6515 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6516 region->init_req(2, src_dest_conjoint);
6517
6518 record_for_igvn(region);
6519 return _gvn.transform(region);
6520 }
6521
6522 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6523 // Return node representing slow path of predicate check.
6524 // the pseudo code we want to emulate with this predicate is:
6525 // for encryption:
6526 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6527 // for decryption:
6528 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6529 // note cipher==plain is more conservative than the original java code but that's OK
6530 //
6531
6532 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6533 // The receiver was checked for NULL already.
6534 Node* objCTR = argument(0);
6535
6536 // Load embeddedCipher field of CipherBlockChaining object.
6537 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6538
6539 // get AESCrypt klass for instanceOf check
6540 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6541 // will have same classloader as CipherBlockChaining object
6542 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6543 assert(tinst != NULL, "CTRobj is null");
6544 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6545
6546 // we want to do an instanceof comparison against the AESCrypt class
6547 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6548 if (!klass_AESCrypt->is_loaded()) {
6549 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6550 Node* ctrl = control();
6551 set_control(top()); // no regular fast path
6552 return ctrl;
6553 }
6554
6555 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6556 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6557 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6558 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6559 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6560
6561 return instof_false; // even if it is NULL
6562 }
6563
6564 //------------------------------inline_ghash_processBlocks
6565 bool LibraryCallKit::inline_ghash_processBlocks() {
6566 address stubAddr;
6567 const char *stubName;
6568 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6569
6570 stubAddr = StubRoutines::ghash_processBlocks();
6571 stubName = "ghash_processBlocks";
6572
6573 Node* data = argument(0);
6574 Node* offset = argument(1);
6575 Node* len = argument(2);
6576 Node* state = argument(3);
6577 Node* subkeyH = argument(4);
6578
6579 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6580 assert(state_start, "state is NULL");
6581 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6582 assert(subkeyH_start, "subkeyH is NULL");
6583 Node* data_start = array_element_address(data, offset, T_BYTE);
6584 assert(data_start, "data is NULL");
6585
6586 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6587 OptoRuntime::ghash_processBlocks_Type(),
6588 stubAddr, stubName, TypePtr::BOTTOM,
6589 state_start, subkeyH_start, data_start, len);
6590 return true;
6591 }
6592
6593 //------------------------------inline_sha_implCompress-----------------------
6594 //
6595 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6596 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6597 //
6598 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6599 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6600 //
6601 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6602 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6603 //
6604 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6605 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6606
6607 Node* sha_obj = argument(0);
6608 Node* src = argument(1); // type oop
6609 Node* ofs = argument(2); // type int
6610
6611 const Type* src_type = src->Value(&_gvn);
6612 const TypeAryPtr* top_src = src_type->isa_aryptr();
6613 if (top_src == NULL || top_src->klass() == NULL) {
6614 // failed array check
6615 return false;
6616 }
6617 // Figure out the size and type of the elements we will be copying.
6618 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6619 if (src_elem != T_BYTE) {
6620 return false;
6621 }
6622 // 'src_start' points to src array + offset
6623 Node* src_start = array_element_address(src, ofs, src_elem);
6624 Node* state = NULL;
6625 address stubAddr;
6626 const char *stubName;
6627
6628 switch(id) {
6629 case vmIntrinsics::_sha_implCompress:
6630 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6631 state = get_state_from_sha_object(sha_obj);
6632 stubAddr = StubRoutines::sha1_implCompress();
6633 stubName = "sha1_implCompress";
6634 break;
6635 case vmIntrinsics::_sha2_implCompress:
6636 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6637 state = get_state_from_sha_object(sha_obj);
6638 stubAddr = StubRoutines::sha256_implCompress();
6639 stubName = "sha256_implCompress";
6640 break;
6641 case vmIntrinsics::_sha5_implCompress:
6642 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6643 state = get_state_from_sha5_object(sha_obj);
6644 stubAddr = StubRoutines::sha512_implCompress();
6645 stubName = "sha512_implCompress";
6646 break;
6647 default:
6648 fatal_unexpected_iid(id);
6649 return false;
6650 }
6651 if (state == NULL) return false;
6652
6653 // Call the stub.
6654 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6655 stubAddr, stubName, TypePtr::BOTTOM,
6656 src_start, state);
6657
6658 return true;
6659 }
6660
6661 //------------------------------inline_digestBase_implCompressMB-----------------------
6662 //
6663 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6664 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6665 //
6666 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6667 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6668 "need SHA1/SHA256/SHA512 instruction support");
6669 assert((uint)predicate < 3, "sanity");
6670 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6671
6672 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6673 Node* src = argument(1); // byte[] array
6674 Node* ofs = argument(2); // type int
6675 Node* limit = argument(3); // type int
6676
6677 const Type* src_type = src->Value(&_gvn);
6678 const TypeAryPtr* top_src = src_type->isa_aryptr();
6679 if (top_src == NULL || top_src->klass() == NULL) {
6680 // failed array check
6681 return false;
6682 }
6683 // Figure out the size and type of the elements we will be copying.
6684 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6685 if (src_elem != T_BYTE) {
6686 return false;
6687 }
6688 // 'src_start' points to src array + offset
6689 Node* src_start = array_element_address(src, ofs, src_elem);
6690
6691 const char* klass_SHA_name = NULL;
6692 const char* stub_name = NULL;
6693 address stub_addr = NULL;
6694 bool long_state = false;
6695
6696 switch (predicate) {
6697 case 0:
6698 if (UseSHA1Intrinsics) {
6699 klass_SHA_name = "sun/security/provider/SHA";
6700 stub_name = "sha1_implCompressMB";
6701 stub_addr = StubRoutines::sha1_implCompressMB();
6702 }
6703 break;
6704 case 1:
6705 if (UseSHA256Intrinsics) {
6706 klass_SHA_name = "sun/security/provider/SHA2";
6707 stub_name = "sha256_implCompressMB";
6708 stub_addr = StubRoutines::sha256_implCompressMB();
6709 }
6710 break;
6711 case 2:
6712 if (UseSHA512Intrinsics) {
6713 klass_SHA_name = "sun/security/provider/SHA5";
6714 stub_name = "sha512_implCompressMB";
6715 stub_addr = StubRoutines::sha512_implCompressMB();
6716 long_state = true;
6717 }
6718 break;
6719 default:
6720 fatal("unknown SHA intrinsic predicate: %d", predicate);
6721 }
6722 if (klass_SHA_name != NULL) {
6723 // get DigestBase klass to lookup for SHA klass
6724 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6725 assert(tinst != NULL, "digestBase_obj is not instance???");
6726 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6727
6728 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6729 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6730 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6731 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6732 }
6733 return false;
6734 }
6735 //------------------------------inline_sha_implCompressMB-----------------------
6736 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6737 bool long_state, address stubAddr, const char *stubName,
6738 Node* src_start, Node* ofs, Node* limit) {
6739 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6740 const TypeOopPtr* xtype = aklass->as_instance_type();
6741 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6742 sha_obj = _gvn.transform(sha_obj);
6743
6744 Node* state;
6745 if (long_state) {
6746 state = get_state_from_sha5_object(sha_obj);
6747 } else {
6748 state = get_state_from_sha_object(sha_obj);
6749 }
6750 if (state == NULL) return false;
6751
6752 // Call the stub.
6753 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6754 OptoRuntime::digestBase_implCompressMB_Type(),
6755 stubAddr, stubName, TypePtr::BOTTOM,
6756 src_start, state, ofs, limit);
6757 // return ofs (int)
6758 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6759 set_result(result);
6760
6761 return true;
6762 }
6763
6764 //------------------------------get_state_from_sha_object-----------------------
6765 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6766 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6767 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6768 if (sha_state == NULL) return (Node *) NULL;
6769
6770 // now have the array, need to get the start address of the state array
6771 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6772 return state;
6773 }
6774
6775 //------------------------------get_state_from_sha5_object-----------------------
6776 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6777 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6778 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6779 if (sha_state == NULL) return (Node *) NULL;
6780
6781 // now have the array, need to get the start address of the state array
6782 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6783 return state;
6784 }
6785
6786 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6787 // Return node representing slow path of predicate check.
6788 // the pseudo code we want to emulate with this predicate is:
6789 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6790 //
6791 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6792 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6793 "need SHA1/SHA256/SHA512 instruction support");
6794 assert((uint)predicate < 3, "sanity");
6795
6796 // The receiver was checked for NULL already.
6797 Node* digestBaseObj = argument(0);
6798
6799 // get DigestBase klass for instanceOf check
6800 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6801 assert(tinst != NULL, "digestBaseObj is null");
6802 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6803
6804 const char* klass_SHA_name = NULL;
6805 switch (predicate) {
6806 case 0:
6807 if (UseSHA1Intrinsics) {
6808 // we want to do an instanceof comparison against the SHA class
6809 klass_SHA_name = "sun/security/provider/SHA";
6810 }
6811 break;
6812 case 1:
6813 if (UseSHA256Intrinsics) {
6814 // we want to do an instanceof comparison against the SHA2 class
6815 klass_SHA_name = "sun/security/provider/SHA2";
6816 }
6817 break;
6818 case 2:
6819 if (UseSHA512Intrinsics) {
6820 // we want to do an instanceof comparison against the SHA5 class
6821 klass_SHA_name = "sun/security/provider/SHA5";
6822 }
6823 break;
6824 default:
6825 fatal("unknown SHA intrinsic predicate: %d", predicate);
6826 }
6827
6828 ciKlass* klass_SHA = NULL;
6829 if (klass_SHA_name != NULL) {
6830 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6831 }
6832 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6833 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6834 Node* ctrl = control();
6835 set_control(top()); // no intrinsic path
6836 return ctrl;
6837 }
6838 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6839
6840 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6841 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6842 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6843 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6844
6845 return instof_false; // even if it is NULL
6846 }
6847
6848 //-------------inline_fma-----------------------------------
6849 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6850 Node *a = NULL;
6851 Node *b = NULL;
6852 Node *c = NULL;
6853 Node* result = NULL;
6854 switch (id) {
6855 case vmIntrinsics::_fmaD:
6856 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6857 // no receiver since it is static method
6858 a = round_double_node(argument(0));
6859 b = round_double_node(argument(2));
6860 c = round_double_node(argument(4));
6861 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6862 break;
6863 case vmIntrinsics::_fmaF:
6864 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6865 a = argument(0);
6866 b = argument(1);
6867 c = argument(2);
6868 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6869 break;
6870 default:
6871 fatal_unexpected_iid(id); break;
6872 }
6873 set_result(result);
6874 return true;
6875 }
6876
6877 bool LibraryCallKit::inline_profileBoolean() {
6878 Node* counts = argument(1);
6879 const TypeAryPtr* ary = NULL;
6880 ciArray* aobj = NULL;
6881 if (counts->is_Con()
6882 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6883 && (aobj = ary->const_oop()->as_array()) != NULL
6884 && (aobj->length() == 2)) {
6885 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6886 jint false_cnt = aobj->element_value(0).as_int();
6887 jint true_cnt = aobj->element_value(1).as_int();
6888
6889 if (C->log() != NULL) {
6890 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6891 false_cnt, true_cnt);
6892 }
6893
6894 if (false_cnt + true_cnt == 0) {
6895 // According to profile, never executed.
6896 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6897 Deoptimization::Action_reinterpret);
6898 return true;
6899 }
6900
6901 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6902 // is a number of each value occurrences.
6903 Node* result = argument(0);
6904 if (false_cnt == 0 || true_cnt == 0) {
6905 // According to profile, one value has been never seen.
6906 int expected_val = (false_cnt == 0) ? 1 : 0;
6907
6908 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6909 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6910
6911 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6912 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6913 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6914
6915 { // Slow path: uncommon trap for never seen value and then reexecute
6916 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6917 // the value has been seen at least once.
6918 PreserveJVMState pjvms(this);
6919 PreserveReexecuteState preexecs(this);
6920 jvms()->set_should_reexecute(true);
6921
6922 set_control(slow_path);
6923 set_i_o(i_o());
6924
6925 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6926 Deoptimization::Action_reinterpret);
6927 }
6928 // The guard for never seen value enables sharpening of the result and
6929 // returning a constant. It allows to eliminate branches on the same value
6930 // later on.
6931 set_control(fast_path);
6932 result = intcon(expected_val);
6933 }
6934 // Stop profiling.
6935 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6936 // By replacing method body with profile data (represented as ProfileBooleanNode
6937 // on IR level) we effectively disable profiling.
6938 // It enables full speed execution once optimized code is generated.
6939 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6940 C->record_for_igvn(profile);
6941 set_result(profile);
6942 return true;
6943 } else {
6944 // Continue profiling.
6945 // Profile data isn't available at the moment. So, execute method's bytecode version.
6946 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6947 // is compiled and counters aren't available since corresponding MethodHandle
6948 // isn't a compile-time constant.
6949 return false;
6950 }
6951 }
6952
6953 bool LibraryCallKit::inline_isCompileConstant() {
6954 Node* n = argument(0);
6955 set_result(n->is_Con() ? intcon(1) : intcon(0));
6956 return true;
6957 }