1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "prims/unsafe.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #ifdef TRACE_HAVE_INTRINSICS
56 #include "trace/traceMacros.hpp"
57 #endif
58
59 class LibraryIntrinsic : public InlineCallGenerator {
60 // Extend the set of intrinsics known to the runtime:
61 public:
62 private:
63 bool _is_virtual;
64 bool _does_virtual_dispatch;
65 int8_t _predicates_count; // Intrinsic is predicated by several conditions
66 int8_t _last_predicate; // Last generated predicate
67 vmIntrinsics::ID _intrinsic_id;
68
69 public:
70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
71 : InlineCallGenerator(m),
72 _is_virtual(is_virtual),
73 _does_virtual_dispatch(does_virtual_dispatch),
74 _predicates_count((int8_t)predicates_count),
75 _last_predicate((int8_t)-1),
76 _intrinsic_id(id)
77 {
78 }
79 virtual bool is_intrinsic() const { return true; }
80 virtual bool is_virtual() const { return _is_virtual; }
81 virtual bool is_predicated() const { return _predicates_count > 0; }
82 virtual int predicates_count() const { return _predicates_count; }
83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
84 virtual JVMState* generate(JVMState* jvms);
85 virtual Node* generate_predicate(JVMState* jvms, int predicate);
86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
87 };
88
89
90 // Local helper class for LibraryIntrinsic:
91 class LibraryCallKit : public GraphKit {
92 private:
93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
94 Node* _result; // the result node, if any
95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
96
97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
98
99 public:
100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
101 : GraphKit(jvms),
102 _intrinsic(intrinsic),
103 _result(NULL)
104 {
105 // Check if this is a root compile. In that case we don't have a caller.
106 if (!jvms->has_method()) {
107 _reexecute_sp = sp();
108 } else {
109 // Find out how many arguments the interpreter needs when deoptimizing
110 // and save the stack pointer value so it can used by uncommon_trap.
111 // We find the argument count by looking at the declared signature.
112 bool ignored_will_link;
113 ciSignature* declared_signature = NULL;
114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
117 }
118 }
119
120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
121
122 ciMethod* caller() const { return jvms()->method(); }
123 int bci() const { return jvms()->bci(); }
124 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
126 ciMethod* callee() const { return _intrinsic->method(); }
127
128 bool try_to_inline(int predicate);
129 Node* try_to_predicate(int predicate);
130
131 void push_result() {
132 // Push the result onto the stack.
133 if (!stopped() && result() != NULL) {
134 BasicType bt = result()->bottom_type()->basic_type();
135 push_node(bt, result());
136 }
137 }
138
139 private:
140 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
142 }
143
144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
145 void set_result(RegionNode* region, PhiNode* value);
146 Node* result() { return _result; }
147
148 virtual int reexecute_sp() { return _reexecute_sp; }
149
150 // Helper functions to inline natives
151 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
152 Node* generate_slow_guard(Node* test, RegionNode* region);
153 Node* generate_fair_guard(Node* test, RegionNode* region);
154 Node* generate_negative_guard(Node* index, RegionNode* region,
155 // resulting CastII of index:
156 Node* *pos_index = NULL);
157 Node* generate_limit_guard(Node* offset, Node* subseq_length,
158 Node* array_length,
159 RegionNode* region);
160 void generate_string_range_check(Node* array, Node* offset,
161 Node* length, bool char_count);
162 Node* generate_current_thread(Node* &tls_output);
163 Node* load_mirror_from_klass(Node* klass);
164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path,
166 int offset);
167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
175 RegionNode* region, int null_path) {
176 int offset = java_lang_Class::array_klass_offset_in_bytes();
177 return load_klass_from_mirror_common(mirror, never_see_null,
178 region, null_path,
179 offset);
180 }
181 Node* generate_access_flags_guard(Node* kls,
182 int modifier_mask, int modifier_bits,
183 RegionNode* region);
184 Node* generate_interface_guard(Node* kls, RegionNode* region);
185
186 enum ArrayKind {
187 AnyArray,
188 NonArray,
189 ObjectArray,
190 NonObjectArray,
191 TypeArray,
192 ValueArray
193 };
194
195 Node* generate_array_guard(Node* kls, RegionNode* region) {
196 return generate_array_guard_common(kls, region, AnyArray);
197 }
198 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
199 return generate_array_guard_common(kls, region, NonArray);
200 }
201 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
202 return generate_array_guard_common(kls, region, ObjectArray);
203 }
204 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
205 return generate_array_guard_common(kls, region, NonObjectArray);
206 }
207 Node* generate_typeArray_guard(Node* kls, RegionNode* region) {
208 return generate_array_guard_common(kls, region, TypeArray);
209 }
210 Node* generate_valueArray_guard(Node* kls, RegionNode* region) {
211 return generate_array_guard_common(kls, region, ValueArray);
212 }
213 Node* generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind);
214 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
215 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
216 bool is_virtual = false, bool is_static = false);
217 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
218 return generate_method_call(method_id, false, true);
219 }
220 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
221 return generate_method_call(method_id, true, false);
222 }
223 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
224 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
225
226 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
227 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
228 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
229 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
230 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
231 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
232 bool inline_string_indexOfChar();
233 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
234 bool inline_string_toBytesU();
235 bool inline_string_getCharsU();
236 bool inline_string_copy(bool compress);
237 bool inline_string_char_access(bool is_store);
238 Node* round_double_node(Node* n);
239 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
240 bool inline_math_native(vmIntrinsics::ID id);
241 bool inline_math(vmIntrinsics::ID id);
242 template <typename OverflowOp>
243 bool inline_math_overflow(Node* arg1, Node* arg2);
244 void inline_math_mathExact(Node* math, Node* test);
245 bool inline_math_addExactI(bool is_increment);
246 bool inline_math_addExactL(bool is_increment);
247 bool inline_math_multiplyExactI();
248 bool inline_math_multiplyExactL();
249 bool inline_math_multiplyHigh();
250 bool inline_math_negateExactI();
251 bool inline_math_negateExactL();
252 bool inline_math_subtractExactI(bool is_decrement);
253 bool inline_math_subtractExactL(bool is_decrement);
254 bool inline_min_max(vmIntrinsics::ID id);
255 bool inline_notify(vmIntrinsics::ID id);
256 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
257 // This returns Type::AnyPtr, RawPtr, or OopPtr.
258 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
259 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
260 // Helper for inline_unsafe_access.
261 // Generates the guards that check whether the result of
262 // Unsafe.getObject should be recorded in an SATB log buffer.
263 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
264
265 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
266 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
267 static bool klass_needs_init_guard(Node* kls);
268 bool inline_unsafe_allocate();
269 bool inline_unsafe_newArray(bool uninitialized);
270 bool inline_unsafe_copyMemory();
271 bool inline_native_currentThread();
272
273 bool inline_native_time_funcs(address method, const char* funcName);
274 #ifdef TRACE_HAVE_INTRINSICS
275 bool inline_native_classID();
276 bool inline_native_getBufferWriter();
277 #endif
278 bool inline_native_isInterrupted();
279 bool inline_native_Class_query(vmIntrinsics::ID id);
280 bool inline_native_subtype_check();
281 bool inline_native_getLength();
282 bool inline_array_copyOf(bool is_copyOfRange);
283 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
284 bool inline_preconditions_checkIndex();
285 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
286 bool inline_native_clone(bool is_virtual);
287 bool inline_native_Reflection_getCallerClass();
288 // Helper function for inlining native object hash method
289 bool inline_native_hashcode(bool is_virtual, bool is_static);
290 bool inline_native_getClass();
291
292 // Helper functions for inlining arraycopy
293 bool inline_arraycopy();
294 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
295 RegionNode* slow_region);
296 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
297 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
298 uint new_idx);
299
300 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
301 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
302 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
303 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
304 bool inline_unsafe_fence(vmIntrinsics::ID id);
305 bool inline_onspinwait();
306 bool inline_fp_conversions(vmIntrinsics::ID id);
307 bool inline_number_methods(vmIntrinsics::ID id);
308 bool inline_reference_get();
309 bool inline_Class_cast();
310 bool inline_aescrypt_Block(vmIntrinsics::ID id);
311 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
312 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
313 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
314 Node* inline_counterMode_AESCrypt_predicate();
315 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
316 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
317 bool inline_ghash_processBlocks();
318 bool inline_sha_implCompress(vmIntrinsics::ID id);
319 bool inline_digestBase_implCompressMB(int predicate);
320 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
321 bool long_state, address stubAddr, const char *stubName,
322 Node* src_start, Node* ofs, Node* limit);
323 Node* get_state_from_sha_object(Node *sha_object);
324 Node* get_state_from_sha5_object(Node *sha_object);
325 Node* inline_digestBase_implCompressMB_predicate(int predicate);
326 bool inline_encodeISOArray();
327 bool inline_updateCRC32();
328 bool inline_updateBytesCRC32();
329 bool inline_updateByteBufferCRC32();
330 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
331 bool inline_updateBytesCRC32C();
332 bool inline_updateDirectByteBufferCRC32C();
333 bool inline_updateBytesAdler32();
334 bool inline_updateByteBufferAdler32();
335 bool inline_multiplyToLen();
336 bool inline_hasNegatives();
337 bool inline_squareToLen();
338 bool inline_mulAdd();
339 bool inline_montgomeryMultiply();
340 bool inline_montgomerySquare();
341 bool inline_vectorizedMismatch();
342 bool inline_fma(vmIntrinsics::ID id);
343
344 bool inline_profileBoolean();
345 bool inline_isCompileConstant();
346 };
347
348 //---------------------------make_vm_intrinsic----------------------------
349 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
350 vmIntrinsics::ID id = m->intrinsic_id();
351 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
352
353 if (!m->is_loaded()) {
354 // Do not attempt to inline unloaded methods.
355 return NULL;
356 }
357
358 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
359 bool is_available = false;
360
361 {
362 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
363 // the compiler must transition to '_thread_in_vm' state because both
364 // methods access VM-internal data.
365 VM_ENTRY_MARK;
366 methodHandle mh(THREAD, m->get_Method());
367 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
368 !C->directive()->is_intrinsic_disabled(mh) &&
369 !vmIntrinsics::is_disabled_by_flags(mh);
370
371 }
372
373 if (is_available) {
374 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
375 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
376 return new LibraryIntrinsic(m, is_virtual,
377 vmIntrinsics::predicates_needed(id),
378 vmIntrinsics::does_virtual_dispatch(id),
379 (vmIntrinsics::ID) id);
380 } else {
381 return NULL;
382 }
383 }
384
385 //----------------------register_library_intrinsics-----------------------
386 // Initialize this file's data structures, for each Compile instance.
387 void Compile::register_library_intrinsics() {
388 // Nothing to do here.
389 }
390
391 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
392 LibraryCallKit kit(jvms, this);
393 Compile* C = kit.C;
394 int nodes = C->unique();
395 #ifndef PRODUCT
396 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
397 char buf[1000];
398 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
399 tty->print_cr("Intrinsic %s", str);
400 }
401 #endif
402 ciMethod* callee = kit.callee();
403 const int bci = kit.bci();
404
405 // Try to inline the intrinsic.
406 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
407 kit.try_to_inline(_last_predicate)) {
408 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
409 : "(intrinsic)";
410 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
411 if (C->print_intrinsics() || C->print_inlining()) {
412 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
413 }
414 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
415 if (C->log()) {
416 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
417 vmIntrinsics::name_at(intrinsic_id()),
418 (is_virtual() ? " virtual='1'" : ""),
419 C->unique() - nodes);
420 }
421 // Push the result from the inlined method onto the stack.
422 kit.push_result();
423 C->print_inlining_update(this);
424 return kit.transfer_exceptions_into_jvms();
425 }
426
427 // The intrinsic bailed out
428 if (jvms->has_method()) {
429 // Not a root compile.
430 const char* msg;
431 if (callee->intrinsic_candidate()) {
432 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
433 } else {
434 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
435 : "failed to inline (intrinsic), method not annotated";
436 }
437 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
438 if (C->print_intrinsics() || C->print_inlining()) {
439 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
440 }
441 } else {
442 // Root compile
443 ResourceMark rm;
444 stringStream msg_stream;
445 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
446 vmIntrinsics::name_at(intrinsic_id()),
447 is_virtual() ? " (virtual)" : "", bci);
448 const char *msg = msg_stream.as_string();
449 log_debug(jit, inlining)("%s", msg);
450 if (C->print_intrinsics() || C->print_inlining()) {
451 tty->print("%s", msg);
452 }
453 }
454 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
455 C->print_inlining_update(this);
456 return NULL;
457 }
458
459 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
460 LibraryCallKit kit(jvms, this);
461 Compile* C = kit.C;
462 int nodes = C->unique();
463 _last_predicate = predicate;
464 #ifndef PRODUCT
465 assert(is_predicated() && predicate < predicates_count(), "sanity");
466 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
467 char buf[1000];
468 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
469 tty->print_cr("Predicate for intrinsic %s", str);
470 }
471 #endif
472 ciMethod* callee = kit.callee();
473 const int bci = kit.bci();
474
475 Node* slow_ctl = kit.try_to_predicate(predicate);
476 if (!kit.failing()) {
477 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
478 : "(intrinsic, predicate)";
479 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
480 if (C->print_intrinsics() || C->print_inlining()) {
481 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
482 }
483 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
484 if (C->log()) {
485 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
486 vmIntrinsics::name_at(intrinsic_id()),
487 (is_virtual() ? " virtual='1'" : ""),
488 C->unique() - nodes);
489 }
490 return slow_ctl; // Could be NULL if the check folds.
491 }
492
493 // The intrinsic bailed out
494 if (jvms->has_method()) {
495 // Not a root compile.
496 const char* msg = "failed to generate predicate for intrinsic";
497 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
498 if (C->print_intrinsics() || C->print_inlining()) {
499 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
500 }
501 } else {
502 // Root compile
503 ResourceMark rm;
504 stringStream msg_stream;
505 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
506 vmIntrinsics::name_at(intrinsic_id()),
507 is_virtual() ? " (virtual)" : "", bci);
508 const char *msg = msg_stream.as_string();
509 log_debug(jit, inlining)("%s", msg);
510 if (C->print_intrinsics() || C->print_inlining()) {
511 C->print_inlining_stream()->print("%s", msg);
512 }
513 }
514 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
515 return NULL;
516 }
517
518 bool LibraryCallKit::try_to_inline(int predicate) {
519 // Handle symbolic names for otherwise undistinguished boolean switches:
520 const bool is_store = true;
521 const bool is_compress = true;
522 const bool is_static = true;
523 const bool is_volatile = true;
524
525 if (!jvms()->has_method()) {
526 // Root JVMState has a null method.
527 assert(map()->memory()->Opcode() == Op_Parm, "");
528 // Insert the memory aliasing node
529 set_all_memory(reset_memory());
530 }
531 assert(merged_memory(), "");
532
533
534 switch (intrinsic_id()) {
535 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
536 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
537 case vmIntrinsics::_getClass: return inline_native_getClass();
538
539 case vmIntrinsics::_dsin:
540 case vmIntrinsics::_dcos:
541 case vmIntrinsics::_dtan:
542 case vmIntrinsics::_dabs:
543 case vmIntrinsics::_datan2:
544 case vmIntrinsics::_dsqrt:
545 case vmIntrinsics::_dexp:
546 case vmIntrinsics::_dlog:
547 case vmIntrinsics::_dlog10:
548 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
549
550 case vmIntrinsics::_min:
551 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
552
553 case vmIntrinsics::_notify:
554 case vmIntrinsics::_notifyAll:
555 if (InlineNotify) {
556 return inline_notify(intrinsic_id());
557 }
558 return false;
559
560 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
561 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
562 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
563 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
564 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
565 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
566 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
567 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
568 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
569 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
570 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
571 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
572 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
573
574 case vmIntrinsics::_arraycopy: return inline_arraycopy();
575
576 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
577 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
578 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
579 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
580
581 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
582 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
583 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
584 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
585 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
586 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
587 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
588
589 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
590 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
591
592 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
593 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
594 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
595 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
596
597 case vmIntrinsics::_compressStringC:
598 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
599 case vmIntrinsics::_inflateStringC:
600 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
601
602 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
603 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
604 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
605 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
606 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
607 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
608 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
609 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
610 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
611
612 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
613 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
614 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
615 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
616 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
617 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
618 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
619 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
620 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
621
622 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
623 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
624 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
625 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
626 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
627 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
628 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
629 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
630 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
631
632 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
633 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
634 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
635 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
636 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
637 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
638 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
639 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
640 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
641
642 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
643 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
644 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
645 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
646
647 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
648 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
649 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
650 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
651
652 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
653 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
654 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
655 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
656 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
657 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
658 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
659 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
660 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
661
662 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
663 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
664 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
665 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
666 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
667 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
668 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
669 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
670 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
671
672 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
673 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
674 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
675 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
676 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
677 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
678 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
679 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
680 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
681
682 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
683 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
684 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
685 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
686 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
687 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
688 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
689 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
690 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
691
692 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
693 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
694 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
695 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
696 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
697
698 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
699 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
700 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
701 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
702 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
703 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
704 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
705 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
706 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
707 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
708 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
709 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
710 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
711 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
712 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
713 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
714 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
715 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
716 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
717 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
718
719 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
720 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
721 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
722 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
723 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
724 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
725 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
726 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
727 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
728 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
729 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
730 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
731 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
732 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
733 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
734
735 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
736 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
737 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
738 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
739
740 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
741 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
742 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
743 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
744 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
745
746 case vmIntrinsics::_loadFence:
747 case vmIntrinsics::_storeFence:
748 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
749
750 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
751
752 case vmIntrinsics::_currentThread: return inline_native_currentThread();
753 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
754
755 #ifdef TRACE_HAVE_INTRINSICS
756 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
757 case vmIntrinsics::_getClassId: return inline_native_classID();
758 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
759 #endif
760 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
761 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
762 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
763 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
764 case vmIntrinsics::_getLength: return inline_native_getLength();
765 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
766 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
767 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
768 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
769 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
770 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
771
772 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
773 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
774
775 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
776
777 case vmIntrinsics::_isInstance:
778 case vmIntrinsics::_getModifiers:
779 case vmIntrinsics::_isInterface:
780 case vmIntrinsics::_isArray:
781 case vmIntrinsics::_isPrimitive:
782 case vmIntrinsics::_getSuperclass:
783 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
784
785 case vmIntrinsics::_floatToRawIntBits:
786 case vmIntrinsics::_floatToIntBits:
787 case vmIntrinsics::_intBitsToFloat:
788 case vmIntrinsics::_doubleToRawLongBits:
789 case vmIntrinsics::_doubleToLongBits:
790 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
791
792 case vmIntrinsics::_numberOfLeadingZeros_i:
793 case vmIntrinsics::_numberOfLeadingZeros_l:
794 case vmIntrinsics::_numberOfTrailingZeros_i:
795 case vmIntrinsics::_numberOfTrailingZeros_l:
796 case vmIntrinsics::_bitCount_i:
797 case vmIntrinsics::_bitCount_l:
798 case vmIntrinsics::_reverseBytes_i:
799 case vmIntrinsics::_reverseBytes_l:
800 case vmIntrinsics::_reverseBytes_s:
801 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
802
803 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
804
805 case vmIntrinsics::_Reference_get: return inline_reference_get();
806
807 case vmIntrinsics::_Class_cast: return inline_Class_cast();
808
809 case vmIntrinsics::_aescrypt_encryptBlock:
810 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
811
812 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
813 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
814 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
815
816 case vmIntrinsics::_counterMode_AESCrypt:
817 return inline_counterMode_AESCrypt(intrinsic_id());
818
819 case vmIntrinsics::_sha_implCompress:
820 case vmIntrinsics::_sha2_implCompress:
821 case vmIntrinsics::_sha5_implCompress:
822 return inline_sha_implCompress(intrinsic_id());
823
824 case vmIntrinsics::_digestBase_implCompressMB:
825 return inline_digestBase_implCompressMB(predicate);
826
827 case vmIntrinsics::_multiplyToLen:
828 return inline_multiplyToLen();
829
830 case vmIntrinsics::_squareToLen:
831 return inline_squareToLen();
832
833 case vmIntrinsics::_mulAdd:
834 return inline_mulAdd();
835
836 case vmIntrinsics::_montgomeryMultiply:
837 return inline_montgomeryMultiply();
838 case vmIntrinsics::_montgomerySquare:
839 return inline_montgomerySquare();
840
841 case vmIntrinsics::_vectorizedMismatch:
842 return inline_vectorizedMismatch();
843
844 case vmIntrinsics::_ghash_processBlocks:
845 return inline_ghash_processBlocks();
846
847 case vmIntrinsics::_encodeISOArray:
848 case vmIntrinsics::_encodeByteISOArray:
849 return inline_encodeISOArray();
850
851 case vmIntrinsics::_updateCRC32:
852 return inline_updateCRC32();
853 case vmIntrinsics::_updateBytesCRC32:
854 return inline_updateBytesCRC32();
855 case vmIntrinsics::_updateByteBufferCRC32:
856 return inline_updateByteBufferCRC32();
857
858 case vmIntrinsics::_updateBytesCRC32C:
859 return inline_updateBytesCRC32C();
860 case vmIntrinsics::_updateDirectByteBufferCRC32C:
861 return inline_updateDirectByteBufferCRC32C();
862
863 case vmIntrinsics::_updateBytesAdler32:
864 return inline_updateBytesAdler32();
865 case vmIntrinsics::_updateByteBufferAdler32:
866 return inline_updateByteBufferAdler32();
867
868 case vmIntrinsics::_profileBoolean:
869 return inline_profileBoolean();
870 case vmIntrinsics::_isCompileConstant:
871 return inline_isCompileConstant();
872
873 case vmIntrinsics::_hasNegatives:
874 return inline_hasNegatives();
875
876 case vmIntrinsics::_fmaD:
877 case vmIntrinsics::_fmaF:
878 return inline_fma(intrinsic_id());
879
880 default:
881 // If you get here, it may be that someone has added a new intrinsic
882 // to the list in vmSymbols.hpp without implementing it here.
883 #ifndef PRODUCT
884 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
885 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
886 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
887 }
888 #endif
889 return false;
890 }
891 }
892
893 Node* LibraryCallKit::try_to_predicate(int predicate) {
894 if (!jvms()->has_method()) {
895 // Root JVMState has a null method.
896 assert(map()->memory()->Opcode() == Op_Parm, "");
897 // Insert the memory aliasing node
898 set_all_memory(reset_memory());
899 }
900 assert(merged_memory(), "");
901
902 switch (intrinsic_id()) {
903 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
904 return inline_cipherBlockChaining_AESCrypt_predicate(false);
905 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
906 return inline_cipherBlockChaining_AESCrypt_predicate(true);
907 case vmIntrinsics::_counterMode_AESCrypt:
908 return inline_counterMode_AESCrypt_predicate();
909 case vmIntrinsics::_digestBase_implCompressMB:
910 return inline_digestBase_implCompressMB_predicate(predicate);
911
912 default:
913 // If you get here, it may be that someone has added a new intrinsic
914 // to the list in vmSymbols.hpp without implementing it here.
915 #ifndef PRODUCT
916 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
917 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
918 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
919 }
920 #endif
921 Node* slow_ctl = control();
922 set_control(top()); // No fast path instrinsic
923 return slow_ctl;
924 }
925 }
926
927 //------------------------------set_result-------------------------------
928 // Helper function for finishing intrinsics.
929 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
930 record_for_igvn(region);
931 set_control(_gvn.transform(region));
932 set_result( _gvn.transform(value));
933 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
934 }
935
936 //------------------------------generate_guard---------------------------
937 // Helper function for generating guarded fast-slow graph structures.
938 // The given 'test', if true, guards a slow path. If the test fails
939 // then a fast path can be taken. (We generally hope it fails.)
940 // In all cases, GraphKit::control() is updated to the fast path.
941 // The returned value represents the control for the slow path.
942 // The return value is never 'top'; it is either a valid control
943 // or NULL if it is obvious that the slow path can never be taken.
944 // Also, if region and the slow control are not NULL, the slow edge
945 // is appended to the region.
946 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
947 if (stopped()) {
948 // Already short circuited.
949 return NULL;
950 }
951
952 // Build an if node and its projections.
953 // If test is true we take the slow path, which we assume is uncommon.
954 if (_gvn.type(test) == TypeInt::ZERO) {
955 // The slow branch is never taken. No need to build this guard.
956 return NULL;
957 }
958
959 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
960
961 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
962 if (if_slow == top()) {
963 // The slow branch is never taken. No need to build this guard.
964 return NULL;
965 }
966
967 if (region != NULL)
968 region->add_req(if_slow);
969
970 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
971 set_control(if_fast);
972
973 return if_slow;
974 }
975
976 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
977 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
978 }
979 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
980 return generate_guard(test, region, PROB_FAIR);
981 }
982
983 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
984 Node* *pos_index) {
985 if (stopped())
986 return NULL; // already stopped
987 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
988 return NULL; // index is already adequately typed
989 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
990 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
991 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
992 if (is_neg != NULL && pos_index != NULL) {
993 // Emulate effect of Parse::adjust_map_after_if.
994 Node* ccast = new CastIINode(index, TypeInt::POS);
995 ccast->set_req(0, control());
996 (*pos_index) = _gvn.transform(ccast);
997 }
998 return is_neg;
999 }
1000
1001 // Make sure that 'position' is a valid limit index, in [0..length].
1002 // There are two equivalent plans for checking this:
1003 // A. (offset + copyLength) unsigned<= arrayLength
1004 // B. offset <= (arrayLength - copyLength)
1005 // We require that all of the values above, except for the sum and
1006 // difference, are already known to be non-negative.
1007 // Plan A is robust in the face of overflow, if offset and copyLength
1008 // are both hugely positive.
1009 //
1010 // Plan B is less direct and intuitive, but it does not overflow at
1011 // all, since the difference of two non-negatives is always
1012 // representable. Whenever Java methods must perform the equivalent
1013 // check they generally use Plan B instead of Plan A.
1014 // For the moment we use Plan A.
1015 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1016 Node* subseq_length,
1017 Node* array_length,
1018 RegionNode* region) {
1019 if (stopped())
1020 return NULL; // already stopped
1021 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1022 if (zero_offset && subseq_length->eqv_uncast(array_length))
1023 return NULL; // common case of whole-array copy
1024 Node* last = subseq_length;
1025 if (!zero_offset) // last += offset
1026 last = _gvn.transform(new AddINode(last, offset));
1027 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1028 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1029 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1030 return is_over;
1031 }
1032
1033 // Emit range checks for the given String.value byte array
1034 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1035 if (stopped()) {
1036 return; // already stopped
1037 }
1038 RegionNode* bailout = new RegionNode(1);
1039 record_for_igvn(bailout);
1040 if (char_count) {
1041 // Convert char count to byte count
1042 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1043 }
1044
1045 // Offset and count must not be negative
1046 generate_negative_guard(offset, bailout);
1047 generate_negative_guard(count, bailout);
1048 // Offset + count must not exceed length of array
1049 generate_limit_guard(offset, count, load_array_length(array), bailout);
1050
1051 if (bailout->req() > 1) {
1052 PreserveJVMState pjvms(this);
1053 set_control(_gvn.transform(bailout));
1054 uncommon_trap(Deoptimization::Reason_intrinsic,
1055 Deoptimization::Action_maybe_recompile);
1056 }
1057 }
1058
1059 //--------------------------generate_current_thread--------------------
1060 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1061 ciKlass* thread_klass = env()->Thread_klass();
1062 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1063 Node* thread = _gvn.transform(new ThreadLocalNode());
1064 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1065 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1066 tls_output = thread;
1067 return threadObj;
1068 }
1069
1070
1071 //------------------------------make_string_method_node------------------------
1072 // Helper method for String intrinsic functions. This version is called with
1073 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1074 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1075 // containing the lengths of str1 and str2.
1076 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1077 Node* result = NULL;
1078 switch (opcode) {
1079 case Op_StrIndexOf:
1080 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1081 str1_start, cnt1, str2_start, cnt2, ae);
1082 break;
1083 case Op_StrComp:
1084 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1085 str1_start, cnt1, str2_start, cnt2, ae);
1086 break;
1087 case Op_StrEquals:
1088 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1089 // Use the constant length if there is one because optimized match rule may exist.
1090 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1091 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1092 break;
1093 default:
1094 ShouldNotReachHere();
1095 return NULL;
1096 }
1097
1098 // All these intrinsics have checks.
1099 C->set_has_split_ifs(true); // Has chance for split-if optimization
1100
1101 return _gvn.transform(result);
1102 }
1103
1104 //------------------------------inline_string_compareTo------------------------
1105 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1106 Node* arg1 = argument(0);
1107 Node* arg2 = argument(1);
1108
1109 // Get start addr and length of first argument
1110 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1111 Node* arg1_cnt = load_array_length(arg1);
1112
1113 // Get start addr and length of second argument
1114 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1115 Node* arg2_cnt = load_array_length(arg2);
1116
1117 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1118 set_result(result);
1119 return true;
1120 }
1121
1122 //------------------------------inline_string_equals------------------------
1123 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1124 Node* arg1 = argument(0);
1125 Node* arg2 = argument(1);
1126
1127 // paths (plus control) merge
1128 RegionNode* region = new RegionNode(3);
1129 Node* phi = new PhiNode(region, TypeInt::BOOL);
1130
1131 if (!stopped()) {
1132 // Get start addr and length of first argument
1133 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1134 Node* arg1_cnt = load_array_length(arg1);
1135
1136 // Get start addr and length of second argument
1137 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1138 Node* arg2_cnt = load_array_length(arg2);
1139
1140 // Check for arg1_cnt != arg2_cnt
1141 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1142 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1143 Node* if_ne = generate_slow_guard(bol, NULL);
1144 if (if_ne != NULL) {
1145 phi->init_req(2, intcon(0));
1146 region->init_req(2, if_ne);
1147 }
1148
1149 // Check for count == 0 is done by assembler code for StrEquals.
1150
1151 if (!stopped()) {
1152 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1153 phi->init_req(1, equals);
1154 region->init_req(1, control());
1155 }
1156 }
1157
1158 // post merge
1159 set_control(_gvn.transform(region));
1160 record_for_igvn(region);
1161
1162 set_result(_gvn.transform(phi));
1163 return true;
1164 }
1165
1166 //------------------------------inline_array_equals----------------------------
1167 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1168 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1169 Node* arg1 = argument(0);
1170 Node* arg2 = argument(1);
1171
1172 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1173 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1174 return true;
1175 }
1176
1177 //------------------------------inline_hasNegatives------------------------------
1178 bool LibraryCallKit::inline_hasNegatives() {
1179 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1180 return false;
1181 }
1182
1183 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1184 // no receiver since it is static method
1185 Node* ba = argument(0);
1186 Node* offset = argument(1);
1187 Node* len = argument(2);
1188
1189 // Range checks
1190 generate_string_range_check(ba, offset, len, false);
1191 if (stopped()) {
1192 return true;
1193 }
1194 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1195 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1196 set_result(_gvn.transform(result));
1197 return true;
1198 }
1199
1200 bool LibraryCallKit::inline_preconditions_checkIndex() {
1201 Node* index = argument(0);
1202 Node* length = argument(1);
1203 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1204 return false;
1205 }
1206
1207 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1208 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1209
1210 {
1211 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1212 uncommon_trap(Deoptimization::Reason_intrinsic,
1213 Deoptimization::Action_make_not_entrant);
1214 }
1215
1216 if (stopped()) {
1217 return false;
1218 }
1219
1220 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1221 BoolTest::mask btest = BoolTest::lt;
1222 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1223 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1224 _gvn.set_type(rc, rc->Value(&_gvn));
1225 if (!rc_bool->is_Con()) {
1226 record_for_igvn(rc);
1227 }
1228 set_control(_gvn.transform(new IfTrueNode(rc)));
1229 {
1230 PreserveJVMState pjvms(this);
1231 set_control(_gvn.transform(new IfFalseNode(rc)));
1232 uncommon_trap(Deoptimization::Reason_range_check,
1233 Deoptimization::Action_make_not_entrant);
1234 }
1235
1236 if (stopped()) {
1237 return false;
1238 }
1239
1240 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1241 result->set_req(0, control());
1242 result = _gvn.transform(result);
1243 set_result(result);
1244 replace_in_map(index, result);
1245 return true;
1246 }
1247
1248 //------------------------------inline_string_indexOf------------------------
1249 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1250 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1251 return false;
1252 }
1253 Node* src = argument(0);
1254 Node* tgt = argument(1);
1255
1256 // Make the merge point
1257 RegionNode* result_rgn = new RegionNode(4);
1258 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1259
1260 // Get start addr and length of source string
1261 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1262 Node* src_count = load_array_length(src);
1263
1264 // Get start addr and length of substring
1265 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1266 Node* tgt_count = load_array_length(tgt);
1267
1268 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1269 // Divide src size by 2 if String is UTF16 encoded
1270 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1271 }
1272 if (ae == StrIntrinsicNode::UU) {
1273 // Divide substring size by 2 if String is UTF16 encoded
1274 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1275 }
1276
1277 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1278 if (result != NULL) {
1279 result_phi->init_req(3, result);
1280 result_rgn->init_req(3, control());
1281 }
1282 set_control(_gvn.transform(result_rgn));
1283 record_for_igvn(result_rgn);
1284 set_result(_gvn.transform(result_phi));
1285
1286 return true;
1287 }
1288
1289 //-----------------------------inline_string_indexOf-----------------------
1290 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1291 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1292 return false;
1293 }
1294 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1295 return false;
1296 }
1297 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1298 Node* src = argument(0); // byte[]
1299 Node* src_count = argument(1); // char count
1300 Node* tgt = argument(2); // byte[]
1301 Node* tgt_count = argument(3); // char count
1302 Node* from_index = argument(4); // char index
1303
1304 // Multiply byte array index by 2 if String is UTF16 encoded
1305 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1306 src_count = _gvn.transform(new SubINode(src_count, from_index));
1307 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1308 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1309
1310 // Range checks
1311 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1312 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1313 if (stopped()) {
1314 return true;
1315 }
1316
1317 RegionNode* region = new RegionNode(5);
1318 Node* phi = new PhiNode(region, TypeInt::INT);
1319
1320 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1321 if (result != NULL) {
1322 // The result is index relative to from_index if substring was found, -1 otherwise.
1323 // Generate code which will fold into cmove.
1324 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1325 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1326
1327 Node* if_lt = generate_slow_guard(bol, NULL);
1328 if (if_lt != NULL) {
1329 // result == -1
1330 phi->init_req(3, result);
1331 region->init_req(3, if_lt);
1332 }
1333 if (!stopped()) {
1334 result = _gvn.transform(new AddINode(result, from_index));
1335 phi->init_req(4, result);
1336 region->init_req(4, control());
1337 }
1338 }
1339
1340 set_control(_gvn.transform(region));
1341 record_for_igvn(region);
1342 set_result(_gvn.transform(phi));
1343
1344 return true;
1345 }
1346
1347 // Create StrIndexOfNode with fast path checks
1348 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1349 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1350 // Check for substr count > string count
1351 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1352 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1353 Node* if_gt = generate_slow_guard(bol, NULL);
1354 if (if_gt != NULL) {
1355 phi->init_req(1, intcon(-1));
1356 region->init_req(1, if_gt);
1357 }
1358 if (!stopped()) {
1359 // Check for substr count == 0
1360 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1361 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1362 Node* if_zero = generate_slow_guard(bol, NULL);
1363 if (if_zero != NULL) {
1364 phi->init_req(2, intcon(0));
1365 region->init_req(2, if_zero);
1366 }
1367 }
1368 if (!stopped()) {
1369 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1370 }
1371 return NULL;
1372 }
1373
1374 //-----------------------------inline_string_indexOfChar-----------------------
1375 bool LibraryCallKit::inline_string_indexOfChar() {
1376 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1377 return false;
1378 }
1379 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1380 return false;
1381 }
1382 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1383 Node* src = argument(0); // byte[]
1384 Node* tgt = argument(1); // tgt is int ch
1385 Node* from_index = argument(2);
1386 Node* max = argument(3);
1387
1388 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1389 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1390 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1391
1392 // Range checks
1393 generate_string_range_check(src, src_offset, src_count, true);
1394 if (stopped()) {
1395 return true;
1396 }
1397
1398 RegionNode* region = new RegionNode(3);
1399 Node* phi = new PhiNode(region, TypeInt::INT);
1400
1401 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1402 C->set_has_split_ifs(true); // Has chance for split-if optimization
1403 _gvn.transform(result);
1404
1405 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1406 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1407
1408 Node* if_lt = generate_slow_guard(bol, NULL);
1409 if (if_lt != NULL) {
1410 // result == -1
1411 phi->init_req(2, result);
1412 region->init_req(2, if_lt);
1413 }
1414 if (!stopped()) {
1415 result = _gvn.transform(new AddINode(result, from_index));
1416 phi->init_req(1, result);
1417 region->init_req(1, control());
1418 }
1419 set_control(_gvn.transform(region));
1420 record_for_igvn(region);
1421 set_result(_gvn.transform(phi));
1422
1423 return true;
1424 }
1425 //---------------------------inline_string_copy---------------------
1426 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1427 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1428 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1429 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1430 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1431 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1432 bool LibraryCallKit::inline_string_copy(bool compress) {
1433 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1434 return false;
1435 }
1436 int nargs = 5; // 2 oops, 3 ints
1437 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1438
1439 Node* src = argument(0);
1440 Node* src_offset = argument(1);
1441 Node* dst = argument(2);
1442 Node* dst_offset = argument(3);
1443 Node* length = argument(4);
1444
1445 // Check for allocation before we add nodes that would confuse
1446 // tightly_coupled_allocation()
1447 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1448
1449 // Figure out the size and type of the elements we will be copying.
1450 const Type* src_type = src->Value(&_gvn);
1451 const Type* dst_type = dst->Value(&_gvn);
1452 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1453 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1454 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1455 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1456 "Unsupported array types for inline_string_copy");
1457
1458 // Convert char[] offsets to byte[] offsets
1459 bool convert_src = (compress && src_elem == T_BYTE);
1460 bool convert_dst = (!compress && dst_elem == T_BYTE);
1461 if (convert_src) {
1462 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1463 } else if (convert_dst) {
1464 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1465 }
1466
1467 // Range checks
1468 generate_string_range_check(src, src_offset, length, convert_src);
1469 generate_string_range_check(dst, dst_offset, length, convert_dst);
1470 if (stopped()) {
1471 return true;
1472 }
1473
1474 Node* src_start = array_element_address(src, src_offset, src_elem);
1475 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1476 // 'src_start' points to src array + scaled offset
1477 // 'dst_start' points to dst array + scaled offset
1478 Node* count = NULL;
1479 if (compress) {
1480 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1481 } else {
1482 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1483 }
1484
1485 if (alloc != NULL) {
1486 if (alloc->maybe_set_complete(&_gvn)) {
1487 // "You break it, you buy it."
1488 InitializeNode* init = alloc->initialization();
1489 assert(init->is_complete(), "we just did this");
1490 init->set_complete_with_arraycopy();
1491 assert(dst->is_CheckCastPP(), "sanity");
1492 assert(dst->in(0)->in(0) == init, "dest pinned");
1493 }
1494 // Do not let stores that initialize this object be reordered with
1495 // a subsequent store that would make this object accessible by
1496 // other threads.
1497 // Record what AllocateNode this StoreStore protects so that
1498 // escape analysis can go from the MemBarStoreStoreNode to the
1499 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1500 // based on the escape status of the AllocateNode.
1501 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1502 }
1503 if (compress) {
1504 set_result(_gvn.transform(count));
1505 }
1506 return true;
1507 }
1508
1509 #ifdef _LP64
1510 #define XTOP ,top() /*additional argument*/
1511 #else //_LP64
1512 #define XTOP /*no additional argument*/
1513 #endif //_LP64
1514
1515 //------------------------inline_string_toBytesU--------------------------
1516 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1517 bool LibraryCallKit::inline_string_toBytesU() {
1518 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1519 return false;
1520 }
1521 // Get the arguments.
1522 Node* value = argument(0);
1523 Node* offset = argument(1);
1524 Node* length = argument(2);
1525
1526 Node* newcopy = NULL;
1527
1528 // Set the original stack and the reexecute bit for the interpreter to reexecute
1529 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1530 { PreserveReexecuteState preexecs(this);
1531 jvms()->set_should_reexecute(true);
1532
1533 // Check if a null path was taken unconditionally.
1534 value = null_check(value);
1535
1536 RegionNode* bailout = new RegionNode(1);
1537 record_for_igvn(bailout);
1538
1539 // Range checks
1540 generate_negative_guard(offset, bailout);
1541 generate_negative_guard(length, bailout);
1542 generate_limit_guard(offset, length, load_array_length(value), bailout);
1543 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1544 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1545
1546 if (bailout->req() > 1) {
1547 PreserveJVMState pjvms(this);
1548 set_control(_gvn.transform(bailout));
1549 uncommon_trap(Deoptimization::Reason_intrinsic,
1550 Deoptimization::Action_maybe_recompile);
1551 }
1552 if (stopped()) {
1553 return true;
1554 }
1555
1556 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1557 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1558 newcopy = new_array(klass_node, size, 0); // no arguments to push
1559 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1560
1561 // Calculate starting addresses.
1562 Node* src_start = array_element_address(value, offset, T_CHAR);
1563 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1564
1565 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1566 const TypeInt* toffset = gvn().type(offset)->is_int();
1567 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1568
1569 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1570 const char* copyfunc_name = "arraycopy";
1571 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1572 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1573 OptoRuntime::fast_arraycopy_Type(),
1574 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1575 src_start, dst_start, ConvI2X(length) XTOP);
1576 // Do not let reads from the cloned object float above the arraycopy.
1577 if (alloc != NULL) {
1578 if (alloc->maybe_set_complete(&_gvn)) {
1579 // "You break it, you buy it."
1580 InitializeNode* init = alloc->initialization();
1581 assert(init->is_complete(), "we just did this");
1582 init->set_complete_with_arraycopy();
1583 assert(newcopy->is_CheckCastPP(), "sanity");
1584 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1585 }
1586 // Do not let stores that initialize this object be reordered with
1587 // a subsequent store that would make this object accessible by
1588 // other threads.
1589 // Record what AllocateNode this StoreStore protects so that
1590 // escape analysis can go from the MemBarStoreStoreNode to the
1591 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1592 // based on the escape status of the AllocateNode.
1593 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1594 } else {
1595 insert_mem_bar(Op_MemBarCPUOrder);
1596 }
1597 } // original reexecute is set back here
1598
1599 C->set_has_split_ifs(true); // Has chance for split-if optimization
1600 if (!stopped()) {
1601 set_result(newcopy);
1602 }
1603 return true;
1604 }
1605
1606 //------------------------inline_string_getCharsU--------------------------
1607 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1608 bool LibraryCallKit::inline_string_getCharsU() {
1609 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1610 return false;
1611 }
1612
1613 // Get the arguments.
1614 Node* src = argument(0);
1615 Node* src_begin = argument(1);
1616 Node* src_end = argument(2); // exclusive offset (i < src_end)
1617 Node* dst = argument(3);
1618 Node* dst_begin = argument(4);
1619
1620 // Check for allocation before we add nodes that would confuse
1621 // tightly_coupled_allocation()
1622 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1623
1624 // Check if a null path was taken unconditionally.
1625 src = null_check(src);
1626 dst = null_check(dst);
1627 if (stopped()) {
1628 return true;
1629 }
1630
1631 // Get length and convert char[] offset to byte[] offset
1632 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1633 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1634
1635 // Range checks
1636 generate_string_range_check(src, src_begin, length, true);
1637 generate_string_range_check(dst, dst_begin, length, false);
1638 if (stopped()) {
1639 return true;
1640 }
1641
1642 if (!stopped()) {
1643 // Calculate starting addresses.
1644 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1645 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1646
1647 // Check if array addresses are aligned to HeapWordSize
1648 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1649 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1650 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1651 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1652
1653 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1654 const char* copyfunc_name = "arraycopy";
1655 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1656 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1657 OptoRuntime::fast_arraycopy_Type(),
1658 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1659 src_start, dst_start, ConvI2X(length) XTOP);
1660 // Do not let reads from the cloned object float above the arraycopy.
1661 if (alloc != NULL) {
1662 if (alloc->maybe_set_complete(&_gvn)) {
1663 // "You break it, you buy it."
1664 InitializeNode* init = alloc->initialization();
1665 assert(init->is_complete(), "we just did this");
1666 init->set_complete_with_arraycopy();
1667 assert(dst->is_CheckCastPP(), "sanity");
1668 assert(dst->in(0)->in(0) == init, "dest pinned");
1669 }
1670 // Do not let stores that initialize this object be reordered with
1671 // a subsequent store that would make this object accessible by
1672 // other threads.
1673 // Record what AllocateNode this StoreStore protects so that
1674 // escape analysis can go from the MemBarStoreStoreNode to the
1675 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1676 // based on the escape status of the AllocateNode.
1677 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1678 } else {
1679 insert_mem_bar(Op_MemBarCPUOrder);
1680 }
1681 }
1682
1683 C->set_has_split_ifs(true); // Has chance for split-if optimization
1684 return true;
1685 }
1686
1687 //----------------------inline_string_char_access----------------------------
1688 // Store/Load char to/from byte[] array.
1689 // static void StringUTF16.putChar(byte[] val, int index, int c)
1690 // static char StringUTF16.getChar(byte[] val, int index)
1691 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1692 Node* value = argument(0);
1693 Node* index = argument(1);
1694 Node* ch = is_store ? argument(2) : NULL;
1695
1696 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1697 // correctly requires matched array shapes.
1698 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1699 "sanity: byte[] and char[] bases agree");
1700 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1701 "sanity: byte[] and char[] scales agree");
1702
1703 // Bail when getChar over constants is requested: constant folding would
1704 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1705 // Java method would constant fold nicely instead.
1706 if (!is_store && value->is_Con() && index->is_Con()) {
1707 return false;
1708 }
1709
1710 Node* adr = array_element_address(value, index, T_CHAR);
1711 if (adr->is_top()) {
1712 return false;
1713 }
1714 if (is_store) {
1715 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1716 false, false, true /* mismatched */);
1717 } else {
1718 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1719 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1720 set_result(ch);
1721 }
1722 return true;
1723 }
1724
1725 //--------------------------round_double_node--------------------------------
1726 // Round a double node if necessary.
1727 Node* LibraryCallKit::round_double_node(Node* n) {
1728 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1729 n = _gvn.transform(new RoundDoubleNode(0, n));
1730 return n;
1731 }
1732
1733 //------------------------------inline_math-----------------------------------
1734 // public static double Math.abs(double)
1735 // public static double Math.sqrt(double)
1736 // public static double Math.log(double)
1737 // public static double Math.log10(double)
1738 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1739 Node* arg = round_double_node(argument(0));
1740 Node* n = NULL;
1741 switch (id) {
1742 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1743 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1744 default: fatal_unexpected_iid(id); break;
1745 }
1746 set_result(_gvn.transform(n));
1747 return true;
1748 }
1749
1750 //------------------------------runtime_math-----------------------------
1751 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1752 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1753 "must be (DD)D or (D)D type");
1754
1755 // Inputs
1756 Node* a = round_double_node(argument(0));
1757 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1758
1759 const TypePtr* no_memory_effects = NULL;
1760 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1761 no_memory_effects,
1762 a, top(), b, b ? top() : NULL);
1763 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1764 #ifdef ASSERT
1765 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1766 assert(value_top == top(), "second value must be top");
1767 #endif
1768
1769 set_result(value);
1770 return true;
1771 }
1772
1773 //------------------------------inline_math_native-----------------------------
1774 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1775 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1776 switch (id) {
1777 // These intrinsics are not properly supported on all hardware
1778 case vmIntrinsics::_dsin:
1779 return StubRoutines::dsin() != NULL ?
1780 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1781 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1782 case vmIntrinsics::_dcos:
1783 return StubRoutines::dcos() != NULL ?
1784 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1785 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1786 case vmIntrinsics::_dtan:
1787 return StubRoutines::dtan() != NULL ?
1788 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1789 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1790 case vmIntrinsics::_dlog:
1791 return StubRoutines::dlog() != NULL ?
1792 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1793 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1794 case vmIntrinsics::_dlog10:
1795 return StubRoutines::dlog10() != NULL ?
1796 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1797 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1798
1799 // These intrinsics are supported on all hardware
1800 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1801 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1802
1803 case vmIntrinsics::_dexp:
1804 return StubRoutines::dexp() != NULL ?
1805 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1806 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1807 case vmIntrinsics::_dpow:
1808 return StubRoutines::dpow() != NULL ?
1809 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1810 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1811 #undef FN_PTR
1812
1813 // These intrinsics are not yet correctly implemented
1814 case vmIntrinsics::_datan2:
1815 return false;
1816
1817 default:
1818 fatal_unexpected_iid(id);
1819 return false;
1820 }
1821 }
1822
1823 static bool is_simple_name(Node* n) {
1824 return (n->req() == 1 // constant
1825 || (n->is_Type() && n->as_Type()->type()->singleton())
1826 || n->is_Proj() // parameter or return value
1827 || n->is_Phi() // local of some sort
1828 );
1829 }
1830
1831 //----------------------------inline_notify-----------------------------------*
1832 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1833 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1834 address func;
1835 if (id == vmIntrinsics::_notify) {
1836 func = OptoRuntime::monitor_notify_Java();
1837 } else {
1838 func = OptoRuntime::monitor_notifyAll_Java();
1839 }
1840 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1841 make_slow_call_ex(call, env()->Throwable_klass(), false);
1842 return true;
1843 }
1844
1845
1846 //----------------------------inline_min_max-----------------------------------
1847 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1848 set_result(generate_min_max(id, argument(0), argument(1)));
1849 return true;
1850 }
1851
1852 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1853 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1854 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1855 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1856 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1857
1858 {
1859 PreserveJVMState pjvms(this);
1860 PreserveReexecuteState preexecs(this);
1861 jvms()->set_should_reexecute(true);
1862
1863 set_control(slow_path);
1864 set_i_o(i_o());
1865
1866 uncommon_trap(Deoptimization::Reason_intrinsic,
1867 Deoptimization::Action_none);
1868 }
1869
1870 set_control(fast_path);
1871 set_result(math);
1872 }
1873
1874 template <typename OverflowOp>
1875 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1876 typedef typename OverflowOp::MathOp MathOp;
1877
1878 MathOp* mathOp = new MathOp(arg1, arg2);
1879 Node* operation = _gvn.transform( mathOp );
1880 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1881 inline_math_mathExact(operation, ofcheck);
1882 return true;
1883 }
1884
1885 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1886 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1887 }
1888
1889 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1890 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1891 }
1892
1893 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1894 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1895 }
1896
1897 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1898 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1899 }
1900
1901 bool LibraryCallKit::inline_math_negateExactI() {
1902 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1903 }
1904
1905 bool LibraryCallKit::inline_math_negateExactL() {
1906 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1907 }
1908
1909 bool LibraryCallKit::inline_math_multiplyExactI() {
1910 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1911 }
1912
1913 bool LibraryCallKit::inline_math_multiplyExactL() {
1914 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1915 }
1916
1917 bool LibraryCallKit::inline_math_multiplyHigh() {
1918 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1919 return true;
1920 }
1921
1922 Node*
1923 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1924 // These are the candidate return value:
1925 Node* xvalue = x0;
1926 Node* yvalue = y0;
1927
1928 if (xvalue == yvalue) {
1929 return xvalue;
1930 }
1931
1932 bool want_max = (id == vmIntrinsics::_max);
1933
1934 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1935 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1936 if (txvalue == NULL || tyvalue == NULL) return top();
1937 // This is not really necessary, but it is consistent with a
1938 // hypothetical MaxINode::Value method:
1939 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1940
1941 // %%% This folding logic should (ideally) be in a different place.
1942 // Some should be inside IfNode, and there to be a more reliable
1943 // transformation of ?: style patterns into cmoves. We also want
1944 // more powerful optimizations around cmove and min/max.
1945
1946 // Try to find a dominating comparison of these guys.
1947 // It can simplify the index computation for Arrays.copyOf
1948 // and similar uses of System.arraycopy.
1949 // First, compute the normalized version of CmpI(x, y).
1950 int cmp_op = Op_CmpI;
1951 Node* xkey = xvalue;
1952 Node* ykey = yvalue;
1953 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1954 if (ideal_cmpxy->is_Cmp()) {
1955 // E.g., if we have CmpI(length - offset, count),
1956 // it might idealize to CmpI(length, count + offset)
1957 cmp_op = ideal_cmpxy->Opcode();
1958 xkey = ideal_cmpxy->in(1);
1959 ykey = ideal_cmpxy->in(2);
1960 }
1961
1962 // Start by locating any relevant comparisons.
1963 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1964 Node* cmpxy = NULL;
1965 Node* cmpyx = NULL;
1966 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1967 Node* cmp = start_from->fast_out(k);
1968 if (cmp->outcnt() > 0 && // must have prior uses
1969 cmp->in(0) == NULL && // must be context-independent
1970 cmp->Opcode() == cmp_op) { // right kind of compare
1971 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1972 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1973 }
1974 }
1975
1976 const int NCMPS = 2;
1977 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1978 int cmpn;
1979 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1980 if (cmps[cmpn] != NULL) break; // find a result
1981 }
1982 if (cmpn < NCMPS) {
1983 // Look for a dominating test that tells us the min and max.
1984 int depth = 0; // Limit search depth for speed
1985 Node* dom = control();
1986 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1987 if (++depth >= 100) break;
1988 Node* ifproj = dom;
1989 if (!ifproj->is_Proj()) continue;
1990 Node* iff = ifproj->in(0);
1991 if (!iff->is_If()) continue;
1992 Node* bol = iff->in(1);
1993 if (!bol->is_Bool()) continue;
1994 Node* cmp = bol->in(1);
1995 if (cmp == NULL) continue;
1996 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1997 if (cmps[cmpn] == cmp) break;
1998 if (cmpn == NCMPS) continue;
1999 BoolTest::mask btest = bol->as_Bool()->_test._test;
2000 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2001 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2002 // At this point, we know that 'x btest y' is true.
2003 switch (btest) {
2004 case BoolTest::eq:
2005 // They are proven equal, so we can collapse the min/max.
2006 // Either value is the answer. Choose the simpler.
2007 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2008 return yvalue;
2009 return xvalue;
2010 case BoolTest::lt: // x < y
2011 case BoolTest::le: // x <= y
2012 return (want_max ? yvalue : xvalue);
2013 case BoolTest::gt: // x > y
2014 case BoolTest::ge: // x >= y
2015 return (want_max ? xvalue : yvalue);
2016 default:
2017 break;
2018 }
2019 }
2020 }
2021
2022 // We failed to find a dominating test.
2023 // Let's pick a test that might GVN with prior tests.
2024 Node* best_bol = NULL;
2025 BoolTest::mask best_btest = BoolTest::illegal;
2026 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2027 Node* cmp = cmps[cmpn];
2028 if (cmp == NULL) continue;
2029 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2030 Node* bol = cmp->fast_out(j);
2031 if (!bol->is_Bool()) continue;
2032 BoolTest::mask btest = bol->as_Bool()->_test._test;
2033 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2034 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2035 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2036 best_bol = bol->as_Bool();
2037 best_btest = btest;
2038 }
2039 }
2040 }
2041
2042 Node* answer_if_true = NULL;
2043 Node* answer_if_false = NULL;
2044 switch (best_btest) {
2045 default:
2046 if (cmpxy == NULL)
2047 cmpxy = ideal_cmpxy;
2048 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2049 // and fall through:
2050 case BoolTest::lt: // x < y
2051 case BoolTest::le: // x <= y
2052 answer_if_true = (want_max ? yvalue : xvalue);
2053 answer_if_false = (want_max ? xvalue : yvalue);
2054 break;
2055 case BoolTest::gt: // x > y
2056 case BoolTest::ge: // x >= y
2057 answer_if_true = (want_max ? xvalue : yvalue);
2058 answer_if_false = (want_max ? yvalue : xvalue);
2059 break;
2060 }
2061
2062 jint hi, lo;
2063 if (want_max) {
2064 // We can sharpen the minimum.
2065 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2066 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2067 } else {
2068 // We can sharpen the maximum.
2069 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2070 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2071 }
2072
2073 // Use a flow-free graph structure, to avoid creating excess control edges
2074 // which could hinder other optimizations.
2075 // Since Math.min/max is often used with arraycopy, we want
2076 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2077 Node* cmov = CMoveNode::make(NULL, best_bol,
2078 answer_if_false, answer_if_true,
2079 TypeInt::make(lo, hi, widen));
2080
2081 return _gvn.transform(cmov);
2082
2083 /*
2084 // This is not as desirable as it may seem, since Min and Max
2085 // nodes do not have a full set of optimizations.
2086 // And they would interfere, anyway, with 'if' optimizations
2087 // and with CMoveI canonical forms.
2088 switch (id) {
2089 case vmIntrinsics::_min:
2090 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2091 case vmIntrinsics::_max:
2092 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2093 default:
2094 ShouldNotReachHere();
2095 }
2096 */
2097 }
2098
2099 inline int
2100 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2101 const TypePtr* base_type = TypePtr::NULL_PTR;
2102 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2103 if (base_type == NULL) {
2104 // Unknown type.
2105 return Type::AnyPtr;
2106 } else if (base_type == TypePtr::NULL_PTR) {
2107 // Since this is a NULL+long form, we have to switch to a rawptr.
2108 base = _gvn.transform(new CastX2PNode(offset));
2109 offset = MakeConX(0);
2110 return Type::RawPtr;
2111 } else if (base_type->base() == Type::RawPtr) {
2112 return Type::RawPtr;
2113 } else if (base_type->isa_oopptr()) {
2114 // Base is never null => always a heap address.
2115 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2116 return Type::OopPtr;
2117 }
2118 // Offset is small => always a heap address.
2119 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2120 if (offset_type != NULL &&
2121 base_type->offset() == 0 && // (should always be?)
2122 offset_type->_lo >= 0 &&
2123 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2124 return Type::OopPtr;
2125 } else if (type == T_OBJECT) {
2126 // off heap access to an oop doesn't make any sense. Has to be on
2127 // heap.
2128 return Type::OopPtr;
2129 }
2130 // Otherwise, it might either be oop+off or NULL+addr.
2131 return Type::AnyPtr;
2132 } else {
2133 // No information:
2134 return Type::AnyPtr;
2135 }
2136 }
2137
2138 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2139 Node* uncasted_base = base;
2140 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2141 if (kind == Type::RawPtr) {
2142 return basic_plus_adr(top(), uncasted_base, offset);
2143 } else if (kind == Type::AnyPtr) {
2144 assert(base == uncasted_base, "unexpected base change");
2145 if (can_cast) {
2146 if (!_gvn.type(base)->speculative_maybe_null() &&
2147 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2148 // According to profiling, this access is always on
2149 // heap. Casting the base to not null and thus avoiding membars
2150 // around the access should allow better optimizations
2151 Node* null_ctl = top();
2152 base = null_check_oop(base, &null_ctl, true, true, true);
2153 assert(null_ctl->is_top(), "no null control here");
2154 return basic_plus_adr(base, offset);
2155 } else if (_gvn.type(base)->speculative_always_null() &&
2156 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2157 // According to profiling, this access is always off
2158 // heap.
2159 base = null_assert(base);
2160 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2161 offset = MakeConX(0);
2162 return basic_plus_adr(top(), raw_base, offset);
2163 }
2164 }
2165 // We don't know if it's an on heap or off heap access. Fall back
2166 // to raw memory access.
2167 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2168 return basic_plus_adr(top(), raw, offset);
2169 } else {
2170 assert(base == uncasted_base, "unexpected base change");
2171 // We know it's an on heap access so base can't be null
2172 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2173 base = must_be_not_null(base, true);
2174 }
2175 return basic_plus_adr(base, offset);
2176 }
2177 }
2178
2179 //--------------------------inline_number_methods-----------------------------
2180 // inline int Integer.numberOfLeadingZeros(int)
2181 // inline int Long.numberOfLeadingZeros(long)
2182 //
2183 // inline int Integer.numberOfTrailingZeros(int)
2184 // inline int Long.numberOfTrailingZeros(long)
2185 //
2186 // inline int Integer.bitCount(int)
2187 // inline int Long.bitCount(long)
2188 //
2189 // inline char Character.reverseBytes(char)
2190 // inline short Short.reverseBytes(short)
2191 // inline int Integer.reverseBytes(int)
2192 // inline long Long.reverseBytes(long)
2193 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2194 Node* arg = argument(0);
2195 Node* n = NULL;
2196 switch (id) {
2197 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2198 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2199 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2200 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2201 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2202 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2203 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2204 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2205 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2206 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2207 default: fatal_unexpected_iid(id); break;
2208 }
2209 set_result(_gvn.transform(n));
2210 return true;
2211 }
2212
2213 //----------------------------inline_unsafe_access----------------------------
2214
2215 // Helper that guards and inserts a pre-barrier.
2216 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2217 Node* pre_val, bool need_mem_bar) {
2218 // We could be accessing the referent field of a reference object. If so, when G1
2219 // is enabled, we need to log the value in the referent field in an SATB buffer.
2220 // This routine performs some compile time filters and generates suitable
2221 // runtime filters that guard the pre-barrier code.
2222 // Also add memory barrier for non volatile load from the referent field
2223 // to prevent commoning of loads across safepoint.
2224 if (!UseG1GC && !need_mem_bar)
2225 return;
2226
2227 // Some compile time checks.
2228
2229 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2230 const TypeX* otype = offset->find_intptr_t_type();
2231 if (otype != NULL && otype->is_con() &&
2232 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2233 // Constant offset but not the reference_offset so just return
2234 return;
2235 }
2236
2237 // We only need to generate the runtime guards for instances.
2238 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2239 if (btype != NULL) {
2240 if (btype->isa_aryptr()) {
2241 // Array type so nothing to do
2242 return;
2243 }
2244
2245 const TypeInstPtr* itype = btype->isa_instptr();
2246 if (itype != NULL) {
2247 // Can the klass of base_oop be statically determined to be
2248 // _not_ a sub-class of Reference and _not_ Object?
2249 ciKlass* klass = itype->klass();
2250 if ( klass->is_loaded() &&
2251 !klass->is_subtype_of(env()->Reference_klass()) &&
2252 !env()->Object_klass()->is_subtype_of(klass)) {
2253 return;
2254 }
2255 }
2256 }
2257
2258 // The compile time filters did not reject base_oop/offset so
2259 // we need to generate the following runtime filters
2260 //
2261 // if (offset == java_lang_ref_Reference::_reference_offset) {
2262 // if (instance_of(base, java.lang.ref.Reference)) {
2263 // pre_barrier(_, pre_val, ...);
2264 // }
2265 // }
2266
2267 float likely = PROB_LIKELY( 0.999);
2268 float unlikely = PROB_UNLIKELY(0.999);
2269
2270 IdealKit ideal(this);
2271 #define __ ideal.
2272
2273 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2274
2275 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2276 // Update graphKit memory and control from IdealKit.
2277 sync_kit(ideal);
2278
2279 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2280 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2281
2282 // Update IdealKit memory and control from graphKit.
2283 __ sync_kit(this);
2284
2285 Node* one = __ ConI(1);
2286 // is_instof == 0 if base_oop == NULL
2287 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2288
2289 // Update graphKit from IdeakKit.
2290 sync_kit(ideal);
2291
2292 // Use the pre-barrier to record the value in the referent field
2293 pre_barrier(false /* do_load */,
2294 __ ctrl(),
2295 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2296 pre_val /* pre_val */,
2297 T_OBJECT);
2298 if (need_mem_bar) {
2299 // Add memory barrier to prevent commoning reads from this field
2300 // across safepoint since GC can change its value.
2301 insert_mem_bar(Op_MemBarCPUOrder);
2302 }
2303 // Update IdealKit from graphKit.
2304 __ sync_kit(this);
2305
2306 } __ end_if(); // _ref_type != ref_none
2307 } __ end_if(); // offset == referent_offset
2308
2309 // Final sync IdealKit and GraphKit.
2310 final_sync(ideal);
2311 #undef __
2312 }
2313
2314
2315 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2316 // Attempt to infer a sharper value type from the offset and base type.
2317 ciKlass* sharpened_klass = NULL;
2318
2319 // See if it is an instance field, with an object type.
2320 if (alias_type->field() != NULL) {
2321 if (alias_type->field()->type()->is_klass()) {
2322 sharpened_klass = alias_type->field()->type()->as_klass();
2323 }
2324 }
2325
2326 // See if it is a narrow oop array.
2327 if (adr_type->isa_aryptr()) {
2328 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2329 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2330 if (elem_type != NULL) {
2331 sharpened_klass = elem_type->klass();
2332 }
2333 }
2334 }
2335
2336 // The sharpened class might be unloaded if there is no class loader
2337 // contraint in place.
2338 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2339 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2340
2341 #ifndef PRODUCT
2342 if (C->print_intrinsics() || C->print_inlining()) {
2343 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2344 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2345 }
2346 #endif
2347 // Sharpen the value type.
2348 return tjp;
2349 }
2350 return NULL;
2351 }
2352
2353 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2354 if (callee()->is_static()) return false; // caller must have the capability!
2355 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2356 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2357 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2358
2359 #ifndef PRODUCT
2360 {
2361 ResourceMark rm;
2362 // Check the signatures.
2363 ciSignature* sig = callee()->signature();
2364 #ifdef ASSERT
2365 if (!is_store) {
2366 // Object getObject(Object base, int/long offset), etc.
2367 BasicType rtype = sig->return_type()->basic_type();
2368 assert(rtype == type, "getter must return the expected value");
2369 assert(sig->count() == 2, "oop getter has 2 arguments");
2370 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2371 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2372 } else {
2373 // void putObject(Object base, int/long offset, Object x), etc.
2374 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2375 assert(sig->count() == 3, "oop putter has 3 arguments");
2376 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2377 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2378 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2379 assert(vtype == type, "putter must accept the expected value");
2380 }
2381 #endif // ASSERT
2382 }
2383 #endif //PRODUCT
2384
2385 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2386
2387 Node* receiver = argument(0); // type: oop
2388
2389 // Build address expression.
2390 Node* adr;
2391 Node* heap_base_oop = top();
2392 Node* offset = top();
2393 Node* val;
2394
2395 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2396 Node* base = argument(1); // type: oop
2397 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2398 offset = argument(2); // type: long
2399 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2400 // to be plain byte offsets, which are also the same as those accepted
2401 // by oopDesc::field_base.
2402 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2403 "fieldOffset must be byte-scaled");
2404 // 32-bit machines ignore the high half!
2405 offset = ConvL2X(offset);
2406 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2407
2408 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2409 heap_base_oop = base;
2410 } else if (type == T_OBJECT) {
2411 return false; // off-heap oop accesses are not supported
2412 }
2413
2414 // Can base be NULL? Otherwise, always on-heap access.
2415 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2416
2417 val = is_store ? argument(4) : NULL;
2418
2419 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2420
2421 // Try to categorize the address.
2422 Compile::AliasType* alias_type = C->alias_type(adr_type);
2423 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2424
2425 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2426 alias_type->adr_type() == TypeAryPtr::RANGE) {
2427 return false; // not supported
2428 }
2429
2430 bool mismatched = false;
2431 BasicType bt = alias_type->basic_type();
2432 if (bt != T_ILLEGAL) {
2433 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2434 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2435 // Alias type doesn't differentiate between byte[] and boolean[]).
2436 // Use address type to get the element type.
2437 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2438 }
2439 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2440 // accessing an array field with getObject is not a mismatch
2441 bt = T_OBJECT;
2442 }
2443 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2444 // Don't intrinsify mismatched object accesses
2445 return false;
2446 }
2447 mismatched = (bt != type);
2448 } else if (alias_type->adr_type()->isa_oopptr()) {
2449 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2450 }
2451
2452 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2453
2454 // First guess at the value type.
2455 const Type *value_type = Type::get_const_basic_type(type);
2456
2457 // We will need memory barriers unless we can determine a unique
2458 // alias category for this reference. (Note: If for some reason
2459 // the barriers get omitted and the unsafe reference begins to "pollute"
2460 // the alias analysis of the rest of the graph, either Compile::can_alias
2461 // or Compile::must_alias will throw a diagnostic assert.)
2462 bool need_mem_bar = false;
2463 switch (kind) {
2464 case Relaxed:
2465 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2466 break;
2467 case Opaque:
2468 // Opaque uses CPUOrder membars for protection against code movement.
2469 case Acquire:
2470 case Release:
2471 case Volatile:
2472 need_mem_bar = true;
2473 break;
2474 default:
2475 ShouldNotReachHere();
2476 }
2477
2478 // Some accesses require access atomicity for all types, notably longs and doubles.
2479 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2480 bool requires_atomic_access = false;
2481 switch (kind) {
2482 case Relaxed:
2483 requires_atomic_access = AlwaysAtomicAccesses;
2484 break;
2485 case Opaque:
2486 // Opaque accesses are atomic.
2487 case Acquire:
2488 case Release:
2489 case Volatile:
2490 requires_atomic_access = true;
2491 break;
2492 default:
2493 ShouldNotReachHere();
2494 }
2495
2496 // Figure out the memory ordering.
2497 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2498 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2499
2500 // If we are reading the value of the referent field of a Reference
2501 // object (either by using Unsafe directly or through reflection)
2502 // then, if G1 is enabled, we need to record the referent in an
2503 // SATB log buffer using the pre-barrier mechanism.
2504 // Also we need to add memory barrier to prevent commoning reads
2505 // from this field across safepoint since GC can change its value.
2506 bool need_read_barrier = !is_store &&
2507 offset != top() && heap_base_oop != top();
2508
2509 if (!is_store && type == T_OBJECT) {
2510 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2511 if (tjp != NULL) {
2512 value_type = tjp;
2513 }
2514 }
2515
2516 receiver = null_check(receiver);
2517 if (stopped()) {
2518 return true;
2519 }
2520 // Heap pointers get a null-check from the interpreter,
2521 // as a courtesy. However, this is not guaranteed by Unsafe,
2522 // and it is not possible to fully distinguish unintended nulls
2523 // from intended ones in this API.
2524
2525 // We need to emit leading and trailing CPU membars (see below) in
2526 // addition to memory membars for special access modes. This is a little
2527 // too strong, but avoids the need to insert per-alias-type
2528 // volatile membars (for stores; compare Parse::do_put_xxx), which
2529 // we cannot do effectively here because we probably only have a
2530 // rough approximation of type.
2531
2532 switch(kind) {
2533 case Relaxed:
2534 case Opaque:
2535 case Acquire:
2536 break;
2537 case Release:
2538 case Volatile:
2539 if (is_store) {
2540 insert_mem_bar(Op_MemBarRelease);
2541 } else {
2542 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2543 insert_mem_bar(Op_MemBarVolatile);
2544 }
2545 }
2546 break;
2547 default:
2548 ShouldNotReachHere();
2549 }
2550
2551 // Memory barrier to prevent normal and 'unsafe' accesses from
2552 // bypassing each other. Happens after null checks, so the
2553 // exception paths do not take memory state from the memory barrier,
2554 // so there's no problems making a strong assert about mixing users
2555 // of safe & unsafe memory.
2556 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2557
2558 if (!is_store) {
2559 Node* p = NULL;
2560 // Try to constant fold a load from a constant field
2561 ciField* field = alias_type->field();
2562 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2563 // final or stable field
2564 p = make_constant_from_field(field, heap_base_oop);
2565 }
2566 if (p == NULL) {
2567 // To be valid, unsafe loads may depend on other conditions than
2568 // the one that guards them: pin the Load node
2569 LoadNode::ControlDependency dep = LoadNode::Pinned;
2570 Node* ctrl = control();
2571 if (adr_type->isa_instptr()) {
2572 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2573 intptr_t offset = Type::OffsetBot;
2574 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2575 if (offset >= 0) {
2576 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2577 if (offset < s) {
2578 // Guaranteed to be a valid access, no need to pin it
2579 dep = LoadNode::DependsOnlyOnTest;
2580 ctrl = NULL;
2581 }
2582 }
2583 }
2584 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2585 // load value
2586 switch (type) {
2587 case T_BOOLEAN:
2588 {
2589 // Normalize the value returned by getBoolean in the following cases
2590 if (mismatched ||
2591 heap_base_oop == top() || // - heap_base_oop is NULL or
2592 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2593 // and the unsafe access is made to large offset
2594 // (i.e., larger than the maximum offset necessary for any
2595 // field access)
2596 ) {
2597 IdealKit ideal = IdealKit(this);
2598 #define __ ideal.
2599 IdealVariable normalized_result(ideal);
2600 __ declarations_done();
2601 __ set(normalized_result, p);
2602 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2603 __ set(normalized_result, ideal.ConI(1));
2604 ideal.end_if();
2605 final_sync(ideal);
2606 p = __ value(normalized_result);
2607 #undef __
2608 }
2609 }
2610 case T_CHAR:
2611 case T_BYTE:
2612 case T_SHORT:
2613 case T_INT:
2614 case T_LONG:
2615 case T_FLOAT:
2616 case T_DOUBLE:
2617 break;
2618 case T_OBJECT:
2619 if (need_read_barrier) {
2620 // We do not require a mem bar inside pre_barrier if need_mem_bar
2621 // is set: the barriers would be emitted by us.
2622 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2623 }
2624 break;
2625 case T_ADDRESS:
2626 // Cast to an int type.
2627 p = _gvn.transform(new CastP2XNode(NULL, p));
2628 p = ConvX2UL(p);
2629 break;
2630 default:
2631 fatal("unexpected type %d: %s", type, type2name(type));
2632 break;
2633 }
2634 }
2635 // The load node has the control of the preceding MemBarCPUOrder. All
2636 // following nodes will have the control of the MemBarCPUOrder inserted at
2637 // the end of this method. So, pushing the load onto the stack at a later
2638 // point is fine.
2639 set_result(p);
2640 } else {
2641 // place effect of store into memory
2642 switch (type) {
2643 case T_DOUBLE:
2644 val = dstore_rounding(val);
2645 break;
2646 case T_ADDRESS:
2647 // Repackage the long as a pointer.
2648 val = ConvL2X(val);
2649 val = _gvn.transform(new CastX2PNode(val));
2650 break;
2651 default:
2652 break;
2653 }
2654
2655 if (type == T_OBJECT) {
2656 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2657 } else {
2658 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2659 }
2660 }
2661
2662 switch(kind) {
2663 case Relaxed:
2664 case Opaque:
2665 case Release:
2666 break;
2667 case Acquire:
2668 case Volatile:
2669 if (!is_store) {
2670 insert_mem_bar(Op_MemBarAcquire);
2671 } else {
2672 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2673 insert_mem_bar(Op_MemBarVolatile);
2674 }
2675 }
2676 break;
2677 default:
2678 ShouldNotReachHere();
2679 }
2680
2681 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2682
2683 return true;
2684 }
2685
2686 //----------------------------inline_unsafe_load_store----------------------------
2687 // This method serves a couple of different customers (depending on LoadStoreKind):
2688 //
2689 // LS_cmp_swap:
2690 //
2691 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2692 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2693 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2694 //
2695 // LS_cmp_swap_weak:
2696 //
2697 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2698 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2699 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2700 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2701 //
2702 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2703 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2704 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2705 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2706 //
2707 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2708 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2709 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2710 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2711 //
2712 // LS_cmp_exchange:
2713 //
2714 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2715 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2716 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2717 //
2718 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2719 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2720 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2721 //
2722 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2723 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2724 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2725 //
2726 // LS_get_add:
2727 //
2728 // int getAndAddInt( Object o, long offset, int delta)
2729 // long getAndAddLong(Object o, long offset, long delta)
2730 //
2731 // LS_get_set:
2732 //
2733 // int getAndSet(Object o, long offset, int newValue)
2734 // long getAndSet(Object o, long offset, long newValue)
2735 // Object getAndSet(Object o, long offset, Object newValue)
2736 //
2737 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2738 // This basic scheme here is the same as inline_unsafe_access, but
2739 // differs in enough details that combining them would make the code
2740 // overly confusing. (This is a true fact! I originally combined
2741 // them, but even I was confused by it!) As much code/comments as
2742 // possible are retained from inline_unsafe_access though to make
2743 // the correspondences clearer. - dl
2744
2745 if (callee()->is_static()) return false; // caller must have the capability!
2746
2747 #ifndef PRODUCT
2748 BasicType rtype;
2749 {
2750 ResourceMark rm;
2751 // Check the signatures.
2752 ciSignature* sig = callee()->signature();
2753 rtype = sig->return_type()->basic_type();
2754 switch(kind) {
2755 case LS_get_add:
2756 case LS_get_set: {
2757 // Check the signatures.
2758 #ifdef ASSERT
2759 assert(rtype == type, "get and set must return the expected type");
2760 assert(sig->count() == 3, "get and set has 3 arguments");
2761 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2762 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2763 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2764 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2765 #endif // ASSERT
2766 break;
2767 }
2768 case LS_cmp_swap:
2769 case LS_cmp_swap_weak: {
2770 // Check the signatures.
2771 #ifdef ASSERT
2772 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2773 assert(sig->count() == 4, "CAS has 4 arguments");
2774 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2775 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2776 #endif // ASSERT
2777 break;
2778 }
2779 case LS_cmp_exchange: {
2780 // Check the signatures.
2781 #ifdef ASSERT
2782 assert(rtype == type, "CAS must return the expected type");
2783 assert(sig->count() == 4, "CAS has 4 arguments");
2784 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2785 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2786 #endif // ASSERT
2787 break;
2788 }
2789 default:
2790 ShouldNotReachHere();
2791 }
2792 }
2793 #endif //PRODUCT
2794
2795 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2796
2797 // Get arguments:
2798 Node* receiver = NULL;
2799 Node* base = NULL;
2800 Node* offset = NULL;
2801 Node* oldval = NULL;
2802 Node* newval = NULL;
2803 switch(kind) {
2804 case LS_cmp_swap:
2805 case LS_cmp_swap_weak:
2806 case LS_cmp_exchange: {
2807 const bool two_slot_type = type2size[type] == 2;
2808 receiver = argument(0); // type: oop
2809 base = argument(1); // type: oop
2810 offset = argument(2); // type: long
2811 oldval = argument(4); // type: oop, int, or long
2812 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2813 break;
2814 }
2815 case LS_get_add:
2816 case LS_get_set: {
2817 receiver = argument(0); // type: oop
2818 base = argument(1); // type: oop
2819 offset = argument(2); // type: long
2820 oldval = NULL;
2821 newval = argument(4); // type: oop, int, or long
2822 break;
2823 }
2824 default:
2825 ShouldNotReachHere();
2826 }
2827
2828 // Build field offset expression.
2829 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2830 // to be plain byte offsets, which are also the same as those accepted
2831 // by oopDesc::field_base.
2832 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2833 // 32-bit machines ignore the high half of long offsets
2834 offset = ConvL2X(offset);
2835 Node* adr = make_unsafe_address(base, offset, type, false);
2836 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2837
2838 Compile::AliasType* alias_type = C->alias_type(adr_type);
2839 BasicType bt = alias_type->basic_type();
2840 if (bt != T_ILLEGAL &&
2841 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2842 // Don't intrinsify mismatched object accesses.
2843 return false;
2844 }
2845
2846 // For CAS, unlike inline_unsafe_access, there seems no point in
2847 // trying to refine types. Just use the coarse types here.
2848 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2849 const Type *value_type = Type::get_const_basic_type(type);
2850
2851 switch (kind) {
2852 case LS_get_set:
2853 case LS_cmp_exchange: {
2854 if (type == T_OBJECT) {
2855 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2856 if (tjp != NULL) {
2857 value_type = tjp;
2858 }
2859 }
2860 break;
2861 }
2862 case LS_cmp_swap:
2863 case LS_cmp_swap_weak:
2864 case LS_get_add:
2865 break;
2866 default:
2867 ShouldNotReachHere();
2868 }
2869
2870 // Null check receiver.
2871 receiver = null_check(receiver);
2872 if (stopped()) {
2873 return true;
2874 }
2875
2876 int alias_idx = C->get_alias_index(adr_type);
2877
2878 // Memory-model-wise, a LoadStore acts like a little synchronized
2879 // block, so needs barriers on each side. These don't translate
2880 // into actual barriers on most machines, but we still need rest of
2881 // compiler to respect ordering.
2882
2883 switch (access_kind) {
2884 case Relaxed:
2885 case Acquire:
2886 break;
2887 case Release:
2888 insert_mem_bar(Op_MemBarRelease);
2889 break;
2890 case Volatile:
2891 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2892 insert_mem_bar(Op_MemBarVolatile);
2893 } else {
2894 insert_mem_bar(Op_MemBarRelease);
2895 }
2896 break;
2897 default:
2898 ShouldNotReachHere();
2899 }
2900 insert_mem_bar(Op_MemBarCPUOrder);
2901
2902 // Figure out the memory ordering.
2903 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2904
2905 // 4984716: MemBars must be inserted before this
2906 // memory node in order to avoid a false
2907 // dependency which will confuse the scheduler.
2908 Node *mem = memory(alias_idx);
2909
2910 // For now, we handle only those cases that actually exist: ints,
2911 // longs, and Object. Adding others should be straightforward.
2912 Node* load_store = NULL;
2913 switch(type) {
2914 case T_BYTE:
2915 switch(kind) {
2916 case LS_get_add:
2917 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2918 break;
2919 case LS_get_set:
2920 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2921 break;
2922 case LS_cmp_swap_weak:
2923 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2924 break;
2925 case LS_cmp_swap:
2926 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2927 break;
2928 case LS_cmp_exchange:
2929 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2930 break;
2931 default:
2932 ShouldNotReachHere();
2933 }
2934 break;
2935 case T_SHORT:
2936 switch(kind) {
2937 case LS_get_add:
2938 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2939 break;
2940 case LS_get_set:
2941 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2942 break;
2943 case LS_cmp_swap_weak:
2944 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2945 break;
2946 case LS_cmp_swap:
2947 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2948 break;
2949 case LS_cmp_exchange:
2950 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2951 break;
2952 default:
2953 ShouldNotReachHere();
2954 }
2955 break;
2956 case T_INT:
2957 switch(kind) {
2958 case LS_get_add:
2959 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2960 break;
2961 case LS_get_set:
2962 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2963 break;
2964 case LS_cmp_swap_weak:
2965 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2966 break;
2967 case LS_cmp_swap:
2968 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2969 break;
2970 case LS_cmp_exchange:
2971 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2972 break;
2973 default:
2974 ShouldNotReachHere();
2975 }
2976 break;
2977 case T_LONG:
2978 switch(kind) {
2979 case LS_get_add:
2980 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2981 break;
2982 case LS_get_set:
2983 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2984 break;
2985 case LS_cmp_swap_weak:
2986 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2987 break;
2988 case LS_cmp_swap:
2989 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2990 break;
2991 case LS_cmp_exchange:
2992 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2993 break;
2994 default:
2995 ShouldNotReachHere();
2996 }
2997 break;
2998 case T_OBJECT:
2999 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3000 // could be delayed during Parse (for example, in adjust_map_after_if()).
3001 // Execute transformation here to avoid barrier generation in such case.
3002 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3003 newval = _gvn.makecon(TypePtr::NULL_PTR);
3004
3005 // Reference stores need a store barrier.
3006 switch(kind) {
3007 case LS_get_set: {
3008 // If pre-barrier must execute before the oop store, old value will require do_load here.
3009 if (!can_move_pre_barrier()) {
3010 pre_barrier(true /* do_load*/,
3011 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3012 NULL /* pre_val*/,
3013 T_OBJECT);
3014 } // Else move pre_barrier to use load_store value, see below.
3015 break;
3016 }
3017 case LS_cmp_swap_weak:
3018 case LS_cmp_swap:
3019 case LS_cmp_exchange: {
3020 // Same as for newval above:
3021 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3022 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3023 }
3024 // The only known value which might get overwritten is oldval.
3025 pre_barrier(false /* do_load */,
3026 control(), NULL, NULL, max_juint, NULL, NULL,
3027 oldval /* pre_val */,
3028 T_OBJECT);
3029 break;
3030 }
3031 default:
3032 ShouldNotReachHere();
3033 }
3034
3035 #ifdef _LP64
3036 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3037 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3038
3039 switch(kind) {
3040 case LS_get_set:
3041 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3042 break;
3043 case LS_cmp_swap_weak: {
3044 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3045 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3046 break;
3047 }
3048 case LS_cmp_swap: {
3049 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3050 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3051 break;
3052 }
3053 case LS_cmp_exchange: {
3054 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3055 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3056 break;
3057 }
3058 default:
3059 ShouldNotReachHere();
3060 }
3061 } else
3062 #endif
3063 switch (kind) {
3064 case LS_get_set:
3065 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3066 break;
3067 case LS_cmp_swap_weak:
3068 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3069 break;
3070 case LS_cmp_swap:
3071 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3072 break;
3073 case LS_cmp_exchange:
3074 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3075 break;
3076 default:
3077 ShouldNotReachHere();
3078 }
3079
3080 // Emit the post barrier only when the actual store happened. This makes sense
3081 // to check only for LS_cmp_* that can fail to set the value.
3082 // LS_cmp_exchange does not produce any branches by default, so there is no
3083 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3084 // CompareAndExchange and move branches here, it would make sense to conditionalize
3085 // post_barriers for LS_cmp_exchange as well.
3086 //
3087 // CAS success path is marked more likely since we anticipate this is a performance
3088 // critical path, while CAS failure path can use the penalty for going through unlikely
3089 // path as backoff. Which is still better than doing a store barrier there.
3090 switch (kind) {
3091 case LS_get_set:
3092 case LS_cmp_exchange: {
3093 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3094 break;
3095 }
3096 case LS_cmp_swap_weak:
3097 case LS_cmp_swap: {
3098 IdealKit ideal(this);
3099 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3100 sync_kit(ideal);
3101 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3102 ideal.sync_kit(this);
3103 } ideal.end_if();
3104 final_sync(ideal);
3105 break;
3106 }
3107 default:
3108 ShouldNotReachHere();
3109 }
3110 break;
3111 default:
3112 fatal("unexpected type %d: %s", type, type2name(type));
3113 break;
3114 }
3115
3116 // SCMemProjNodes represent the memory state of a LoadStore. Their
3117 // main role is to prevent LoadStore nodes from being optimized away
3118 // when their results aren't used.
3119 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3120 set_memory(proj, alias_idx);
3121
3122 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3123 #ifdef _LP64
3124 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3125 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3126 }
3127 #endif
3128 if (can_move_pre_barrier() && kind == LS_get_set) {
3129 // Don't need to load pre_val. The old value is returned by load_store.
3130 // The pre_barrier can execute after the xchg as long as no safepoint
3131 // gets inserted between them.
3132 pre_barrier(false /* do_load */,
3133 control(), NULL, NULL, max_juint, NULL, NULL,
3134 load_store /* pre_val */,
3135 T_OBJECT);
3136 }
3137 }
3138
3139 // Add the trailing membar surrounding the access
3140 insert_mem_bar(Op_MemBarCPUOrder);
3141
3142 switch (access_kind) {
3143 case Relaxed:
3144 case Release:
3145 break; // do nothing
3146 case Acquire:
3147 case Volatile:
3148 insert_mem_bar(Op_MemBarAcquire);
3149 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3150 break;
3151 default:
3152 ShouldNotReachHere();
3153 }
3154
3155 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3156 set_result(load_store);
3157 return true;
3158 }
3159
3160 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3161 MemNode::MemOrd mo = MemNode::unset;
3162 switch(kind) {
3163 case Opaque:
3164 case Relaxed: mo = MemNode::unordered; break;
3165 case Acquire: mo = MemNode::acquire; break;
3166 case Release: mo = MemNode::release; break;
3167 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3168 default:
3169 ShouldNotReachHere();
3170 }
3171 guarantee(mo != MemNode::unset, "Should select memory ordering");
3172 return mo;
3173 }
3174
3175 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3176 MemNode::MemOrd mo = MemNode::unset;
3177 switch(kind) {
3178 case Opaque:
3179 case Relaxed: mo = MemNode::unordered; break;
3180 case Acquire: mo = MemNode::acquire; break;
3181 case Release: mo = MemNode::release; break;
3182 case Volatile: mo = MemNode::seqcst; break;
3183 default:
3184 ShouldNotReachHere();
3185 }
3186 guarantee(mo != MemNode::unset, "Should select memory ordering");
3187 return mo;
3188 }
3189
3190 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3191 // Regardless of form, don't allow previous ld/st to move down,
3192 // then issue acquire, release, or volatile mem_bar.
3193 insert_mem_bar(Op_MemBarCPUOrder);
3194 switch(id) {
3195 case vmIntrinsics::_loadFence:
3196 insert_mem_bar(Op_LoadFence);
3197 return true;
3198 case vmIntrinsics::_storeFence:
3199 insert_mem_bar(Op_StoreFence);
3200 return true;
3201 case vmIntrinsics::_fullFence:
3202 insert_mem_bar(Op_MemBarVolatile);
3203 return true;
3204 default:
3205 fatal_unexpected_iid(id);
3206 return false;
3207 }
3208 }
3209
3210 bool LibraryCallKit::inline_onspinwait() {
3211 insert_mem_bar(Op_OnSpinWait);
3212 return true;
3213 }
3214
3215 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3216 if (!kls->is_Con()) {
3217 return true;
3218 }
3219 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3220 if (klsptr == NULL) {
3221 return true;
3222 }
3223 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3224 // don't need a guard for a klass that is already initialized
3225 return !ik->is_initialized();
3226 }
3227
3228 //----------------------------inline_unsafe_allocate---------------------------
3229 // public native Object Unsafe.allocateInstance(Class<?> cls);
3230 bool LibraryCallKit::inline_unsafe_allocate() {
3231 if (callee()->is_static()) return false; // caller must have the capability!
3232
3233 null_check_receiver(); // null-check, then ignore
3234 Node* cls = null_check(argument(1));
3235 if (stopped()) return true;
3236
3237 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3238 kls = null_check(kls);
3239 if (stopped()) return true; // argument was like int.class
3240
3241 Node* test = NULL;
3242 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3243 // Note: The argument might still be an illegal value like
3244 // Serializable.class or Object[].class. The runtime will handle it.
3245 // But we must make an explicit check for initialization.
3246 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3247 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3248 // can generate code to load it as unsigned byte.
3249 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3250 Node* bits = intcon(InstanceKlass::fully_initialized);
3251 test = _gvn.transform(new SubINode(inst, bits));
3252 // The 'test' is non-zero if we need to take a slow path.
3253 }
3254
3255 Node* obj = new_instance(kls, test);
3256 set_result(obj);
3257 return true;
3258 }
3259
3260 //------------------------inline_native_time_funcs--------------
3261 // inline code for System.currentTimeMillis() and System.nanoTime()
3262 // these have the same type and signature
3263 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3264 const TypeFunc* tf = OptoRuntime::void_long_Type();
3265 const TypePtr* no_memory_effects = NULL;
3266 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3267 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3268 #ifdef ASSERT
3269 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3270 assert(value_top == top(), "second value must be top");
3271 #endif
3272 set_result(value);
3273 return true;
3274 }
3275
3276 #ifdef TRACE_HAVE_INTRINSICS
3277
3278 /*
3279 * oop -> myklass
3280 * myklass->trace_id |= USED
3281 * return myklass->trace_id & ~0x3
3282 */
3283 bool LibraryCallKit::inline_native_classID() {
3284 Node* cls = null_check(argument(0), T_OBJECT);
3285 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3286 kls = null_check(kls, T_OBJECT);
3287
3288 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3289 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3290 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3291
3292 Node* clsused = longcon(0x01l); // set the class bit
3293 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3294 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3295 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3296
3297 #ifdef TRACE_ID_META_BITS
3298 Node* mbits = longcon(~TRACE_ID_META_BITS);
3299 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3300 #endif
3301 #ifdef TRACE_ID_CLASS_SHIFT
3302 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3303 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3304 #endif
3305
3306 set_result(tvalue);
3307 return true;
3308
3309 }
3310
3311 bool LibraryCallKit::inline_native_getBufferWriter() {
3312 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3313
3314 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3315 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3316 );
3317
3318 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3319
3320 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3321 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3322
3323 IfNode* iff_jobj_null =
3324 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3325
3326 enum { _normal_path = 1,
3327 _null_path = 2,
3328 PATH_LIMIT };
3329
3330 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3331 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3332
3333 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3334 result_rgn->init_req(_null_path, jobj_is_null);
3335 result_val->init_req(_null_path, null());
3336
3337 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3338 result_rgn->init_req(_normal_path, jobj_is_not_null);
3339
3340 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3341 result_val->init_req(_normal_path, res);
3342
3343 set_result(result_rgn, result_val);
3344
3345 return true;
3346 }
3347
3348 #endif
3349
3350 //------------------------inline_native_currentThread------------------
3351 bool LibraryCallKit::inline_native_currentThread() {
3352 Node* junk = NULL;
3353 set_result(generate_current_thread(junk));
3354 return true;
3355 }
3356
3357 //------------------------inline_native_isInterrupted------------------
3358 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3359 bool LibraryCallKit::inline_native_isInterrupted() {
3360 // Add a fast path to t.isInterrupted(clear_int):
3361 // (t == Thread.current() &&
3362 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3363 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3364 // So, in the common case that the interrupt bit is false,
3365 // we avoid making a call into the VM. Even if the interrupt bit
3366 // is true, if the clear_int argument is false, we avoid the VM call.
3367 // However, if the receiver is not currentThread, we must call the VM,
3368 // because there must be some locking done around the operation.
3369
3370 // We only go to the fast case code if we pass two guards.
3371 // Paths which do not pass are accumulated in the slow_region.
3372
3373 enum {
3374 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3375 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3376 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3377 PATH_LIMIT
3378 };
3379
3380 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3381 // out of the function.
3382 insert_mem_bar(Op_MemBarCPUOrder);
3383
3384 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3385 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3386
3387 RegionNode* slow_region = new RegionNode(1);
3388 record_for_igvn(slow_region);
3389
3390 // (a) Receiving thread must be the current thread.
3391 Node* rec_thr = argument(0);
3392 Node* tls_ptr = NULL;
3393 Node* cur_thr = generate_current_thread(tls_ptr);
3394 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3395 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3396
3397 generate_slow_guard(bol_thr, slow_region);
3398
3399 // (b) Interrupt bit on TLS must be false.
3400 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3401 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3402 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3403
3404 // Set the control input on the field _interrupted read to prevent it floating up.
3405 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3406 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3407 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3408
3409 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3410
3411 // First fast path: if (!TLS._interrupted) return false;
3412 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3413 result_rgn->init_req(no_int_result_path, false_bit);
3414 result_val->init_req(no_int_result_path, intcon(0));
3415
3416 // drop through to next case
3417 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3418
3419 #ifndef _WINDOWS
3420 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3421 Node* clr_arg = argument(1);
3422 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3423 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3424 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3425
3426 // Second fast path: ... else if (!clear_int) return true;
3427 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3428 result_rgn->init_req(no_clear_result_path, false_arg);
3429 result_val->init_req(no_clear_result_path, intcon(1));
3430
3431 // drop through to next case
3432 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3433 #else
3434 // To return true on Windows you must read the _interrupted field
3435 // and check the event state i.e. take the slow path.
3436 #endif // _WINDOWS
3437
3438 // (d) Otherwise, go to the slow path.
3439 slow_region->add_req(control());
3440 set_control( _gvn.transform(slow_region));
3441
3442 if (stopped()) {
3443 // There is no slow path.
3444 result_rgn->init_req(slow_result_path, top());
3445 result_val->init_req(slow_result_path, top());
3446 } else {
3447 // non-virtual because it is a private non-static
3448 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3449
3450 Node* slow_val = set_results_for_java_call(slow_call);
3451 // this->control() comes from set_results_for_java_call
3452
3453 Node* fast_io = slow_call->in(TypeFunc::I_O);
3454 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3455
3456 // These two phis are pre-filled with copies of of the fast IO and Memory
3457 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3458 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3459
3460 result_rgn->init_req(slow_result_path, control());
3461 result_io ->init_req(slow_result_path, i_o());
3462 result_mem->init_req(slow_result_path, reset_memory());
3463 result_val->init_req(slow_result_path, slow_val);
3464
3465 set_all_memory(_gvn.transform(result_mem));
3466 set_i_o( _gvn.transform(result_io));
3467 }
3468
3469 C->set_has_split_ifs(true); // Has chance for split-if optimization
3470 set_result(result_rgn, result_val);
3471 return true;
3472 }
3473
3474 //---------------------------load_mirror_from_klass----------------------------
3475 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3476 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3477 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3478 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3479 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3480 }
3481
3482 //-----------------------load_klass_from_mirror_common-------------------------
3483 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3484 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3485 // and branch to the given path on the region.
3486 // If never_see_null, take an uncommon trap on null, so we can optimistically
3487 // compile for the non-null case.
3488 // If the region is NULL, force never_see_null = true.
3489 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3490 bool never_see_null,
3491 RegionNode* region,
3492 int null_path,
3493 int offset) {
3494 if (region == NULL) never_see_null = true;
3495 Node* p = basic_plus_adr(mirror, offset);
3496 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3497 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3498 Node* null_ctl = top();
3499 kls = null_check_oop(kls, &null_ctl, never_see_null);
3500 if (region != NULL) {
3501 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3502 region->init_req(null_path, null_ctl);
3503 } else {
3504 assert(null_ctl == top(), "no loose ends");
3505 }
3506 return kls;
3507 }
3508
3509 //--------------------(inline_native_Class_query helpers)---------------------
3510 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3511 // Fall through if (mods & mask) == bits, take the guard otherwise.
3512 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3513 // Branch around if the given klass has the given modifier bit set.
3514 // Like generate_guard, adds a new path onto the region.
3515 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3516 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3517 Node* mask = intcon(modifier_mask);
3518 Node* bits = intcon(modifier_bits);
3519 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3520 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3521 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3522 return generate_fair_guard(bol, region);
3523 }
3524 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3525 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3526 }
3527
3528 //-------------------------inline_native_Class_query-------------------
3529 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3530 const Type* return_type = TypeInt::BOOL;
3531 Node* prim_return_value = top(); // what happens if it's a primitive class?
3532 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3533 bool expect_prim = false; // most of these guys expect to work on refs
3534
3535 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3536
3537 Node* mirror = argument(0);
3538 Node* obj = top();
3539
3540 switch (id) {
3541 case vmIntrinsics::_isInstance:
3542 // nothing is an instance of a primitive type
3543 prim_return_value = intcon(0);
3544 obj = argument(1);
3545 break;
3546 case vmIntrinsics::_getModifiers:
3547 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3548 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3549 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3550 break;
3551 case vmIntrinsics::_isInterface:
3552 prim_return_value = intcon(0);
3553 break;
3554 case vmIntrinsics::_isArray:
3555 prim_return_value = intcon(0);
3556 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3557 break;
3558 case vmIntrinsics::_isPrimitive:
3559 prim_return_value = intcon(1);
3560 expect_prim = true; // obviously
3561 break;
3562 case vmIntrinsics::_getSuperclass:
3563 prim_return_value = null();
3564 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3565 break;
3566 case vmIntrinsics::_getClassAccessFlags:
3567 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3568 return_type = TypeInt::INT; // not bool! 6297094
3569 break;
3570 default:
3571 fatal_unexpected_iid(id);
3572 break;
3573 }
3574
3575 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3576 if (mirror_con == NULL) return false; // cannot happen?
3577
3578 #ifndef PRODUCT
3579 if (C->print_intrinsics() || C->print_inlining()) {
3580 ciType* k = mirror_con->java_mirror_type();
3581 if (k) {
3582 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3583 k->print_name();
3584 tty->cr();
3585 }
3586 }
3587 #endif
3588
3589 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3590 RegionNode* region = new RegionNode(PATH_LIMIT);
3591 record_for_igvn(region);
3592 PhiNode* phi = new PhiNode(region, return_type);
3593
3594 // The mirror will never be null of Reflection.getClassAccessFlags, however
3595 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3596 // if it is. See bug 4774291.
3597
3598 // For Reflection.getClassAccessFlags(), the null check occurs in
3599 // the wrong place; see inline_unsafe_access(), above, for a similar
3600 // situation.
3601 mirror = null_check(mirror);
3602 // If mirror or obj is dead, only null-path is taken.
3603 if (stopped()) return true;
3604
3605 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3606
3607 // Now load the mirror's klass metaobject, and null-check it.
3608 // Side-effects region with the control path if the klass is null.
3609 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3610 // If kls is null, we have a primitive mirror.
3611 phi->init_req(_prim_path, prim_return_value);
3612 if (stopped()) { set_result(region, phi); return true; }
3613 bool safe_for_replace = (region->in(_prim_path) == top());
3614
3615 Node* p; // handy temp
3616 Node* null_ctl;
3617
3618 // Now that we have the non-null klass, we can perform the real query.
3619 // For constant classes, the query will constant-fold in LoadNode::Value.
3620 Node* query_value = top();
3621 switch (id) {
3622 case vmIntrinsics::_isInstance:
3623 // nothing is an instance of a primitive type
3624 query_value = gen_instanceof(obj, kls, safe_for_replace);
3625 break;
3626
3627 case vmIntrinsics::_getModifiers:
3628 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3629 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3630 break;
3631
3632 case vmIntrinsics::_isInterface:
3633 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3634 if (generate_interface_guard(kls, region) != NULL)
3635 // A guard was added. If the guard is taken, it was an interface.
3636 phi->add_req(intcon(1));
3637 // If we fall through, it's a plain class.
3638 query_value = intcon(0);
3639 break;
3640
3641 case vmIntrinsics::_isArray:
3642 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3643 if (generate_array_guard(kls, region) != NULL)
3644 // A guard was added. If the guard is taken, it was an array.
3645 phi->add_req(intcon(1));
3646 // If we fall through, it's a plain class.
3647 query_value = intcon(0);
3648 break;
3649
3650 case vmIntrinsics::_isPrimitive:
3651 query_value = intcon(0); // "normal" path produces false
3652 break;
3653
3654 case vmIntrinsics::_getSuperclass:
3655 // The rules here are somewhat unfortunate, but we can still do better
3656 // with random logic than with a JNI call.
3657 // Interfaces store null or Object as _super, but must report null.
3658 // Arrays store an intermediate super as _super, but must report Object.
3659 // Other types can report the actual _super.
3660 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3661 if (generate_interface_guard(kls, region) != NULL)
3662 // A guard was added. If the guard is taken, it was an interface.
3663 phi->add_req(null());
3664 if (generate_array_guard(kls, region) != NULL)
3665 // A guard was added. If the guard is taken, it was an array.
3666 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3667 // If we fall through, it's a plain class. Get its _super.
3668 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3669 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::BOTTOM));
3670 null_ctl = top();
3671 kls = null_check_oop(kls, &null_ctl);
3672 if (null_ctl != top()) {
3673 // If the guard is taken, Object.superClass is null (both klass and mirror).
3674 region->add_req(null_ctl);
3675 phi ->add_req(null());
3676 }
3677 if (!stopped()) {
3678 query_value = load_mirror_from_klass(kls);
3679 }
3680 break;
3681
3682 case vmIntrinsics::_getClassAccessFlags:
3683 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3684 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3685 break;
3686
3687 default:
3688 fatal_unexpected_iid(id);
3689 break;
3690 }
3691
3692 // Fall-through is the normal case of a query to a real class.
3693 phi->init_req(1, query_value);
3694 region->init_req(1, control());
3695
3696 C->set_has_split_ifs(true); // Has chance for split-if optimization
3697 set_result(region, phi);
3698 return true;
3699 }
3700
3701 //-------------------------inline_Class_cast-------------------
3702 bool LibraryCallKit::inline_Class_cast() {
3703 Node* mirror = argument(0); // Class
3704 Node* obj = argument(1);
3705 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3706 if (mirror_con == NULL) {
3707 return false; // dead path (mirror->is_top()).
3708 }
3709 if (obj == NULL || obj->is_top()) {
3710 return false; // dead path
3711 }
3712 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3713
3714 // First, see if Class.cast() can be folded statically.
3715 // java_mirror_type() returns non-null for compile-time Class constants.
3716 ciType* tm = mirror_con->java_mirror_type();
3717 if (tm != NULL && tm->is_klass() &&
3718 tp != NULL && tp->klass() != NULL) {
3719 if (!tp->klass()->is_loaded()) {
3720 // Don't use intrinsic when class is not loaded.
3721 return false;
3722 } else {
3723 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3724 if (static_res == Compile::SSC_always_true) {
3725 // isInstance() is true - fold the code.
3726 set_result(obj);
3727 return true;
3728 } else if (static_res == Compile::SSC_always_false) {
3729 // Don't use intrinsic, have to throw ClassCastException.
3730 // If the reference is null, the non-intrinsic bytecode will
3731 // be optimized appropriately.
3732 return false;
3733 }
3734 }
3735 }
3736
3737 // Bailout intrinsic and do normal inlining if exception path is frequent.
3738 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3739 return false;
3740 }
3741
3742 // Generate dynamic checks.
3743 // Class.cast() is java implementation of _checkcast bytecode.
3744 // Do checkcast (Parse::do_checkcast()) optimizations here.
3745
3746 mirror = null_check(mirror);
3747 // If mirror is dead, only null-path is taken.
3748 if (stopped()) {
3749 return true;
3750 }
3751
3752 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3753 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3754 RegionNode* region = new RegionNode(PATH_LIMIT);
3755 record_for_igvn(region);
3756
3757 // Now load the mirror's klass metaobject, and null-check it.
3758 // If kls is null, we have a primitive mirror and
3759 // nothing is an instance of a primitive type.
3760 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3761
3762 Node* res = top();
3763 if (!stopped()) {
3764 Node* bad_type_ctrl = top();
3765 // Do checkcast optimizations.
3766 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3767 region->init_req(_bad_type_path, bad_type_ctrl);
3768 }
3769 if (region->in(_prim_path) != top() ||
3770 region->in(_bad_type_path) != top()) {
3771 // Let Interpreter throw ClassCastException.
3772 PreserveJVMState pjvms(this);
3773 set_control(_gvn.transform(region));
3774 uncommon_trap(Deoptimization::Reason_intrinsic,
3775 Deoptimization::Action_maybe_recompile);
3776 }
3777 if (!stopped()) {
3778 set_result(res);
3779 }
3780 return true;
3781 }
3782
3783
3784 //--------------------------inline_native_subtype_check------------------------
3785 // This intrinsic takes the JNI calls out of the heart of
3786 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3787 bool LibraryCallKit::inline_native_subtype_check() {
3788 // Pull both arguments off the stack.
3789 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3790 args[0] = argument(0);
3791 args[1] = argument(1);
3792 Node* klasses[2]; // corresponding Klasses: superk, subk
3793 klasses[0] = klasses[1] = top();
3794
3795 enum {
3796 // A full decision tree on {superc is prim, subc is prim}:
3797 _prim_0_path = 1, // {P,N} => false
3798 // {P,P} & superc!=subc => false
3799 _prim_same_path, // {P,P} & superc==subc => true
3800 _prim_1_path, // {N,P} => false
3801 _ref_subtype_path, // {N,N} & subtype check wins => true
3802 _both_ref_path, // {N,N} & subtype check loses => false
3803 PATH_LIMIT
3804 };
3805
3806 RegionNode* region = new RegionNode(PATH_LIMIT);
3807 Node* phi = new PhiNode(region, TypeInt::BOOL);
3808 record_for_igvn(region);
3809
3810 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3811 const TypeKlassPtr* kls_type = TypeKlassPtr::BOTTOM;
3812 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3813
3814 // First null-check both mirrors and load each mirror's klass metaobject.
3815 int which_arg;
3816 for (which_arg = 0; which_arg <= 1; which_arg++) {
3817 Node* arg = args[which_arg];
3818 arg = null_check(arg);
3819 if (stopped()) break;
3820 args[which_arg] = arg;
3821
3822 Node* p = basic_plus_adr(arg, class_klass_offset);
3823 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3824 klasses[which_arg] = _gvn.transform(kls);
3825 }
3826
3827 // Having loaded both klasses, test each for null.
3828 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3829 for (which_arg = 0; which_arg <= 1; which_arg++) {
3830 Node* kls = klasses[which_arg];
3831 Node* null_ctl = top();
3832 kls = null_check_oop(kls, &null_ctl, never_see_null);
3833 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3834 region->init_req(prim_path, null_ctl);
3835 if (stopped()) break;
3836 klasses[which_arg] = kls;
3837 }
3838
3839 if (!stopped()) {
3840 // now we have two reference types, in klasses[0..1]
3841 Node* subk = klasses[1]; // the argument to isAssignableFrom
3842 Node* superk = klasses[0]; // the receiver
3843 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3844 // now we have a successful reference subtype check
3845 region->set_req(_ref_subtype_path, control());
3846 }
3847
3848 // If both operands are primitive (both klasses null), then
3849 // we must return true when they are identical primitives.
3850 // It is convenient to test this after the first null klass check.
3851 set_control(region->in(_prim_0_path)); // go back to first null check
3852 if (!stopped()) {
3853 // Since superc is primitive, make a guard for the superc==subc case.
3854 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3855 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3856 generate_guard(bol_eq, region, PROB_FAIR);
3857 if (region->req() == PATH_LIMIT+1) {
3858 // A guard was added. If the added guard is taken, superc==subc.
3859 region->swap_edges(PATH_LIMIT, _prim_same_path);
3860 region->del_req(PATH_LIMIT);
3861 }
3862 region->set_req(_prim_0_path, control()); // Not equal after all.
3863 }
3864
3865 // these are the only paths that produce 'true':
3866 phi->set_req(_prim_same_path, intcon(1));
3867 phi->set_req(_ref_subtype_path, intcon(1));
3868
3869 // pull together the cases:
3870 assert(region->req() == PATH_LIMIT, "sane region");
3871 for (uint i = 1; i < region->req(); i++) {
3872 Node* ctl = region->in(i);
3873 if (ctl == NULL || ctl == top()) {
3874 region->set_req(i, top());
3875 phi ->set_req(i, top());
3876 } else if (phi->in(i) == NULL) {
3877 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3878 }
3879 }
3880
3881 set_control(_gvn.transform(region));
3882 set_result(_gvn.transform(phi));
3883 return true;
3884 }
3885
3886 //---------------------generate_array_guard_common------------------------
3887 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind) {
3888
3889 if (stopped()) {
3890 return NULL;
3891 }
3892
3893 // Like generate_guard, adds a new path onto the region.
3894 jint layout_con = 0;
3895 Node* layout_val = get_layout_helper(kls, layout_con);
3896 if (layout_val == NULL) {
3897 bool query = 0;
3898 switch(kind) {
3899 case ObjectArray: query = Klass::layout_helper_is_objArray(layout_con); break;
3900 case NonObjectArray: query = !Klass::layout_helper_is_objArray(layout_con); break;
3901 case TypeArray: query = Klass::layout_helper_is_typeArray(layout_con); break;
3902 case ValueArray: query = Klass::layout_helper_is_valueArray(layout_con); break;
3903 case AnyArray: query = Klass::layout_helper_is_array(layout_con); break;
3904 case NonArray: query = !Klass::layout_helper_is_array(layout_con); break;
3905 default:
3906 ShouldNotReachHere();
3907 }
3908 if (!query) {
3909 return NULL; // never a branch
3910 } else { // always a branch
3911 Node* always_branch = control();
3912 if (region != NULL)
3913 region->add_req(always_branch);
3914 set_control(top());
3915 return always_branch;
3916 }
3917 }
3918 unsigned int value = 0;
3919 BoolTest::mask btest = BoolTest::illegal;
3920 switch(kind) {
3921 case ObjectArray:
3922 case NonObjectArray: {
3923 value = Klass::_lh_array_tag_obj_value;
3924 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3925 btest = kind == ObjectArray ? BoolTest::eq : BoolTest::ne;
3926 break;
3927 }
3928 case TypeArray: {
3929 value = Klass::_lh_array_tag_type_value;
3930 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3931 btest = BoolTest::eq;
3932 break;
3933 }
3934 case ValueArray: {
3935 value = Klass::_lh_array_tag_vt_value;
3936 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3937 btest = BoolTest::eq;
3938 break;
3939 }
3940 case AnyArray: value = Klass::_lh_neutral_value; btest = BoolTest::lt; break;
3941 case NonArray: value = Klass::_lh_neutral_value; btest = BoolTest::gt; break;
3942 default:
3943 ShouldNotReachHere();
3944 }
3945 // Now test the correct condition.
3946 jint nval = (jint)value;
3947 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3948 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3949 return generate_fair_guard(bol, region);
3950 }
3951
3952
3953 //-----------------------inline_native_newArray--------------------------
3954 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3955 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3956 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3957 Node* mirror;
3958 Node* count_val;
3959 if (uninitialized) {
3960 mirror = argument(1);
3961 count_val = argument(2);
3962 } else {
3963 mirror = argument(0);
3964 count_val = argument(1);
3965 }
3966
3967 mirror = null_check(mirror);
3968 // If mirror or obj is dead, only null-path is taken.
3969 if (stopped()) return true;
3970
3971 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3972 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3973 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3974 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3975 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3976
3977 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3978 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3979 result_reg, _slow_path);
3980 Node* normal_ctl = control();
3981 Node* no_array_ctl = result_reg->in(_slow_path);
3982
3983 // Generate code for the slow case. We make a call to newArray().
3984 set_control(no_array_ctl);
3985 if (!stopped()) {
3986 // Either the input type is void.class, or else the
3987 // array klass has not yet been cached. Either the
3988 // ensuing call will throw an exception, or else it
3989 // will cache the array klass for next time.
3990 PreserveJVMState pjvms(this);
3991 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3992 Node* slow_result = set_results_for_java_call(slow_call);
3993 // this->control() comes from set_results_for_java_call
3994 result_reg->set_req(_slow_path, control());
3995 result_val->set_req(_slow_path, slow_result);
3996 result_io ->set_req(_slow_path, i_o());
3997 result_mem->set_req(_slow_path, reset_memory());
3998 }
3999
4000 set_control(normal_ctl);
4001 if (!stopped()) {
4002 // Normal case: The array type has been cached in the java.lang.Class.
4003 // The following call works fine even if the array type is polymorphic.
4004 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4005 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
4006 result_reg->init_req(_normal_path, control());
4007 result_val->init_req(_normal_path, obj);
4008 result_io ->init_req(_normal_path, i_o());
4009 result_mem->init_req(_normal_path, reset_memory());
4010
4011 if (uninitialized) {
4012 // Mark the allocation so that zeroing is skipped
4013 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4014 alloc->maybe_set_complete(&_gvn);
4015 }
4016 }
4017
4018 // Return the combined state.
4019 set_i_o( _gvn.transform(result_io) );
4020 set_all_memory( _gvn.transform(result_mem));
4021
4022 C->set_has_split_ifs(true); // Has chance for split-if optimization
4023 set_result(result_reg, result_val);
4024 return true;
4025 }
4026
4027 //----------------------inline_native_getLength--------------------------
4028 // public static native int java.lang.reflect.Array.getLength(Object array);
4029 bool LibraryCallKit::inline_native_getLength() {
4030 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4031
4032 Node* array = null_check(argument(0));
4033 // If array is dead, only null-path is taken.
4034 if (stopped()) return true;
4035
4036 // Deoptimize if it is a non-array.
4037 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4038
4039 if (non_array != NULL) {
4040 PreserveJVMState pjvms(this);
4041 set_control(non_array);
4042 uncommon_trap(Deoptimization::Reason_intrinsic,
4043 Deoptimization::Action_maybe_recompile);
4044 }
4045
4046 // If control is dead, only non-array-path is taken.
4047 if (stopped()) return true;
4048
4049 // The works fine even if the array type is polymorphic.
4050 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4051 Node* result = load_array_length(array);
4052
4053 C->set_has_split_ifs(true); // Has chance for split-if optimization
4054 set_result(result);
4055 return true;
4056 }
4057
4058 //------------------------inline_array_copyOf----------------------------
4059 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4060 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4061 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4062 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4063
4064 // Get the arguments.
4065 Node* original = argument(0);
4066 Node* start = is_copyOfRange? argument(1): intcon(0);
4067 Node* end = is_copyOfRange? argument(2): argument(1);
4068 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4069
4070 Node* newcopy = NULL;
4071
4072 // Set the original stack and the reexecute bit for the interpreter to reexecute
4073 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4074 { PreserveReexecuteState preexecs(this);
4075 jvms()->set_should_reexecute(true);
4076
4077 array_type_mirror = null_check(array_type_mirror);
4078 original = null_check(original);
4079
4080 // Check if a null path was taken unconditionally.
4081 if (stopped()) return true;
4082
4083 Node* orig_length = load_array_length(original);
4084
4085 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4086 klass_node = null_check(klass_node);
4087
4088 RegionNode* bailout = new RegionNode(1);
4089 record_for_igvn(bailout);
4090
4091 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4092 // Bail out if that is so.
4093 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4094 if (not_objArray != NULL) {
4095 // Improve the klass node's type from the new optimistic assumption:
4096 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4097 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
4098 Node* cast = new CastPPNode(klass_node, akls);
4099 cast->init_req(0, control());
4100 klass_node = _gvn.transform(cast);
4101 }
4102
4103 // Bail out if either start or end is negative.
4104 generate_negative_guard(start, bailout, &start);
4105 generate_negative_guard(end, bailout, &end);
4106
4107 Node* length = end;
4108 if (_gvn.type(start) != TypeInt::ZERO) {
4109 length = _gvn.transform(new SubINode(end, start));
4110 }
4111
4112 // Bail out if length is negative.
4113 // Without this the new_array would throw
4114 // NegativeArraySizeException but IllegalArgumentException is what
4115 // should be thrown
4116 generate_negative_guard(length, bailout, &length);
4117
4118 if (bailout->req() > 1) {
4119 PreserveJVMState pjvms(this);
4120 set_control(_gvn.transform(bailout));
4121 uncommon_trap(Deoptimization::Reason_intrinsic,
4122 Deoptimization::Action_maybe_recompile);
4123 }
4124
4125 if (!stopped()) {
4126 // How many elements will we copy from the original?
4127 // The answer is MinI(orig_length - start, length).
4128 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4129 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4130
4131 // Generate a direct call to the right arraycopy function(s).
4132 // We know the copy is disjoint but we might not know if the
4133 // oop stores need checking.
4134 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4135 // This will fail a store-check if x contains any non-nulls.
4136
4137 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4138 // loads/stores but it is legal only if we're sure the
4139 // Arrays.copyOf would succeed. So we need all input arguments
4140 // to the copyOf to be validated, including that the copy to the
4141 // new array won't trigger an ArrayStoreException. That subtype
4142 // check can be optimized if we know something on the type of
4143 // the input array from type speculation.
4144 if (_gvn.type(klass_node)->singleton()) {
4145 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4146 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4147
4148 int test = C->static_subtype_check(superk, subk);
4149 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4150 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4151 if (t_original->speculative_type() != NULL) {
4152 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4153 }
4154 }
4155 }
4156
4157 bool validated = false;
4158 // Reason_class_check rather than Reason_intrinsic because we
4159 // want to intrinsify even if this traps.
4160 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4161 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4162 klass_node);
4163
4164 if (not_subtype_ctrl != top()) {
4165 PreserveJVMState pjvms(this);
4166 set_control(not_subtype_ctrl);
4167 uncommon_trap(Deoptimization::Reason_class_check,
4168 Deoptimization::Action_make_not_entrant);
4169 assert(stopped(), "Should be stopped");
4170 }
4171 validated = true;
4172 }
4173
4174 if (!stopped()) {
4175 newcopy = new_array(klass_node, length, 0); // no arguments to push
4176
4177 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4178 load_object_klass(original), klass_node);
4179 if (!is_copyOfRange) {
4180 ac->set_copyof(validated);
4181 } else {
4182 ac->set_copyofrange(validated);
4183 }
4184 Node* n = _gvn.transform(ac);
4185 if (n == ac) {
4186 ac->connect_outputs(this);
4187 } else {
4188 assert(validated, "shouldn't transform if all arguments not validated");
4189 set_all_memory(n);
4190 }
4191 }
4192 }
4193 } // original reexecute is set back here
4194
4195 C->set_has_split_ifs(true); // Has chance for split-if optimization
4196 if (!stopped()) {
4197 set_result(newcopy);
4198 }
4199 return true;
4200 }
4201
4202
4203 //----------------------generate_virtual_guard---------------------------
4204 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4205 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4206 RegionNode* slow_region) {
4207 ciMethod* method = callee();
4208 int vtable_index = method->vtable_index();
4209 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4210 "bad index %d", vtable_index);
4211 // Get the Method* out of the appropriate vtable entry.
4212 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4213 vtable_index*vtableEntry::size_in_bytes() +
4214 vtableEntry::method_offset_in_bytes();
4215 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4216 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4217
4218 // Compare the target method with the expected method (e.g., Object.hashCode).
4219 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4220
4221 Node* native_call = makecon(native_call_addr);
4222 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4223 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4224
4225 return generate_slow_guard(test_native, slow_region);
4226 }
4227
4228 //-----------------------generate_method_call----------------------------
4229 // Use generate_method_call to make a slow-call to the real
4230 // method if the fast path fails. An alternative would be to
4231 // use a stub like OptoRuntime::slow_arraycopy_Java.
4232 // This only works for expanding the current library call,
4233 // not another intrinsic. (E.g., don't use this for making an
4234 // arraycopy call inside of the copyOf intrinsic.)
4235 CallJavaNode*
4236 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4237 // When compiling the intrinsic method itself, do not use this technique.
4238 guarantee(callee() != C->method(), "cannot make slow-call to self");
4239
4240 ciMethod* method = callee();
4241 // ensure the JVMS we have will be correct for this call
4242 guarantee(method_id == method->intrinsic_id(), "must match");
4243
4244 const TypeFunc* tf = TypeFunc::make(method);
4245 CallJavaNode* slow_call;
4246 if (is_static) {
4247 assert(!is_virtual, "");
4248 slow_call = new CallStaticJavaNode(C, tf,
4249 SharedRuntime::get_resolve_static_call_stub(),
4250 method, bci());
4251 } else if (is_virtual) {
4252 null_check_receiver();
4253 int vtable_index = Method::invalid_vtable_index;
4254 if (UseInlineCaches) {
4255 // Suppress the vtable call
4256 } else {
4257 // hashCode and clone are not a miranda methods,
4258 // so the vtable index is fixed.
4259 // No need to use the linkResolver to get it.
4260 vtable_index = method->vtable_index();
4261 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4262 "bad index %d", vtable_index);
4263 }
4264 slow_call = new CallDynamicJavaNode(tf,
4265 SharedRuntime::get_resolve_virtual_call_stub(),
4266 method, vtable_index, bci());
4267 } else { // neither virtual nor static: opt_virtual
4268 null_check_receiver();
4269 slow_call = new CallStaticJavaNode(C, tf,
4270 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4271 method, bci());
4272 slow_call->set_optimized_virtual(true);
4273 }
4274 set_arguments_for_java_call(slow_call);
4275 set_edges_for_java_call(slow_call);
4276 return slow_call;
4277 }
4278
4279
4280 /**
4281 * Build special case code for calls to hashCode on an object. This call may
4282 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4283 * slightly different code.
4284 */
4285 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4286 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4287 assert(!(is_virtual && is_static), "either virtual, special, or static");
4288
4289 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4290
4291 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4292 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4293 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4294 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4295 Node* obj = NULL;
4296 if (!is_static) {
4297 // Check for hashing null object
4298 obj = null_check_receiver();
4299 if (stopped()) return true; // unconditionally null
4300 result_reg->init_req(_null_path, top());
4301 result_val->init_req(_null_path, top());
4302 } else {
4303 // Do a null check, and return zero if null.
4304 // System.identityHashCode(null) == 0
4305 obj = argument(0);
4306 Node* null_ctl = top();
4307 obj = null_check_oop(obj, &null_ctl);
4308 result_reg->init_req(_null_path, null_ctl);
4309 result_val->init_req(_null_path, _gvn.intcon(0));
4310 }
4311
4312 // Unconditionally null? Then return right away.
4313 if (stopped()) {
4314 set_control( result_reg->in(_null_path));
4315 if (!stopped())
4316 set_result(result_val->in(_null_path));
4317 return true;
4318 }
4319
4320 // We only go to the fast case code if we pass a number of guards. The
4321 // paths which do not pass are accumulated in the slow_region.
4322 RegionNode* slow_region = new RegionNode(1);
4323 record_for_igvn(slow_region);
4324
4325 // If this is a virtual call, we generate a funny guard. We pull out
4326 // the vtable entry corresponding to hashCode() from the target object.
4327 // If the target method which we are calling happens to be the native
4328 // Object hashCode() method, we pass the guard. We do not need this
4329 // guard for non-virtual calls -- the caller is known to be the native
4330 // Object hashCode().
4331 if (is_virtual) {
4332 // After null check, get the object's klass.
4333 Node* obj_klass = load_object_klass(obj);
4334 generate_virtual_guard(obj_klass, slow_region);
4335 }
4336
4337 // Get the header out of the object, use LoadMarkNode when available
4338 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4339 // The control of the load must be NULL. Otherwise, the load can move before
4340 // the null check after castPP removal.
4341 Node* no_ctrl = NULL;
4342 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4343
4344 // Test the header to see if it is unlocked.
4345 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4346 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4347 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4348 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4349 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4350
4351 generate_slow_guard(test_unlocked, slow_region);
4352
4353 // Get the hash value and check to see that it has been properly assigned.
4354 // We depend on hash_mask being at most 32 bits and avoid the use of
4355 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4356 // vm: see markOop.hpp.
4357 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4358 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4359 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4360 // This hack lets the hash bits live anywhere in the mark object now, as long
4361 // as the shift drops the relevant bits into the low 32 bits. Note that
4362 // Java spec says that HashCode is an int so there's no point in capturing
4363 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4364 hshifted_header = ConvX2I(hshifted_header);
4365 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4366
4367 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4368 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4369 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4370
4371 generate_slow_guard(test_assigned, slow_region);
4372
4373 Node* init_mem = reset_memory();
4374 // fill in the rest of the null path:
4375 result_io ->init_req(_null_path, i_o());
4376 result_mem->init_req(_null_path, init_mem);
4377
4378 result_val->init_req(_fast_path, hash_val);
4379 result_reg->init_req(_fast_path, control());
4380 result_io ->init_req(_fast_path, i_o());
4381 result_mem->init_req(_fast_path, init_mem);
4382
4383 // Generate code for the slow case. We make a call to hashCode().
4384 set_control(_gvn.transform(slow_region));
4385 if (!stopped()) {
4386 // No need for PreserveJVMState, because we're using up the present state.
4387 set_all_memory(init_mem);
4388 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4389 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4390 Node* slow_result = set_results_for_java_call(slow_call);
4391 // this->control() comes from set_results_for_java_call
4392 result_reg->init_req(_slow_path, control());
4393 result_val->init_req(_slow_path, slow_result);
4394 result_io ->set_req(_slow_path, i_o());
4395 result_mem ->set_req(_slow_path, reset_memory());
4396 }
4397
4398 // Return the combined state.
4399 set_i_o( _gvn.transform(result_io) );
4400 set_all_memory( _gvn.transform(result_mem));
4401
4402 set_result(result_reg, result_val);
4403 return true;
4404 }
4405
4406 //---------------------------inline_native_getClass----------------------------
4407 // public final native Class<?> java.lang.Object.getClass();
4408 //
4409 // Build special case code for calls to getClass on an object.
4410 bool LibraryCallKit::inline_native_getClass() {
4411 Node* obj = null_check_receiver();
4412 if (stopped()) return true;
4413 set_result(load_mirror_from_klass(load_object_klass(obj)));
4414 return true;
4415 }
4416
4417 //-----------------inline_native_Reflection_getCallerClass---------------------
4418 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4419 //
4420 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4421 //
4422 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4423 // in that it must skip particular security frames and checks for
4424 // caller sensitive methods.
4425 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4426 #ifndef PRODUCT
4427 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4428 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4429 }
4430 #endif
4431
4432 if (!jvms()->has_method()) {
4433 #ifndef PRODUCT
4434 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4435 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4436 }
4437 #endif
4438 return false;
4439 }
4440
4441 // Walk back up the JVM state to find the caller at the required
4442 // depth.
4443 JVMState* caller_jvms = jvms();
4444
4445 // Cf. JVM_GetCallerClass
4446 // NOTE: Start the loop at depth 1 because the current JVM state does
4447 // not include the Reflection.getCallerClass() frame.
4448 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4449 ciMethod* m = caller_jvms->method();
4450 switch (n) {
4451 case 0:
4452 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4453 break;
4454 case 1:
4455 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4456 if (!m->caller_sensitive()) {
4457 #ifndef PRODUCT
4458 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4459 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4460 }
4461 #endif
4462 return false; // bail-out; let JVM_GetCallerClass do the work
4463 }
4464 break;
4465 default:
4466 if (!m->is_ignored_by_security_stack_walk()) {
4467 // We have reached the desired frame; return the holder class.
4468 // Acquire method holder as java.lang.Class and push as constant.
4469 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4470 ciInstance* caller_mirror = caller_klass->java_mirror();
4471 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4472
4473 #ifndef PRODUCT
4474 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4475 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());
4476 tty->print_cr(" JVM state at this point:");
4477 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4478 ciMethod* m = jvms()->of_depth(i)->method();
4479 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4480 }
4481 }
4482 #endif
4483 return true;
4484 }
4485 break;
4486 }
4487 }
4488
4489 #ifndef PRODUCT
4490 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4491 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4492 tty->print_cr(" JVM state at this point:");
4493 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4494 ciMethod* m = jvms()->of_depth(i)->method();
4495 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4496 }
4497 }
4498 #endif
4499
4500 return false; // bail-out; let JVM_GetCallerClass do the work
4501 }
4502
4503 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4504 Node* arg = argument(0);
4505 Node* result = NULL;
4506
4507 switch (id) {
4508 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4509 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4510 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4511 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4512
4513 case vmIntrinsics::_doubleToLongBits: {
4514 // two paths (plus control) merge in a wood
4515 RegionNode *r = new RegionNode(3);
4516 Node *phi = new PhiNode(r, TypeLong::LONG);
4517
4518 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4519 // Build the boolean node
4520 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4521
4522 // Branch either way.
4523 // NaN case is less traveled, which makes all the difference.
4524 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4525 Node *opt_isnan = _gvn.transform(ifisnan);
4526 assert( opt_isnan->is_If(), "Expect an IfNode");
4527 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4528 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4529
4530 set_control(iftrue);
4531
4532 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4533 Node *slow_result = longcon(nan_bits); // return NaN
4534 phi->init_req(1, _gvn.transform( slow_result ));
4535 r->init_req(1, iftrue);
4536
4537 // Else fall through
4538 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4539 set_control(iffalse);
4540
4541 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4542 r->init_req(2, iffalse);
4543
4544 // Post merge
4545 set_control(_gvn.transform(r));
4546 record_for_igvn(r);
4547
4548 C->set_has_split_ifs(true); // Has chance for split-if optimization
4549 result = phi;
4550 assert(result->bottom_type()->isa_long(), "must be");
4551 break;
4552 }
4553
4554 case vmIntrinsics::_floatToIntBits: {
4555 // two paths (plus control) merge in a wood
4556 RegionNode *r = new RegionNode(3);
4557 Node *phi = new PhiNode(r, TypeInt::INT);
4558
4559 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4560 // Build the boolean node
4561 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4562
4563 // Branch either way.
4564 // NaN case is less traveled, which makes all the difference.
4565 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4566 Node *opt_isnan = _gvn.transform(ifisnan);
4567 assert( opt_isnan->is_If(), "Expect an IfNode");
4568 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4569 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4570
4571 set_control(iftrue);
4572
4573 static const jint nan_bits = 0x7fc00000;
4574 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4575 phi->init_req(1, _gvn.transform( slow_result ));
4576 r->init_req(1, iftrue);
4577
4578 // Else fall through
4579 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4580 set_control(iffalse);
4581
4582 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4583 r->init_req(2, iffalse);
4584
4585 // Post merge
4586 set_control(_gvn.transform(r));
4587 record_for_igvn(r);
4588
4589 C->set_has_split_ifs(true); // Has chance for split-if optimization
4590 result = phi;
4591 assert(result->bottom_type()->isa_int(), "must be");
4592 break;
4593 }
4594
4595 default:
4596 fatal_unexpected_iid(id);
4597 break;
4598 }
4599 set_result(_gvn.transform(result));
4600 return true;
4601 }
4602
4603 //----------------------inline_unsafe_copyMemory-------------------------
4604 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4605 bool LibraryCallKit::inline_unsafe_copyMemory() {
4606 if (callee()->is_static()) return false; // caller must have the capability!
4607 null_check_receiver(); // null-check receiver
4608 if (stopped()) return true;
4609
4610 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4611
4612 Node* src_ptr = argument(1); // type: oop
4613 Node* src_off = ConvL2X(argument(2)); // type: long
4614 Node* dst_ptr = argument(4); // type: oop
4615 Node* dst_off = ConvL2X(argument(5)); // type: long
4616 Node* size = ConvL2X(argument(7)); // type: long
4617
4618 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4619 "fieldOffset must be byte-scaled");
4620
4621 Node* src = make_unsafe_address(src_ptr, src_off);
4622 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4623
4624 // Conservatively insert a memory barrier on all memory slices.
4625 // Do not let writes of the copy source or destination float below the copy.
4626 insert_mem_bar(Op_MemBarCPUOrder);
4627
4628 // Call it. Note that the length argument is not scaled.
4629 make_runtime_call(RC_LEAF|RC_NO_FP,
4630 OptoRuntime::fast_arraycopy_Type(),
4631 StubRoutines::unsafe_arraycopy(),
4632 "unsafe_arraycopy",
4633 TypeRawPtr::BOTTOM,
4634 src, dst, size XTOP);
4635
4636 // Do not let reads of the copy destination float above the copy.
4637 insert_mem_bar(Op_MemBarCPUOrder);
4638
4639 return true;
4640 }
4641
4642 //------------------------clone_coping-----------------------------------
4643 // Helper function for inline_native_clone.
4644 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4645 assert(obj_size != NULL, "");
4646 Node* raw_obj = alloc_obj->in(1);
4647 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4648
4649 AllocateNode* alloc = NULL;
4650 if (ReduceBulkZeroing) {
4651 // We will be completely responsible for initializing this object -
4652 // mark Initialize node as complete.
4653 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4654 // The object was just allocated - there should be no any stores!
4655 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4656 // Mark as complete_with_arraycopy so that on AllocateNode
4657 // expansion, we know this AllocateNode is initialized by an array
4658 // copy and a StoreStore barrier exists after the array copy.
4659 alloc->initialization()->set_complete_with_arraycopy();
4660 }
4661
4662 // Copy the fastest available way.
4663 // TODO: generate fields copies for small objects instead.
4664 Node* src = obj;
4665 Node* dest = alloc_obj;
4666 Node* size = _gvn.transform(obj_size);
4667
4668 // Exclude the header but include array length to copy by 8 bytes words.
4669 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4670 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4671 instanceOopDesc::base_offset_in_bytes();
4672 // base_off:
4673 // 8 - 32-bit VM
4674 // 12 - 64-bit VM, compressed klass
4675 // 16 - 64-bit VM, normal klass
4676 if (base_off % BytesPerLong != 0) {
4677 assert(UseCompressedClassPointers, "");
4678 if (is_array) {
4679 // Exclude length to copy by 8 bytes words.
4680 base_off += sizeof(int);
4681 } else {
4682 // Include klass to copy by 8 bytes words.
4683 base_off = instanceOopDesc::klass_offset_in_bytes();
4684 }
4685 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4686 }
4687 src = basic_plus_adr(src, base_off);
4688 dest = basic_plus_adr(dest, base_off);
4689
4690 // Compute the length also, if needed:
4691 Node* countx = size;
4692 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4693 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4694
4695 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4696
4697 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4698 ac->set_clonebasic();
4699 Node* n = _gvn.transform(ac);
4700 if (n == ac) {
4701 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4702 } else {
4703 set_all_memory(n);
4704 }
4705
4706 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4707 if (card_mark) {
4708 assert(!is_array, "");
4709 // Put in store barrier for any and all oops we are sticking
4710 // into this object. (We could avoid this if we could prove
4711 // that the object type contains no oop fields at all.)
4712 Node* no_particular_value = NULL;
4713 Node* no_particular_field = NULL;
4714 int raw_adr_idx = Compile::AliasIdxRaw;
4715 post_barrier(control(),
4716 memory(raw_adr_type),
4717 alloc_obj,
4718 no_particular_field,
4719 raw_adr_idx,
4720 no_particular_value,
4721 T_OBJECT,
4722 false);
4723 }
4724
4725 // Do not let reads from the cloned object float above the arraycopy.
4726 if (alloc != NULL) {
4727 // Do not let stores that initialize this object be reordered with
4728 // a subsequent store that would make this object accessible by
4729 // other threads.
4730 // Record what AllocateNode this StoreStore protects so that
4731 // escape analysis can go from the MemBarStoreStoreNode to the
4732 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4733 // based on the escape status of the AllocateNode.
4734 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4735 } else {
4736 insert_mem_bar(Op_MemBarCPUOrder);
4737 }
4738 }
4739
4740 //------------------------inline_native_clone----------------------------
4741 // protected native Object java.lang.Object.clone();
4742 //
4743 // Here are the simple edge cases:
4744 // null receiver => normal trap
4745 // virtual and clone was overridden => slow path to out-of-line clone
4746 // not cloneable or finalizer => slow path to out-of-line Object.clone
4747 //
4748 // The general case has two steps, allocation and copying.
4749 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4750 //
4751 // Copying also has two cases, oop arrays and everything else.
4752 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4753 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4754 //
4755 // These steps fold up nicely if and when the cloned object's klass
4756 // can be sharply typed as an object array, a type array, or an instance.
4757 //
4758 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4759 PhiNode* result_val;
4760
4761 // Set the reexecute bit for the interpreter to reexecute
4762 // the bytecode that invokes Object.clone if deoptimization happens.
4763 { PreserveReexecuteState preexecs(this);
4764 jvms()->set_should_reexecute(true);
4765
4766 Node* obj = null_check_receiver();
4767 if (stopped()) return true;
4768
4769 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4770
4771 // If we are going to clone an instance, we need its exact type to
4772 // know the number and types of fields to convert the clone to
4773 // loads/stores. Maybe a speculative type can help us.
4774 if (!obj_type->klass_is_exact() &&
4775 obj_type->speculative_type() != NULL &&
4776 obj_type->speculative_type()->is_instance_klass()) {
4777 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4778 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4779 !spec_ik->has_injected_fields()) {
4780 ciKlass* k = obj_type->klass();
4781 if (!k->is_instance_klass() ||
4782 k->as_instance_klass()->is_interface() ||
4783 k->as_instance_klass()->has_subklass()) {
4784 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4785 }
4786 }
4787 }
4788
4789 Node* obj_klass = load_object_klass(obj);
4790 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4791 const TypeOopPtr* toop = ((tklass != NULL)
4792 ? tklass->as_instance_type()
4793 : TypeInstPtr::NOTNULL);
4794
4795 // Conservatively insert a memory barrier on all memory slices.
4796 // Do not let writes into the original float below the clone.
4797 insert_mem_bar(Op_MemBarCPUOrder);
4798
4799 // paths into result_reg:
4800 enum {
4801 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4802 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4803 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4804 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4805 PATH_LIMIT
4806 };
4807 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4808 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4809 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4810 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4811 record_for_igvn(result_reg);
4812
4813 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4814 int raw_adr_idx = Compile::AliasIdxRaw;
4815
4816 // We only go to the fast case code if we pass a number of guards.
4817 // The paths which do not pass are accumulated in the slow_region.
4818 RegionNode* slow_region = new RegionNode(1);
4819 record_for_igvn(slow_region);
4820
4821 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4822 if (array_ctl != NULL) {
4823 // It's an array.
4824 PreserveJVMState pjvms(this);
4825 set_control(array_ctl);
4826
4827 if (!use_ReduceInitialCardMarks()) {
4828 // Value type array may have object field that would require a
4829 // write barrier. Conservatively, go to slow path.
4830 generate_valueArray_guard(obj_klass, slow_region);
4831 }
4832
4833 if (!stopped()) {
4834 Node* obj_length = load_array_length(obj);
4835 Node* obj_size = NULL;
4836 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4837
4838 if (!use_ReduceInitialCardMarks()) {
4839 // If it is an oop array, it requires very special treatment,
4840 // because card marking is required on each card of the array.
4841 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4842 if (is_obja != NULL) {
4843 PreserveJVMState pjvms2(this);
4844 set_control(is_obja);
4845 // Generate a direct call to the right arraycopy function(s).
4846 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4847 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4848 ac->set_cloneoop();
4849 Node* n = _gvn.transform(ac);
4850 assert(n == ac, "cannot disappear");
4851 ac->connect_outputs(this);
4852
4853 result_reg->init_req(_objArray_path, control());
4854 result_val->init_req(_objArray_path, alloc_obj);
4855 result_i_o ->set_req(_objArray_path, i_o());
4856 result_mem ->set_req(_objArray_path, reset_memory());
4857 }
4858 }
4859
4860 // Otherwise, there are no card marks to worry about.
4861 // (We can dispense with card marks if we know the allocation
4862 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4863 // causes the non-eden paths to take compensating steps to
4864 // simulate a fresh allocation, so that no further
4865 // card marks are required in compiled code to initialize
4866 // the object.)
4867
4868 if (!stopped()) {
4869 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4870
4871 // Present the results of the copy.
4872 result_reg->init_req(_array_path, control());
4873 result_val->init_req(_array_path, alloc_obj);
4874 result_i_o ->set_req(_array_path, i_o());
4875 result_mem ->set_req(_array_path, reset_memory());
4876 }
4877 }
4878 }
4879
4880 if (!stopped()) {
4881 // It's an instance (we did array above). Make the slow-path tests.
4882 // If this is a virtual call, we generate a funny guard. We grab
4883 // the vtable entry corresponding to clone() from the target object.
4884 // If the target method which we are calling happens to be the
4885 // Object clone() method, we pass the guard. We do not need this
4886 // guard for non-virtual calls; the caller is known to be the native
4887 // Object clone().
4888 if (is_virtual) {
4889 generate_virtual_guard(obj_klass, slow_region);
4890 }
4891
4892 // The object must be easily cloneable and must not have a finalizer.
4893 // Both of these conditions may be checked in a single test.
4894 // We could optimize the test further, but we don't care.
4895 generate_access_flags_guard(obj_klass,
4896 // Test both conditions:
4897 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4898 // Must be cloneable but not finalizer:
4899 JVM_ACC_IS_CLONEABLE_FAST,
4900 slow_region);
4901 }
4902
4903 if (!stopped()) {
4904 // It's an instance, and it passed the slow-path tests.
4905 PreserveJVMState pjvms(this);
4906 Node* obj_size = NULL;
4907 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4908 // is reexecuted if deoptimization occurs and there could be problems when merging
4909 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4910 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4911
4912 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4913
4914 // Present the results of the slow call.
4915 result_reg->init_req(_instance_path, control());
4916 result_val->init_req(_instance_path, alloc_obj);
4917 result_i_o ->set_req(_instance_path, i_o());
4918 result_mem ->set_req(_instance_path, reset_memory());
4919 }
4920
4921 // Generate code for the slow case. We make a call to clone().
4922 set_control(_gvn.transform(slow_region));
4923 if (!stopped()) {
4924 PreserveJVMState pjvms(this);
4925 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4926 Node* slow_result = set_results_for_java_call(slow_call);
4927 // this->control() comes from set_results_for_java_call
4928 result_reg->init_req(_slow_path, control());
4929 result_val->init_req(_slow_path, slow_result);
4930 result_i_o ->set_req(_slow_path, i_o());
4931 result_mem ->set_req(_slow_path, reset_memory());
4932 }
4933
4934 // Return the combined state.
4935 set_control( _gvn.transform(result_reg));
4936 set_i_o( _gvn.transform(result_i_o));
4937 set_all_memory( _gvn.transform(result_mem));
4938 } // original reexecute is set back here
4939
4940 set_result(_gvn.transform(result_val));
4941 return true;
4942 }
4943
4944 // If we have a tighly coupled allocation, the arraycopy may take care
4945 // of the array initialization. If one of the guards we insert between
4946 // the allocation and the arraycopy causes a deoptimization, an
4947 // unitialized array will escape the compiled method. To prevent that
4948 // we set the JVM state for uncommon traps between the allocation and
4949 // the arraycopy to the state before the allocation so, in case of
4950 // deoptimization, we'll reexecute the allocation and the
4951 // initialization.
4952 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4953 if (alloc != NULL) {
4954 ciMethod* trap_method = alloc->jvms()->method();
4955 int trap_bci = alloc->jvms()->bci();
4956
4957 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4958 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4959 // Make sure there's no store between the allocation and the
4960 // arraycopy otherwise visible side effects could be rexecuted
4961 // in case of deoptimization and cause incorrect execution.
4962 bool no_interfering_store = true;
4963 Node* mem = alloc->in(TypeFunc::Memory);
4964 if (mem->is_MergeMem()) {
4965 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4966 Node* n = mms.memory();
4967 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4968 assert(n->is_Store(), "what else?");
4969 no_interfering_store = false;
4970 break;
4971 }
4972 }
4973 } else {
4974 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4975 Node* n = mms.memory();
4976 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4977 assert(n->is_Store(), "what else?");
4978 no_interfering_store = false;
4979 break;
4980 }
4981 }
4982 }
4983
4984 if (no_interfering_store) {
4985 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4986 uint size = alloc->req();
4987 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4988 old_jvms->set_map(sfpt);
4989 for (uint i = 0; i < size; i++) {
4990 sfpt->init_req(i, alloc->in(i));
4991 }
4992 // re-push array length for deoptimization
4993 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4994 old_jvms->set_sp(old_jvms->sp()+1);
4995 old_jvms->set_monoff(old_jvms->monoff()+1);
4996 old_jvms->set_scloff(old_jvms->scloff()+1);
4997 old_jvms->set_endoff(old_jvms->endoff()+1);
4998 old_jvms->set_should_reexecute(true);
4999
5000 sfpt->set_i_o(map()->i_o());
5001 sfpt->set_memory(map()->memory());
5002 sfpt->set_control(map()->control());
5003
5004 JVMState* saved_jvms = jvms();
5005 saved_reexecute_sp = _reexecute_sp;
5006
5007 set_jvms(sfpt->jvms());
5008 _reexecute_sp = jvms()->sp();
5009
5010 return saved_jvms;
5011 }
5012 }
5013 }
5014 return NULL;
5015 }
5016
5017 // In case of a deoptimization, we restart execution at the
5018 // allocation, allocating a new array. We would leave an uninitialized
5019 // array in the heap that GCs wouldn't expect. Move the allocation
5020 // after the traps so we don't allocate the array if we
5021 // deoptimize. This is possible because tightly_coupled_allocation()
5022 // guarantees there's no observer of the allocated array at this point
5023 // and the control flow is simple enough.
5024 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5025 int saved_reexecute_sp, uint new_idx) {
5026 if (saved_jvms != NULL && !stopped()) {
5027 assert(alloc != NULL, "only with a tightly coupled allocation");
5028 // restore JVM state to the state at the arraycopy
5029 saved_jvms->map()->set_control(map()->control());
5030 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5031 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5032 // If we've improved the types of some nodes (null check) while
5033 // emitting the guards, propagate them to the current state
5034 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5035 set_jvms(saved_jvms);
5036 _reexecute_sp = saved_reexecute_sp;
5037
5038 // Remove the allocation from above the guards
5039 CallProjections callprojs;
5040 alloc->extract_projections(&callprojs, true);
5041 InitializeNode* init = alloc->initialization();
5042 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5043 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5044 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5045 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5046
5047 // move the allocation here (after the guards)
5048 _gvn.hash_delete(alloc);
5049 alloc->set_req(TypeFunc::Control, control());
5050 alloc->set_req(TypeFunc::I_O, i_o());
5051 Node *mem = reset_memory();
5052 set_all_memory(mem);
5053 alloc->set_req(TypeFunc::Memory, mem);
5054 set_control(init->proj_out(TypeFunc::Control));
5055 set_i_o(callprojs.fallthrough_ioproj);
5056
5057 // Update memory as done in GraphKit::set_output_for_allocation()
5058 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5059 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5060 if (ary_type->isa_aryptr() && length_type != NULL) {
5061 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5062 }
5063 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5064 int elemidx = C->get_alias_index(telemref);
5065 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
5066 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
5067
5068 Node* allocx = _gvn.transform(alloc);
5069 assert(allocx == alloc, "where has the allocation gone?");
5070 assert(dest->is_CheckCastPP(), "not an allocation result?");
5071
5072 _gvn.hash_delete(dest);
5073 dest->set_req(0, control());
5074 Node* destx = _gvn.transform(dest);
5075 assert(destx == dest, "where has the allocation result gone?");
5076 }
5077 }
5078
5079
5080 //------------------------------inline_arraycopy-----------------------
5081 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5082 // Object dest, int destPos,
5083 // int length);
5084 bool LibraryCallKit::inline_arraycopy() {
5085 // Get the arguments.
5086 Node* src = argument(0); // type: oop
5087 Node* src_offset = argument(1); // type: int
5088 Node* dest = argument(2); // type: oop
5089 Node* dest_offset = argument(3); // type: int
5090 Node* length = argument(4); // type: int
5091
5092 uint new_idx = C->unique();
5093
5094 // Check for allocation before we add nodes that would confuse
5095 // tightly_coupled_allocation()
5096 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5097
5098 int saved_reexecute_sp = -1;
5099 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5100 // See arraycopy_restore_alloc_state() comment
5101 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5102 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5103 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5104 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5105
5106 // The following tests must be performed
5107 // (1) src and dest are arrays.
5108 // (2) src and dest arrays must have elements of the same BasicType
5109 // (3) src and dest must not be null.
5110 // (4) src_offset must not be negative.
5111 // (5) dest_offset must not be negative.
5112 // (6) length must not be negative.
5113 // (7) src_offset + length must not exceed length of src.
5114 // (8) dest_offset + length must not exceed length of dest.
5115 // (9) each element of an oop array must be assignable
5116
5117 // (3) src and dest must not be null.
5118 // always do this here because we need the JVM state for uncommon traps
5119 Node* null_ctl = top();
5120 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5121 assert(null_ctl->is_top(), "no null control here");
5122 dest = null_check(dest, T_ARRAY);
5123
5124 if (!can_emit_guards) {
5125 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5126 // guards but the arraycopy node could still take advantage of a
5127 // tightly allocated allocation. tightly_coupled_allocation() is
5128 // called again to make sure it takes the null check above into
5129 // account: the null check is mandatory and if it caused an
5130 // uncommon trap to be emitted then the allocation can't be
5131 // considered tightly coupled in this context.
5132 alloc = tightly_coupled_allocation(dest, NULL);
5133 }
5134
5135 bool validated = false;
5136
5137 const Type* src_type = _gvn.type(src);
5138 const Type* dest_type = _gvn.type(dest);
5139 const TypeAryPtr* top_src = src_type->isa_aryptr();
5140 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5141
5142 // Do we have the type of src?
5143 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5144 // Do we have the type of dest?
5145 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5146 // Is the type for src from speculation?
5147 bool src_spec = false;
5148 // Is the type for dest from speculation?
5149 bool dest_spec = false;
5150
5151 if ((!has_src || !has_dest) && can_emit_guards) {
5152 // We don't have sufficient type information, let's see if
5153 // speculative types can help. We need to have types for both src
5154 // and dest so that it pays off.
5155
5156 // Do we already have or could we have type information for src
5157 bool could_have_src = has_src;
5158 // Do we already have or could we have type information for dest
5159 bool could_have_dest = has_dest;
5160
5161 ciKlass* src_k = NULL;
5162 if (!has_src) {
5163 src_k = src_type->speculative_type_not_null();
5164 if (src_k != NULL && src_k->is_array_klass()) {
5165 could_have_src = true;
5166 }
5167 }
5168
5169 ciKlass* dest_k = NULL;
5170 if (!has_dest) {
5171 dest_k = dest_type->speculative_type_not_null();
5172 if (dest_k != NULL && dest_k->is_array_klass()) {
5173 could_have_dest = true;
5174 }
5175 }
5176
5177 if (could_have_src && could_have_dest) {
5178 // This is going to pay off so emit the required guards
5179 if (!has_src) {
5180 src = maybe_cast_profiled_obj(src, src_k, true);
5181 src_type = _gvn.type(src);
5182 top_src = src_type->isa_aryptr();
5183 has_src = (top_src != NULL && top_src->klass() != NULL);
5184 src_spec = true;
5185 }
5186 if (!has_dest) {
5187 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5188 dest_type = _gvn.type(dest);
5189 top_dest = dest_type->isa_aryptr();
5190 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5191 dest_spec = true;
5192 }
5193 }
5194 }
5195
5196 if (has_src && has_dest && can_emit_guards) {
5197 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5198 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5199 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5200 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5201
5202 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5203 // If both arrays are object arrays then having the exact types
5204 // for both will remove the need for a subtype check at runtime
5205 // before the call and may make it possible to pick a faster copy
5206 // routine (without a subtype check on every element)
5207 // Do we have the exact type of src?
5208 bool could_have_src = src_spec;
5209 // Do we have the exact type of dest?
5210 bool could_have_dest = dest_spec;
5211 ciKlass* src_k = top_src->klass();
5212 ciKlass* dest_k = top_dest->klass();
5213 if (!src_spec) {
5214 src_k = src_type->speculative_type_not_null();
5215 if (src_k != NULL && src_k->is_array_klass()) {
5216 could_have_src = true;
5217 }
5218 }
5219 if (!dest_spec) {
5220 dest_k = dest_type->speculative_type_not_null();
5221 if (dest_k != NULL && dest_k->is_array_klass()) {
5222 could_have_dest = true;
5223 }
5224 }
5225 if (could_have_src && could_have_dest) {
5226 // If we can have both exact types, emit the missing guards
5227 if (could_have_src && !src_spec) {
5228 src = maybe_cast_profiled_obj(src, src_k, true);
5229 }
5230 if (could_have_dest && !dest_spec) {
5231 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5232 }
5233 }
5234 }
5235 }
5236
5237 ciMethod* trap_method = method();
5238 int trap_bci = bci();
5239 if (saved_jvms != NULL) {
5240 trap_method = alloc->jvms()->method();
5241 trap_bci = alloc->jvms()->bci();
5242 }
5243
5244 bool negative_length_guard_generated = false;
5245
5246 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5247 can_emit_guards &&
5248 !src->is_top() && !dest->is_top()) {
5249 // validate arguments: enables transformation the ArrayCopyNode
5250 validated = true;
5251
5252 RegionNode* slow_region = new RegionNode(1);
5253 record_for_igvn(slow_region);
5254
5255 // (1) src and dest are arrays.
5256 generate_non_array_guard(load_object_klass(src), slow_region);
5257 generate_non_array_guard(load_object_klass(dest), slow_region);
5258
5259 // (2) src and dest arrays must have elements of the same BasicType
5260 // done at macro expansion or at Ideal transformation time
5261
5262 // (4) src_offset must not be negative.
5263 generate_negative_guard(src_offset, slow_region);
5264
5265 // (5) dest_offset must not be negative.
5266 generate_negative_guard(dest_offset, slow_region);
5267
5268 // (7) src_offset + length must not exceed length of src.
5269 generate_limit_guard(src_offset, length,
5270 load_array_length(src),
5271 slow_region);
5272
5273 // (8) dest_offset + length must not exceed length of dest.
5274 generate_limit_guard(dest_offset, length,
5275 load_array_length(dest),
5276 slow_region);
5277
5278 // (6) length must not be negative.
5279 // This is also checked in generate_arraycopy() during macro expansion, but
5280 // we also have to check it here for the case where the ArrayCopyNode will
5281 // be eliminated by Escape Analysis.
5282 if (EliminateAllocations) {
5283 generate_negative_guard(length, slow_region);
5284 negative_length_guard_generated = true;
5285 }
5286
5287 // (9) each element of an oop array must be assignable
5288 Node* src_klass = load_object_klass(src);
5289 Node* dest_klass = load_object_klass(dest);
5290 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5291
5292 if (not_subtype_ctrl != top()) {
5293 PreserveJVMState pjvms(this);
5294 set_control(not_subtype_ctrl);
5295 uncommon_trap(Deoptimization::Reason_intrinsic,
5296 Deoptimization::Action_make_not_entrant);
5297 assert(stopped(), "Should be stopped");
5298 }
5299 {
5300 PreserveJVMState pjvms(this);
5301 set_control(_gvn.transform(slow_region));
5302 uncommon_trap(Deoptimization::Reason_intrinsic,
5303 Deoptimization::Action_make_not_entrant);
5304 assert(stopped(), "Should be stopped");
5305 }
5306
5307 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5308 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5309 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5310 }
5311
5312 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5313
5314 if (stopped()) {
5315 return true;
5316 }
5317
5318 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5319 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5320 // so the compiler has a chance to eliminate them: during macro expansion,
5321 // we have to set their control (CastPP nodes are eliminated).
5322 load_object_klass(src), load_object_klass(dest),
5323 load_array_length(src), load_array_length(dest));
5324
5325 ac->set_arraycopy(validated);
5326
5327 Node* n = _gvn.transform(ac);
5328 if (n == ac) {
5329 ac->connect_outputs(this);
5330 } else {
5331 assert(validated, "shouldn't transform if all arguments not validated");
5332 set_all_memory(n);
5333 }
5334
5335 return true;
5336 }
5337
5338
5339 // Helper function which determines if an arraycopy immediately follows
5340 // an allocation, with no intervening tests or other escapes for the object.
5341 AllocateArrayNode*
5342 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5343 RegionNode* slow_region) {
5344 if (stopped()) return NULL; // no fast path
5345 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5346
5347 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5348 if (alloc == NULL) return NULL;
5349
5350 Node* rawmem = memory(Compile::AliasIdxRaw);
5351 // Is the allocation's memory state untouched?
5352 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5353 // Bail out if there have been raw-memory effects since the allocation.
5354 // (Example: There might have been a call or safepoint.)
5355 return NULL;
5356 }
5357 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5358 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5359 return NULL;
5360 }
5361
5362 // There must be no unexpected observers of this allocation.
5363 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5364 Node* obs = ptr->fast_out(i);
5365 if (obs != this->map()) {
5366 return NULL;
5367 }
5368 }
5369
5370 // This arraycopy must unconditionally follow the allocation of the ptr.
5371 Node* alloc_ctl = ptr->in(0);
5372 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5373
5374 Node* ctl = control();
5375 while (ctl != alloc_ctl) {
5376 // There may be guards which feed into the slow_region.
5377 // Any other control flow means that we might not get a chance
5378 // to finish initializing the allocated object.
5379 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5380 IfNode* iff = ctl->in(0)->as_If();
5381 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5382 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5383 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5384 ctl = iff->in(0); // This test feeds the known slow_region.
5385 continue;
5386 }
5387 // One more try: Various low-level checks bottom out in
5388 // uncommon traps. If the debug-info of the trap omits
5389 // any reference to the allocation, as we've already
5390 // observed, then there can be no objection to the trap.
5391 bool found_trap = false;
5392 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5393 Node* obs = not_ctl->fast_out(j);
5394 if (obs->in(0) == not_ctl && obs->is_Call() &&
5395 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5396 found_trap = true; break;
5397 }
5398 }
5399 if (found_trap) {
5400 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5401 continue;
5402 }
5403 }
5404 return NULL;
5405 }
5406
5407 // If we get this far, we have an allocation which immediately
5408 // precedes the arraycopy, and we can take over zeroing the new object.
5409 // The arraycopy will finish the initialization, and provide
5410 // a new control state to which we will anchor the destination pointer.
5411
5412 return alloc;
5413 }
5414
5415 //-------------inline_encodeISOArray-----------------------------------
5416 // encode char[] to byte[] in ISO_8859_1
5417 bool LibraryCallKit::inline_encodeISOArray() {
5418 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5419 // no receiver since it is static method
5420 Node *src = argument(0);
5421 Node *src_offset = argument(1);
5422 Node *dst = argument(2);
5423 Node *dst_offset = argument(3);
5424 Node *length = argument(4);
5425
5426 const Type* src_type = src->Value(&_gvn);
5427 const Type* dst_type = dst->Value(&_gvn);
5428 const TypeAryPtr* top_src = src_type->isa_aryptr();
5429 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5430 if (top_src == NULL || top_src->klass() == NULL ||
5431 top_dest == NULL || top_dest->klass() == NULL) {
5432 // failed array check
5433 return false;
5434 }
5435
5436 // Figure out the size and type of the elements we will be copying.
5437 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5438 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5439 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5440 return false;
5441 }
5442
5443 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5444 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5445 // 'src_start' points to src array + scaled offset
5446 // 'dst_start' points to dst array + scaled offset
5447
5448 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5449 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5450 enc = _gvn.transform(enc);
5451 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5452 set_memory(res_mem, mtype);
5453 set_result(enc);
5454 return true;
5455 }
5456
5457 //-------------inline_multiplyToLen-----------------------------------
5458 bool LibraryCallKit::inline_multiplyToLen() {
5459 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5460
5461 address stubAddr = StubRoutines::multiplyToLen();
5462 if (stubAddr == NULL) {
5463 return false; // Intrinsic's stub is not implemented on this platform
5464 }
5465 const char* stubName = "multiplyToLen";
5466
5467 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5468
5469 // no receiver because it is a static method
5470 Node* x = argument(0);
5471 Node* xlen = argument(1);
5472 Node* y = argument(2);
5473 Node* ylen = argument(3);
5474 Node* z = argument(4);
5475
5476 const Type* x_type = x->Value(&_gvn);
5477 const Type* y_type = y->Value(&_gvn);
5478 const TypeAryPtr* top_x = x_type->isa_aryptr();
5479 const TypeAryPtr* top_y = y_type->isa_aryptr();
5480 if (top_x == NULL || top_x->klass() == NULL ||
5481 top_y == NULL || top_y->klass() == NULL) {
5482 // failed array check
5483 return false;
5484 }
5485
5486 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5487 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5488 if (x_elem != T_INT || y_elem != T_INT) {
5489 return false;
5490 }
5491
5492 // Set the original stack and the reexecute bit for the interpreter to reexecute
5493 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5494 // on the return from z array allocation in runtime.
5495 { PreserveReexecuteState preexecs(this);
5496 jvms()->set_should_reexecute(true);
5497
5498 Node* x_start = array_element_address(x, intcon(0), x_elem);
5499 Node* y_start = array_element_address(y, intcon(0), y_elem);
5500 // 'x_start' points to x array + scaled xlen
5501 // 'y_start' points to y array + scaled ylen
5502
5503 // Allocate the result array
5504 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5505 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5506 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5507
5508 IdealKit ideal(this);
5509
5510 #define __ ideal.
5511 Node* one = __ ConI(1);
5512 Node* zero = __ ConI(0);
5513 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5514 __ set(need_alloc, zero);
5515 __ set(z_alloc, z);
5516 __ if_then(z, BoolTest::eq, null()); {
5517 __ increment (need_alloc, one);
5518 } __ else_(); {
5519 // Update graphKit memory and control from IdealKit.
5520 sync_kit(ideal);
5521 Node* zlen_arg = load_array_length(z);
5522 // Update IdealKit memory and control from graphKit.
5523 __ sync_kit(this);
5524 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5525 __ increment (need_alloc, one);
5526 } __ end_if();
5527 } __ end_if();
5528
5529 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5530 // Update graphKit memory and control from IdealKit.
5531 sync_kit(ideal);
5532 Node * narr = new_array(klass_node, zlen, 1);
5533 // Update IdealKit memory and control from graphKit.
5534 __ sync_kit(this);
5535 __ set(z_alloc, narr);
5536 } __ end_if();
5537
5538 sync_kit(ideal);
5539 z = __ value(z_alloc);
5540 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5541 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5542 // Final sync IdealKit and GraphKit.
5543 final_sync(ideal);
5544 #undef __
5545
5546 Node* z_start = array_element_address(z, intcon(0), T_INT);
5547
5548 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5549 OptoRuntime::multiplyToLen_Type(),
5550 stubAddr, stubName, TypePtr::BOTTOM,
5551 x_start, xlen, y_start, ylen, z_start, zlen);
5552 } // original reexecute is set back here
5553
5554 C->set_has_split_ifs(true); // Has chance for split-if optimization
5555 set_result(z);
5556 return true;
5557 }
5558
5559 //-------------inline_squareToLen------------------------------------
5560 bool LibraryCallKit::inline_squareToLen() {
5561 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5562
5563 address stubAddr = StubRoutines::squareToLen();
5564 if (stubAddr == NULL) {
5565 return false; // Intrinsic's stub is not implemented on this platform
5566 }
5567 const char* stubName = "squareToLen";
5568
5569 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5570
5571 Node* x = argument(0);
5572 Node* len = argument(1);
5573 Node* z = argument(2);
5574 Node* zlen = argument(3);
5575
5576 const Type* x_type = x->Value(&_gvn);
5577 const Type* z_type = z->Value(&_gvn);
5578 const TypeAryPtr* top_x = x_type->isa_aryptr();
5579 const TypeAryPtr* top_z = z_type->isa_aryptr();
5580 if (top_x == NULL || top_x->klass() == NULL ||
5581 top_z == NULL || top_z->klass() == NULL) {
5582 // failed array check
5583 return false;
5584 }
5585
5586 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5587 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5588 if (x_elem != T_INT || z_elem != T_INT) {
5589 return false;
5590 }
5591
5592
5593 Node* x_start = array_element_address(x, intcon(0), x_elem);
5594 Node* z_start = array_element_address(z, intcon(0), z_elem);
5595
5596 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5597 OptoRuntime::squareToLen_Type(),
5598 stubAddr, stubName, TypePtr::BOTTOM,
5599 x_start, len, z_start, zlen);
5600
5601 set_result(z);
5602 return true;
5603 }
5604
5605 //-------------inline_mulAdd------------------------------------------
5606 bool LibraryCallKit::inline_mulAdd() {
5607 assert(UseMulAddIntrinsic, "not implemented on this platform");
5608
5609 address stubAddr = StubRoutines::mulAdd();
5610 if (stubAddr == NULL) {
5611 return false; // Intrinsic's stub is not implemented on this platform
5612 }
5613 const char* stubName = "mulAdd";
5614
5615 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5616
5617 Node* out = argument(0);
5618 Node* in = argument(1);
5619 Node* offset = argument(2);
5620 Node* len = argument(3);
5621 Node* k = argument(4);
5622
5623 const Type* out_type = out->Value(&_gvn);
5624 const Type* in_type = in->Value(&_gvn);
5625 const TypeAryPtr* top_out = out_type->isa_aryptr();
5626 const TypeAryPtr* top_in = in_type->isa_aryptr();
5627 if (top_out == NULL || top_out->klass() == NULL ||
5628 top_in == NULL || top_in->klass() == NULL) {
5629 // failed array check
5630 return false;
5631 }
5632
5633 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5634 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5635 if (out_elem != T_INT || in_elem != T_INT) {
5636 return false;
5637 }
5638
5639 Node* outlen = load_array_length(out);
5640 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5641 Node* out_start = array_element_address(out, intcon(0), out_elem);
5642 Node* in_start = array_element_address(in, intcon(0), in_elem);
5643
5644 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5645 OptoRuntime::mulAdd_Type(),
5646 stubAddr, stubName, TypePtr::BOTTOM,
5647 out_start,in_start, new_offset, len, k);
5648 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5649 set_result(result);
5650 return true;
5651 }
5652
5653 //-------------inline_montgomeryMultiply-----------------------------------
5654 bool LibraryCallKit::inline_montgomeryMultiply() {
5655 address stubAddr = StubRoutines::montgomeryMultiply();
5656 if (stubAddr == NULL) {
5657 return false; // Intrinsic's stub is not implemented on this platform
5658 }
5659
5660 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5661 const char* stubName = "montgomery_multiply";
5662
5663 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5664
5665 Node* a = argument(0);
5666 Node* b = argument(1);
5667 Node* n = argument(2);
5668 Node* len = argument(3);
5669 Node* inv = argument(4);
5670 Node* m = argument(6);
5671
5672 const Type* a_type = a->Value(&_gvn);
5673 const TypeAryPtr* top_a = a_type->isa_aryptr();
5674 const Type* b_type = b->Value(&_gvn);
5675 const TypeAryPtr* top_b = b_type->isa_aryptr();
5676 const Type* n_type = a->Value(&_gvn);
5677 const TypeAryPtr* top_n = n_type->isa_aryptr();
5678 const Type* m_type = a->Value(&_gvn);
5679 const TypeAryPtr* top_m = m_type->isa_aryptr();
5680 if (top_a == NULL || top_a->klass() == NULL ||
5681 top_b == NULL || top_b->klass() == NULL ||
5682 top_n == NULL || top_n->klass() == NULL ||
5683 top_m == NULL || top_m->klass() == NULL) {
5684 // failed array check
5685 return false;
5686 }
5687
5688 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5689 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5690 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5691 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5692 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5693 return false;
5694 }
5695
5696 // Make the call
5697 {
5698 Node* a_start = array_element_address(a, intcon(0), a_elem);
5699 Node* b_start = array_element_address(b, intcon(0), b_elem);
5700 Node* n_start = array_element_address(n, intcon(0), n_elem);
5701 Node* m_start = array_element_address(m, intcon(0), m_elem);
5702
5703 Node* call = make_runtime_call(RC_LEAF,
5704 OptoRuntime::montgomeryMultiply_Type(),
5705 stubAddr, stubName, TypePtr::BOTTOM,
5706 a_start, b_start, n_start, len, inv, top(),
5707 m_start);
5708 set_result(m);
5709 }
5710
5711 return true;
5712 }
5713
5714 bool LibraryCallKit::inline_montgomerySquare() {
5715 address stubAddr = StubRoutines::montgomerySquare();
5716 if (stubAddr == NULL) {
5717 return false; // Intrinsic's stub is not implemented on this platform
5718 }
5719
5720 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5721 const char* stubName = "montgomery_square";
5722
5723 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5724
5725 Node* a = argument(0);
5726 Node* n = argument(1);
5727 Node* len = argument(2);
5728 Node* inv = argument(3);
5729 Node* m = argument(5);
5730
5731 const Type* a_type = a->Value(&_gvn);
5732 const TypeAryPtr* top_a = a_type->isa_aryptr();
5733 const Type* n_type = a->Value(&_gvn);
5734 const TypeAryPtr* top_n = n_type->isa_aryptr();
5735 const Type* m_type = a->Value(&_gvn);
5736 const TypeAryPtr* top_m = m_type->isa_aryptr();
5737 if (top_a == NULL || top_a->klass() == NULL ||
5738 top_n == NULL || top_n->klass() == NULL ||
5739 top_m == NULL || top_m->klass() == NULL) {
5740 // failed array check
5741 return false;
5742 }
5743
5744 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5745 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5746 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5747 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5748 return false;
5749 }
5750
5751 // Make the call
5752 {
5753 Node* a_start = array_element_address(a, intcon(0), a_elem);
5754 Node* n_start = array_element_address(n, intcon(0), n_elem);
5755 Node* m_start = array_element_address(m, intcon(0), m_elem);
5756
5757 Node* call = make_runtime_call(RC_LEAF,
5758 OptoRuntime::montgomerySquare_Type(),
5759 stubAddr, stubName, TypePtr::BOTTOM,
5760 a_start, n_start, len, inv, top(),
5761 m_start);
5762 set_result(m);
5763 }
5764
5765 return true;
5766 }
5767
5768 //-------------inline_vectorizedMismatch------------------------------
5769 bool LibraryCallKit::inline_vectorizedMismatch() {
5770 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5771
5772 address stubAddr = StubRoutines::vectorizedMismatch();
5773 if (stubAddr == NULL) {
5774 return false; // Intrinsic's stub is not implemented on this platform
5775 }
5776 const char* stubName = "vectorizedMismatch";
5777 int size_l = callee()->signature()->size();
5778 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5779
5780 Node* obja = argument(0);
5781 Node* aoffset = argument(1);
5782 Node* objb = argument(3);
5783 Node* boffset = argument(4);
5784 Node* length = argument(6);
5785 Node* scale = argument(7);
5786
5787 const Type* a_type = obja->Value(&_gvn);
5788 const Type* b_type = objb->Value(&_gvn);
5789 const TypeAryPtr* top_a = a_type->isa_aryptr();
5790 const TypeAryPtr* top_b = b_type->isa_aryptr();
5791 if (top_a == NULL || top_a->klass() == NULL ||
5792 top_b == NULL || top_b->klass() == NULL) {
5793 // failed array check
5794 return false;
5795 }
5796
5797 Node* call;
5798 jvms()->set_should_reexecute(true);
5799
5800 Node* obja_adr = make_unsafe_address(obja, aoffset);
5801 Node* objb_adr = make_unsafe_address(objb, boffset);
5802
5803 call = make_runtime_call(RC_LEAF,
5804 OptoRuntime::vectorizedMismatch_Type(),
5805 stubAddr, stubName, TypePtr::BOTTOM,
5806 obja_adr, objb_adr, length, scale);
5807
5808 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5809 set_result(result);
5810 return true;
5811 }
5812
5813 /**
5814 * Calculate CRC32 for byte.
5815 * int java.util.zip.CRC32.update(int crc, int b)
5816 */
5817 bool LibraryCallKit::inline_updateCRC32() {
5818 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5819 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5820 // no receiver since it is static method
5821 Node* crc = argument(0); // type: int
5822 Node* b = argument(1); // type: int
5823
5824 /*
5825 * int c = ~ crc;
5826 * b = timesXtoThe32[(b ^ c) & 0xFF];
5827 * b = b ^ (c >>> 8);
5828 * crc = ~b;
5829 */
5830
5831 Node* M1 = intcon(-1);
5832 crc = _gvn.transform(new XorINode(crc, M1));
5833 Node* result = _gvn.transform(new XorINode(crc, b));
5834 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5835
5836 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5837 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5838 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5839 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5840
5841 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5842 result = _gvn.transform(new XorINode(crc, result));
5843 result = _gvn.transform(new XorINode(result, M1));
5844 set_result(result);
5845 return true;
5846 }
5847
5848 /**
5849 * Calculate CRC32 for byte[] array.
5850 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5851 */
5852 bool LibraryCallKit::inline_updateBytesCRC32() {
5853 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5854 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5855 // no receiver since it is static method
5856 Node* crc = argument(0); // type: int
5857 Node* src = argument(1); // type: oop
5858 Node* offset = argument(2); // type: int
5859 Node* length = argument(3); // type: int
5860
5861 const Type* src_type = src->Value(&_gvn);
5862 const TypeAryPtr* top_src = src_type->isa_aryptr();
5863 if (top_src == NULL || top_src->klass() == NULL) {
5864 // failed array check
5865 return false;
5866 }
5867
5868 // Figure out the size and type of the elements we will be copying.
5869 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5870 if (src_elem != T_BYTE) {
5871 return false;
5872 }
5873
5874 // 'src_start' points to src array + scaled offset
5875 Node* src_start = array_element_address(src, offset, src_elem);
5876
5877 // We assume that range check is done by caller.
5878 // TODO: generate range check (offset+length < src.length) in debug VM.
5879
5880 // Call the stub.
5881 address stubAddr = StubRoutines::updateBytesCRC32();
5882 const char *stubName = "updateBytesCRC32";
5883
5884 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5885 stubAddr, stubName, TypePtr::BOTTOM,
5886 crc, src_start, length);
5887 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5888 set_result(result);
5889 return true;
5890 }
5891
5892 /**
5893 * Calculate CRC32 for ByteBuffer.
5894 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5895 */
5896 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5897 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5898 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5899 // no receiver since it is static method
5900 Node* crc = argument(0); // type: int
5901 Node* src = argument(1); // type: long
5902 Node* offset = argument(3); // type: int
5903 Node* length = argument(4); // type: int
5904
5905 src = ConvL2X(src); // adjust Java long to machine word
5906 Node* base = _gvn.transform(new CastX2PNode(src));
5907 offset = ConvI2X(offset);
5908
5909 // 'src_start' points to src array + scaled offset
5910 Node* src_start = basic_plus_adr(top(), base, offset);
5911
5912 // Call the stub.
5913 address stubAddr = StubRoutines::updateBytesCRC32();
5914 const char *stubName = "updateBytesCRC32";
5915
5916 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5917 stubAddr, stubName, TypePtr::BOTTOM,
5918 crc, src_start, length);
5919 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5920 set_result(result);
5921 return true;
5922 }
5923
5924 //------------------------------get_table_from_crc32c_class-----------------------
5925 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5926 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5927 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5928
5929 return table;
5930 }
5931
5932 //------------------------------inline_updateBytesCRC32C-----------------------
5933 //
5934 // Calculate CRC32C for byte[] array.
5935 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5936 //
5937 bool LibraryCallKit::inline_updateBytesCRC32C() {
5938 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5939 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5940 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5941 // no receiver since it is a static method
5942 Node* crc = argument(0); // type: int
5943 Node* src = argument(1); // type: oop
5944 Node* offset = argument(2); // type: int
5945 Node* end = argument(3); // type: int
5946
5947 Node* length = _gvn.transform(new SubINode(end, offset));
5948
5949 const Type* src_type = src->Value(&_gvn);
5950 const TypeAryPtr* top_src = src_type->isa_aryptr();
5951 if (top_src == NULL || top_src->klass() == NULL) {
5952 // failed array check
5953 return false;
5954 }
5955
5956 // Figure out the size and type of the elements we will be copying.
5957 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5958 if (src_elem != T_BYTE) {
5959 return false;
5960 }
5961
5962 // 'src_start' points to src array + scaled offset
5963 Node* src_start = array_element_address(src, offset, src_elem);
5964
5965 // static final int[] byteTable in class CRC32C
5966 Node* table = get_table_from_crc32c_class(callee()->holder());
5967 Node* table_start = array_element_address(table, intcon(0), T_INT);
5968
5969 // We assume that range check is done by caller.
5970 // TODO: generate range check (offset+length < src.length) in debug VM.
5971
5972 // Call the stub.
5973 address stubAddr = StubRoutines::updateBytesCRC32C();
5974 const char *stubName = "updateBytesCRC32C";
5975
5976 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5977 stubAddr, stubName, TypePtr::BOTTOM,
5978 crc, src_start, length, table_start);
5979 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5980 set_result(result);
5981 return true;
5982 }
5983
5984 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5985 //
5986 // Calculate CRC32C for DirectByteBuffer.
5987 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5988 //
5989 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5990 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5991 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5992 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5993 // no receiver since it is a static method
5994 Node* crc = argument(0); // type: int
5995 Node* src = argument(1); // type: long
5996 Node* offset = argument(3); // type: int
5997 Node* end = argument(4); // type: int
5998
5999 Node* length = _gvn.transform(new SubINode(end, offset));
6000
6001 src = ConvL2X(src); // adjust Java long to machine word
6002 Node* base = _gvn.transform(new CastX2PNode(src));
6003 offset = ConvI2X(offset);
6004
6005 // 'src_start' points to src array + scaled offset
6006 Node* src_start = basic_plus_adr(top(), base, offset);
6007
6008 // static final int[] byteTable in class CRC32C
6009 Node* table = get_table_from_crc32c_class(callee()->holder());
6010 Node* table_start = array_element_address(table, intcon(0), T_INT);
6011
6012 // Call the stub.
6013 address stubAddr = StubRoutines::updateBytesCRC32C();
6014 const char *stubName = "updateBytesCRC32C";
6015
6016 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6017 stubAddr, stubName, TypePtr::BOTTOM,
6018 crc, src_start, length, table_start);
6019 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6020 set_result(result);
6021 return true;
6022 }
6023
6024 //------------------------------inline_updateBytesAdler32----------------------
6025 //
6026 // Calculate Adler32 checksum for byte[] array.
6027 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6028 //
6029 bool LibraryCallKit::inline_updateBytesAdler32() {
6030 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6031 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6032 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6033 // no receiver since it is static method
6034 Node* crc = argument(0); // type: int
6035 Node* src = argument(1); // type: oop
6036 Node* offset = argument(2); // type: int
6037 Node* length = argument(3); // type: int
6038
6039 const Type* src_type = src->Value(&_gvn);
6040 const TypeAryPtr* top_src = src_type->isa_aryptr();
6041 if (top_src == NULL || top_src->klass() == NULL) {
6042 // failed array check
6043 return false;
6044 }
6045
6046 // Figure out the size and type of the elements we will be copying.
6047 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6048 if (src_elem != T_BYTE) {
6049 return false;
6050 }
6051
6052 // 'src_start' points to src array + scaled offset
6053 Node* src_start = array_element_address(src, offset, src_elem);
6054
6055 // We assume that range check is done by caller.
6056 // TODO: generate range check (offset+length < src.length) in debug VM.
6057
6058 // Call the stub.
6059 address stubAddr = StubRoutines::updateBytesAdler32();
6060 const char *stubName = "updateBytesAdler32";
6061
6062 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6063 stubAddr, stubName, TypePtr::BOTTOM,
6064 crc, src_start, length);
6065 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6066 set_result(result);
6067 return true;
6068 }
6069
6070 //------------------------------inline_updateByteBufferAdler32---------------
6071 //
6072 // Calculate Adler32 checksum for DirectByteBuffer.
6073 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6074 //
6075 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6076 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6077 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6078 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6079 // no receiver since it is static method
6080 Node* crc = argument(0); // type: int
6081 Node* src = argument(1); // type: long
6082 Node* offset = argument(3); // type: int
6083 Node* length = argument(4); // type: int
6084
6085 src = ConvL2X(src); // adjust Java long to machine word
6086 Node* base = _gvn.transform(new CastX2PNode(src));
6087 offset = ConvI2X(offset);
6088
6089 // 'src_start' points to src array + scaled offset
6090 Node* src_start = basic_plus_adr(top(), base, offset);
6091
6092 // Call the stub.
6093 address stubAddr = StubRoutines::updateBytesAdler32();
6094 const char *stubName = "updateBytesAdler32";
6095
6096 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6097 stubAddr, stubName, TypePtr::BOTTOM,
6098 crc, src_start, length);
6099
6100 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6101 set_result(result);
6102 return true;
6103 }
6104
6105 //----------------------------inline_reference_get----------------------------
6106 // public T java.lang.ref.Reference.get();
6107 bool LibraryCallKit::inline_reference_get() {
6108 const int referent_offset = java_lang_ref_Reference::referent_offset;
6109 guarantee(referent_offset > 0, "should have already been set");
6110
6111 // Get the argument:
6112 Node* reference_obj = null_check_receiver();
6113 if (stopped()) return true;
6114
6115 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6116
6117 ciInstanceKlass* klass = env()->Object_klass();
6118 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6119
6120 Node* no_ctrl = NULL;
6121 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6122
6123 // Use the pre-barrier to record the value in the referent field
6124 pre_barrier(false /* do_load */,
6125 control(),
6126 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6127 result /* pre_val */,
6128 T_OBJECT);
6129
6130 // Add memory barrier to prevent commoning reads from this field
6131 // across safepoint since GC can change its value.
6132 insert_mem_bar(Op_MemBarCPUOrder);
6133
6134 set_result(result);
6135 return true;
6136 }
6137
6138
6139 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6140 bool is_exact=true, bool is_static=false,
6141 ciInstanceKlass * fromKls=NULL) {
6142 if (fromKls == NULL) {
6143 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6144 assert(tinst != NULL, "obj is null");
6145 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6146 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6147 fromKls = tinst->klass()->as_instance_klass();
6148 } else {
6149 assert(is_static, "only for static field access");
6150 }
6151 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6152 ciSymbol::make(fieldTypeString),
6153 is_static);
6154
6155 assert (field != NULL, "undefined field");
6156 if (field == NULL) return (Node *) NULL;
6157
6158 if (is_static) {
6159 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6160 fromObj = makecon(tip);
6161 }
6162
6163 // Next code copied from Parse::do_get_xxx():
6164
6165 // Compute address and memory type.
6166 int offset = field->offset_in_bytes();
6167 bool is_vol = field->is_volatile();
6168 ciType* field_klass = field->type();
6169 assert(field_klass->is_loaded(), "should be loaded");
6170 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6171 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6172 BasicType bt = field->layout_type();
6173
6174 // Build the resultant type of the load
6175 const Type *type;
6176 if (bt == T_OBJECT) {
6177 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6178 } else {
6179 type = Type::get_const_basic_type(bt);
6180 }
6181
6182 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6183 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6184 }
6185 // Build the load.
6186 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6187 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6188 // If reference is volatile, prevent following memory ops from
6189 // floating up past the volatile read. Also prevents commoning
6190 // another volatile read.
6191 if (is_vol) {
6192 // Memory barrier includes bogus read of value to force load BEFORE membar
6193 insert_mem_bar(Op_MemBarAcquire, loadedField);
6194 }
6195 return loadedField;
6196 }
6197
6198 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6199 bool is_exact = true, bool is_static = false,
6200 ciInstanceKlass * fromKls = NULL) {
6201 if (fromKls == NULL) {
6202 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6203 assert(tinst != NULL, "obj is null");
6204 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6205 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6206 fromKls = tinst->klass()->as_instance_klass();
6207 }
6208 else {
6209 assert(is_static, "only for static field access");
6210 }
6211 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6212 ciSymbol::make(fieldTypeString),
6213 is_static);
6214
6215 assert(field != NULL, "undefined field");
6216 assert(!field->is_volatile(), "not defined for volatile fields");
6217
6218 if (is_static) {
6219 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6220 fromObj = makecon(tip);
6221 }
6222
6223 // Next code copied from Parse::do_get_xxx():
6224
6225 // Compute address and memory type.
6226 int offset = field->offset_in_bytes();
6227 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6228
6229 return adr;
6230 }
6231
6232 //------------------------------inline_aescrypt_Block-----------------------
6233 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6234 address stubAddr = NULL;
6235 const char *stubName;
6236 assert(UseAES, "need AES instruction support");
6237
6238 switch(id) {
6239 case vmIntrinsics::_aescrypt_encryptBlock:
6240 stubAddr = StubRoutines::aescrypt_encryptBlock();
6241 stubName = "aescrypt_encryptBlock";
6242 break;
6243 case vmIntrinsics::_aescrypt_decryptBlock:
6244 stubAddr = StubRoutines::aescrypt_decryptBlock();
6245 stubName = "aescrypt_decryptBlock";
6246 break;
6247 default:
6248 break;
6249 }
6250 if (stubAddr == NULL) return false;
6251
6252 Node* aescrypt_object = argument(0);
6253 Node* src = argument(1);
6254 Node* src_offset = argument(2);
6255 Node* dest = argument(3);
6256 Node* dest_offset = argument(4);
6257
6258 // (1) src and dest are arrays.
6259 const Type* src_type = src->Value(&_gvn);
6260 const Type* dest_type = dest->Value(&_gvn);
6261 const TypeAryPtr* top_src = src_type->isa_aryptr();
6262 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6263 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6264
6265 // for the quick and dirty code we will skip all the checks.
6266 // we are just trying to get the call to be generated.
6267 Node* src_start = src;
6268 Node* dest_start = dest;
6269 if (src_offset != NULL || dest_offset != NULL) {
6270 assert(src_offset != NULL && dest_offset != NULL, "");
6271 src_start = array_element_address(src, src_offset, T_BYTE);
6272 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6273 }
6274
6275 // now need to get the start of its expanded key array
6276 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6277 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6278 if (k_start == NULL) return false;
6279
6280 if (Matcher::pass_original_key_for_aes()) {
6281 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6282 // compatibility issues between Java key expansion and SPARC crypto instructions
6283 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6284 if (original_k_start == NULL) return false;
6285
6286 // Call the stub.
6287 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6288 stubAddr, stubName, TypePtr::BOTTOM,
6289 src_start, dest_start, k_start, original_k_start);
6290 } else {
6291 // Call the stub.
6292 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6293 stubAddr, stubName, TypePtr::BOTTOM,
6294 src_start, dest_start, k_start);
6295 }
6296
6297 return true;
6298 }
6299
6300 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6301 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6302 address stubAddr = NULL;
6303 const char *stubName = NULL;
6304
6305 assert(UseAES, "need AES instruction support");
6306
6307 switch(id) {
6308 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6309 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6310 stubName = "cipherBlockChaining_encryptAESCrypt";
6311 break;
6312 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6313 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6314 stubName = "cipherBlockChaining_decryptAESCrypt";
6315 break;
6316 default:
6317 break;
6318 }
6319 if (stubAddr == NULL) return false;
6320
6321 Node* cipherBlockChaining_object = argument(0);
6322 Node* src = argument(1);
6323 Node* src_offset = argument(2);
6324 Node* len = argument(3);
6325 Node* dest = argument(4);
6326 Node* dest_offset = argument(5);
6327
6328 // (1) src and dest are arrays.
6329 const Type* src_type = src->Value(&_gvn);
6330 const Type* dest_type = dest->Value(&_gvn);
6331 const TypeAryPtr* top_src = src_type->isa_aryptr();
6332 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6333 assert (top_src != NULL && top_src->klass() != NULL
6334 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6335
6336 // checks are the responsibility of the caller
6337 Node* src_start = src;
6338 Node* dest_start = dest;
6339 if (src_offset != NULL || dest_offset != NULL) {
6340 assert(src_offset != NULL && dest_offset != NULL, "");
6341 src_start = array_element_address(src, src_offset, T_BYTE);
6342 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6343 }
6344
6345 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6346 // (because of the predicated logic executed earlier).
6347 // so we cast it here safely.
6348 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6349
6350 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6351 if (embeddedCipherObj == NULL) return false;
6352
6353 // cast it to what we know it will be at runtime
6354 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6355 assert(tinst != NULL, "CBC obj is null");
6356 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6357 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6358 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6359
6360 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6361 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6362 const TypeOopPtr* xtype = aklass->as_instance_type();
6363 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6364 aescrypt_object = _gvn.transform(aescrypt_object);
6365
6366 // we need to get the start of the aescrypt_object's expanded key array
6367 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6368 if (k_start == NULL) return false;
6369
6370 // similarly, get the start address of the r vector
6371 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6372 if (objRvec == NULL) return false;
6373 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6374
6375 Node* cbcCrypt;
6376 if (Matcher::pass_original_key_for_aes()) {
6377 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6378 // compatibility issues between Java key expansion and SPARC crypto instructions
6379 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6380 if (original_k_start == NULL) return false;
6381
6382 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6383 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6384 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6385 stubAddr, stubName, TypePtr::BOTTOM,
6386 src_start, dest_start, k_start, r_start, len, original_k_start);
6387 } else {
6388 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6389 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6390 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6391 stubAddr, stubName, TypePtr::BOTTOM,
6392 src_start, dest_start, k_start, r_start, len);
6393 }
6394
6395 // return cipher length (int)
6396 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6397 set_result(retvalue);
6398 return true;
6399 }
6400
6401 //------------------------------inline_counterMode_AESCrypt-----------------------
6402 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6403 assert(UseAES, "need AES instruction support");
6404 if (!UseAESCTRIntrinsics) return false;
6405
6406 address stubAddr = NULL;
6407 const char *stubName = NULL;
6408 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6409 stubAddr = StubRoutines::counterMode_AESCrypt();
6410 stubName = "counterMode_AESCrypt";
6411 }
6412 if (stubAddr == NULL) return false;
6413
6414 Node* counterMode_object = argument(0);
6415 Node* src = argument(1);
6416 Node* src_offset = argument(2);
6417 Node* len = argument(3);
6418 Node* dest = argument(4);
6419 Node* dest_offset = argument(5);
6420
6421 // (1) src and dest are arrays.
6422 const Type* src_type = src->Value(&_gvn);
6423 const Type* dest_type = dest->Value(&_gvn);
6424 const TypeAryPtr* top_src = src_type->isa_aryptr();
6425 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6426 assert(top_src != NULL && top_src->klass() != NULL &&
6427 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6428
6429 // checks are the responsibility of the caller
6430 Node* src_start = src;
6431 Node* dest_start = dest;
6432 if (src_offset != NULL || dest_offset != NULL) {
6433 assert(src_offset != NULL && dest_offset != NULL, "");
6434 src_start = array_element_address(src, src_offset, T_BYTE);
6435 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6436 }
6437
6438 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6439 // (because of the predicated logic executed earlier).
6440 // so we cast it here safely.
6441 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6442 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6443 if (embeddedCipherObj == NULL) return false;
6444 // cast it to what we know it will be at runtime
6445 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6446 assert(tinst != NULL, "CTR obj is null");
6447 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6448 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6449 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6450 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6451 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6452 const TypeOopPtr* xtype = aklass->as_instance_type();
6453 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6454 aescrypt_object = _gvn.transform(aescrypt_object);
6455 // we need to get the start of the aescrypt_object's expanded key array
6456 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6457 if (k_start == NULL) return false;
6458 // similarly, get the start address of the r vector
6459 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6460 if (obj_counter == NULL) return false;
6461 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6462
6463 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6464 if (saved_encCounter == NULL) return false;
6465 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6466 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6467
6468 Node* ctrCrypt;
6469 if (Matcher::pass_original_key_for_aes()) {
6470 // no SPARC version for AES/CTR intrinsics now.
6471 return false;
6472 }
6473 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6474 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6475 OptoRuntime::counterMode_aescrypt_Type(),
6476 stubAddr, stubName, TypePtr::BOTTOM,
6477 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6478
6479 // return cipher length (int)
6480 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6481 set_result(retvalue);
6482 return true;
6483 }
6484
6485 //------------------------------get_key_start_from_aescrypt_object-----------------------
6486 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6487 #if defined(PPC64) || defined(S390)
6488 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6489 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6490 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6491 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6492 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6493 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6494 if (objSessionK == NULL) {
6495 return (Node *) NULL;
6496 }
6497 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6498 #else
6499 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6500 #endif // PPC64
6501 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6502 if (objAESCryptKey == NULL) return (Node *) NULL;
6503
6504 // now have the array, need to get the start address of the K array
6505 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6506 return k_start;
6507 }
6508
6509 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6510 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6511 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6512 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6513 if (objAESCryptKey == NULL) return (Node *) NULL;
6514
6515 // now have the array, need to get the start address of the lastKey array
6516 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6517 return original_k_start;
6518 }
6519
6520 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6521 // Return node representing slow path of predicate check.
6522 // the pseudo code we want to emulate with this predicate is:
6523 // for encryption:
6524 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6525 // for decryption:
6526 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6527 // note cipher==plain is more conservative than the original java code but that's OK
6528 //
6529 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6530 // The receiver was checked for NULL already.
6531 Node* objCBC = argument(0);
6532
6533 // Load embeddedCipher field of CipherBlockChaining object.
6534 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6535
6536 // get AESCrypt klass for instanceOf check
6537 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6538 // will have same classloader as CipherBlockChaining object
6539 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6540 assert(tinst != NULL, "CBCobj is null");
6541 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6542
6543 // we want to do an instanceof comparison against the AESCrypt class
6544 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6545 if (!klass_AESCrypt->is_loaded()) {
6546 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6547 Node* ctrl = control();
6548 set_control(top()); // no regular fast path
6549 return ctrl;
6550 }
6551 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6552
6553 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6554 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6555 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6556
6557 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6558
6559 // for encryption, we are done
6560 if (!decrypting)
6561 return instof_false; // even if it is NULL
6562
6563 // for decryption, we need to add a further check to avoid
6564 // taking the intrinsic path when cipher and plain are the same
6565 // see the original java code for why.
6566 RegionNode* region = new RegionNode(3);
6567 region->init_req(1, instof_false);
6568 Node* src = argument(1);
6569 Node* dest = argument(4);
6570 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6571 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6572 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6573 region->init_req(2, src_dest_conjoint);
6574
6575 record_for_igvn(region);
6576 return _gvn.transform(region);
6577 }
6578
6579 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6580 // Return node representing slow path of predicate check.
6581 // the pseudo code we want to emulate with this predicate is:
6582 // for encryption:
6583 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6584 // for decryption:
6585 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6586 // note cipher==plain is more conservative than the original java code but that's OK
6587 //
6588
6589 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6590 // The receiver was checked for NULL already.
6591 Node* objCTR = argument(0);
6592
6593 // Load embeddedCipher field of CipherBlockChaining object.
6594 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6595
6596 // get AESCrypt klass for instanceOf check
6597 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6598 // will have same classloader as CipherBlockChaining object
6599 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6600 assert(tinst != NULL, "CTRobj is null");
6601 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6602
6603 // we want to do an instanceof comparison against the AESCrypt class
6604 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6605 if (!klass_AESCrypt->is_loaded()) {
6606 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6607 Node* ctrl = control();
6608 set_control(top()); // no regular fast path
6609 return ctrl;
6610 }
6611
6612 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6613 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6614 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6615 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6616 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6617
6618 return instof_false; // even if it is NULL
6619 }
6620
6621 //------------------------------inline_ghash_processBlocks
6622 bool LibraryCallKit::inline_ghash_processBlocks() {
6623 address stubAddr;
6624 const char *stubName;
6625 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6626
6627 stubAddr = StubRoutines::ghash_processBlocks();
6628 stubName = "ghash_processBlocks";
6629
6630 Node* data = argument(0);
6631 Node* offset = argument(1);
6632 Node* len = argument(2);
6633 Node* state = argument(3);
6634 Node* subkeyH = argument(4);
6635
6636 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6637 assert(state_start, "state is NULL");
6638 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6639 assert(subkeyH_start, "subkeyH is NULL");
6640 Node* data_start = array_element_address(data, offset, T_BYTE);
6641 assert(data_start, "data is NULL");
6642
6643 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6644 OptoRuntime::ghash_processBlocks_Type(),
6645 stubAddr, stubName, TypePtr::BOTTOM,
6646 state_start, subkeyH_start, data_start, len);
6647 return true;
6648 }
6649
6650 //------------------------------inline_sha_implCompress-----------------------
6651 //
6652 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6653 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6654 //
6655 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6656 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6657 //
6658 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6659 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6660 //
6661 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6662 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6663
6664 Node* sha_obj = argument(0);
6665 Node* src = argument(1); // type oop
6666 Node* ofs = argument(2); // type int
6667
6668 const Type* src_type = src->Value(&_gvn);
6669 const TypeAryPtr* top_src = src_type->isa_aryptr();
6670 if (top_src == NULL || top_src->klass() == NULL) {
6671 // failed array check
6672 return false;
6673 }
6674 // Figure out the size and type of the elements we will be copying.
6675 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6676 if (src_elem != T_BYTE) {
6677 return false;
6678 }
6679 // 'src_start' points to src array + offset
6680 Node* src_start = array_element_address(src, ofs, src_elem);
6681 Node* state = NULL;
6682 address stubAddr;
6683 const char *stubName;
6684
6685 switch(id) {
6686 case vmIntrinsics::_sha_implCompress:
6687 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6688 state = get_state_from_sha_object(sha_obj);
6689 stubAddr = StubRoutines::sha1_implCompress();
6690 stubName = "sha1_implCompress";
6691 break;
6692 case vmIntrinsics::_sha2_implCompress:
6693 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6694 state = get_state_from_sha_object(sha_obj);
6695 stubAddr = StubRoutines::sha256_implCompress();
6696 stubName = "sha256_implCompress";
6697 break;
6698 case vmIntrinsics::_sha5_implCompress:
6699 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6700 state = get_state_from_sha5_object(sha_obj);
6701 stubAddr = StubRoutines::sha512_implCompress();
6702 stubName = "sha512_implCompress";
6703 break;
6704 default:
6705 fatal_unexpected_iid(id);
6706 return false;
6707 }
6708 if (state == NULL) return false;
6709
6710 // Call the stub.
6711 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6712 stubAddr, stubName, TypePtr::BOTTOM,
6713 src_start, state);
6714
6715 return true;
6716 }
6717
6718 //------------------------------inline_digestBase_implCompressMB-----------------------
6719 //
6720 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6721 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6722 //
6723 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6724 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6725 "need SHA1/SHA256/SHA512 instruction support");
6726 assert((uint)predicate < 3, "sanity");
6727 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6728
6729 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6730 Node* src = argument(1); // byte[] array
6731 Node* ofs = argument(2); // type int
6732 Node* limit = argument(3); // type int
6733
6734 const Type* src_type = src->Value(&_gvn);
6735 const TypeAryPtr* top_src = src_type->isa_aryptr();
6736 if (top_src == NULL || top_src->klass() == NULL) {
6737 // failed array check
6738 return false;
6739 }
6740 // Figure out the size and type of the elements we will be copying.
6741 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6742 if (src_elem != T_BYTE) {
6743 return false;
6744 }
6745 // 'src_start' points to src array + offset
6746 Node* src_start = array_element_address(src, ofs, src_elem);
6747
6748 const char* klass_SHA_name = NULL;
6749 const char* stub_name = NULL;
6750 address stub_addr = NULL;
6751 bool long_state = false;
6752
6753 switch (predicate) {
6754 case 0:
6755 if (UseSHA1Intrinsics) {
6756 klass_SHA_name = "sun/security/provider/SHA";
6757 stub_name = "sha1_implCompressMB";
6758 stub_addr = StubRoutines::sha1_implCompressMB();
6759 }
6760 break;
6761 case 1:
6762 if (UseSHA256Intrinsics) {
6763 klass_SHA_name = "sun/security/provider/SHA2";
6764 stub_name = "sha256_implCompressMB";
6765 stub_addr = StubRoutines::sha256_implCompressMB();
6766 }
6767 break;
6768 case 2:
6769 if (UseSHA512Intrinsics) {
6770 klass_SHA_name = "sun/security/provider/SHA5";
6771 stub_name = "sha512_implCompressMB";
6772 stub_addr = StubRoutines::sha512_implCompressMB();
6773 long_state = true;
6774 }
6775 break;
6776 default:
6777 fatal("unknown SHA intrinsic predicate: %d", predicate);
6778 }
6779 if (klass_SHA_name != NULL) {
6780 // get DigestBase klass to lookup for SHA klass
6781 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6782 assert(tinst != NULL, "digestBase_obj is not instance???");
6783 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6784
6785 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6786 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6787 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6788 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6789 }
6790 return false;
6791 }
6792 //------------------------------inline_sha_implCompressMB-----------------------
6793 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6794 bool long_state, address stubAddr, const char *stubName,
6795 Node* src_start, Node* ofs, Node* limit) {
6796 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6797 const TypeOopPtr* xtype = aklass->as_instance_type();
6798 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6799 sha_obj = _gvn.transform(sha_obj);
6800
6801 Node* state;
6802 if (long_state) {
6803 state = get_state_from_sha5_object(sha_obj);
6804 } else {
6805 state = get_state_from_sha_object(sha_obj);
6806 }
6807 if (state == NULL) return false;
6808
6809 // Call the stub.
6810 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6811 OptoRuntime::digestBase_implCompressMB_Type(),
6812 stubAddr, stubName, TypePtr::BOTTOM,
6813 src_start, state, ofs, limit);
6814 // return ofs (int)
6815 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6816 set_result(result);
6817
6818 return true;
6819 }
6820
6821 //------------------------------get_state_from_sha_object-----------------------
6822 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6823 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6824 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6825 if (sha_state == NULL) return (Node *) NULL;
6826
6827 // now have the array, need to get the start address of the state array
6828 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6829 return state;
6830 }
6831
6832 //------------------------------get_state_from_sha5_object-----------------------
6833 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6834 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6835 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6836 if (sha_state == NULL) return (Node *) NULL;
6837
6838 // now have the array, need to get the start address of the state array
6839 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6840 return state;
6841 }
6842
6843 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6844 // Return node representing slow path of predicate check.
6845 // the pseudo code we want to emulate with this predicate is:
6846 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6847 //
6848 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6849 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6850 "need SHA1/SHA256/SHA512 instruction support");
6851 assert((uint)predicate < 3, "sanity");
6852
6853 // The receiver was checked for NULL already.
6854 Node* digestBaseObj = argument(0);
6855
6856 // get DigestBase klass for instanceOf check
6857 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6858 assert(tinst != NULL, "digestBaseObj is null");
6859 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6860
6861 const char* klass_SHA_name = NULL;
6862 switch (predicate) {
6863 case 0:
6864 if (UseSHA1Intrinsics) {
6865 // we want to do an instanceof comparison against the SHA class
6866 klass_SHA_name = "sun/security/provider/SHA";
6867 }
6868 break;
6869 case 1:
6870 if (UseSHA256Intrinsics) {
6871 // we want to do an instanceof comparison against the SHA2 class
6872 klass_SHA_name = "sun/security/provider/SHA2";
6873 }
6874 break;
6875 case 2:
6876 if (UseSHA512Intrinsics) {
6877 // we want to do an instanceof comparison against the SHA5 class
6878 klass_SHA_name = "sun/security/provider/SHA5";
6879 }
6880 break;
6881 default:
6882 fatal("unknown SHA intrinsic predicate: %d", predicate);
6883 }
6884
6885 ciKlass* klass_SHA = NULL;
6886 if (klass_SHA_name != NULL) {
6887 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6888 }
6889 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6890 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6891 Node* ctrl = control();
6892 set_control(top()); // no intrinsic path
6893 return ctrl;
6894 }
6895 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6896
6897 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6898 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6899 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6900 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6901
6902 return instof_false; // even if it is NULL
6903 }
6904
6905 //-------------inline_fma-----------------------------------
6906 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6907 Node *a = NULL;
6908 Node *b = NULL;
6909 Node *c = NULL;
6910 Node* result = NULL;
6911 switch (id) {
6912 case vmIntrinsics::_fmaD:
6913 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6914 // no receiver since it is static method
6915 a = round_double_node(argument(0));
6916 b = round_double_node(argument(2));
6917 c = round_double_node(argument(4));
6918 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6919 break;
6920 case vmIntrinsics::_fmaF:
6921 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6922 a = argument(0);
6923 b = argument(1);
6924 c = argument(2);
6925 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6926 break;
6927 default:
6928 fatal_unexpected_iid(id); break;
6929 }
6930 set_result(result);
6931 return true;
6932 }
6933
6934 bool LibraryCallKit::inline_profileBoolean() {
6935 Node* counts = argument(1);
6936 const TypeAryPtr* ary = NULL;
6937 ciArray* aobj = NULL;
6938 if (counts->is_Con()
6939 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6940 && (aobj = ary->const_oop()->as_array()) != NULL
6941 && (aobj->length() == 2)) {
6942 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6943 jint false_cnt = aobj->element_value(0).as_int();
6944 jint true_cnt = aobj->element_value(1).as_int();
6945
6946 if (C->log() != NULL) {
6947 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6948 false_cnt, true_cnt);
6949 }
6950
6951 if (false_cnt + true_cnt == 0) {
6952 // According to profile, never executed.
6953 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6954 Deoptimization::Action_reinterpret);
6955 return true;
6956 }
6957
6958 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6959 // is a number of each value occurrences.
6960 Node* result = argument(0);
6961 if (false_cnt == 0 || true_cnt == 0) {
6962 // According to profile, one value has been never seen.
6963 int expected_val = (false_cnt == 0) ? 1 : 0;
6964
6965 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6966 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6967
6968 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6969 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6970 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6971
6972 { // Slow path: uncommon trap for never seen value and then reexecute
6973 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6974 // the value has been seen at least once.
6975 PreserveJVMState pjvms(this);
6976 PreserveReexecuteState preexecs(this);
6977 jvms()->set_should_reexecute(true);
6978
6979 set_control(slow_path);
6980 set_i_o(i_o());
6981
6982 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6983 Deoptimization::Action_reinterpret);
6984 }
6985 // The guard for never seen value enables sharpening of the result and
6986 // returning a constant. It allows to eliminate branches on the same value
6987 // later on.
6988 set_control(fast_path);
6989 result = intcon(expected_val);
6990 }
6991 // Stop profiling.
6992 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6993 // By replacing method body with profile data (represented as ProfileBooleanNode
6994 // on IR level) we effectively disable profiling.
6995 // It enables full speed execution once optimized code is generated.
6996 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6997 C->record_for_igvn(profile);
6998 set_result(profile);
6999 return true;
7000 } else {
7001 // Continue profiling.
7002 // Profile data isn't available at the moment. So, execute method's bytecode version.
7003 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7004 // is compiled and counters aren't available since corresponding MethodHandle
7005 // isn't a compile-time constant.
7006 return false;
7007 }
7008 }
7009
7010 bool LibraryCallKit::inline_isCompileConstant() {
7011 Node* n = argument(0);
7012 set_result(n->is_Con() ? intcon(1) : intcon(0));
7013 return true;
7014 }