1 /*
2 * Copyright (c) 1999, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "prims/nativeLookup.hpp"
52 #include "prims/unsafe.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #ifdef TRACE_HAVE_INTRINSICS
55 #include "trace/traceMacros.hpp"
56 #endif
57
58 class LibraryIntrinsic : public InlineCallGenerator {
59 // Extend the set of intrinsics known to the runtime:
60 public:
61 private:
62 bool _is_virtual;
63 bool _does_virtual_dispatch;
64 int8_t _predicates_count; // Intrinsic is predicated by several conditions
65 int8_t _last_predicate; // Last generated predicate
66 vmIntrinsics::ID _intrinsic_id;
67
68 public:
69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
70 : InlineCallGenerator(m),
71 _is_virtual(is_virtual),
72 _does_virtual_dispatch(does_virtual_dispatch),
73 _predicates_count((int8_t)predicates_count),
74 _last_predicate((int8_t)-1),
75 _intrinsic_id(id)
76 {
77 }
78 virtual bool is_intrinsic() const { return true; }
79 virtual bool is_virtual() const { return _is_virtual; }
80 virtual bool is_predicated() const { return _predicates_count > 0; }
81 virtual int predicates_count() const { return _predicates_count; }
82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
83 virtual JVMState* generate(JVMState* jvms);
84 virtual Node* generate_predicate(JVMState* jvms, int predicate);
85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
86 };
87
88
89 // Local helper class for LibraryIntrinsic:
90 class LibraryCallKit : public GraphKit {
91 private:
92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
93 Node* _result; // the result node, if any
94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
95
96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
97
98 public:
99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
100 : GraphKit(jvms),
101 _intrinsic(intrinsic),
102 _result(NULL)
103 {
104 // Check if this is a root compile. In that case we don't have a caller.
105 if (!jvms->has_method()) {
106 _reexecute_sp = sp();
107 } else {
108 // Find out how many arguments the interpreter needs when deoptimizing
109 // and save the stack pointer value so it can used by uncommon_trap.
110 // We find the argument count by looking at the declared signature.
111 bool ignored_will_link;
112 ciSignature* declared_signature = NULL;
113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
116 }
117 }
118
119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
120
121 ciMethod* caller() const { return jvms()->method(); }
122 int bci() const { return jvms()->bci(); }
123 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
125 ciMethod* callee() const { return _intrinsic->method(); }
126
127 bool try_to_inline(int predicate);
128 Node* try_to_predicate(int predicate);
129
130 void push_result() {
131 // Push the result onto the stack.
132 if (!stopped() && result() != NULL) {
133 BasicType bt = result()->bottom_type()->basic_type();
134 push_node(bt, result());
135 }
136 }
137
138 private:
139 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
141 }
142
143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
144 void set_result(RegionNode* region, PhiNode* value);
145 Node* result() { return _result; }
146
147 virtual int reexecute_sp() { return _reexecute_sp; }
148
149 // Helper functions to inline natives
150 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
151 Node* generate_slow_guard(Node* test, RegionNode* region);
152 Node* generate_fair_guard(Node* test, RegionNode* region);
153 Node* generate_negative_guard(Node* index, RegionNode* region,
154 // resulting CastII of index:
155 Node* *pos_index = NULL);
156 Node* generate_limit_guard(Node* offset, Node* subseq_length,
157 Node* array_length,
158 RegionNode* region);
159 void generate_string_range_check(Node* array, Node* offset,
160 Node* length, bool char_count);
161 Node* generate_current_thread(Node* &tls_output);
162 Node* load_mirror_from_klass(Node* klass);
163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path,
165 int offset);
166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
174 RegionNode* region, int null_path) {
175 int offset = java_lang_Class::array_klass_offset_in_bytes();
176 return load_klass_from_mirror_common(mirror, never_see_null,
177 region, null_path,
178 offset);
179 }
180 Node* generate_access_flags_guard(Node* kls,
181 int modifier_mask, int modifier_bits,
182 RegionNode* region);
183 Node* generate_interface_guard(Node* kls, RegionNode* region);
184 Node* generate_array_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, false, false);
186 }
187 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, false, true);
189 }
190 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, true, false);
192 }
193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, true, true);
195 }
196 Node* generate_array_guard_common(Node* kls, RegionNode* region,
197 bool obj_array, bool not_array);
198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
200 bool is_virtual = false, bool is_static = false);
201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
202 return generate_method_call(method_id, false, true);
203 }
204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
205 return generate_method_call(method_id, true, false);
206 }
207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209
210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_indexOfChar();
217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_toBytesU();
219 bool inline_string_getCharsU();
220 bool inline_string_copy(bool compress);
221 bool inline_string_char_access(bool is_store);
222 Node* round_double_node(Node* n);
223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_math_native(vmIntrinsics::ID id);
225 bool inline_math(vmIntrinsics::ID id);
226 template <typename OverflowOp>
227 bool inline_math_overflow(Node* arg1, Node* arg2);
228 void inline_math_mathExact(Node* math, Node* test);
229 bool inline_math_addExactI(bool is_increment);
230 bool inline_math_addExactL(bool is_increment);
231 bool inline_math_multiplyExactI();
232 bool inline_math_multiplyExactL();
233 bool inline_math_negateExactI();
234 bool inline_math_negateExactL();
235 bool inline_math_subtractExactI(bool is_decrement);
236 bool inline_math_subtractExactL(bool is_decrement);
237 bool inline_min_max(vmIntrinsics::ID id);
238 bool inline_notify(vmIntrinsics::ID id);
239 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
240 // This returns Type::AnyPtr, RawPtr, or OopPtr.
241 int classify_unsafe_addr(Node* &base, Node* &offset);
242 Node* make_unsafe_address(Node* base, Node* offset);
243 // Helper for inline_unsafe_access.
244 // Generates the guards that check whether the result of
245 // Unsafe.getObject should be recorded in an SATB log buffer.
246 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
247
248 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
249 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
250 static bool klass_needs_init_guard(Node* kls);
251 bool inline_unsafe_allocate();
252 bool inline_unsafe_newArray(bool uninitialized);
253 bool inline_unsafe_copyMemory();
254 bool inline_native_currentThread();
255
256 bool inline_native_time_funcs(address method, const char* funcName);
257 #ifdef TRACE_HAVE_INTRINSICS
258 bool inline_native_classID();
259 bool inline_native_getBufferWriter();
260 #endif
261 bool inline_native_isInterrupted();
262 bool inline_native_Class_query(vmIntrinsics::ID id);
263 bool inline_native_subtype_check();
264 bool inline_native_getLength();
265 bool inline_array_copyOf(bool is_copyOfRange);
266 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
267 bool inline_preconditions_checkIndex();
268 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
269 bool inline_native_clone(bool is_virtual);
270 bool inline_native_Reflection_getCallerClass();
271 // Helper function for inlining native object hash method
272 bool inline_native_hashcode(bool is_virtual, bool is_static);
273 bool inline_native_getClass();
274
275 // Helper functions for inlining arraycopy
276 bool inline_arraycopy();
277 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
278 RegionNode* slow_region);
279 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
280 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
281
282 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
283 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
284 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
285 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
286 bool inline_unsafe_fence(vmIntrinsics::ID id);
287 bool inline_onspinwait();
288 bool inline_fp_conversions(vmIntrinsics::ID id);
289 bool inline_number_methods(vmIntrinsics::ID id);
290 bool inline_reference_get();
291 bool inline_Class_cast();
292 bool inline_aescrypt_Block(vmIntrinsics::ID id);
293 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
294 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
295 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
296 Node* inline_counterMode_AESCrypt_predicate();
297 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
298 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
299 bool inline_ghash_processBlocks();
300 bool inline_sha_implCompress(vmIntrinsics::ID id);
301 bool inline_digestBase_implCompressMB(int predicate);
302 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
303 bool long_state, address stubAddr, const char *stubName,
304 Node* src_start, Node* ofs, Node* limit);
305 Node* get_state_from_sha_object(Node *sha_object);
306 Node* get_state_from_sha5_object(Node *sha_object);
307 Node* inline_digestBase_implCompressMB_predicate(int predicate);
308 bool inline_encodeISOArray();
309 bool inline_updateCRC32();
310 bool inline_updateBytesCRC32();
311 bool inline_updateByteBufferCRC32();
312 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
313 bool inline_updateBytesCRC32C();
314 bool inline_updateDirectByteBufferCRC32C();
315 bool inline_updateBytesAdler32();
316 bool inline_updateByteBufferAdler32();
317 bool inline_multiplyToLen();
318 bool inline_hasNegatives();
319 bool inline_squareToLen();
320 bool inline_mulAdd();
321 bool inline_montgomeryMultiply();
322 bool inline_montgomerySquare();
323 bool inline_vectorizedMismatch();
324 bool inline_fma(vmIntrinsics::ID id);
325
326 bool inline_profileBoolean();
327 bool inline_isCompileConstant();
328 };
329
330 //---------------------------make_vm_intrinsic----------------------------
331 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
332 vmIntrinsics::ID id = m->intrinsic_id();
333 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
334
335 if (!m->is_loaded()) {
336 // Do not attempt to inline unloaded methods.
337 return NULL;
338 }
339
340 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
341 bool is_available = false;
342
343 {
344 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
345 // the compiler must transition to '_thread_in_vm' state because both
346 // methods access VM-internal data.
347 VM_ENTRY_MARK;
348 methodHandle mh(THREAD, m->get_Method());
349 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
350 !C->directive()->is_intrinsic_disabled(mh) &&
351 !vmIntrinsics::is_disabled_by_flags(mh);
352
353 }
354
355 if (is_available) {
356 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
357 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
358 return new LibraryIntrinsic(m, is_virtual,
359 vmIntrinsics::predicates_needed(id),
360 vmIntrinsics::does_virtual_dispatch(id),
361 (vmIntrinsics::ID) id);
362 } else {
363 return NULL;
364 }
365 }
366
367 //----------------------register_library_intrinsics-----------------------
368 // Initialize this file's data structures, for each Compile instance.
369 void Compile::register_library_intrinsics() {
370 // Nothing to do here.
371 }
372
373 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
374 LibraryCallKit kit(jvms, this);
375 Compile* C = kit.C;
376 int nodes = C->unique();
377 #ifndef PRODUCT
378 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
379 char buf[1000];
380 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
381 tty->print_cr("Intrinsic %s", str);
382 }
383 #endif
384 ciMethod* callee = kit.callee();
385 const int bci = kit.bci();
386
387 // Try to inline the intrinsic.
388 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
389 kit.try_to_inline(_last_predicate)) {
390 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
391 : "(intrinsic)";
392 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
393 if (C->print_intrinsics() || C->print_inlining()) {
394 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
395 }
396 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
397 if (C->log()) {
398 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
399 vmIntrinsics::name_at(intrinsic_id()),
400 (is_virtual() ? " virtual='1'" : ""),
401 C->unique() - nodes);
402 }
403 // Push the result from the inlined method onto the stack.
404 kit.push_result();
405 C->print_inlining_update(this);
406 return kit.transfer_exceptions_into_jvms();
407 }
408
409 // The intrinsic bailed out
410 if (jvms->has_method()) {
411 // Not a root compile.
412 const char* msg;
413 if (callee->intrinsic_candidate()) {
414 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
415 } else {
416 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
417 : "failed to inline (intrinsic), method not annotated";
418 }
419 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
420 if (C->print_intrinsics() || C->print_inlining()) {
421 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
422 }
423 } else {
424 // Root compile
425 ResourceMark rm;
426 stringStream msg_stream;
427 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
428 vmIntrinsics::name_at(intrinsic_id()),
429 is_virtual() ? " (virtual)" : "", bci);
430 const char *msg = msg_stream.as_string();
431 log_debug(jit, inlining)("%s", msg);
432 if (C->print_intrinsics() || C->print_inlining()) {
433 tty->print("%s", msg);
434 }
435 }
436 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
437 C->print_inlining_update(this);
438 return NULL;
439 }
440
441 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
442 LibraryCallKit kit(jvms, this);
443 Compile* C = kit.C;
444 int nodes = C->unique();
445 _last_predicate = predicate;
446 #ifndef PRODUCT
447 assert(is_predicated() && predicate < predicates_count(), "sanity");
448 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
449 char buf[1000];
450 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
451 tty->print_cr("Predicate for intrinsic %s", str);
452 }
453 #endif
454 ciMethod* callee = kit.callee();
455 const int bci = kit.bci();
456
457 Node* slow_ctl = kit.try_to_predicate(predicate);
458 if (!kit.failing()) {
459 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
460 : "(intrinsic, predicate)";
461 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
462 if (C->print_intrinsics() || C->print_inlining()) {
463 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
464 }
465 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
466 if (C->log()) {
467 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
468 vmIntrinsics::name_at(intrinsic_id()),
469 (is_virtual() ? " virtual='1'" : ""),
470 C->unique() - nodes);
471 }
472 return slow_ctl; // Could be NULL if the check folds.
473 }
474
475 // The intrinsic bailed out
476 if (jvms->has_method()) {
477 // Not a root compile.
478 const char* msg = "failed to generate predicate for intrinsic";
479 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
480 if (C->print_intrinsics() || C->print_inlining()) {
481 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
482 }
483 } else {
484 // Root compile
485 ResourceMark rm;
486 stringStream msg_stream;
487 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
488 vmIntrinsics::name_at(intrinsic_id()),
489 is_virtual() ? " (virtual)" : "", bci);
490 const char *msg = msg_stream.as_string();
491 log_debug(jit, inlining)("%s", msg);
492 if (C->print_intrinsics() || C->print_inlining()) {
493 C->print_inlining_stream()->print("%s", msg);
494 }
495 }
496 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
497 return NULL;
498 }
499
500 bool LibraryCallKit::try_to_inline(int predicate) {
501 // Handle symbolic names for otherwise undistinguished boolean switches:
502 const bool is_store = true;
503 const bool is_compress = true;
504 const bool is_static = true;
505 const bool is_volatile = true;
506
507 if (!jvms()->has_method()) {
508 // Root JVMState has a null method.
509 assert(map()->memory()->Opcode() == Op_Parm, "");
510 // Insert the memory aliasing node
511 set_all_memory(reset_memory());
512 }
513 assert(merged_memory(), "");
514
515
516 switch (intrinsic_id()) {
517 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
518 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
519 case vmIntrinsics::_getClass: return inline_native_getClass();
520
521 case vmIntrinsics::_dsin:
522 case vmIntrinsics::_dcos:
523 case vmIntrinsics::_dtan:
524 case vmIntrinsics::_dabs:
525 case vmIntrinsics::_datan2:
526 case vmIntrinsics::_dsqrt:
527 case vmIntrinsics::_dexp:
528 case vmIntrinsics::_dlog:
529 case vmIntrinsics::_dlog10:
530 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
531
532 case vmIntrinsics::_min:
533 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
534
535 case vmIntrinsics::_notify:
536 case vmIntrinsics::_notifyAll:
537 if (InlineNotify) {
538 return inline_notify(intrinsic_id());
539 }
540 return false;
541
542 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
543 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
544 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
545 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
546 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
547 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
548 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
549 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
550 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
551 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
552 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
553 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
554
555 case vmIntrinsics::_arraycopy: return inline_arraycopy();
556
557 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
558 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
559 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
560 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
561
562 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
563 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
564 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
565 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
566 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
567 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
568 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
569
570 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
571 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
572
573 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
574 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
575 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
576 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
577
578 case vmIntrinsics::_compressStringC:
579 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
580 case vmIntrinsics::_inflateStringC:
581 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
582
583 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
584 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
585 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
586 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
587 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
588 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
589 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
590 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
591 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
592
593 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
594 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
595 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
596 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
597 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
598 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
599 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
600 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
601 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
602
603 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
604 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
605 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
606 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
607 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
608 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
609 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
610 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
611 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
612
613 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
614 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
615 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
616 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
617 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
618 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
619 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
620 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
621 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
622
623 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
624 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
625 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
626 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
627
628 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
629 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
630 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
631 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
632
633 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
634 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
635 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
636 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
637 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
638 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
639 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
640 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
641 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
642
643 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
644 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
645 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
646 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
647 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
648 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
649 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
650 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
651 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
652
653 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
654 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
655 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
656 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
657 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
658 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
659 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
660 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
661 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
662
663 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
664 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
665 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
666 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
667 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
668 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
669 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
670 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
671 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
672
673 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
674 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
675 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
676 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
677 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
678
679 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
680 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
681 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
682 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
683 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
684 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
685 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
686 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
687 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
688 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
689 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
690 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
691 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
692 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
693 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
694 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
695 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
696 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
697 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
698 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
699
700 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
701 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
702 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
703 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
704 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
705 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
706 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
707 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
708 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
709 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
710 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
711 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
712 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
713 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
714 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
715
716 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
717 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
718 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
719 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
720
721 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
722 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
723 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
724 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
725 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
726
727 case vmIntrinsics::_loadFence:
728 case vmIntrinsics::_storeFence:
729 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
730
731 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
732
733 case vmIntrinsics::_currentThread: return inline_native_currentThread();
734 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
735
736 #ifdef TRACE_HAVE_INTRINSICS
737 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
738 case vmIntrinsics::_getClassId: return inline_native_classID();
739 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
740 #endif
741 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
742 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
743 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
744 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
745 case vmIntrinsics::_getLength: return inline_native_getLength();
746 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
747 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
748 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
749 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
750 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
751 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
752
753 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
754 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
755
756 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
757
758 case vmIntrinsics::_isInstance:
759 case vmIntrinsics::_getModifiers:
760 case vmIntrinsics::_isInterface:
761 case vmIntrinsics::_isArray:
762 case vmIntrinsics::_isPrimitive:
763 case vmIntrinsics::_getSuperclass:
764 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
765
766 case vmIntrinsics::_floatToRawIntBits:
767 case vmIntrinsics::_floatToIntBits:
768 case vmIntrinsics::_intBitsToFloat:
769 case vmIntrinsics::_doubleToRawLongBits:
770 case vmIntrinsics::_doubleToLongBits:
771 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
772
773 case vmIntrinsics::_numberOfLeadingZeros_i:
774 case vmIntrinsics::_numberOfLeadingZeros_l:
775 case vmIntrinsics::_numberOfTrailingZeros_i:
776 case vmIntrinsics::_numberOfTrailingZeros_l:
777 case vmIntrinsics::_bitCount_i:
778 case vmIntrinsics::_bitCount_l:
779 case vmIntrinsics::_reverseBytes_i:
780 case vmIntrinsics::_reverseBytes_l:
781 case vmIntrinsics::_reverseBytes_s:
782 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
783
784 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
785
786 case vmIntrinsics::_Reference_get: return inline_reference_get();
787
788 case vmIntrinsics::_Class_cast: return inline_Class_cast();
789
790 case vmIntrinsics::_aescrypt_encryptBlock:
791 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
792
793 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
794 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
795 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
796
797 case vmIntrinsics::_counterMode_AESCrypt:
798 return inline_counterMode_AESCrypt(intrinsic_id());
799
800 case vmIntrinsics::_sha_implCompress:
801 case vmIntrinsics::_sha2_implCompress:
802 case vmIntrinsics::_sha5_implCompress:
803 return inline_sha_implCompress(intrinsic_id());
804
805 case vmIntrinsics::_digestBase_implCompressMB:
806 return inline_digestBase_implCompressMB(predicate);
807
808 case vmIntrinsics::_multiplyToLen:
809 return inline_multiplyToLen();
810
811 case vmIntrinsics::_squareToLen:
812 return inline_squareToLen();
813
814 case vmIntrinsics::_mulAdd:
815 return inline_mulAdd();
816
817 case vmIntrinsics::_montgomeryMultiply:
818 return inline_montgomeryMultiply();
819 case vmIntrinsics::_montgomerySquare:
820 return inline_montgomerySquare();
821
822 case vmIntrinsics::_vectorizedMismatch:
823 return inline_vectorizedMismatch();
824
825 case vmIntrinsics::_ghash_processBlocks:
826 return inline_ghash_processBlocks();
827
828 case vmIntrinsics::_encodeISOArray:
829 case vmIntrinsics::_encodeByteISOArray:
830 return inline_encodeISOArray();
831
832 case vmIntrinsics::_updateCRC32:
833 return inline_updateCRC32();
834 case vmIntrinsics::_updateBytesCRC32:
835 return inline_updateBytesCRC32();
836 case vmIntrinsics::_updateByteBufferCRC32:
837 return inline_updateByteBufferCRC32();
838
839 case vmIntrinsics::_updateBytesCRC32C:
840 return inline_updateBytesCRC32C();
841 case vmIntrinsics::_updateDirectByteBufferCRC32C:
842 return inline_updateDirectByteBufferCRC32C();
843
844 case vmIntrinsics::_updateBytesAdler32:
845 return inline_updateBytesAdler32();
846 case vmIntrinsics::_updateByteBufferAdler32:
847 return inline_updateByteBufferAdler32();
848
849 case vmIntrinsics::_profileBoolean:
850 return inline_profileBoolean();
851 case vmIntrinsics::_isCompileConstant:
852 return inline_isCompileConstant();
853
854 case vmIntrinsics::_hasNegatives:
855 return inline_hasNegatives();
856
857 case vmIntrinsics::_fmaD:
858 case vmIntrinsics::_fmaF:
859 return inline_fma(intrinsic_id());
860
861 default:
862 // If you get here, it may be that someone has added a new intrinsic
863 // to the list in vmSymbols.hpp without implementing it here.
864 #ifndef PRODUCT
865 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
866 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
867 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
868 }
869 #endif
870 return false;
871 }
872 }
873
874 Node* LibraryCallKit::try_to_predicate(int predicate) {
875 if (!jvms()->has_method()) {
876 // Root JVMState has a null method.
877 assert(map()->memory()->Opcode() == Op_Parm, "");
878 // Insert the memory aliasing node
879 set_all_memory(reset_memory());
880 }
881 assert(merged_memory(), "");
882
883 switch (intrinsic_id()) {
884 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
885 return inline_cipherBlockChaining_AESCrypt_predicate(false);
886 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
887 return inline_cipherBlockChaining_AESCrypt_predicate(true);
888 case vmIntrinsics::_counterMode_AESCrypt:
889 return inline_counterMode_AESCrypt_predicate();
890 case vmIntrinsics::_digestBase_implCompressMB:
891 return inline_digestBase_implCompressMB_predicate(predicate);
892
893 default:
894 // If you get here, it may be that someone has added a new intrinsic
895 // to the list in vmSymbols.hpp without implementing it here.
896 #ifndef PRODUCT
897 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
898 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
899 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
900 }
901 #endif
902 Node* slow_ctl = control();
903 set_control(top()); // No fast path instrinsic
904 return slow_ctl;
905 }
906 }
907
908 //------------------------------set_result-------------------------------
909 // Helper function for finishing intrinsics.
910 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
911 record_for_igvn(region);
912 set_control(_gvn.transform(region));
913 set_result( _gvn.transform(value));
914 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
915 }
916
917 //------------------------------generate_guard---------------------------
918 // Helper function for generating guarded fast-slow graph structures.
919 // The given 'test', if true, guards a slow path. If the test fails
920 // then a fast path can be taken. (We generally hope it fails.)
921 // In all cases, GraphKit::control() is updated to the fast path.
922 // The returned value represents the control for the slow path.
923 // The return value is never 'top'; it is either a valid control
924 // or NULL if it is obvious that the slow path can never be taken.
925 // Also, if region and the slow control are not NULL, the slow edge
926 // is appended to the region.
927 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
928 if (stopped()) {
929 // Already short circuited.
930 return NULL;
931 }
932
933 // Build an if node and its projections.
934 // If test is true we take the slow path, which we assume is uncommon.
935 if (_gvn.type(test) == TypeInt::ZERO) {
936 // The slow branch is never taken. No need to build this guard.
937 return NULL;
938 }
939
940 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
941
942 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
943 if (if_slow == top()) {
944 // The slow branch is never taken. No need to build this guard.
945 return NULL;
946 }
947
948 if (region != NULL)
949 region->add_req(if_slow);
950
951 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
952 set_control(if_fast);
953
954 return if_slow;
955 }
956
957 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
958 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
959 }
960 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
961 return generate_guard(test, region, PROB_FAIR);
962 }
963
964 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
965 Node* *pos_index) {
966 if (stopped())
967 return NULL; // already stopped
968 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
969 return NULL; // index is already adequately typed
970 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
971 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
972 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
973 if (is_neg != NULL && pos_index != NULL) {
974 // Emulate effect of Parse::adjust_map_after_if.
975 Node* ccast = new CastIINode(index, TypeInt::POS);
976 ccast->set_req(0, control());
977 (*pos_index) = _gvn.transform(ccast);
978 }
979 return is_neg;
980 }
981
982 // Make sure that 'position' is a valid limit index, in [0..length].
983 // There are two equivalent plans for checking this:
984 // A. (offset + copyLength) unsigned<= arrayLength
985 // B. offset <= (arrayLength - copyLength)
986 // We require that all of the values above, except for the sum and
987 // difference, are already known to be non-negative.
988 // Plan A is robust in the face of overflow, if offset and copyLength
989 // are both hugely positive.
990 //
991 // Plan B is less direct and intuitive, but it does not overflow at
992 // all, since the difference of two non-negatives is always
993 // representable. Whenever Java methods must perform the equivalent
994 // check they generally use Plan B instead of Plan A.
995 // For the moment we use Plan A.
996 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
997 Node* subseq_length,
998 Node* array_length,
999 RegionNode* region) {
1000 if (stopped())
1001 return NULL; // already stopped
1002 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1003 if (zero_offset && subseq_length->eqv_uncast(array_length))
1004 return NULL; // common case of whole-array copy
1005 Node* last = subseq_length;
1006 if (!zero_offset) // last += offset
1007 last = _gvn.transform(new AddINode(last, offset));
1008 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1009 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1010 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1011 return is_over;
1012 }
1013
1014 // Emit range checks for the given String.value byte array
1015 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1016 if (stopped()) {
1017 return; // already stopped
1018 }
1019 RegionNode* bailout = new RegionNode(1);
1020 record_for_igvn(bailout);
1021 if (char_count) {
1022 // Convert char count to byte count
1023 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1024 }
1025
1026 // Offset and count must not be negative
1027 generate_negative_guard(offset, bailout);
1028 generate_negative_guard(count, bailout);
1029 // Offset + count must not exceed length of array
1030 generate_limit_guard(offset, count, load_array_length(array), bailout);
1031
1032 if (bailout->req() > 1) {
1033 PreserveJVMState pjvms(this);
1034 set_control(_gvn.transform(bailout));
1035 uncommon_trap(Deoptimization::Reason_intrinsic,
1036 Deoptimization::Action_maybe_recompile);
1037 }
1038 }
1039
1040 //--------------------------generate_current_thread--------------------
1041 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1042 ciKlass* thread_klass = env()->Thread_klass();
1043 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1044 Node* thread = _gvn.transform(new ThreadLocalNode());
1045 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1046 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1047 tls_output = thread;
1048 return threadObj;
1049 }
1050
1051
1052 //------------------------------make_string_method_node------------------------
1053 // Helper method for String intrinsic functions. This version is called with
1054 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1055 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1056 // containing the lengths of str1 and str2.
1057 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1058 Node* result = NULL;
1059 switch (opcode) {
1060 case Op_StrIndexOf:
1061 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1062 str1_start, cnt1, str2_start, cnt2, ae);
1063 break;
1064 case Op_StrComp:
1065 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1066 str1_start, cnt1, str2_start, cnt2, ae);
1067 break;
1068 case Op_StrEquals:
1069 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1070 // Use the constant length if there is one because optimized match rule may exist.
1071 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1072 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1073 break;
1074 default:
1075 ShouldNotReachHere();
1076 return NULL;
1077 }
1078
1079 // All these intrinsics have checks.
1080 C->set_has_split_ifs(true); // Has chance for split-if optimization
1081
1082 return _gvn.transform(result);
1083 }
1084
1085 //------------------------------inline_string_compareTo------------------------
1086 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1087 Node* arg1 = argument(0);
1088 Node* arg2 = argument(1);
1089
1090 // Get start addr and length of first argument
1091 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1092 Node* arg1_cnt = load_array_length(arg1);
1093
1094 // Get start addr and length of second argument
1095 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1096 Node* arg2_cnt = load_array_length(arg2);
1097
1098 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1099 set_result(result);
1100 return true;
1101 }
1102
1103 //------------------------------inline_string_equals------------------------
1104 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1105 Node* arg1 = argument(0);
1106 Node* arg2 = argument(1);
1107
1108 // paths (plus control) merge
1109 RegionNode* region = new RegionNode(3);
1110 Node* phi = new PhiNode(region, TypeInt::BOOL);
1111
1112 if (!stopped()) {
1113 // Get start addr and length of first argument
1114 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1115 Node* arg1_cnt = load_array_length(arg1);
1116
1117 // Get start addr and length of second argument
1118 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1119 Node* arg2_cnt = load_array_length(arg2);
1120
1121 // Check for arg1_cnt != arg2_cnt
1122 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1123 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1124 Node* if_ne = generate_slow_guard(bol, NULL);
1125 if (if_ne != NULL) {
1126 phi->init_req(2, intcon(0));
1127 region->init_req(2, if_ne);
1128 }
1129
1130 // Check for count == 0 is done by assembler code for StrEquals.
1131
1132 if (!stopped()) {
1133 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1134 phi->init_req(1, equals);
1135 region->init_req(1, control());
1136 }
1137 }
1138
1139 // post merge
1140 set_control(_gvn.transform(region));
1141 record_for_igvn(region);
1142
1143 set_result(_gvn.transform(phi));
1144 return true;
1145 }
1146
1147 //------------------------------inline_array_equals----------------------------
1148 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1149 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1150 Node* arg1 = argument(0);
1151 Node* arg2 = argument(1);
1152
1153 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1154 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1155 return true;
1156 }
1157
1158 //------------------------------inline_hasNegatives------------------------------
1159 bool LibraryCallKit::inline_hasNegatives() {
1160 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1161 return false;
1162 }
1163
1164 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1165 // no receiver since it is static method
1166 Node* ba = argument(0);
1167 Node* offset = argument(1);
1168 Node* len = argument(2);
1169
1170 // Range checks
1171 generate_string_range_check(ba, offset, len, false);
1172 if (stopped()) {
1173 return true;
1174 }
1175 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1176 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1177 set_result(_gvn.transform(result));
1178 return true;
1179 }
1180
1181 bool LibraryCallKit::inline_preconditions_checkIndex() {
1182 Node* index = argument(0);
1183 Node* length = argument(1);
1184 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1185 return false;
1186 }
1187
1188 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1189 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1190
1191 {
1192 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1193 uncommon_trap(Deoptimization::Reason_intrinsic,
1194 Deoptimization::Action_make_not_entrant);
1195 }
1196
1197 if (stopped()) {
1198 return false;
1199 }
1200
1201 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1202 BoolTest::mask btest = BoolTest::lt;
1203 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1204 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1205 _gvn.set_type(rc, rc->Value(&_gvn));
1206 if (!rc_bool->is_Con()) {
1207 record_for_igvn(rc);
1208 }
1209 set_control(_gvn.transform(new IfTrueNode(rc)));
1210 {
1211 PreserveJVMState pjvms(this);
1212 set_control(_gvn.transform(new IfFalseNode(rc)));
1213 uncommon_trap(Deoptimization::Reason_range_check,
1214 Deoptimization::Action_make_not_entrant);
1215 }
1216
1217 if (stopped()) {
1218 return false;
1219 }
1220
1221 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1222 result->set_req(0, control());
1223 result = _gvn.transform(result);
1224 set_result(result);
1225 replace_in_map(index, result);
1226 return true;
1227 }
1228
1229 //------------------------------inline_string_indexOf------------------------
1230 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1231 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1232 return false;
1233 }
1234 Node* src = argument(0);
1235 Node* tgt = argument(1);
1236
1237 // Make the merge point
1238 RegionNode* result_rgn = new RegionNode(4);
1239 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1240
1241 // Get start addr and length of source string
1242 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1243 Node* src_count = load_array_length(src);
1244
1245 // Get start addr and length of substring
1246 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1247 Node* tgt_count = load_array_length(tgt);
1248
1249 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1250 // Divide src size by 2 if String is UTF16 encoded
1251 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1252 }
1253 if (ae == StrIntrinsicNode::UU) {
1254 // Divide substring size by 2 if String is UTF16 encoded
1255 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1256 }
1257
1258 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1259 if (result != NULL) {
1260 result_phi->init_req(3, result);
1261 result_rgn->init_req(3, control());
1262 }
1263 set_control(_gvn.transform(result_rgn));
1264 record_for_igvn(result_rgn);
1265 set_result(_gvn.transform(result_phi));
1266
1267 return true;
1268 }
1269
1270 //-----------------------------inline_string_indexOf-----------------------
1271 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1272 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1273 return false;
1274 }
1275 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1276 return false;
1277 }
1278 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1279 Node* src = argument(0); // byte[]
1280 Node* src_count = argument(1); // char count
1281 Node* tgt = argument(2); // byte[]
1282 Node* tgt_count = argument(3); // char count
1283 Node* from_index = argument(4); // char index
1284
1285 // Multiply byte array index by 2 if String is UTF16 encoded
1286 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1287 src_count = _gvn.transform(new SubINode(src_count, from_index));
1288 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1289 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1290
1291 // Range checks
1292 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1293 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1294 if (stopped()) {
1295 return true;
1296 }
1297
1298 RegionNode* region = new RegionNode(5);
1299 Node* phi = new PhiNode(region, TypeInt::INT);
1300
1301 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1302 if (result != NULL) {
1303 // The result is index relative to from_index if substring was found, -1 otherwise.
1304 // Generate code which will fold into cmove.
1305 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1306 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1307
1308 Node* if_lt = generate_slow_guard(bol, NULL);
1309 if (if_lt != NULL) {
1310 // result == -1
1311 phi->init_req(3, result);
1312 region->init_req(3, if_lt);
1313 }
1314 if (!stopped()) {
1315 result = _gvn.transform(new AddINode(result, from_index));
1316 phi->init_req(4, result);
1317 region->init_req(4, control());
1318 }
1319 }
1320
1321 set_control(_gvn.transform(region));
1322 record_for_igvn(region);
1323 set_result(_gvn.transform(phi));
1324
1325 return true;
1326 }
1327
1328 // Create StrIndexOfNode with fast path checks
1329 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1330 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1331 // Check for substr count > string count
1332 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1333 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1334 Node* if_gt = generate_slow_guard(bol, NULL);
1335 if (if_gt != NULL) {
1336 phi->init_req(1, intcon(-1));
1337 region->init_req(1, if_gt);
1338 }
1339 if (!stopped()) {
1340 // Check for substr count == 0
1341 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1342 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1343 Node* if_zero = generate_slow_guard(bol, NULL);
1344 if (if_zero != NULL) {
1345 phi->init_req(2, intcon(0));
1346 region->init_req(2, if_zero);
1347 }
1348 }
1349 if (!stopped()) {
1350 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1351 }
1352 return NULL;
1353 }
1354
1355 //-----------------------------inline_string_indexOfChar-----------------------
1356 bool LibraryCallKit::inline_string_indexOfChar() {
1357 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1358 return false;
1359 }
1360 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1361 return false;
1362 }
1363 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1364 Node* src = argument(0); // byte[]
1365 Node* tgt = argument(1); // tgt is int ch
1366 Node* from_index = argument(2);
1367 Node* max = argument(3);
1368
1369 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1370 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1371 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1372
1373 // Range checks
1374 generate_string_range_check(src, src_offset, src_count, true);
1375 if (stopped()) {
1376 return true;
1377 }
1378
1379 RegionNode* region = new RegionNode(3);
1380 Node* phi = new PhiNode(region, TypeInt::INT);
1381
1382 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1383 C->set_has_split_ifs(true); // Has chance for split-if optimization
1384 _gvn.transform(result);
1385
1386 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1387 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1388
1389 Node* if_lt = generate_slow_guard(bol, NULL);
1390 if (if_lt != NULL) {
1391 // result == -1
1392 phi->init_req(2, result);
1393 region->init_req(2, if_lt);
1394 }
1395 if (!stopped()) {
1396 result = _gvn.transform(new AddINode(result, from_index));
1397 phi->init_req(1, result);
1398 region->init_req(1, control());
1399 }
1400 set_control(_gvn.transform(region));
1401 record_for_igvn(region);
1402 set_result(_gvn.transform(phi));
1403
1404 return true;
1405 }
1406 //---------------------------inline_string_copy---------------------
1407 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1408 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1409 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1410 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1411 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1412 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1413 bool LibraryCallKit::inline_string_copy(bool compress) {
1414 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1415 return false;
1416 }
1417 int nargs = 5; // 2 oops, 3 ints
1418 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1419
1420 Node* src = argument(0);
1421 Node* src_offset = argument(1);
1422 Node* dst = argument(2);
1423 Node* dst_offset = argument(3);
1424 Node* length = argument(4);
1425
1426 // Check for allocation before we add nodes that would confuse
1427 // tightly_coupled_allocation()
1428 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1429
1430 // Figure out the size and type of the elements we will be copying.
1431 const Type* src_type = src->Value(&_gvn);
1432 const Type* dst_type = dst->Value(&_gvn);
1433 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1434 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1435 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1436 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1437 "Unsupported array types for inline_string_copy");
1438
1439 // Convert char[] offsets to byte[] offsets
1440 bool convert_src = (compress && src_elem == T_BYTE);
1441 bool convert_dst = (!compress && dst_elem == T_BYTE);
1442 if (convert_src) {
1443 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1444 } else if (convert_dst) {
1445 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1446 }
1447
1448 // Range checks
1449 generate_string_range_check(src, src_offset, length, convert_src);
1450 generate_string_range_check(dst, dst_offset, length, convert_dst);
1451 if (stopped()) {
1452 return true;
1453 }
1454
1455 Node* src_start = array_element_address(src, src_offset, src_elem);
1456 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1457 // 'src_start' points to src array + scaled offset
1458 // 'dst_start' points to dst array + scaled offset
1459 Node* count = NULL;
1460 if (compress) {
1461 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1462 } else {
1463 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1464 }
1465
1466 if (alloc != NULL) {
1467 if (alloc->maybe_set_complete(&_gvn)) {
1468 // "You break it, you buy it."
1469 InitializeNode* init = alloc->initialization();
1470 assert(init->is_complete(), "we just did this");
1471 init->set_complete_with_arraycopy();
1472 assert(dst->is_CheckCastPP(), "sanity");
1473 assert(dst->in(0)->in(0) == init, "dest pinned");
1474 }
1475 // Do not let stores that initialize this object be reordered with
1476 // a subsequent store that would make this object accessible by
1477 // other threads.
1478 // Record what AllocateNode this StoreStore protects so that
1479 // escape analysis can go from the MemBarStoreStoreNode to the
1480 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1481 // based on the escape status of the AllocateNode.
1482 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1483 }
1484 if (compress) {
1485 set_result(_gvn.transform(count));
1486 }
1487 return true;
1488 }
1489
1490 #ifdef _LP64
1491 #define XTOP ,top() /*additional argument*/
1492 #else //_LP64
1493 #define XTOP /*no additional argument*/
1494 #endif //_LP64
1495
1496 //------------------------inline_string_toBytesU--------------------------
1497 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1498 bool LibraryCallKit::inline_string_toBytesU() {
1499 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1500 return false;
1501 }
1502 // Get the arguments.
1503 Node* value = argument(0);
1504 Node* offset = argument(1);
1505 Node* length = argument(2);
1506
1507 Node* newcopy = NULL;
1508
1509 // Set the original stack and the reexecute bit for the interpreter to reexecute
1510 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1511 { PreserveReexecuteState preexecs(this);
1512 jvms()->set_should_reexecute(true);
1513
1514 // Check if a null path was taken unconditionally.
1515 value = null_check(value);
1516
1517 RegionNode* bailout = new RegionNode(1);
1518 record_for_igvn(bailout);
1519
1520 // Range checks
1521 generate_negative_guard(offset, bailout);
1522 generate_negative_guard(length, bailout);
1523 generate_limit_guard(offset, length, load_array_length(value), bailout);
1524 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1525 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1526
1527 if (bailout->req() > 1) {
1528 PreserveJVMState pjvms(this);
1529 set_control(_gvn.transform(bailout));
1530 uncommon_trap(Deoptimization::Reason_intrinsic,
1531 Deoptimization::Action_maybe_recompile);
1532 }
1533 if (stopped()) {
1534 return true;
1535 }
1536
1537 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1538 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1539 newcopy = new_array(klass_node, size, 0); // no arguments to push
1540 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1541
1542 // Calculate starting addresses.
1543 Node* src_start = array_element_address(value, offset, T_CHAR);
1544 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1545
1546 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1547 const TypeInt* toffset = gvn().type(offset)->is_int();
1548 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1549
1550 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1551 const char* copyfunc_name = "arraycopy";
1552 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1553 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1554 OptoRuntime::fast_arraycopy_Type(),
1555 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1556 src_start, dst_start, ConvI2X(length) XTOP);
1557 // Do not let reads from the cloned object float above the arraycopy.
1558 if (alloc != NULL) {
1559 if (alloc->maybe_set_complete(&_gvn)) {
1560 // "You break it, you buy it."
1561 InitializeNode* init = alloc->initialization();
1562 assert(init->is_complete(), "we just did this");
1563 init->set_complete_with_arraycopy();
1564 assert(newcopy->is_CheckCastPP(), "sanity");
1565 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1566 }
1567 // Do not let stores that initialize this object be reordered with
1568 // a subsequent store that would make this object accessible by
1569 // other threads.
1570 // Record what AllocateNode this StoreStore protects so that
1571 // escape analysis can go from the MemBarStoreStoreNode to the
1572 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1573 // based on the escape status of the AllocateNode.
1574 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1575 } else {
1576 insert_mem_bar(Op_MemBarCPUOrder);
1577 }
1578 } // original reexecute is set back here
1579
1580 C->set_has_split_ifs(true); // Has chance for split-if optimization
1581 if (!stopped()) {
1582 set_result(newcopy);
1583 }
1584 return true;
1585 }
1586
1587 //------------------------inline_string_getCharsU--------------------------
1588 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1589 bool LibraryCallKit::inline_string_getCharsU() {
1590 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1591 return false;
1592 }
1593
1594 // Get the arguments.
1595 Node* src = argument(0);
1596 Node* src_begin = argument(1);
1597 Node* src_end = argument(2); // exclusive offset (i < src_end)
1598 Node* dst = argument(3);
1599 Node* dst_begin = argument(4);
1600
1601 // Check for allocation before we add nodes that would confuse
1602 // tightly_coupled_allocation()
1603 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1604
1605 // Check if a null path was taken unconditionally.
1606 src = null_check(src);
1607 dst = null_check(dst);
1608 if (stopped()) {
1609 return true;
1610 }
1611
1612 // Get length and convert char[] offset to byte[] offset
1613 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1614 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1615
1616 // Range checks
1617 generate_string_range_check(src, src_begin, length, true);
1618 generate_string_range_check(dst, dst_begin, length, false);
1619 if (stopped()) {
1620 return true;
1621 }
1622
1623 if (!stopped()) {
1624 // Calculate starting addresses.
1625 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1626 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1627
1628 // Check if array addresses are aligned to HeapWordSize
1629 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1630 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1631 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1632 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1633
1634 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1635 const char* copyfunc_name = "arraycopy";
1636 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1637 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1638 OptoRuntime::fast_arraycopy_Type(),
1639 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1640 src_start, dst_start, ConvI2X(length) XTOP);
1641 // Do not let reads from the cloned object float above the arraycopy.
1642 if (alloc != NULL) {
1643 if (alloc->maybe_set_complete(&_gvn)) {
1644 // "You break it, you buy it."
1645 InitializeNode* init = alloc->initialization();
1646 assert(init->is_complete(), "we just did this");
1647 init->set_complete_with_arraycopy();
1648 assert(dst->is_CheckCastPP(), "sanity");
1649 assert(dst->in(0)->in(0) == init, "dest pinned");
1650 }
1651 // Do not let stores that initialize this object be reordered with
1652 // a subsequent store that would make this object accessible by
1653 // other threads.
1654 // Record what AllocateNode this StoreStore protects so that
1655 // escape analysis can go from the MemBarStoreStoreNode to the
1656 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1657 // based on the escape status of the AllocateNode.
1658 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1659 } else {
1660 insert_mem_bar(Op_MemBarCPUOrder);
1661 }
1662 }
1663
1664 C->set_has_split_ifs(true); // Has chance for split-if optimization
1665 return true;
1666 }
1667
1668 //----------------------inline_string_char_access----------------------------
1669 // Store/Load char to/from byte[] array.
1670 // static void StringUTF16.putChar(byte[] val, int index, int c)
1671 // static char StringUTF16.getChar(byte[] val, int index)
1672 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1673 Node* value = argument(0);
1674 Node* index = argument(1);
1675 Node* ch = is_store ? argument(2) : NULL;
1676
1677 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1678 // correctly requires matched array shapes.
1679 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1680 "sanity: byte[] and char[] bases agree");
1681 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1682 "sanity: byte[] and char[] scales agree");
1683
1684 // Bail when getChar over constants is requested: constant folding would
1685 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1686 // Java method would constant fold nicely instead.
1687 if (!is_store && value->is_Con() && index->is_Con()) {
1688 return false;
1689 }
1690
1691 Node* adr = array_element_address(value, index, T_CHAR);
1692 if (is_store) {
1693 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1694 false, false, true /* mismatched */);
1695 } else {
1696 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1697 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1698 set_result(ch);
1699 }
1700 return true;
1701 }
1702
1703 //--------------------------round_double_node--------------------------------
1704 // Round a double node if necessary.
1705 Node* LibraryCallKit::round_double_node(Node* n) {
1706 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1707 n = _gvn.transform(new RoundDoubleNode(0, n));
1708 return n;
1709 }
1710
1711 //------------------------------inline_math-----------------------------------
1712 // public static double Math.abs(double)
1713 // public static double Math.sqrt(double)
1714 // public static double Math.log(double)
1715 // public static double Math.log10(double)
1716 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1717 Node* arg = round_double_node(argument(0));
1718 Node* n = NULL;
1719 switch (id) {
1720 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1721 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1722 default: fatal_unexpected_iid(id); break;
1723 }
1724 set_result(_gvn.transform(n));
1725 return true;
1726 }
1727
1728 //------------------------------runtime_math-----------------------------
1729 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1730 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1731 "must be (DD)D or (D)D type");
1732
1733 // Inputs
1734 Node* a = round_double_node(argument(0));
1735 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1736
1737 const TypePtr* no_memory_effects = NULL;
1738 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1739 no_memory_effects,
1740 a, top(), b, b ? top() : NULL);
1741 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1742 #ifdef ASSERT
1743 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1744 assert(value_top == top(), "second value must be top");
1745 #endif
1746
1747 set_result(value);
1748 return true;
1749 }
1750
1751 //------------------------------inline_math_native-----------------------------
1752 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1753 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1754 switch (id) {
1755 // These intrinsics are not properly supported on all hardware
1756 case vmIntrinsics::_dsin:
1757 return StubRoutines::dsin() != NULL ?
1758 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1759 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1760 case vmIntrinsics::_dcos:
1761 return StubRoutines::dcos() != NULL ?
1762 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1763 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1764 case vmIntrinsics::_dtan:
1765 return StubRoutines::dtan() != NULL ?
1766 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1767 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1768 case vmIntrinsics::_dlog:
1769 return StubRoutines::dlog() != NULL ?
1770 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1771 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1772 case vmIntrinsics::_dlog10:
1773 return StubRoutines::dlog10() != NULL ?
1774 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1775 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1776
1777 // These intrinsics are supported on all hardware
1778 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1779 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1780
1781 case vmIntrinsics::_dexp:
1782 return StubRoutines::dexp() != NULL ?
1783 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1784 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1785 case vmIntrinsics::_dpow:
1786 return StubRoutines::dpow() != NULL ?
1787 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1788 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1789 #undef FN_PTR
1790
1791 // These intrinsics are not yet correctly implemented
1792 case vmIntrinsics::_datan2:
1793 return false;
1794
1795 default:
1796 fatal_unexpected_iid(id);
1797 return false;
1798 }
1799 }
1800
1801 static bool is_simple_name(Node* n) {
1802 return (n->req() == 1 // constant
1803 || (n->is_Type() && n->as_Type()->type()->singleton())
1804 || n->is_Proj() // parameter or return value
1805 || n->is_Phi() // local of some sort
1806 );
1807 }
1808
1809 //----------------------------inline_notify-----------------------------------*
1810 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1811 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1812 address func;
1813 if (id == vmIntrinsics::_notify) {
1814 func = OptoRuntime::monitor_notify_Java();
1815 } else {
1816 func = OptoRuntime::monitor_notifyAll_Java();
1817 }
1818 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1819 make_slow_call_ex(call, env()->Throwable_klass(), false);
1820 return true;
1821 }
1822
1823
1824 //----------------------------inline_min_max-----------------------------------
1825 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1826 set_result(generate_min_max(id, argument(0), argument(1)));
1827 return true;
1828 }
1829
1830 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1831 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1832 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1833 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1834 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1835
1836 {
1837 PreserveJVMState pjvms(this);
1838 PreserveReexecuteState preexecs(this);
1839 jvms()->set_should_reexecute(true);
1840
1841 set_control(slow_path);
1842 set_i_o(i_o());
1843
1844 uncommon_trap(Deoptimization::Reason_intrinsic,
1845 Deoptimization::Action_none);
1846 }
1847
1848 set_control(fast_path);
1849 set_result(math);
1850 }
1851
1852 template <typename OverflowOp>
1853 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1854 typedef typename OverflowOp::MathOp MathOp;
1855
1856 MathOp* mathOp = new MathOp(arg1, arg2);
1857 Node* operation = _gvn.transform( mathOp );
1858 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1859 inline_math_mathExact(operation, ofcheck);
1860 return true;
1861 }
1862
1863 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1864 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1865 }
1866
1867 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1868 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1869 }
1870
1871 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1872 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1873 }
1874
1875 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1876 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1877 }
1878
1879 bool LibraryCallKit::inline_math_negateExactI() {
1880 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1881 }
1882
1883 bool LibraryCallKit::inline_math_negateExactL() {
1884 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1885 }
1886
1887 bool LibraryCallKit::inline_math_multiplyExactI() {
1888 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1889 }
1890
1891 bool LibraryCallKit::inline_math_multiplyExactL() {
1892 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1893 }
1894
1895 Node*
1896 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1897 // These are the candidate return value:
1898 Node* xvalue = x0;
1899 Node* yvalue = y0;
1900
1901 if (xvalue == yvalue) {
1902 return xvalue;
1903 }
1904
1905 bool want_max = (id == vmIntrinsics::_max);
1906
1907 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1908 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1909 if (txvalue == NULL || tyvalue == NULL) return top();
1910 // This is not really necessary, but it is consistent with a
1911 // hypothetical MaxINode::Value method:
1912 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1913
1914 // %%% This folding logic should (ideally) be in a different place.
1915 // Some should be inside IfNode, and there to be a more reliable
1916 // transformation of ?: style patterns into cmoves. We also want
1917 // more powerful optimizations around cmove and min/max.
1918
1919 // Try to find a dominating comparison of these guys.
1920 // It can simplify the index computation for Arrays.copyOf
1921 // and similar uses of System.arraycopy.
1922 // First, compute the normalized version of CmpI(x, y).
1923 int cmp_op = Op_CmpI;
1924 Node* xkey = xvalue;
1925 Node* ykey = yvalue;
1926 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1927 if (ideal_cmpxy->is_Cmp()) {
1928 // E.g., if we have CmpI(length - offset, count),
1929 // it might idealize to CmpI(length, count + offset)
1930 cmp_op = ideal_cmpxy->Opcode();
1931 xkey = ideal_cmpxy->in(1);
1932 ykey = ideal_cmpxy->in(2);
1933 }
1934
1935 // Start by locating any relevant comparisons.
1936 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1937 Node* cmpxy = NULL;
1938 Node* cmpyx = NULL;
1939 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1940 Node* cmp = start_from->fast_out(k);
1941 if (cmp->outcnt() > 0 && // must have prior uses
1942 cmp->in(0) == NULL && // must be context-independent
1943 cmp->Opcode() == cmp_op) { // right kind of compare
1944 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1945 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1946 }
1947 }
1948
1949 const int NCMPS = 2;
1950 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1951 int cmpn;
1952 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1953 if (cmps[cmpn] != NULL) break; // find a result
1954 }
1955 if (cmpn < NCMPS) {
1956 // Look for a dominating test that tells us the min and max.
1957 int depth = 0; // Limit search depth for speed
1958 Node* dom = control();
1959 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1960 if (++depth >= 100) break;
1961 Node* ifproj = dom;
1962 if (!ifproj->is_Proj()) continue;
1963 Node* iff = ifproj->in(0);
1964 if (!iff->is_If()) continue;
1965 Node* bol = iff->in(1);
1966 if (!bol->is_Bool()) continue;
1967 Node* cmp = bol->in(1);
1968 if (cmp == NULL) continue;
1969 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1970 if (cmps[cmpn] == cmp) break;
1971 if (cmpn == NCMPS) continue;
1972 BoolTest::mask btest = bol->as_Bool()->_test._test;
1973 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1974 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1975 // At this point, we know that 'x btest y' is true.
1976 switch (btest) {
1977 case BoolTest::eq:
1978 // They are proven equal, so we can collapse the min/max.
1979 // Either value is the answer. Choose the simpler.
1980 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1981 return yvalue;
1982 return xvalue;
1983 case BoolTest::lt: // x < y
1984 case BoolTest::le: // x <= y
1985 return (want_max ? yvalue : xvalue);
1986 case BoolTest::gt: // x > y
1987 case BoolTest::ge: // x >= y
1988 return (want_max ? xvalue : yvalue);
1989 }
1990 }
1991 }
1992
1993 // We failed to find a dominating test.
1994 // Let's pick a test that might GVN with prior tests.
1995 Node* best_bol = NULL;
1996 BoolTest::mask best_btest = BoolTest::illegal;
1997 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1998 Node* cmp = cmps[cmpn];
1999 if (cmp == NULL) continue;
2000 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2001 Node* bol = cmp->fast_out(j);
2002 if (!bol->is_Bool()) continue;
2003 BoolTest::mask btest = bol->as_Bool()->_test._test;
2004 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2005 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2006 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2007 best_bol = bol->as_Bool();
2008 best_btest = btest;
2009 }
2010 }
2011 }
2012
2013 Node* answer_if_true = NULL;
2014 Node* answer_if_false = NULL;
2015 switch (best_btest) {
2016 default:
2017 if (cmpxy == NULL)
2018 cmpxy = ideal_cmpxy;
2019 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2020 // and fall through:
2021 case BoolTest::lt: // x < y
2022 case BoolTest::le: // x <= y
2023 answer_if_true = (want_max ? yvalue : xvalue);
2024 answer_if_false = (want_max ? xvalue : yvalue);
2025 break;
2026 case BoolTest::gt: // x > y
2027 case BoolTest::ge: // x >= y
2028 answer_if_true = (want_max ? xvalue : yvalue);
2029 answer_if_false = (want_max ? yvalue : xvalue);
2030 break;
2031 }
2032
2033 jint hi, lo;
2034 if (want_max) {
2035 // We can sharpen the minimum.
2036 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2037 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2038 } else {
2039 // We can sharpen the maximum.
2040 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2041 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2042 }
2043
2044 // Use a flow-free graph structure, to avoid creating excess control edges
2045 // which could hinder other optimizations.
2046 // Since Math.min/max is often used with arraycopy, we want
2047 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2048 Node* cmov = CMoveNode::make(NULL, best_bol,
2049 answer_if_false, answer_if_true,
2050 TypeInt::make(lo, hi, widen));
2051
2052 return _gvn.transform(cmov);
2053
2054 /*
2055 // This is not as desirable as it may seem, since Min and Max
2056 // nodes do not have a full set of optimizations.
2057 // And they would interfere, anyway, with 'if' optimizations
2058 // and with CMoveI canonical forms.
2059 switch (id) {
2060 case vmIntrinsics::_min:
2061 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2062 case vmIntrinsics::_max:
2063 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2064 default:
2065 ShouldNotReachHere();
2066 }
2067 */
2068 }
2069
2070 inline int
2071 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2072 const TypePtr* base_type = TypePtr::NULL_PTR;
2073 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2074 if (base_type == NULL) {
2075 // Unknown type.
2076 return Type::AnyPtr;
2077 } else if (base_type == TypePtr::NULL_PTR) {
2078 // Since this is a NULL+long form, we have to switch to a rawptr.
2079 base = _gvn.transform(new CastX2PNode(offset));
2080 offset = MakeConX(0);
2081 return Type::RawPtr;
2082 } else if (base_type->base() == Type::RawPtr) {
2083 return Type::RawPtr;
2084 } else if (base_type->isa_oopptr()) {
2085 // Base is never null => always a heap address.
2086 if (base_type->ptr() == TypePtr::NotNull) {
2087 return Type::OopPtr;
2088 }
2089 // Offset is small => always a heap address.
2090 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2091 if (offset_type != NULL &&
2092 base_type->offset() == 0 && // (should always be?)
2093 offset_type->_lo >= 0 &&
2094 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2095 return Type::OopPtr;
2096 }
2097 // Otherwise, it might either be oop+off or NULL+addr.
2098 return Type::AnyPtr;
2099 } else {
2100 // No information:
2101 return Type::AnyPtr;
2102 }
2103 }
2104
2105 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2106 int kind = classify_unsafe_addr(base, offset);
2107 if (kind == Type::RawPtr) {
2108 return basic_plus_adr(top(), base, offset);
2109 } else {
2110 return basic_plus_adr(base, offset);
2111 }
2112 }
2113
2114 //--------------------------inline_number_methods-----------------------------
2115 // inline int Integer.numberOfLeadingZeros(int)
2116 // inline int Long.numberOfLeadingZeros(long)
2117 //
2118 // inline int Integer.numberOfTrailingZeros(int)
2119 // inline int Long.numberOfTrailingZeros(long)
2120 //
2121 // inline int Integer.bitCount(int)
2122 // inline int Long.bitCount(long)
2123 //
2124 // inline char Character.reverseBytes(char)
2125 // inline short Short.reverseBytes(short)
2126 // inline int Integer.reverseBytes(int)
2127 // inline long Long.reverseBytes(long)
2128 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2129 Node* arg = argument(0);
2130 Node* n = NULL;
2131 switch (id) {
2132 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2133 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2134 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2135 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2136 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2137 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2138 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2139 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2140 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2141 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2142 default: fatal_unexpected_iid(id); break;
2143 }
2144 set_result(_gvn.transform(n));
2145 return true;
2146 }
2147
2148 //----------------------------inline_unsafe_access----------------------------
2149
2150 // Helper that guards and inserts a pre-barrier.
2151 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2152 Node* pre_val, bool need_mem_bar) {
2153 // We could be accessing the referent field of a reference object. If so, when G1
2154 // is enabled, we need to log the value in the referent field in an SATB buffer.
2155 // This routine performs some compile time filters and generates suitable
2156 // runtime filters that guard the pre-barrier code.
2157 // Also add memory barrier for non volatile load from the referent field
2158 // to prevent commoning of loads across safepoint.
2159 if (!UseG1GC && !need_mem_bar)
2160 return;
2161
2162 // Some compile time checks.
2163
2164 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2165 const TypeX* otype = offset->find_intptr_t_type();
2166 if (otype != NULL && otype->is_con() &&
2167 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2168 // Constant offset but not the reference_offset so just return
2169 return;
2170 }
2171
2172 // We only need to generate the runtime guards for instances.
2173 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2174 if (btype != NULL) {
2175 if (btype->isa_aryptr()) {
2176 // Array type so nothing to do
2177 return;
2178 }
2179
2180 const TypeInstPtr* itype = btype->isa_instptr();
2181 if (itype != NULL) {
2182 // Can the klass of base_oop be statically determined to be
2183 // _not_ a sub-class of Reference and _not_ Object?
2184 ciKlass* klass = itype->klass();
2185 if ( klass->is_loaded() &&
2186 !klass->is_subtype_of(env()->Reference_klass()) &&
2187 !env()->Object_klass()->is_subtype_of(klass)) {
2188 return;
2189 }
2190 }
2191 }
2192
2193 // The compile time filters did not reject base_oop/offset so
2194 // we need to generate the following runtime filters
2195 //
2196 // if (offset == java_lang_ref_Reference::_reference_offset) {
2197 // if (instance_of(base, java.lang.ref.Reference)) {
2198 // pre_barrier(_, pre_val, ...);
2199 // }
2200 // }
2201
2202 float likely = PROB_LIKELY( 0.999);
2203 float unlikely = PROB_UNLIKELY(0.999);
2204
2205 IdealKit ideal(this);
2206 #define __ ideal.
2207
2208 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2209
2210 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2211 // Update graphKit memory and control from IdealKit.
2212 sync_kit(ideal);
2213
2214 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2215 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2216
2217 // Update IdealKit memory and control from graphKit.
2218 __ sync_kit(this);
2219
2220 Node* one = __ ConI(1);
2221 // is_instof == 0 if base_oop == NULL
2222 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2223
2224 // Update graphKit from IdeakKit.
2225 sync_kit(ideal);
2226
2227 // Use the pre-barrier to record the value in the referent field
2228 pre_barrier(false /* do_load */,
2229 __ ctrl(),
2230 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2231 pre_val /* pre_val */,
2232 T_OBJECT);
2233 if (need_mem_bar) {
2234 // Add memory barrier to prevent commoning reads from this field
2235 // across safepoint since GC can change its value.
2236 insert_mem_bar(Op_MemBarCPUOrder);
2237 }
2238 // Update IdealKit from graphKit.
2239 __ sync_kit(this);
2240
2241 } __ end_if(); // _ref_type != ref_none
2242 } __ end_if(); // offset == referent_offset
2243
2244 // Final sync IdealKit and GraphKit.
2245 final_sync(ideal);
2246 #undef __
2247 }
2248
2249
2250 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2251 // Attempt to infer a sharper value type from the offset and base type.
2252 ciKlass* sharpened_klass = NULL;
2253
2254 // See if it is an instance field, with an object type.
2255 if (alias_type->field() != NULL) {
2256 if (alias_type->field()->type()->is_klass()) {
2257 sharpened_klass = alias_type->field()->type()->as_klass();
2258 }
2259 }
2260
2261 // See if it is a narrow oop array.
2262 if (adr_type->isa_aryptr()) {
2263 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2264 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2265 if (elem_type != NULL) {
2266 sharpened_klass = elem_type->klass();
2267 }
2268 }
2269 }
2270
2271 // The sharpened class might be unloaded if there is no class loader
2272 // contraint in place.
2273 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2274 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2275
2276 #ifndef PRODUCT
2277 if (C->print_intrinsics() || C->print_inlining()) {
2278 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2279 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2280 }
2281 #endif
2282 // Sharpen the value type.
2283 return tjp;
2284 }
2285 return NULL;
2286 }
2287
2288 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2289 if (callee()->is_static()) return false; // caller must have the capability!
2290 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2291 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2292 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2293
2294 #ifndef PRODUCT
2295 {
2296 ResourceMark rm;
2297 // Check the signatures.
2298 ciSignature* sig = callee()->signature();
2299 #ifdef ASSERT
2300 if (!is_store) {
2301 // Object getObject(Object base, int/long offset), etc.
2302 BasicType rtype = sig->return_type()->basic_type();
2303 assert(rtype == type, "getter must return the expected value");
2304 assert(sig->count() == 2, "oop getter has 2 arguments");
2305 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2306 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2307 } else {
2308 // void putObject(Object base, int/long offset, Object x), etc.
2309 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2310 assert(sig->count() == 3, "oop putter has 3 arguments");
2311 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2312 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2313 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2314 assert(vtype == type, "putter must accept the expected value");
2315 }
2316 #endif // ASSERT
2317 }
2318 #endif //PRODUCT
2319
2320 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2321
2322 Node* receiver = argument(0); // type: oop
2323
2324 // Build address expression.
2325 Node* adr;
2326 Node* heap_base_oop = top();
2327 Node* offset = top();
2328 Node* val;
2329
2330 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2331 Node* base = argument(1); // type: oop
2332 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2333 offset = argument(2); // type: long
2334 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2335 // to be plain byte offsets, which are also the same as those accepted
2336 // by oopDesc::field_base.
2337 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2338 "fieldOffset must be byte-scaled");
2339 // 32-bit machines ignore the high half!
2340 offset = ConvL2X(offset);
2341 adr = make_unsafe_address(base, offset);
2342 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2343 heap_base_oop = base;
2344 } else if (type == T_OBJECT) {
2345 return false; // off-heap oop accesses are not supported
2346 }
2347
2348 // Can base be NULL? Otherwise, always on-heap access.
2349 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2350
2351 val = is_store ? argument(4) : NULL;
2352
2353 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2354
2355 // Try to categorize the address.
2356 Compile::AliasType* alias_type = C->alias_type(adr_type);
2357 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2358
2359 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2360 alias_type->adr_type() == TypeAryPtr::RANGE) {
2361 return false; // not supported
2362 }
2363
2364 bool mismatched = false;
2365 BasicType bt = alias_type->basic_type();
2366 if (bt != T_ILLEGAL) {
2367 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2368 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2369 // Alias type doesn't differentiate between byte[] and boolean[]).
2370 // Use address type to get the element type.
2371 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2372 }
2373 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2374 // accessing an array field with getObject is not a mismatch
2375 bt = T_OBJECT;
2376 }
2377 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2378 // Don't intrinsify mismatched object accesses
2379 return false;
2380 }
2381 mismatched = (bt != type);
2382 } else if (alias_type->adr_type()->isa_oopptr()) {
2383 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2384 }
2385
2386 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2387
2388 // First guess at the value type.
2389 const Type *value_type = Type::get_const_basic_type(type);
2390
2391 // We will need memory barriers unless we can determine a unique
2392 // alias category for this reference. (Note: If for some reason
2393 // the barriers get omitted and the unsafe reference begins to "pollute"
2394 // the alias analysis of the rest of the graph, either Compile::can_alias
2395 // or Compile::must_alias will throw a diagnostic assert.)
2396 bool need_mem_bar;
2397 switch (kind) {
2398 case Relaxed:
2399 need_mem_bar = mismatched || can_access_non_heap;
2400 break;
2401 case Opaque:
2402 // Opaque uses CPUOrder membars for protection against code movement.
2403 case Acquire:
2404 case Release:
2405 case Volatile:
2406 need_mem_bar = true;
2407 break;
2408 default:
2409 ShouldNotReachHere();
2410 }
2411
2412 // Some accesses require access atomicity for all types, notably longs and doubles.
2413 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2414 bool requires_atomic_access = false;
2415 switch (kind) {
2416 case Relaxed:
2417 requires_atomic_access = AlwaysAtomicAccesses;
2418 break;
2419 case Opaque:
2420 // Opaque accesses are atomic.
2421 case Acquire:
2422 case Release:
2423 case Volatile:
2424 requires_atomic_access = true;
2425 break;
2426 default:
2427 ShouldNotReachHere();
2428 }
2429
2430 // Figure out the memory ordering.
2431 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2432 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2433
2434 // If we are reading the value of the referent field of a Reference
2435 // object (either by using Unsafe directly or through reflection)
2436 // then, if G1 is enabled, we need to record the referent in an
2437 // SATB log buffer using the pre-barrier mechanism.
2438 // Also we need to add memory barrier to prevent commoning reads
2439 // from this field across safepoint since GC can change its value.
2440 bool need_read_barrier = !is_store &&
2441 offset != top() && heap_base_oop != top();
2442
2443 if (!is_store && type == T_OBJECT) {
2444 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2445 if (tjp != NULL) {
2446 value_type = tjp;
2447 }
2448 }
2449
2450 receiver = null_check(receiver);
2451 if (stopped()) {
2452 return true;
2453 }
2454 // Heap pointers get a null-check from the interpreter,
2455 // as a courtesy. However, this is not guaranteed by Unsafe,
2456 // and it is not possible to fully distinguish unintended nulls
2457 // from intended ones in this API.
2458
2459 // We need to emit leading and trailing CPU membars (see below) in
2460 // addition to memory membars for special access modes. This is a little
2461 // too strong, but avoids the need to insert per-alias-type
2462 // volatile membars (for stores; compare Parse::do_put_xxx), which
2463 // we cannot do effectively here because we probably only have a
2464 // rough approximation of type.
2465
2466 switch(kind) {
2467 case Relaxed:
2468 case Opaque:
2469 case Acquire:
2470 break;
2471 case Release:
2472 case Volatile:
2473 if (is_store) {
2474 insert_mem_bar(Op_MemBarRelease);
2475 } else {
2476 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2477 insert_mem_bar(Op_MemBarVolatile);
2478 }
2479 }
2480 break;
2481 default:
2482 ShouldNotReachHere();
2483 }
2484
2485 // Memory barrier to prevent normal and 'unsafe' accesses from
2486 // bypassing each other. Happens after null checks, so the
2487 // exception paths do not take memory state from the memory barrier,
2488 // so there's no problems making a strong assert about mixing users
2489 // of safe & unsafe memory.
2490 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2491
2492 if (!is_store) {
2493 Node* p = NULL;
2494 // Try to constant fold a load from a constant field
2495 ciField* field = alias_type->field();
2496 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2497 // final or stable field
2498 p = make_constant_from_field(field, heap_base_oop);
2499 }
2500 if (p == NULL) {
2501 // To be valid, unsafe loads may depend on other conditions than
2502 // the one that guards them: pin the Load node
2503 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched);
2504 // load value
2505 switch (type) {
2506 case T_BOOLEAN:
2507 {
2508 // Normalize the value returned by getBoolean in the following cases
2509 if (mismatched ||
2510 heap_base_oop == top() || // - heap_base_oop is NULL or
2511 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2512 // and the unsafe access is made to large offset
2513 // (i.e., larger than the maximum offset necessary for any
2514 // field access)
2515 ) {
2516 IdealKit ideal = IdealKit(this);
2517 #define __ ideal.
2518 IdealVariable normalized_result(ideal);
2519 __ declarations_done();
2520 __ set(normalized_result, p);
2521 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2522 __ set(normalized_result, ideal.ConI(1));
2523 ideal.end_if();
2524 final_sync(ideal);
2525 p = __ value(normalized_result);
2526 #undef __
2527 }
2528 }
2529 case T_CHAR:
2530 case T_BYTE:
2531 case T_SHORT:
2532 case T_INT:
2533 case T_LONG:
2534 case T_FLOAT:
2535 case T_DOUBLE:
2536 break;
2537 case T_OBJECT:
2538 if (need_read_barrier) {
2539 // We do not require a mem bar inside pre_barrier if need_mem_bar
2540 // is set: the barriers would be emitted by us.
2541 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2542 }
2543 break;
2544 case T_ADDRESS:
2545 // Cast to an int type.
2546 p = _gvn.transform(new CastP2XNode(NULL, p));
2547 p = ConvX2UL(p);
2548 break;
2549 default:
2550 fatal("unexpected type %d: %s", type, type2name(type));
2551 break;
2552 }
2553 }
2554 // The load node has the control of the preceding MemBarCPUOrder. All
2555 // following nodes will have the control of the MemBarCPUOrder inserted at
2556 // the end of this method. So, pushing the load onto the stack at a later
2557 // point is fine.
2558 set_result(p);
2559 } else {
2560 // place effect of store into memory
2561 switch (type) {
2562 case T_DOUBLE:
2563 val = dstore_rounding(val);
2564 break;
2565 case T_ADDRESS:
2566 // Repackage the long as a pointer.
2567 val = ConvL2X(val);
2568 val = _gvn.transform(new CastX2PNode(val));
2569 break;
2570 }
2571
2572 if (type == T_OBJECT) {
2573 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2574 } else {
2575 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2576 }
2577 }
2578
2579 switch(kind) {
2580 case Relaxed:
2581 case Opaque:
2582 case Release:
2583 break;
2584 case Acquire:
2585 case Volatile:
2586 if (!is_store) {
2587 insert_mem_bar(Op_MemBarAcquire);
2588 } else {
2589 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2590 insert_mem_bar(Op_MemBarVolatile);
2591 }
2592 }
2593 break;
2594 default:
2595 ShouldNotReachHere();
2596 }
2597
2598 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2599
2600 return true;
2601 }
2602
2603 //----------------------------inline_unsafe_load_store----------------------------
2604 // This method serves a couple of different customers (depending on LoadStoreKind):
2605 //
2606 // LS_cmp_swap:
2607 //
2608 // boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2609 // boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2610 // boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2611 //
2612 // LS_cmp_swap_weak:
2613 //
2614 // boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x);
2615 // boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x);
2616 // boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x);
2617 //
2618 // boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x);
2619 // boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x);
2620 // boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x);
2621 //
2622 // boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x);
2623 // boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x);
2624 // boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x);
2625 //
2626 // LS_cmp_exchange:
2627 //
2628 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2629 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2630 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2631 //
2632 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2633 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2634 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2635 //
2636 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2637 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2638 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2639 //
2640 // LS_get_add:
2641 //
2642 // int getAndAddInt( Object o, long offset, int delta)
2643 // long getAndAddLong(Object o, long offset, long delta)
2644 //
2645 // LS_get_set:
2646 //
2647 // int getAndSet(Object o, long offset, int newValue)
2648 // long getAndSet(Object o, long offset, long newValue)
2649 // Object getAndSet(Object o, long offset, Object newValue)
2650 //
2651 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2652 // This basic scheme here is the same as inline_unsafe_access, but
2653 // differs in enough details that combining them would make the code
2654 // overly confusing. (This is a true fact! I originally combined
2655 // them, but even I was confused by it!) As much code/comments as
2656 // possible are retained from inline_unsafe_access though to make
2657 // the correspondences clearer. - dl
2658
2659 if (callee()->is_static()) return false; // caller must have the capability!
2660
2661 #ifndef PRODUCT
2662 BasicType rtype;
2663 {
2664 ResourceMark rm;
2665 // Check the signatures.
2666 ciSignature* sig = callee()->signature();
2667 rtype = sig->return_type()->basic_type();
2668 switch(kind) {
2669 case LS_get_add:
2670 case LS_get_set: {
2671 // Check the signatures.
2672 #ifdef ASSERT
2673 assert(rtype == type, "get and set must return the expected type");
2674 assert(sig->count() == 3, "get and set has 3 arguments");
2675 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2676 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2677 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2678 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2679 #endif // ASSERT
2680 break;
2681 }
2682 case LS_cmp_swap:
2683 case LS_cmp_swap_weak: {
2684 // Check the signatures.
2685 #ifdef ASSERT
2686 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2687 assert(sig->count() == 4, "CAS has 4 arguments");
2688 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2689 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2690 #endif // ASSERT
2691 break;
2692 }
2693 case LS_cmp_exchange: {
2694 // Check the signatures.
2695 #ifdef ASSERT
2696 assert(rtype == type, "CAS must return the expected type");
2697 assert(sig->count() == 4, "CAS has 4 arguments");
2698 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2699 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2700 #endif // ASSERT
2701 break;
2702 }
2703 default:
2704 ShouldNotReachHere();
2705 }
2706 }
2707 #endif //PRODUCT
2708
2709 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2710
2711 // Get arguments:
2712 Node* receiver = NULL;
2713 Node* base = NULL;
2714 Node* offset = NULL;
2715 Node* oldval = NULL;
2716 Node* newval = NULL;
2717 switch(kind) {
2718 case LS_cmp_swap:
2719 case LS_cmp_swap_weak:
2720 case LS_cmp_exchange: {
2721 const bool two_slot_type = type2size[type] == 2;
2722 receiver = argument(0); // type: oop
2723 base = argument(1); // type: oop
2724 offset = argument(2); // type: long
2725 oldval = argument(4); // type: oop, int, or long
2726 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2727 break;
2728 }
2729 case LS_get_add:
2730 case LS_get_set: {
2731 receiver = argument(0); // type: oop
2732 base = argument(1); // type: oop
2733 offset = argument(2); // type: long
2734 oldval = NULL;
2735 newval = argument(4); // type: oop, int, or long
2736 break;
2737 }
2738 default:
2739 ShouldNotReachHere();
2740 }
2741
2742 // Build field offset expression.
2743 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2744 // to be plain byte offsets, which are also the same as those accepted
2745 // by oopDesc::field_base.
2746 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2747 // 32-bit machines ignore the high half of long offsets
2748 offset = ConvL2X(offset);
2749 Node* adr = make_unsafe_address(base, offset);
2750 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2751
2752 Compile::AliasType* alias_type = C->alias_type(adr_type);
2753 BasicType bt = alias_type->basic_type();
2754 if (bt != T_ILLEGAL &&
2755 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2756 // Don't intrinsify mismatched object accesses.
2757 return false;
2758 }
2759
2760 // For CAS, unlike inline_unsafe_access, there seems no point in
2761 // trying to refine types. Just use the coarse types here.
2762 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2763 const Type *value_type = Type::get_const_basic_type(type);
2764
2765 switch (kind) {
2766 case LS_get_set:
2767 case LS_cmp_exchange: {
2768 if (type == T_OBJECT) {
2769 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2770 if (tjp != NULL) {
2771 value_type = tjp;
2772 }
2773 }
2774 break;
2775 }
2776 case LS_cmp_swap:
2777 case LS_cmp_swap_weak:
2778 case LS_get_add:
2779 break;
2780 default:
2781 ShouldNotReachHere();
2782 }
2783
2784 // Null check receiver.
2785 receiver = null_check(receiver);
2786 if (stopped()) {
2787 return true;
2788 }
2789
2790 int alias_idx = C->get_alias_index(adr_type);
2791
2792 // Memory-model-wise, a LoadStore acts like a little synchronized
2793 // block, so needs barriers on each side. These don't translate
2794 // into actual barriers on most machines, but we still need rest of
2795 // compiler to respect ordering.
2796
2797 switch (access_kind) {
2798 case Relaxed:
2799 case Acquire:
2800 break;
2801 case Release:
2802 insert_mem_bar(Op_MemBarRelease);
2803 break;
2804 case Volatile:
2805 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2806 insert_mem_bar(Op_MemBarVolatile);
2807 } else {
2808 insert_mem_bar(Op_MemBarRelease);
2809 }
2810 break;
2811 default:
2812 ShouldNotReachHere();
2813 }
2814 insert_mem_bar(Op_MemBarCPUOrder);
2815
2816 // Figure out the memory ordering.
2817 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2818
2819 // 4984716: MemBars must be inserted before this
2820 // memory node in order to avoid a false
2821 // dependency which will confuse the scheduler.
2822 Node *mem = memory(alias_idx);
2823
2824 // For now, we handle only those cases that actually exist: ints,
2825 // longs, and Object. Adding others should be straightforward.
2826 Node* load_store = NULL;
2827 switch(type) {
2828 case T_BYTE:
2829 switch(kind) {
2830 case LS_get_add:
2831 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2832 break;
2833 case LS_get_set:
2834 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2835 break;
2836 case LS_cmp_swap_weak:
2837 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2838 break;
2839 case LS_cmp_swap:
2840 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2841 break;
2842 case LS_cmp_exchange:
2843 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2844 break;
2845 default:
2846 ShouldNotReachHere();
2847 }
2848 break;
2849 case T_SHORT:
2850 switch(kind) {
2851 case LS_get_add:
2852 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2853 break;
2854 case LS_get_set:
2855 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2856 break;
2857 case LS_cmp_swap_weak:
2858 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2859 break;
2860 case LS_cmp_swap:
2861 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2862 break;
2863 case LS_cmp_exchange:
2864 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2865 break;
2866 default:
2867 ShouldNotReachHere();
2868 }
2869 break;
2870 case T_INT:
2871 switch(kind) {
2872 case LS_get_add:
2873 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2874 break;
2875 case LS_get_set:
2876 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2877 break;
2878 case LS_cmp_swap_weak:
2879 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2880 break;
2881 case LS_cmp_swap:
2882 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2883 break;
2884 case LS_cmp_exchange:
2885 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2886 break;
2887 default:
2888 ShouldNotReachHere();
2889 }
2890 break;
2891 case T_LONG:
2892 switch(kind) {
2893 case LS_get_add:
2894 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2895 break;
2896 case LS_get_set:
2897 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2898 break;
2899 case LS_cmp_swap_weak:
2900 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2901 break;
2902 case LS_cmp_swap:
2903 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2904 break;
2905 case LS_cmp_exchange:
2906 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2907 break;
2908 default:
2909 ShouldNotReachHere();
2910 }
2911 break;
2912 case T_OBJECT:
2913 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2914 // could be delayed during Parse (for example, in adjust_map_after_if()).
2915 // Execute transformation here to avoid barrier generation in such case.
2916 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2917 newval = _gvn.makecon(TypePtr::NULL_PTR);
2918
2919 // Reference stores need a store barrier.
2920 switch(kind) {
2921 case LS_get_set: {
2922 // If pre-barrier must execute before the oop store, old value will require do_load here.
2923 if (!can_move_pre_barrier()) {
2924 pre_barrier(true /* do_load*/,
2925 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2926 NULL /* pre_val*/,
2927 T_OBJECT);
2928 } // Else move pre_barrier to use load_store value, see below.
2929 break;
2930 }
2931 case LS_cmp_swap_weak:
2932 case LS_cmp_swap:
2933 case LS_cmp_exchange: {
2934 // Same as for newval above:
2935 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2936 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2937 }
2938 // The only known value which might get overwritten is oldval.
2939 pre_barrier(false /* do_load */,
2940 control(), NULL, NULL, max_juint, NULL, NULL,
2941 oldval /* pre_val */,
2942 T_OBJECT);
2943 break;
2944 }
2945 default:
2946 ShouldNotReachHere();
2947 }
2948
2949 #ifdef _LP64
2950 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2951 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2952
2953 switch(kind) {
2954 case LS_get_set:
2955 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
2956 break;
2957 case LS_cmp_swap_weak: {
2958 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2959 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2960 break;
2961 }
2962 case LS_cmp_swap: {
2963 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2964 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2965 break;
2966 }
2967 case LS_cmp_exchange: {
2968 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2969 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
2970 break;
2971 }
2972 default:
2973 ShouldNotReachHere();
2974 }
2975 } else
2976 #endif
2977 switch (kind) {
2978 case LS_get_set:
2979 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2980 break;
2981 case LS_cmp_swap_weak:
2982 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2983 break;
2984 case LS_cmp_swap:
2985 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2986 break;
2987 case LS_cmp_exchange:
2988 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
2989 break;
2990 default:
2991 ShouldNotReachHere();
2992 }
2993
2994 // Emit the post barrier only when the actual store happened. This makes sense
2995 // to check only for LS_cmp_* that can fail to set the value.
2996 // LS_cmp_exchange does not produce any branches by default, so there is no
2997 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
2998 // CompareAndExchange and move branches here, it would make sense to conditionalize
2999 // post_barriers for LS_cmp_exchange as well.
3000 //
3001 // CAS success path is marked more likely since we anticipate this is a performance
3002 // critical path, while CAS failure path can use the penalty for going through unlikely
3003 // path as backoff. Which is still better than doing a store barrier there.
3004 switch (kind) {
3005 case LS_get_set:
3006 case LS_cmp_exchange: {
3007 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3008 break;
3009 }
3010 case LS_cmp_swap_weak:
3011 case LS_cmp_swap: {
3012 IdealKit ideal(this);
3013 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3014 sync_kit(ideal);
3015 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3016 ideal.sync_kit(this);
3017 } ideal.end_if();
3018 final_sync(ideal);
3019 break;
3020 }
3021 default:
3022 ShouldNotReachHere();
3023 }
3024 break;
3025 default:
3026 fatal("unexpected type %d: %s", type, type2name(type));
3027 break;
3028 }
3029
3030 // SCMemProjNodes represent the memory state of a LoadStore. Their
3031 // main role is to prevent LoadStore nodes from being optimized away
3032 // when their results aren't used.
3033 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3034 set_memory(proj, alias_idx);
3035
3036 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3037 #ifdef _LP64
3038 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3039 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3040 }
3041 #endif
3042 if (can_move_pre_barrier()) {
3043 // Don't need to load pre_val. The old value is returned by load_store.
3044 // The pre_barrier can execute after the xchg as long as no safepoint
3045 // gets inserted between them.
3046 pre_barrier(false /* do_load */,
3047 control(), NULL, NULL, max_juint, NULL, NULL,
3048 load_store /* pre_val */,
3049 T_OBJECT);
3050 }
3051 }
3052
3053 // Add the trailing membar surrounding the access
3054 insert_mem_bar(Op_MemBarCPUOrder);
3055
3056 switch (access_kind) {
3057 case Relaxed:
3058 case Release:
3059 break; // do nothing
3060 case Acquire:
3061 case Volatile:
3062 insert_mem_bar(Op_MemBarAcquire);
3063 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3064 break;
3065 default:
3066 ShouldNotReachHere();
3067 }
3068
3069 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3070 set_result(load_store);
3071 return true;
3072 }
3073
3074 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3075 MemNode::MemOrd mo = MemNode::unset;
3076 switch(kind) {
3077 case Opaque:
3078 case Relaxed: mo = MemNode::unordered; break;
3079 case Acquire: mo = MemNode::acquire; break;
3080 case Release: mo = MemNode::release; break;
3081 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3082 default:
3083 ShouldNotReachHere();
3084 }
3085 guarantee(mo != MemNode::unset, "Should select memory ordering");
3086 return mo;
3087 }
3088
3089 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3090 MemNode::MemOrd mo = MemNode::unset;
3091 switch(kind) {
3092 case Opaque:
3093 case Relaxed: mo = MemNode::unordered; break;
3094 case Acquire: mo = MemNode::acquire; break;
3095 case Release: mo = MemNode::release; break;
3096 case Volatile: mo = MemNode::seqcst; break;
3097 default:
3098 ShouldNotReachHere();
3099 }
3100 guarantee(mo != MemNode::unset, "Should select memory ordering");
3101 return mo;
3102 }
3103
3104 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3105 // Regardless of form, don't allow previous ld/st to move down,
3106 // then issue acquire, release, or volatile mem_bar.
3107 insert_mem_bar(Op_MemBarCPUOrder);
3108 switch(id) {
3109 case vmIntrinsics::_loadFence:
3110 insert_mem_bar(Op_LoadFence);
3111 return true;
3112 case vmIntrinsics::_storeFence:
3113 insert_mem_bar(Op_StoreFence);
3114 return true;
3115 case vmIntrinsics::_fullFence:
3116 insert_mem_bar(Op_MemBarVolatile);
3117 return true;
3118 default:
3119 fatal_unexpected_iid(id);
3120 return false;
3121 }
3122 }
3123
3124 bool LibraryCallKit::inline_onspinwait() {
3125 insert_mem_bar(Op_OnSpinWait);
3126 return true;
3127 }
3128
3129 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3130 if (!kls->is_Con()) {
3131 return true;
3132 }
3133 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3134 if (klsptr == NULL) {
3135 return true;
3136 }
3137 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3138 // don't need a guard for a klass that is already initialized
3139 return !ik->is_initialized();
3140 }
3141
3142 //----------------------------inline_unsafe_allocate---------------------------
3143 // public native Object Unsafe.allocateInstance(Class<?> cls);
3144 bool LibraryCallKit::inline_unsafe_allocate() {
3145 if (callee()->is_static()) return false; // caller must have the capability!
3146
3147 null_check_receiver(); // null-check, then ignore
3148 Node* cls = null_check(argument(1));
3149 if (stopped()) return true;
3150
3151 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3152 kls = null_check(kls);
3153 if (stopped()) return true; // argument was like int.class
3154
3155 Node* test = NULL;
3156 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3157 // Note: The argument might still be an illegal value like
3158 // Serializable.class or Object[].class. The runtime will handle it.
3159 // But we must make an explicit check for initialization.
3160 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3161 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3162 // can generate code to load it as unsigned byte.
3163 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3164 Node* bits = intcon(InstanceKlass::fully_initialized);
3165 test = _gvn.transform(new SubINode(inst, bits));
3166 // The 'test' is non-zero if we need to take a slow path.
3167 }
3168
3169 Node* obj = new_instance(kls, test);
3170 set_result(obj);
3171 return true;
3172 }
3173
3174 //------------------------inline_native_time_funcs--------------
3175 // inline code for System.currentTimeMillis() and System.nanoTime()
3176 // these have the same type and signature
3177 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3178 const TypeFunc* tf = OptoRuntime::void_long_Type();
3179 const TypePtr* no_memory_effects = NULL;
3180 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3181 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3182 #ifdef ASSERT
3183 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3184 assert(value_top == top(), "second value must be top");
3185 #endif
3186 set_result(value);
3187 return true;
3188 }
3189
3190 #ifdef TRACE_HAVE_INTRINSICS
3191
3192 /*
3193 * oop -> myklass
3194 * myklass->trace_id |= USED
3195 * return myklass->trace_id & ~0x3
3196 */
3197 bool LibraryCallKit::inline_native_classID() {
3198 Node* cls = null_check(argument(0), T_OBJECT);
3199 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3200 kls = null_check(kls, T_OBJECT);
3201
3202 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3203 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3204 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3205
3206 Node* clsused = longcon(0x01l); // set the class bit
3207 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3208 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3209 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3210
3211 #ifdef TRACE_ID_META_BITS
3212 Node* mbits = longcon(~TRACE_ID_META_BITS);
3213 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3214 #endif
3215 #ifdef TRACE_ID_CLASS_SHIFT
3216 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3217 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3218 #endif
3219
3220 set_result(tvalue);
3221 return true;
3222
3223 }
3224
3225 bool LibraryCallKit::inline_native_getBufferWriter() {
3226 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3227
3228 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3229 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3230 );
3231
3232 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3233
3234 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3235 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3236
3237 IfNode* iff_jobj_null =
3238 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3239
3240 enum { _normal_path = 1,
3241 _null_path = 2,
3242 PATH_LIMIT };
3243
3244 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3245 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3246
3247 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3248 result_rgn->init_req(_null_path, jobj_is_null);
3249 result_val->init_req(_null_path, null());
3250
3251 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3252 result_rgn->init_req(_normal_path, jobj_is_not_null);
3253
3254 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3255 result_val->init_req(_normal_path, res);
3256
3257 set_result(result_rgn, result_val);
3258
3259 return true;
3260 }
3261
3262 #endif
3263
3264 //------------------------inline_native_currentThread------------------
3265 bool LibraryCallKit::inline_native_currentThread() {
3266 Node* junk = NULL;
3267 set_result(generate_current_thread(junk));
3268 return true;
3269 }
3270
3271 //------------------------inline_native_isInterrupted------------------
3272 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3273 bool LibraryCallKit::inline_native_isInterrupted() {
3274 // Add a fast path to t.isInterrupted(clear_int):
3275 // (t == Thread.current() &&
3276 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3277 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3278 // So, in the common case that the interrupt bit is false,
3279 // we avoid making a call into the VM. Even if the interrupt bit
3280 // is true, if the clear_int argument is false, we avoid the VM call.
3281 // However, if the receiver is not currentThread, we must call the VM,
3282 // because there must be some locking done around the operation.
3283
3284 // We only go to the fast case code if we pass two guards.
3285 // Paths which do not pass are accumulated in the slow_region.
3286
3287 enum {
3288 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3289 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3290 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3291 PATH_LIMIT
3292 };
3293
3294 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3295 // out of the function.
3296 insert_mem_bar(Op_MemBarCPUOrder);
3297
3298 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3299 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3300
3301 RegionNode* slow_region = new RegionNode(1);
3302 record_for_igvn(slow_region);
3303
3304 // (a) Receiving thread must be the current thread.
3305 Node* rec_thr = argument(0);
3306 Node* tls_ptr = NULL;
3307 Node* cur_thr = generate_current_thread(tls_ptr);
3308 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3309 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3310
3311 generate_slow_guard(bol_thr, slow_region);
3312
3313 // (b) Interrupt bit on TLS must be false.
3314 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3315 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3316 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3317
3318 // Set the control input on the field _interrupted read to prevent it floating up.
3319 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3320 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3321 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3322
3323 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3324
3325 // First fast path: if (!TLS._interrupted) return false;
3326 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3327 result_rgn->init_req(no_int_result_path, false_bit);
3328 result_val->init_req(no_int_result_path, intcon(0));
3329
3330 // drop through to next case
3331 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3332
3333 #ifndef _WINDOWS
3334 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3335 Node* clr_arg = argument(1);
3336 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3337 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3338 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3339
3340 // Second fast path: ... else if (!clear_int) return true;
3341 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3342 result_rgn->init_req(no_clear_result_path, false_arg);
3343 result_val->init_req(no_clear_result_path, intcon(1));
3344
3345 // drop through to next case
3346 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3347 #else
3348 // To return true on Windows you must read the _interrupted field
3349 // and check the event state i.e. take the slow path.
3350 #endif // _WINDOWS
3351
3352 // (d) Otherwise, go to the slow path.
3353 slow_region->add_req(control());
3354 set_control( _gvn.transform(slow_region));
3355
3356 if (stopped()) {
3357 // There is no slow path.
3358 result_rgn->init_req(slow_result_path, top());
3359 result_val->init_req(slow_result_path, top());
3360 } else {
3361 // non-virtual because it is a private non-static
3362 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3363
3364 Node* slow_val = set_results_for_java_call(slow_call);
3365 // this->control() comes from set_results_for_java_call
3366
3367 Node* fast_io = slow_call->in(TypeFunc::I_O);
3368 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3369
3370 // These two phis are pre-filled with copies of of the fast IO and Memory
3371 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3372 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3373
3374 result_rgn->init_req(slow_result_path, control());
3375 result_io ->init_req(slow_result_path, i_o());
3376 result_mem->init_req(slow_result_path, reset_memory());
3377 result_val->init_req(slow_result_path, slow_val);
3378
3379 set_all_memory(_gvn.transform(result_mem));
3380 set_i_o( _gvn.transform(result_io));
3381 }
3382
3383 C->set_has_split_ifs(true); // Has chance for split-if optimization
3384 set_result(result_rgn, result_val);
3385 return true;
3386 }
3387
3388 //---------------------------load_mirror_from_klass----------------------------
3389 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3390 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3391 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3392 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3393 }
3394
3395 //-----------------------load_klass_from_mirror_common-------------------------
3396 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3397 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3398 // and branch to the given path on the region.
3399 // If never_see_null, take an uncommon trap on null, so we can optimistically
3400 // compile for the non-null case.
3401 // If the region is NULL, force never_see_null = true.
3402 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3403 bool never_see_null,
3404 RegionNode* region,
3405 int null_path,
3406 int offset) {
3407 if (region == NULL) never_see_null = true;
3408 Node* p = basic_plus_adr(mirror, offset);
3409 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3410 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3411 Node* null_ctl = top();
3412 kls = null_check_oop(kls, &null_ctl, never_see_null);
3413 if (region != NULL) {
3414 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3415 region->init_req(null_path, null_ctl);
3416 } else {
3417 assert(null_ctl == top(), "no loose ends");
3418 }
3419 return kls;
3420 }
3421
3422 //--------------------(inline_native_Class_query helpers)---------------------
3423 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3424 // Fall through if (mods & mask) == bits, take the guard otherwise.
3425 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3426 // Branch around if the given klass has the given modifier bit set.
3427 // Like generate_guard, adds a new path onto the region.
3428 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3429 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3430 Node* mask = intcon(modifier_mask);
3431 Node* bits = intcon(modifier_bits);
3432 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3433 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3434 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3435 return generate_fair_guard(bol, region);
3436 }
3437 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3438 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3439 }
3440
3441 //-------------------------inline_native_Class_query-------------------
3442 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3443 const Type* return_type = TypeInt::BOOL;
3444 Node* prim_return_value = top(); // what happens if it's a primitive class?
3445 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3446 bool expect_prim = false; // most of these guys expect to work on refs
3447
3448 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3449
3450 Node* mirror = argument(0);
3451 Node* obj = top();
3452
3453 switch (id) {
3454 case vmIntrinsics::_isInstance:
3455 // nothing is an instance of a primitive type
3456 prim_return_value = intcon(0);
3457 obj = argument(1);
3458 break;
3459 case vmIntrinsics::_getModifiers:
3460 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3461 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3462 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3463 break;
3464 case vmIntrinsics::_isInterface:
3465 prim_return_value = intcon(0);
3466 break;
3467 case vmIntrinsics::_isArray:
3468 prim_return_value = intcon(0);
3469 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3470 break;
3471 case vmIntrinsics::_isPrimitive:
3472 prim_return_value = intcon(1);
3473 expect_prim = true; // obviously
3474 break;
3475 case vmIntrinsics::_getSuperclass:
3476 prim_return_value = null();
3477 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3478 break;
3479 case vmIntrinsics::_getClassAccessFlags:
3480 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3481 return_type = TypeInt::INT; // not bool! 6297094
3482 break;
3483 default:
3484 fatal_unexpected_iid(id);
3485 break;
3486 }
3487
3488 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3489 if (mirror_con == NULL) return false; // cannot happen?
3490
3491 #ifndef PRODUCT
3492 if (C->print_intrinsics() || C->print_inlining()) {
3493 ciType* k = mirror_con->java_mirror_type();
3494 if (k) {
3495 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3496 k->print_name();
3497 tty->cr();
3498 }
3499 }
3500 #endif
3501
3502 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3503 RegionNode* region = new RegionNode(PATH_LIMIT);
3504 record_for_igvn(region);
3505 PhiNode* phi = new PhiNode(region, return_type);
3506
3507 // The mirror will never be null of Reflection.getClassAccessFlags, however
3508 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3509 // if it is. See bug 4774291.
3510
3511 // For Reflection.getClassAccessFlags(), the null check occurs in
3512 // the wrong place; see inline_unsafe_access(), above, for a similar
3513 // situation.
3514 mirror = null_check(mirror);
3515 // If mirror or obj is dead, only null-path is taken.
3516 if (stopped()) return true;
3517
3518 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3519
3520 // Now load the mirror's klass metaobject, and null-check it.
3521 // Side-effects region with the control path if the klass is null.
3522 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3523 // If kls is null, we have a primitive mirror.
3524 phi->init_req(_prim_path, prim_return_value);
3525 if (stopped()) { set_result(region, phi); return true; }
3526 bool safe_for_replace = (region->in(_prim_path) == top());
3527
3528 Node* p; // handy temp
3529 Node* null_ctl;
3530
3531 // Now that we have the non-null klass, we can perform the real query.
3532 // For constant classes, the query will constant-fold in LoadNode::Value.
3533 Node* query_value = top();
3534 switch (id) {
3535 case vmIntrinsics::_isInstance:
3536 // nothing is an instance of a primitive type
3537 query_value = gen_instanceof(obj, kls, safe_for_replace);
3538 break;
3539
3540 case vmIntrinsics::_getModifiers:
3541 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3542 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3543 break;
3544
3545 case vmIntrinsics::_isInterface:
3546 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3547 if (generate_interface_guard(kls, region) != NULL)
3548 // A guard was added. If the guard is taken, it was an interface.
3549 phi->add_req(intcon(1));
3550 // If we fall through, it's a plain class.
3551 query_value = intcon(0);
3552 break;
3553
3554 case vmIntrinsics::_isArray:
3555 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3556 if (generate_array_guard(kls, region) != NULL)
3557 // A guard was added. If the guard is taken, it was an array.
3558 phi->add_req(intcon(1));
3559 // If we fall through, it's a plain class.
3560 query_value = intcon(0);
3561 break;
3562
3563 case vmIntrinsics::_isPrimitive:
3564 query_value = intcon(0); // "normal" path produces false
3565 break;
3566
3567 case vmIntrinsics::_getSuperclass:
3568 // The rules here are somewhat unfortunate, but we can still do better
3569 // with random logic than with a JNI call.
3570 // Interfaces store null or Object as _super, but must report null.
3571 // Arrays store an intermediate super as _super, but must report Object.
3572 // Other types can report the actual _super.
3573 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3574 if (generate_interface_guard(kls, region) != NULL)
3575 // A guard was added. If the guard is taken, it was an interface.
3576 phi->add_req(null());
3577 if (generate_array_guard(kls, region) != NULL)
3578 // A guard was added. If the guard is taken, it was an array.
3579 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3580 // If we fall through, it's a plain class. Get its _super.
3581 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3582 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3583 null_ctl = top();
3584 kls = null_check_oop(kls, &null_ctl);
3585 if (null_ctl != top()) {
3586 // If the guard is taken, Object.superClass is null (both klass and mirror).
3587 region->add_req(null_ctl);
3588 phi ->add_req(null());
3589 }
3590 if (!stopped()) {
3591 query_value = load_mirror_from_klass(kls);
3592 }
3593 break;
3594
3595 case vmIntrinsics::_getClassAccessFlags:
3596 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3597 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3598 break;
3599
3600 default:
3601 fatal_unexpected_iid(id);
3602 break;
3603 }
3604
3605 // Fall-through is the normal case of a query to a real class.
3606 phi->init_req(1, query_value);
3607 region->init_req(1, control());
3608
3609 C->set_has_split_ifs(true); // Has chance for split-if optimization
3610 set_result(region, phi);
3611 return true;
3612 }
3613
3614 //-------------------------inline_Class_cast-------------------
3615 bool LibraryCallKit::inline_Class_cast() {
3616 Node* mirror = argument(0); // Class
3617 Node* obj = argument(1);
3618 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3619 if (mirror_con == NULL) {
3620 return false; // dead path (mirror->is_top()).
3621 }
3622 if (obj == NULL || obj->is_top()) {
3623 return false; // dead path
3624 }
3625 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3626
3627 // First, see if Class.cast() can be folded statically.
3628 // java_mirror_type() returns non-null for compile-time Class constants.
3629 ciType* tm = mirror_con->java_mirror_type();
3630 if (tm != NULL && tm->is_klass() &&
3631 tp != NULL && tp->klass() != NULL) {
3632 if (!tp->klass()->is_loaded()) {
3633 // Don't use intrinsic when class is not loaded.
3634 return false;
3635 } else {
3636 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3637 if (static_res == Compile::SSC_always_true) {
3638 // isInstance() is true - fold the code.
3639 set_result(obj);
3640 return true;
3641 } else if (static_res == Compile::SSC_always_false) {
3642 // Don't use intrinsic, have to throw ClassCastException.
3643 // If the reference is null, the non-intrinsic bytecode will
3644 // be optimized appropriately.
3645 return false;
3646 }
3647 }
3648 }
3649
3650 // Bailout intrinsic and do normal inlining if exception path is frequent.
3651 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3652 return false;
3653 }
3654
3655 // Generate dynamic checks.
3656 // Class.cast() is java implementation of _checkcast bytecode.
3657 // Do checkcast (Parse::do_checkcast()) optimizations here.
3658
3659 mirror = null_check(mirror);
3660 // If mirror is dead, only null-path is taken.
3661 if (stopped()) {
3662 return true;
3663 }
3664
3665 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3666 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3667 RegionNode* region = new RegionNode(PATH_LIMIT);
3668 record_for_igvn(region);
3669
3670 // Now load the mirror's klass metaobject, and null-check it.
3671 // If kls is null, we have a primitive mirror and
3672 // nothing is an instance of a primitive type.
3673 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3674
3675 Node* res = top();
3676 if (!stopped()) {
3677 Node* bad_type_ctrl = top();
3678 // Do checkcast optimizations.
3679 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3680 region->init_req(_bad_type_path, bad_type_ctrl);
3681 }
3682 if (region->in(_prim_path) != top() ||
3683 region->in(_bad_type_path) != top()) {
3684 // Let Interpreter throw ClassCastException.
3685 PreserveJVMState pjvms(this);
3686 set_control(_gvn.transform(region));
3687 uncommon_trap(Deoptimization::Reason_intrinsic,
3688 Deoptimization::Action_maybe_recompile);
3689 }
3690 if (!stopped()) {
3691 set_result(res);
3692 }
3693 return true;
3694 }
3695
3696
3697 //--------------------------inline_native_subtype_check------------------------
3698 // This intrinsic takes the JNI calls out of the heart of
3699 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3700 bool LibraryCallKit::inline_native_subtype_check() {
3701 // Pull both arguments off the stack.
3702 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3703 args[0] = argument(0);
3704 args[1] = argument(1);
3705 Node* klasses[2]; // corresponding Klasses: superk, subk
3706 klasses[0] = klasses[1] = top();
3707
3708 enum {
3709 // A full decision tree on {superc is prim, subc is prim}:
3710 _prim_0_path = 1, // {P,N} => false
3711 // {P,P} & superc!=subc => false
3712 _prim_same_path, // {P,P} & superc==subc => true
3713 _prim_1_path, // {N,P} => false
3714 _ref_subtype_path, // {N,N} & subtype check wins => true
3715 _both_ref_path, // {N,N} & subtype check loses => false
3716 PATH_LIMIT
3717 };
3718
3719 RegionNode* region = new RegionNode(PATH_LIMIT);
3720 Node* phi = new PhiNode(region, TypeInt::BOOL);
3721 record_for_igvn(region);
3722
3723 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3724 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3725 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3726
3727 // First null-check both mirrors and load each mirror's klass metaobject.
3728 int which_arg;
3729 for (which_arg = 0; which_arg <= 1; which_arg++) {
3730 Node* arg = args[which_arg];
3731 arg = null_check(arg);
3732 if (stopped()) break;
3733 args[which_arg] = arg;
3734
3735 Node* p = basic_plus_adr(arg, class_klass_offset);
3736 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3737 klasses[which_arg] = _gvn.transform(kls);
3738 }
3739
3740 // Having loaded both klasses, test each for null.
3741 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3742 for (which_arg = 0; which_arg <= 1; which_arg++) {
3743 Node* kls = klasses[which_arg];
3744 Node* null_ctl = top();
3745 kls = null_check_oop(kls, &null_ctl, never_see_null);
3746 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3747 region->init_req(prim_path, null_ctl);
3748 if (stopped()) break;
3749 klasses[which_arg] = kls;
3750 }
3751
3752 if (!stopped()) {
3753 // now we have two reference types, in klasses[0..1]
3754 Node* subk = klasses[1]; // the argument to isAssignableFrom
3755 Node* superk = klasses[0]; // the receiver
3756 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3757 // now we have a successful reference subtype check
3758 region->set_req(_ref_subtype_path, control());
3759 }
3760
3761 // If both operands are primitive (both klasses null), then
3762 // we must return true when they are identical primitives.
3763 // It is convenient to test this after the first null klass check.
3764 set_control(region->in(_prim_0_path)); // go back to first null check
3765 if (!stopped()) {
3766 // Since superc is primitive, make a guard for the superc==subc case.
3767 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3768 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3769 generate_guard(bol_eq, region, PROB_FAIR);
3770 if (region->req() == PATH_LIMIT+1) {
3771 // A guard was added. If the added guard is taken, superc==subc.
3772 region->swap_edges(PATH_LIMIT, _prim_same_path);
3773 region->del_req(PATH_LIMIT);
3774 }
3775 region->set_req(_prim_0_path, control()); // Not equal after all.
3776 }
3777
3778 // these are the only paths that produce 'true':
3779 phi->set_req(_prim_same_path, intcon(1));
3780 phi->set_req(_ref_subtype_path, intcon(1));
3781
3782 // pull together the cases:
3783 assert(region->req() == PATH_LIMIT, "sane region");
3784 for (uint i = 1; i < region->req(); i++) {
3785 Node* ctl = region->in(i);
3786 if (ctl == NULL || ctl == top()) {
3787 region->set_req(i, top());
3788 phi ->set_req(i, top());
3789 } else if (phi->in(i) == NULL) {
3790 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3791 }
3792 }
3793
3794 set_control(_gvn.transform(region));
3795 set_result(_gvn.transform(phi));
3796 return true;
3797 }
3798
3799 //---------------------generate_array_guard_common------------------------
3800 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3801 bool obj_array, bool not_array) {
3802
3803 if (stopped()) {
3804 return NULL;
3805 }
3806
3807 // If obj_array/non_array==false/false:
3808 // Branch around if the given klass is in fact an array (either obj or prim).
3809 // If obj_array/non_array==false/true:
3810 // Branch around if the given klass is not an array klass of any kind.
3811 // If obj_array/non_array==true/true:
3812 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3813 // If obj_array/non_array==true/false:
3814 // Branch around if the kls is an oop array (Object[] or subtype)
3815 //
3816 // Like generate_guard, adds a new path onto the region.
3817 jint layout_con = 0;
3818 Node* layout_val = get_layout_helper(kls, layout_con);
3819 if (layout_val == NULL) {
3820 bool query = (obj_array
3821 ? Klass::layout_helper_is_objArray(layout_con)
3822 : Klass::layout_helper_is_array(layout_con));
3823 if (query == not_array) {
3824 return NULL; // never a branch
3825 } else { // always a branch
3826 Node* always_branch = control();
3827 if (region != NULL)
3828 region->add_req(always_branch);
3829 set_control(top());
3830 return always_branch;
3831 }
3832 }
3833 // Now test the correct condition.
3834 jint nval = (obj_array
3835 ? (jint)(Klass::_lh_array_tag_type_value
3836 << Klass::_lh_array_tag_shift)
3837 : Klass::_lh_neutral_value);
3838 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3839 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3840 // invert the test if we are looking for a non-array
3841 if (not_array) btest = BoolTest(btest).negate();
3842 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3843 return generate_fair_guard(bol, region);
3844 }
3845
3846
3847 //-----------------------inline_native_newArray--------------------------
3848 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3849 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3850 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3851 Node* mirror;
3852 Node* count_val;
3853 if (uninitialized) {
3854 mirror = argument(1);
3855 count_val = argument(2);
3856 } else {
3857 mirror = argument(0);
3858 count_val = argument(1);
3859 }
3860
3861 mirror = null_check(mirror);
3862 // If mirror or obj is dead, only null-path is taken.
3863 if (stopped()) return true;
3864
3865 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3866 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3867 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3868 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3869 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3870
3871 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3872 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3873 result_reg, _slow_path);
3874 Node* normal_ctl = control();
3875 Node* no_array_ctl = result_reg->in(_slow_path);
3876
3877 // Generate code for the slow case. We make a call to newArray().
3878 set_control(no_array_ctl);
3879 if (!stopped()) {
3880 // Either the input type is void.class, or else the
3881 // array klass has not yet been cached. Either the
3882 // ensuing call will throw an exception, or else it
3883 // will cache the array klass for next time.
3884 PreserveJVMState pjvms(this);
3885 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3886 Node* slow_result = set_results_for_java_call(slow_call);
3887 // this->control() comes from set_results_for_java_call
3888 result_reg->set_req(_slow_path, control());
3889 result_val->set_req(_slow_path, slow_result);
3890 result_io ->set_req(_slow_path, i_o());
3891 result_mem->set_req(_slow_path, reset_memory());
3892 }
3893
3894 set_control(normal_ctl);
3895 if (!stopped()) {
3896 // Normal case: The array type has been cached in the java.lang.Class.
3897 // The following call works fine even if the array type is polymorphic.
3898 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3899 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3900 result_reg->init_req(_normal_path, control());
3901 result_val->init_req(_normal_path, obj);
3902 result_io ->init_req(_normal_path, i_o());
3903 result_mem->init_req(_normal_path, reset_memory());
3904
3905 if (uninitialized) {
3906 // Mark the allocation so that zeroing is skipped
3907 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3908 alloc->maybe_set_complete(&_gvn);
3909 }
3910 }
3911
3912 // Return the combined state.
3913 set_i_o( _gvn.transform(result_io) );
3914 set_all_memory( _gvn.transform(result_mem));
3915
3916 C->set_has_split_ifs(true); // Has chance for split-if optimization
3917 set_result(result_reg, result_val);
3918 return true;
3919 }
3920
3921 //----------------------inline_native_getLength--------------------------
3922 // public static native int java.lang.reflect.Array.getLength(Object array);
3923 bool LibraryCallKit::inline_native_getLength() {
3924 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3925
3926 Node* array = null_check(argument(0));
3927 // If array is dead, only null-path is taken.
3928 if (stopped()) return true;
3929
3930 // Deoptimize if it is a non-array.
3931 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3932
3933 if (non_array != NULL) {
3934 PreserveJVMState pjvms(this);
3935 set_control(non_array);
3936 uncommon_trap(Deoptimization::Reason_intrinsic,
3937 Deoptimization::Action_maybe_recompile);
3938 }
3939
3940 // If control is dead, only non-array-path is taken.
3941 if (stopped()) return true;
3942
3943 // The works fine even if the array type is polymorphic.
3944 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3945 Node* result = load_array_length(array);
3946
3947 C->set_has_split_ifs(true); // Has chance for split-if optimization
3948 set_result(result);
3949 return true;
3950 }
3951
3952 //------------------------inline_array_copyOf----------------------------
3953 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3954 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3955 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3956 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3957
3958 // Get the arguments.
3959 Node* original = argument(0);
3960 Node* start = is_copyOfRange? argument(1): intcon(0);
3961 Node* end = is_copyOfRange? argument(2): argument(1);
3962 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3963
3964 Node* newcopy = NULL;
3965
3966 // Set the original stack and the reexecute bit for the interpreter to reexecute
3967 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3968 { PreserveReexecuteState preexecs(this);
3969 jvms()->set_should_reexecute(true);
3970
3971 array_type_mirror = null_check(array_type_mirror);
3972 original = null_check(original);
3973
3974 // Check if a null path was taken unconditionally.
3975 if (stopped()) return true;
3976
3977 Node* orig_length = load_array_length(original);
3978
3979 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3980 klass_node = null_check(klass_node);
3981
3982 RegionNode* bailout = new RegionNode(1);
3983 record_for_igvn(bailout);
3984
3985 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3986 // Bail out if that is so.
3987 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3988 if (not_objArray != NULL) {
3989 // Improve the klass node's type from the new optimistic assumption:
3990 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3991 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3992 Node* cast = new CastPPNode(klass_node, akls);
3993 cast->init_req(0, control());
3994 klass_node = _gvn.transform(cast);
3995 }
3996
3997 // Bail out if either start or end is negative.
3998 generate_negative_guard(start, bailout, &start);
3999 generate_negative_guard(end, bailout, &end);
4000
4001 Node* length = end;
4002 if (_gvn.type(start) != TypeInt::ZERO) {
4003 length = _gvn.transform(new SubINode(end, start));
4004 }
4005
4006 // Bail out if length is negative.
4007 // Without this the new_array would throw
4008 // NegativeArraySizeException but IllegalArgumentException is what
4009 // should be thrown
4010 generate_negative_guard(length, bailout, &length);
4011
4012 if (bailout->req() > 1) {
4013 PreserveJVMState pjvms(this);
4014 set_control(_gvn.transform(bailout));
4015 uncommon_trap(Deoptimization::Reason_intrinsic,
4016 Deoptimization::Action_maybe_recompile);
4017 }
4018
4019 if (!stopped()) {
4020 // How many elements will we copy from the original?
4021 // The answer is MinI(orig_length - start, length).
4022 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4023 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4024
4025 // Generate a direct call to the right arraycopy function(s).
4026 // We know the copy is disjoint but we might not know if the
4027 // oop stores need checking.
4028 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4029 // This will fail a store-check if x contains any non-nulls.
4030
4031 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4032 // loads/stores but it is legal only if we're sure the
4033 // Arrays.copyOf would succeed. So we need all input arguments
4034 // to the copyOf to be validated, including that the copy to the
4035 // new array won't trigger an ArrayStoreException. That subtype
4036 // check can be optimized if we know something on the type of
4037 // the input array from type speculation.
4038 if (_gvn.type(klass_node)->singleton()) {
4039 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4040 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4041
4042 int test = C->static_subtype_check(superk, subk);
4043 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4044 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4045 if (t_original->speculative_type() != NULL) {
4046 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4047 }
4048 }
4049 }
4050
4051 bool validated = false;
4052 // Reason_class_check rather than Reason_intrinsic because we
4053 // want to intrinsify even if this traps.
4054 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4055 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4056 klass_node);
4057
4058 if (not_subtype_ctrl != top()) {
4059 PreserveJVMState pjvms(this);
4060 set_control(not_subtype_ctrl);
4061 uncommon_trap(Deoptimization::Reason_class_check,
4062 Deoptimization::Action_make_not_entrant);
4063 assert(stopped(), "Should be stopped");
4064 }
4065 validated = true;
4066 }
4067
4068 if (!stopped()) {
4069 newcopy = new_array(klass_node, length, 0); // no arguments to push
4070
4071 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4072 load_object_klass(original), klass_node);
4073 if (!is_copyOfRange) {
4074 ac->set_copyof(validated);
4075 } else {
4076 ac->set_copyofrange(validated);
4077 }
4078 Node* n = _gvn.transform(ac);
4079 if (n == ac) {
4080 ac->connect_outputs(this);
4081 } else {
4082 assert(validated, "shouldn't transform if all arguments not validated");
4083 set_all_memory(n);
4084 }
4085 }
4086 }
4087 } // original reexecute is set back here
4088
4089 C->set_has_split_ifs(true); // Has chance for split-if optimization
4090 if (!stopped()) {
4091 set_result(newcopy);
4092 }
4093 return true;
4094 }
4095
4096
4097 //----------------------generate_virtual_guard---------------------------
4098 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4099 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4100 RegionNode* slow_region) {
4101 ciMethod* method = callee();
4102 int vtable_index = method->vtable_index();
4103 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4104 "bad index %d", vtable_index);
4105 // Get the Method* out of the appropriate vtable entry.
4106 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4107 vtable_index*vtableEntry::size_in_bytes() +
4108 vtableEntry::method_offset_in_bytes();
4109 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4110 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4111
4112 // Compare the target method with the expected method (e.g., Object.hashCode).
4113 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4114
4115 Node* native_call = makecon(native_call_addr);
4116 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4117 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4118
4119 return generate_slow_guard(test_native, slow_region);
4120 }
4121
4122 //-----------------------generate_method_call----------------------------
4123 // Use generate_method_call to make a slow-call to the real
4124 // method if the fast path fails. An alternative would be to
4125 // use a stub like OptoRuntime::slow_arraycopy_Java.
4126 // This only works for expanding the current library call,
4127 // not another intrinsic. (E.g., don't use this for making an
4128 // arraycopy call inside of the copyOf intrinsic.)
4129 CallJavaNode*
4130 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4131 // When compiling the intrinsic method itself, do not use this technique.
4132 guarantee(callee() != C->method(), "cannot make slow-call to self");
4133
4134 ciMethod* method = callee();
4135 // ensure the JVMS we have will be correct for this call
4136 guarantee(method_id == method->intrinsic_id(), "must match");
4137
4138 const TypeFunc* tf = TypeFunc::make(method);
4139 CallJavaNode* slow_call;
4140 if (is_static) {
4141 assert(!is_virtual, "");
4142 slow_call = new CallStaticJavaNode(C, tf,
4143 SharedRuntime::get_resolve_static_call_stub(),
4144 method, bci());
4145 } else if (is_virtual) {
4146 null_check_receiver();
4147 int vtable_index = Method::invalid_vtable_index;
4148 if (UseInlineCaches) {
4149 // Suppress the vtable call
4150 } else {
4151 // hashCode and clone are not a miranda methods,
4152 // so the vtable index is fixed.
4153 // No need to use the linkResolver to get it.
4154 vtable_index = method->vtable_index();
4155 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4156 "bad index %d", vtable_index);
4157 }
4158 slow_call = new CallDynamicJavaNode(tf,
4159 SharedRuntime::get_resolve_virtual_call_stub(),
4160 method, vtable_index, bci());
4161 } else { // neither virtual nor static: opt_virtual
4162 null_check_receiver();
4163 slow_call = new CallStaticJavaNode(C, tf,
4164 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4165 method, bci());
4166 slow_call->set_optimized_virtual(true);
4167 }
4168 set_arguments_for_java_call(slow_call);
4169 set_edges_for_java_call(slow_call);
4170 return slow_call;
4171 }
4172
4173
4174 /**
4175 * Build special case code for calls to hashCode on an object. This call may
4176 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4177 * slightly different code.
4178 */
4179 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4180 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4181 assert(!(is_virtual && is_static), "either virtual, special, or static");
4182
4183 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4184
4185 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4186 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4187 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4188 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4189 Node* obj = NULL;
4190 if (!is_static) {
4191 // Check for hashing null object
4192 obj = null_check_receiver();
4193 if (stopped()) return true; // unconditionally null
4194 result_reg->init_req(_null_path, top());
4195 result_val->init_req(_null_path, top());
4196 } else {
4197 // Do a null check, and return zero if null.
4198 // System.identityHashCode(null) == 0
4199 obj = argument(0);
4200 Node* null_ctl = top();
4201 obj = null_check_oop(obj, &null_ctl);
4202 result_reg->init_req(_null_path, null_ctl);
4203 result_val->init_req(_null_path, _gvn.intcon(0));
4204 }
4205
4206 // Unconditionally null? Then return right away.
4207 if (stopped()) {
4208 set_control( result_reg->in(_null_path));
4209 if (!stopped())
4210 set_result(result_val->in(_null_path));
4211 return true;
4212 }
4213
4214 // We only go to the fast case code if we pass a number of guards. The
4215 // paths which do not pass are accumulated in the slow_region.
4216 RegionNode* slow_region = new RegionNode(1);
4217 record_for_igvn(slow_region);
4218
4219 // If this is a virtual call, we generate a funny guard. We pull out
4220 // the vtable entry corresponding to hashCode() from the target object.
4221 // If the target method which we are calling happens to be the native
4222 // Object hashCode() method, we pass the guard. We do not need this
4223 // guard for non-virtual calls -- the caller is known to be the native
4224 // Object hashCode().
4225 if (is_virtual) {
4226 // After null check, get the object's klass.
4227 Node* obj_klass = load_object_klass(obj);
4228 generate_virtual_guard(obj_klass, slow_region);
4229 }
4230
4231 // Get the header out of the object, use LoadMarkNode when available
4232 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4233 // The control of the load must be NULL. Otherwise, the load can move before
4234 // the null check after castPP removal.
4235 Node* no_ctrl = NULL;
4236 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4237
4238 // Test the header to see if it is unlocked.
4239 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4240 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4241 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4242 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4243 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4244
4245 generate_slow_guard(test_unlocked, slow_region);
4246
4247 // Get the hash value and check to see that it has been properly assigned.
4248 // We depend on hash_mask being at most 32 bits and avoid the use of
4249 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4250 // vm: see markOop.hpp.
4251 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4252 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4253 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4254 // This hack lets the hash bits live anywhere in the mark object now, as long
4255 // as the shift drops the relevant bits into the low 32 bits. Note that
4256 // Java spec says that HashCode is an int so there's no point in capturing
4257 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4258 hshifted_header = ConvX2I(hshifted_header);
4259 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4260
4261 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4262 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4263 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4264
4265 generate_slow_guard(test_assigned, slow_region);
4266
4267 Node* init_mem = reset_memory();
4268 // fill in the rest of the null path:
4269 result_io ->init_req(_null_path, i_o());
4270 result_mem->init_req(_null_path, init_mem);
4271
4272 result_val->init_req(_fast_path, hash_val);
4273 result_reg->init_req(_fast_path, control());
4274 result_io ->init_req(_fast_path, i_o());
4275 result_mem->init_req(_fast_path, init_mem);
4276
4277 // Generate code for the slow case. We make a call to hashCode().
4278 set_control(_gvn.transform(slow_region));
4279 if (!stopped()) {
4280 // No need for PreserveJVMState, because we're using up the present state.
4281 set_all_memory(init_mem);
4282 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4283 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4284 Node* slow_result = set_results_for_java_call(slow_call);
4285 // this->control() comes from set_results_for_java_call
4286 result_reg->init_req(_slow_path, control());
4287 result_val->init_req(_slow_path, slow_result);
4288 result_io ->set_req(_slow_path, i_o());
4289 result_mem ->set_req(_slow_path, reset_memory());
4290 }
4291
4292 // Return the combined state.
4293 set_i_o( _gvn.transform(result_io) );
4294 set_all_memory( _gvn.transform(result_mem));
4295
4296 set_result(result_reg, result_val);
4297 return true;
4298 }
4299
4300 //---------------------------inline_native_getClass----------------------------
4301 // public final native Class<?> java.lang.Object.getClass();
4302 //
4303 // Build special case code for calls to getClass on an object.
4304 bool LibraryCallKit::inline_native_getClass() {
4305 Node* obj = null_check_receiver();
4306 if (stopped()) return true;
4307 set_result(load_mirror_from_klass(load_object_klass(obj)));
4308 return true;
4309 }
4310
4311 //-----------------inline_native_Reflection_getCallerClass---------------------
4312 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4313 //
4314 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4315 //
4316 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4317 // in that it must skip particular security frames and checks for
4318 // caller sensitive methods.
4319 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4320 #ifndef PRODUCT
4321 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4322 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4323 }
4324 #endif
4325
4326 if (!jvms()->has_method()) {
4327 #ifndef PRODUCT
4328 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4329 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4330 }
4331 #endif
4332 return false;
4333 }
4334
4335 // Walk back up the JVM state to find the caller at the required
4336 // depth.
4337 JVMState* caller_jvms = jvms();
4338
4339 // Cf. JVM_GetCallerClass
4340 // NOTE: Start the loop at depth 1 because the current JVM state does
4341 // not include the Reflection.getCallerClass() frame.
4342 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4343 ciMethod* m = caller_jvms->method();
4344 switch (n) {
4345 case 0:
4346 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4347 break;
4348 case 1:
4349 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4350 if (!m->caller_sensitive()) {
4351 #ifndef PRODUCT
4352 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4353 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4354 }
4355 #endif
4356 return false; // bail-out; let JVM_GetCallerClass do the work
4357 }
4358 break;
4359 default:
4360 if (!m->is_ignored_by_security_stack_walk()) {
4361 // We have reached the desired frame; return the holder class.
4362 // Acquire method holder as java.lang.Class and push as constant.
4363 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4364 ciInstance* caller_mirror = caller_klass->java_mirror();
4365 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4366
4367 #ifndef PRODUCT
4368 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4369 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());
4370 tty->print_cr(" JVM state at this point:");
4371 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4372 ciMethod* m = jvms()->of_depth(i)->method();
4373 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4374 }
4375 }
4376 #endif
4377 return true;
4378 }
4379 break;
4380 }
4381 }
4382
4383 #ifndef PRODUCT
4384 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4385 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4386 tty->print_cr(" JVM state at this point:");
4387 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4388 ciMethod* m = jvms()->of_depth(i)->method();
4389 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4390 }
4391 }
4392 #endif
4393
4394 return false; // bail-out; let JVM_GetCallerClass do the work
4395 }
4396
4397 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4398 Node* arg = argument(0);
4399 Node* result = NULL;
4400
4401 switch (id) {
4402 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4403 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4404 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4405 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4406
4407 case vmIntrinsics::_doubleToLongBits: {
4408 // two paths (plus control) merge in a wood
4409 RegionNode *r = new RegionNode(3);
4410 Node *phi = new PhiNode(r, TypeLong::LONG);
4411
4412 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4413 // Build the boolean node
4414 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4415
4416 // Branch either way.
4417 // NaN case is less traveled, which makes all the difference.
4418 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4419 Node *opt_isnan = _gvn.transform(ifisnan);
4420 assert( opt_isnan->is_If(), "Expect an IfNode");
4421 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4422 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4423
4424 set_control(iftrue);
4425
4426 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4427 Node *slow_result = longcon(nan_bits); // return NaN
4428 phi->init_req(1, _gvn.transform( slow_result ));
4429 r->init_req(1, iftrue);
4430
4431 // Else fall through
4432 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4433 set_control(iffalse);
4434
4435 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4436 r->init_req(2, iffalse);
4437
4438 // Post merge
4439 set_control(_gvn.transform(r));
4440 record_for_igvn(r);
4441
4442 C->set_has_split_ifs(true); // Has chance for split-if optimization
4443 result = phi;
4444 assert(result->bottom_type()->isa_long(), "must be");
4445 break;
4446 }
4447
4448 case vmIntrinsics::_floatToIntBits: {
4449 // two paths (plus control) merge in a wood
4450 RegionNode *r = new RegionNode(3);
4451 Node *phi = new PhiNode(r, TypeInt::INT);
4452
4453 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4454 // Build the boolean node
4455 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4456
4457 // Branch either way.
4458 // NaN case is less traveled, which makes all the difference.
4459 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4460 Node *opt_isnan = _gvn.transform(ifisnan);
4461 assert( opt_isnan->is_If(), "Expect an IfNode");
4462 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4463 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4464
4465 set_control(iftrue);
4466
4467 static const jint nan_bits = 0x7fc00000;
4468 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4469 phi->init_req(1, _gvn.transform( slow_result ));
4470 r->init_req(1, iftrue);
4471
4472 // Else fall through
4473 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4474 set_control(iffalse);
4475
4476 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4477 r->init_req(2, iffalse);
4478
4479 // Post merge
4480 set_control(_gvn.transform(r));
4481 record_for_igvn(r);
4482
4483 C->set_has_split_ifs(true); // Has chance for split-if optimization
4484 result = phi;
4485 assert(result->bottom_type()->isa_int(), "must be");
4486 break;
4487 }
4488
4489 default:
4490 fatal_unexpected_iid(id);
4491 break;
4492 }
4493 set_result(_gvn.transform(result));
4494 return true;
4495 }
4496
4497 //----------------------inline_unsafe_copyMemory-------------------------
4498 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4499 bool LibraryCallKit::inline_unsafe_copyMemory() {
4500 if (callee()->is_static()) return false; // caller must have the capability!
4501 null_check_receiver(); // null-check receiver
4502 if (stopped()) return true;
4503
4504 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4505
4506 Node* src_ptr = argument(1); // type: oop
4507 Node* src_off = ConvL2X(argument(2)); // type: long
4508 Node* dst_ptr = argument(4); // type: oop
4509 Node* dst_off = ConvL2X(argument(5)); // type: long
4510 Node* size = ConvL2X(argument(7)); // type: long
4511
4512 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4513 "fieldOffset must be byte-scaled");
4514
4515 Node* src = make_unsafe_address(src_ptr, src_off);
4516 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4517
4518 // Conservatively insert a memory barrier on all memory slices.
4519 // Do not let writes of the copy source or destination float below the copy.
4520 insert_mem_bar(Op_MemBarCPUOrder);
4521
4522 // Call it. Note that the length argument is not scaled.
4523 make_runtime_call(RC_LEAF|RC_NO_FP,
4524 OptoRuntime::fast_arraycopy_Type(),
4525 StubRoutines::unsafe_arraycopy(),
4526 "unsafe_arraycopy",
4527 TypeRawPtr::BOTTOM,
4528 src, dst, size XTOP);
4529
4530 // Do not let reads of the copy destination float above the copy.
4531 insert_mem_bar(Op_MemBarCPUOrder);
4532
4533 return true;
4534 }
4535
4536 //------------------------clone_coping-----------------------------------
4537 // Helper function for inline_native_clone.
4538 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4539 assert(obj_size != NULL, "");
4540 Node* raw_obj = alloc_obj->in(1);
4541 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4542
4543 AllocateNode* alloc = NULL;
4544 if (ReduceBulkZeroing) {
4545 // We will be completely responsible for initializing this object -
4546 // mark Initialize node as complete.
4547 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4548 // The object was just allocated - there should be no any stores!
4549 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4550 // Mark as complete_with_arraycopy so that on AllocateNode
4551 // expansion, we know this AllocateNode is initialized by an array
4552 // copy and a StoreStore barrier exists after the array copy.
4553 alloc->initialization()->set_complete_with_arraycopy();
4554 }
4555
4556 // Copy the fastest available way.
4557 // TODO: generate fields copies for small objects instead.
4558 Node* src = obj;
4559 Node* dest = alloc_obj;
4560 Node* size = _gvn.transform(obj_size);
4561
4562 // Exclude the header but include array length to copy by 8 bytes words.
4563 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4564 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4565 instanceOopDesc::base_offset_in_bytes();
4566 // base_off:
4567 // 8 - 32-bit VM
4568 // 12 - 64-bit VM, compressed klass
4569 // 16 - 64-bit VM, normal klass
4570 if (base_off % BytesPerLong != 0) {
4571 assert(UseCompressedClassPointers, "");
4572 if (is_array) {
4573 // Exclude length to copy by 8 bytes words.
4574 base_off += sizeof(int);
4575 } else {
4576 // Include klass to copy by 8 bytes words.
4577 base_off = instanceOopDesc::klass_offset_in_bytes();
4578 }
4579 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4580 }
4581 src = basic_plus_adr(src, base_off);
4582 dest = basic_plus_adr(dest, base_off);
4583
4584 // Compute the length also, if needed:
4585 Node* countx = size;
4586 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4587 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4588
4589 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4590
4591 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4592 ac->set_clonebasic();
4593 Node* n = _gvn.transform(ac);
4594 if (n == ac) {
4595 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4596 } else {
4597 set_all_memory(n);
4598 }
4599
4600 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4601 if (card_mark) {
4602 assert(!is_array, "");
4603 // Put in store barrier for any and all oops we are sticking
4604 // into this object. (We could avoid this if we could prove
4605 // that the object type contains no oop fields at all.)
4606 Node* no_particular_value = NULL;
4607 Node* no_particular_field = NULL;
4608 int raw_adr_idx = Compile::AliasIdxRaw;
4609 post_barrier(control(),
4610 memory(raw_adr_type),
4611 alloc_obj,
4612 no_particular_field,
4613 raw_adr_idx,
4614 no_particular_value,
4615 T_OBJECT,
4616 false);
4617 }
4618
4619 // Do not let reads from the cloned object float above the arraycopy.
4620 if (alloc != NULL) {
4621 // Do not let stores that initialize this object be reordered with
4622 // a subsequent store that would make this object accessible by
4623 // other threads.
4624 // Record what AllocateNode this StoreStore protects so that
4625 // escape analysis can go from the MemBarStoreStoreNode to the
4626 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4627 // based on the escape status of the AllocateNode.
4628 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4629 } else {
4630 insert_mem_bar(Op_MemBarCPUOrder);
4631 }
4632 }
4633
4634 //------------------------inline_native_clone----------------------------
4635 // protected native Object java.lang.Object.clone();
4636 //
4637 // Here are the simple edge cases:
4638 // null receiver => normal trap
4639 // virtual and clone was overridden => slow path to out-of-line clone
4640 // not cloneable or finalizer => slow path to out-of-line Object.clone
4641 //
4642 // The general case has two steps, allocation and copying.
4643 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4644 //
4645 // Copying also has two cases, oop arrays and everything else.
4646 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4647 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4648 //
4649 // These steps fold up nicely if and when the cloned object's klass
4650 // can be sharply typed as an object array, a type array, or an instance.
4651 //
4652 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4653 PhiNode* result_val;
4654
4655 // Set the reexecute bit for the interpreter to reexecute
4656 // the bytecode that invokes Object.clone if deoptimization happens.
4657 { PreserveReexecuteState preexecs(this);
4658 jvms()->set_should_reexecute(true);
4659
4660 Node* obj = null_check_receiver();
4661 if (stopped()) return true;
4662
4663 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4664
4665 // If we are going to clone an instance, we need its exact type to
4666 // know the number and types of fields to convert the clone to
4667 // loads/stores. Maybe a speculative type can help us.
4668 if (!obj_type->klass_is_exact() &&
4669 obj_type->speculative_type() != NULL &&
4670 obj_type->speculative_type()->is_instance_klass()) {
4671 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4672 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4673 !spec_ik->has_injected_fields()) {
4674 ciKlass* k = obj_type->klass();
4675 if (!k->is_instance_klass() ||
4676 k->as_instance_klass()->is_interface() ||
4677 k->as_instance_klass()->has_subklass()) {
4678 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4679 }
4680 }
4681 }
4682
4683 Node* obj_klass = load_object_klass(obj);
4684 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4685 const TypeOopPtr* toop = ((tklass != NULL)
4686 ? tklass->as_instance_type()
4687 : TypeInstPtr::NOTNULL);
4688
4689 // Conservatively insert a memory barrier on all memory slices.
4690 // Do not let writes into the original float below the clone.
4691 insert_mem_bar(Op_MemBarCPUOrder);
4692
4693 // paths into result_reg:
4694 enum {
4695 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4696 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4697 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4698 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4699 PATH_LIMIT
4700 };
4701 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4702 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4703 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4704 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4705 record_for_igvn(result_reg);
4706
4707 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4708 int raw_adr_idx = Compile::AliasIdxRaw;
4709
4710 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4711 if (array_ctl != NULL) {
4712 // It's an array.
4713 PreserveJVMState pjvms(this);
4714 set_control(array_ctl);
4715 Node* obj_length = load_array_length(obj);
4716 Node* obj_size = NULL;
4717 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4718
4719 if (!use_ReduceInitialCardMarks()) {
4720 // If it is an oop array, it requires very special treatment,
4721 // because card marking is required on each card of the array.
4722 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4723 if (is_obja != NULL) {
4724 PreserveJVMState pjvms2(this);
4725 set_control(is_obja);
4726 // Generate a direct call to the right arraycopy function(s).
4727 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4728 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4729 ac->set_cloneoop();
4730 Node* n = _gvn.transform(ac);
4731 assert(n == ac, "cannot disappear");
4732 ac->connect_outputs(this);
4733
4734 result_reg->init_req(_objArray_path, control());
4735 result_val->init_req(_objArray_path, alloc_obj);
4736 result_i_o ->set_req(_objArray_path, i_o());
4737 result_mem ->set_req(_objArray_path, reset_memory());
4738 }
4739 }
4740 // Otherwise, there are no card marks to worry about.
4741 // (We can dispense with card marks if we know the allocation
4742 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4743 // causes the non-eden paths to take compensating steps to
4744 // simulate a fresh allocation, so that no further
4745 // card marks are required in compiled code to initialize
4746 // the object.)
4747
4748 if (!stopped()) {
4749 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4750
4751 // Present the results of the copy.
4752 result_reg->init_req(_array_path, control());
4753 result_val->init_req(_array_path, alloc_obj);
4754 result_i_o ->set_req(_array_path, i_o());
4755 result_mem ->set_req(_array_path, reset_memory());
4756 }
4757 }
4758
4759 // We only go to the instance fast case code if we pass a number of guards.
4760 // The paths which do not pass are accumulated in the slow_region.
4761 RegionNode* slow_region = new RegionNode(1);
4762 record_for_igvn(slow_region);
4763 if (!stopped()) {
4764 // It's an instance (we did array above). Make the slow-path tests.
4765 // If this is a virtual call, we generate a funny guard. We grab
4766 // the vtable entry corresponding to clone() from the target object.
4767 // If the target method which we are calling happens to be the
4768 // Object clone() method, we pass the guard. We do not need this
4769 // guard for non-virtual calls; the caller is known to be the native
4770 // Object clone().
4771 if (is_virtual) {
4772 generate_virtual_guard(obj_klass, slow_region);
4773 }
4774
4775 // The object must be easily cloneable and must not have a finalizer.
4776 // Both of these conditions may be checked in a single test.
4777 // We could optimize the test further, but we don't care.
4778 generate_access_flags_guard(obj_klass,
4779 // Test both conditions:
4780 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4781 // Must be cloneable but not finalizer:
4782 JVM_ACC_IS_CLONEABLE_FAST,
4783 slow_region);
4784 }
4785
4786 if (!stopped()) {
4787 // It's an instance, and it passed the slow-path tests.
4788 PreserveJVMState pjvms(this);
4789 Node* obj_size = NULL;
4790 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4791 // is reexecuted if deoptimization occurs and there could be problems when merging
4792 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4793 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4794
4795 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4796
4797 // Present the results of the slow call.
4798 result_reg->init_req(_instance_path, control());
4799 result_val->init_req(_instance_path, alloc_obj);
4800 result_i_o ->set_req(_instance_path, i_o());
4801 result_mem ->set_req(_instance_path, reset_memory());
4802 }
4803
4804 // Generate code for the slow case. We make a call to clone().
4805 set_control(_gvn.transform(slow_region));
4806 if (!stopped()) {
4807 PreserveJVMState pjvms(this);
4808 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4809 Node* slow_result = set_results_for_java_call(slow_call);
4810 // this->control() comes from set_results_for_java_call
4811 result_reg->init_req(_slow_path, control());
4812 result_val->init_req(_slow_path, slow_result);
4813 result_i_o ->set_req(_slow_path, i_o());
4814 result_mem ->set_req(_slow_path, reset_memory());
4815 }
4816
4817 // Return the combined state.
4818 set_control( _gvn.transform(result_reg));
4819 set_i_o( _gvn.transform(result_i_o));
4820 set_all_memory( _gvn.transform(result_mem));
4821 } // original reexecute is set back here
4822
4823 set_result(_gvn.transform(result_val));
4824 return true;
4825 }
4826
4827 // If we have a tighly coupled allocation, the arraycopy may take care
4828 // of the array initialization. If one of the guards we insert between
4829 // the allocation and the arraycopy causes a deoptimization, an
4830 // unitialized array will escape the compiled method. To prevent that
4831 // we set the JVM state for uncommon traps between the allocation and
4832 // the arraycopy to the state before the allocation so, in case of
4833 // deoptimization, we'll reexecute the allocation and the
4834 // initialization.
4835 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4836 if (alloc != NULL) {
4837 ciMethod* trap_method = alloc->jvms()->method();
4838 int trap_bci = alloc->jvms()->bci();
4839
4840 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4841 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4842 // Make sure there's no store between the allocation and the
4843 // arraycopy otherwise visible side effects could be rexecuted
4844 // in case of deoptimization and cause incorrect execution.
4845 bool no_interfering_store = true;
4846 Node* mem = alloc->in(TypeFunc::Memory);
4847 if (mem->is_MergeMem()) {
4848 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4849 Node* n = mms.memory();
4850 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4851 assert(n->is_Store(), "what else?");
4852 no_interfering_store = false;
4853 break;
4854 }
4855 }
4856 } else {
4857 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4858 Node* n = mms.memory();
4859 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4860 assert(n->is_Store(), "what else?");
4861 no_interfering_store = false;
4862 break;
4863 }
4864 }
4865 }
4866
4867 if (no_interfering_store) {
4868 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4869 uint size = alloc->req();
4870 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4871 old_jvms->set_map(sfpt);
4872 for (uint i = 0; i < size; i++) {
4873 sfpt->init_req(i, alloc->in(i));
4874 }
4875 // re-push array length for deoptimization
4876 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4877 old_jvms->set_sp(old_jvms->sp()+1);
4878 old_jvms->set_monoff(old_jvms->monoff()+1);
4879 old_jvms->set_scloff(old_jvms->scloff()+1);
4880 old_jvms->set_endoff(old_jvms->endoff()+1);
4881 old_jvms->set_should_reexecute(true);
4882
4883 sfpt->set_i_o(map()->i_o());
4884 sfpt->set_memory(map()->memory());
4885 sfpt->set_control(map()->control());
4886
4887 JVMState* saved_jvms = jvms();
4888 saved_reexecute_sp = _reexecute_sp;
4889
4890 set_jvms(sfpt->jvms());
4891 _reexecute_sp = jvms()->sp();
4892
4893 return saved_jvms;
4894 }
4895 }
4896 }
4897 return NULL;
4898 }
4899
4900 // In case of a deoptimization, we restart execution at the
4901 // allocation, allocating a new array. We would leave an uninitialized
4902 // array in the heap that GCs wouldn't expect. Move the allocation
4903 // after the traps so we don't allocate the array if we
4904 // deoptimize. This is possible because tightly_coupled_allocation()
4905 // guarantees there's no observer of the allocated array at this point
4906 // and the control flow is simple enough.
4907 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4908 if (saved_jvms != NULL && !stopped()) {
4909 assert(alloc != NULL, "only with a tightly coupled allocation");
4910 // restore JVM state to the state at the arraycopy
4911 saved_jvms->map()->set_control(map()->control());
4912 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4913 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4914 // If we've improved the types of some nodes (null check) while
4915 // emitting the guards, propagate them to the current state
4916 map()->replaced_nodes().apply(saved_jvms->map());
4917 set_jvms(saved_jvms);
4918 _reexecute_sp = saved_reexecute_sp;
4919
4920 // Remove the allocation from above the guards
4921 CallProjections callprojs;
4922 alloc->extract_projections(&callprojs, true);
4923 InitializeNode* init = alloc->initialization();
4924 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4925 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4926 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4927 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4928
4929 // move the allocation here (after the guards)
4930 _gvn.hash_delete(alloc);
4931 alloc->set_req(TypeFunc::Control, control());
4932 alloc->set_req(TypeFunc::I_O, i_o());
4933 Node *mem = reset_memory();
4934 set_all_memory(mem);
4935 alloc->set_req(TypeFunc::Memory, mem);
4936 set_control(init->proj_out(TypeFunc::Control));
4937 set_i_o(callprojs.fallthrough_ioproj);
4938
4939 // Update memory as done in GraphKit::set_output_for_allocation()
4940 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4941 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4942 if (ary_type->isa_aryptr() && length_type != NULL) {
4943 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4944 }
4945 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4946 int elemidx = C->get_alias_index(telemref);
4947 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4948 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4949
4950 Node* allocx = _gvn.transform(alloc);
4951 assert(allocx == alloc, "where has the allocation gone?");
4952 assert(dest->is_CheckCastPP(), "not an allocation result?");
4953
4954 _gvn.hash_delete(dest);
4955 dest->set_req(0, control());
4956 Node* destx = _gvn.transform(dest);
4957 assert(destx == dest, "where has the allocation result gone?");
4958 }
4959 }
4960
4961
4962 //------------------------------inline_arraycopy-----------------------
4963 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4964 // Object dest, int destPos,
4965 // int length);
4966 bool LibraryCallKit::inline_arraycopy() {
4967 // Get the arguments.
4968 Node* src = argument(0); // type: oop
4969 Node* src_offset = argument(1); // type: int
4970 Node* dest = argument(2); // type: oop
4971 Node* dest_offset = argument(3); // type: int
4972 Node* length = argument(4); // type: int
4973
4974
4975 // Check for allocation before we add nodes that would confuse
4976 // tightly_coupled_allocation()
4977 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4978
4979 int saved_reexecute_sp = -1;
4980 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4981 // See arraycopy_restore_alloc_state() comment
4982 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4983 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4984 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4985 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4986
4987 // The following tests must be performed
4988 // (1) src and dest are arrays.
4989 // (2) src and dest arrays must have elements of the same BasicType
4990 // (3) src and dest must not be null.
4991 // (4) src_offset must not be negative.
4992 // (5) dest_offset must not be negative.
4993 // (6) length must not be negative.
4994 // (7) src_offset + length must not exceed length of src.
4995 // (8) dest_offset + length must not exceed length of dest.
4996 // (9) each element of an oop array must be assignable
4997
4998 // (3) src and dest must not be null.
4999 // always do this here because we need the JVM state for uncommon traps
5000 Node* null_ctl = top();
5001 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5002 assert(null_ctl->is_top(), "no null control here");
5003 dest = null_check(dest, T_ARRAY);
5004
5005 if (!can_emit_guards) {
5006 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5007 // guards but the arraycopy node could still take advantage of a
5008 // tightly allocated allocation. tightly_coupled_allocation() is
5009 // called again to make sure it takes the null check above into
5010 // account: the null check is mandatory and if it caused an
5011 // uncommon trap to be emitted then the allocation can't be
5012 // considered tightly coupled in this context.
5013 alloc = tightly_coupled_allocation(dest, NULL);
5014 }
5015
5016 bool validated = false;
5017
5018 const Type* src_type = _gvn.type(src);
5019 const Type* dest_type = _gvn.type(dest);
5020 const TypeAryPtr* top_src = src_type->isa_aryptr();
5021 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5022
5023 // Do we have the type of src?
5024 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5025 // Do we have the type of dest?
5026 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5027 // Is the type for src from speculation?
5028 bool src_spec = false;
5029 // Is the type for dest from speculation?
5030 bool dest_spec = false;
5031
5032 if ((!has_src || !has_dest) && can_emit_guards) {
5033 // We don't have sufficient type information, let's see if
5034 // speculative types can help. We need to have types for both src
5035 // and dest so that it pays off.
5036
5037 // Do we already have or could we have type information for src
5038 bool could_have_src = has_src;
5039 // Do we already have or could we have type information for dest
5040 bool could_have_dest = has_dest;
5041
5042 ciKlass* src_k = NULL;
5043 if (!has_src) {
5044 src_k = src_type->speculative_type_not_null();
5045 if (src_k != NULL && src_k->is_array_klass()) {
5046 could_have_src = true;
5047 }
5048 }
5049
5050 ciKlass* dest_k = NULL;
5051 if (!has_dest) {
5052 dest_k = dest_type->speculative_type_not_null();
5053 if (dest_k != NULL && dest_k->is_array_klass()) {
5054 could_have_dest = true;
5055 }
5056 }
5057
5058 if (could_have_src && could_have_dest) {
5059 // This is going to pay off so emit the required guards
5060 if (!has_src) {
5061 src = maybe_cast_profiled_obj(src, src_k, true);
5062 src_type = _gvn.type(src);
5063 top_src = src_type->isa_aryptr();
5064 has_src = (top_src != NULL && top_src->klass() != NULL);
5065 src_spec = true;
5066 }
5067 if (!has_dest) {
5068 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5069 dest_type = _gvn.type(dest);
5070 top_dest = dest_type->isa_aryptr();
5071 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5072 dest_spec = true;
5073 }
5074 }
5075 }
5076
5077 if (has_src && has_dest && can_emit_guards) {
5078 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5079 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5080 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5081 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5082
5083 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5084 // If both arrays are object arrays then having the exact types
5085 // for both will remove the need for a subtype check at runtime
5086 // before the call and may make it possible to pick a faster copy
5087 // routine (without a subtype check on every element)
5088 // Do we have the exact type of src?
5089 bool could_have_src = src_spec;
5090 // Do we have the exact type of dest?
5091 bool could_have_dest = dest_spec;
5092 ciKlass* src_k = top_src->klass();
5093 ciKlass* dest_k = top_dest->klass();
5094 if (!src_spec) {
5095 src_k = src_type->speculative_type_not_null();
5096 if (src_k != NULL && src_k->is_array_klass()) {
5097 could_have_src = true;
5098 }
5099 }
5100 if (!dest_spec) {
5101 dest_k = dest_type->speculative_type_not_null();
5102 if (dest_k != NULL && dest_k->is_array_klass()) {
5103 could_have_dest = true;
5104 }
5105 }
5106 if (could_have_src && could_have_dest) {
5107 // If we can have both exact types, emit the missing guards
5108 if (could_have_src && !src_spec) {
5109 src = maybe_cast_profiled_obj(src, src_k, true);
5110 }
5111 if (could_have_dest && !dest_spec) {
5112 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5113 }
5114 }
5115 }
5116 }
5117
5118 ciMethod* trap_method = method();
5119 int trap_bci = bci();
5120 if (saved_jvms != NULL) {
5121 trap_method = alloc->jvms()->method();
5122 trap_bci = alloc->jvms()->bci();
5123 }
5124
5125 bool negative_length_guard_generated = false;
5126
5127 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5128 can_emit_guards &&
5129 !src->is_top() && !dest->is_top()) {
5130 // validate arguments: enables transformation the ArrayCopyNode
5131 validated = true;
5132
5133 RegionNode* slow_region = new RegionNode(1);
5134 record_for_igvn(slow_region);
5135
5136 // (1) src and dest are arrays.
5137 generate_non_array_guard(load_object_klass(src), slow_region);
5138 generate_non_array_guard(load_object_klass(dest), slow_region);
5139
5140 // (2) src and dest arrays must have elements of the same BasicType
5141 // done at macro expansion or at Ideal transformation time
5142
5143 // (4) src_offset must not be negative.
5144 generate_negative_guard(src_offset, slow_region);
5145
5146 // (5) dest_offset must not be negative.
5147 generate_negative_guard(dest_offset, slow_region);
5148
5149 // (7) src_offset + length must not exceed length of src.
5150 generate_limit_guard(src_offset, length,
5151 load_array_length(src),
5152 slow_region);
5153
5154 // (8) dest_offset + length must not exceed length of dest.
5155 generate_limit_guard(dest_offset, length,
5156 load_array_length(dest),
5157 slow_region);
5158
5159 // (6) length must not be negative.
5160 // This is also checked in generate_arraycopy() during macro expansion, but
5161 // we also have to check it here for the case where the ArrayCopyNode will
5162 // be eliminated by Escape Analysis.
5163 if (EliminateAllocations) {
5164 generate_negative_guard(length, slow_region);
5165 negative_length_guard_generated = true;
5166 }
5167
5168 // (9) each element of an oop array must be assignable
5169 Node* src_klass = load_object_klass(src);
5170 Node* dest_klass = load_object_klass(dest);
5171 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5172
5173 if (not_subtype_ctrl != top()) {
5174 PreserveJVMState pjvms(this);
5175 set_control(not_subtype_ctrl);
5176 uncommon_trap(Deoptimization::Reason_intrinsic,
5177 Deoptimization::Action_make_not_entrant);
5178 assert(stopped(), "Should be stopped");
5179 }
5180 {
5181 PreserveJVMState pjvms(this);
5182 set_control(_gvn.transform(slow_region));
5183 uncommon_trap(Deoptimization::Reason_intrinsic,
5184 Deoptimization::Action_make_not_entrant);
5185 assert(stopped(), "Should be stopped");
5186 }
5187 }
5188
5189 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5190
5191 if (stopped()) {
5192 return true;
5193 }
5194
5195 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5196 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5197 // so the compiler has a chance to eliminate them: during macro expansion,
5198 // we have to set their control (CastPP nodes are eliminated).
5199 load_object_klass(src), load_object_klass(dest),
5200 load_array_length(src), load_array_length(dest));
5201
5202 ac->set_arraycopy(validated);
5203
5204 Node* n = _gvn.transform(ac);
5205 if (n == ac) {
5206 ac->connect_outputs(this);
5207 } else {
5208 assert(validated, "shouldn't transform if all arguments not validated");
5209 set_all_memory(n);
5210 }
5211
5212 return true;
5213 }
5214
5215
5216 // Helper function which determines if an arraycopy immediately follows
5217 // an allocation, with no intervening tests or other escapes for the object.
5218 AllocateArrayNode*
5219 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5220 RegionNode* slow_region) {
5221 if (stopped()) return NULL; // no fast path
5222 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5223
5224 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5225 if (alloc == NULL) return NULL;
5226
5227 Node* rawmem = memory(Compile::AliasIdxRaw);
5228 // Is the allocation's memory state untouched?
5229 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5230 // Bail out if there have been raw-memory effects since the allocation.
5231 // (Example: There might have been a call or safepoint.)
5232 return NULL;
5233 }
5234 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5235 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5236 return NULL;
5237 }
5238
5239 // There must be no unexpected observers of this allocation.
5240 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5241 Node* obs = ptr->fast_out(i);
5242 if (obs != this->map()) {
5243 return NULL;
5244 }
5245 }
5246
5247 // This arraycopy must unconditionally follow the allocation of the ptr.
5248 Node* alloc_ctl = ptr->in(0);
5249 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5250
5251 Node* ctl = control();
5252 while (ctl != alloc_ctl) {
5253 // There may be guards which feed into the slow_region.
5254 // Any other control flow means that we might not get a chance
5255 // to finish initializing the allocated object.
5256 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5257 IfNode* iff = ctl->in(0)->as_If();
5258 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5259 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5260 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5261 ctl = iff->in(0); // This test feeds the known slow_region.
5262 continue;
5263 }
5264 // One more try: Various low-level checks bottom out in
5265 // uncommon traps. If the debug-info of the trap omits
5266 // any reference to the allocation, as we've already
5267 // observed, then there can be no objection to the trap.
5268 bool found_trap = false;
5269 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5270 Node* obs = not_ctl->fast_out(j);
5271 if (obs->in(0) == not_ctl && obs->is_Call() &&
5272 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5273 found_trap = true; break;
5274 }
5275 }
5276 if (found_trap) {
5277 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5278 continue;
5279 }
5280 }
5281 return NULL;
5282 }
5283
5284 // If we get this far, we have an allocation which immediately
5285 // precedes the arraycopy, and we can take over zeroing the new object.
5286 // The arraycopy will finish the initialization, and provide
5287 // a new control state to which we will anchor the destination pointer.
5288
5289 return alloc;
5290 }
5291
5292 //-------------inline_encodeISOArray-----------------------------------
5293 // encode char[] to byte[] in ISO_8859_1
5294 bool LibraryCallKit::inline_encodeISOArray() {
5295 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5296 // no receiver since it is static method
5297 Node *src = argument(0);
5298 Node *src_offset = argument(1);
5299 Node *dst = argument(2);
5300 Node *dst_offset = argument(3);
5301 Node *length = argument(4);
5302
5303 const Type* src_type = src->Value(&_gvn);
5304 const Type* dst_type = dst->Value(&_gvn);
5305 const TypeAryPtr* top_src = src_type->isa_aryptr();
5306 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5307 if (top_src == NULL || top_src->klass() == NULL ||
5308 top_dest == NULL || top_dest->klass() == NULL) {
5309 // failed array check
5310 return false;
5311 }
5312
5313 // Figure out the size and type of the elements we will be copying.
5314 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5315 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5316 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5317 return false;
5318 }
5319
5320 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5321 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5322 // 'src_start' points to src array + scaled offset
5323 // 'dst_start' points to dst array + scaled offset
5324
5325 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5326 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5327 enc = _gvn.transform(enc);
5328 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5329 set_memory(res_mem, mtype);
5330 set_result(enc);
5331 return true;
5332 }
5333
5334 //-------------inline_multiplyToLen-----------------------------------
5335 bool LibraryCallKit::inline_multiplyToLen() {
5336 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5337
5338 address stubAddr = StubRoutines::multiplyToLen();
5339 if (stubAddr == NULL) {
5340 return false; // Intrinsic's stub is not implemented on this platform
5341 }
5342 const char* stubName = "multiplyToLen";
5343
5344 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5345
5346 // no receiver because it is a static method
5347 Node* x = argument(0);
5348 Node* xlen = argument(1);
5349 Node* y = argument(2);
5350 Node* ylen = argument(3);
5351 Node* z = argument(4);
5352
5353 const Type* x_type = x->Value(&_gvn);
5354 const Type* y_type = y->Value(&_gvn);
5355 const TypeAryPtr* top_x = x_type->isa_aryptr();
5356 const TypeAryPtr* top_y = y_type->isa_aryptr();
5357 if (top_x == NULL || top_x->klass() == NULL ||
5358 top_y == NULL || top_y->klass() == NULL) {
5359 // failed array check
5360 return false;
5361 }
5362
5363 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5364 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5365 if (x_elem != T_INT || y_elem != T_INT) {
5366 return false;
5367 }
5368
5369 // Set the original stack and the reexecute bit for the interpreter to reexecute
5370 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5371 // on the return from z array allocation in runtime.
5372 { PreserveReexecuteState preexecs(this);
5373 jvms()->set_should_reexecute(true);
5374
5375 Node* x_start = array_element_address(x, intcon(0), x_elem);
5376 Node* y_start = array_element_address(y, intcon(0), y_elem);
5377 // 'x_start' points to x array + scaled xlen
5378 // 'y_start' points to y array + scaled ylen
5379
5380 // Allocate the result array
5381 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5382 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5383 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5384
5385 IdealKit ideal(this);
5386
5387 #define __ ideal.
5388 Node* one = __ ConI(1);
5389 Node* zero = __ ConI(0);
5390 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5391 __ set(need_alloc, zero);
5392 __ set(z_alloc, z);
5393 __ if_then(z, BoolTest::eq, null()); {
5394 __ increment (need_alloc, one);
5395 } __ else_(); {
5396 // Update graphKit memory and control from IdealKit.
5397 sync_kit(ideal);
5398 Node* zlen_arg = load_array_length(z);
5399 // Update IdealKit memory and control from graphKit.
5400 __ sync_kit(this);
5401 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5402 __ increment (need_alloc, one);
5403 } __ end_if();
5404 } __ end_if();
5405
5406 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5407 // Update graphKit memory and control from IdealKit.
5408 sync_kit(ideal);
5409 Node * narr = new_array(klass_node, zlen, 1);
5410 // Update IdealKit memory and control from graphKit.
5411 __ sync_kit(this);
5412 __ set(z_alloc, narr);
5413 } __ end_if();
5414
5415 sync_kit(ideal);
5416 z = __ value(z_alloc);
5417 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5418 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5419 // Final sync IdealKit and GraphKit.
5420 final_sync(ideal);
5421 #undef __
5422
5423 Node* z_start = array_element_address(z, intcon(0), T_INT);
5424
5425 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5426 OptoRuntime::multiplyToLen_Type(),
5427 stubAddr, stubName, TypePtr::BOTTOM,
5428 x_start, xlen, y_start, ylen, z_start, zlen);
5429 } // original reexecute is set back here
5430
5431 C->set_has_split_ifs(true); // Has chance for split-if optimization
5432 set_result(z);
5433 return true;
5434 }
5435
5436 //-------------inline_squareToLen------------------------------------
5437 bool LibraryCallKit::inline_squareToLen() {
5438 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5439
5440 address stubAddr = StubRoutines::squareToLen();
5441 if (stubAddr == NULL) {
5442 return false; // Intrinsic's stub is not implemented on this platform
5443 }
5444 const char* stubName = "squareToLen";
5445
5446 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5447
5448 Node* x = argument(0);
5449 Node* len = argument(1);
5450 Node* z = argument(2);
5451 Node* zlen = argument(3);
5452
5453 const Type* x_type = x->Value(&_gvn);
5454 const Type* z_type = z->Value(&_gvn);
5455 const TypeAryPtr* top_x = x_type->isa_aryptr();
5456 const TypeAryPtr* top_z = z_type->isa_aryptr();
5457 if (top_x == NULL || top_x->klass() == NULL ||
5458 top_z == NULL || top_z->klass() == NULL) {
5459 // failed array check
5460 return false;
5461 }
5462
5463 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5464 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5465 if (x_elem != T_INT || z_elem != T_INT) {
5466 return false;
5467 }
5468
5469
5470 Node* x_start = array_element_address(x, intcon(0), x_elem);
5471 Node* z_start = array_element_address(z, intcon(0), z_elem);
5472
5473 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5474 OptoRuntime::squareToLen_Type(),
5475 stubAddr, stubName, TypePtr::BOTTOM,
5476 x_start, len, z_start, zlen);
5477
5478 set_result(z);
5479 return true;
5480 }
5481
5482 //-------------inline_mulAdd------------------------------------------
5483 bool LibraryCallKit::inline_mulAdd() {
5484 assert(UseMulAddIntrinsic, "not implemented on this platform");
5485
5486 address stubAddr = StubRoutines::mulAdd();
5487 if (stubAddr == NULL) {
5488 return false; // Intrinsic's stub is not implemented on this platform
5489 }
5490 const char* stubName = "mulAdd";
5491
5492 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5493
5494 Node* out = argument(0);
5495 Node* in = argument(1);
5496 Node* offset = argument(2);
5497 Node* len = argument(3);
5498 Node* k = argument(4);
5499
5500 const Type* out_type = out->Value(&_gvn);
5501 const Type* in_type = in->Value(&_gvn);
5502 const TypeAryPtr* top_out = out_type->isa_aryptr();
5503 const TypeAryPtr* top_in = in_type->isa_aryptr();
5504 if (top_out == NULL || top_out->klass() == NULL ||
5505 top_in == NULL || top_in->klass() == NULL) {
5506 // failed array check
5507 return false;
5508 }
5509
5510 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5511 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5512 if (out_elem != T_INT || in_elem != T_INT) {
5513 return false;
5514 }
5515
5516 Node* outlen = load_array_length(out);
5517 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5518 Node* out_start = array_element_address(out, intcon(0), out_elem);
5519 Node* in_start = array_element_address(in, intcon(0), in_elem);
5520
5521 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5522 OptoRuntime::mulAdd_Type(),
5523 stubAddr, stubName, TypePtr::BOTTOM,
5524 out_start,in_start, new_offset, len, k);
5525 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5526 set_result(result);
5527 return true;
5528 }
5529
5530 //-------------inline_montgomeryMultiply-----------------------------------
5531 bool LibraryCallKit::inline_montgomeryMultiply() {
5532 address stubAddr = StubRoutines::montgomeryMultiply();
5533 if (stubAddr == NULL) {
5534 return false; // Intrinsic's stub is not implemented on this platform
5535 }
5536
5537 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5538 const char* stubName = "montgomery_multiply";
5539
5540 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5541
5542 Node* a = argument(0);
5543 Node* b = argument(1);
5544 Node* n = argument(2);
5545 Node* len = argument(3);
5546 Node* inv = argument(4);
5547 Node* m = argument(6);
5548
5549 const Type* a_type = a->Value(&_gvn);
5550 const TypeAryPtr* top_a = a_type->isa_aryptr();
5551 const Type* b_type = b->Value(&_gvn);
5552 const TypeAryPtr* top_b = b_type->isa_aryptr();
5553 const Type* n_type = a->Value(&_gvn);
5554 const TypeAryPtr* top_n = n_type->isa_aryptr();
5555 const Type* m_type = a->Value(&_gvn);
5556 const TypeAryPtr* top_m = m_type->isa_aryptr();
5557 if (top_a == NULL || top_a->klass() == NULL ||
5558 top_b == NULL || top_b->klass() == NULL ||
5559 top_n == NULL || top_n->klass() == NULL ||
5560 top_m == NULL || top_m->klass() == NULL) {
5561 // failed array check
5562 return false;
5563 }
5564
5565 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5566 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5567 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5568 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5569 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5570 return false;
5571 }
5572
5573 // Make the call
5574 {
5575 Node* a_start = array_element_address(a, intcon(0), a_elem);
5576 Node* b_start = array_element_address(b, intcon(0), b_elem);
5577 Node* n_start = array_element_address(n, intcon(0), n_elem);
5578 Node* m_start = array_element_address(m, intcon(0), m_elem);
5579
5580 Node* call = make_runtime_call(RC_LEAF,
5581 OptoRuntime::montgomeryMultiply_Type(),
5582 stubAddr, stubName, TypePtr::BOTTOM,
5583 a_start, b_start, n_start, len, inv, top(),
5584 m_start);
5585 set_result(m);
5586 }
5587
5588 return true;
5589 }
5590
5591 bool LibraryCallKit::inline_montgomerySquare() {
5592 address stubAddr = StubRoutines::montgomerySquare();
5593 if (stubAddr == NULL) {
5594 return false; // Intrinsic's stub is not implemented on this platform
5595 }
5596
5597 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5598 const char* stubName = "montgomery_square";
5599
5600 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5601
5602 Node* a = argument(0);
5603 Node* n = argument(1);
5604 Node* len = argument(2);
5605 Node* inv = argument(3);
5606 Node* m = argument(5);
5607
5608 const Type* a_type = a->Value(&_gvn);
5609 const TypeAryPtr* top_a = a_type->isa_aryptr();
5610 const Type* n_type = a->Value(&_gvn);
5611 const TypeAryPtr* top_n = n_type->isa_aryptr();
5612 const Type* m_type = a->Value(&_gvn);
5613 const TypeAryPtr* top_m = m_type->isa_aryptr();
5614 if (top_a == NULL || top_a->klass() == NULL ||
5615 top_n == NULL || top_n->klass() == NULL ||
5616 top_m == NULL || top_m->klass() == NULL) {
5617 // failed array check
5618 return false;
5619 }
5620
5621 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5622 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5623 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5624 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5625 return false;
5626 }
5627
5628 // Make the call
5629 {
5630 Node* a_start = array_element_address(a, intcon(0), a_elem);
5631 Node* n_start = array_element_address(n, intcon(0), n_elem);
5632 Node* m_start = array_element_address(m, intcon(0), m_elem);
5633
5634 Node* call = make_runtime_call(RC_LEAF,
5635 OptoRuntime::montgomerySquare_Type(),
5636 stubAddr, stubName, TypePtr::BOTTOM,
5637 a_start, n_start, len, inv, top(),
5638 m_start);
5639 set_result(m);
5640 }
5641
5642 return true;
5643 }
5644
5645 //-------------inline_vectorizedMismatch------------------------------
5646 bool LibraryCallKit::inline_vectorizedMismatch() {
5647 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5648
5649 address stubAddr = StubRoutines::vectorizedMismatch();
5650 if (stubAddr == NULL) {
5651 return false; // Intrinsic's stub is not implemented on this platform
5652 }
5653 const char* stubName = "vectorizedMismatch";
5654 int size_l = callee()->signature()->size();
5655 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5656
5657 Node* obja = argument(0);
5658 Node* aoffset = argument(1);
5659 Node* objb = argument(3);
5660 Node* boffset = argument(4);
5661 Node* length = argument(6);
5662 Node* scale = argument(7);
5663
5664 const Type* a_type = obja->Value(&_gvn);
5665 const Type* b_type = objb->Value(&_gvn);
5666 const TypeAryPtr* top_a = a_type->isa_aryptr();
5667 const TypeAryPtr* top_b = b_type->isa_aryptr();
5668 if (top_a == NULL || top_a->klass() == NULL ||
5669 top_b == NULL || top_b->klass() == NULL) {
5670 // failed array check
5671 return false;
5672 }
5673
5674 Node* call;
5675 jvms()->set_should_reexecute(true);
5676
5677 Node* obja_adr = make_unsafe_address(obja, aoffset);
5678 Node* objb_adr = make_unsafe_address(objb, boffset);
5679
5680 call = make_runtime_call(RC_LEAF,
5681 OptoRuntime::vectorizedMismatch_Type(),
5682 stubAddr, stubName, TypePtr::BOTTOM,
5683 obja_adr, objb_adr, length, scale);
5684
5685 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5686 set_result(result);
5687 return true;
5688 }
5689
5690 /**
5691 * Calculate CRC32 for byte.
5692 * int java.util.zip.CRC32.update(int crc, int b)
5693 */
5694 bool LibraryCallKit::inline_updateCRC32() {
5695 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5696 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5697 // no receiver since it is static method
5698 Node* crc = argument(0); // type: int
5699 Node* b = argument(1); // type: int
5700
5701 /*
5702 * int c = ~ crc;
5703 * b = timesXtoThe32[(b ^ c) & 0xFF];
5704 * b = b ^ (c >>> 8);
5705 * crc = ~b;
5706 */
5707
5708 Node* M1 = intcon(-1);
5709 crc = _gvn.transform(new XorINode(crc, M1));
5710 Node* result = _gvn.transform(new XorINode(crc, b));
5711 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5712
5713 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5714 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5715 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5716 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5717
5718 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5719 result = _gvn.transform(new XorINode(crc, result));
5720 result = _gvn.transform(new XorINode(result, M1));
5721 set_result(result);
5722 return true;
5723 }
5724
5725 /**
5726 * Calculate CRC32 for byte[] array.
5727 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5728 */
5729 bool LibraryCallKit::inline_updateBytesCRC32() {
5730 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5731 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5732 // no receiver since it is static method
5733 Node* crc = argument(0); // type: int
5734 Node* src = argument(1); // type: oop
5735 Node* offset = argument(2); // type: int
5736 Node* length = argument(3); // type: int
5737
5738 const Type* src_type = src->Value(&_gvn);
5739 const TypeAryPtr* top_src = src_type->isa_aryptr();
5740 if (top_src == NULL || top_src->klass() == NULL) {
5741 // failed array check
5742 return false;
5743 }
5744
5745 // Figure out the size and type of the elements we will be copying.
5746 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5747 if (src_elem != T_BYTE) {
5748 return false;
5749 }
5750
5751 // 'src_start' points to src array + scaled offset
5752 Node* src_start = array_element_address(src, offset, src_elem);
5753
5754 // We assume that range check is done by caller.
5755 // TODO: generate range check (offset+length < src.length) in debug VM.
5756
5757 // Call the stub.
5758 address stubAddr = StubRoutines::updateBytesCRC32();
5759 const char *stubName = "updateBytesCRC32";
5760
5761 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5762 stubAddr, stubName, TypePtr::BOTTOM,
5763 crc, src_start, length);
5764 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5765 set_result(result);
5766 return true;
5767 }
5768
5769 /**
5770 * Calculate CRC32 for ByteBuffer.
5771 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5772 */
5773 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5774 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5775 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5776 // no receiver since it is static method
5777 Node* crc = argument(0); // type: int
5778 Node* src = argument(1); // type: long
5779 Node* offset = argument(3); // type: int
5780 Node* length = argument(4); // type: int
5781
5782 src = ConvL2X(src); // adjust Java long to machine word
5783 Node* base = _gvn.transform(new CastX2PNode(src));
5784 offset = ConvI2X(offset);
5785
5786 // 'src_start' points to src array + scaled offset
5787 Node* src_start = basic_plus_adr(top(), base, offset);
5788
5789 // Call the stub.
5790 address stubAddr = StubRoutines::updateBytesCRC32();
5791 const char *stubName = "updateBytesCRC32";
5792
5793 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5794 stubAddr, stubName, TypePtr::BOTTOM,
5795 crc, src_start, length);
5796 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5797 set_result(result);
5798 return true;
5799 }
5800
5801 //------------------------------get_table_from_crc32c_class-----------------------
5802 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5803 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5804 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5805
5806 return table;
5807 }
5808
5809 //------------------------------inline_updateBytesCRC32C-----------------------
5810 //
5811 // Calculate CRC32C for byte[] array.
5812 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5813 //
5814 bool LibraryCallKit::inline_updateBytesCRC32C() {
5815 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5816 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5817 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5818 // no receiver since it is a static method
5819 Node* crc = argument(0); // type: int
5820 Node* src = argument(1); // type: oop
5821 Node* offset = argument(2); // type: int
5822 Node* end = argument(3); // type: int
5823
5824 Node* length = _gvn.transform(new SubINode(end, offset));
5825
5826 const Type* src_type = src->Value(&_gvn);
5827 const TypeAryPtr* top_src = src_type->isa_aryptr();
5828 if (top_src == NULL || top_src->klass() == NULL) {
5829 // failed array check
5830 return false;
5831 }
5832
5833 // Figure out the size and type of the elements we will be copying.
5834 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5835 if (src_elem != T_BYTE) {
5836 return false;
5837 }
5838
5839 // 'src_start' points to src array + scaled offset
5840 Node* src_start = array_element_address(src, offset, src_elem);
5841
5842 // static final int[] byteTable in class CRC32C
5843 Node* table = get_table_from_crc32c_class(callee()->holder());
5844 Node* table_start = array_element_address(table, intcon(0), T_INT);
5845
5846 // We assume that range check is done by caller.
5847 // TODO: generate range check (offset+length < src.length) in debug VM.
5848
5849 // Call the stub.
5850 address stubAddr = StubRoutines::updateBytesCRC32C();
5851 const char *stubName = "updateBytesCRC32C";
5852
5853 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5854 stubAddr, stubName, TypePtr::BOTTOM,
5855 crc, src_start, length, table_start);
5856 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5857 set_result(result);
5858 return true;
5859 }
5860
5861 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5862 //
5863 // Calculate CRC32C for DirectByteBuffer.
5864 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5865 //
5866 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5867 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5868 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5869 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5870 // no receiver since it is a static method
5871 Node* crc = argument(0); // type: int
5872 Node* src = argument(1); // type: long
5873 Node* offset = argument(3); // type: int
5874 Node* end = argument(4); // type: int
5875
5876 Node* length = _gvn.transform(new SubINode(end, offset));
5877
5878 src = ConvL2X(src); // adjust Java long to machine word
5879 Node* base = _gvn.transform(new CastX2PNode(src));
5880 offset = ConvI2X(offset);
5881
5882 // 'src_start' points to src array + scaled offset
5883 Node* src_start = basic_plus_adr(top(), base, offset);
5884
5885 // static final int[] byteTable in class CRC32C
5886 Node* table = get_table_from_crc32c_class(callee()->holder());
5887 Node* table_start = array_element_address(table, intcon(0), T_INT);
5888
5889 // Call the stub.
5890 address stubAddr = StubRoutines::updateBytesCRC32C();
5891 const char *stubName = "updateBytesCRC32C";
5892
5893 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5894 stubAddr, stubName, TypePtr::BOTTOM,
5895 crc, src_start, length, table_start);
5896 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5897 set_result(result);
5898 return true;
5899 }
5900
5901 //------------------------------inline_updateBytesAdler32----------------------
5902 //
5903 // Calculate Adler32 checksum for byte[] array.
5904 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5905 //
5906 bool LibraryCallKit::inline_updateBytesAdler32() {
5907 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5908 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5909 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5910 // no receiver since it is static method
5911 Node* crc = argument(0); // type: int
5912 Node* src = argument(1); // type: oop
5913 Node* offset = argument(2); // type: int
5914 Node* length = argument(3); // type: int
5915
5916 const Type* src_type = src->Value(&_gvn);
5917 const TypeAryPtr* top_src = src_type->isa_aryptr();
5918 if (top_src == NULL || top_src->klass() == NULL) {
5919 // failed array check
5920 return false;
5921 }
5922
5923 // Figure out the size and type of the elements we will be copying.
5924 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5925 if (src_elem != T_BYTE) {
5926 return false;
5927 }
5928
5929 // 'src_start' points to src array + scaled offset
5930 Node* src_start = array_element_address(src, offset, src_elem);
5931
5932 // We assume that range check is done by caller.
5933 // TODO: generate range check (offset+length < src.length) in debug VM.
5934
5935 // Call the stub.
5936 address stubAddr = StubRoutines::updateBytesAdler32();
5937 const char *stubName = "updateBytesAdler32";
5938
5939 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5940 stubAddr, stubName, TypePtr::BOTTOM,
5941 crc, src_start, length);
5942 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5943 set_result(result);
5944 return true;
5945 }
5946
5947 //------------------------------inline_updateByteBufferAdler32---------------
5948 //
5949 // Calculate Adler32 checksum for DirectByteBuffer.
5950 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5951 //
5952 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5953 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5954 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5955 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5956 // no receiver since it is static method
5957 Node* crc = argument(0); // type: int
5958 Node* src = argument(1); // type: long
5959 Node* offset = argument(3); // type: int
5960 Node* length = argument(4); // type: int
5961
5962 src = ConvL2X(src); // adjust Java long to machine word
5963 Node* base = _gvn.transform(new CastX2PNode(src));
5964 offset = ConvI2X(offset);
5965
5966 // 'src_start' points to src array + scaled offset
5967 Node* src_start = basic_plus_adr(top(), base, offset);
5968
5969 // Call the stub.
5970 address stubAddr = StubRoutines::updateBytesAdler32();
5971 const char *stubName = "updateBytesAdler32";
5972
5973 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5974 stubAddr, stubName, TypePtr::BOTTOM,
5975 crc, src_start, length);
5976
5977 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5978 set_result(result);
5979 return true;
5980 }
5981
5982 //----------------------------inline_reference_get----------------------------
5983 // public T java.lang.ref.Reference.get();
5984 bool LibraryCallKit::inline_reference_get() {
5985 const int referent_offset = java_lang_ref_Reference::referent_offset;
5986 guarantee(referent_offset > 0, "should have already been set");
5987
5988 // Get the argument:
5989 Node* reference_obj = null_check_receiver();
5990 if (stopped()) return true;
5991
5992 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5993
5994 ciInstanceKlass* klass = env()->Object_klass();
5995 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5996
5997 Node* no_ctrl = NULL;
5998 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5999
6000 // Use the pre-barrier to record the value in the referent field
6001 pre_barrier(false /* do_load */,
6002 control(),
6003 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6004 result /* pre_val */,
6005 T_OBJECT);
6006
6007 // Add memory barrier to prevent commoning reads from this field
6008 // across safepoint since GC can change its value.
6009 insert_mem_bar(Op_MemBarCPUOrder);
6010
6011 set_result(result);
6012 return true;
6013 }
6014
6015
6016 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6017 bool is_exact=true, bool is_static=false,
6018 ciInstanceKlass * fromKls=NULL) {
6019 if (fromKls == NULL) {
6020 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6021 assert(tinst != NULL, "obj is null");
6022 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6023 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6024 fromKls = tinst->klass()->as_instance_klass();
6025 } else {
6026 assert(is_static, "only for static field access");
6027 }
6028 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6029 ciSymbol::make(fieldTypeString),
6030 is_static);
6031
6032 assert (field != NULL, "undefined field");
6033 if (field == NULL) return (Node *) NULL;
6034
6035 if (is_static) {
6036 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6037 fromObj = makecon(tip);
6038 }
6039
6040 // Next code copied from Parse::do_get_xxx():
6041
6042 // Compute address and memory type.
6043 int offset = field->offset_in_bytes();
6044 bool is_vol = field->is_volatile();
6045 ciType* field_klass = field->type();
6046 assert(field_klass->is_loaded(), "should be loaded");
6047 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6048 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6049 BasicType bt = field->layout_type();
6050
6051 // Build the resultant type of the load
6052 const Type *type;
6053 if (bt == T_OBJECT) {
6054 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6055 } else {
6056 type = Type::get_const_basic_type(bt);
6057 }
6058
6059 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6060 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6061 }
6062 // Build the load.
6063 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6064 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6065 // If reference is volatile, prevent following memory ops from
6066 // floating up past the volatile read. Also prevents commoning
6067 // another volatile read.
6068 if (is_vol) {
6069 // Memory barrier includes bogus read of value to force load BEFORE membar
6070 insert_mem_bar(Op_MemBarAcquire, loadedField);
6071 }
6072 return loadedField;
6073 }
6074
6075 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6076 bool is_exact = true, bool is_static = false,
6077 ciInstanceKlass * fromKls = NULL) {
6078 if (fromKls == NULL) {
6079 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6080 assert(tinst != NULL, "obj is null");
6081 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6082 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6083 fromKls = tinst->klass()->as_instance_klass();
6084 }
6085 else {
6086 assert(is_static, "only for static field access");
6087 }
6088 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6089 ciSymbol::make(fieldTypeString),
6090 is_static);
6091
6092 assert(field != NULL, "undefined field");
6093 assert(!field->is_volatile(), "not defined for volatile fields");
6094
6095 if (is_static) {
6096 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6097 fromObj = makecon(tip);
6098 }
6099
6100 // Next code copied from Parse::do_get_xxx():
6101
6102 // Compute address and memory type.
6103 int offset = field->offset_in_bytes();
6104 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6105
6106 return adr;
6107 }
6108
6109 //------------------------------inline_aescrypt_Block-----------------------
6110 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6111 address stubAddr = NULL;
6112 const char *stubName;
6113 assert(UseAES, "need AES instruction support");
6114
6115 switch(id) {
6116 case vmIntrinsics::_aescrypt_encryptBlock:
6117 stubAddr = StubRoutines::aescrypt_encryptBlock();
6118 stubName = "aescrypt_encryptBlock";
6119 break;
6120 case vmIntrinsics::_aescrypt_decryptBlock:
6121 stubAddr = StubRoutines::aescrypt_decryptBlock();
6122 stubName = "aescrypt_decryptBlock";
6123 break;
6124 }
6125 if (stubAddr == NULL) return false;
6126
6127 Node* aescrypt_object = argument(0);
6128 Node* src = argument(1);
6129 Node* src_offset = argument(2);
6130 Node* dest = argument(3);
6131 Node* dest_offset = argument(4);
6132
6133 // (1) src and dest are arrays.
6134 const Type* src_type = src->Value(&_gvn);
6135 const Type* dest_type = dest->Value(&_gvn);
6136 const TypeAryPtr* top_src = src_type->isa_aryptr();
6137 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6138 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6139
6140 // for the quick and dirty code we will skip all the checks.
6141 // we are just trying to get the call to be generated.
6142 Node* src_start = src;
6143 Node* dest_start = dest;
6144 if (src_offset != NULL || dest_offset != NULL) {
6145 assert(src_offset != NULL && dest_offset != NULL, "");
6146 src_start = array_element_address(src, src_offset, T_BYTE);
6147 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6148 }
6149
6150 // now need to get the start of its expanded key array
6151 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6152 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6153 if (k_start == NULL) return false;
6154
6155 if (Matcher::pass_original_key_for_aes()) {
6156 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6157 // compatibility issues between Java key expansion and SPARC crypto instructions
6158 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6159 if (original_k_start == NULL) return false;
6160
6161 // Call the stub.
6162 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6163 stubAddr, stubName, TypePtr::BOTTOM,
6164 src_start, dest_start, k_start, original_k_start);
6165 } else {
6166 // Call the stub.
6167 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6168 stubAddr, stubName, TypePtr::BOTTOM,
6169 src_start, dest_start, k_start);
6170 }
6171
6172 return true;
6173 }
6174
6175 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6176 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6177 address stubAddr = NULL;
6178 const char *stubName = NULL;
6179
6180 assert(UseAES, "need AES instruction support");
6181
6182 switch(id) {
6183 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6184 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6185 stubName = "cipherBlockChaining_encryptAESCrypt";
6186 break;
6187 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6188 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6189 stubName = "cipherBlockChaining_decryptAESCrypt";
6190 break;
6191 }
6192 if (stubAddr == NULL) return false;
6193
6194 Node* cipherBlockChaining_object = argument(0);
6195 Node* src = argument(1);
6196 Node* src_offset = argument(2);
6197 Node* len = argument(3);
6198 Node* dest = argument(4);
6199 Node* dest_offset = argument(5);
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
6207 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6208
6209 // checks are the responsibility of the caller
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 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6219 // (because of the predicated logic executed earlier).
6220 // so we cast it here safely.
6221 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6222
6223 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6224 if (embeddedCipherObj == NULL) return false;
6225
6226 // cast it to what we know it will be at runtime
6227 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6228 assert(tinst != NULL, "CBC obj is null");
6229 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6230 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6231 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6232
6233 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6234 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6235 const TypeOopPtr* xtype = aklass->as_instance_type();
6236 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6237 aescrypt_object = _gvn.transform(aescrypt_object);
6238
6239 // we need to get the start of the aescrypt_object's expanded key array
6240 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6241 if (k_start == NULL) return false;
6242
6243 // similarly, get the start address of the r vector
6244 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6245 if (objRvec == NULL) return false;
6246 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6247
6248 Node* cbcCrypt;
6249 if (Matcher::pass_original_key_for_aes()) {
6250 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6251 // compatibility issues between Java key expansion and SPARC crypto instructions
6252 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6253 if (original_k_start == NULL) return false;
6254
6255 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6256 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6257 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6258 stubAddr, stubName, TypePtr::BOTTOM,
6259 src_start, dest_start, k_start, r_start, len, original_k_start);
6260 } else {
6261 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6262 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6263 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6264 stubAddr, stubName, TypePtr::BOTTOM,
6265 src_start, dest_start, k_start, r_start, len);
6266 }
6267
6268 // return cipher length (int)
6269 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6270 set_result(retvalue);
6271 return true;
6272 }
6273
6274 //------------------------------inline_counterMode_AESCrypt-----------------------
6275 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6276 assert(UseAES, "need AES instruction support");
6277 if (!UseAESCTRIntrinsics) return false;
6278
6279 address stubAddr = NULL;
6280 const char *stubName = NULL;
6281 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6282 stubAddr = StubRoutines::counterMode_AESCrypt();
6283 stubName = "counterMode_AESCrypt";
6284 }
6285 if (stubAddr == NULL) return false;
6286
6287 Node* counterMode_object = argument(0);
6288 Node* src = argument(1);
6289 Node* src_offset = argument(2);
6290 Node* len = argument(3);
6291 Node* dest = argument(4);
6292 Node* dest_offset = argument(5);
6293
6294 // (1) src and dest are arrays.
6295 const Type* src_type = src->Value(&_gvn);
6296 const Type* dest_type = dest->Value(&_gvn);
6297 const TypeAryPtr* top_src = src_type->isa_aryptr();
6298 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6299 assert(top_src != NULL && top_src->klass() != NULL &&
6300 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6301
6302 // checks are the responsibility of the caller
6303 Node* src_start = src;
6304 Node* dest_start = dest;
6305 if (src_offset != NULL || dest_offset != NULL) {
6306 assert(src_offset != NULL && dest_offset != NULL, "");
6307 src_start = array_element_address(src, src_offset, T_BYTE);
6308 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6309 }
6310
6311 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6312 // (because of the predicated logic executed earlier).
6313 // so we cast it here safely.
6314 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6315 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6316 if (embeddedCipherObj == NULL) return false;
6317 // cast it to what we know it will be at runtime
6318 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6319 assert(tinst != NULL, "CTR obj is null");
6320 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6321 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6322 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6323 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6324 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6325 const TypeOopPtr* xtype = aklass->as_instance_type();
6326 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6327 aescrypt_object = _gvn.transform(aescrypt_object);
6328 // we need to get the start of the aescrypt_object's expanded key array
6329 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6330 if (k_start == NULL) return false;
6331 // similarly, get the start address of the r vector
6332 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6333 if (obj_counter == NULL) return false;
6334 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6335
6336 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6337 if (saved_encCounter == NULL) return false;
6338 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6339 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6340
6341 Node* ctrCrypt;
6342 if (Matcher::pass_original_key_for_aes()) {
6343 // no SPARC version for AES/CTR intrinsics now.
6344 return false;
6345 }
6346 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6347 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6348 OptoRuntime::counterMode_aescrypt_Type(),
6349 stubAddr, stubName, TypePtr::BOTTOM,
6350 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6351
6352 // return cipher length (int)
6353 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6354 set_result(retvalue);
6355 return true;
6356 }
6357
6358 //------------------------------get_key_start_from_aescrypt_object-----------------------
6359 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6360 #if defined(PPC64) || defined(S390)
6361 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6362 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6363 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6364 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6365 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6366 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6367 if (objSessionK == NULL) {
6368 return (Node *) NULL;
6369 }
6370 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6371 #else
6372 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6373 #endif // PPC64
6374 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6375 if (objAESCryptKey == NULL) return (Node *) NULL;
6376
6377 // now have the array, need to get the start address of the K array
6378 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6379 return k_start;
6380 }
6381
6382 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6383 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6384 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6385 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6386 if (objAESCryptKey == NULL) return (Node *) NULL;
6387
6388 // now have the array, need to get the start address of the lastKey array
6389 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6390 return original_k_start;
6391 }
6392
6393 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6394 // Return node representing slow path of predicate check.
6395 // the pseudo code we want to emulate with this predicate is:
6396 // for encryption:
6397 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6398 // for decryption:
6399 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6400 // note cipher==plain is more conservative than the original java code but that's OK
6401 //
6402 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6403 // The receiver was checked for NULL already.
6404 Node* objCBC = argument(0);
6405
6406 // Load embeddedCipher field of CipherBlockChaining object.
6407 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6408
6409 // get AESCrypt klass for instanceOf check
6410 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6411 // will have same classloader as CipherBlockChaining object
6412 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6413 assert(tinst != NULL, "CBCobj is null");
6414 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6415
6416 // we want to do an instanceof comparison against the AESCrypt class
6417 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6418 if (!klass_AESCrypt->is_loaded()) {
6419 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6420 Node* ctrl = control();
6421 set_control(top()); // no regular fast path
6422 return ctrl;
6423 }
6424 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6425
6426 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6427 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6428 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6429
6430 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6431
6432 // for encryption, we are done
6433 if (!decrypting)
6434 return instof_false; // even if it is NULL
6435
6436 // for decryption, we need to add a further check to avoid
6437 // taking the intrinsic path when cipher and plain are the same
6438 // see the original java code for why.
6439 RegionNode* region = new RegionNode(3);
6440 region->init_req(1, instof_false);
6441 Node* src = argument(1);
6442 Node* dest = argument(4);
6443 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6444 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6445 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6446 region->init_req(2, src_dest_conjoint);
6447
6448 record_for_igvn(region);
6449 return _gvn.transform(region);
6450 }
6451
6452 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6453 // Return node representing slow path of predicate check.
6454 // the pseudo code we want to emulate with this predicate is:
6455 // for encryption:
6456 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6457 // for decryption:
6458 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6459 // note cipher==plain is more conservative than the original java code but that's OK
6460 //
6461
6462 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6463 // The receiver was checked for NULL already.
6464 Node* objCTR = argument(0);
6465
6466 // Load embeddedCipher field of CipherBlockChaining object.
6467 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6468
6469 // get AESCrypt klass for instanceOf check
6470 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6471 // will have same classloader as CipherBlockChaining object
6472 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6473 assert(tinst != NULL, "CTRobj is null");
6474 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6475
6476 // we want to do an instanceof comparison against the AESCrypt class
6477 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6478 if (!klass_AESCrypt->is_loaded()) {
6479 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6480 Node* ctrl = control();
6481 set_control(top()); // no regular fast path
6482 return ctrl;
6483 }
6484
6485 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6486 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6487 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6488 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6489 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6490
6491 return instof_false; // even if it is NULL
6492 }
6493
6494 //------------------------------inline_ghash_processBlocks
6495 bool LibraryCallKit::inline_ghash_processBlocks() {
6496 address stubAddr;
6497 const char *stubName;
6498 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6499
6500 stubAddr = StubRoutines::ghash_processBlocks();
6501 stubName = "ghash_processBlocks";
6502
6503 Node* data = argument(0);
6504 Node* offset = argument(1);
6505 Node* len = argument(2);
6506 Node* state = argument(3);
6507 Node* subkeyH = argument(4);
6508
6509 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6510 assert(state_start, "state is NULL");
6511 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6512 assert(subkeyH_start, "subkeyH is NULL");
6513 Node* data_start = array_element_address(data, offset, T_BYTE);
6514 assert(data_start, "data is NULL");
6515
6516 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6517 OptoRuntime::ghash_processBlocks_Type(),
6518 stubAddr, stubName, TypePtr::BOTTOM,
6519 state_start, subkeyH_start, data_start, len);
6520 return true;
6521 }
6522
6523 //------------------------------inline_sha_implCompress-----------------------
6524 //
6525 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6526 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6527 //
6528 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6529 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6530 //
6531 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6532 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6533 //
6534 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6535 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6536
6537 Node* sha_obj = argument(0);
6538 Node* src = argument(1); // type oop
6539 Node* ofs = argument(2); // type int
6540
6541 const Type* src_type = src->Value(&_gvn);
6542 const TypeAryPtr* top_src = src_type->isa_aryptr();
6543 if (top_src == NULL || top_src->klass() == NULL) {
6544 // failed array check
6545 return false;
6546 }
6547 // Figure out the size and type of the elements we will be copying.
6548 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6549 if (src_elem != T_BYTE) {
6550 return false;
6551 }
6552 // 'src_start' points to src array + offset
6553 Node* src_start = array_element_address(src, ofs, src_elem);
6554 Node* state = NULL;
6555 address stubAddr;
6556 const char *stubName;
6557
6558 switch(id) {
6559 case vmIntrinsics::_sha_implCompress:
6560 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6561 state = get_state_from_sha_object(sha_obj);
6562 stubAddr = StubRoutines::sha1_implCompress();
6563 stubName = "sha1_implCompress";
6564 break;
6565 case vmIntrinsics::_sha2_implCompress:
6566 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6567 state = get_state_from_sha_object(sha_obj);
6568 stubAddr = StubRoutines::sha256_implCompress();
6569 stubName = "sha256_implCompress";
6570 break;
6571 case vmIntrinsics::_sha5_implCompress:
6572 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6573 state = get_state_from_sha5_object(sha_obj);
6574 stubAddr = StubRoutines::sha512_implCompress();
6575 stubName = "sha512_implCompress";
6576 break;
6577 default:
6578 fatal_unexpected_iid(id);
6579 return false;
6580 }
6581 if (state == NULL) return false;
6582
6583 // Call the stub.
6584 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6585 stubAddr, stubName, TypePtr::BOTTOM,
6586 src_start, state);
6587
6588 return true;
6589 }
6590
6591 //------------------------------inline_digestBase_implCompressMB-----------------------
6592 //
6593 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6594 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6595 //
6596 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6597 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6598 "need SHA1/SHA256/SHA512 instruction support");
6599 assert((uint)predicate < 3, "sanity");
6600 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6601
6602 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6603 Node* src = argument(1); // byte[] array
6604 Node* ofs = argument(2); // type int
6605 Node* limit = argument(3); // type int
6606
6607 const Type* src_type = src->Value(&_gvn);
6608 const TypeAryPtr* top_src = src_type->isa_aryptr();
6609 if (top_src == NULL || top_src->klass() == NULL) {
6610 // failed array check
6611 return false;
6612 }
6613 // Figure out the size and type of the elements we will be copying.
6614 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6615 if (src_elem != T_BYTE) {
6616 return false;
6617 }
6618 // 'src_start' points to src array + offset
6619 Node* src_start = array_element_address(src, ofs, src_elem);
6620
6621 const char* klass_SHA_name = NULL;
6622 const char* stub_name = NULL;
6623 address stub_addr = NULL;
6624 bool long_state = false;
6625
6626 switch (predicate) {
6627 case 0:
6628 if (UseSHA1Intrinsics) {
6629 klass_SHA_name = "sun/security/provider/SHA";
6630 stub_name = "sha1_implCompressMB";
6631 stub_addr = StubRoutines::sha1_implCompressMB();
6632 }
6633 break;
6634 case 1:
6635 if (UseSHA256Intrinsics) {
6636 klass_SHA_name = "sun/security/provider/SHA2";
6637 stub_name = "sha256_implCompressMB";
6638 stub_addr = StubRoutines::sha256_implCompressMB();
6639 }
6640 break;
6641 case 2:
6642 if (UseSHA512Intrinsics) {
6643 klass_SHA_name = "sun/security/provider/SHA5";
6644 stub_name = "sha512_implCompressMB";
6645 stub_addr = StubRoutines::sha512_implCompressMB();
6646 long_state = true;
6647 }
6648 break;
6649 default:
6650 fatal("unknown SHA intrinsic predicate: %d", predicate);
6651 }
6652 if (klass_SHA_name != NULL) {
6653 // get DigestBase klass to lookup for SHA klass
6654 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6655 assert(tinst != NULL, "digestBase_obj is not instance???");
6656 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6657
6658 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6659 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6660 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6661 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6662 }
6663 return false;
6664 }
6665 //------------------------------inline_sha_implCompressMB-----------------------
6666 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6667 bool long_state, address stubAddr, const char *stubName,
6668 Node* src_start, Node* ofs, Node* limit) {
6669 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6670 const TypeOopPtr* xtype = aklass->as_instance_type();
6671 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6672 sha_obj = _gvn.transform(sha_obj);
6673
6674 Node* state;
6675 if (long_state) {
6676 state = get_state_from_sha5_object(sha_obj);
6677 } else {
6678 state = get_state_from_sha_object(sha_obj);
6679 }
6680 if (state == NULL) return false;
6681
6682 // Call the stub.
6683 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6684 OptoRuntime::digestBase_implCompressMB_Type(),
6685 stubAddr, stubName, TypePtr::BOTTOM,
6686 src_start, state, ofs, limit);
6687 // return ofs (int)
6688 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6689 set_result(result);
6690
6691 return true;
6692 }
6693
6694 //------------------------------get_state_from_sha_object-----------------------
6695 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6696 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6697 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6698 if (sha_state == NULL) return (Node *) NULL;
6699
6700 // now have the array, need to get the start address of the state array
6701 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6702 return state;
6703 }
6704
6705 //------------------------------get_state_from_sha5_object-----------------------
6706 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6707 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6708 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6709 if (sha_state == NULL) return (Node *) NULL;
6710
6711 // now have the array, need to get the start address of the state array
6712 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6713 return state;
6714 }
6715
6716 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6717 // Return node representing slow path of predicate check.
6718 // the pseudo code we want to emulate with this predicate is:
6719 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6720 //
6721 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6722 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6723 "need SHA1/SHA256/SHA512 instruction support");
6724 assert((uint)predicate < 3, "sanity");
6725
6726 // The receiver was checked for NULL already.
6727 Node* digestBaseObj = argument(0);
6728
6729 // get DigestBase klass for instanceOf check
6730 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6731 assert(tinst != NULL, "digestBaseObj is null");
6732 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6733
6734 const char* klass_SHA_name = NULL;
6735 switch (predicate) {
6736 case 0:
6737 if (UseSHA1Intrinsics) {
6738 // we want to do an instanceof comparison against the SHA class
6739 klass_SHA_name = "sun/security/provider/SHA";
6740 }
6741 break;
6742 case 1:
6743 if (UseSHA256Intrinsics) {
6744 // we want to do an instanceof comparison against the SHA2 class
6745 klass_SHA_name = "sun/security/provider/SHA2";
6746 }
6747 break;
6748 case 2:
6749 if (UseSHA512Intrinsics) {
6750 // we want to do an instanceof comparison against the SHA5 class
6751 klass_SHA_name = "sun/security/provider/SHA5";
6752 }
6753 break;
6754 default:
6755 fatal("unknown SHA intrinsic predicate: %d", predicate);
6756 }
6757
6758 ciKlass* klass_SHA = NULL;
6759 if (klass_SHA_name != NULL) {
6760 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6761 }
6762 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6763 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6764 Node* ctrl = control();
6765 set_control(top()); // no intrinsic path
6766 return ctrl;
6767 }
6768 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6769
6770 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6771 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6772 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6773 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6774
6775 return instof_false; // even if it is NULL
6776 }
6777
6778 //-------------inline_fma-----------------------------------
6779 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6780 Node *a = NULL;
6781 Node *b = NULL;
6782 Node *c = NULL;
6783 Node* result = NULL;
6784 switch (id) {
6785 case vmIntrinsics::_fmaD:
6786 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6787 // no receiver since it is static method
6788 a = round_double_node(argument(0));
6789 b = round_double_node(argument(2));
6790 c = round_double_node(argument(4));
6791 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6792 break;
6793 case vmIntrinsics::_fmaF:
6794 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6795 a = argument(0);
6796 b = argument(1);
6797 c = argument(2);
6798 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6799 break;
6800 default:
6801 fatal_unexpected_iid(id); break;
6802 }
6803 set_result(result);
6804 return true;
6805 }
6806
6807 bool LibraryCallKit::inline_profileBoolean() {
6808 Node* counts = argument(1);
6809 const TypeAryPtr* ary = NULL;
6810 ciArray* aobj = NULL;
6811 if (counts->is_Con()
6812 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6813 && (aobj = ary->const_oop()->as_array()) != NULL
6814 && (aobj->length() == 2)) {
6815 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6816 jint false_cnt = aobj->element_value(0).as_int();
6817 jint true_cnt = aobj->element_value(1).as_int();
6818
6819 if (C->log() != NULL) {
6820 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6821 false_cnt, true_cnt);
6822 }
6823
6824 if (false_cnt + true_cnt == 0) {
6825 // According to profile, never executed.
6826 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6827 Deoptimization::Action_reinterpret);
6828 return true;
6829 }
6830
6831 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6832 // is a number of each value occurrences.
6833 Node* result = argument(0);
6834 if (false_cnt == 0 || true_cnt == 0) {
6835 // According to profile, one value has been never seen.
6836 int expected_val = (false_cnt == 0) ? 1 : 0;
6837
6838 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6839 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6840
6841 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6842 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6843 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6844
6845 { // Slow path: uncommon trap for never seen value and then reexecute
6846 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6847 // the value has been seen at least once.
6848 PreserveJVMState pjvms(this);
6849 PreserveReexecuteState preexecs(this);
6850 jvms()->set_should_reexecute(true);
6851
6852 set_control(slow_path);
6853 set_i_o(i_o());
6854
6855 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6856 Deoptimization::Action_reinterpret);
6857 }
6858 // The guard for never seen value enables sharpening of the result and
6859 // returning a constant. It allows to eliminate branches on the same value
6860 // later on.
6861 set_control(fast_path);
6862 result = intcon(expected_val);
6863 }
6864 // Stop profiling.
6865 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6866 // By replacing method body with profile data (represented as ProfileBooleanNode
6867 // on IR level) we effectively disable profiling.
6868 // It enables full speed execution once optimized code is generated.
6869 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6870 C->record_for_igvn(profile);
6871 set_result(profile);
6872 return true;
6873 } else {
6874 // Continue profiling.
6875 // Profile data isn't available at the moment. So, execute method's bytecode version.
6876 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6877 // is compiled and counters aren't available since corresponding MethodHandle
6878 // isn't a compile-time constant.
6879 return false;
6880 }
6881 }
6882
6883 bool LibraryCallKit::inline_isCompileConstant() {
6884 Node* n = argument(0);
6885 set_result(n->is_Con() ? intcon(1) : intcon(0));
6886 return true;
6887 }