1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/subnode.hpp"
51 #include "prims/nativeLookup.hpp"
52 #include "prims/unsafe.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #ifdef TRACE_HAVE_INTRINSICS
55 #include "trace/traceMacros.hpp"
56 #endif
57
58 class LibraryIntrinsic : public InlineCallGenerator {
59 // Extend the set of intrinsics known to the runtime:
60 public:
61 private:
62 bool _is_virtual;
63 bool _does_virtual_dispatch;
64 int8_t _predicates_count; // Intrinsic is predicated by several conditions
65 int8_t _last_predicate; // Last generated predicate
66 vmIntrinsics::ID _intrinsic_id;
67
68 public:
69 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
70 : InlineCallGenerator(m),
71 _is_virtual(is_virtual),
72 _does_virtual_dispatch(does_virtual_dispatch),
73 _predicates_count((int8_t)predicates_count),
74 _last_predicate((int8_t)-1),
75 _intrinsic_id(id)
76 {
77 }
78 virtual bool is_intrinsic() const { return true; }
79 virtual bool is_virtual() const { return _is_virtual; }
80 virtual bool is_predicated() const { return _predicates_count > 0; }
81 virtual int predicates_count() const { return _predicates_count; }
82 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
83 virtual JVMState* generate(JVMState* jvms);
84 virtual Node* generate_predicate(JVMState* jvms, int predicate);
85 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
86 };
87
88
89 // Local helper class for LibraryIntrinsic:
90 class LibraryCallKit : public GraphKit {
91 private:
92 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
93 Node* _result; // the result node, if any
94 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
95
96 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
97
98 public:
99 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
100 : GraphKit(jvms),
101 _intrinsic(intrinsic),
102 _result(NULL)
103 {
104 // Check if this is a root compile. In that case we don't have a caller.
105 if (!jvms->has_method()) {
106 _reexecute_sp = sp();
107 } else {
108 // Find out how many arguments the interpreter needs when deoptimizing
109 // and save the stack pointer value so it can used by uncommon_trap.
110 // We find the argument count by looking at the declared signature.
111 bool ignored_will_link;
112 ciSignature* declared_signature = NULL;
113 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
114 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
115 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
116 }
117 }
118
119 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
120
121 ciMethod* caller() const { return jvms()->method(); }
122 int bci() const { return jvms()->bci(); }
123 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
124 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
125 ciMethod* callee() const { return _intrinsic->method(); }
126
127 bool try_to_inline(int predicate);
128 Node* try_to_predicate(int predicate);
129
130 void push_result() {
131 // Push the result onto the stack.
132 if (!stopped() && result() != NULL) {
133 BasicType bt = result()->bottom_type()->basic_type();
134 push_node(bt, result());
135 }
136 }
137
138 private:
139 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
140 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
141 }
142
143 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
144 void set_result(RegionNode* region, PhiNode* value);
145 Node* result() { return _result; }
146
147 virtual int reexecute_sp() { return _reexecute_sp; }
148
149 // Helper functions to inline natives
150 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
151 Node* generate_slow_guard(Node* test, RegionNode* region);
152 Node* generate_fair_guard(Node* test, RegionNode* region);
153 Node* generate_negative_guard(Node* index, RegionNode* region,
154 // resulting CastII of index:
155 Node* *pos_index = NULL);
156 Node* generate_limit_guard(Node* offset, Node* subseq_length,
157 Node* array_length,
158 RegionNode* region);
159 void generate_string_range_check(Node* array, Node* offset,
160 Node* length, bool char_count);
161 Node* generate_current_thread(Node* &tls_output);
162 Node* load_mirror_from_klass(Node* klass);
163 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
164 RegionNode* region, int null_path,
165 int offset);
166 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
167 RegionNode* region, int null_path) {
168 int offset = java_lang_Class::klass_offset_in_bytes();
169 return load_klass_from_mirror_common(mirror, never_see_null,
170 region, null_path,
171 offset);
172 }
173 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
174 RegionNode* region, int null_path) {
175 int offset = java_lang_Class::array_klass_offset_in_bytes();
176 return load_klass_from_mirror_common(mirror, never_see_null,
177 region, null_path,
178 offset);
179 }
180 Node* generate_access_flags_guard(Node* kls,
181 int modifier_mask, int modifier_bits,
182 RegionNode* region);
183 Node* generate_interface_guard(Node* kls, RegionNode* region);
184 Node* generate_array_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, false, false);
186 }
187 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, false, true);
189 }
190 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
191 return generate_array_guard_common(kls, region, true, false);
192 }
193 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
194 return generate_array_guard_common(kls, region, true, true);
195 }
196 Node* generate_array_guard_common(Node* kls, RegionNode* region,
197 bool obj_array, bool not_array);
198 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
199 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
200 bool is_virtual = false, bool is_static = false);
201 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
202 return generate_method_call(method_id, false, true);
203 }
204 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
205 return generate_method_call(method_id, true, false);
206 }
207 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
208 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209
210 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
211 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
214 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
215 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
216 bool inline_string_indexOfChar();
217 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
218 bool inline_string_toBytesU();
219 bool inline_string_getCharsU();
220 bool inline_string_copy(bool compress);
221 bool inline_string_char_access(bool is_store);
222 Node* round_double_node(Node* n);
223 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
224 bool inline_math_native(vmIntrinsics::ID id);
225 bool inline_math(vmIntrinsics::ID id);
226 template <typename OverflowOp>
227 bool inline_math_overflow(Node* arg1, Node* arg2);
228 void inline_math_mathExact(Node* math, Node* test);
229 bool inline_math_addExactI(bool is_increment);
230 bool inline_math_addExactL(bool is_increment);
231 bool inline_math_multiplyExactI();
232 bool inline_math_multiplyExactL();
233 bool inline_math_negateExactI();
234 bool inline_math_negateExactL();
235 bool inline_math_subtractExactI(bool is_decrement);
236 bool inline_math_subtractExactL(bool is_decrement);
237 bool inline_min_max(vmIntrinsics::ID id);
238 bool inline_notify(vmIntrinsics::ID id);
239 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
240 // This returns Type::AnyPtr, RawPtr, or OopPtr.
241 int classify_unsafe_addr(Node* &base, Node* &offset);
242 Node* make_unsafe_address(Node* base, Node* offset);
243 // Helper for inline_unsafe_access.
244 // Generates the guards that check whether the result of
245 // Unsafe.getObject should be recorded in an SATB log buffer.
246 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
247
248 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
249 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
250 static bool klass_needs_init_guard(Node* kls);
251 bool inline_unsafe_allocate();
252 bool inline_unsafe_newArray(bool uninitialized);
253 bool inline_unsafe_copyMemory();
254 bool inline_native_currentThread();
255
256 bool inline_native_time_funcs(address method, const char* funcName);
257 #ifdef TRACE_HAVE_INTRINSICS
258 bool inline_native_classID();
259 bool inline_native_getBufferWriter();
260 #endif
261 bool inline_native_isInterrupted();
262 bool inline_native_Class_query(vmIntrinsics::ID id);
263 bool inline_native_subtype_check();
264 bool inline_native_getLength();
265 bool inline_array_copyOf(bool is_copyOfRange);
266 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
267 bool inline_preconditions_checkIndex();
268 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
269 bool inline_native_clone(bool is_virtual);
270 bool inline_native_Reflection_getCallerClass();
271 // Helper function for inlining native object hash method
272 bool inline_native_hashcode(bool is_virtual, bool is_static);
273 bool inline_native_getClass();
274
275 // Helper functions for inlining arraycopy
276 bool inline_arraycopy();
277 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
278 RegionNode* slow_region);
279 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
280 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
281 uint new_idx);
282
283 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
284 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
285 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
286 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
287 bool inline_unsafe_fence(vmIntrinsics::ID id);
288 bool inline_onspinwait();
289 bool inline_fp_conversions(vmIntrinsics::ID id);
290 bool inline_number_methods(vmIntrinsics::ID id);
291 bool inline_reference_get();
292 bool inline_Class_cast();
293 bool inline_aescrypt_Block(vmIntrinsics::ID id);
294 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
295 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
296 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
297 Node* inline_counterMode_AESCrypt_predicate();
298 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
299 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
300 bool inline_ghash_processBlocks();
301 bool inline_sha_implCompress(vmIntrinsics::ID id);
302 bool inline_digestBase_implCompressMB(int predicate);
303 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
304 bool long_state, address stubAddr, const char *stubName,
305 Node* src_start, Node* ofs, Node* limit);
306 Node* get_state_from_sha_object(Node *sha_object);
307 Node* get_state_from_sha5_object(Node *sha_object);
308 Node* inline_digestBase_implCompressMB_predicate(int predicate);
309 bool inline_encodeISOArray();
310 bool inline_updateCRC32();
311 bool inline_updateBytesCRC32();
312 bool inline_updateByteBufferCRC32();
313 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
314 bool inline_updateBytesCRC32C();
315 bool inline_updateDirectByteBufferCRC32C();
316 bool inline_updateBytesAdler32();
317 bool inline_updateByteBufferAdler32();
318 bool inline_multiplyToLen();
319 bool inline_hasNegatives();
320 bool inline_squareToLen();
321 bool inline_mulAdd();
322 bool inline_montgomeryMultiply();
323 bool inline_montgomerySquare();
324 bool inline_vectorizedMismatch();
325 bool inline_fma(vmIntrinsics::ID id);
326
327 bool inline_profileBoolean();
328 bool inline_isCompileConstant();
329 };
330
331 //---------------------------make_vm_intrinsic----------------------------
332 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
333 vmIntrinsics::ID id = m->intrinsic_id();
334 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
335
336 if (!m->is_loaded()) {
337 // Do not attempt to inline unloaded methods.
338 return NULL;
339 }
340
341 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
342 bool is_available = false;
343
344 {
345 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
346 // the compiler must transition to '_thread_in_vm' state because both
347 // methods access VM-internal data.
348 VM_ENTRY_MARK;
349 methodHandle mh(THREAD, m->get_Method());
350 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
351 !C->directive()->is_intrinsic_disabled(mh) &&
352 !vmIntrinsics::is_disabled_by_flags(mh);
353
354 }
355
356 if (is_available) {
357 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
358 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
359 return new LibraryIntrinsic(m, is_virtual,
360 vmIntrinsics::predicates_needed(id),
361 vmIntrinsics::does_virtual_dispatch(id),
362 (vmIntrinsics::ID) id);
363 } else {
364 return NULL;
365 }
366 }
367
368 //----------------------register_library_intrinsics-----------------------
369 // Initialize this file's data structures, for each Compile instance.
370 void Compile::register_library_intrinsics() {
371 // Nothing to do here.
372 }
373
374 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
375 LibraryCallKit kit(jvms, this);
376 Compile* C = kit.C;
377 int nodes = C->unique();
378 #ifndef PRODUCT
379 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
380 char buf[1000];
381 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
382 tty->print_cr("Intrinsic %s", str);
383 }
384 #endif
385 ciMethod* callee = kit.callee();
386 const int bci = kit.bci();
387
388 // Try to inline the intrinsic.
389 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
390 kit.try_to_inline(_last_predicate)) {
391 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
392 : "(intrinsic)";
393 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
394 if (C->print_intrinsics() || C->print_inlining()) {
395 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
396 }
397 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
398 if (C->log()) {
399 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
400 vmIntrinsics::name_at(intrinsic_id()),
401 (is_virtual() ? " virtual='1'" : ""),
402 C->unique() - nodes);
403 }
404 // Push the result from the inlined method onto the stack.
405 kit.push_result();
406 C->print_inlining_update(this);
407 return kit.transfer_exceptions_into_jvms();
408 }
409
410 // The intrinsic bailed out
411 if (jvms->has_method()) {
412 // Not a root compile.
413 const char* msg;
414 if (callee->intrinsic_candidate()) {
415 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
416 } else {
417 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
418 : "failed to inline (intrinsic), method not annotated";
419 }
420 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
421 if (C->print_intrinsics() || C->print_inlining()) {
422 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
423 }
424 } else {
425 // Root compile
426 ResourceMark rm;
427 stringStream msg_stream;
428 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
429 vmIntrinsics::name_at(intrinsic_id()),
430 is_virtual() ? " (virtual)" : "", bci);
431 const char *msg = msg_stream.as_string();
432 log_debug(jit, inlining)("%s", msg);
433 if (C->print_intrinsics() || C->print_inlining()) {
434 tty->print("%s", msg);
435 }
436 }
437 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
438 C->print_inlining_update(this);
439 return NULL;
440 }
441
442 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
443 LibraryCallKit kit(jvms, this);
444 Compile* C = kit.C;
445 int nodes = C->unique();
446 _last_predicate = predicate;
447 #ifndef PRODUCT
448 assert(is_predicated() && predicate < predicates_count(), "sanity");
449 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
450 char buf[1000];
451 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
452 tty->print_cr("Predicate for intrinsic %s", str);
453 }
454 #endif
455 ciMethod* callee = kit.callee();
456 const int bci = kit.bci();
457
458 Node* slow_ctl = kit.try_to_predicate(predicate);
459 if (!kit.failing()) {
460 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
461 : "(intrinsic, predicate)";
462 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
463 if (C->print_intrinsics() || C->print_inlining()) {
464 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
465 }
466 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
467 if (C->log()) {
468 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
469 vmIntrinsics::name_at(intrinsic_id()),
470 (is_virtual() ? " virtual='1'" : ""),
471 C->unique() - nodes);
472 }
473 return slow_ctl; // Could be NULL if the check folds.
474 }
475
476 // The intrinsic bailed out
477 if (jvms->has_method()) {
478 // Not a root compile.
479 const char* msg = "failed to generate predicate for intrinsic";
480 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
481 if (C->print_intrinsics() || C->print_inlining()) {
482 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
483 }
484 } else {
485 // Root compile
486 ResourceMark rm;
487 stringStream msg_stream;
488 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
489 vmIntrinsics::name_at(intrinsic_id()),
490 is_virtual() ? " (virtual)" : "", bci);
491 const char *msg = msg_stream.as_string();
492 log_debug(jit, inlining)("%s", msg);
493 if (C->print_intrinsics() || C->print_inlining()) {
494 C->print_inlining_stream()->print("%s", msg);
495 }
496 }
497 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
498 return NULL;
499 }
500
501 bool LibraryCallKit::try_to_inline(int predicate) {
502 // Handle symbolic names for otherwise undistinguished boolean switches:
503 const bool is_store = true;
504 const bool is_compress = true;
505 const bool is_static = true;
506 const bool is_volatile = true;
507
508 if (!jvms()->has_method()) {
509 // Root JVMState has a null method.
510 assert(map()->memory()->Opcode() == Op_Parm, "");
511 // Insert the memory aliasing node
512 set_all_memory(reset_memory());
513 }
514 assert(merged_memory(), "");
515
516
517 switch (intrinsic_id()) {
518 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
519 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
520 case vmIntrinsics::_getClass: return inline_native_getClass();
521
522 case vmIntrinsics::_dsin:
523 case vmIntrinsics::_dcos:
524 case vmIntrinsics::_dtan:
525 case vmIntrinsics::_dabs:
526 case vmIntrinsics::_datan2:
527 case vmIntrinsics::_dsqrt:
528 case vmIntrinsics::_dexp:
529 case vmIntrinsics::_dlog:
530 case vmIntrinsics::_dlog10:
531 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
532
533 case vmIntrinsics::_min:
534 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
535
536 case vmIntrinsics::_notify:
537 case vmIntrinsics::_notifyAll:
538 if (InlineNotify) {
539 return inline_notify(intrinsic_id());
540 }
541 return false;
542
543 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
544 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
545 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
546 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
547 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
548 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
549 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
550 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
551 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
552 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
553 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
554 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
555
556 case vmIntrinsics::_arraycopy: return inline_arraycopy();
557
558 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
559 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
560 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
561 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
562
563 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
564 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
565 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
566 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
567 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
568 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
569 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
570
571 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
572 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
573
574 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
575 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
576 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
577 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
578
579 case vmIntrinsics::_compressStringC:
580 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
581 case vmIntrinsics::_inflateStringC:
582 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
583
584 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
585 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
586 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
587 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
588 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
589 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
590 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
591 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
592 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
593
594 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
595 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
596 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
597 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
598 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
599 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
600 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
601 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
602 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
603
604 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
605 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
606 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
607 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
608 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
609 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
610 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
611 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
612 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
613
614 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
615 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
616 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
617 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
618 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
619 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
620 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
621 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
622 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
623
624 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
625 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
626 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
627 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
628
629 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
630 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
631 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
632 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
633
634 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
635 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
636 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
637 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
638 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
639 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
640 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
641 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
642 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
643
644 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
645 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
646 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
647 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
648 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
649 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
650 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
651 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
652 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
653
654 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
655 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
656 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
657 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
658 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
659 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
660 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
661 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
662 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
663
664 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
665 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
666 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
667 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
668 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
669 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
670 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
671 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
672 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
673
674 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
675 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
676 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
677 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
678 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
679
680 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
681 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
682 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
683 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
684 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
685 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
686 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
687 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
688 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
689 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
690 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
691 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
692 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
693 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
694 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
695 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
696 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
697 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
698 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
699 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
700
701 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
702 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
703 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
704 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
705 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
706 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
707 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
708 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
709 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
710 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
711 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
712 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
713 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
714 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
715 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
716
717 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
718 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
719 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
720 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
721
722 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
723 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
724 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
725 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
726 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
727
728 case vmIntrinsics::_loadFence:
729 case vmIntrinsics::_storeFence:
730 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
731
732 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
733
734 case vmIntrinsics::_currentThread: return inline_native_currentThread();
735 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
736
737 #ifdef TRACE_HAVE_INTRINSICS
738 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
739 case vmIntrinsics::_getClassId: return inline_native_classID();
740 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
741 #endif
742 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
743 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
744 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
745 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
746 case vmIntrinsics::_getLength: return inline_native_getLength();
747 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
748 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
749 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
750 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
751 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
752 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
753
754 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
755 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
756
757 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
758
759 case vmIntrinsics::_isInstance:
760 case vmIntrinsics::_getModifiers:
761 case vmIntrinsics::_isInterface:
762 case vmIntrinsics::_isArray:
763 case vmIntrinsics::_isPrimitive:
764 case vmIntrinsics::_getSuperclass:
765 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
766
767 case vmIntrinsics::_floatToRawIntBits:
768 case vmIntrinsics::_floatToIntBits:
769 case vmIntrinsics::_intBitsToFloat:
770 case vmIntrinsics::_doubleToRawLongBits:
771 case vmIntrinsics::_doubleToLongBits:
772 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
773
774 case vmIntrinsics::_numberOfLeadingZeros_i:
775 case vmIntrinsics::_numberOfLeadingZeros_l:
776 case vmIntrinsics::_numberOfTrailingZeros_i:
777 case vmIntrinsics::_numberOfTrailingZeros_l:
778 case vmIntrinsics::_bitCount_i:
779 case vmIntrinsics::_bitCount_l:
780 case vmIntrinsics::_reverseBytes_i:
781 case vmIntrinsics::_reverseBytes_l:
782 case vmIntrinsics::_reverseBytes_s:
783 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
784
785 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
786
787 case vmIntrinsics::_Reference_get: return inline_reference_get();
788
789 case vmIntrinsics::_Class_cast: return inline_Class_cast();
790
791 case vmIntrinsics::_aescrypt_encryptBlock:
792 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
793
794 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
795 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
796 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
797
798 case vmIntrinsics::_counterMode_AESCrypt:
799 return inline_counterMode_AESCrypt(intrinsic_id());
800
801 case vmIntrinsics::_sha_implCompress:
802 case vmIntrinsics::_sha2_implCompress:
803 case vmIntrinsics::_sha5_implCompress:
804 return inline_sha_implCompress(intrinsic_id());
805
806 case vmIntrinsics::_digestBase_implCompressMB:
807 return inline_digestBase_implCompressMB(predicate);
808
809 case vmIntrinsics::_multiplyToLen:
810 return inline_multiplyToLen();
811
812 case vmIntrinsics::_squareToLen:
813 return inline_squareToLen();
814
815 case vmIntrinsics::_mulAdd:
816 return inline_mulAdd();
817
818 case vmIntrinsics::_montgomeryMultiply:
819 return inline_montgomeryMultiply();
820 case vmIntrinsics::_montgomerySquare:
821 return inline_montgomerySquare();
822
823 case vmIntrinsics::_vectorizedMismatch:
824 return inline_vectorizedMismatch();
825
826 case vmIntrinsics::_ghash_processBlocks:
827 return inline_ghash_processBlocks();
828
829 case vmIntrinsics::_encodeISOArray:
830 case vmIntrinsics::_encodeByteISOArray:
831 return inline_encodeISOArray();
832
833 case vmIntrinsics::_updateCRC32:
834 return inline_updateCRC32();
835 case vmIntrinsics::_updateBytesCRC32:
836 return inline_updateBytesCRC32();
837 case vmIntrinsics::_updateByteBufferCRC32:
838 return inline_updateByteBufferCRC32();
839
840 case vmIntrinsics::_updateBytesCRC32C:
841 return inline_updateBytesCRC32C();
842 case vmIntrinsics::_updateDirectByteBufferCRC32C:
843 return inline_updateDirectByteBufferCRC32C();
844
845 case vmIntrinsics::_updateBytesAdler32:
846 return inline_updateBytesAdler32();
847 case vmIntrinsics::_updateByteBufferAdler32:
848 return inline_updateByteBufferAdler32();
849
850 case vmIntrinsics::_profileBoolean:
851 return inline_profileBoolean();
852 case vmIntrinsics::_isCompileConstant:
853 return inline_isCompileConstant();
854
855 case vmIntrinsics::_hasNegatives:
856 return inline_hasNegatives();
857
858 case vmIntrinsics::_fmaD:
859 case vmIntrinsics::_fmaF:
860 return inline_fma(intrinsic_id());
861
862 default:
863 // If you get here, it may be that someone has added a new intrinsic
864 // to the list in vmSymbols.hpp without implementing it here.
865 #ifndef PRODUCT
866 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
867 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
868 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
869 }
870 #endif
871 return false;
872 }
873 }
874
875 Node* LibraryCallKit::try_to_predicate(int predicate) {
876 if (!jvms()->has_method()) {
877 // Root JVMState has a null method.
878 assert(map()->memory()->Opcode() == Op_Parm, "");
879 // Insert the memory aliasing node
880 set_all_memory(reset_memory());
881 }
882 assert(merged_memory(), "");
883
884 switch (intrinsic_id()) {
885 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
886 return inline_cipherBlockChaining_AESCrypt_predicate(false);
887 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
888 return inline_cipherBlockChaining_AESCrypt_predicate(true);
889 case vmIntrinsics::_counterMode_AESCrypt:
890 return inline_counterMode_AESCrypt_predicate();
891 case vmIntrinsics::_digestBase_implCompressMB:
892 return inline_digestBase_implCompressMB_predicate(predicate);
893
894 default:
895 // If you get here, it may be that someone has added a new intrinsic
896 // to the list in vmSymbols.hpp without implementing it here.
897 #ifndef PRODUCT
898 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
899 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
900 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
901 }
902 #endif
903 Node* slow_ctl = control();
904 set_control(top()); // No fast path instrinsic
905 return slow_ctl;
906 }
907 }
908
909 //------------------------------set_result-------------------------------
910 // Helper function for finishing intrinsics.
911 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
912 record_for_igvn(region);
913 set_control(_gvn.transform(region));
914 set_result( _gvn.transform(value));
915 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
916 }
917
918 //------------------------------generate_guard---------------------------
919 // Helper function for generating guarded fast-slow graph structures.
920 // The given 'test', if true, guards a slow path. If the test fails
921 // then a fast path can be taken. (We generally hope it fails.)
922 // In all cases, GraphKit::control() is updated to the fast path.
923 // The returned value represents the control for the slow path.
924 // The return value is never 'top'; it is either a valid control
925 // or NULL if it is obvious that the slow path can never be taken.
926 // Also, if region and the slow control are not NULL, the slow edge
927 // is appended to the region.
928 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
929 if (stopped()) {
930 // Already short circuited.
931 return NULL;
932 }
933
934 // Build an if node and its projections.
935 // If test is true we take the slow path, which we assume is uncommon.
936 if (_gvn.type(test) == TypeInt::ZERO) {
937 // The slow branch is never taken. No need to build this guard.
938 return NULL;
939 }
940
941 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
942
943 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
944 if (if_slow == top()) {
945 // The slow branch is never taken. No need to build this guard.
946 return NULL;
947 }
948
949 if (region != NULL)
950 region->add_req(if_slow);
951
952 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
953 set_control(if_fast);
954
955 return if_slow;
956 }
957
958 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
959 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
960 }
961 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
962 return generate_guard(test, region, PROB_FAIR);
963 }
964
965 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
966 Node* *pos_index) {
967 if (stopped())
968 return NULL; // already stopped
969 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
970 return NULL; // index is already adequately typed
971 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
972 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
973 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
974 if (is_neg != NULL && pos_index != NULL) {
975 // Emulate effect of Parse::adjust_map_after_if.
976 Node* ccast = new CastIINode(index, TypeInt::POS);
977 ccast->set_req(0, control());
978 (*pos_index) = _gvn.transform(ccast);
979 }
980 return is_neg;
981 }
982
983 // Make sure that 'position' is a valid limit index, in [0..length].
984 // There are two equivalent plans for checking this:
985 // A. (offset + copyLength) unsigned<= arrayLength
986 // B. offset <= (arrayLength - copyLength)
987 // We require that all of the values above, except for the sum and
988 // difference, are already known to be non-negative.
989 // Plan A is robust in the face of overflow, if offset and copyLength
990 // are both hugely positive.
991 //
992 // Plan B is less direct and intuitive, but it does not overflow at
993 // all, since the difference of two non-negatives is always
994 // representable. Whenever Java methods must perform the equivalent
995 // check they generally use Plan B instead of Plan A.
996 // For the moment we use Plan A.
997 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
998 Node* subseq_length,
999 Node* array_length,
1000 RegionNode* region) {
1001 if (stopped())
1002 return NULL; // already stopped
1003 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1004 if (zero_offset && subseq_length->eqv_uncast(array_length))
1005 return NULL; // common case of whole-array copy
1006 Node* last = subseq_length;
1007 if (!zero_offset) // last += offset
1008 last = _gvn.transform(new AddINode(last, offset));
1009 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1010 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1011 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1012 return is_over;
1013 }
1014
1015 // Emit range checks for the given String.value byte array
1016 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1017 if (stopped()) {
1018 return; // already stopped
1019 }
1020 RegionNode* bailout = new RegionNode(1);
1021 record_for_igvn(bailout);
1022 if (char_count) {
1023 // Convert char count to byte count
1024 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1025 }
1026
1027 // Offset and count must not be negative
1028 generate_negative_guard(offset, bailout);
1029 generate_negative_guard(count, bailout);
1030 // Offset + count must not exceed length of array
1031 generate_limit_guard(offset, count, load_array_length(array), bailout);
1032
1033 if (bailout->req() > 1) {
1034 PreserveJVMState pjvms(this);
1035 set_control(_gvn.transform(bailout));
1036 uncommon_trap(Deoptimization::Reason_intrinsic,
1037 Deoptimization::Action_maybe_recompile);
1038 }
1039 }
1040
1041 //--------------------------generate_current_thread--------------------
1042 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1043 ciKlass* thread_klass = env()->Thread_klass();
1044 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1045 Node* thread = _gvn.transform(new ThreadLocalNode());
1046 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1047 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1048 tls_output = thread;
1049 return threadObj;
1050 }
1051
1052
1053 //------------------------------make_string_method_node------------------------
1054 // Helper method for String intrinsic functions. This version is called with
1055 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1056 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1057 // containing the lengths of str1 and str2.
1058 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1059 Node* result = NULL;
1060 switch (opcode) {
1061 case Op_StrIndexOf:
1062 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1063 str1_start, cnt1, str2_start, cnt2, ae);
1064 break;
1065 case Op_StrComp:
1066 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1067 str1_start, cnt1, str2_start, cnt2, ae);
1068 break;
1069 case Op_StrEquals:
1070 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1071 // Use the constant length if there is one because optimized match rule may exist.
1072 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1073 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1074 break;
1075 default:
1076 ShouldNotReachHere();
1077 return NULL;
1078 }
1079
1080 // All these intrinsics have checks.
1081 C->set_has_split_ifs(true); // Has chance for split-if optimization
1082
1083 return _gvn.transform(result);
1084 }
1085
1086 //------------------------------inline_string_compareTo------------------------
1087 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1088 Node* arg1 = argument(0);
1089 Node* arg2 = argument(1);
1090
1091 // Get start addr and length of first argument
1092 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1093 Node* arg1_cnt = load_array_length(arg1);
1094
1095 // Get start addr and length of second argument
1096 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1097 Node* arg2_cnt = load_array_length(arg2);
1098
1099 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1100 set_result(result);
1101 return true;
1102 }
1103
1104 //------------------------------inline_string_equals------------------------
1105 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1106 Node* arg1 = argument(0);
1107 Node* arg2 = argument(1);
1108
1109 // paths (plus control) merge
1110 RegionNode* region = new RegionNode(3);
1111 Node* phi = new PhiNode(region, TypeInt::BOOL);
1112
1113 if (!stopped()) {
1114 // Get start addr and length of first argument
1115 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1116 Node* arg1_cnt = load_array_length(arg1);
1117
1118 // Get start addr and length of second argument
1119 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1120 Node* arg2_cnt = load_array_length(arg2);
1121
1122 // Check for arg1_cnt != arg2_cnt
1123 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1124 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1125 Node* if_ne = generate_slow_guard(bol, NULL);
1126 if (if_ne != NULL) {
1127 phi->init_req(2, intcon(0));
1128 region->init_req(2, if_ne);
1129 }
1130
1131 // Check for count == 0 is done by assembler code for StrEquals.
1132
1133 if (!stopped()) {
1134 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1135 phi->init_req(1, equals);
1136 region->init_req(1, control());
1137 }
1138 }
1139
1140 // post merge
1141 set_control(_gvn.transform(region));
1142 record_for_igvn(region);
1143
1144 set_result(_gvn.transform(phi));
1145 return true;
1146 }
1147
1148 //------------------------------inline_array_equals----------------------------
1149 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1150 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1151 Node* arg1 = argument(0);
1152 Node* arg2 = argument(1);
1153
1154 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1155 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1156 return true;
1157 }
1158
1159 //------------------------------inline_hasNegatives------------------------------
1160 bool LibraryCallKit::inline_hasNegatives() {
1161 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1162 return false;
1163 }
1164
1165 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1166 // no receiver since it is static method
1167 Node* ba = argument(0);
1168 Node* offset = argument(1);
1169 Node* len = argument(2);
1170
1171 // Range checks
1172 generate_string_range_check(ba, offset, len, false);
1173 if (stopped()) {
1174 return true;
1175 }
1176 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1177 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1178 set_result(_gvn.transform(result));
1179 return true;
1180 }
1181
1182 bool LibraryCallKit::inline_preconditions_checkIndex() {
1183 Node* index = argument(0);
1184 Node* length = argument(1);
1185 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1186 return false;
1187 }
1188
1189 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1190 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1191
1192 {
1193 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1194 uncommon_trap(Deoptimization::Reason_intrinsic,
1195 Deoptimization::Action_make_not_entrant);
1196 }
1197
1198 if (stopped()) {
1199 return false;
1200 }
1201
1202 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1203 BoolTest::mask btest = BoolTest::lt;
1204 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1205 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1206 _gvn.set_type(rc, rc->Value(&_gvn));
1207 if (!rc_bool->is_Con()) {
1208 record_for_igvn(rc);
1209 }
1210 set_control(_gvn.transform(new IfTrueNode(rc)));
1211 {
1212 PreserveJVMState pjvms(this);
1213 set_control(_gvn.transform(new IfFalseNode(rc)));
1214 uncommon_trap(Deoptimization::Reason_range_check,
1215 Deoptimization::Action_make_not_entrant);
1216 }
1217
1218 if (stopped()) {
1219 return false;
1220 }
1221
1222 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1223 result->set_req(0, control());
1224 result = _gvn.transform(result);
1225 set_result(result);
1226 replace_in_map(index, result);
1227 return true;
1228 }
1229
1230 //------------------------------inline_string_indexOf------------------------
1231 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1232 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1233 return false;
1234 }
1235 Node* src = argument(0);
1236 Node* tgt = argument(1);
1237
1238 // Make the merge point
1239 RegionNode* result_rgn = new RegionNode(4);
1240 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1241
1242 // Get start addr and length of source string
1243 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1244 Node* src_count = load_array_length(src);
1245
1246 // Get start addr and length of substring
1247 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1248 Node* tgt_count = load_array_length(tgt);
1249
1250 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1251 // Divide src size by 2 if String is UTF16 encoded
1252 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1253 }
1254 if (ae == StrIntrinsicNode::UU) {
1255 // Divide substring size by 2 if String is UTF16 encoded
1256 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1257 }
1258
1259 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1260 if (result != NULL) {
1261 result_phi->init_req(3, result);
1262 result_rgn->init_req(3, control());
1263 }
1264 set_control(_gvn.transform(result_rgn));
1265 record_for_igvn(result_rgn);
1266 set_result(_gvn.transform(result_phi));
1267
1268 return true;
1269 }
1270
1271 //-----------------------------inline_string_indexOf-----------------------
1272 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1273 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1274 return false;
1275 }
1276 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1277 return false;
1278 }
1279 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1280 Node* src = argument(0); // byte[]
1281 Node* src_count = argument(1); // char count
1282 Node* tgt = argument(2); // byte[]
1283 Node* tgt_count = argument(3); // char count
1284 Node* from_index = argument(4); // char index
1285
1286 // Multiply byte array index by 2 if String is UTF16 encoded
1287 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1288 src_count = _gvn.transform(new SubINode(src_count, from_index));
1289 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1290 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1291
1292 // Range checks
1293 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1294 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1295 if (stopped()) {
1296 return true;
1297 }
1298
1299 RegionNode* region = new RegionNode(5);
1300 Node* phi = new PhiNode(region, TypeInt::INT);
1301
1302 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1303 if (result != NULL) {
1304 // The result is index relative to from_index if substring was found, -1 otherwise.
1305 // Generate code which will fold into cmove.
1306 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1307 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1308
1309 Node* if_lt = generate_slow_guard(bol, NULL);
1310 if (if_lt != NULL) {
1311 // result == -1
1312 phi->init_req(3, result);
1313 region->init_req(3, if_lt);
1314 }
1315 if (!stopped()) {
1316 result = _gvn.transform(new AddINode(result, from_index));
1317 phi->init_req(4, result);
1318 region->init_req(4, control());
1319 }
1320 }
1321
1322 set_control(_gvn.transform(region));
1323 record_for_igvn(region);
1324 set_result(_gvn.transform(phi));
1325
1326 return true;
1327 }
1328
1329 // Create StrIndexOfNode with fast path checks
1330 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1331 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1332 // Check for substr count > string count
1333 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1334 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1335 Node* if_gt = generate_slow_guard(bol, NULL);
1336 if (if_gt != NULL) {
1337 phi->init_req(1, intcon(-1));
1338 region->init_req(1, if_gt);
1339 }
1340 if (!stopped()) {
1341 // Check for substr count == 0
1342 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1343 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1344 Node* if_zero = generate_slow_guard(bol, NULL);
1345 if (if_zero != NULL) {
1346 phi->init_req(2, intcon(0));
1347 region->init_req(2, if_zero);
1348 }
1349 }
1350 if (!stopped()) {
1351 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1352 }
1353 return NULL;
1354 }
1355
1356 //-----------------------------inline_string_indexOfChar-----------------------
1357 bool LibraryCallKit::inline_string_indexOfChar() {
1358 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1359 return false;
1360 }
1361 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1362 return false;
1363 }
1364 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1365 Node* src = argument(0); // byte[]
1366 Node* tgt = argument(1); // tgt is int ch
1367 Node* from_index = argument(2);
1368 Node* max = argument(3);
1369
1370 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1371 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1372 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1373
1374 // Range checks
1375 generate_string_range_check(src, src_offset, src_count, true);
1376 if (stopped()) {
1377 return true;
1378 }
1379
1380 RegionNode* region = new RegionNode(3);
1381 Node* phi = new PhiNode(region, TypeInt::INT);
1382
1383 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1384 C->set_has_split_ifs(true); // Has chance for split-if optimization
1385 _gvn.transform(result);
1386
1387 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1388 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1389
1390 Node* if_lt = generate_slow_guard(bol, NULL);
1391 if (if_lt != NULL) {
1392 // result == -1
1393 phi->init_req(2, result);
1394 region->init_req(2, if_lt);
1395 }
1396 if (!stopped()) {
1397 result = _gvn.transform(new AddINode(result, from_index));
1398 phi->init_req(1, result);
1399 region->init_req(1, control());
1400 }
1401 set_control(_gvn.transform(region));
1402 record_for_igvn(region);
1403 set_result(_gvn.transform(phi));
1404
1405 return true;
1406 }
1407 //---------------------------inline_string_copy---------------------
1408 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1409 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1410 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1411 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1412 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1413 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1414 bool LibraryCallKit::inline_string_copy(bool compress) {
1415 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1416 return false;
1417 }
1418 int nargs = 5; // 2 oops, 3 ints
1419 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1420
1421 Node* src = argument(0);
1422 Node* src_offset = argument(1);
1423 Node* dst = argument(2);
1424 Node* dst_offset = argument(3);
1425 Node* length = argument(4);
1426
1427 // Check for allocation before we add nodes that would confuse
1428 // tightly_coupled_allocation()
1429 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1430
1431 // Figure out the size and type of the elements we will be copying.
1432 const Type* src_type = src->Value(&_gvn);
1433 const Type* dst_type = dst->Value(&_gvn);
1434 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1435 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1436 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1437 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1438 "Unsupported array types for inline_string_copy");
1439
1440 // Convert char[] offsets to byte[] offsets
1441 bool convert_src = (compress && src_elem == T_BYTE);
1442 bool convert_dst = (!compress && dst_elem == T_BYTE);
1443 if (convert_src) {
1444 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1445 } else if (convert_dst) {
1446 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1447 }
1448
1449 // Range checks
1450 generate_string_range_check(src, src_offset, length, convert_src);
1451 generate_string_range_check(dst, dst_offset, length, convert_dst);
1452 if (stopped()) {
1453 return true;
1454 }
1455
1456 Node* src_start = array_element_address(src, src_offset, src_elem);
1457 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1458 // 'src_start' points to src array + scaled offset
1459 // 'dst_start' points to dst array + scaled offset
1460 Node* count = NULL;
1461 if (compress) {
1462 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1463 } else {
1464 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1465 }
1466
1467 if (alloc != NULL) {
1468 if (alloc->maybe_set_complete(&_gvn)) {
1469 // "You break it, you buy it."
1470 InitializeNode* init = alloc->initialization();
1471 assert(init->is_complete(), "we just did this");
1472 init->set_complete_with_arraycopy();
1473 assert(dst->is_CheckCastPP(), "sanity");
1474 assert(dst->in(0)->in(0) == init, "dest pinned");
1475 }
1476 // Do not let stores that initialize this object be reordered with
1477 // a subsequent store that would make this object accessible by
1478 // other threads.
1479 // Record what AllocateNode this StoreStore protects so that
1480 // escape analysis can go from the MemBarStoreStoreNode to the
1481 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1482 // based on the escape status of the AllocateNode.
1483 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1484 }
1485 if (compress) {
1486 set_result(_gvn.transform(count));
1487 }
1488 return true;
1489 }
1490
1491 #ifdef _LP64
1492 #define XTOP ,top() /*additional argument*/
1493 #else //_LP64
1494 #define XTOP /*no additional argument*/
1495 #endif //_LP64
1496
1497 //------------------------inline_string_toBytesU--------------------------
1498 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1499 bool LibraryCallKit::inline_string_toBytesU() {
1500 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1501 return false;
1502 }
1503 // Get the arguments.
1504 Node* value = argument(0);
1505 Node* offset = argument(1);
1506 Node* length = argument(2);
1507
1508 Node* newcopy = NULL;
1509
1510 // Set the original stack and the reexecute bit for the interpreter to reexecute
1511 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1512 { PreserveReexecuteState preexecs(this);
1513 jvms()->set_should_reexecute(true);
1514
1515 // Check if a null path was taken unconditionally.
1516 value = null_check(value);
1517
1518 RegionNode* bailout = new RegionNode(1);
1519 record_for_igvn(bailout);
1520
1521 // Range checks
1522 generate_negative_guard(offset, bailout);
1523 generate_negative_guard(length, bailout);
1524 generate_limit_guard(offset, length, load_array_length(value), bailout);
1525 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1526 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1527
1528 if (bailout->req() > 1) {
1529 PreserveJVMState pjvms(this);
1530 set_control(_gvn.transform(bailout));
1531 uncommon_trap(Deoptimization::Reason_intrinsic,
1532 Deoptimization::Action_maybe_recompile);
1533 }
1534 if (stopped()) {
1535 return true;
1536 }
1537
1538 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1539 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1540 newcopy = new_array(klass_node, size, 0); // no arguments to push
1541 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1542
1543 // Calculate starting addresses.
1544 Node* src_start = array_element_address(value, offset, T_CHAR);
1545 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1546
1547 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1548 const TypeInt* toffset = gvn().type(offset)->is_int();
1549 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1550
1551 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1552 const char* copyfunc_name = "arraycopy";
1553 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1554 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1555 OptoRuntime::fast_arraycopy_Type(),
1556 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1557 src_start, dst_start, ConvI2X(length) XTOP);
1558 // Do not let reads from the cloned object float above the arraycopy.
1559 if (alloc != NULL) {
1560 if (alloc->maybe_set_complete(&_gvn)) {
1561 // "You break it, you buy it."
1562 InitializeNode* init = alloc->initialization();
1563 assert(init->is_complete(), "we just did this");
1564 init->set_complete_with_arraycopy();
1565 assert(newcopy->is_CheckCastPP(), "sanity");
1566 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1567 }
1568 // Do not let stores that initialize this object be reordered with
1569 // a subsequent store that would make this object accessible by
1570 // other threads.
1571 // Record what AllocateNode this StoreStore protects so that
1572 // escape analysis can go from the MemBarStoreStoreNode to the
1573 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1574 // based on the escape status of the AllocateNode.
1575 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1576 } else {
1577 insert_mem_bar(Op_MemBarCPUOrder);
1578 }
1579 } // original reexecute is set back here
1580
1581 C->set_has_split_ifs(true); // Has chance for split-if optimization
1582 if (!stopped()) {
1583 set_result(newcopy);
1584 }
1585 return true;
1586 }
1587
1588 //------------------------inline_string_getCharsU--------------------------
1589 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1590 bool LibraryCallKit::inline_string_getCharsU() {
1591 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1592 return false;
1593 }
1594
1595 // Get the arguments.
1596 Node* src = argument(0);
1597 Node* src_begin = argument(1);
1598 Node* src_end = argument(2); // exclusive offset (i < src_end)
1599 Node* dst = argument(3);
1600 Node* dst_begin = argument(4);
1601
1602 // Check for allocation before we add nodes that would confuse
1603 // tightly_coupled_allocation()
1604 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1605
1606 // Check if a null path was taken unconditionally.
1607 src = null_check(src);
1608 dst = null_check(dst);
1609 if (stopped()) {
1610 return true;
1611 }
1612
1613 // Get length and convert char[] offset to byte[] offset
1614 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1615 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1616
1617 // Range checks
1618 generate_string_range_check(src, src_begin, length, true);
1619 generate_string_range_check(dst, dst_begin, length, false);
1620 if (stopped()) {
1621 return true;
1622 }
1623
1624 if (!stopped()) {
1625 // Calculate starting addresses.
1626 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1627 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1628
1629 // Check if array addresses are aligned to HeapWordSize
1630 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1631 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1632 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1633 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1634
1635 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1636 const char* copyfunc_name = "arraycopy";
1637 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1638 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1639 OptoRuntime::fast_arraycopy_Type(),
1640 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1641 src_start, dst_start, ConvI2X(length) XTOP);
1642 // Do not let reads from the cloned object float above the arraycopy.
1643 if (alloc != NULL) {
1644 if (alloc->maybe_set_complete(&_gvn)) {
1645 // "You break it, you buy it."
1646 InitializeNode* init = alloc->initialization();
1647 assert(init->is_complete(), "we just did this");
1648 init->set_complete_with_arraycopy();
1649 assert(dst->is_CheckCastPP(), "sanity");
1650 assert(dst->in(0)->in(0) == init, "dest pinned");
1651 }
1652 // Do not let stores that initialize this object be reordered with
1653 // a subsequent store that would make this object accessible by
1654 // other threads.
1655 // Record what AllocateNode this StoreStore protects so that
1656 // escape analysis can go from the MemBarStoreStoreNode to the
1657 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1658 // based on the escape status of the AllocateNode.
1659 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1660 } else {
1661 insert_mem_bar(Op_MemBarCPUOrder);
1662 }
1663 }
1664
1665 C->set_has_split_ifs(true); // Has chance for split-if optimization
1666 return true;
1667 }
1668
1669 //----------------------inline_string_char_access----------------------------
1670 // Store/Load char to/from byte[] array.
1671 // static void StringUTF16.putChar(byte[] val, int index, int c)
1672 // static char StringUTF16.getChar(byte[] val, int index)
1673 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1674 Node* value = argument(0);
1675 Node* index = argument(1);
1676 Node* ch = is_store ? argument(2) : NULL;
1677
1678 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1679 // correctly requires matched array shapes.
1680 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1681 "sanity: byte[] and char[] bases agree");
1682 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1683 "sanity: byte[] and char[] scales agree");
1684
1685 // Bail when getChar over constants is requested: constant folding would
1686 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1687 // Java method would constant fold nicely instead.
1688 if (!is_store && value->is_Con() && index->is_Con()) {
1689 return false;
1690 }
1691
1692 Node* adr = array_element_address(value, index, T_CHAR);
1693 if (adr->is_top()) {
1694 return false;
1695 }
1696 if (is_store) {
1697 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1698 false, false, true /* mismatched */);
1699 } else {
1700 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1701 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1702 set_result(ch);
1703 }
1704 return true;
1705 }
1706
1707 //--------------------------round_double_node--------------------------------
1708 // Round a double node if necessary.
1709 Node* LibraryCallKit::round_double_node(Node* n) {
1710 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1711 n = _gvn.transform(new RoundDoubleNode(0, n));
1712 return n;
1713 }
1714
1715 //------------------------------inline_math-----------------------------------
1716 // public static double Math.abs(double)
1717 // public static double Math.sqrt(double)
1718 // public static double Math.log(double)
1719 // public static double Math.log10(double)
1720 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1721 Node* arg = round_double_node(argument(0));
1722 Node* n = NULL;
1723 switch (id) {
1724 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1725 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1726 default: fatal_unexpected_iid(id); break;
1727 }
1728 set_result(_gvn.transform(n));
1729 return true;
1730 }
1731
1732 //------------------------------runtime_math-----------------------------
1733 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1734 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1735 "must be (DD)D or (D)D type");
1736
1737 // Inputs
1738 Node* a = round_double_node(argument(0));
1739 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1740
1741 const TypePtr* no_memory_effects = NULL;
1742 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1743 no_memory_effects,
1744 a, top(), b, b ? top() : NULL);
1745 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1746 #ifdef ASSERT
1747 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1748 assert(value_top == top(), "second value must be top");
1749 #endif
1750
1751 set_result(value);
1752 return true;
1753 }
1754
1755 //------------------------------inline_math_native-----------------------------
1756 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1757 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1758 switch (id) {
1759 // These intrinsics are not properly supported on all hardware
1760 case vmIntrinsics::_dsin:
1761 return StubRoutines::dsin() != NULL ?
1762 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1763 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1764 case vmIntrinsics::_dcos:
1765 return StubRoutines::dcos() != NULL ?
1766 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1767 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1768 case vmIntrinsics::_dtan:
1769 return StubRoutines::dtan() != NULL ?
1770 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1771 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1772 case vmIntrinsics::_dlog:
1773 return StubRoutines::dlog() != NULL ?
1774 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1775 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1776 case vmIntrinsics::_dlog10:
1777 return StubRoutines::dlog10() != NULL ?
1778 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1779 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1780
1781 // These intrinsics are supported on all hardware
1782 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1783 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1784
1785 case vmIntrinsics::_dexp:
1786 return StubRoutines::dexp() != NULL ?
1787 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1788 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1789 case vmIntrinsics::_dpow:
1790 return StubRoutines::dpow() != NULL ?
1791 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1792 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1793 #undef FN_PTR
1794
1795 // These intrinsics are not yet correctly implemented
1796 case vmIntrinsics::_datan2:
1797 return false;
1798
1799 default:
1800 fatal_unexpected_iid(id);
1801 return false;
1802 }
1803 }
1804
1805 static bool is_simple_name(Node* n) {
1806 return (n->req() == 1 // constant
1807 || (n->is_Type() && n->as_Type()->type()->singleton())
1808 || n->is_Proj() // parameter or return value
1809 || n->is_Phi() // local of some sort
1810 );
1811 }
1812
1813 //----------------------------inline_notify-----------------------------------*
1814 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1815 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1816 address func;
1817 if (id == vmIntrinsics::_notify) {
1818 func = OptoRuntime::monitor_notify_Java();
1819 } else {
1820 func = OptoRuntime::monitor_notifyAll_Java();
1821 }
1822 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1823 make_slow_call_ex(call, env()->Throwable_klass(), false);
1824 return true;
1825 }
1826
1827
1828 //----------------------------inline_min_max-----------------------------------
1829 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1830 set_result(generate_min_max(id, argument(0), argument(1)));
1831 return true;
1832 }
1833
1834 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1835 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1836 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1837 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1838 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1839
1840 {
1841 PreserveJVMState pjvms(this);
1842 PreserveReexecuteState preexecs(this);
1843 jvms()->set_should_reexecute(true);
1844
1845 set_control(slow_path);
1846 set_i_o(i_o());
1847
1848 uncommon_trap(Deoptimization::Reason_intrinsic,
1849 Deoptimization::Action_none);
1850 }
1851
1852 set_control(fast_path);
1853 set_result(math);
1854 }
1855
1856 template <typename OverflowOp>
1857 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1858 typedef typename OverflowOp::MathOp MathOp;
1859
1860 MathOp* mathOp = new MathOp(arg1, arg2);
1861 Node* operation = _gvn.transform( mathOp );
1862 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1863 inline_math_mathExact(operation, ofcheck);
1864 return true;
1865 }
1866
1867 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1868 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1869 }
1870
1871 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1872 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1873 }
1874
1875 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1876 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1877 }
1878
1879 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1880 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1881 }
1882
1883 bool LibraryCallKit::inline_math_negateExactI() {
1884 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1885 }
1886
1887 bool LibraryCallKit::inline_math_negateExactL() {
1888 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1889 }
1890
1891 bool LibraryCallKit::inline_math_multiplyExactI() {
1892 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1893 }
1894
1895 bool LibraryCallKit::inline_math_multiplyExactL() {
1896 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1897 }
1898
1899 Node*
1900 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1901 // These are the candidate return value:
1902 Node* xvalue = x0;
1903 Node* yvalue = y0;
1904
1905 if (xvalue == yvalue) {
1906 return xvalue;
1907 }
1908
1909 bool want_max = (id == vmIntrinsics::_max);
1910
1911 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1912 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1913 if (txvalue == NULL || tyvalue == NULL) return top();
1914 // This is not really necessary, but it is consistent with a
1915 // hypothetical MaxINode::Value method:
1916 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1917
1918 // %%% This folding logic should (ideally) be in a different place.
1919 // Some should be inside IfNode, and there to be a more reliable
1920 // transformation of ?: style patterns into cmoves. We also want
1921 // more powerful optimizations around cmove and min/max.
1922
1923 // Try to find a dominating comparison of these guys.
1924 // It can simplify the index computation for Arrays.copyOf
1925 // and similar uses of System.arraycopy.
1926 // First, compute the normalized version of CmpI(x, y).
1927 int cmp_op = Op_CmpI;
1928 Node* xkey = xvalue;
1929 Node* ykey = yvalue;
1930 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1931 if (ideal_cmpxy->is_Cmp()) {
1932 // E.g., if we have CmpI(length - offset, count),
1933 // it might idealize to CmpI(length, count + offset)
1934 cmp_op = ideal_cmpxy->Opcode();
1935 xkey = ideal_cmpxy->in(1);
1936 ykey = ideal_cmpxy->in(2);
1937 }
1938
1939 // Start by locating any relevant comparisons.
1940 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1941 Node* cmpxy = NULL;
1942 Node* cmpyx = NULL;
1943 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1944 Node* cmp = start_from->fast_out(k);
1945 if (cmp->outcnt() > 0 && // must have prior uses
1946 cmp->in(0) == NULL && // must be context-independent
1947 cmp->Opcode() == cmp_op) { // right kind of compare
1948 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1949 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1950 }
1951 }
1952
1953 const int NCMPS = 2;
1954 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1955 int cmpn;
1956 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1957 if (cmps[cmpn] != NULL) break; // find a result
1958 }
1959 if (cmpn < NCMPS) {
1960 // Look for a dominating test that tells us the min and max.
1961 int depth = 0; // Limit search depth for speed
1962 Node* dom = control();
1963 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1964 if (++depth >= 100) break;
1965 Node* ifproj = dom;
1966 if (!ifproj->is_Proj()) continue;
1967 Node* iff = ifproj->in(0);
1968 if (!iff->is_If()) continue;
1969 Node* bol = iff->in(1);
1970 if (!bol->is_Bool()) continue;
1971 Node* cmp = bol->in(1);
1972 if (cmp == NULL) continue;
1973 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1974 if (cmps[cmpn] == cmp) break;
1975 if (cmpn == NCMPS) continue;
1976 BoolTest::mask btest = bol->as_Bool()->_test._test;
1977 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1978 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1979 // At this point, we know that 'x btest y' is true.
1980 switch (btest) {
1981 case BoolTest::eq:
1982 // They are proven equal, so we can collapse the min/max.
1983 // Either value is the answer. Choose the simpler.
1984 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1985 return yvalue;
1986 return xvalue;
1987 case BoolTest::lt: // x < y
1988 case BoolTest::le: // x <= y
1989 return (want_max ? yvalue : xvalue);
1990 case BoolTest::gt: // x > y
1991 case BoolTest::ge: // x >= y
1992 return (want_max ? xvalue : yvalue);
1993 }
1994 }
1995 }
1996
1997 // We failed to find a dominating test.
1998 // Let's pick a test that might GVN with prior tests.
1999 Node* best_bol = NULL;
2000 BoolTest::mask best_btest = BoolTest::illegal;
2001 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2002 Node* cmp = cmps[cmpn];
2003 if (cmp == NULL) continue;
2004 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2005 Node* bol = cmp->fast_out(j);
2006 if (!bol->is_Bool()) continue;
2007 BoolTest::mask btest = bol->as_Bool()->_test._test;
2008 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2009 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2010 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2011 best_bol = bol->as_Bool();
2012 best_btest = btest;
2013 }
2014 }
2015 }
2016
2017 Node* answer_if_true = NULL;
2018 Node* answer_if_false = NULL;
2019 switch (best_btest) {
2020 default:
2021 if (cmpxy == NULL)
2022 cmpxy = ideal_cmpxy;
2023 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2024 // and fall through:
2025 case BoolTest::lt: // x < y
2026 case BoolTest::le: // x <= y
2027 answer_if_true = (want_max ? yvalue : xvalue);
2028 answer_if_false = (want_max ? xvalue : yvalue);
2029 break;
2030 case BoolTest::gt: // x > y
2031 case BoolTest::ge: // x >= y
2032 answer_if_true = (want_max ? xvalue : yvalue);
2033 answer_if_false = (want_max ? yvalue : xvalue);
2034 break;
2035 }
2036
2037 jint hi, lo;
2038 if (want_max) {
2039 // We can sharpen the minimum.
2040 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2041 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2042 } else {
2043 // We can sharpen the maximum.
2044 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2045 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2046 }
2047
2048 // Use a flow-free graph structure, to avoid creating excess control edges
2049 // which could hinder other optimizations.
2050 // Since Math.min/max is often used with arraycopy, we want
2051 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2052 Node* cmov = CMoveNode::make(NULL, best_bol,
2053 answer_if_false, answer_if_true,
2054 TypeInt::make(lo, hi, widen));
2055
2056 return _gvn.transform(cmov);
2057
2058 /*
2059 // This is not as desirable as it may seem, since Min and Max
2060 // nodes do not have a full set of optimizations.
2061 // And they would interfere, anyway, with 'if' optimizations
2062 // and with CMoveI canonical forms.
2063 switch (id) {
2064 case vmIntrinsics::_min:
2065 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2066 case vmIntrinsics::_max:
2067 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2068 default:
2069 ShouldNotReachHere();
2070 }
2071 */
2072 }
2073
2074 inline int
2075 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2076 const TypePtr* base_type = TypePtr::NULL_PTR;
2077 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2078 if (base_type == NULL) {
2079 // Unknown type.
2080 return Type::AnyPtr;
2081 } else if (base_type == TypePtr::NULL_PTR) {
2082 // Since this is a NULL+long form, we have to switch to a rawptr.
2083 base = _gvn.transform(new CastX2PNode(offset));
2084 offset = MakeConX(0);
2085 return Type::RawPtr;
2086 } else if (base_type->base() == Type::RawPtr) {
2087 return Type::RawPtr;
2088 } else if (base_type->isa_oopptr()) {
2089 // Base is never null => always a heap address.
2090 if (base_type->ptr() == TypePtr::NotNull) {
2091 return Type::OopPtr;
2092 }
2093 // Offset is small => always a heap address.
2094 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2095 if (offset_type != NULL &&
2096 base_type->offset() == 0 && // (should always be?)
2097 offset_type->_lo >= 0 &&
2098 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2099 return Type::OopPtr;
2100 }
2101 // Otherwise, it might either be oop+off or NULL+addr.
2102 return Type::AnyPtr;
2103 } else {
2104 // No information:
2105 return Type::AnyPtr;
2106 }
2107 }
2108
2109 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2110 int kind = classify_unsafe_addr(base, offset);
2111 if (kind == Type::RawPtr) {
2112 return basic_plus_adr(top(), base, offset);
2113 } else {
2114 return basic_plus_adr(base, offset);
2115 }
2116 }
2117
2118 //--------------------------inline_number_methods-----------------------------
2119 // inline int Integer.numberOfLeadingZeros(int)
2120 // inline int Long.numberOfLeadingZeros(long)
2121 //
2122 // inline int Integer.numberOfTrailingZeros(int)
2123 // inline int Long.numberOfTrailingZeros(long)
2124 //
2125 // inline int Integer.bitCount(int)
2126 // inline int Long.bitCount(long)
2127 //
2128 // inline char Character.reverseBytes(char)
2129 // inline short Short.reverseBytes(short)
2130 // inline int Integer.reverseBytes(int)
2131 // inline long Long.reverseBytes(long)
2132 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2133 Node* arg = argument(0);
2134 Node* n = NULL;
2135 switch (id) {
2136 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2137 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2138 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2139 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2140 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2141 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2142 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2143 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2144 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2145 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2146 default: fatal_unexpected_iid(id); break;
2147 }
2148 set_result(_gvn.transform(n));
2149 return true;
2150 }
2151
2152 //----------------------------inline_unsafe_access----------------------------
2153
2154 // Helper that guards and inserts a pre-barrier.
2155 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2156 Node* pre_val, bool need_mem_bar) {
2157 // We could be accessing the referent field of a reference object. If so, when G1
2158 // is enabled, we need to log the value in the referent field in an SATB buffer.
2159 // This routine performs some compile time filters and generates suitable
2160 // runtime filters that guard the pre-barrier code.
2161 // Also add memory barrier for non volatile load from the referent field
2162 // to prevent commoning of loads across safepoint.
2163 if (!UseG1GC && !need_mem_bar)
2164 return;
2165
2166 // Some compile time checks.
2167
2168 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2169 const TypeX* otype = offset->find_intptr_t_type();
2170 if (otype != NULL && otype->is_con() &&
2171 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2172 // Constant offset but not the reference_offset so just return
2173 return;
2174 }
2175
2176 // We only need to generate the runtime guards for instances.
2177 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2178 if (btype != NULL) {
2179 if (btype->isa_aryptr()) {
2180 // Array type so nothing to do
2181 return;
2182 }
2183
2184 const TypeInstPtr* itype = btype->isa_instptr();
2185 if (itype != NULL) {
2186 // Can the klass of base_oop be statically determined to be
2187 // _not_ a sub-class of Reference and _not_ Object?
2188 ciKlass* klass = itype->klass();
2189 if ( klass->is_loaded() &&
2190 !klass->is_subtype_of(env()->Reference_klass()) &&
2191 !env()->Object_klass()->is_subtype_of(klass)) {
2192 return;
2193 }
2194 }
2195 }
2196
2197 // The compile time filters did not reject base_oop/offset so
2198 // we need to generate the following runtime filters
2199 //
2200 // if (offset == java_lang_ref_Reference::_reference_offset) {
2201 // if (instance_of(base, java.lang.ref.Reference)) {
2202 // pre_barrier(_, pre_val, ...);
2203 // }
2204 // }
2205
2206 float likely = PROB_LIKELY( 0.999);
2207 float unlikely = PROB_UNLIKELY(0.999);
2208
2209 IdealKit ideal(this);
2210 #define __ ideal.
2211
2212 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2213
2214 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2215 // Update graphKit memory and control from IdealKit.
2216 sync_kit(ideal);
2217
2218 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2219 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2220
2221 // Update IdealKit memory and control from graphKit.
2222 __ sync_kit(this);
2223
2224 Node* one = __ ConI(1);
2225 // is_instof == 0 if base_oop == NULL
2226 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2227
2228 // Update graphKit from IdeakKit.
2229 sync_kit(ideal);
2230
2231 // Use the pre-barrier to record the value in the referent field
2232 pre_barrier(false /* do_load */,
2233 __ ctrl(),
2234 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2235 pre_val /* pre_val */,
2236 T_OBJECT);
2237 if (need_mem_bar) {
2238 // Add memory barrier to prevent commoning reads from this field
2239 // across safepoint since GC can change its value.
2240 insert_mem_bar(Op_MemBarCPUOrder);
2241 }
2242 // Update IdealKit from graphKit.
2243 __ sync_kit(this);
2244
2245 } __ end_if(); // _ref_type != ref_none
2246 } __ end_if(); // offset == referent_offset
2247
2248 // Final sync IdealKit and GraphKit.
2249 final_sync(ideal);
2250 #undef __
2251 }
2252
2253
2254 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2255 // Attempt to infer a sharper value type from the offset and base type.
2256 ciKlass* sharpened_klass = NULL;
2257
2258 // See if it is an instance field, with an object type.
2259 if (alias_type->field() != NULL) {
2260 if (alias_type->field()->type()->is_klass()) {
2261 sharpened_klass = alias_type->field()->type()->as_klass();
2262 }
2263 }
2264
2265 // See if it is a narrow oop array.
2266 if (adr_type->isa_aryptr()) {
2267 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2268 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2269 if (elem_type != NULL) {
2270 sharpened_klass = elem_type->klass();
2271 }
2272 }
2273 }
2274
2275 // The sharpened class might be unloaded if there is no class loader
2276 // contraint in place.
2277 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2278 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2279
2280 #ifndef PRODUCT
2281 if (C->print_intrinsics() || C->print_inlining()) {
2282 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2283 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2284 }
2285 #endif
2286 // Sharpen the value type.
2287 return tjp;
2288 }
2289 return NULL;
2290 }
2291
2292 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2293 if (callee()->is_static()) return false; // caller must have the capability!
2294 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2295 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2296 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2297
2298 #ifndef PRODUCT
2299 {
2300 ResourceMark rm;
2301 // Check the signatures.
2302 ciSignature* sig = callee()->signature();
2303 #ifdef ASSERT
2304 if (!is_store) {
2305 // Object getObject(Object base, int/long offset), etc.
2306 BasicType rtype = sig->return_type()->basic_type();
2307 assert(rtype == type, "getter must return the expected value");
2308 assert(sig->count() == 2, "oop getter has 2 arguments");
2309 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2310 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2311 } else {
2312 // void putObject(Object base, int/long offset, Object x), etc.
2313 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2314 assert(sig->count() == 3, "oop putter has 3 arguments");
2315 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2316 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2317 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2318 assert(vtype == type, "putter must accept the expected value");
2319 }
2320 #endif // ASSERT
2321 }
2322 #endif //PRODUCT
2323
2324 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2325
2326 Node* receiver = argument(0); // type: oop
2327
2328 // Build address expression.
2329 Node* adr;
2330 Node* heap_base_oop = top();
2331 Node* offset = top();
2332 Node* val;
2333
2334 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2335 Node* base = argument(1); // type: oop
2336 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2337 offset = argument(2); // type: long
2338 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2339 // to be plain byte offsets, which are also the same as those accepted
2340 // by oopDesc::field_base.
2341 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2342 "fieldOffset must be byte-scaled");
2343 // 32-bit machines ignore the high half!
2344 offset = ConvL2X(offset);
2345 adr = make_unsafe_address(base, offset);
2346 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2347 heap_base_oop = base;
2348 } else if (type == T_OBJECT) {
2349 return false; // off-heap oop accesses are not supported
2350 }
2351
2352 // Can base be NULL? Otherwise, always on-heap access.
2353 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2354
2355 val = is_store ? argument(4) : NULL;
2356
2357 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2358
2359 // Try to categorize the address.
2360 Compile::AliasType* alias_type = C->alias_type(adr_type);
2361 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2362
2363 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2364 alias_type->adr_type() == TypeAryPtr::RANGE) {
2365 return false; // not supported
2366 }
2367
2368 bool mismatched = false;
2369 BasicType bt = alias_type->basic_type();
2370 if (bt != T_ILLEGAL) {
2371 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2372 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2373 // Alias type doesn't differentiate between byte[] and boolean[]).
2374 // Use address type to get the element type.
2375 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2376 }
2377 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2378 // accessing an array field with getObject is not a mismatch
2379 bt = T_OBJECT;
2380 }
2381 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2382 // Don't intrinsify mismatched object accesses
2383 return false;
2384 }
2385 mismatched = (bt != type);
2386 } else if (alias_type->adr_type()->isa_oopptr()) {
2387 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2388 }
2389
2390 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2391
2392 // First guess at the value type.
2393 const Type *value_type = Type::get_const_basic_type(type);
2394
2395 // We will need memory barriers unless we can determine a unique
2396 // alias category for this reference. (Note: If for some reason
2397 // the barriers get omitted and the unsafe reference begins to "pollute"
2398 // the alias analysis of the rest of the graph, either Compile::can_alias
2399 // or Compile::must_alias will throw a diagnostic assert.)
2400 bool need_mem_bar = false;
2401 switch (kind) {
2402 case Relaxed:
2403 need_mem_bar = mismatched && !adr_type->isa_aryptr();
2404 break;
2405 case Opaque:
2406 // Opaque uses CPUOrder membars for protection against code movement.
2407 case Acquire:
2408 case Release:
2409 case Volatile:
2410 need_mem_bar = true;
2411 break;
2412 default:
2413 ShouldNotReachHere();
2414 }
2415
2416 // Some accesses require access atomicity for all types, notably longs and doubles.
2417 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2418 bool requires_atomic_access = false;
2419 switch (kind) {
2420 case Relaxed:
2421 requires_atomic_access = AlwaysAtomicAccesses;
2422 break;
2423 case Opaque:
2424 // Opaque accesses are atomic.
2425 case Acquire:
2426 case Release:
2427 case Volatile:
2428 requires_atomic_access = true;
2429 break;
2430 default:
2431 ShouldNotReachHere();
2432 }
2433
2434 // Figure out the memory ordering.
2435 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2436 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2437
2438 // If we are reading the value of the referent field of a Reference
2439 // object (either by using Unsafe directly or through reflection)
2440 // then, if G1 is enabled, we need to record the referent in an
2441 // SATB log buffer using the pre-barrier mechanism.
2442 // Also we need to add memory barrier to prevent commoning reads
2443 // from this field across safepoint since GC can change its value.
2444 bool need_read_barrier = !is_store &&
2445 offset != top() && heap_base_oop != top();
2446
2447 if (!is_store && type == T_OBJECT) {
2448 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2449 if (tjp != NULL) {
2450 value_type = tjp;
2451 }
2452 }
2453
2454 receiver = null_check(receiver);
2455 if (stopped()) {
2456 return true;
2457 }
2458 // Heap pointers get a null-check from the interpreter,
2459 // as a courtesy. However, this is not guaranteed by Unsafe,
2460 // and it is not possible to fully distinguish unintended nulls
2461 // from intended ones in this API.
2462
2463 // We need to emit leading and trailing CPU membars (see below) in
2464 // addition to memory membars for special access modes. This is a little
2465 // too strong, but avoids the need to insert per-alias-type
2466 // volatile membars (for stores; compare Parse::do_put_xxx), which
2467 // we cannot do effectively here because we probably only have a
2468 // rough approximation of type.
2469
2470 switch(kind) {
2471 case Relaxed:
2472 case Opaque:
2473 case Acquire:
2474 break;
2475 case Release:
2476 case Volatile:
2477 if (is_store) {
2478 insert_mem_bar(Op_MemBarRelease);
2479 } else {
2480 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2481 insert_mem_bar(Op_MemBarVolatile);
2482 }
2483 }
2484 break;
2485 default:
2486 ShouldNotReachHere();
2487 }
2488
2489 // Memory barrier to prevent normal and 'unsafe' accesses from
2490 // bypassing each other. Happens after null checks, so the
2491 // exception paths do not take memory state from the memory barrier,
2492 // so there's no problems making a strong assert about mixing users
2493 // of safe & unsafe memory.
2494 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2495
2496 if (!is_store) {
2497 Node* p = NULL;
2498 // Try to constant fold a load from a constant field
2499 ciField* field = alias_type->field();
2500 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2501 // final or stable field
2502 p = make_constant_from_field(field, heap_base_oop);
2503 }
2504 if (p == NULL) {
2505 // To be valid, unsafe loads may depend on other conditions than
2506 // the one that guards them: pin the Load node
2507 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched);
2508 // load value
2509 switch (type) {
2510 case T_BOOLEAN:
2511 {
2512 // Normalize the value returned by getBoolean in the following cases
2513 if (mismatched ||
2514 heap_base_oop == top() || // - heap_base_oop is NULL or
2515 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2516 // and the unsafe access is made to large offset
2517 // (i.e., larger than the maximum offset necessary for any
2518 // field access)
2519 ) {
2520 IdealKit ideal = IdealKit(this);
2521 #define __ ideal.
2522 IdealVariable normalized_result(ideal);
2523 __ declarations_done();
2524 __ set(normalized_result, p);
2525 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2526 __ set(normalized_result, ideal.ConI(1));
2527 ideal.end_if();
2528 final_sync(ideal);
2529 p = __ value(normalized_result);
2530 #undef __
2531 }
2532 }
2533 case T_CHAR:
2534 case T_BYTE:
2535 case T_SHORT:
2536 case T_INT:
2537 case T_LONG:
2538 case T_FLOAT:
2539 case T_DOUBLE:
2540 break;
2541 case T_OBJECT:
2542 if (need_read_barrier) {
2543 // We do not require a mem bar inside pre_barrier if need_mem_bar
2544 // is set: the barriers would be emitted by us.
2545 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2546 }
2547 break;
2548 case T_ADDRESS:
2549 // Cast to an int type.
2550 p = _gvn.transform(new CastP2XNode(NULL, p));
2551 p = ConvX2UL(p);
2552 break;
2553 default:
2554 fatal("unexpected type %d: %s", type, type2name(type));
2555 break;
2556 }
2557 }
2558 // The load node has the control of the preceding MemBarCPUOrder. All
2559 // following nodes will have the control of the MemBarCPUOrder inserted at
2560 // the end of this method. So, pushing the load onto the stack at a later
2561 // point is fine.
2562 set_result(p);
2563 } else {
2564 // place effect of store into memory
2565 switch (type) {
2566 case T_DOUBLE:
2567 val = dstore_rounding(val);
2568 break;
2569 case T_ADDRESS:
2570 // Repackage the long as a pointer.
2571 val = ConvL2X(val);
2572 val = _gvn.transform(new CastX2PNode(val));
2573 break;
2574 }
2575
2576 if (type == T_OBJECT) {
2577 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2578 } else {
2579 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2580 }
2581 }
2582
2583 switch(kind) {
2584 case Relaxed:
2585 case Opaque:
2586 case Release:
2587 break;
2588 case Acquire:
2589 case Volatile:
2590 if (!is_store) {
2591 insert_mem_bar(Op_MemBarAcquire);
2592 } else {
2593 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2594 insert_mem_bar(Op_MemBarVolatile);
2595 }
2596 }
2597 break;
2598 default:
2599 ShouldNotReachHere();
2600 }
2601
2602 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2603
2604 return true;
2605 }
2606
2607 //----------------------------inline_unsafe_load_store----------------------------
2608 // This method serves a couple of different customers (depending on LoadStoreKind):
2609 //
2610 // LS_cmp_swap:
2611 //
2612 // boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2613 // boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2614 // boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2615 //
2616 // LS_cmp_swap_weak:
2617 //
2618 // boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x);
2619 // boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x);
2620 // boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x);
2621 //
2622 // boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x);
2623 // boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x);
2624 // boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x);
2625 //
2626 // boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x);
2627 // boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x);
2628 // boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x);
2629 //
2630 // LS_cmp_exchange:
2631 //
2632 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2633 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2634 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2635 //
2636 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2637 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2638 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2639 //
2640 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2641 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2642 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2643 //
2644 // LS_get_add:
2645 //
2646 // int getAndAddInt( Object o, long offset, int delta)
2647 // long getAndAddLong(Object o, long offset, long delta)
2648 //
2649 // LS_get_set:
2650 //
2651 // int getAndSet(Object o, long offset, int newValue)
2652 // long getAndSet(Object o, long offset, long newValue)
2653 // Object getAndSet(Object o, long offset, Object newValue)
2654 //
2655 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2656 // This basic scheme here is the same as inline_unsafe_access, but
2657 // differs in enough details that combining them would make the code
2658 // overly confusing. (This is a true fact! I originally combined
2659 // them, but even I was confused by it!) As much code/comments as
2660 // possible are retained from inline_unsafe_access though to make
2661 // the correspondences clearer. - dl
2662
2663 if (callee()->is_static()) return false; // caller must have the capability!
2664
2665 #ifndef PRODUCT
2666 BasicType rtype;
2667 {
2668 ResourceMark rm;
2669 // Check the signatures.
2670 ciSignature* sig = callee()->signature();
2671 rtype = sig->return_type()->basic_type();
2672 switch(kind) {
2673 case LS_get_add:
2674 case LS_get_set: {
2675 // Check the signatures.
2676 #ifdef ASSERT
2677 assert(rtype == type, "get and set must return the expected type");
2678 assert(sig->count() == 3, "get and set has 3 arguments");
2679 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2680 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2681 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2682 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2683 #endif // ASSERT
2684 break;
2685 }
2686 case LS_cmp_swap:
2687 case LS_cmp_swap_weak: {
2688 // Check the signatures.
2689 #ifdef ASSERT
2690 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2691 assert(sig->count() == 4, "CAS has 4 arguments");
2692 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2693 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2694 #endif // ASSERT
2695 break;
2696 }
2697 case LS_cmp_exchange: {
2698 // Check the signatures.
2699 #ifdef ASSERT
2700 assert(rtype == type, "CAS must return the expected type");
2701 assert(sig->count() == 4, "CAS has 4 arguments");
2702 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2703 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2704 #endif // ASSERT
2705 break;
2706 }
2707 default:
2708 ShouldNotReachHere();
2709 }
2710 }
2711 #endif //PRODUCT
2712
2713 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2714
2715 // Get arguments:
2716 Node* receiver = NULL;
2717 Node* base = NULL;
2718 Node* offset = NULL;
2719 Node* oldval = NULL;
2720 Node* newval = NULL;
2721 switch(kind) {
2722 case LS_cmp_swap:
2723 case LS_cmp_swap_weak:
2724 case LS_cmp_exchange: {
2725 const bool two_slot_type = type2size[type] == 2;
2726 receiver = argument(0); // type: oop
2727 base = argument(1); // type: oop
2728 offset = argument(2); // type: long
2729 oldval = argument(4); // type: oop, int, or long
2730 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2731 break;
2732 }
2733 case LS_get_add:
2734 case LS_get_set: {
2735 receiver = argument(0); // type: oop
2736 base = argument(1); // type: oop
2737 offset = argument(2); // type: long
2738 oldval = NULL;
2739 newval = argument(4); // type: oop, int, or long
2740 break;
2741 }
2742 default:
2743 ShouldNotReachHere();
2744 }
2745
2746 // Build field offset expression.
2747 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2748 // to be plain byte offsets, which are also the same as those accepted
2749 // by oopDesc::field_base.
2750 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2751 // 32-bit machines ignore the high half of long offsets
2752 offset = ConvL2X(offset);
2753 Node* adr = make_unsafe_address(base, offset);
2754 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2755
2756 Compile::AliasType* alias_type = C->alias_type(adr_type);
2757 BasicType bt = alias_type->basic_type();
2758 if (bt != T_ILLEGAL &&
2759 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2760 // Don't intrinsify mismatched object accesses.
2761 return false;
2762 }
2763
2764 // For CAS, unlike inline_unsafe_access, there seems no point in
2765 // trying to refine types. Just use the coarse types here.
2766 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2767 const Type *value_type = Type::get_const_basic_type(type);
2768
2769 switch (kind) {
2770 case LS_get_set:
2771 case LS_cmp_exchange: {
2772 if (type == T_OBJECT) {
2773 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2774 if (tjp != NULL) {
2775 value_type = tjp;
2776 }
2777 }
2778 break;
2779 }
2780 case LS_cmp_swap:
2781 case LS_cmp_swap_weak:
2782 case LS_get_add:
2783 break;
2784 default:
2785 ShouldNotReachHere();
2786 }
2787
2788 // Null check receiver.
2789 receiver = null_check(receiver);
2790 if (stopped()) {
2791 return true;
2792 }
2793
2794 int alias_idx = C->get_alias_index(adr_type);
2795
2796 // Memory-model-wise, a LoadStore acts like a little synchronized
2797 // block, so needs barriers on each side. These don't translate
2798 // into actual barriers on most machines, but we still need rest of
2799 // compiler to respect ordering.
2800
2801 switch (access_kind) {
2802 case Relaxed:
2803 case Acquire:
2804 break;
2805 case Release:
2806 insert_mem_bar(Op_MemBarRelease);
2807 break;
2808 case Volatile:
2809 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2810 insert_mem_bar(Op_MemBarVolatile);
2811 } else {
2812 insert_mem_bar(Op_MemBarRelease);
2813 }
2814 break;
2815 default:
2816 ShouldNotReachHere();
2817 }
2818 insert_mem_bar(Op_MemBarCPUOrder);
2819
2820 // Figure out the memory ordering.
2821 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2822
2823 // 4984716: MemBars must be inserted before this
2824 // memory node in order to avoid a false
2825 // dependency which will confuse the scheduler.
2826 Node *mem = memory(alias_idx);
2827
2828 // For now, we handle only those cases that actually exist: ints,
2829 // longs, and Object. Adding others should be straightforward.
2830 Node* load_store = NULL;
2831 switch(type) {
2832 case T_BYTE:
2833 switch(kind) {
2834 case LS_get_add:
2835 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2836 break;
2837 case LS_get_set:
2838 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2839 break;
2840 case LS_cmp_swap_weak:
2841 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2842 break;
2843 case LS_cmp_swap:
2844 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2845 break;
2846 case LS_cmp_exchange:
2847 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2848 break;
2849 default:
2850 ShouldNotReachHere();
2851 }
2852 break;
2853 case T_SHORT:
2854 switch(kind) {
2855 case LS_get_add:
2856 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2857 break;
2858 case LS_get_set:
2859 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2860 break;
2861 case LS_cmp_swap_weak:
2862 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2863 break;
2864 case LS_cmp_swap:
2865 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2866 break;
2867 case LS_cmp_exchange:
2868 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2869 break;
2870 default:
2871 ShouldNotReachHere();
2872 }
2873 break;
2874 case T_INT:
2875 switch(kind) {
2876 case LS_get_add:
2877 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2878 break;
2879 case LS_get_set:
2880 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2881 break;
2882 case LS_cmp_swap_weak:
2883 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2884 break;
2885 case LS_cmp_swap:
2886 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2887 break;
2888 case LS_cmp_exchange:
2889 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2890 break;
2891 default:
2892 ShouldNotReachHere();
2893 }
2894 break;
2895 case T_LONG:
2896 switch(kind) {
2897 case LS_get_add:
2898 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2899 break;
2900 case LS_get_set:
2901 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2902 break;
2903 case LS_cmp_swap_weak:
2904 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2905 break;
2906 case LS_cmp_swap:
2907 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2908 break;
2909 case LS_cmp_exchange:
2910 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2911 break;
2912 default:
2913 ShouldNotReachHere();
2914 }
2915 break;
2916 case T_OBJECT:
2917 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2918 // could be delayed during Parse (for example, in adjust_map_after_if()).
2919 // Execute transformation here to avoid barrier generation in such case.
2920 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2921 newval = _gvn.makecon(TypePtr::NULL_PTR);
2922
2923 // Reference stores need a store barrier.
2924 switch(kind) {
2925 case LS_get_set: {
2926 // If pre-barrier must execute before the oop store, old value will require do_load here.
2927 if (!can_move_pre_barrier()) {
2928 pre_barrier(true /* do_load*/,
2929 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2930 NULL /* pre_val*/,
2931 T_OBJECT);
2932 } // Else move pre_barrier to use load_store value, see below.
2933 break;
2934 }
2935 case LS_cmp_swap_weak:
2936 case LS_cmp_swap:
2937 case LS_cmp_exchange: {
2938 // Same as for newval above:
2939 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2940 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2941 }
2942 // The only known value which might get overwritten is oldval.
2943 pre_barrier(false /* do_load */,
2944 control(), NULL, NULL, max_juint, NULL, NULL,
2945 oldval /* pre_val */,
2946 T_OBJECT);
2947 break;
2948 }
2949 default:
2950 ShouldNotReachHere();
2951 }
2952
2953 #ifdef _LP64
2954 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2955 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2956
2957 switch(kind) {
2958 case LS_get_set:
2959 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
2960 break;
2961 case LS_cmp_swap_weak: {
2962 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2963 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2964 break;
2965 }
2966 case LS_cmp_swap: {
2967 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2968 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
2969 break;
2970 }
2971 case LS_cmp_exchange: {
2972 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2973 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
2974 break;
2975 }
2976 default:
2977 ShouldNotReachHere();
2978 }
2979 } else
2980 #endif
2981 switch (kind) {
2982 case LS_get_set:
2983 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2984 break;
2985 case LS_cmp_swap_weak:
2986 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2987 break;
2988 case LS_cmp_swap:
2989 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
2990 break;
2991 case LS_cmp_exchange:
2992 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
2993 break;
2994 default:
2995 ShouldNotReachHere();
2996 }
2997
2998 // Emit the post barrier only when the actual store happened. This makes sense
2999 // to check only for LS_cmp_* that can fail to set the value.
3000 // LS_cmp_exchange does not produce any branches by default, so there is no
3001 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3002 // CompareAndExchange and move branches here, it would make sense to conditionalize
3003 // post_barriers for LS_cmp_exchange as well.
3004 //
3005 // CAS success path is marked more likely since we anticipate this is a performance
3006 // critical path, while CAS failure path can use the penalty for going through unlikely
3007 // path as backoff. Which is still better than doing a store barrier there.
3008 switch (kind) {
3009 case LS_get_set:
3010 case LS_cmp_exchange: {
3011 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3012 break;
3013 }
3014 case LS_cmp_swap_weak:
3015 case LS_cmp_swap: {
3016 IdealKit ideal(this);
3017 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3018 sync_kit(ideal);
3019 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3020 ideal.sync_kit(this);
3021 } ideal.end_if();
3022 final_sync(ideal);
3023 break;
3024 }
3025 default:
3026 ShouldNotReachHere();
3027 }
3028 break;
3029 default:
3030 fatal("unexpected type %d: %s", type, type2name(type));
3031 break;
3032 }
3033
3034 // SCMemProjNodes represent the memory state of a LoadStore. Their
3035 // main role is to prevent LoadStore nodes from being optimized away
3036 // when their results aren't used.
3037 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3038 set_memory(proj, alias_idx);
3039
3040 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3041 #ifdef _LP64
3042 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3043 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3044 }
3045 #endif
3046 if (can_move_pre_barrier() && kind == LS_get_set) {
3047 // Don't need to load pre_val. The old value is returned by load_store.
3048 // The pre_barrier can execute after the xchg as long as no safepoint
3049 // gets inserted between them.
3050 pre_barrier(false /* do_load */,
3051 control(), NULL, NULL, max_juint, NULL, NULL,
3052 load_store /* pre_val */,
3053 T_OBJECT);
3054 }
3055 }
3056
3057 // Add the trailing membar surrounding the access
3058 insert_mem_bar(Op_MemBarCPUOrder);
3059
3060 switch (access_kind) {
3061 case Relaxed:
3062 case Release:
3063 break; // do nothing
3064 case Acquire:
3065 case Volatile:
3066 insert_mem_bar(Op_MemBarAcquire);
3067 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3068 break;
3069 default:
3070 ShouldNotReachHere();
3071 }
3072
3073 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3074 set_result(load_store);
3075 return true;
3076 }
3077
3078 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3079 MemNode::MemOrd mo = MemNode::unset;
3080 switch(kind) {
3081 case Opaque:
3082 case Relaxed: mo = MemNode::unordered; break;
3083 case Acquire: mo = MemNode::acquire; break;
3084 case Release: mo = MemNode::release; break;
3085 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3086 default:
3087 ShouldNotReachHere();
3088 }
3089 guarantee(mo != MemNode::unset, "Should select memory ordering");
3090 return mo;
3091 }
3092
3093 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3094 MemNode::MemOrd mo = MemNode::unset;
3095 switch(kind) {
3096 case Opaque:
3097 case Relaxed: mo = MemNode::unordered; break;
3098 case Acquire: mo = MemNode::acquire; break;
3099 case Release: mo = MemNode::release; break;
3100 case Volatile: mo = MemNode::seqcst; break;
3101 default:
3102 ShouldNotReachHere();
3103 }
3104 guarantee(mo != MemNode::unset, "Should select memory ordering");
3105 return mo;
3106 }
3107
3108 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3109 // Regardless of form, don't allow previous ld/st to move down,
3110 // then issue acquire, release, or volatile mem_bar.
3111 insert_mem_bar(Op_MemBarCPUOrder);
3112 switch(id) {
3113 case vmIntrinsics::_loadFence:
3114 insert_mem_bar(Op_LoadFence);
3115 return true;
3116 case vmIntrinsics::_storeFence:
3117 insert_mem_bar(Op_StoreFence);
3118 return true;
3119 case vmIntrinsics::_fullFence:
3120 insert_mem_bar(Op_MemBarVolatile);
3121 return true;
3122 default:
3123 fatal_unexpected_iid(id);
3124 return false;
3125 }
3126 }
3127
3128 bool LibraryCallKit::inline_onspinwait() {
3129 insert_mem_bar(Op_OnSpinWait);
3130 return true;
3131 }
3132
3133 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3134 if (!kls->is_Con()) {
3135 return true;
3136 }
3137 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3138 if (klsptr == NULL) {
3139 return true;
3140 }
3141 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3142 // don't need a guard for a klass that is already initialized
3143 return !ik->is_initialized();
3144 }
3145
3146 //----------------------------inline_unsafe_allocate---------------------------
3147 // public native Object Unsafe.allocateInstance(Class<?> cls);
3148 bool LibraryCallKit::inline_unsafe_allocate() {
3149 if (callee()->is_static()) return false; // caller must have the capability!
3150
3151 null_check_receiver(); // null-check, then ignore
3152 Node* cls = null_check(argument(1));
3153 if (stopped()) return true;
3154
3155 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3156 kls = null_check(kls);
3157 if (stopped()) return true; // argument was like int.class
3158
3159 Node* test = NULL;
3160 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3161 // Note: The argument might still be an illegal value like
3162 // Serializable.class or Object[].class. The runtime will handle it.
3163 // But we must make an explicit check for initialization.
3164 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3165 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3166 // can generate code to load it as unsigned byte.
3167 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3168 Node* bits = intcon(InstanceKlass::fully_initialized);
3169 test = _gvn.transform(new SubINode(inst, bits));
3170 // The 'test' is non-zero if we need to take a slow path.
3171 }
3172
3173 Node* obj = new_instance(kls, test);
3174 set_result(obj);
3175 return true;
3176 }
3177
3178 //------------------------inline_native_time_funcs--------------
3179 // inline code for System.currentTimeMillis() and System.nanoTime()
3180 // these have the same type and signature
3181 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3182 const TypeFunc* tf = OptoRuntime::void_long_Type();
3183 const TypePtr* no_memory_effects = NULL;
3184 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3185 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3186 #ifdef ASSERT
3187 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3188 assert(value_top == top(), "second value must be top");
3189 #endif
3190 set_result(value);
3191 return true;
3192 }
3193
3194 #ifdef TRACE_HAVE_INTRINSICS
3195
3196 /*
3197 * oop -> myklass
3198 * myklass->trace_id |= USED
3199 * return myklass->trace_id & ~0x3
3200 */
3201 bool LibraryCallKit::inline_native_classID() {
3202 Node* cls = null_check(argument(0), T_OBJECT);
3203 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3204 kls = null_check(kls, T_OBJECT);
3205
3206 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3207 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3208 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3209
3210 Node* clsused = longcon(0x01l); // set the class bit
3211 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3212 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3213 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3214
3215 #ifdef TRACE_ID_META_BITS
3216 Node* mbits = longcon(~TRACE_ID_META_BITS);
3217 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3218 #endif
3219 #ifdef TRACE_ID_CLASS_SHIFT
3220 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3221 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3222 #endif
3223
3224 set_result(tvalue);
3225 return true;
3226
3227 }
3228
3229 bool LibraryCallKit::inline_native_getBufferWriter() {
3230 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3231
3232 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3233 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3234 );
3235
3236 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3237
3238 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3239 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3240
3241 IfNode* iff_jobj_null =
3242 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3243
3244 enum { _normal_path = 1,
3245 _null_path = 2,
3246 PATH_LIMIT };
3247
3248 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3249 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3250
3251 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3252 result_rgn->init_req(_null_path, jobj_is_null);
3253 result_val->init_req(_null_path, null());
3254
3255 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3256 result_rgn->init_req(_normal_path, jobj_is_not_null);
3257
3258 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3259 result_val->init_req(_normal_path, res);
3260
3261 set_result(result_rgn, result_val);
3262
3263 return true;
3264 }
3265
3266 #endif
3267
3268 //------------------------inline_native_currentThread------------------
3269 bool LibraryCallKit::inline_native_currentThread() {
3270 Node* junk = NULL;
3271 set_result(generate_current_thread(junk));
3272 return true;
3273 }
3274
3275 //------------------------inline_native_isInterrupted------------------
3276 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3277 bool LibraryCallKit::inline_native_isInterrupted() {
3278 // Add a fast path to t.isInterrupted(clear_int):
3279 // (t == Thread.current() &&
3280 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3281 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3282 // So, in the common case that the interrupt bit is false,
3283 // we avoid making a call into the VM. Even if the interrupt bit
3284 // is true, if the clear_int argument is false, we avoid the VM call.
3285 // However, if the receiver is not currentThread, we must call the VM,
3286 // because there must be some locking done around the operation.
3287
3288 // We only go to the fast case code if we pass two guards.
3289 // Paths which do not pass are accumulated in the slow_region.
3290
3291 enum {
3292 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3293 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3294 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3295 PATH_LIMIT
3296 };
3297
3298 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3299 // out of the function.
3300 insert_mem_bar(Op_MemBarCPUOrder);
3301
3302 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3303 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3304
3305 RegionNode* slow_region = new RegionNode(1);
3306 record_for_igvn(slow_region);
3307
3308 // (a) Receiving thread must be the current thread.
3309 Node* rec_thr = argument(0);
3310 Node* tls_ptr = NULL;
3311 Node* cur_thr = generate_current_thread(tls_ptr);
3312 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3313 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3314
3315 generate_slow_guard(bol_thr, slow_region);
3316
3317 // (b) Interrupt bit on TLS must be false.
3318 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3319 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3320 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3321
3322 // Set the control input on the field _interrupted read to prevent it floating up.
3323 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3324 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3325 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3326
3327 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3328
3329 // First fast path: if (!TLS._interrupted) return false;
3330 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3331 result_rgn->init_req(no_int_result_path, false_bit);
3332 result_val->init_req(no_int_result_path, intcon(0));
3333
3334 // drop through to next case
3335 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3336
3337 #ifndef _WINDOWS
3338 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3339 Node* clr_arg = argument(1);
3340 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3341 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3342 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3343
3344 // Second fast path: ... else if (!clear_int) return true;
3345 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3346 result_rgn->init_req(no_clear_result_path, false_arg);
3347 result_val->init_req(no_clear_result_path, intcon(1));
3348
3349 // drop through to next case
3350 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3351 #else
3352 // To return true on Windows you must read the _interrupted field
3353 // and check the event state i.e. take the slow path.
3354 #endif // _WINDOWS
3355
3356 // (d) Otherwise, go to the slow path.
3357 slow_region->add_req(control());
3358 set_control( _gvn.transform(slow_region));
3359
3360 if (stopped()) {
3361 // There is no slow path.
3362 result_rgn->init_req(slow_result_path, top());
3363 result_val->init_req(slow_result_path, top());
3364 } else {
3365 // non-virtual because it is a private non-static
3366 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3367
3368 Node* slow_val = set_results_for_java_call(slow_call);
3369 // this->control() comes from set_results_for_java_call
3370
3371 Node* fast_io = slow_call->in(TypeFunc::I_O);
3372 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3373
3374 // These two phis are pre-filled with copies of of the fast IO and Memory
3375 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3376 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3377
3378 result_rgn->init_req(slow_result_path, control());
3379 result_io ->init_req(slow_result_path, i_o());
3380 result_mem->init_req(slow_result_path, reset_memory());
3381 result_val->init_req(slow_result_path, slow_val);
3382
3383 set_all_memory(_gvn.transform(result_mem));
3384 set_i_o( _gvn.transform(result_io));
3385 }
3386
3387 C->set_has_split_ifs(true); // Has chance for split-if optimization
3388 set_result(result_rgn, result_val);
3389 return true;
3390 }
3391
3392 //---------------------------load_mirror_from_klass----------------------------
3393 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3394 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3395 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3396 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3397 }
3398
3399 //-----------------------load_klass_from_mirror_common-------------------------
3400 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3401 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3402 // and branch to the given path on the region.
3403 // If never_see_null, take an uncommon trap on null, so we can optimistically
3404 // compile for the non-null case.
3405 // If the region is NULL, force never_see_null = true.
3406 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3407 bool never_see_null,
3408 RegionNode* region,
3409 int null_path,
3410 int offset) {
3411 if (region == NULL) never_see_null = true;
3412 Node* p = basic_plus_adr(mirror, offset);
3413 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3414 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3415 Node* null_ctl = top();
3416 kls = null_check_oop(kls, &null_ctl, never_see_null);
3417 if (region != NULL) {
3418 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3419 region->init_req(null_path, null_ctl);
3420 } else {
3421 assert(null_ctl == top(), "no loose ends");
3422 }
3423 return kls;
3424 }
3425
3426 //--------------------(inline_native_Class_query helpers)---------------------
3427 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3428 // Fall through if (mods & mask) == bits, take the guard otherwise.
3429 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3430 // Branch around if the given klass has the given modifier bit set.
3431 // Like generate_guard, adds a new path onto the region.
3432 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3433 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3434 Node* mask = intcon(modifier_mask);
3435 Node* bits = intcon(modifier_bits);
3436 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3437 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3438 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3439 return generate_fair_guard(bol, region);
3440 }
3441 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3442 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3443 }
3444
3445 //-------------------------inline_native_Class_query-------------------
3446 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3447 const Type* return_type = TypeInt::BOOL;
3448 Node* prim_return_value = top(); // what happens if it's a primitive class?
3449 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3450 bool expect_prim = false; // most of these guys expect to work on refs
3451
3452 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3453
3454 Node* mirror = argument(0);
3455 Node* obj = top();
3456
3457 switch (id) {
3458 case vmIntrinsics::_isInstance:
3459 // nothing is an instance of a primitive type
3460 prim_return_value = intcon(0);
3461 obj = argument(1);
3462 break;
3463 case vmIntrinsics::_getModifiers:
3464 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3465 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3466 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3467 break;
3468 case vmIntrinsics::_isInterface:
3469 prim_return_value = intcon(0);
3470 break;
3471 case vmIntrinsics::_isArray:
3472 prim_return_value = intcon(0);
3473 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3474 break;
3475 case vmIntrinsics::_isPrimitive:
3476 prim_return_value = intcon(1);
3477 expect_prim = true; // obviously
3478 break;
3479 case vmIntrinsics::_getSuperclass:
3480 prim_return_value = null();
3481 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3482 break;
3483 case vmIntrinsics::_getClassAccessFlags:
3484 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3485 return_type = TypeInt::INT; // not bool! 6297094
3486 break;
3487 default:
3488 fatal_unexpected_iid(id);
3489 break;
3490 }
3491
3492 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3493 if (mirror_con == NULL) return false; // cannot happen?
3494
3495 #ifndef PRODUCT
3496 if (C->print_intrinsics() || C->print_inlining()) {
3497 ciType* k = mirror_con->java_mirror_type();
3498 if (k) {
3499 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3500 k->print_name();
3501 tty->cr();
3502 }
3503 }
3504 #endif
3505
3506 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3507 RegionNode* region = new RegionNode(PATH_LIMIT);
3508 record_for_igvn(region);
3509 PhiNode* phi = new PhiNode(region, return_type);
3510
3511 // The mirror will never be null of Reflection.getClassAccessFlags, however
3512 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3513 // if it is. See bug 4774291.
3514
3515 // For Reflection.getClassAccessFlags(), the null check occurs in
3516 // the wrong place; see inline_unsafe_access(), above, for a similar
3517 // situation.
3518 mirror = null_check(mirror);
3519 // If mirror or obj is dead, only null-path is taken.
3520 if (stopped()) return true;
3521
3522 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3523
3524 // Now load the mirror's klass metaobject, and null-check it.
3525 // Side-effects region with the control path if the klass is null.
3526 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3527 // If kls is null, we have a primitive mirror.
3528 phi->init_req(_prim_path, prim_return_value);
3529 if (stopped()) { set_result(region, phi); return true; }
3530 bool safe_for_replace = (region->in(_prim_path) == top());
3531
3532 Node* p; // handy temp
3533 Node* null_ctl;
3534
3535 // Now that we have the non-null klass, we can perform the real query.
3536 // For constant classes, the query will constant-fold in LoadNode::Value.
3537 Node* query_value = top();
3538 switch (id) {
3539 case vmIntrinsics::_isInstance:
3540 // nothing is an instance of a primitive type
3541 query_value = gen_instanceof(obj, kls, safe_for_replace);
3542 break;
3543
3544 case vmIntrinsics::_getModifiers:
3545 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3546 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3547 break;
3548
3549 case vmIntrinsics::_isInterface:
3550 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3551 if (generate_interface_guard(kls, region) != NULL)
3552 // A guard was added. If the guard is taken, it was an interface.
3553 phi->add_req(intcon(1));
3554 // If we fall through, it's a plain class.
3555 query_value = intcon(0);
3556 break;
3557
3558 case vmIntrinsics::_isArray:
3559 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3560 if (generate_array_guard(kls, region) != NULL)
3561 // A guard was added. If the guard is taken, it was an array.
3562 phi->add_req(intcon(1));
3563 // If we fall through, it's a plain class.
3564 query_value = intcon(0);
3565 break;
3566
3567 case vmIntrinsics::_isPrimitive:
3568 query_value = intcon(0); // "normal" path produces false
3569 break;
3570
3571 case vmIntrinsics::_getSuperclass:
3572 // The rules here are somewhat unfortunate, but we can still do better
3573 // with random logic than with a JNI call.
3574 // Interfaces store null or Object as _super, but must report null.
3575 // Arrays store an intermediate super as _super, but must report Object.
3576 // Other types can report the actual _super.
3577 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3578 if (generate_interface_guard(kls, region) != NULL)
3579 // A guard was added. If the guard is taken, it was an interface.
3580 phi->add_req(null());
3581 if (generate_array_guard(kls, region) != NULL)
3582 // A guard was added. If the guard is taken, it was an array.
3583 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3584 // If we fall through, it's a plain class. Get its _super.
3585 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3586 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3587 null_ctl = top();
3588 kls = null_check_oop(kls, &null_ctl);
3589 if (null_ctl != top()) {
3590 // If the guard is taken, Object.superClass is null (both klass and mirror).
3591 region->add_req(null_ctl);
3592 phi ->add_req(null());
3593 }
3594 if (!stopped()) {
3595 query_value = load_mirror_from_klass(kls);
3596 }
3597 break;
3598
3599 case vmIntrinsics::_getClassAccessFlags:
3600 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3601 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3602 break;
3603
3604 default:
3605 fatal_unexpected_iid(id);
3606 break;
3607 }
3608
3609 // Fall-through is the normal case of a query to a real class.
3610 phi->init_req(1, query_value);
3611 region->init_req(1, control());
3612
3613 C->set_has_split_ifs(true); // Has chance for split-if optimization
3614 set_result(region, phi);
3615 return true;
3616 }
3617
3618 //-------------------------inline_Class_cast-------------------
3619 bool LibraryCallKit::inline_Class_cast() {
3620 Node* mirror = argument(0); // Class
3621 Node* obj = argument(1);
3622 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3623 if (mirror_con == NULL) {
3624 return false; // dead path (mirror->is_top()).
3625 }
3626 if (obj == NULL || obj->is_top()) {
3627 return false; // dead path
3628 }
3629 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3630
3631 // First, see if Class.cast() can be folded statically.
3632 // java_mirror_type() returns non-null for compile-time Class constants.
3633 ciType* tm = mirror_con->java_mirror_type();
3634 if (tm != NULL && tm->is_klass() &&
3635 tp != NULL && tp->klass() != NULL) {
3636 if (!tp->klass()->is_loaded()) {
3637 // Don't use intrinsic when class is not loaded.
3638 return false;
3639 } else {
3640 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3641 if (static_res == Compile::SSC_always_true) {
3642 // isInstance() is true - fold the code.
3643 set_result(obj);
3644 return true;
3645 } else if (static_res == Compile::SSC_always_false) {
3646 // Don't use intrinsic, have to throw ClassCastException.
3647 // If the reference is null, the non-intrinsic bytecode will
3648 // be optimized appropriately.
3649 return false;
3650 }
3651 }
3652 }
3653
3654 // Bailout intrinsic and do normal inlining if exception path is frequent.
3655 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3656 return false;
3657 }
3658
3659 // Generate dynamic checks.
3660 // Class.cast() is java implementation of _checkcast bytecode.
3661 // Do checkcast (Parse::do_checkcast()) optimizations here.
3662
3663 mirror = null_check(mirror);
3664 // If mirror is dead, only null-path is taken.
3665 if (stopped()) {
3666 return true;
3667 }
3668
3669 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3670 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3671 RegionNode* region = new RegionNode(PATH_LIMIT);
3672 record_for_igvn(region);
3673
3674 // Now load the mirror's klass metaobject, and null-check it.
3675 // If kls is null, we have a primitive mirror and
3676 // nothing is an instance of a primitive type.
3677 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3678
3679 Node* res = top();
3680 if (!stopped()) {
3681 Node* bad_type_ctrl = top();
3682 // Do checkcast optimizations.
3683 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3684 region->init_req(_bad_type_path, bad_type_ctrl);
3685 }
3686 if (region->in(_prim_path) != top() ||
3687 region->in(_bad_type_path) != top()) {
3688 // Let Interpreter throw ClassCastException.
3689 PreserveJVMState pjvms(this);
3690 set_control(_gvn.transform(region));
3691 uncommon_trap(Deoptimization::Reason_intrinsic,
3692 Deoptimization::Action_maybe_recompile);
3693 }
3694 if (!stopped()) {
3695 set_result(res);
3696 }
3697 return true;
3698 }
3699
3700
3701 //--------------------------inline_native_subtype_check------------------------
3702 // This intrinsic takes the JNI calls out of the heart of
3703 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3704 bool LibraryCallKit::inline_native_subtype_check() {
3705 // Pull both arguments off the stack.
3706 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3707 args[0] = argument(0);
3708 args[1] = argument(1);
3709 Node* klasses[2]; // corresponding Klasses: superk, subk
3710 klasses[0] = klasses[1] = top();
3711
3712 enum {
3713 // A full decision tree on {superc is prim, subc is prim}:
3714 _prim_0_path = 1, // {P,N} => false
3715 // {P,P} & superc!=subc => false
3716 _prim_same_path, // {P,P} & superc==subc => true
3717 _prim_1_path, // {N,P} => false
3718 _ref_subtype_path, // {N,N} & subtype check wins => true
3719 _both_ref_path, // {N,N} & subtype check loses => false
3720 PATH_LIMIT
3721 };
3722
3723 RegionNode* region = new RegionNode(PATH_LIMIT);
3724 Node* phi = new PhiNode(region, TypeInt::BOOL);
3725 record_for_igvn(region);
3726
3727 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3728 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3729 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3730
3731 // First null-check both mirrors and load each mirror's klass metaobject.
3732 int which_arg;
3733 for (which_arg = 0; which_arg <= 1; which_arg++) {
3734 Node* arg = args[which_arg];
3735 arg = null_check(arg);
3736 if (stopped()) break;
3737 args[which_arg] = arg;
3738
3739 Node* p = basic_plus_adr(arg, class_klass_offset);
3740 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3741 klasses[which_arg] = _gvn.transform(kls);
3742 }
3743
3744 // Having loaded both klasses, test each for null.
3745 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3746 for (which_arg = 0; which_arg <= 1; which_arg++) {
3747 Node* kls = klasses[which_arg];
3748 Node* null_ctl = top();
3749 kls = null_check_oop(kls, &null_ctl, never_see_null);
3750 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3751 region->init_req(prim_path, null_ctl);
3752 if (stopped()) break;
3753 klasses[which_arg] = kls;
3754 }
3755
3756 if (!stopped()) {
3757 // now we have two reference types, in klasses[0..1]
3758 Node* subk = klasses[1]; // the argument to isAssignableFrom
3759 Node* superk = klasses[0]; // the receiver
3760 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3761 // now we have a successful reference subtype check
3762 region->set_req(_ref_subtype_path, control());
3763 }
3764
3765 // If both operands are primitive (both klasses null), then
3766 // we must return true when they are identical primitives.
3767 // It is convenient to test this after the first null klass check.
3768 set_control(region->in(_prim_0_path)); // go back to first null check
3769 if (!stopped()) {
3770 // Since superc is primitive, make a guard for the superc==subc case.
3771 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3772 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3773 generate_guard(bol_eq, region, PROB_FAIR);
3774 if (region->req() == PATH_LIMIT+1) {
3775 // A guard was added. If the added guard is taken, superc==subc.
3776 region->swap_edges(PATH_LIMIT, _prim_same_path);
3777 region->del_req(PATH_LIMIT);
3778 }
3779 region->set_req(_prim_0_path, control()); // Not equal after all.
3780 }
3781
3782 // these are the only paths that produce 'true':
3783 phi->set_req(_prim_same_path, intcon(1));
3784 phi->set_req(_ref_subtype_path, intcon(1));
3785
3786 // pull together the cases:
3787 assert(region->req() == PATH_LIMIT, "sane region");
3788 for (uint i = 1; i < region->req(); i++) {
3789 Node* ctl = region->in(i);
3790 if (ctl == NULL || ctl == top()) {
3791 region->set_req(i, top());
3792 phi ->set_req(i, top());
3793 } else if (phi->in(i) == NULL) {
3794 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3795 }
3796 }
3797
3798 set_control(_gvn.transform(region));
3799 set_result(_gvn.transform(phi));
3800 return true;
3801 }
3802
3803 //---------------------generate_array_guard_common------------------------
3804 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3805 bool obj_array, bool not_array) {
3806
3807 if (stopped()) {
3808 return NULL;
3809 }
3810
3811 // If obj_array/non_array==false/false:
3812 // Branch around if the given klass is in fact an array (either obj or prim).
3813 // If obj_array/non_array==false/true:
3814 // Branch around if the given klass is not an array klass of any kind.
3815 // If obj_array/non_array==true/true:
3816 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3817 // If obj_array/non_array==true/false:
3818 // Branch around if the kls is an oop array (Object[] or subtype)
3819 //
3820 // Like generate_guard, adds a new path onto the region.
3821 jint layout_con = 0;
3822 Node* layout_val = get_layout_helper(kls, layout_con);
3823 if (layout_val == NULL) {
3824 bool query = (obj_array
3825 ? Klass::layout_helper_is_objArray(layout_con)
3826 : Klass::layout_helper_is_array(layout_con));
3827 if (query == not_array) {
3828 return NULL; // never a branch
3829 } else { // always a branch
3830 Node* always_branch = control();
3831 if (region != NULL)
3832 region->add_req(always_branch);
3833 set_control(top());
3834 return always_branch;
3835 }
3836 }
3837 // Now test the correct condition.
3838 jint nval = (obj_array
3839 ? (jint)(Klass::_lh_array_tag_type_value
3840 << Klass::_lh_array_tag_shift)
3841 : Klass::_lh_neutral_value);
3842 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3843 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3844 // invert the test if we are looking for a non-array
3845 if (not_array) btest = BoolTest(btest).negate();
3846 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3847 return generate_fair_guard(bol, region);
3848 }
3849
3850
3851 //-----------------------inline_native_newArray--------------------------
3852 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3853 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3854 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3855 Node* mirror;
3856 Node* count_val;
3857 if (uninitialized) {
3858 mirror = argument(1);
3859 count_val = argument(2);
3860 } else {
3861 mirror = argument(0);
3862 count_val = argument(1);
3863 }
3864
3865 mirror = null_check(mirror);
3866 // If mirror or obj is dead, only null-path is taken.
3867 if (stopped()) return true;
3868
3869 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3870 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3871 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3872 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3873 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3874
3875 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3876 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3877 result_reg, _slow_path);
3878 Node* normal_ctl = control();
3879 Node* no_array_ctl = result_reg->in(_slow_path);
3880
3881 // Generate code for the slow case. We make a call to newArray().
3882 set_control(no_array_ctl);
3883 if (!stopped()) {
3884 // Either the input type is void.class, or else the
3885 // array klass has not yet been cached. Either the
3886 // ensuing call will throw an exception, or else it
3887 // will cache the array klass for next time.
3888 PreserveJVMState pjvms(this);
3889 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3890 Node* slow_result = set_results_for_java_call(slow_call);
3891 // this->control() comes from set_results_for_java_call
3892 result_reg->set_req(_slow_path, control());
3893 result_val->set_req(_slow_path, slow_result);
3894 result_io ->set_req(_slow_path, i_o());
3895 result_mem->set_req(_slow_path, reset_memory());
3896 }
3897
3898 set_control(normal_ctl);
3899 if (!stopped()) {
3900 // Normal case: The array type has been cached in the java.lang.Class.
3901 // The following call works fine even if the array type is polymorphic.
3902 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3903 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3904 result_reg->init_req(_normal_path, control());
3905 result_val->init_req(_normal_path, obj);
3906 result_io ->init_req(_normal_path, i_o());
3907 result_mem->init_req(_normal_path, reset_memory());
3908
3909 if (uninitialized) {
3910 // Mark the allocation so that zeroing is skipped
3911 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3912 alloc->maybe_set_complete(&_gvn);
3913 }
3914 }
3915
3916 // Return the combined state.
3917 set_i_o( _gvn.transform(result_io) );
3918 set_all_memory( _gvn.transform(result_mem));
3919
3920 C->set_has_split_ifs(true); // Has chance for split-if optimization
3921 set_result(result_reg, result_val);
3922 return true;
3923 }
3924
3925 //----------------------inline_native_getLength--------------------------
3926 // public static native int java.lang.reflect.Array.getLength(Object array);
3927 bool LibraryCallKit::inline_native_getLength() {
3928 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3929
3930 Node* array = null_check(argument(0));
3931 // If array is dead, only null-path is taken.
3932 if (stopped()) return true;
3933
3934 // Deoptimize if it is a non-array.
3935 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3936
3937 if (non_array != NULL) {
3938 PreserveJVMState pjvms(this);
3939 set_control(non_array);
3940 uncommon_trap(Deoptimization::Reason_intrinsic,
3941 Deoptimization::Action_maybe_recompile);
3942 }
3943
3944 // If control is dead, only non-array-path is taken.
3945 if (stopped()) return true;
3946
3947 // The works fine even if the array type is polymorphic.
3948 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3949 Node* result = load_array_length(array);
3950
3951 C->set_has_split_ifs(true); // Has chance for split-if optimization
3952 set_result(result);
3953 return true;
3954 }
3955
3956 //------------------------inline_array_copyOf----------------------------
3957 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3958 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3959 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3960 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3961
3962 // Get the arguments.
3963 Node* original = argument(0);
3964 Node* start = is_copyOfRange? argument(1): intcon(0);
3965 Node* end = is_copyOfRange? argument(2): argument(1);
3966 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3967
3968 Node* newcopy = NULL;
3969
3970 // Set the original stack and the reexecute bit for the interpreter to reexecute
3971 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3972 { PreserveReexecuteState preexecs(this);
3973 jvms()->set_should_reexecute(true);
3974
3975 array_type_mirror = null_check(array_type_mirror);
3976 original = null_check(original);
3977
3978 // Check if a null path was taken unconditionally.
3979 if (stopped()) return true;
3980
3981 Node* orig_length = load_array_length(original);
3982
3983 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3984 klass_node = null_check(klass_node);
3985
3986 RegionNode* bailout = new RegionNode(1);
3987 record_for_igvn(bailout);
3988
3989 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3990 // Bail out if that is so.
3991 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3992 if (not_objArray != NULL) {
3993 // Improve the klass node's type from the new optimistic assumption:
3994 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3995 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3996 Node* cast = new CastPPNode(klass_node, akls);
3997 cast->init_req(0, control());
3998 klass_node = _gvn.transform(cast);
3999 }
4000
4001 // Bail out if either start or end is negative.
4002 generate_negative_guard(start, bailout, &start);
4003 generate_negative_guard(end, bailout, &end);
4004
4005 Node* length = end;
4006 if (_gvn.type(start) != TypeInt::ZERO) {
4007 length = _gvn.transform(new SubINode(end, start));
4008 }
4009
4010 // Bail out if length is negative.
4011 // Without this the new_array would throw
4012 // NegativeArraySizeException but IllegalArgumentException is what
4013 // should be thrown
4014 generate_negative_guard(length, bailout, &length);
4015
4016 if (bailout->req() > 1) {
4017 PreserveJVMState pjvms(this);
4018 set_control(_gvn.transform(bailout));
4019 uncommon_trap(Deoptimization::Reason_intrinsic,
4020 Deoptimization::Action_maybe_recompile);
4021 }
4022
4023 if (!stopped()) {
4024 // How many elements will we copy from the original?
4025 // The answer is MinI(orig_length - start, length).
4026 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4027 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4028
4029 // Generate a direct call to the right arraycopy function(s).
4030 // We know the copy is disjoint but we might not know if the
4031 // oop stores need checking.
4032 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4033 // This will fail a store-check if x contains any non-nulls.
4034
4035 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4036 // loads/stores but it is legal only if we're sure the
4037 // Arrays.copyOf would succeed. So we need all input arguments
4038 // to the copyOf to be validated, including that the copy to the
4039 // new array won't trigger an ArrayStoreException. That subtype
4040 // check can be optimized if we know something on the type of
4041 // the input array from type speculation.
4042 if (_gvn.type(klass_node)->singleton()) {
4043 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4044 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4045
4046 int test = C->static_subtype_check(superk, subk);
4047 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4048 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4049 if (t_original->speculative_type() != NULL) {
4050 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4051 }
4052 }
4053 }
4054
4055 bool validated = false;
4056 // Reason_class_check rather than Reason_intrinsic because we
4057 // want to intrinsify even if this traps.
4058 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4059 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4060 klass_node);
4061
4062 if (not_subtype_ctrl != top()) {
4063 PreserveJVMState pjvms(this);
4064 set_control(not_subtype_ctrl);
4065 uncommon_trap(Deoptimization::Reason_class_check,
4066 Deoptimization::Action_make_not_entrant);
4067 assert(stopped(), "Should be stopped");
4068 }
4069 validated = true;
4070 }
4071
4072 if (!stopped()) {
4073 newcopy = new_array(klass_node, length, 0); // no arguments to push
4074
4075 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4076 load_object_klass(original), klass_node);
4077 if (!is_copyOfRange) {
4078 ac->set_copyof(validated);
4079 } else {
4080 ac->set_copyofrange(validated);
4081 }
4082 Node* n = _gvn.transform(ac);
4083 if (n == ac) {
4084 ac->connect_outputs(this);
4085 } else {
4086 assert(validated, "shouldn't transform if all arguments not validated");
4087 set_all_memory(n);
4088 }
4089 }
4090 }
4091 } // original reexecute is set back here
4092
4093 C->set_has_split_ifs(true); // Has chance for split-if optimization
4094 if (!stopped()) {
4095 set_result(newcopy);
4096 }
4097 return true;
4098 }
4099
4100
4101 //----------------------generate_virtual_guard---------------------------
4102 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4103 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4104 RegionNode* slow_region) {
4105 ciMethod* method = callee();
4106 int vtable_index = method->vtable_index();
4107 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4108 "bad index %d", vtable_index);
4109 // Get the Method* out of the appropriate vtable entry.
4110 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4111 vtable_index*vtableEntry::size_in_bytes() +
4112 vtableEntry::method_offset_in_bytes();
4113 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4114 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4115
4116 // Compare the target method with the expected method (e.g., Object.hashCode).
4117 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4118
4119 Node* native_call = makecon(native_call_addr);
4120 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4121 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4122
4123 return generate_slow_guard(test_native, slow_region);
4124 }
4125
4126 //-----------------------generate_method_call----------------------------
4127 // Use generate_method_call to make a slow-call to the real
4128 // method if the fast path fails. An alternative would be to
4129 // use a stub like OptoRuntime::slow_arraycopy_Java.
4130 // This only works for expanding the current library call,
4131 // not another intrinsic. (E.g., don't use this for making an
4132 // arraycopy call inside of the copyOf intrinsic.)
4133 CallJavaNode*
4134 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4135 // When compiling the intrinsic method itself, do not use this technique.
4136 guarantee(callee() != C->method(), "cannot make slow-call to self");
4137
4138 ciMethod* method = callee();
4139 // ensure the JVMS we have will be correct for this call
4140 guarantee(method_id == method->intrinsic_id(), "must match");
4141
4142 const TypeFunc* tf = TypeFunc::make(method);
4143 CallJavaNode* slow_call;
4144 if (is_static) {
4145 assert(!is_virtual, "");
4146 slow_call = new CallStaticJavaNode(C, tf,
4147 SharedRuntime::get_resolve_static_call_stub(),
4148 method, bci());
4149 } else if (is_virtual) {
4150 null_check_receiver();
4151 int vtable_index = Method::invalid_vtable_index;
4152 if (UseInlineCaches) {
4153 // Suppress the vtable call
4154 } else {
4155 // hashCode and clone are not a miranda methods,
4156 // so the vtable index is fixed.
4157 // No need to use the linkResolver to get it.
4158 vtable_index = method->vtable_index();
4159 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4160 "bad index %d", vtable_index);
4161 }
4162 slow_call = new CallDynamicJavaNode(tf,
4163 SharedRuntime::get_resolve_virtual_call_stub(),
4164 method, vtable_index, bci());
4165 } else { // neither virtual nor static: opt_virtual
4166 null_check_receiver();
4167 slow_call = new CallStaticJavaNode(C, tf,
4168 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4169 method, bci());
4170 slow_call->set_optimized_virtual(true);
4171 }
4172 set_arguments_for_java_call(slow_call);
4173 set_edges_for_java_call(slow_call);
4174 return slow_call;
4175 }
4176
4177
4178 /**
4179 * Build special case code for calls to hashCode on an object. This call may
4180 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4181 * slightly different code.
4182 */
4183 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4184 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4185 assert(!(is_virtual && is_static), "either virtual, special, or static");
4186
4187 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4188
4189 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4190 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4191 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4192 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4193 Node* obj = NULL;
4194 if (!is_static) {
4195 // Check for hashing null object
4196 obj = null_check_receiver();
4197 if (stopped()) return true; // unconditionally null
4198 result_reg->init_req(_null_path, top());
4199 result_val->init_req(_null_path, top());
4200 } else {
4201 // Do a null check, and return zero if null.
4202 // System.identityHashCode(null) == 0
4203 obj = argument(0);
4204 Node* null_ctl = top();
4205 obj = null_check_oop(obj, &null_ctl);
4206 result_reg->init_req(_null_path, null_ctl);
4207 result_val->init_req(_null_path, _gvn.intcon(0));
4208 }
4209
4210 // Unconditionally null? Then return right away.
4211 if (stopped()) {
4212 set_control( result_reg->in(_null_path));
4213 if (!stopped())
4214 set_result(result_val->in(_null_path));
4215 return true;
4216 }
4217
4218 // We only go to the fast case code if we pass a number of guards. The
4219 // paths which do not pass are accumulated in the slow_region.
4220 RegionNode* slow_region = new RegionNode(1);
4221 record_for_igvn(slow_region);
4222
4223 // If this is a virtual call, we generate a funny guard. We pull out
4224 // the vtable entry corresponding to hashCode() from the target object.
4225 // If the target method which we are calling happens to be the native
4226 // Object hashCode() method, we pass the guard. We do not need this
4227 // guard for non-virtual calls -- the caller is known to be the native
4228 // Object hashCode().
4229 if (is_virtual) {
4230 // After null check, get the object's klass.
4231 Node* obj_klass = load_object_klass(obj);
4232 generate_virtual_guard(obj_klass, slow_region);
4233 }
4234
4235 // Get the header out of the object, use LoadMarkNode when available
4236 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4237 // The control of the load must be NULL. Otherwise, the load can move before
4238 // the null check after castPP removal.
4239 Node* no_ctrl = NULL;
4240 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4241
4242 // Test the header to see if it is unlocked.
4243 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4244 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4245 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4246 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4247 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4248
4249 generate_slow_guard(test_unlocked, slow_region);
4250
4251 // Get the hash value and check to see that it has been properly assigned.
4252 // We depend on hash_mask being at most 32 bits and avoid the use of
4253 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4254 // vm: see markOop.hpp.
4255 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4256 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4257 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4258 // This hack lets the hash bits live anywhere in the mark object now, as long
4259 // as the shift drops the relevant bits into the low 32 bits. Note that
4260 // Java spec says that HashCode is an int so there's no point in capturing
4261 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4262 hshifted_header = ConvX2I(hshifted_header);
4263 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4264
4265 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4266 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4267 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4268
4269 generate_slow_guard(test_assigned, slow_region);
4270
4271 Node* init_mem = reset_memory();
4272 // fill in the rest of the null path:
4273 result_io ->init_req(_null_path, i_o());
4274 result_mem->init_req(_null_path, init_mem);
4275
4276 result_val->init_req(_fast_path, hash_val);
4277 result_reg->init_req(_fast_path, control());
4278 result_io ->init_req(_fast_path, i_o());
4279 result_mem->init_req(_fast_path, init_mem);
4280
4281 // Generate code for the slow case. We make a call to hashCode().
4282 set_control(_gvn.transform(slow_region));
4283 if (!stopped()) {
4284 // No need for PreserveJVMState, because we're using up the present state.
4285 set_all_memory(init_mem);
4286 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4287 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4288 Node* slow_result = set_results_for_java_call(slow_call);
4289 // this->control() comes from set_results_for_java_call
4290 result_reg->init_req(_slow_path, control());
4291 result_val->init_req(_slow_path, slow_result);
4292 result_io ->set_req(_slow_path, i_o());
4293 result_mem ->set_req(_slow_path, reset_memory());
4294 }
4295
4296 // Return the combined state.
4297 set_i_o( _gvn.transform(result_io) );
4298 set_all_memory( _gvn.transform(result_mem));
4299
4300 set_result(result_reg, result_val);
4301 return true;
4302 }
4303
4304 //---------------------------inline_native_getClass----------------------------
4305 // public final native Class<?> java.lang.Object.getClass();
4306 //
4307 // Build special case code for calls to getClass on an object.
4308 bool LibraryCallKit::inline_native_getClass() {
4309 Node* obj = null_check_receiver();
4310 if (stopped()) return true;
4311 set_result(load_mirror_from_klass(load_object_klass(obj)));
4312 return true;
4313 }
4314
4315 //-----------------inline_native_Reflection_getCallerClass---------------------
4316 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4317 //
4318 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4319 //
4320 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4321 // in that it must skip particular security frames and checks for
4322 // caller sensitive methods.
4323 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4324 #ifndef PRODUCT
4325 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4326 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4327 }
4328 #endif
4329
4330 if (!jvms()->has_method()) {
4331 #ifndef PRODUCT
4332 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4333 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4334 }
4335 #endif
4336 return false;
4337 }
4338
4339 // Walk back up the JVM state to find the caller at the required
4340 // depth.
4341 JVMState* caller_jvms = jvms();
4342
4343 // Cf. JVM_GetCallerClass
4344 // NOTE: Start the loop at depth 1 because the current JVM state does
4345 // not include the Reflection.getCallerClass() frame.
4346 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4347 ciMethod* m = caller_jvms->method();
4348 switch (n) {
4349 case 0:
4350 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4351 break;
4352 case 1:
4353 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4354 if (!m->caller_sensitive()) {
4355 #ifndef PRODUCT
4356 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4357 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4358 }
4359 #endif
4360 return false; // bail-out; let JVM_GetCallerClass do the work
4361 }
4362 break;
4363 default:
4364 if (!m->is_ignored_by_security_stack_walk()) {
4365 // We have reached the desired frame; return the holder class.
4366 // Acquire method holder as java.lang.Class and push as constant.
4367 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4368 ciInstance* caller_mirror = caller_klass->java_mirror();
4369 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4370
4371 #ifndef PRODUCT
4372 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4373 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());
4374 tty->print_cr(" JVM state at this point:");
4375 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4376 ciMethod* m = jvms()->of_depth(i)->method();
4377 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4378 }
4379 }
4380 #endif
4381 return true;
4382 }
4383 break;
4384 }
4385 }
4386
4387 #ifndef PRODUCT
4388 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4389 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4390 tty->print_cr(" JVM state at this point:");
4391 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4392 ciMethod* m = jvms()->of_depth(i)->method();
4393 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4394 }
4395 }
4396 #endif
4397
4398 return false; // bail-out; let JVM_GetCallerClass do the work
4399 }
4400
4401 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4402 Node* arg = argument(0);
4403 Node* result = NULL;
4404
4405 switch (id) {
4406 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4407 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4408 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4409 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4410
4411 case vmIntrinsics::_doubleToLongBits: {
4412 // two paths (plus control) merge in a wood
4413 RegionNode *r = new RegionNode(3);
4414 Node *phi = new PhiNode(r, TypeLong::LONG);
4415
4416 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4417 // Build the boolean node
4418 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4419
4420 // Branch either way.
4421 // NaN case is less traveled, which makes all the difference.
4422 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4423 Node *opt_isnan = _gvn.transform(ifisnan);
4424 assert( opt_isnan->is_If(), "Expect an IfNode");
4425 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4426 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4427
4428 set_control(iftrue);
4429
4430 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4431 Node *slow_result = longcon(nan_bits); // return NaN
4432 phi->init_req(1, _gvn.transform( slow_result ));
4433 r->init_req(1, iftrue);
4434
4435 // Else fall through
4436 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4437 set_control(iffalse);
4438
4439 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4440 r->init_req(2, iffalse);
4441
4442 // Post merge
4443 set_control(_gvn.transform(r));
4444 record_for_igvn(r);
4445
4446 C->set_has_split_ifs(true); // Has chance for split-if optimization
4447 result = phi;
4448 assert(result->bottom_type()->isa_long(), "must be");
4449 break;
4450 }
4451
4452 case vmIntrinsics::_floatToIntBits: {
4453 // two paths (plus control) merge in a wood
4454 RegionNode *r = new RegionNode(3);
4455 Node *phi = new PhiNode(r, TypeInt::INT);
4456
4457 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4458 // Build the boolean node
4459 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4460
4461 // Branch either way.
4462 // NaN case is less traveled, which makes all the difference.
4463 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4464 Node *opt_isnan = _gvn.transform(ifisnan);
4465 assert( opt_isnan->is_If(), "Expect an IfNode");
4466 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4467 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4468
4469 set_control(iftrue);
4470
4471 static const jint nan_bits = 0x7fc00000;
4472 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4473 phi->init_req(1, _gvn.transform( slow_result ));
4474 r->init_req(1, iftrue);
4475
4476 // Else fall through
4477 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4478 set_control(iffalse);
4479
4480 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4481 r->init_req(2, iffalse);
4482
4483 // Post merge
4484 set_control(_gvn.transform(r));
4485 record_for_igvn(r);
4486
4487 C->set_has_split_ifs(true); // Has chance for split-if optimization
4488 result = phi;
4489 assert(result->bottom_type()->isa_int(), "must be");
4490 break;
4491 }
4492
4493 default:
4494 fatal_unexpected_iid(id);
4495 break;
4496 }
4497 set_result(_gvn.transform(result));
4498 return true;
4499 }
4500
4501 //----------------------inline_unsafe_copyMemory-------------------------
4502 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4503 bool LibraryCallKit::inline_unsafe_copyMemory() {
4504 if (callee()->is_static()) return false; // caller must have the capability!
4505 null_check_receiver(); // null-check receiver
4506 if (stopped()) return true;
4507
4508 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4509
4510 Node* src_ptr = argument(1); // type: oop
4511 Node* src_off = ConvL2X(argument(2)); // type: long
4512 Node* dst_ptr = argument(4); // type: oop
4513 Node* dst_off = ConvL2X(argument(5)); // type: long
4514 Node* size = ConvL2X(argument(7)); // type: long
4515
4516 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4517 "fieldOffset must be byte-scaled");
4518
4519 Node* src = make_unsafe_address(src_ptr, src_off);
4520 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4521
4522 // Conservatively insert a memory barrier on all memory slices.
4523 // Do not let writes of the copy source or destination float below the copy.
4524 insert_mem_bar(Op_MemBarCPUOrder);
4525
4526 // Call it. Note that the length argument is not scaled.
4527 make_runtime_call(RC_LEAF|RC_NO_FP,
4528 OptoRuntime::fast_arraycopy_Type(),
4529 StubRoutines::unsafe_arraycopy(),
4530 "unsafe_arraycopy",
4531 TypeRawPtr::BOTTOM,
4532 src, dst, size XTOP);
4533
4534 // Do not let reads of the copy destination float above the copy.
4535 insert_mem_bar(Op_MemBarCPUOrder);
4536
4537 return true;
4538 }
4539
4540 //------------------------clone_coping-----------------------------------
4541 // Helper function for inline_native_clone.
4542 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4543 assert(obj_size != NULL, "");
4544 Node* raw_obj = alloc_obj->in(1);
4545 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4546
4547 AllocateNode* alloc = NULL;
4548 if (ReduceBulkZeroing) {
4549 // We will be completely responsible for initializing this object -
4550 // mark Initialize node as complete.
4551 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4552 // The object was just allocated - there should be no any stores!
4553 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4554 // Mark as complete_with_arraycopy so that on AllocateNode
4555 // expansion, we know this AllocateNode is initialized by an array
4556 // copy and a StoreStore barrier exists after the array copy.
4557 alloc->initialization()->set_complete_with_arraycopy();
4558 }
4559
4560 // Copy the fastest available way.
4561 // TODO: generate fields copies for small objects instead.
4562 Node* src = obj;
4563 Node* dest = alloc_obj;
4564 Node* size = _gvn.transform(obj_size);
4565
4566 // Exclude the header but include array length to copy by 8 bytes words.
4567 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4568 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4569 instanceOopDesc::base_offset_in_bytes();
4570 // base_off:
4571 // 8 - 32-bit VM
4572 // 12 - 64-bit VM, compressed klass
4573 // 16 - 64-bit VM, normal klass
4574 if (base_off % BytesPerLong != 0) {
4575 assert(UseCompressedClassPointers, "");
4576 if (is_array) {
4577 // Exclude length to copy by 8 bytes words.
4578 base_off += sizeof(int);
4579 } else {
4580 // Include klass to copy by 8 bytes words.
4581 base_off = instanceOopDesc::klass_offset_in_bytes();
4582 }
4583 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4584 }
4585 src = basic_plus_adr(src, base_off);
4586 dest = basic_plus_adr(dest, base_off);
4587
4588 // Compute the length also, if needed:
4589 Node* countx = size;
4590 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4591 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4592
4593 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4594
4595 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4596 ac->set_clonebasic();
4597 Node* n = _gvn.transform(ac);
4598 if (n == ac) {
4599 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4600 } else {
4601 set_all_memory(n);
4602 }
4603
4604 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4605 if (card_mark) {
4606 assert(!is_array, "");
4607 // Put in store barrier for any and all oops we are sticking
4608 // into this object. (We could avoid this if we could prove
4609 // that the object type contains no oop fields at all.)
4610 Node* no_particular_value = NULL;
4611 Node* no_particular_field = NULL;
4612 int raw_adr_idx = Compile::AliasIdxRaw;
4613 post_barrier(control(),
4614 memory(raw_adr_type),
4615 alloc_obj,
4616 no_particular_field,
4617 raw_adr_idx,
4618 no_particular_value,
4619 T_OBJECT,
4620 false);
4621 }
4622
4623 // Do not let reads from the cloned object float above the arraycopy.
4624 if (alloc != NULL) {
4625 // Do not let stores that initialize this object be reordered with
4626 // a subsequent store that would make this object accessible by
4627 // other threads.
4628 // Record what AllocateNode this StoreStore protects so that
4629 // escape analysis can go from the MemBarStoreStoreNode to the
4630 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4631 // based on the escape status of the AllocateNode.
4632 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4633 } else {
4634 insert_mem_bar(Op_MemBarCPUOrder);
4635 }
4636 }
4637
4638 //------------------------inline_native_clone----------------------------
4639 // protected native Object java.lang.Object.clone();
4640 //
4641 // Here are the simple edge cases:
4642 // null receiver => normal trap
4643 // virtual and clone was overridden => slow path to out-of-line clone
4644 // not cloneable or finalizer => slow path to out-of-line Object.clone
4645 //
4646 // The general case has two steps, allocation and copying.
4647 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4648 //
4649 // Copying also has two cases, oop arrays and everything else.
4650 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4651 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4652 //
4653 // These steps fold up nicely if and when the cloned object's klass
4654 // can be sharply typed as an object array, a type array, or an instance.
4655 //
4656 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4657 PhiNode* result_val;
4658
4659 // Set the reexecute bit for the interpreter to reexecute
4660 // the bytecode that invokes Object.clone if deoptimization happens.
4661 { PreserveReexecuteState preexecs(this);
4662 jvms()->set_should_reexecute(true);
4663
4664 Node* obj = null_check_receiver();
4665 if (stopped()) return true;
4666
4667 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4668
4669 // If we are going to clone an instance, we need its exact type to
4670 // know the number and types of fields to convert the clone to
4671 // loads/stores. Maybe a speculative type can help us.
4672 if (!obj_type->klass_is_exact() &&
4673 obj_type->speculative_type() != NULL &&
4674 obj_type->speculative_type()->is_instance_klass()) {
4675 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4676 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4677 !spec_ik->has_injected_fields()) {
4678 ciKlass* k = obj_type->klass();
4679 if (!k->is_instance_klass() ||
4680 k->as_instance_klass()->is_interface() ||
4681 k->as_instance_klass()->has_subklass()) {
4682 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4683 }
4684 }
4685 }
4686
4687 Node* obj_klass = load_object_klass(obj);
4688 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4689 const TypeOopPtr* toop = ((tklass != NULL)
4690 ? tklass->as_instance_type()
4691 : TypeInstPtr::NOTNULL);
4692
4693 // Conservatively insert a memory barrier on all memory slices.
4694 // Do not let writes into the original float below the clone.
4695 insert_mem_bar(Op_MemBarCPUOrder);
4696
4697 // paths into result_reg:
4698 enum {
4699 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4700 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4701 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4702 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4703 PATH_LIMIT
4704 };
4705 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4706 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4707 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4708 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4709 record_for_igvn(result_reg);
4710
4711 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4712 int raw_adr_idx = Compile::AliasIdxRaw;
4713
4714 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4715 if (array_ctl != NULL) {
4716 // It's an array.
4717 PreserveJVMState pjvms(this);
4718 set_control(array_ctl);
4719 Node* obj_length = load_array_length(obj);
4720 Node* obj_size = NULL;
4721 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4722
4723 if (!use_ReduceInitialCardMarks()) {
4724 // If it is an oop array, it requires very special treatment,
4725 // because card marking is required on each card of the array.
4726 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4727 if (is_obja != NULL) {
4728 PreserveJVMState pjvms2(this);
4729 set_control(is_obja);
4730 // Generate a direct call to the right arraycopy function(s).
4731 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4732 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4733 ac->set_cloneoop();
4734 Node* n = _gvn.transform(ac);
4735 assert(n == ac, "cannot disappear");
4736 ac->connect_outputs(this);
4737
4738 result_reg->init_req(_objArray_path, control());
4739 result_val->init_req(_objArray_path, alloc_obj);
4740 result_i_o ->set_req(_objArray_path, i_o());
4741 result_mem ->set_req(_objArray_path, reset_memory());
4742 }
4743 }
4744 // Otherwise, there are no card marks to worry about.
4745 // (We can dispense with card marks if we know the allocation
4746 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4747 // causes the non-eden paths to take compensating steps to
4748 // simulate a fresh allocation, so that no further
4749 // card marks are required in compiled code to initialize
4750 // the object.)
4751
4752 if (!stopped()) {
4753 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4754
4755 // Present the results of the copy.
4756 result_reg->init_req(_array_path, control());
4757 result_val->init_req(_array_path, alloc_obj);
4758 result_i_o ->set_req(_array_path, i_o());
4759 result_mem ->set_req(_array_path, reset_memory());
4760 }
4761 }
4762
4763 // We only go to the instance fast case code if we pass a number of guards.
4764 // The paths which do not pass are accumulated in the slow_region.
4765 RegionNode* slow_region = new RegionNode(1);
4766 record_for_igvn(slow_region);
4767 if (!stopped()) {
4768 // It's an instance (we did array above). Make the slow-path tests.
4769 // If this is a virtual call, we generate a funny guard. We grab
4770 // the vtable entry corresponding to clone() from the target object.
4771 // If the target method which we are calling happens to be the
4772 // Object clone() method, we pass the guard. We do not need this
4773 // guard for non-virtual calls; the caller is known to be the native
4774 // Object clone().
4775 if (is_virtual) {
4776 generate_virtual_guard(obj_klass, slow_region);
4777 }
4778
4779 // The object must be easily cloneable and must not have a finalizer.
4780 // Both of these conditions may be checked in a single test.
4781 // We could optimize the test further, but we don't care.
4782 generate_access_flags_guard(obj_klass,
4783 // Test both conditions:
4784 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4785 // Must be cloneable but not finalizer:
4786 JVM_ACC_IS_CLONEABLE_FAST,
4787 slow_region);
4788 }
4789
4790 if (!stopped()) {
4791 // It's an instance, and it passed the slow-path tests.
4792 PreserveJVMState pjvms(this);
4793 Node* obj_size = NULL;
4794 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4795 // is reexecuted if deoptimization occurs and there could be problems when merging
4796 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4797 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4798
4799 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4800
4801 // Present the results of the slow call.
4802 result_reg->init_req(_instance_path, control());
4803 result_val->init_req(_instance_path, alloc_obj);
4804 result_i_o ->set_req(_instance_path, i_o());
4805 result_mem ->set_req(_instance_path, reset_memory());
4806 }
4807
4808 // Generate code for the slow case. We make a call to clone().
4809 set_control(_gvn.transform(slow_region));
4810 if (!stopped()) {
4811 PreserveJVMState pjvms(this);
4812 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4813 Node* slow_result = set_results_for_java_call(slow_call);
4814 // this->control() comes from set_results_for_java_call
4815 result_reg->init_req(_slow_path, control());
4816 result_val->init_req(_slow_path, slow_result);
4817 result_i_o ->set_req(_slow_path, i_o());
4818 result_mem ->set_req(_slow_path, reset_memory());
4819 }
4820
4821 // Return the combined state.
4822 set_control( _gvn.transform(result_reg));
4823 set_i_o( _gvn.transform(result_i_o));
4824 set_all_memory( _gvn.transform(result_mem));
4825 } // original reexecute is set back here
4826
4827 set_result(_gvn.transform(result_val));
4828 return true;
4829 }
4830
4831 // If we have a tighly coupled allocation, the arraycopy may take care
4832 // of the array initialization. If one of the guards we insert between
4833 // the allocation and the arraycopy causes a deoptimization, an
4834 // unitialized array will escape the compiled method. To prevent that
4835 // we set the JVM state for uncommon traps between the allocation and
4836 // the arraycopy to the state before the allocation so, in case of
4837 // deoptimization, we'll reexecute the allocation and the
4838 // initialization.
4839 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4840 if (alloc != NULL) {
4841 ciMethod* trap_method = alloc->jvms()->method();
4842 int trap_bci = alloc->jvms()->bci();
4843
4844 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4845 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4846 // Make sure there's no store between the allocation and the
4847 // arraycopy otherwise visible side effects could be rexecuted
4848 // in case of deoptimization and cause incorrect execution.
4849 bool no_interfering_store = true;
4850 Node* mem = alloc->in(TypeFunc::Memory);
4851 if (mem->is_MergeMem()) {
4852 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4853 Node* n = mms.memory();
4854 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4855 assert(n->is_Store(), "what else?");
4856 no_interfering_store = false;
4857 break;
4858 }
4859 }
4860 } else {
4861 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4862 Node* n = mms.memory();
4863 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4864 assert(n->is_Store(), "what else?");
4865 no_interfering_store = false;
4866 break;
4867 }
4868 }
4869 }
4870
4871 if (no_interfering_store) {
4872 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4873 uint size = alloc->req();
4874 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4875 old_jvms->set_map(sfpt);
4876 for (uint i = 0; i < size; i++) {
4877 sfpt->init_req(i, alloc->in(i));
4878 }
4879 // re-push array length for deoptimization
4880 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4881 old_jvms->set_sp(old_jvms->sp()+1);
4882 old_jvms->set_monoff(old_jvms->monoff()+1);
4883 old_jvms->set_scloff(old_jvms->scloff()+1);
4884 old_jvms->set_endoff(old_jvms->endoff()+1);
4885 old_jvms->set_should_reexecute(true);
4886
4887 sfpt->set_i_o(map()->i_o());
4888 sfpt->set_memory(map()->memory());
4889 sfpt->set_control(map()->control());
4890
4891 JVMState* saved_jvms = jvms();
4892 saved_reexecute_sp = _reexecute_sp;
4893
4894 set_jvms(sfpt->jvms());
4895 _reexecute_sp = jvms()->sp();
4896
4897 return saved_jvms;
4898 }
4899 }
4900 }
4901 return NULL;
4902 }
4903
4904 // In case of a deoptimization, we restart execution at the
4905 // allocation, allocating a new array. We would leave an uninitialized
4906 // array in the heap that GCs wouldn't expect. Move the allocation
4907 // after the traps so we don't allocate the array if we
4908 // deoptimize. This is possible because tightly_coupled_allocation()
4909 // guarantees there's no observer of the allocated array at this point
4910 // and the control flow is simple enough.
4911 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4912 int saved_reexecute_sp, uint new_idx) {
4913 if (saved_jvms != NULL && !stopped()) {
4914 assert(alloc != NULL, "only with a tightly coupled allocation");
4915 // restore JVM state to the state at the arraycopy
4916 saved_jvms->map()->set_control(map()->control());
4917 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4918 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4919 // If we've improved the types of some nodes (null check) while
4920 // emitting the guards, propagate them to the current state
4921 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4922 set_jvms(saved_jvms);
4923 _reexecute_sp = saved_reexecute_sp;
4924
4925 // Remove the allocation from above the guards
4926 CallProjections callprojs;
4927 alloc->extract_projections(&callprojs, true);
4928 InitializeNode* init = alloc->initialization();
4929 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4930 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4931 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4932 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4933
4934 // move the allocation here (after the guards)
4935 _gvn.hash_delete(alloc);
4936 alloc->set_req(TypeFunc::Control, control());
4937 alloc->set_req(TypeFunc::I_O, i_o());
4938 Node *mem = reset_memory();
4939 set_all_memory(mem);
4940 alloc->set_req(TypeFunc::Memory, mem);
4941 set_control(init->proj_out(TypeFunc::Control));
4942 set_i_o(callprojs.fallthrough_ioproj);
4943
4944 // Update memory as done in GraphKit::set_output_for_allocation()
4945 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4946 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4947 if (ary_type->isa_aryptr() && length_type != NULL) {
4948 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4949 }
4950 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4951 int elemidx = C->get_alias_index(telemref);
4952 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4953 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4954
4955 Node* allocx = _gvn.transform(alloc);
4956 assert(allocx == alloc, "where has the allocation gone?");
4957 assert(dest->is_CheckCastPP(), "not an allocation result?");
4958
4959 _gvn.hash_delete(dest);
4960 dest->set_req(0, control());
4961 Node* destx = _gvn.transform(dest);
4962 assert(destx == dest, "where has the allocation result gone?");
4963 }
4964 }
4965
4966
4967 //------------------------------inline_arraycopy-----------------------
4968 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4969 // Object dest, int destPos,
4970 // int length);
4971 bool LibraryCallKit::inline_arraycopy() {
4972 // Get the arguments.
4973 Node* src = argument(0); // type: oop
4974 Node* src_offset = argument(1); // type: int
4975 Node* dest = argument(2); // type: oop
4976 Node* dest_offset = argument(3); // type: int
4977 Node* length = argument(4); // type: int
4978
4979 uint new_idx = C->unique();
4980
4981 // Check for allocation before we add nodes that would confuse
4982 // tightly_coupled_allocation()
4983 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4984
4985 int saved_reexecute_sp = -1;
4986 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4987 // See arraycopy_restore_alloc_state() comment
4988 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4989 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4990 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4991 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4992
4993 // The following tests must be performed
4994 // (1) src and dest are arrays.
4995 // (2) src and dest arrays must have elements of the same BasicType
4996 // (3) src and dest must not be null.
4997 // (4) src_offset must not be negative.
4998 // (5) dest_offset must not be negative.
4999 // (6) length must not be negative.
5000 // (7) src_offset + length must not exceed length of src.
5001 // (8) dest_offset + length must not exceed length of dest.
5002 // (9) each element of an oop array must be assignable
5003
5004 // (3) src and dest must not be null.
5005 // always do this here because we need the JVM state for uncommon traps
5006 Node* null_ctl = top();
5007 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5008 assert(null_ctl->is_top(), "no null control here");
5009 dest = null_check(dest, T_ARRAY);
5010
5011 if (!can_emit_guards) {
5012 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5013 // guards but the arraycopy node could still take advantage of a
5014 // tightly allocated allocation. tightly_coupled_allocation() is
5015 // called again to make sure it takes the null check above into
5016 // account: the null check is mandatory and if it caused an
5017 // uncommon trap to be emitted then the allocation can't be
5018 // considered tightly coupled in this context.
5019 alloc = tightly_coupled_allocation(dest, NULL);
5020 }
5021
5022 bool validated = false;
5023
5024 const Type* src_type = _gvn.type(src);
5025 const Type* dest_type = _gvn.type(dest);
5026 const TypeAryPtr* top_src = src_type->isa_aryptr();
5027 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5028
5029 // Do we have the type of src?
5030 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5031 // Do we have the type of dest?
5032 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5033 // Is the type for src from speculation?
5034 bool src_spec = false;
5035 // Is the type for dest from speculation?
5036 bool dest_spec = false;
5037
5038 if ((!has_src || !has_dest) && can_emit_guards) {
5039 // We don't have sufficient type information, let's see if
5040 // speculative types can help. We need to have types for both src
5041 // and dest so that it pays off.
5042
5043 // Do we already have or could we have type information for src
5044 bool could_have_src = has_src;
5045 // Do we already have or could we have type information for dest
5046 bool could_have_dest = has_dest;
5047
5048 ciKlass* src_k = NULL;
5049 if (!has_src) {
5050 src_k = src_type->speculative_type_not_null();
5051 if (src_k != NULL && src_k->is_array_klass()) {
5052 could_have_src = true;
5053 }
5054 }
5055
5056 ciKlass* dest_k = NULL;
5057 if (!has_dest) {
5058 dest_k = dest_type->speculative_type_not_null();
5059 if (dest_k != NULL && dest_k->is_array_klass()) {
5060 could_have_dest = true;
5061 }
5062 }
5063
5064 if (could_have_src && could_have_dest) {
5065 // This is going to pay off so emit the required guards
5066 if (!has_src) {
5067 src = maybe_cast_profiled_obj(src, src_k, true);
5068 src_type = _gvn.type(src);
5069 top_src = src_type->isa_aryptr();
5070 has_src = (top_src != NULL && top_src->klass() != NULL);
5071 src_spec = true;
5072 }
5073 if (!has_dest) {
5074 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5075 dest_type = _gvn.type(dest);
5076 top_dest = dest_type->isa_aryptr();
5077 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5078 dest_spec = true;
5079 }
5080 }
5081 }
5082
5083 if (has_src && has_dest && can_emit_guards) {
5084 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5085 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5086 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5087 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5088
5089 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5090 // If both arrays are object arrays then having the exact types
5091 // for both will remove the need for a subtype check at runtime
5092 // before the call and may make it possible to pick a faster copy
5093 // routine (without a subtype check on every element)
5094 // Do we have the exact type of src?
5095 bool could_have_src = src_spec;
5096 // Do we have the exact type of dest?
5097 bool could_have_dest = dest_spec;
5098 ciKlass* src_k = top_src->klass();
5099 ciKlass* dest_k = top_dest->klass();
5100 if (!src_spec) {
5101 src_k = src_type->speculative_type_not_null();
5102 if (src_k != NULL && src_k->is_array_klass()) {
5103 could_have_src = true;
5104 }
5105 }
5106 if (!dest_spec) {
5107 dest_k = dest_type->speculative_type_not_null();
5108 if (dest_k != NULL && dest_k->is_array_klass()) {
5109 could_have_dest = true;
5110 }
5111 }
5112 if (could_have_src && could_have_dest) {
5113 // If we can have both exact types, emit the missing guards
5114 if (could_have_src && !src_spec) {
5115 src = maybe_cast_profiled_obj(src, src_k, true);
5116 }
5117 if (could_have_dest && !dest_spec) {
5118 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5119 }
5120 }
5121 }
5122 }
5123
5124 ciMethod* trap_method = method();
5125 int trap_bci = bci();
5126 if (saved_jvms != NULL) {
5127 trap_method = alloc->jvms()->method();
5128 trap_bci = alloc->jvms()->bci();
5129 }
5130
5131 bool negative_length_guard_generated = false;
5132
5133 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5134 can_emit_guards &&
5135 !src->is_top() && !dest->is_top()) {
5136 // validate arguments: enables transformation the ArrayCopyNode
5137 validated = true;
5138
5139 RegionNode* slow_region = new RegionNode(1);
5140 record_for_igvn(slow_region);
5141
5142 // (1) src and dest are arrays.
5143 generate_non_array_guard(load_object_klass(src), slow_region);
5144 generate_non_array_guard(load_object_klass(dest), slow_region);
5145
5146 // (2) src and dest arrays must have elements of the same BasicType
5147 // done at macro expansion or at Ideal transformation time
5148
5149 // (4) src_offset must not be negative.
5150 generate_negative_guard(src_offset, slow_region);
5151
5152 // (5) dest_offset must not be negative.
5153 generate_negative_guard(dest_offset, slow_region);
5154
5155 // (7) src_offset + length must not exceed length of src.
5156 generate_limit_guard(src_offset, length,
5157 load_array_length(src),
5158 slow_region);
5159
5160 // (8) dest_offset + length must not exceed length of dest.
5161 generate_limit_guard(dest_offset, length,
5162 load_array_length(dest),
5163 slow_region);
5164
5165 // (6) length must not be negative.
5166 // This is also checked in generate_arraycopy() during macro expansion, but
5167 // we also have to check it here for the case where the ArrayCopyNode will
5168 // be eliminated by Escape Analysis.
5169 if (EliminateAllocations) {
5170 generate_negative_guard(length, slow_region);
5171 negative_length_guard_generated = true;
5172 }
5173
5174 // (9) each element of an oop array must be assignable
5175 Node* src_klass = load_object_klass(src);
5176 Node* dest_klass = load_object_klass(dest);
5177 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5178
5179 if (not_subtype_ctrl != top()) {
5180 PreserveJVMState pjvms(this);
5181 set_control(not_subtype_ctrl);
5182 uncommon_trap(Deoptimization::Reason_intrinsic,
5183 Deoptimization::Action_make_not_entrant);
5184 assert(stopped(), "Should be stopped");
5185 }
5186 {
5187 PreserveJVMState pjvms(this);
5188 set_control(_gvn.transform(slow_region));
5189 uncommon_trap(Deoptimization::Reason_intrinsic,
5190 Deoptimization::Action_make_not_entrant);
5191 assert(stopped(), "Should be stopped");
5192 }
5193 }
5194
5195 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5196
5197 if (stopped()) {
5198 return true;
5199 }
5200
5201 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5202 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5203 // so the compiler has a chance to eliminate them: during macro expansion,
5204 // we have to set their control (CastPP nodes are eliminated).
5205 load_object_klass(src), load_object_klass(dest),
5206 load_array_length(src), load_array_length(dest));
5207
5208 ac->set_arraycopy(validated);
5209
5210 Node* n = _gvn.transform(ac);
5211 if (n == ac) {
5212 ac->connect_outputs(this);
5213 } else {
5214 assert(validated, "shouldn't transform if all arguments not validated");
5215 set_all_memory(n);
5216 }
5217
5218 return true;
5219 }
5220
5221
5222 // Helper function which determines if an arraycopy immediately follows
5223 // an allocation, with no intervening tests or other escapes for the object.
5224 AllocateArrayNode*
5225 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5226 RegionNode* slow_region) {
5227 if (stopped()) return NULL; // no fast path
5228 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5229
5230 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5231 if (alloc == NULL) return NULL;
5232
5233 Node* rawmem = memory(Compile::AliasIdxRaw);
5234 // Is the allocation's memory state untouched?
5235 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5236 // Bail out if there have been raw-memory effects since the allocation.
5237 // (Example: There might have been a call or safepoint.)
5238 return NULL;
5239 }
5240 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5241 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5242 return NULL;
5243 }
5244
5245 // There must be no unexpected observers of this allocation.
5246 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5247 Node* obs = ptr->fast_out(i);
5248 if (obs != this->map()) {
5249 return NULL;
5250 }
5251 }
5252
5253 // This arraycopy must unconditionally follow the allocation of the ptr.
5254 Node* alloc_ctl = ptr->in(0);
5255 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5256
5257 Node* ctl = control();
5258 while (ctl != alloc_ctl) {
5259 // There may be guards which feed into the slow_region.
5260 // Any other control flow means that we might not get a chance
5261 // to finish initializing the allocated object.
5262 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5263 IfNode* iff = ctl->in(0)->as_If();
5264 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5265 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5266 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5267 ctl = iff->in(0); // This test feeds the known slow_region.
5268 continue;
5269 }
5270 // One more try: Various low-level checks bottom out in
5271 // uncommon traps. If the debug-info of the trap omits
5272 // any reference to the allocation, as we've already
5273 // observed, then there can be no objection to the trap.
5274 bool found_trap = false;
5275 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5276 Node* obs = not_ctl->fast_out(j);
5277 if (obs->in(0) == not_ctl && obs->is_Call() &&
5278 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5279 found_trap = true; break;
5280 }
5281 }
5282 if (found_trap) {
5283 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5284 continue;
5285 }
5286 }
5287 return NULL;
5288 }
5289
5290 // If we get this far, we have an allocation which immediately
5291 // precedes the arraycopy, and we can take over zeroing the new object.
5292 // The arraycopy will finish the initialization, and provide
5293 // a new control state to which we will anchor the destination pointer.
5294
5295 return alloc;
5296 }
5297
5298 //-------------inline_encodeISOArray-----------------------------------
5299 // encode char[] to byte[] in ISO_8859_1
5300 bool LibraryCallKit::inline_encodeISOArray() {
5301 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5302 // no receiver since it is static method
5303 Node *src = argument(0);
5304 Node *src_offset = argument(1);
5305 Node *dst = argument(2);
5306 Node *dst_offset = argument(3);
5307 Node *length = argument(4);
5308
5309 const Type* src_type = src->Value(&_gvn);
5310 const Type* dst_type = dst->Value(&_gvn);
5311 const TypeAryPtr* top_src = src_type->isa_aryptr();
5312 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5313 if (top_src == NULL || top_src->klass() == NULL ||
5314 top_dest == NULL || top_dest->klass() == NULL) {
5315 // failed array check
5316 return false;
5317 }
5318
5319 // Figure out the size and type of the elements we will be copying.
5320 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5321 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5322 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5323 return false;
5324 }
5325
5326 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5327 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5328 // 'src_start' points to src array + scaled offset
5329 // 'dst_start' points to dst array + scaled offset
5330
5331 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5332 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5333 enc = _gvn.transform(enc);
5334 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5335 set_memory(res_mem, mtype);
5336 set_result(enc);
5337 return true;
5338 }
5339
5340 //-------------inline_multiplyToLen-----------------------------------
5341 bool LibraryCallKit::inline_multiplyToLen() {
5342 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5343
5344 address stubAddr = StubRoutines::multiplyToLen();
5345 if (stubAddr == NULL) {
5346 return false; // Intrinsic's stub is not implemented on this platform
5347 }
5348 const char* stubName = "multiplyToLen";
5349
5350 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5351
5352 // no receiver because it is a static method
5353 Node* x = argument(0);
5354 Node* xlen = argument(1);
5355 Node* y = argument(2);
5356 Node* ylen = argument(3);
5357 Node* z = argument(4);
5358
5359 const Type* x_type = x->Value(&_gvn);
5360 const Type* y_type = y->Value(&_gvn);
5361 const TypeAryPtr* top_x = x_type->isa_aryptr();
5362 const TypeAryPtr* top_y = y_type->isa_aryptr();
5363 if (top_x == NULL || top_x->klass() == NULL ||
5364 top_y == NULL || top_y->klass() == NULL) {
5365 // failed array check
5366 return false;
5367 }
5368
5369 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5370 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5371 if (x_elem != T_INT || y_elem != T_INT) {
5372 return false;
5373 }
5374
5375 // Set the original stack and the reexecute bit for the interpreter to reexecute
5376 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5377 // on the return from z array allocation in runtime.
5378 { PreserveReexecuteState preexecs(this);
5379 jvms()->set_should_reexecute(true);
5380
5381 Node* x_start = array_element_address(x, intcon(0), x_elem);
5382 Node* y_start = array_element_address(y, intcon(0), y_elem);
5383 // 'x_start' points to x array + scaled xlen
5384 // 'y_start' points to y array + scaled ylen
5385
5386 // Allocate the result array
5387 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5388 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5389 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5390
5391 IdealKit ideal(this);
5392
5393 #define __ ideal.
5394 Node* one = __ ConI(1);
5395 Node* zero = __ ConI(0);
5396 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5397 __ set(need_alloc, zero);
5398 __ set(z_alloc, z);
5399 __ if_then(z, BoolTest::eq, null()); {
5400 __ increment (need_alloc, one);
5401 } __ else_(); {
5402 // Update graphKit memory and control from IdealKit.
5403 sync_kit(ideal);
5404 Node* zlen_arg = load_array_length(z);
5405 // Update IdealKit memory and control from graphKit.
5406 __ sync_kit(this);
5407 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5408 __ increment (need_alloc, one);
5409 } __ end_if();
5410 } __ end_if();
5411
5412 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5413 // Update graphKit memory and control from IdealKit.
5414 sync_kit(ideal);
5415 Node * narr = new_array(klass_node, zlen, 1);
5416 // Update IdealKit memory and control from graphKit.
5417 __ sync_kit(this);
5418 __ set(z_alloc, narr);
5419 } __ end_if();
5420
5421 sync_kit(ideal);
5422 z = __ value(z_alloc);
5423 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5424 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5425 // Final sync IdealKit and GraphKit.
5426 final_sync(ideal);
5427 #undef __
5428
5429 Node* z_start = array_element_address(z, intcon(0), T_INT);
5430
5431 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5432 OptoRuntime::multiplyToLen_Type(),
5433 stubAddr, stubName, TypePtr::BOTTOM,
5434 x_start, xlen, y_start, ylen, z_start, zlen);
5435 } // original reexecute is set back here
5436
5437 C->set_has_split_ifs(true); // Has chance for split-if optimization
5438 set_result(z);
5439 return true;
5440 }
5441
5442 //-------------inline_squareToLen------------------------------------
5443 bool LibraryCallKit::inline_squareToLen() {
5444 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5445
5446 address stubAddr = StubRoutines::squareToLen();
5447 if (stubAddr == NULL) {
5448 return false; // Intrinsic's stub is not implemented on this platform
5449 }
5450 const char* stubName = "squareToLen";
5451
5452 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5453
5454 Node* x = argument(0);
5455 Node* len = argument(1);
5456 Node* z = argument(2);
5457 Node* zlen = argument(3);
5458
5459 const Type* x_type = x->Value(&_gvn);
5460 const Type* z_type = z->Value(&_gvn);
5461 const TypeAryPtr* top_x = x_type->isa_aryptr();
5462 const TypeAryPtr* top_z = z_type->isa_aryptr();
5463 if (top_x == NULL || top_x->klass() == NULL ||
5464 top_z == NULL || top_z->klass() == NULL) {
5465 // failed array check
5466 return false;
5467 }
5468
5469 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5470 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5471 if (x_elem != T_INT || z_elem != T_INT) {
5472 return false;
5473 }
5474
5475
5476 Node* x_start = array_element_address(x, intcon(0), x_elem);
5477 Node* z_start = array_element_address(z, intcon(0), z_elem);
5478
5479 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5480 OptoRuntime::squareToLen_Type(),
5481 stubAddr, stubName, TypePtr::BOTTOM,
5482 x_start, len, z_start, zlen);
5483
5484 set_result(z);
5485 return true;
5486 }
5487
5488 //-------------inline_mulAdd------------------------------------------
5489 bool LibraryCallKit::inline_mulAdd() {
5490 assert(UseMulAddIntrinsic, "not implemented on this platform");
5491
5492 address stubAddr = StubRoutines::mulAdd();
5493 if (stubAddr == NULL) {
5494 return false; // Intrinsic's stub is not implemented on this platform
5495 }
5496 const char* stubName = "mulAdd";
5497
5498 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5499
5500 Node* out = argument(0);
5501 Node* in = argument(1);
5502 Node* offset = argument(2);
5503 Node* len = argument(3);
5504 Node* k = argument(4);
5505
5506 const Type* out_type = out->Value(&_gvn);
5507 const Type* in_type = in->Value(&_gvn);
5508 const TypeAryPtr* top_out = out_type->isa_aryptr();
5509 const TypeAryPtr* top_in = in_type->isa_aryptr();
5510 if (top_out == NULL || top_out->klass() == NULL ||
5511 top_in == NULL || top_in->klass() == NULL) {
5512 // failed array check
5513 return false;
5514 }
5515
5516 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5517 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5518 if (out_elem != T_INT || in_elem != T_INT) {
5519 return false;
5520 }
5521
5522 Node* outlen = load_array_length(out);
5523 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5524 Node* out_start = array_element_address(out, intcon(0), out_elem);
5525 Node* in_start = array_element_address(in, intcon(0), in_elem);
5526
5527 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5528 OptoRuntime::mulAdd_Type(),
5529 stubAddr, stubName, TypePtr::BOTTOM,
5530 out_start,in_start, new_offset, len, k);
5531 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5532 set_result(result);
5533 return true;
5534 }
5535
5536 //-------------inline_montgomeryMultiply-----------------------------------
5537 bool LibraryCallKit::inline_montgomeryMultiply() {
5538 address stubAddr = StubRoutines::montgomeryMultiply();
5539 if (stubAddr == NULL) {
5540 return false; // Intrinsic's stub is not implemented on this platform
5541 }
5542
5543 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5544 const char* stubName = "montgomery_multiply";
5545
5546 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5547
5548 Node* a = argument(0);
5549 Node* b = argument(1);
5550 Node* n = argument(2);
5551 Node* len = argument(3);
5552 Node* inv = argument(4);
5553 Node* m = argument(6);
5554
5555 const Type* a_type = a->Value(&_gvn);
5556 const TypeAryPtr* top_a = a_type->isa_aryptr();
5557 const Type* b_type = b->Value(&_gvn);
5558 const TypeAryPtr* top_b = b_type->isa_aryptr();
5559 const Type* n_type = a->Value(&_gvn);
5560 const TypeAryPtr* top_n = n_type->isa_aryptr();
5561 const Type* m_type = a->Value(&_gvn);
5562 const TypeAryPtr* top_m = m_type->isa_aryptr();
5563 if (top_a == NULL || top_a->klass() == NULL ||
5564 top_b == NULL || top_b->klass() == NULL ||
5565 top_n == NULL || top_n->klass() == NULL ||
5566 top_m == NULL || top_m->klass() == NULL) {
5567 // failed array check
5568 return false;
5569 }
5570
5571 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5572 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5573 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5574 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5575 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5576 return false;
5577 }
5578
5579 // Make the call
5580 {
5581 Node* a_start = array_element_address(a, intcon(0), a_elem);
5582 Node* b_start = array_element_address(b, intcon(0), b_elem);
5583 Node* n_start = array_element_address(n, intcon(0), n_elem);
5584 Node* m_start = array_element_address(m, intcon(0), m_elem);
5585
5586 Node* call = make_runtime_call(RC_LEAF,
5587 OptoRuntime::montgomeryMultiply_Type(),
5588 stubAddr, stubName, TypePtr::BOTTOM,
5589 a_start, b_start, n_start, len, inv, top(),
5590 m_start);
5591 set_result(m);
5592 }
5593
5594 return true;
5595 }
5596
5597 bool LibraryCallKit::inline_montgomerySquare() {
5598 address stubAddr = StubRoutines::montgomerySquare();
5599 if (stubAddr == NULL) {
5600 return false; // Intrinsic's stub is not implemented on this platform
5601 }
5602
5603 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5604 const char* stubName = "montgomery_square";
5605
5606 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5607
5608 Node* a = argument(0);
5609 Node* n = argument(1);
5610 Node* len = argument(2);
5611 Node* inv = argument(3);
5612 Node* m = argument(5);
5613
5614 const Type* a_type = a->Value(&_gvn);
5615 const TypeAryPtr* top_a = a_type->isa_aryptr();
5616 const Type* n_type = a->Value(&_gvn);
5617 const TypeAryPtr* top_n = n_type->isa_aryptr();
5618 const Type* m_type = a->Value(&_gvn);
5619 const TypeAryPtr* top_m = m_type->isa_aryptr();
5620 if (top_a == NULL || top_a->klass() == NULL ||
5621 top_n == NULL || top_n->klass() == NULL ||
5622 top_m == NULL || top_m->klass() == NULL) {
5623 // failed array check
5624 return false;
5625 }
5626
5627 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5628 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5629 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5630 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5631 return false;
5632 }
5633
5634 // Make the call
5635 {
5636 Node* a_start = array_element_address(a, intcon(0), a_elem);
5637 Node* n_start = array_element_address(n, intcon(0), n_elem);
5638 Node* m_start = array_element_address(m, intcon(0), m_elem);
5639
5640 Node* call = make_runtime_call(RC_LEAF,
5641 OptoRuntime::montgomerySquare_Type(),
5642 stubAddr, stubName, TypePtr::BOTTOM,
5643 a_start, n_start, len, inv, top(),
5644 m_start);
5645 set_result(m);
5646 }
5647
5648 return true;
5649 }
5650
5651 //-------------inline_vectorizedMismatch------------------------------
5652 bool LibraryCallKit::inline_vectorizedMismatch() {
5653 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5654
5655 address stubAddr = StubRoutines::vectorizedMismatch();
5656 if (stubAddr == NULL) {
5657 return false; // Intrinsic's stub is not implemented on this platform
5658 }
5659 const char* stubName = "vectorizedMismatch";
5660 int size_l = callee()->signature()->size();
5661 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5662
5663 Node* obja = argument(0);
5664 Node* aoffset = argument(1);
5665 Node* objb = argument(3);
5666 Node* boffset = argument(4);
5667 Node* length = argument(6);
5668 Node* scale = argument(7);
5669
5670 const Type* a_type = obja->Value(&_gvn);
5671 const Type* b_type = objb->Value(&_gvn);
5672 const TypeAryPtr* top_a = a_type->isa_aryptr();
5673 const TypeAryPtr* top_b = b_type->isa_aryptr();
5674 if (top_a == NULL || top_a->klass() == NULL ||
5675 top_b == NULL || top_b->klass() == NULL) {
5676 // failed array check
5677 return false;
5678 }
5679
5680 Node* call;
5681 jvms()->set_should_reexecute(true);
5682
5683 Node* obja_adr = make_unsafe_address(obja, aoffset);
5684 Node* objb_adr = make_unsafe_address(objb, boffset);
5685
5686 call = make_runtime_call(RC_LEAF,
5687 OptoRuntime::vectorizedMismatch_Type(),
5688 stubAddr, stubName, TypePtr::BOTTOM,
5689 obja_adr, objb_adr, length, scale);
5690
5691 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5692 set_result(result);
5693 return true;
5694 }
5695
5696 /**
5697 * Calculate CRC32 for byte.
5698 * int java.util.zip.CRC32.update(int crc, int b)
5699 */
5700 bool LibraryCallKit::inline_updateCRC32() {
5701 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5702 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5703 // no receiver since it is static method
5704 Node* crc = argument(0); // type: int
5705 Node* b = argument(1); // type: int
5706
5707 /*
5708 * int c = ~ crc;
5709 * b = timesXtoThe32[(b ^ c) & 0xFF];
5710 * b = b ^ (c >>> 8);
5711 * crc = ~b;
5712 */
5713
5714 Node* M1 = intcon(-1);
5715 crc = _gvn.transform(new XorINode(crc, M1));
5716 Node* result = _gvn.transform(new XorINode(crc, b));
5717 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5718
5719 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5720 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5721 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5722 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5723
5724 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5725 result = _gvn.transform(new XorINode(crc, result));
5726 result = _gvn.transform(new XorINode(result, M1));
5727 set_result(result);
5728 return true;
5729 }
5730
5731 /**
5732 * Calculate CRC32 for byte[] array.
5733 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5734 */
5735 bool LibraryCallKit::inline_updateBytesCRC32() {
5736 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5737 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5738 // no receiver since it is static method
5739 Node* crc = argument(0); // type: int
5740 Node* src = argument(1); // type: oop
5741 Node* offset = argument(2); // type: int
5742 Node* length = argument(3); // type: int
5743
5744 const Type* src_type = src->Value(&_gvn);
5745 const TypeAryPtr* top_src = src_type->isa_aryptr();
5746 if (top_src == NULL || top_src->klass() == NULL) {
5747 // failed array check
5748 return false;
5749 }
5750
5751 // Figure out the size and type of the elements we will be copying.
5752 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5753 if (src_elem != T_BYTE) {
5754 return false;
5755 }
5756
5757 // 'src_start' points to src array + scaled offset
5758 Node* src_start = array_element_address(src, offset, src_elem);
5759
5760 // We assume that range check is done by caller.
5761 // TODO: generate range check (offset+length < src.length) in debug VM.
5762
5763 // Call the stub.
5764 address stubAddr = StubRoutines::updateBytesCRC32();
5765 const char *stubName = "updateBytesCRC32";
5766
5767 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5768 stubAddr, stubName, TypePtr::BOTTOM,
5769 crc, src_start, length);
5770 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5771 set_result(result);
5772 return true;
5773 }
5774
5775 /**
5776 * Calculate CRC32 for ByteBuffer.
5777 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5778 */
5779 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5780 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5781 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5782 // no receiver since it is static method
5783 Node* crc = argument(0); // type: int
5784 Node* src = argument(1); // type: long
5785 Node* offset = argument(3); // type: int
5786 Node* length = argument(4); // type: int
5787
5788 src = ConvL2X(src); // adjust Java long to machine word
5789 Node* base = _gvn.transform(new CastX2PNode(src));
5790 offset = ConvI2X(offset);
5791
5792 // 'src_start' points to src array + scaled offset
5793 Node* src_start = basic_plus_adr(top(), base, offset);
5794
5795 // Call the stub.
5796 address stubAddr = StubRoutines::updateBytesCRC32();
5797 const char *stubName = "updateBytesCRC32";
5798
5799 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5800 stubAddr, stubName, TypePtr::BOTTOM,
5801 crc, src_start, length);
5802 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5803 set_result(result);
5804 return true;
5805 }
5806
5807 //------------------------------get_table_from_crc32c_class-----------------------
5808 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5809 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5810 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5811
5812 return table;
5813 }
5814
5815 //------------------------------inline_updateBytesCRC32C-----------------------
5816 //
5817 // Calculate CRC32C for byte[] array.
5818 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5819 //
5820 bool LibraryCallKit::inline_updateBytesCRC32C() {
5821 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5822 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5823 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5824 // no receiver since it is a static method
5825 Node* crc = argument(0); // type: int
5826 Node* src = argument(1); // type: oop
5827 Node* offset = argument(2); // type: int
5828 Node* end = argument(3); // type: int
5829
5830 Node* length = _gvn.transform(new SubINode(end, offset));
5831
5832 const Type* src_type = src->Value(&_gvn);
5833 const TypeAryPtr* top_src = src_type->isa_aryptr();
5834 if (top_src == NULL || top_src->klass() == NULL) {
5835 // failed array check
5836 return false;
5837 }
5838
5839 // Figure out the size and type of the elements we will be copying.
5840 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5841 if (src_elem != T_BYTE) {
5842 return false;
5843 }
5844
5845 // 'src_start' points to src array + scaled offset
5846 Node* src_start = array_element_address(src, offset, src_elem);
5847
5848 // static final int[] byteTable in class CRC32C
5849 Node* table = get_table_from_crc32c_class(callee()->holder());
5850 Node* table_start = array_element_address(table, intcon(0), T_INT);
5851
5852 // We assume that range check is done by caller.
5853 // TODO: generate range check (offset+length < src.length) in debug VM.
5854
5855 // Call the stub.
5856 address stubAddr = StubRoutines::updateBytesCRC32C();
5857 const char *stubName = "updateBytesCRC32C";
5858
5859 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5860 stubAddr, stubName, TypePtr::BOTTOM,
5861 crc, src_start, length, table_start);
5862 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5863 set_result(result);
5864 return true;
5865 }
5866
5867 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5868 //
5869 // Calculate CRC32C for DirectByteBuffer.
5870 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5871 //
5872 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5873 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5874 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5875 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5876 // no receiver since it is a static method
5877 Node* crc = argument(0); // type: int
5878 Node* src = argument(1); // type: long
5879 Node* offset = argument(3); // type: int
5880 Node* end = argument(4); // type: int
5881
5882 Node* length = _gvn.transform(new SubINode(end, offset));
5883
5884 src = ConvL2X(src); // adjust Java long to machine word
5885 Node* base = _gvn.transform(new CastX2PNode(src));
5886 offset = ConvI2X(offset);
5887
5888 // 'src_start' points to src array + scaled offset
5889 Node* src_start = basic_plus_adr(top(), base, offset);
5890
5891 // static final int[] byteTable in class CRC32C
5892 Node* table = get_table_from_crc32c_class(callee()->holder());
5893 Node* table_start = array_element_address(table, intcon(0), T_INT);
5894
5895 // Call the stub.
5896 address stubAddr = StubRoutines::updateBytesCRC32C();
5897 const char *stubName = "updateBytesCRC32C";
5898
5899 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5900 stubAddr, stubName, TypePtr::BOTTOM,
5901 crc, src_start, length, table_start);
5902 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5903 set_result(result);
5904 return true;
5905 }
5906
5907 //------------------------------inline_updateBytesAdler32----------------------
5908 //
5909 // Calculate Adler32 checksum for byte[] array.
5910 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5911 //
5912 bool LibraryCallKit::inline_updateBytesAdler32() {
5913 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5914 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5915 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5916 // no receiver since it is static method
5917 Node* crc = argument(0); // type: int
5918 Node* src = argument(1); // type: oop
5919 Node* offset = argument(2); // type: int
5920 Node* length = argument(3); // type: int
5921
5922 const Type* src_type = src->Value(&_gvn);
5923 const TypeAryPtr* top_src = src_type->isa_aryptr();
5924 if (top_src == NULL || top_src->klass() == NULL) {
5925 // failed array check
5926 return false;
5927 }
5928
5929 // Figure out the size and type of the elements we will be copying.
5930 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5931 if (src_elem != T_BYTE) {
5932 return false;
5933 }
5934
5935 // 'src_start' points to src array + scaled offset
5936 Node* src_start = array_element_address(src, offset, src_elem);
5937
5938 // We assume that range check is done by caller.
5939 // TODO: generate range check (offset+length < src.length) in debug VM.
5940
5941 // Call the stub.
5942 address stubAddr = StubRoutines::updateBytesAdler32();
5943 const char *stubName = "updateBytesAdler32";
5944
5945 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5946 stubAddr, stubName, TypePtr::BOTTOM,
5947 crc, src_start, length);
5948 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5949 set_result(result);
5950 return true;
5951 }
5952
5953 //------------------------------inline_updateByteBufferAdler32---------------
5954 //
5955 // Calculate Adler32 checksum for DirectByteBuffer.
5956 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5957 //
5958 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5959 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5960 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5961 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5962 // no receiver since it is static method
5963 Node* crc = argument(0); // type: int
5964 Node* src = argument(1); // type: long
5965 Node* offset = argument(3); // type: int
5966 Node* length = argument(4); // type: int
5967
5968 src = ConvL2X(src); // adjust Java long to machine word
5969 Node* base = _gvn.transform(new CastX2PNode(src));
5970 offset = ConvI2X(offset);
5971
5972 // 'src_start' points to src array + scaled offset
5973 Node* src_start = basic_plus_adr(top(), base, offset);
5974
5975 // Call the stub.
5976 address stubAddr = StubRoutines::updateBytesAdler32();
5977 const char *stubName = "updateBytesAdler32";
5978
5979 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5980 stubAddr, stubName, TypePtr::BOTTOM,
5981 crc, src_start, length);
5982
5983 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5984 set_result(result);
5985 return true;
5986 }
5987
5988 //----------------------------inline_reference_get----------------------------
5989 // public T java.lang.ref.Reference.get();
5990 bool LibraryCallKit::inline_reference_get() {
5991 const int referent_offset = java_lang_ref_Reference::referent_offset;
5992 guarantee(referent_offset > 0, "should have already been set");
5993
5994 // Get the argument:
5995 Node* reference_obj = null_check_receiver();
5996 if (stopped()) return true;
5997
5998 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5999
6000 ciInstanceKlass* klass = env()->Object_klass();
6001 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6002
6003 Node* no_ctrl = NULL;
6004 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6005
6006 // Use the pre-barrier to record the value in the referent field
6007 pre_barrier(false /* do_load */,
6008 control(),
6009 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6010 result /* pre_val */,
6011 T_OBJECT);
6012
6013 // Add memory barrier to prevent commoning reads from this field
6014 // across safepoint since GC can change its value.
6015 insert_mem_bar(Op_MemBarCPUOrder);
6016
6017 set_result(result);
6018 return true;
6019 }
6020
6021
6022 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6023 bool is_exact=true, bool is_static=false,
6024 ciInstanceKlass * fromKls=NULL) {
6025 if (fromKls == NULL) {
6026 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6027 assert(tinst != NULL, "obj is null");
6028 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6029 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6030 fromKls = tinst->klass()->as_instance_klass();
6031 } else {
6032 assert(is_static, "only for static field access");
6033 }
6034 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6035 ciSymbol::make(fieldTypeString),
6036 is_static);
6037
6038 assert (field != NULL, "undefined field");
6039 if (field == NULL) return (Node *) NULL;
6040
6041 if (is_static) {
6042 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6043 fromObj = makecon(tip);
6044 }
6045
6046 // Next code copied from Parse::do_get_xxx():
6047
6048 // Compute address and memory type.
6049 int offset = field->offset_in_bytes();
6050 bool is_vol = field->is_volatile();
6051 ciType* field_klass = field->type();
6052 assert(field_klass->is_loaded(), "should be loaded");
6053 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6054 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6055 BasicType bt = field->layout_type();
6056
6057 // Build the resultant type of the load
6058 const Type *type;
6059 if (bt == T_OBJECT) {
6060 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6061 } else {
6062 type = Type::get_const_basic_type(bt);
6063 }
6064
6065 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6066 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6067 }
6068 // Build the load.
6069 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6070 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6071 // If reference is volatile, prevent following memory ops from
6072 // floating up past the volatile read. Also prevents commoning
6073 // another volatile read.
6074 if (is_vol) {
6075 // Memory barrier includes bogus read of value to force load BEFORE membar
6076 insert_mem_bar(Op_MemBarAcquire, loadedField);
6077 }
6078 return loadedField;
6079 }
6080
6081 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6082 bool is_exact = true, bool is_static = false,
6083 ciInstanceKlass * fromKls = NULL) {
6084 if (fromKls == NULL) {
6085 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6086 assert(tinst != NULL, "obj is null");
6087 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6088 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6089 fromKls = tinst->klass()->as_instance_klass();
6090 }
6091 else {
6092 assert(is_static, "only for static field access");
6093 }
6094 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6095 ciSymbol::make(fieldTypeString),
6096 is_static);
6097
6098 assert(field != NULL, "undefined field");
6099 assert(!field->is_volatile(), "not defined for volatile fields");
6100
6101 if (is_static) {
6102 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6103 fromObj = makecon(tip);
6104 }
6105
6106 // Next code copied from Parse::do_get_xxx():
6107
6108 // Compute address and memory type.
6109 int offset = field->offset_in_bytes();
6110 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6111
6112 return adr;
6113 }
6114
6115 //------------------------------inline_aescrypt_Block-----------------------
6116 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6117 address stubAddr = NULL;
6118 const char *stubName;
6119 assert(UseAES, "need AES instruction support");
6120
6121 switch(id) {
6122 case vmIntrinsics::_aescrypt_encryptBlock:
6123 stubAddr = StubRoutines::aescrypt_encryptBlock();
6124 stubName = "aescrypt_encryptBlock";
6125 break;
6126 case vmIntrinsics::_aescrypt_decryptBlock:
6127 stubAddr = StubRoutines::aescrypt_decryptBlock();
6128 stubName = "aescrypt_decryptBlock";
6129 break;
6130 }
6131 if (stubAddr == NULL) return false;
6132
6133 Node* aescrypt_object = argument(0);
6134 Node* src = argument(1);
6135 Node* src_offset = argument(2);
6136 Node* dest = argument(3);
6137 Node* dest_offset = argument(4);
6138
6139 // (1) src and dest are arrays.
6140 const Type* src_type = src->Value(&_gvn);
6141 const Type* dest_type = dest->Value(&_gvn);
6142 const TypeAryPtr* top_src = src_type->isa_aryptr();
6143 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6144 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6145
6146 // for the quick and dirty code we will skip all the checks.
6147 // we are just trying to get the call to be generated.
6148 Node* src_start = src;
6149 Node* dest_start = dest;
6150 if (src_offset != NULL || dest_offset != NULL) {
6151 assert(src_offset != NULL && dest_offset != NULL, "");
6152 src_start = array_element_address(src, src_offset, T_BYTE);
6153 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6154 }
6155
6156 // now need to get the start of its expanded key array
6157 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6158 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6159 if (k_start == NULL) return false;
6160
6161 if (Matcher::pass_original_key_for_aes()) {
6162 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6163 // compatibility issues between Java key expansion and SPARC crypto instructions
6164 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6165 if (original_k_start == NULL) return false;
6166
6167 // Call the stub.
6168 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6169 stubAddr, stubName, TypePtr::BOTTOM,
6170 src_start, dest_start, k_start, original_k_start);
6171 } else {
6172 // Call the stub.
6173 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6174 stubAddr, stubName, TypePtr::BOTTOM,
6175 src_start, dest_start, k_start);
6176 }
6177
6178 return true;
6179 }
6180
6181 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6182 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6183 address stubAddr = NULL;
6184 const char *stubName = NULL;
6185
6186 assert(UseAES, "need AES instruction support");
6187
6188 switch(id) {
6189 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6190 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6191 stubName = "cipherBlockChaining_encryptAESCrypt";
6192 break;
6193 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6194 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6195 stubName = "cipherBlockChaining_decryptAESCrypt";
6196 break;
6197 }
6198 if (stubAddr == NULL) return false;
6199
6200 Node* cipherBlockChaining_object = argument(0);
6201 Node* src = argument(1);
6202 Node* src_offset = argument(2);
6203 Node* len = argument(3);
6204 Node* dest = argument(4);
6205 Node* dest_offset = argument(5);
6206
6207 // (1) src and dest are arrays.
6208 const Type* src_type = src->Value(&_gvn);
6209 const Type* dest_type = dest->Value(&_gvn);
6210 const TypeAryPtr* top_src = src_type->isa_aryptr();
6211 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6212 assert (top_src != NULL && top_src->klass() != NULL
6213 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6214
6215 // checks are the responsibility of the caller
6216 Node* src_start = src;
6217 Node* dest_start = dest;
6218 if (src_offset != NULL || dest_offset != NULL) {
6219 assert(src_offset != NULL && dest_offset != NULL, "");
6220 src_start = array_element_address(src, src_offset, T_BYTE);
6221 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6222 }
6223
6224 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6225 // (because of the predicated logic executed earlier).
6226 // so we cast it here safely.
6227 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6228
6229 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6230 if (embeddedCipherObj == NULL) return false;
6231
6232 // cast it to what we know it will be at runtime
6233 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6234 assert(tinst != NULL, "CBC obj is null");
6235 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6236 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6237 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6238
6239 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6240 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6241 const TypeOopPtr* xtype = aklass->as_instance_type();
6242 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6243 aescrypt_object = _gvn.transform(aescrypt_object);
6244
6245 // we need to get the start of the aescrypt_object's expanded key array
6246 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6247 if (k_start == NULL) return false;
6248
6249 // similarly, get the start address of the r vector
6250 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6251 if (objRvec == NULL) return false;
6252 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6253
6254 Node* cbcCrypt;
6255 if (Matcher::pass_original_key_for_aes()) {
6256 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6257 // compatibility issues between Java key expansion and SPARC crypto instructions
6258 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6259 if (original_k_start == NULL) return false;
6260
6261 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
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, original_k_start);
6266 } else {
6267 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6268 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6269 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6270 stubAddr, stubName, TypePtr::BOTTOM,
6271 src_start, dest_start, k_start, r_start, len);
6272 }
6273
6274 // return cipher length (int)
6275 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6276 set_result(retvalue);
6277 return true;
6278 }
6279
6280 //------------------------------inline_counterMode_AESCrypt-----------------------
6281 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6282 assert(UseAES, "need AES instruction support");
6283 if (!UseAESCTRIntrinsics) return false;
6284
6285 address stubAddr = NULL;
6286 const char *stubName = NULL;
6287 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6288 stubAddr = StubRoutines::counterMode_AESCrypt();
6289 stubName = "counterMode_AESCrypt";
6290 }
6291 if (stubAddr == NULL) return false;
6292
6293 Node* counterMode_object = argument(0);
6294 Node* src = argument(1);
6295 Node* src_offset = argument(2);
6296 Node* len = argument(3);
6297 Node* dest = argument(4);
6298 Node* dest_offset = argument(5);
6299
6300 // (1) src and dest are arrays.
6301 const Type* src_type = src->Value(&_gvn);
6302 const Type* dest_type = dest->Value(&_gvn);
6303 const TypeAryPtr* top_src = src_type->isa_aryptr();
6304 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6305 assert(top_src != NULL && top_src->klass() != NULL &&
6306 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6307
6308 // checks are the responsibility of the caller
6309 Node* src_start = src;
6310 Node* dest_start = dest;
6311 if (src_offset != NULL || dest_offset != NULL) {
6312 assert(src_offset != NULL && dest_offset != NULL, "");
6313 src_start = array_element_address(src, src_offset, T_BYTE);
6314 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6315 }
6316
6317 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6318 // (because of the predicated logic executed earlier).
6319 // so we cast it here safely.
6320 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6321 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6322 if (embeddedCipherObj == NULL) return false;
6323 // cast it to what we know it will be at runtime
6324 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6325 assert(tinst != NULL, "CTR obj is null");
6326 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6327 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6328 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6329 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6330 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6331 const TypeOopPtr* xtype = aklass->as_instance_type();
6332 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6333 aescrypt_object = _gvn.transform(aescrypt_object);
6334 // we need to get the start of the aescrypt_object's expanded key array
6335 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6336 if (k_start == NULL) return false;
6337 // similarly, get the start address of the r vector
6338 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6339 if (obj_counter == NULL) return false;
6340 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6341
6342 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6343 if (saved_encCounter == NULL) return false;
6344 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6345 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6346
6347 Node* ctrCrypt;
6348 if (Matcher::pass_original_key_for_aes()) {
6349 // no SPARC version for AES/CTR intrinsics now.
6350 return false;
6351 }
6352 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6353 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6354 OptoRuntime::counterMode_aescrypt_Type(),
6355 stubAddr, stubName, TypePtr::BOTTOM,
6356 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6357
6358 // return cipher length (int)
6359 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6360 set_result(retvalue);
6361 return true;
6362 }
6363
6364 //------------------------------get_key_start_from_aescrypt_object-----------------------
6365 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6366 #if defined(PPC64) || defined(S390)
6367 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6368 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6369 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6370 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6371 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6372 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6373 if (objSessionK == NULL) {
6374 return (Node *) NULL;
6375 }
6376 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6377 #else
6378 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6379 #endif // PPC64
6380 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6381 if (objAESCryptKey == NULL) return (Node *) NULL;
6382
6383 // now have the array, need to get the start address of the K array
6384 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6385 return k_start;
6386 }
6387
6388 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6389 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6390 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6391 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6392 if (objAESCryptKey == NULL) return (Node *) NULL;
6393
6394 // now have the array, need to get the start address of the lastKey array
6395 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6396 return original_k_start;
6397 }
6398
6399 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6400 // Return node representing slow path of predicate check.
6401 // the pseudo code we want to emulate with this predicate is:
6402 // for encryption:
6403 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6404 // for decryption:
6405 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6406 // note cipher==plain is more conservative than the original java code but that's OK
6407 //
6408 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6409 // The receiver was checked for NULL already.
6410 Node* objCBC = argument(0);
6411
6412 // Load embeddedCipher field of CipherBlockChaining object.
6413 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6414
6415 // get AESCrypt klass for instanceOf check
6416 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6417 // will have same classloader as CipherBlockChaining object
6418 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6419 assert(tinst != NULL, "CBCobj is null");
6420 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6421
6422 // we want to do an instanceof comparison against the AESCrypt class
6423 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6424 if (!klass_AESCrypt->is_loaded()) {
6425 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6426 Node* ctrl = control();
6427 set_control(top()); // no regular fast path
6428 return ctrl;
6429 }
6430 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6431
6432 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6433 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6434 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6435
6436 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6437
6438 // for encryption, we are done
6439 if (!decrypting)
6440 return instof_false; // even if it is NULL
6441
6442 // for decryption, we need to add a further check to avoid
6443 // taking the intrinsic path when cipher and plain are the same
6444 // see the original java code for why.
6445 RegionNode* region = new RegionNode(3);
6446 region->init_req(1, instof_false);
6447 Node* src = argument(1);
6448 Node* dest = argument(4);
6449 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6450 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6451 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6452 region->init_req(2, src_dest_conjoint);
6453
6454 record_for_igvn(region);
6455 return _gvn.transform(region);
6456 }
6457
6458 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6459 // Return node representing slow path of predicate check.
6460 // the pseudo code we want to emulate with this predicate is:
6461 // for encryption:
6462 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6463 // for decryption:
6464 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6465 // note cipher==plain is more conservative than the original java code but that's OK
6466 //
6467
6468 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6469 // The receiver was checked for NULL already.
6470 Node* objCTR = argument(0);
6471
6472 // Load embeddedCipher field of CipherBlockChaining object.
6473 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6474
6475 // get AESCrypt klass for instanceOf check
6476 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6477 // will have same classloader as CipherBlockChaining object
6478 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6479 assert(tinst != NULL, "CTRobj is null");
6480 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6481
6482 // we want to do an instanceof comparison against the AESCrypt class
6483 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6484 if (!klass_AESCrypt->is_loaded()) {
6485 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6486 Node* ctrl = control();
6487 set_control(top()); // no regular fast path
6488 return ctrl;
6489 }
6490
6491 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6492 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6493 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6494 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6495 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6496
6497 return instof_false; // even if it is NULL
6498 }
6499
6500 //------------------------------inline_ghash_processBlocks
6501 bool LibraryCallKit::inline_ghash_processBlocks() {
6502 address stubAddr;
6503 const char *stubName;
6504 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6505
6506 stubAddr = StubRoutines::ghash_processBlocks();
6507 stubName = "ghash_processBlocks";
6508
6509 Node* data = argument(0);
6510 Node* offset = argument(1);
6511 Node* len = argument(2);
6512 Node* state = argument(3);
6513 Node* subkeyH = argument(4);
6514
6515 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6516 assert(state_start, "state is NULL");
6517 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6518 assert(subkeyH_start, "subkeyH is NULL");
6519 Node* data_start = array_element_address(data, offset, T_BYTE);
6520 assert(data_start, "data is NULL");
6521
6522 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6523 OptoRuntime::ghash_processBlocks_Type(),
6524 stubAddr, stubName, TypePtr::BOTTOM,
6525 state_start, subkeyH_start, data_start, len);
6526 return true;
6527 }
6528
6529 //------------------------------inline_sha_implCompress-----------------------
6530 //
6531 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6532 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6533 //
6534 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6535 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6536 //
6537 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6538 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6539 //
6540 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6541 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6542
6543 Node* sha_obj = argument(0);
6544 Node* src = argument(1); // type oop
6545 Node* ofs = argument(2); // type int
6546
6547 const Type* src_type = src->Value(&_gvn);
6548 const TypeAryPtr* top_src = src_type->isa_aryptr();
6549 if (top_src == NULL || top_src->klass() == NULL) {
6550 // failed array check
6551 return false;
6552 }
6553 // Figure out the size and type of the elements we will be copying.
6554 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6555 if (src_elem != T_BYTE) {
6556 return false;
6557 }
6558 // 'src_start' points to src array + offset
6559 Node* src_start = array_element_address(src, ofs, src_elem);
6560 Node* state = NULL;
6561 address stubAddr;
6562 const char *stubName;
6563
6564 switch(id) {
6565 case vmIntrinsics::_sha_implCompress:
6566 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6567 state = get_state_from_sha_object(sha_obj);
6568 stubAddr = StubRoutines::sha1_implCompress();
6569 stubName = "sha1_implCompress";
6570 break;
6571 case vmIntrinsics::_sha2_implCompress:
6572 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6573 state = get_state_from_sha_object(sha_obj);
6574 stubAddr = StubRoutines::sha256_implCompress();
6575 stubName = "sha256_implCompress";
6576 break;
6577 case vmIntrinsics::_sha5_implCompress:
6578 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6579 state = get_state_from_sha5_object(sha_obj);
6580 stubAddr = StubRoutines::sha512_implCompress();
6581 stubName = "sha512_implCompress";
6582 break;
6583 default:
6584 fatal_unexpected_iid(id);
6585 return false;
6586 }
6587 if (state == NULL) return false;
6588
6589 // Call the stub.
6590 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6591 stubAddr, stubName, TypePtr::BOTTOM,
6592 src_start, state);
6593
6594 return true;
6595 }
6596
6597 //------------------------------inline_digestBase_implCompressMB-----------------------
6598 //
6599 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6600 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6601 //
6602 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6603 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6604 "need SHA1/SHA256/SHA512 instruction support");
6605 assert((uint)predicate < 3, "sanity");
6606 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6607
6608 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6609 Node* src = argument(1); // byte[] array
6610 Node* ofs = argument(2); // type int
6611 Node* limit = argument(3); // type int
6612
6613 const Type* src_type = src->Value(&_gvn);
6614 const TypeAryPtr* top_src = src_type->isa_aryptr();
6615 if (top_src == NULL || top_src->klass() == NULL) {
6616 // failed array check
6617 return false;
6618 }
6619 // Figure out the size and type of the elements we will be copying.
6620 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6621 if (src_elem != T_BYTE) {
6622 return false;
6623 }
6624 // 'src_start' points to src array + offset
6625 Node* src_start = array_element_address(src, ofs, src_elem);
6626
6627 const char* klass_SHA_name = NULL;
6628 const char* stub_name = NULL;
6629 address stub_addr = NULL;
6630 bool long_state = false;
6631
6632 switch (predicate) {
6633 case 0:
6634 if (UseSHA1Intrinsics) {
6635 klass_SHA_name = "sun/security/provider/SHA";
6636 stub_name = "sha1_implCompressMB";
6637 stub_addr = StubRoutines::sha1_implCompressMB();
6638 }
6639 break;
6640 case 1:
6641 if (UseSHA256Intrinsics) {
6642 klass_SHA_name = "sun/security/provider/SHA2";
6643 stub_name = "sha256_implCompressMB";
6644 stub_addr = StubRoutines::sha256_implCompressMB();
6645 }
6646 break;
6647 case 2:
6648 if (UseSHA512Intrinsics) {
6649 klass_SHA_name = "sun/security/provider/SHA5";
6650 stub_name = "sha512_implCompressMB";
6651 stub_addr = StubRoutines::sha512_implCompressMB();
6652 long_state = true;
6653 }
6654 break;
6655 default:
6656 fatal("unknown SHA intrinsic predicate: %d", predicate);
6657 }
6658 if (klass_SHA_name != NULL) {
6659 // get DigestBase klass to lookup for SHA klass
6660 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6661 assert(tinst != NULL, "digestBase_obj is not instance???");
6662 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6663
6664 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6665 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6666 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6667 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6668 }
6669 return false;
6670 }
6671 //------------------------------inline_sha_implCompressMB-----------------------
6672 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6673 bool long_state, address stubAddr, const char *stubName,
6674 Node* src_start, Node* ofs, Node* limit) {
6675 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6676 const TypeOopPtr* xtype = aklass->as_instance_type();
6677 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6678 sha_obj = _gvn.transform(sha_obj);
6679
6680 Node* state;
6681 if (long_state) {
6682 state = get_state_from_sha5_object(sha_obj);
6683 } else {
6684 state = get_state_from_sha_object(sha_obj);
6685 }
6686 if (state == NULL) return false;
6687
6688 // Call the stub.
6689 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6690 OptoRuntime::digestBase_implCompressMB_Type(),
6691 stubAddr, stubName, TypePtr::BOTTOM,
6692 src_start, state, ofs, limit);
6693 // return ofs (int)
6694 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6695 set_result(result);
6696
6697 return true;
6698 }
6699
6700 //------------------------------get_state_from_sha_object-----------------------
6701 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6702 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6703 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6704 if (sha_state == NULL) return (Node *) NULL;
6705
6706 // now have the array, need to get the start address of the state array
6707 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6708 return state;
6709 }
6710
6711 //------------------------------get_state_from_sha5_object-----------------------
6712 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6713 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6714 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6715 if (sha_state == NULL) return (Node *) NULL;
6716
6717 // now have the array, need to get the start address of the state array
6718 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6719 return state;
6720 }
6721
6722 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6723 // Return node representing slow path of predicate check.
6724 // the pseudo code we want to emulate with this predicate is:
6725 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6726 //
6727 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6728 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6729 "need SHA1/SHA256/SHA512 instruction support");
6730 assert((uint)predicate < 3, "sanity");
6731
6732 // The receiver was checked for NULL already.
6733 Node* digestBaseObj = argument(0);
6734
6735 // get DigestBase klass for instanceOf check
6736 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6737 assert(tinst != NULL, "digestBaseObj is null");
6738 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6739
6740 const char* klass_SHA_name = NULL;
6741 switch (predicate) {
6742 case 0:
6743 if (UseSHA1Intrinsics) {
6744 // we want to do an instanceof comparison against the SHA class
6745 klass_SHA_name = "sun/security/provider/SHA";
6746 }
6747 break;
6748 case 1:
6749 if (UseSHA256Intrinsics) {
6750 // we want to do an instanceof comparison against the SHA2 class
6751 klass_SHA_name = "sun/security/provider/SHA2";
6752 }
6753 break;
6754 case 2:
6755 if (UseSHA512Intrinsics) {
6756 // we want to do an instanceof comparison against the SHA5 class
6757 klass_SHA_name = "sun/security/provider/SHA5";
6758 }
6759 break;
6760 default:
6761 fatal("unknown SHA intrinsic predicate: %d", predicate);
6762 }
6763
6764 ciKlass* klass_SHA = NULL;
6765 if (klass_SHA_name != NULL) {
6766 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6767 }
6768 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6769 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6770 Node* ctrl = control();
6771 set_control(top()); // no intrinsic path
6772 return ctrl;
6773 }
6774 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6775
6776 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6777 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6778 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6779 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6780
6781 return instof_false; // even if it is NULL
6782 }
6783
6784 //-------------inline_fma-----------------------------------
6785 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6786 Node *a = NULL;
6787 Node *b = NULL;
6788 Node *c = NULL;
6789 Node* result = NULL;
6790 switch (id) {
6791 case vmIntrinsics::_fmaD:
6792 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6793 // no receiver since it is static method
6794 a = round_double_node(argument(0));
6795 b = round_double_node(argument(2));
6796 c = round_double_node(argument(4));
6797 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6798 break;
6799 case vmIntrinsics::_fmaF:
6800 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6801 a = argument(0);
6802 b = argument(1);
6803 c = argument(2);
6804 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6805 break;
6806 default:
6807 fatal_unexpected_iid(id); break;
6808 }
6809 set_result(result);
6810 return true;
6811 }
6812
6813 bool LibraryCallKit::inline_profileBoolean() {
6814 Node* counts = argument(1);
6815 const TypeAryPtr* ary = NULL;
6816 ciArray* aobj = NULL;
6817 if (counts->is_Con()
6818 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6819 && (aobj = ary->const_oop()->as_array()) != NULL
6820 && (aobj->length() == 2)) {
6821 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6822 jint false_cnt = aobj->element_value(0).as_int();
6823 jint true_cnt = aobj->element_value(1).as_int();
6824
6825 if (C->log() != NULL) {
6826 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6827 false_cnt, true_cnt);
6828 }
6829
6830 if (false_cnt + true_cnt == 0) {
6831 // According to profile, never executed.
6832 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6833 Deoptimization::Action_reinterpret);
6834 return true;
6835 }
6836
6837 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6838 // is a number of each value occurrences.
6839 Node* result = argument(0);
6840 if (false_cnt == 0 || true_cnt == 0) {
6841 // According to profile, one value has been never seen.
6842 int expected_val = (false_cnt == 0) ? 1 : 0;
6843
6844 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6845 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6846
6847 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6848 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6849 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6850
6851 { // Slow path: uncommon trap for never seen value and then reexecute
6852 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6853 // the value has been seen at least once.
6854 PreserveJVMState pjvms(this);
6855 PreserveReexecuteState preexecs(this);
6856 jvms()->set_should_reexecute(true);
6857
6858 set_control(slow_path);
6859 set_i_o(i_o());
6860
6861 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6862 Deoptimization::Action_reinterpret);
6863 }
6864 // The guard for never seen value enables sharpening of the result and
6865 // returning a constant. It allows to eliminate branches on the same value
6866 // later on.
6867 set_control(fast_path);
6868 result = intcon(expected_val);
6869 }
6870 // Stop profiling.
6871 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6872 // By replacing method body with profile data (represented as ProfileBooleanNode
6873 // on IR level) we effectively disable profiling.
6874 // It enables full speed execution once optimized code is generated.
6875 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6876 C->record_for_igvn(profile);
6877 set_result(profile);
6878 return true;
6879 } else {
6880 // Continue profiling.
6881 // Profile data isn't available at the moment. So, execute method's bytecode version.
6882 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6883 // is compiled and counters aren't available since corresponding MethodHandle
6884 // isn't a compile-time constant.
6885 return false;
6886 }
6887 }
6888
6889 bool LibraryCallKit::inline_isCompileConstant() {
6890 Node* n = argument(0);
6891 set_result(n->is_Con() ? intcon(1) : intcon(0));
6892 return true;
6893 }