1 /*
2 * Copyright (c) 1999, 2016, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/shenandoahSupport.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "prims/unsafe.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #ifdef TRACE_HAVE_INTRINSICS
56 #include "trace/traceMacros.hpp"
57 #endif
58
59 class LibraryIntrinsic : public InlineCallGenerator {
60 // Extend the set of intrinsics known to the runtime:
61 public:
62 private:
63 bool _is_virtual;
64 bool _does_virtual_dispatch;
65 int8_t _predicates_count; // Intrinsic is predicated by several conditions
66 int8_t _last_predicate; // Last generated predicate
67 vmIntrinsics::ID _intrinsic_id;
68
69 public:
70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
71 : InlineCallGenerator(m),
72 _is_virtual(is_virtual),
73 _does_virtual_dispatch(does_virtual_dispatch),
74 _predicates_count((int8_t)predicates_count),
75 _last_predicate((int8_t)-1),
76 _intrinsic_id(id)
77 {
78 }
79 virtual bool is_intrinsic() const { return true; }
80 virtual bool is_virtual() const { return _is_virtual; }
81 virtual bool is_predicated() const { return _predicates_count > 0; }
82 virtual int predicates_count() const { return _predicates_count; }
83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
84 virtual JVMState* generate(JVMState* jvms);
85 virtual Node* generate_predicate(JVMState* jvms, int predicate);
86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
87 };
88
89
90 // Local helper class for LibraryIntrinsic:
91 class LibraryCallKit : public GraphKit {
92 private:
93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
94 Node* _result; // the result node, if any
95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
96
97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
98
99 public:
100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
101 : GraphKit(jvms),
102 _intrinsic(intrinsic),
103 _result(NULL)
104 {
105 // Check if this is a root compile. In that case we don't have a caller.
106 if (!jvms->has_method()) {
107 _reexecute_sp = sp();
108 } else {
109 // Find out how many arguments the interpreter needs when deoptimizing
110 // and save the stack pointer value so it can used by uncommon_trap.
111 // We find the argument count by looking at the declared signature.
112 bool ignored_will_link;
113 ciSignature* declared_signature = NULL;
114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
117 }
118 }
119
120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
121
122 ciMethod* caller() const { return jvms()->method(); }
123 int bci() const { return jvms()->bci(); }
124 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
126 ciMethod* callee() const { return _intrinsic->method(); }
127
128 bool try_to_inline(int predicate);
129 Node* try_to_predicate(int predicate);
130
131 void push_result() {
132 // Push the result onto the stack.
133 if (!stopped() && result() != NULL) {
134 BasicType bt = result()->bottom_type()->basic_type();
135 push_node(bt, result());
136 }
137 }
138
139 private:
140 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
142 }
143
144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
145 void set_result(RegionNode* region, PhiNode* value);
146 Node* result() { return _result; }
147
148 virtual int reexecute_sp() { return _reexecute_sp; }
149
150 // Helper functions to inline natives
151 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
152 Node* generate_slow_guard(Node* test, RegionNode* region);
153 Node* generate_fair_guard(Node* test, RegionNode* region);
154 Node* generate_negative_guard(Node* index, RegionNode* region,
155 // resulting CastII of index:
156 Node* *pos_index = NULL);
157 Node* generate_limit_guard(Node* offset, Node* subseq_length,
158 Node* array_length,
159 RegionNode* region);
160 void generate_string_range_check(Node* array, Node* offset,
161 Node* length, bool char_count);
162 Node* generate_current_thread(Node* &tls_output);
163 Node* load_mirror_from_klass(Node* klass);
164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path,
166 int offset);
167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
175 RegionNode* region, int null_path) {
176 int offset = java_lang_Class::array_klass_offset_in_bytes();
177 return load_klass_from_mirror_common(mirror, never_see_null,
178 region, null_path,
179 offset);
180 }
181 Node* generate_access_flags_guard(Node* kls,
182 int modifier_mask, int modifier_bits,
183 RegionNode* region);
184 Node* generate_interface_guard(Node* kls, RegionNode* region);
185 Node* generate_array_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, false, false);
187 }
188 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
189 return generate_array_guard_common(kls, region, false, true);
190 }
191 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
192 return generate_array_guard_common(kls, region, true, false);
193 }
194 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
195 return generate_array_guard_common(kls, region, true, true);
196 }
197 Node* generate_array_guard_common(Node* kls, RegionNode* region,
198 bool obj_array, bool not_array);
199 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
200 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
201 bool is_virtual = false, bool is_static = false);
202 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
203 return generate_method_call(method_id, false, true);
204 }
205 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
206 return generate_method_call(method_id, true, false);
207 }
208 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210
211 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
215 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
216 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_indexOfChar();
218 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_toBytesU();
220 bool inline_string_getCharsU();
221 bool inline_string_copy(bool compress);
222 bool inline_string_char_access(bool is_store);
223 Node* round_double_node(Node* n);
224 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_math_native(vmIntrinsics::ID id);
226 bool inline_math(vmIntrinsics::ID id);
227 template <typename OverflowOp>
228 bool inline_math_overflow(Node* arg1, Node* arg2);
229 void inline_math_mathExact(Node* math, Node* test);
230 bool inline_math_addExactI(bool is_increment);
231 bool inline_math_addExactL(bool is_increment);
232 bool inline_math_multiplyExactI();
233 bool inline_math_multiplyExactL();
234 bool inline_math_negateExactI();
235 bool inline_math_negateExactL();
236 bool inline_math_subtractExactI(bool is_decrement);
237 bool inline_math_subtractExactL(bool is_decrement);
238 bool inline_min_max(vmIntrinsics::ID id);
239 bool inline_notify(vmIntrinsics::ID id);
240 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
241 // This returns Type::AnyPtr, RawPtr, or OopPtr.
242 int classify_unsafe_addr(Node* &base, Node* &offset);
243 Node* make_unsafe_address(Node* base, Node* offset, bool is_store);
244 // Helper for inline_unsafe_access.
245 // Generates the guards that check whether the result of
246 // Unsafe.getObject should be recorded in an SATB log buffer.
247 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
248
249 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
250 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
251 static bool klass_needs_init_guard(Node* kls);
252 bool inline_unsafe_allocate();
253 bool inline_unsafe_newArray(bool uninitialized);
254 bool inline_unsafe_copyMemory();
255 bool inline_native_currentThread();
256
257 bool inline_native_time_funcs(address method, const char* funcName);
258 #ifdef TRACE_HAVE_INTRINSICS
259 bool inline_native_classID();
260 bool inline_native_getBufferWriter();
261 #endif
262 bool inline_native_isInterrupted();
263 bool inline_native_Class_query(vmIntrinsics::ID id);
264 bool inline_native_subtype_check();
265 bool inline_native_getLength();
266 bool inline_array_copyOf(bool is_copyOfRange);
267 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
268 bool inline_preconditions_checkIndex();
269 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
270 bool inline_native_clone(bool is_virtual);
271 bool inline_native_Reflection_getCallerClass();
272 // Helper function for inlining native object hash method
273 bool inline_native_hashcode(bool is_virtual, bool is_static);
274 bool inline_native_getClass();
275
276 // Helper functions for inlining arraycopy
277 bool inline_arraycopy();
278 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
279 RegionNode* slow_region);
280 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
281 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
282
283 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
284 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
285 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
286 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
287 bool inline_unsafe_fence(vmIntrinsics::ID id);
288 bool inline_onspinwait();
289 bool inline_fp_conversions(vmIntrinsics::ID id);
290 bool inline_number_methods(vmIntrinsics::ID id);
291 bool inline_reference_get();
292 bool inline_Class_cast();
293 bool inline_aescrypt_Block(vmIntrinsics::ID id);
294 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
295 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
296 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
297 Node* inline_counterMode_AESCrypt_predicate();
298 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
299 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
300 bool inline_ghash_processBlocks();
301 bool inline_sha_implCompress(vmIntrinsics::ID id);
302 bool inline_digestBase_implCompressMB(int predicate);
303 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
304 bool long_state, address stubAddr, const char *stubName,
305 Node* src_start, Node* ofs, Node* limit);
306 Node* get_state_from_sha_object(Node *sha_object);
307 Node* get_state_from_sha5_object(Node *sha_object);
308 Node* inline_digestBase_implCompressMB_predicate(int predicate);
309 bool inline_encodeISOArray();
310 bool inline_updateCRC32();
311 bool inline_updateBytesCRC32();
312 bool inline_updateByteBufferCRC32();
313 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
314 bool inline_updateBytesCRC32C();
315 bool inline_updateDirectByteBufferCRC32C();
316 bool inline_updateBytesAdler32();
317 bool inline_updateByteBufferAdler32();
318 bool inline_multiplyToLen();
319 bool inline_hasNegatives();
320 bool inline_squareToLen();
321 bool inline_mulAdd();
322 bool inline_montgomeryMultiply();
323 bool inline_montgomerySquare();
324 bool inline_vectorizedMismatch();
325 bool inline_fma(vmIntrinsics::ID id);
326
327 bool inline_profileBoolean();
328 bool inline_isCompileConstant();
329 };
330
331 //---------------------------make_vm_intrinsic----------------------------
332 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
333 vmIntrinsics::ID id = m->intrinsic_id();
334 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
335
336 if (!m->is_loaded()) {
337 // Do not attempt to inline unloaded methods.
338 return NULL;
339 }
340
341 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
342 bool is_available = false;
343
344 {
345 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
346 // the compiler must transition to '_thread_in_vm' state because both
347 // methods access VM-internal data.
348 VM_ENTRY_MARK;
349 methodHandle mh(THREAD, m->get_Method());
350 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
351 !C->directive()->is_intrinsic_disabled(mh) &&
352 !vmIntrinsics::is_disabled_by_flags(mh);
353
354 }
355
356 if (is_available) {
357 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
358 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
359 return new LibraryIntrinsic(m, is_virtual,
360 vmIntrinsics::predicates_needed(id),
361 vmIntrinsics::does_virtual_dispatch(id),
362 (vmIntrinsics::ID) id);
363 } else {
364 return NULL;
365 }
366 }
367
368 //----------------------register_library_intrinsics-----------------------
369 // Initialize this file's data structures, for each Compile instance.
370 void Compile::register_library_intrinsics() {
371 // Nothing to do here.
372 }
373
374 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
375 LibraryCallKit kit(jvms, this);
376 Compile* C = kit.C;
377 int nodes = C->unique();
378 #ifndef PRODUCT
379 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
380 char buf[1000];
381 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
382 tty->print_cr("Intrinsic %s", str);
383 }
384 #endif
385 ciMethod* callee = kit.callee();
386 const int bci = kit.bci();
387
388 // Try to inline the intrinsic.
389 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
390 kit.try_to_inline(_last_predicate)) {
391 if (C->print_intrinsics() || C->print_inlining()) {
392 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
393 }
394 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
395 if (C->log()) {
396 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
397 vmIntrinsics::name_at(intrinsic_id()),
398 (is_virtual() ? " virtual='1'" : ""),
399 C->unique() - nodes);
400 }
401 // Push the result from the inlined method onto the stack.
402 kit.push_result();
403 C->print_inlining_update(this);
404 return kit.transfer_exceptions_into_jvms();
405 }
406
407 // The intrinsic bailed out
408 if (C->print_intrinsics() || C->print_inlining()) {
409 if (jvms->has_method()) {
410 // Not a root compile.
411 const char* msg;
412 if (callee->intrinsic_candidate()) {
413 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
414 } else {
415 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
416 : "failed to inline (intrinsic), method not annotated";
417 }
418 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
419 } else {
420 // Root compile
421 tty->print("Did not generate intrinsic %s%s at bci:%d in",
422 vmIntrinsics::name_at(intrinsic_id()),
423 (is_virtual() ? " (virtual)" : ""), bci);
424 }
425 }
426 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
427 C->print_inlining_update(this);
428 return NULL;
429 }
430
431 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
432 LibraryCallKit kit(jvms, this);
433 Compile* C = kit.C;
434 int nodes = C->unique();
435 _last_predicate = predicate;
436 #ifndef PRODUCT
437 assert(is_predicated() && predicate < predicates_count(), "sanity");
438 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
439 char buf[1000];
440 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
441 tty->print_cr("Predicate for intrinsic %s", str);
442 }
443 #endif
444 ciMethod* callee = kit.callee();
445 const int bci = kit.bci();
446
447 Node* slow_ctl = kit.try_to_predicate(predicate);
448 if (!kit.failing()) {
449 if (C->print_intrinsics() || C->print_inlining()) {
450 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
451 }
452 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
453 if (C->log()) {
454 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
455 vmIntrinsics::name_at(intrinsic_id()),
456 (is_virtual() ? " virtual='1'" : ""),
457 C->unique() - nodes);
458 }
459 return slow_ctl; // Could be NULL if the check folds.
460 }
461
462 // The intrinsic bailed out
463 if (C->print_intrinsics() || C->print_inlining()) {
464 if (jvms->has_method()) {
465 // Not a root compile.
466 const char* msg = "failed to generate predicate for intrinsic";
467 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
468 } else {
469 // Root compile
470 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
471 vmIntrinsics::name_at(intrinsic_id()),
472 (is_virtual() ? " (virtual)" : ""), bci);
473 }
474 }
475 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
476 return NULL;
477 }
478
479 bool LibraryCallKit::try_to_inline(int predicate) {
480 // Handle symbolic names for otherwise undistinguished boolean switches:
481 const bool is_store = true;
482 const bool is_compress = true;
483 const bool is_static = true;
484 const bool is_volatile = true;
485
486 if (!jvms()->has_method()) {
487 // Root JVMState has a null method.
488 assert(map()->memory()->Opcode() == Op_Parm, "");
489 // Insert the memory aliasing node
490 set_all_memory(reset_memory());
491 }
492 assert(merged_memory(), "");
493
494
495 switch (intrinsic_id()) {
496 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
497 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
498 case vmIntrinsics::_getClass: return inline_native_getClass();
499
500 case vmIntrinsics::_dsin:
501 case vmIntrinsics::_dcos:
502 case vmIntrinsics::_dtan:
503 case vmIntrinsics::_dabs:
504 case vmIntrinsics::_datan2:
505 case vmIntrinsics::_dsqrt:
506 case vmIntrinsics::_dexp:
507 case vmIntrinsics::_dlog:
508 case vmIntrinsics::_dlog10:
509 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
510
511 case vmIntrinsics::_min:
512 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
513
514 case vmIntrinsics::_notify:
515 case vmIntrinsics::_notifyAll:
516 if (InlineNotify) {
517 return inline_notify(intrinsic_id());
518 }
519 return false;
520
521 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
522 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
523 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
524 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
525 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
526 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
527 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
528 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
529 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
530 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
531 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
532 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
533
534 case vmIntrinsics::_arraycopy: return inline_arraycopy();
535
536 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
537 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
538 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
539 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
540
541 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
542 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
543 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
544 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
545 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
546 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
547 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
548
549 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
550 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
551
552 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
553 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
554 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
555 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
556
557 case vmIntrinsics::_compressStringC:
558 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
559 case vmIntrinsics::_inflateStringC:
560 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
561
562 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
563 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
564 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
565 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
566 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
567 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
568 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
569 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
570 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
571
572 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
573 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
574 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
575 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
576 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
577 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
578 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
579 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
580 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
581
582 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
583 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
584 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
585 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
586 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
587 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
588 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
589 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
590 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
591
592 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
593 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
594 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
595 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
596 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
597 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
598 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
599 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
600 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
601
602 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
603 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
604 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
605 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
606
607 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
608 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
609 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
610 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
611
612 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
613 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
614 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
615 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
616 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
617 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
618 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
619 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
620 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
621
622 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
623 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
624 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
625 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
626 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
627 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
628 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
629 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
630 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
631
632 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
633 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
634 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
635 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
636 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
637 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
638 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
639 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
640 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
641
642 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
643 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
644 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
645 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
646 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
647 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
648 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
649 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
650 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
651
652 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
653 case vmIntrinsics::_compareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
654 case vmIntrinsics::_compareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
655 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
656 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
657
658 case vmIntrinsics::_weakCompareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
659 case vmIntrinsics::_weakCompareAndSwapObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
660 case vmIntrinsics::_weakCompareAndSwapObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
661 case vmIntrinsics::_weakCompareAndSwapObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
662 case vmIntrinsics::_weakCompareAndSwapByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
663 case vmIntrinsics::_weakCompareAndSwapByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
664 case vmIntrinsics::_weakCompareAndSwapByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
665 case vmIntrinsics::_weakCompareAndSwapByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
666 case vmIntrinsics::_weakCompareAndSwapShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
667 case vmIntrinsics::_weakCompareAndSwapShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
668 case vmIntrinsics::_weakCompareAndSwapShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
669 case vmIntrinsics::_weakCompareAndSwapShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
670 case vmIntrinsics::_weakCompareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
671 case vmIntrinsics::_weakCompareAndSwapIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
672 case vmIntrinsics::_weakCompareAndSwapIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
673 case vmIntrinsics::_weakCompareAndSwapIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
674 case vmIntrinsics::_weakCompareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
675 case vmIntrinsics::_weakCompareAndSwapLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
676 case vmIntrinsics::_weakCompareAndSwapLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
677 case vmIntrinsics::_weakCompareAndSwapLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
678
679 case vmIntrinsics::_compareAndExchangeObjectVolatile: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
680 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
681 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
682 case vmIntrinsics::_compareAndExchangeByteVolatile: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
683 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
684 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
685 case vmIntrinsics::_compareAndExchangeShortVolatile: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
686 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
687 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
688 case vmIntrinsics::_compareAndExchangeIntVolatile: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
689 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
690 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
691 case vmIntrinsics::_compareAndExchangeLongVolatile: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
692 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
693 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
694
695 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
696 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
697 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
698 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
699
700 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
701 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
702 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
703 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
704 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
705
706 case vmIntrinsics::_loadFence:
707 case vmIntrinsics::_storeFence:
708 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
709
710 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
711
712 case vmIntrinsics::_currentThread: return inline_native_currentThread();
713 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
714
715 #ifdef TRACE_HAVE_INTRINSICS
716 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
717 case vmIntrinsics::_getClassId: return inline_native_classID();
718 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
719 #endif
720 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
721 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
722 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
723 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
724 case vmIntrinsics::_getLength: return inline_native_getLength();
725 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
726 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
727 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
728 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
729 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
730 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
731
732 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
733 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
734
735 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
736
737 case vmIntrinsics::_isInstance:
738 case vmIntrinsics::_getModifiers:
739 case vmIntrinsics::_isInterface:
740 case vmIntrinsics::_isArray:
741 case vmIntrinsics::_isPrimitive:
742 case vmIntrinsics::_getSuperclass:
743 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
744
745 case vmIntrinsics::_floatToRawIntBits:
746 case vmIntrinsics::_floatToIntBits:
747 case vmIntrinsics::_intBitsToFloat:
748 case vmIntrinsics::_doubleToRawLongBits:
749 case vmIntrinsics::_doubleToLongBits:
750 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
751
752 case vmIntrinsics::_numberOfLeadingZeros_i:
753 case vmIntrinsics::_numberOfLeadingZeros_l:
754 case vmIntrinsics::_numberOfTrailingZeros_i:
755 case vmIntrinsics::_numberOfTrailingZeros_l:
756 case vmIntrinsics::_bitCount_i:
757 case vmIntrinsics::_bitCount_l:
758 case vmIntrinsics::_reverseBytes_i:
759 case vmIntrinsics::_reverseBytes_l:
760 case vmIntrinsics::_reverseBytes_s:
761 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
762
763 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
764
765 case vmIntrinsics::_Reference_get: return inline_reference_get();
766
767 case vmIntrinsics::_Class_cast: return inline_Class_cast();
768
769 case vmIntrinsics::_aescrypt_encryptBlock:
770 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
771
772 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
773 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
774 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
775
776 case vmIntrinsics::_counterMode_AESCrypt:
777 return inline_counterMode_AESCrypt(intrinsic_id());
778
779 case vmIntrinsics::_sha_implCompress:
780 case vmIntrinsics::_sha2_implCompress:
781 case vmIntrinsics::_sha5_implCompress:
782 return inline_sha_implCompress(intrinsic_id());
783
784 case vmIntrinsics::_digestBase_implCompressMB:
785 return inline_digestBase_implCompressMB(predicate);
786
787 case vmIntrinsics::_multiplyToLen:
788 return inline_multiplyToLen();
789
790 case vmIntrinsics::_squareToLen:
791 return inline_squareToLen();
792
793 case vmIntrinsics::_mulAdd:
794 return inline_mulAdd();
795
796 case vmIntrinsics::_montgomeryMultiply:
797 return inline_montgomeryMultiply();
798 case vmIntrinsics::_montgomerySquare:
799 return inline_montgomerySquare();
800
801 case vmIntrinsics::_vectorizedMismatch:
802 return inline_vectorizedMismatch();
803
804 case vmIntrinsics::_ghash_processBlocks:
805 return inline_ghash_processBlocks();
806
807 case vmIntrinsics::_encodeISOArray:
808 case vmIntrinsics::_encodeByteISOArray:
809 return inline_encodeISOArray();
810
811 case vmIntrinsics::_updateCRC32:
812 return inline_updateCRC32();
813 case vmIntrinsics::_updateBytesCRC32:
814 return inline_updateBytesCRC32();
815 case vmIntrinsics::_updateByteBufferCRC32:
816 return inline_updateByteBufferCRC32();
817
818 case vmIntrinsics::_updateBytesCRC32C:
819 return inline_updateBytesCRC32C();
820 case vmIntrinsics::_updateDirectByteBufferCRC32C:
821 return inline_updateDirectByteBufferCRC32C();
822
823 case vmIntrinsics::_updateBytesAdler32:
824 return inline_updateBytesAdler32();
825 case vmIntrinsics::_updateByteBufferAdler32:
826 return inline_updateByteBufferAdler32();
827
828 case vmIntrinsics::_profileBoolean:
829 return inline_profileBoolean();
830 case vmIntrinsics::_isCompileConstant:
831 return inline_isCompileConstant();
832
833 case vmIntrinsics::_hasNegatives:
834 return inline_hasNegatives();
835
836 case vmIntrinsics::_fmaD:
837 case vmIntrinsics::_fmaF:
838 return inline_fma(intrinsic_id());
839
840 default:
841 // If you get here, it may be that someone has added a new intrinsic
842 // to the list in vmSymbols.hpp without implementing it here.
843 #ifndef PRODUCT
844 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
845 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
846 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
847 }
848 #endif
849 return false;
850 }
851 }
852
853 Node* LibraryCallKit::try_to_predicate(int predicate) {
854 if (!jvms()->has_method()) {
855 // Root JVMState has a null method.
856 assert(map()->memory()->Opcode() == Op_Parm, "");
857 // Insert the memory aliasing node
858 set_all_memory(reset_memory());
859 }
860 assert(merged_memory(), "");
861
862 switch (intrinsic_id()) {
863 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
864 return inline_cipherBlockChaining_AESCrypt_predicate(false);
865 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
866 return inline_cipherBlockChaining_AESCrypt_predicate(true);
867 case vmIntrinsics::_counterMode_AESCrypt:
868 return inline_counterMode_AESCrypt_predicate();
869 case vmIntrinsics::_digestBase_implCompressMB:
870 return inline_digestBase_implCompressMB_predicate(predicate);
871
872 default:
873 // If you get here, it may be that someone has added a new intrinsic
874 // to the list in vmSymbols.hpp without implementing it here.
875 #ifndef PRODUCT
876 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
877 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
878 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
879 }
880 #endif
881 Node* slow_ctl = control();
882 set_control(top()); // No fast path instrinsic
883 return slow_ctl;
884 }
885 }
886
887 //------------------------------set_result-------------------------------
888 // Helper function for finishing intrinsics.
889 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
890 record_for_igvn(region);
891 set_control(_gvn.transform(region));
892 set_result( _gvn.transform(value));
893 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
894 }
895
896 //------------------------------generate_guard---------------------------
897 // Helper function for generating guarded fast-slow graph structures.
898 // The given 'test', if true, guards a slow path. If the test fails
899 // then a fast path can be taken. (We generally hope it fails.)
900 // In all cases, GraphKit::control() is updated to the fast path.
901 // The returned value represents the control for the slow path.
902 // The return value is never 'top'; it is either a valid control
903 // or NULL if it is obvious that the slow path can never be taken.
904 // Also, if region and the slow control are not NULL, the slow edge
905 // is appended to the region.
906 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
907 if (stopped()) {
908 // Already short circuited.
909 return NULL;
910 }
911
912 // Build an if node and its projections.
913 // If test is true we take the slow path, which we assume is uncommon.
914 if (_gvn.type(test) == TypeInt::ZERO) {
915 // The slow branch is never taken. No need to build this guard.
916 return NULL;
917 }
918
919 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
920
921 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
922 if (if_slow == top()) {
923 // The slow branch is never taken. No need to build this guard.
924 return NULL;
925 }
926
927 if (region != NULL)
928 region->add_req(if_slow);
929
930 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
931 set_control(if_fast);
932
933 return if_slow;
934 }
935
936 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
937 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
938 }
939 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
940 return generate_guard(test, region, PROB_FAIR);
941 }
942
943 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
944 Node* *pos_index) {
945 if (stopped())
946 return NULL; // already stopped
947 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
948 return NULL; // index is already adequately typed
949 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
950 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
951 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
952 if (is_neg != NULL && pos_index != NULL) {
953 // Emulate effect of Parse::adjust_map_after_if.
954 Node* ccast = new CastIINode(index, TypeInt::POS);
955 ccast->set_req(0, control());
956 (*pos_index) = _gvn.transform(ccast);
957 }
958 return is_neg;
959 }
960
961 // Make sure that 'position' is a valid limit index, in [0..length].
962 // There are two equivalent plans for checking this:
963 // A. (offset + copyLength) unsigned<= arrayLength
964 // B. offset <= (arrayLength - copyLength)
965 // We require that all of the values above, except for the sum and
966 // difference, are already known to be non-negative.
967 // Plan A is robust in the face of overflow, if offset and copyLength
968 // are both hugely positive.
969 //
970 // Plan B is less direct and intuitive, but it does not overflow at
971 // all, since the difference of two non-negatives is always
972 // representable. Whenever Java methods must perform the equivalent
973 // check they generally use Plan B instead of Plan A.
974 // For the moment we use Plan A.
975 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
976 Node* subseq_length,
977 Node* array_length,
978 RegionNode* region) {
979 if (stopped())
980 return NULL; // already stopped
981 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
982 if (zero_offset && subseq_length->eqv_uncast(array_length))
983 return NULL; // common case of whole-array copy
984 Node* last = subseq_length;
985 if (!zero_offset) // last += offset
986 last = _gvn.transform(new AddINode(last, offset));
987 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
988 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
989 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
990 return is_over;
991 }
992
993 // Emit range checks for the given String.value byte array
994 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
995 if (stopped()) {
996 return; // already stopped
997 }
998 RegionNode* bailout = new RegionNode(1);
999 record_for_igvn(bailout);
1000 if (char_count) {
1001 // Convert char count to byte count
1002 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1003 }
1004
1005 // Offset and count must not be negative
1006 generate_negative_guard(offset, bailout);
1007 generate_negative_guard(count, bailout);
1008 // Offset + count must not exceed length of array
1009 generate_limit_guard(offset, count, load_array_length(array), bailout);
1010
1011 if (bailout->req() > 1) {
1012 PreserveJVMState pjvms(this);
1013 set_control(_gvn.transform(bailout));
1014 uncommon_trap(Deoptimization::Reason_intrinsic,
1015 Deoptimization::Action_maybe_recompile);
1016 }
1017 }
1018
1019 //--------------------------generate_current_thread--------------------
1020 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1021 ciKlass* thread_klass = env()->Thread_klass();
1022 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1023 Node* thread = _gvn.transform(new ThreadLocalNode());
1024 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1025 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1026 tls_output = thread;
1027 return threadObj;
1028 }
1029
1030
1031 //------------------------------make_string_method_node------------------------
1032 // Helper method for String intrinsic functions. This version is called with
1033 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1034 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1035 // containing the lengths of str1 and str2.
1036 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1037 Node* result = NULL;
1038 switch (opcode) {
1039 case Op_StrIndexOf:
1040 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1041 str1_start, cnt1, str2_start, cnt2, ae);
1042 break;
1043 case Op_StrComp:
1044 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1045 str1_start, cnt1, str2_start, cnt2, ae);
1046 break;
1047 case Op_StrEquals:
1048 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1049 // Use the constant length if there is one because optimized match rule may exist.
1050 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1051 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1052 break;
1053 default:
1054 ShouldNotReachHere();
1055 return NULL;
1056 }
1057
1058 // All these intrinsics have checks.
1059 C->set_has_split_ifs(true); // Has chance for split-if optimization
1060
1061 return _gvn.transform(result);
1062 }
1063
1064 //------------------------------inline_string_compareTo------------------------
1065 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1066 Node* arg1 = argument(0);
1067 Node* arg2 = argument(1);
1068
1069 arg1 = shenandoah_read_barrier(arg1);
1070 arg2 = shenandoah_read_barrier(arg2);
1071
1072 // Get start addr and length of first argument
1073 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1074 Node* arg1_cnt = load_array_length(arg1);
1075
1076 // Get start addr and length of second argument
1077 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1078 Node* arg2_cnt = load_array_length(arg2);
1079
1080 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1081 set_result(result);
1082 return true;
1083 }
1084
1085 //------------------------------inline_string_equals------------------------
1086 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1087 Node* arg1 = argument(0);
1088 Node* arg2 = argument(1);
1089
1090 // paths (plus control) merge
1091 RegionNode* region = new RegionNode(3);
1092 Node* phi = new PhiNode(region, TypeInt::BOOL);
1093
1094 if (!stopped()) {
1095
1096 arg1 = shenandoah_read_barrier(arg1);
1097 arg2 = shenandoah_read_barrier(arg2);
1098
1099 // Get start addr and length of first argument
1100 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1101 Node* arg1_cnt = load_array_length(arg1);
1102
1103 // Get start addr and length of second argument
1104 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1105 Node* arg2_cnt = load_array_length(arg2);
1106
1107 // Check for arg1_cnt != arg2_cnt
1108 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1109 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1110 Node* if_ne = generate_slow_guard(bol, NULL);
1111 if (if_ne != NULL) {
1112 phi->init_req(2, intcon(0));
1113 region->init_req(2, if_ne);
1114 }
1115
1116 // Check for count == 0 is done by assembler code for StrEquals.
1117
1118 if (!stopped()) {
1119 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1120 phi->init_req(1, equals);
1121 region->init_req(1, control());
1122 }
1123 }
1124
1125 // post merge
1126 set_control(_gvn.transform(region));
1127 record_for_igvn(region);
1128
1129 set_result(_gvn.transform(phi));
1130 return true;
1131 }
1132
1133 //------------------------------inline_array_equals----------------------------
1134 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1135 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1136 Node* arg1 = argument(0);
1137 Node* arg2 = argument(1);
1138
1139 arg1 = shenandoah_read_barrier(arg1);
1140 arg2 = shenandoah_read_barrier(arg2);
1141
1142 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1143 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1144 return true;
1145 }
1146
1147 //------------------------------inline_hasNegatives------------------------------
1148 bool LibraryCallKit::inline_hasNegatives() {
1149 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1150 return false;
1151 }
1152
1153 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1154 // no receiver since it is static method
1155 Node* ba = argument(0);
1156 Node* offset = argument(1);
1157 Node* len = argument(2);
1158
1159 ba = cast_not_null(ba, false);
1160
1161 // Range checks
1162 generate_string_range_check(ba, offset, len, false);
1163 if (stopped()) {
1164 return true;
1165 }
1166
1167 ba = shenandoah_read_barrier(ba);
1168
1169 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1170 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1171 set_result(_gvn.transform(result));
1172 return true;
1173 }
1174
1175 bool LibraryCallKit::inline_preconditions_checkIndex() {
1176 Node* index = argument(0);
1177 Node* length = argument(1);
1178 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1179 return false;
1180 }
1181
1182 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1183 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1184
1185 {
1186 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1187 uncommon_trap(Deoptimization::Reason_intrinsic,
1188 Deoptimization::Action_make_not_entrant);
1189 }
1190
1191 if (stopped()) {
1192 return false;
1193 }
1194
1195 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1196 BoolTest::mask btest = BoolTest::lt;
1197 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1198 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1199 _gvn.set_type(rc, rc->Value(&_gvn));
1200 if (!rc_bool->is_Con()) {
1201 record_for_igvn(rc);
1202 }
1203 set_control(_gvn.transform(new IfTrueNode(rc)));
1204 {
1205 PreserveJVMState pjvms(this);
1206 set_control(_gvn.transform(new IfFalseNode(rc)));
1207 uncommon_trap(Deoptimization::Reason_range_check,
1208 Deoptimization::Action_make_not_entrant);
1209 }
1210
1211 if (stopped()) {
1212 return false;
1213 }
1214
1215 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1216 result->set_req(0, control());
1217 result = _gvn.transform(result);
1218 set_result(result);
1219 replace_in_map(index, result);
1220 return true;
1221 }
1222
1223 //------------------------------inline_string_indexOf------------------------
1224 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1225 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1226 return false;
1227 }
1228 Node* src = argument(0);
1229 Node* tgt = argument(1);
1230
1231 // Make the merge point
1232 RegionNode* result_rgn = new RegionNode(4);
1233 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1234
1235 src = cast_not_null(src, false);
1236 tgt = cast_not_null(tgt, false);
1237
1238 src = shenandoah_read_barrier(src);
1239 tgt = shenandoah_read_barrier(tgt);
1240
1241 // Get start addr and length of source string
1242 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1243 Node* src_count = load_array_length(src);
1244
1245 // Get start addr and length of substring
1246 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1247 Node* tgt_count = load_array_length(tgt);
1248
1249 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1250 // Divide src size by 2 if String is UTF16 encoded
1251 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1252 }
1253 if (ae == StrIntrinsicNode::UU) {
1254 // Divide substring size by 2 if String is UTF16 encoded
1255 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1256 }
1257
1258 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1259 if (result != NULL) {
1260 result_phi->init_req(3, result);
1261 result_rgn->init_req(3, control());
1262 }
1263 set_control(_gvn.transform(result_rgn));
1264 record_for_igvn(result_rgn);
1265 set_result(_gvn.transform(result_phi));
1266
1267 return true;
1268 }
1269
1270 //-----------------------------inline_string_indexOf-----------------------
1271 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1272 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1273 return false;
1274 }
1275 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1276 return false;
1277 }
1278 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1279 Node* src = argument(0); // byte[]
1280 Node* src_count = argument(1); // char count
1281 Node* tgt = argument(2); // byte[]
1282 Node* tgt_count = argument(3); // char count
1283 Node* from_index = argument(4); // char index
1284
1285 src = cast_not_null(src, false);
1286 tgt = cast_not_null(tgt, false);
1287
1288 src = shenandoah_read_barrier(src);
1289 tgt = shenandoah_read_barrier(tgt);
1290
1291 // Multiply byte array index by 2 if String is UTF16 encoded
1292 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1293 src_count = _gvn.transform(new SubINode(src_count, from_index));
1294 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1295 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1296
1297 // Range checks
1298 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1299 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1300 if (stopped()) {
1301 return true;
1302 }
1303
1304 RegionNode* region = new RegionNode(5);
1305 Node* phi = new PhiNode(region, TypeInt::INT);
1306
1307 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1308 if (result != NULL) {
1309 // The result is index relative to from_index if substring was found, -1 otherwise.
1310 // Generate code which will fold into cmove.
1311 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1312 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1313
1314 Node* if_lt = generate_slow_guard(bol, NULL);
1315 if (if_lt != NULL) {
1316 // result == -1
1317 phi->init_req(3, result);
1318 region->init_req(3, if_lt);
1319 }
1320 if (!stopped()) {
1321 result = _gvn.transform(new AddINode(result, from_index));
1322 phi->init_req(4, result);
1323 region->init_req(4, control());
1324 }
1325 }
1326
1327 set_control(_gvn.transform(region));
1328 record_for_igvn(region);
1329 set_result(_gvn.transform(phi));
1330
1331 return true;
1332 }
1333
1334 // Create StrIndexOfNode with fast path checks
1335 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1336 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1337 // Check for substr count > string count
1338 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1339 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1340 Node* if_gt = generate_slow_guard(bol, NULL);
1341 if (if_gt != NULL) {
1342 phi->init_req(1, intcon(-1));
1343 region->init_req(1, if_gt);
1344 }
1345 if (!stopped()) {
1346 // Check for substr count == 0
1347 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1348 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1349 Node* if_zero = generate_slow_guard(bol, NULL);
1350 if (if_zero != NULL) {
1351 phi->init_req(2, intcon(0));
1352 region->init_req(2, if_zero);
1353 }
1354 }
1355 if (!stopped()) {
1356 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1357 }
1358 return NULL;
1359 }
1360
1361 //-----------------------------inline_string_indexOfChar-----------------------
1362 bool LibraryCallKit::inline_string_indexOfChar() {
1363 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1364 return false;
1365 }
1366 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1367 return false;
1368 }
1369 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1370 Node* src = argument(0); // byte[]
1371 Node* tgt = argument(1); // tgt is int ch
1372 Node* from_index = argument(2);
1373 Node* max = argument(3);
1374
1375 src = cast_not_null(src, false);
1376 src = shenandoah_read_barrier(src);
1377
1378 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1379 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1380 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1381
1382 // Range checks
1383 generate_string_range_check(src, src_offset, src_count, true);
1384 if (stopped()) {
1385 return true;
1386 }
1387
1388 RegionNode* region = new RegionNode(3);
1389 Node* phi = new PhiNode(region, TypeInt::INT);
1390
1391 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1392 C->set_has_split_ifs(true); // Has chance for split-if optimization
1393 _gvn.transform(result);
1394
1395 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1396 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1397
1398 Node* if_lt = generate_slow_guard(bol, NULL);
1399 if (if_lt != NULL) {
1400 // result == -1
1401 phi->init_req(2, result);
1402 region->init_req(2, if_lt);
1403 }
1404 if (!stopped()) {
1405 result = _gvn.transform(new AddINode(result, from_index));
1406 phi->init_req(1, result);
1407 region->init_req(1, control());
1408 }
1409 set_control(_gvn.transform(region));
1410 record_for_igvn(region);
1411 set_result(_gvn.transform(phi));
1412
1413 return true;
1414 }
1415 //---------------------------inline_string_copy---------------------
1416 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1417 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1418 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1419 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1420 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1421 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1422 bool LibraryCallKit::inline_string_copy(bool compress) {
1423 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1424 return false;
1425 }
1426 int nargs = 5; // 2 oops, 3 ints
1427 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1428
1429 Node* src = argument(0);
1430 Node* src_offset = argument(1);
1431 Node* dst = argument(2);
1432 Node* dst_offset = argument(3);
1433 Node* length = argument(4);
1434
1435 // Check for allocation before we add nodes that would confuse
1436 // tightly_coupled_allocation()
1437 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1438
1439 // Figure out the size and type of the elements we will be copying.
1440 const Type* src_type = src->Value(&_gvn);
1441 const Type* dst_type = dst->Value(&_gvn);
1442 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1443 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1444 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1445 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1446 "Unsupported array types for inline_string_copy");
1447
1448 src = cast_not_null(src, false);
1449 dst = cast_not_null(dst, false);
1450
1451 // Convert char[] offsets to byte[] offsets
1452 bool convert_src = (compress && src_elem == T_BYTE);
1453 bool convert_dst = (!compress && dst_elem == T_BYTE);
1454 if (convert_src) {
1455 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1456 } else if (convert_dst) {
1457 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1458 }
1459
1460 // Range checks
1461 generate_string_range_check(src, src_offset, length, convert_src);
1462 generate_string_range_check(dst, dst_offset, length, convert_dst);
1463 if (stopped()) {
1464 return true;
1465 }
1466
1467 src = shenandoah_read_barrier(src);
1468 dst = shenandoah_write_barrier(dst);
1469
1470 Node* src_start = array_element_address(src, src_offset, src_elem);
1471 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1472 // 'src_start' points to src array + scaled offset
1473 // 'dst_start' points to dst array + scaled offset
1474 Node* count = NULL;
1475 if (compress) {
1476 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1477 } else {
1478 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1479 }
1480
1481 if (alloc != NULL) {
1482 if (alloc->maybe_set_complete(&_gvn)) {
1483 // "You break it, you buy it."
1484 InitializeNode* init = alloc->initialization();
1485 assert(init->is_complete(), "we just did this");
1486 init->set_complete_with_arraycopy();
1487 assert(dst->is_CheckCastPP(), "sanity");
1488 assert(dst->in(0)->in(0) == init, "dest pinned");
1489 }
1490 // Do not let stores that initialize this object be reordered with
1491 // a subsequent store that would make this object accessible by
1492 // other threads.
1493 // Record what AllocateNode this StoreStore protects so that
1494 // escape analysis can go from the MemBarStoreStoreNode to the
1495 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1496 // based on the escape status of the AllocateNode.
1497 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1498 }
1499 if (compress) {
1500 set_result(_gvn.transform(count));
1501 }
1502 return true;
1503 }
1504
1505 #ifdef _LP64
1506 #define XTOP ,top() /*additional argument*/
1507 #else //_LP64
1508 #define XTOP /*no additional argument*/
1509 #endif //_LP64
1510
1511 //------------------------inline_string_toBytesU--------------------------
1512 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1513 bool LibraryCallKit::inline_string_toBytesU() {
1514 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1515 return false;
1516 }
1517 // Get the arguments.
1518 Node* value = argument(0);
1519 Node* offset = argument(1);
1520 Node* length = argument(2);
1521
1522 Node* newcopy = NULL;
1523
1524 // Set the original stack and the reexecute bit for the interpreter to reexecute
1525 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1526 { PreserveReexecuteState preexecs(this);
1527 jvms()->set_should_reexecute(true);
1528
1529 // Check if a null path was taken unconditionally.
1530 value = null_check(value);
1531
1532 RegionNode* bailout = new RegionNode(1);
1533 record_for_igvn(bailout);
1534
1535 // Range checks
1536 generate_negative_guard(offset, bailout);
1537 generate_negative_guard(length, bailout);
1538 generate_limit_guard(offset, length, load_array_length(value), bailout);
1539 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1540 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1541
1542 if (bailout->req() > 1) {
1543 PreserveJVMState pjvms(this);
1544 set_control(_gvn.transform(bailout));
1545 uncommon_trap(Deoptimization::Reason_intrinsic,
1546 Deoptimization::Action_maybe_recompile);
1547 }
1548 if (stopped()) {
1549 return true;
1550 }
1551
1552 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1553 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1554 newcopy = new_array(klass_node, size, 0); // no arguments to push
1555 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1556
1557 // Calculate starting addresses.
1558 value = shenandoah_read_barrier(value);
1559
1560 Node* src_start = array_element_address(value, offset, T_CHAR);
1561 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1562
1563 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1564 const TypeInt* toffset = gvn().type(offset)->is_int();
1565 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1566
1567 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1568 const char* copyfunc_name = "arraycopy";
1569 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1570 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1571 OptoRuntime::fast_arraycopy_Type(),
1572 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1573 src_start, dst_start, ConvI2X(length) XTOP);
1574 // Do not let reads from the cloned object float above the arraycopy.
1575 if (alloc != NULL) {
1576 if (alloc->maybe_set_complete(&_gvn)) {
1577 // "You break it, you buy it."
1578 InitializeNode* init = alloc->initialization();
1579 assert(init->is_complete(), "we just did this");
1580 init->set_complete_with_arraycopy();
1581 assert(newcopy->is_CheckCastPP(), "sanity");
1582 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1583 }
1584 // Do not let stores that initialize this object be reordered with
1585 // a subsequent store that would make this object accessible by
1586 // other threads.
1587 // Record what AllocateNode this StoreStore protects so that
1588 // escape analysis can go from the MemBarStoreStoreNode to the
1589 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1590 // based on the escape status of the AllocateNode.
1591 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1592 } else {
1593 insert_mem_bar(Op_MemBarCPUOrder);
1594 }
1595 } // original reexecute is set back here
1596
1597 C->set_has_split_ifs(true); // Has chance for split-if optimization
1598 if (!stopped()) {
1599 set_result(newcopy);
1600 }
1601 return true;
1602 }
1603
1604 //------------------------inline_string_getCharsU--------------------------
1605 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1606 bool LibraryCallKit::inline_string_getCharsU() {
1607 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1608 return false;
1609 }
1610
1611 // Get the arguments.
1612 Node* src = argument(0);
1613 Node* src_begin = argument(1);
1614 Node* src_end = argument(2); // exclusive offset (i < src_end)
1615 Node* dst = argument(3);
1616 Node* dst_begin = argument(4);
1617
1618 // Check for allocation before we add nodes that would confuse
1619 // tightly_coupled_allocation()
1620 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1621
1622 // Check if a null path was taken unconditionally.
1623 src = null_check(src);
1624 dst = null_check(dst);
1625 if (stopped()) {
1626 return true;
1627 }
1628
1629 // Get length and convert char[] offset to byte[] offset
1630 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1631 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1632
1633 // Range checks
1634 generate_string_range_check(src, src_begin, length, true);
1635 generate_string_range_check(dst, dst_begin, length, false);
1636 if (stopped()) {
1637 return true;
1638 }
1639
1640 if (!stopped()) {
1641
1642 src = shenandoah_read_barrier(src);
1643 dst = shenandoah_write_barrier(dst);
1644
1645 // Calculate starting addresses.
1646 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1647 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1648
1649 // Check if array addresses are aligned to HeapWordSize
1650 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1651 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1652 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1653 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1654
1655 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1656 const char* copyfunc_name = "arraycopy";
1657 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1658 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1659 OptoRuntime::fast_arraycopy_Type(),
1660 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1661 src_start, dst_start, ConvI2X(length) XTOP);
1662 // Do not let reads from the cloned object float above the arraycopy.
1663 if (alloc != NULL) {
1664 if (alloc->maybe_set_complete(&_gvn)) {
1665 // "You break it, you buy it."
1666 InitializeNode* init = alloc->initialization();
1667 assert(init->is_complete(), "we just did this");
1668 init->set_complete_with_arraycopy();
1669 assert(dst->is_CheckCastPP(), "sanity");
1670 assert(dst->in(0)->in(0) == init, "dest pinned");
1671 }
1672 // Do not let stores that initialize this object be reordered with
1673 // a subsequent store that would make this object accessible by
1674 // other threads.
1675 // Record what AllocateNode this StoreStore protects so that
1676 // escape analysis can go from the MemBarStoreStoreNode to the
1677 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1678 // based on the escape status of the AllocateNode.
1679 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1680 } else {
1681 insert_mem_bar(Op_MemBarCPUOrder);
1682 }
1683 }
1684
1685 C->set_has_split_ifs(true); // Has chance for split-if optimization
1686 return true;
1687 }
1688
1689 //----------------------inline_string_char_access----------------------------
1690 // Store/Load char to/from byte[] array.
1691 // static void StringUTF16.putChar(byte[] val, int index, int c)
1692 // static char StringUTF16.getChar(byte[] val, int index)
1693 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1694 Node* value = argument(0);
1695 Node* index = argument(1);
1696 Node* ch = is_store ? argument(2) : NULL;
1697
1698 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1699 // correctly requires matched array shapes.
1700 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1701 "sanity: byte[] and char[] bases agree");
1702 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1703 "sanity: byte[] and char[] scales agree");
1704
1705 // Bail when getChar over constants is requested: constant folding would
1706 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1707 // Java method would constant fold nicely instead.
1708 if (!is_store && value->is_Con() && index->is_Con()) {
1709 return false;
1710 }
1711
1712 value = cast_not_null(value, false);
1713
1714 if (is_store) {
1715 value = shenandoah_write_barrier(value);
1716 } else {
1717 value = shenandoah_read_barrier(value);
1718 }
1719
1720 Node* adr = array_element_address(value, index, T_CHAR);
1721 if (is_store) {
1722 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1723 false, false, true /* mismatched */);
1724 } else {
1725 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1726 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1727 set_result(ch);
1728 }
1729 return true;
1730 }
1731
1732 //--------------------------round_double_node--------------------------------
1733 // Round a double node if necessary.
1734 Node* LibraryCallKit::round_double_node(Node* n) {
1735 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1736 n = _gvn.transform(new RoundDoubleNode(0, n));
1737 return n;
1738 }
1739
1740 //------------------------------inline_math-----------------------------------
1741 // public static double Math.abs(double)
1742 // public static double Math.sqrt(double)
1743 // public static double Math.log(double)
1744 // public static double Math.log10(double)
1745 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1746 Node* arg = round_double_node(argument(0));
1747 Node* n = NULL;
1748 switch (id) {
1749 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1750 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1751 default: fatal_unexpected_iid(id); break;
1752 }
1753 set_result(_gvn.transform(n));
1754 return true;
1755 }
1756
1757 //------------------------------runtime_math-----------------------------
1758 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1759 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1760 "must be (DD)D or (D)D type");
1761
1762 // Inputs
1763 Node* a = round_double_node(argument(0));
1764 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1765
1766 const TypePtr* no_memory_effects = NULL;
1767 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1768 no_memory_effects,
1769 a, top(), b, b ? top() : NULL);
1770 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1771 #ifdef ASSERT
1772 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1773 assert(value_top == top(), "second value must be top");
1774 #endif
1775
1776 set_result(value);
1777 return true;
1778 }
1779
1780 //------------------------------inline_math_native-----------------------------
1781 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1782 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1783 switch (id) {
1784 // These intrinsics are not properly supported on all hardware
1785 case vmIntrinsics::_dsin:
1786 return StubRoutines::dsin() != NULL ?
1787 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1788 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1789 case vmIntrinsics::_dcos:
1790 return StubRoutines::dcos() != NULL ?
1791 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1792 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1793 case vmIntrinsics::_dtan:
1794 return StubRoutines::dtan() != NULL ?
1795 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1796 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1797 case vmIntrinsics::_dlog:
1798 return StubRoutines::dlog() != NULL ?
1799 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1800 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1801 case vmIntrinsics::_dlog10:
1802 return StubRoutines::dlog10() != NULL ?
1803 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1804 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1805
1806 // These intrinsics are supported on all hardware
1807 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1808 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1809
1810 case vmIntrinsics::_dexp:
1811 return StubRoutines::dexp() != NULL ?
1812 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1813 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1814 case vmIntrinsics::_dpow:
1815 return StubRoutines::dpow() != NULL ?
1816 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1817 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1818 #undef FN_PTR
1819
1820 // These intrinsics are not yet correctly implemented
1821 case vmIntrinsics::_datan2:
1822 return false;
1823
1824 default:
1825 fatal_unexpected_iid(id);
1826 return false;
1827 }
1828 }
1829
1830 static bool is_simple_name(Node* n) {
1831 return (n->req() == 1 // constant
1832 || (n->is_Type() && n->as_Type()->type()->singleton())
1833 || n->is_Proj() // parameter or return value
1834 || n->is_Phi() // local of some sort
1835 );
1836 }
1837
1838 //----------------------------inline_notify-----------------------------------*
1839 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1840 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1841 address func;
1842 if (id == vmIntrinsics::_notify) {
1843 func = OptoRuntime::monitor_notify_Java();
1844 } else {
1845 func = OptoRuntime::monitor_notifyAll_Java();
1846 }
1847 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1848 make_slow_call_ex(call, env()->Throwable_klass(), false);
1849 return true;
1850 }
1851
1852
1853 //----------------------------inline_min_max-----------------------------------
1854 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1855 set_result(generate_min_max(id, argument(0), argument(1)));
1856 return true;
1857 }
1858
1859 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1860 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1861 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1862 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1863 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1864
1865 {
1866 PreserveJVMState pjvms(this);
1867 PreserveReexecuteState preexecs(this);
1868 jvms()->set_should_reexecute(true);
1869
1870 set_control(slow_path);
1871 set_i_o(i_o());
1872
1873 uncommon_trap(Deoptimization::Reason_intrinsic,
1874 Deoptimization::Action_none);
1875 }
1876
1877 set_control(fast_path);
1878 set_result(math);
1879 }
1880
1881 template <typename OverflowOp>
1882 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1883 typedef typename OverflowOp::MathOp MathOp;
1884
1885 MathOp* mathOp = new MathOp(arg1, arg2);
1886 Node* operation = _gvn.transform( mathOp );
1887 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1888 inline_math_mathExact(operation, ofcheck);
1889 return true;
1890 }
1891
1892 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1893 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1894 }
1895
1896 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1897 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1898 }
1899
1900 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1901 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1902 }
1903
1904 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1905 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1906 }
1907
1908 bool LibraryCallKit::inline_math_negateExactI() {
1909 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1910 }
1911
1912 bool LibraryCallKit::inline_math_negateExactL() {
1913 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1914 }
1915
1916 bool LibraryCallKit::inline_math_multiplyExactI() {
1917 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1918 }
1919
1920 bool LibraryCallKit::inline_math_multiplyExactL() {
1921 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1922 }
1923
1924 Node*
1925 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1926 // These are the candidate return value:
1927 Node* xvalue = x0;
1928 Node* yvalue = y0;
1929
1930 if (xvalue == yvalue) {
1931 return xvalue;
1932 }
1933
1934 bool want_max = (id == vmIntrinsics::_max);
1935
1936 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1937 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1938 if (txvalue == NULL || tyvalue == NULL) return top();
1939 // This is not really necessary, but it is consistent with a
1940 // hypothetical MaxINode::Value method:
1941 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1942
1943 // %%% This folding logic should (ideally) be in a different place.
1944 // Some should be inside IfNode, and there to be a more reliable
1945 // transformation of ?: style patterns into cmoves. We also want
1946 // more powerful optimizations around cmove and min/max.
1947
1948 // Try to find a dominating comparison of these guys.
1949 // It can simplify the index computation for Arrays.copyOf
1950 // and similar uses of System.arraycopy.
1951 // First, compute the normalized version of CmpI(x, y).
1952 int cmp_op = Op_CmpI;
1953 Node* xkey = xvalue;
1954 Node* ykey = yvalue;
1955 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1956 if (ideal_cmpxy->is_Cmp()) {
1957 // E.g., if we have CmpI(length - offset, count),
1958 // it might idealize to CmpI(length, count + offset)
1959 cmp_op = ideal_cmpxy->Opcode();
1960 xkey = ideal_cmpxy->in(1);
1961 ykey = ideal_cmpxy->in(2);
1962 }
1963
1964 // Start by locating any relevant comparisons.
1965 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1966 Node* cmpxy = NULL;
1967 Node* cmpyx = NULL;
1968 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1969 Node* cmp = start_from->fast_out(k);
1970 if (cmp->outcnt() > 0 && // must have prior uses
1971 cmp->in(0) == NULL && // must be context-independent
1972 cmp->Opcode() == cmp_op) { // right kind of compare
1973 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1974 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1975 }
1976 }
1977
1978 const int NCMPS = 2;
1979 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1980 int cmpn;
1981 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1982 if (cmps[cmpn] != NULL) break; // find a result
1983 }
1984 if (cmpn < NCMPS) {
1985 // Look for a dominating test that tells us the min and max.
1986 int depth = 0; // Limit search depth for speed
1987 Node* dom = control();
1988 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1989 if (++depth >= 100) break;
1990 Node* ifproj = dom;
1991 if (!ifproj->is_Proj()) continue;
1992 Node* iff = ifproj->in(0);
1993 if (!iff->is_If()) continue;
1994 Node* bol = iff->in(1);
1995 if (!bol->is_Bool()) continue;
1996 Node* cmp = bol->in(1);
1997 if (cmp == NULL) continue;
1998 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1999 if (cmps[cmpn] == cmp) break;
2000 if (cmpn == NCMPS) continue;
2001 BoolTest::mask btest = bol->as_Bool()->_test._test;
2002 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2003 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2004 // At this point, we know that 'x btest y' is true.
2005 switch (btest) {
2006 case BoolTest::eq:
2007 // They are proven equal, so we can collapse the min/max.
2008 // Either value is the answer. Choose the simpler.
2009 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2010 return yvalue;
2011 return xvalue;
2012 case BoolTest::lt: // x < y
2013 case BoolTest::le: // x <= y
2014 return (want_max ? yvalue : xvalue);
2015 case BoolTest::gt: // x > y
2016 case BoolTest::ge: // x >= y
2017 return (want_max ? xvalue : yvalue);
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) {
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 (base_type->ptr() == TypePtr::NotNull) {
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 }
2126 // Otherwise, it might either be oop+off or NULL+addr.
2127 return Type::AnyPtr;
2128 } else {
2129 // No information:
2130 return Type::AnyPtr;
2131 }
2132 }
2133
2134 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset, bool is_store) {
2135 int kind = classify_unsafe_addr(base, offset);
2136 if (kind == Type::RawPtr) {
2137 return basic_plus_adr(top(), base, offset);
2138 } else {
2139 if (UseShenandoahGC &&
2140 ((ShenandoahWriteBarrier && is_store) || (ShenandoahReadBarrier && !is_store))) {
2141 if (kind == Type::OopPtr) {
2142 // A cast without a null check should be sufficient here (we
2143 // know base is an oop with a low offset so it can't be null)
2144 // but if there's a dominating null check with both branches
2145 // taken and the cast is pushed in both branches, the cast
2146 // will become top in the null branch but the control flow
2147 // won't go away. Use a null check instead. Worst case, the
2148 // null check becomes an implicit null check with the follow
2149 // barrier and is essentially free.
2150 Node* ctrl = top();
2151 base = null_check_oop(base, &ctrl, true);
2152 if (is_store) {
2153 base = shenandoah_write_barrier(base);
2154 } else {
2155 base = shenandoah_read_barrier(base);
2156 }
2157 } else if (kind == Type::AnyPtr) {
2158 if (UseShenandoahGC &&
2159 _gvn.type(base)->isa_aryptr()) {
2160 Node* ctrl = top();
2161 base = null_check_oop(base, &ctrl, true);
2162 }
2163
2164 if (is_store) {
2165 base = shenandoah_write_barrier(base);
2166 } else {
2167 base = shenandoah_read_barrier(base);
2168 }
2169 }
2170 }
2171 return basic_plus_adr(base, offset);
2172 }
2173 }
2174
2175 //--------------------------inline_number_methods-----------------------------
2176 // inline int Integer.numberOfLeadingZeros(int)
2177 // inline int Long.numberOfLeadingZeros(long)
2178 //
2179 // inline int Integer.numberOfTrailingZeros(int)
2180 // inline int Long.numberOfTrailingZeros(long)
2181 //
2182 // inline int Integer.bitCount(int)
2183 // inline int Long.bitCount(long)
2184 //
2185 // inline char Character.reverseBytes(char)
2186 // inline short Short.reverseBytes(short)
2187 // inline int Integer.reverseBytes(int)
2188 // inline long Long.reverseBytes(long)
2189 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2190 Node* arg = argument(0);
2191 Node* n = NULL;
2192 switch (id) {
2193 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2194 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2195 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2196 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2197 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2198 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2199 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2200 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2201 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2202 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2203 default: fatal_unexpected_iid(id); break;
2204 }
2205 set_result(_gvn.transform(n));
2206 return true;
2207 }
2208
2209 //----------------------------inline_unsafe_access----------------------------
2210
2211 // Helper that guards and inserts a pre-barrier.
2212 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2213 Node* pre_val, bool need_mem_bar) {
2214 // We could be accessing the referent field of a reference object. If so, when G1
2215 // is enabled, we need to log the value in the referent field in an SATB buffer.
2216 // This routine performs some compile time filters and generates suitable
2217 // runtime filters that guard the pre-barrier code.
2218 // Also add memory barrier for non volatile load from the referent field
2219 // to prevent commoning of loads across safepoint.
2220 if (!(UseG1GC || (UseShenandoahGC && ShenandoahWriteBarrier)) && !need_mem_bar)
2221 return;
2222
2223 // Some compile time checks.
2224
2225 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2226 const TypeX* otype = offset->find_intptr_t_type();
2227 if (otype != NULL && otype->is_con() &&
2228 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2229 // Constant offset but not the reference_offset so just return
2230 return;
2231 }
2232
2233 // We only need to generate the runtime guards for instances.
2234 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2235 if (btype != NULL) {
2236 if (btype->isa_aryptr()) {
2237 // Array type so nothing to do
2238 return;
2239 }
2240
2241 const TypeInstPtr* itype = btype->isa_instptr();
2242 if (itype != NULL) {
2243 // Can the klass of base_oop be statically determined to be
2244 // _not_ a sub-class of Reference and _not_ Object?
2245 ciKlass* klass = itype->klass();
2246 if ( klass->is_loaded() &&
2247 !klass->is_subtype_of(env()->Reference_klass()) &&
2248 !env()->Object_klass()->is_subtype_of(klass)) {
2249 return;
2250 }
2251 }
2252 }
2253
2254 // The compile time filters did not reject base_oop/offset so
2255 // we need to generate the following runtime filters
2256 //
2257 // if (offset == java_lang_ref_Reference::_reference_offset) {
2258 // if (instance_of(base, java.lang.ref.Reference)) {
2259 // pre_barrier(_, pre_val, ...);
2260 // }
2261 // }
2262
2263 float likely = PROB_LIKELY( 0.999);
2264 float unlikely = PROB_UNLIKELY(0.999);
2265
2266 IdealKit ideal(this);
2267 #define __ ideal.
2268
2269 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2270
2271 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2272 // Update graphKit memory and control from IdealKit.
2273 sync_kit(ideal);
2274
2275 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2276 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2277
2278 // Update IdealKit memory and control from graphKit.
2279 __ sync_kit(this);
2280
2281 Node* one = __ ConI(1);
2282 // is_instof == 0 if base_oop == NULL
2283 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2284
2285 // Update graphKit from IdeakKit.
2286 sync_kit(ideal);
2287
2288 // Use the pre-barrier to record the value in the referent field
2289 pre_barrier(false /* do_load */,
2290 __ ctrl(),
2291 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2292 pre_val /* pre_val */,
2293 T_OBJECT);
2294 if (need_mem_bar) {
2295 // Add memory barrier to prevent commoning reads from this field
2296 // across safepoint since GC can change its value.
2297 insert_mem_bar(Op_MemBarCPUOrder);
2298 }
2299 // Update IdealKit from graphKit.
2300 __ sync_kit(this);
2301
2302 } __ end_if(); // _ref_type != ref_none
2303 } __ end_if(); // offset == referent_offset
2304
2305 // Final sync IdealKit and GraphKit.
2306 final_sync(ideal);
2307 #undef __
2308 }
2309
2310
2311 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2312 // Attempt to infer a sharper value type from the offset and base type.
2313 ciKlass* sharpened_klass = NULL;
2314
2315 // See if it is an instance field, with an object type.
2316 if (alias_type->field() != NULL) {
2317 if (alias_type->field()->type()->is_klass()) {
2318 sharpened_klass = alias_type->field()->type()->as_klass();
2319 }
2320 }
2321
2322 // See if it is a narrow oop array.
2323 if (adr_type->isa_aryptr()) {
2324 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2325 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2326 if (elem_type != NULL) {
2327 sharpened_klass = elem_type->klass();
2328 }
2329 }
2330 }
2331
2332 // The sharpened class might be unloaded if there is no class loader
2333 // contraint in place.
2334 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2335 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2336
2337 #ifndef PRODUCT
2338 if (C->print_intrinsics() || C->print_inlining()) {
2339 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2340 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2341 }
2342 #endif
2343 // Sharpen the value type.
2344 return tjp;
2345 }
2346 return NULL;
2347 }
2348
2349 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2350 if (callee()->is_static()) return false; // caller must have the capability!
2351 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2352 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2353 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2354
2355 #ifndef PRODUCT
2356 {
2357 ResourceMark rm;
2358 // Check the signatures.
2359 ciSignature* sig = callee()->signature();
2360 #ifdef ASSERT
2361 if (!is_store) {
2362 // Object getObject(Object base, int/long offset), etc.
2363 BasicType rtype = sig->return_type()->basic_type();
2364 assert(rtype == type, "getter must return the expected value");
2365 assert(sig->count() == 2, "oop getter has 2 arguments");
2366 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2367 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2368 } else {
2369 // void putObject(Object base, int/long offset, Object x), etc.
2370 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2371 assert(sig->count() == 3, "oop putter has 3 arguments");
2372 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2373 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2374 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2375 assert(vtype == type, "putter must accept the expected value");
2376 }
2377 #endif // ASSERT
2378 }
2379 #endif //PRODUCT
2380
2381 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2382
2383 Node* receiver = argument(0); // type: oop
2384
2385 // Build address expression.
2386 Node* adr;
2387 Node* heap_base_oop = top();
2388 Node* offset = top();
2389 Node* val;
2390
2391 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2392 Node* base = argument(1); // type: oop
2393 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2394 offset = argument(2); // type: long
2395 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2396 // to be plain byte offsets, which are also the same as those accepted
2397 // by oopDesc::field_base.
2398 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2399 "fieldOffset must be byte-scaled");
2400 // 32-bit machines ignore the high half!
2401 offset = ConvL2X(offset);
2402 adr = make_unsafe_address(base, offset, is_store);
2403 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2404 heap_base_oop = base;
2405 } else if (type == T_OBJECT) {
2406 return false; // off-heap oop accesses are not supported
2407 }
2408
2409 // Can base be NULL? Otherwise, always on-heap access.
2410 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2411
2412 val = is_store ? argument(4) : NULL;
2413
2414 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2415
2416 if (UseShenandoahGC && adr->is_AddP() &&
2417 adr->in(AddPNode::Base) == adr->in(AddPNode::Address)) {
2418 Node* base = ShenandoahBarrierNode::skip_through_barrier(adr->in(AddPNode::Base));
2419 const TypeInstPtr* base_t = _gvn.type(base)->isa_instptr();
2420 if (base_t != NULL &&
2421 base_t->const_oop() != NULL &&
2422 base_t->klass() == ciEnv::current()->Class_klass() &&
2423 adr_type->is_instptr()->offset() >= base_t->klass()->as_instance_klass()->size_helper() * wordSize) {
2424 adr_type = base_t->add_offset(adr_type->is_instptr()->offset());
2425 }
2426 }
2427
2428 // Try to categorize the address.
2429 Compile::AliasType* alias_type = C->alias_type(adr_type);
2430 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2431
2432 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2433 alias_type->adr_type() == TypeAryPtr::RANGE) {
2434 return false; // not supported
2435 }
2436
2437 bool mismatched = false;
2438 BasicType bt = alias_type->basic_type();
2439 if (bt != T_ILLEGAL) {
2440 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2441 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2442 // Alias type doesn't differentiate between byte[] and boolean[]).
2443 // Use address type to get the element type.
2444 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2445 }
2446 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2447 // accessing an array field with getObject is not a mismatch
2448 bt = T_OBJECT;
2449 }
2450 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2451 // Don't intrinsify mismatched object accesses
2452 return false;
2453 }
2454 mismatched = (bt != type);
2455 } else if (alias_type->adr_type()->isa_oopptr()) {
2456 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2457 }
2458
2459 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2460
2461 // First guess at the value type.
2462 const Type *value_type = Type::get_const_basic_type(type);
2463
2464 // We will need memory barriers unless we can determine a unique
2465 // alias category for this reference. (Note: If for some reason
2466 // the barriers get omitted and the unsafe reference begins to "pollute"
2467 // the alias analysis of the rest of the graph, either Compile::can_alias
2468 // or Compile::must_alias will throw a diagnostic assert.)
2469 bool need_mem_bar;
2470 switch (kind) {
2471 case Relaxed:
2472 need_mem_bar = mismatched || can_access_non_heap;
2473 break;
2474 case Opaque:
2475 // Opaque uses CPUOrder membars for protection against code movement.
2476 case Acquire:
2477 case Release:
2478 case Volatile:
2479 need_mem_bar = true;
2480 break;
2481 default:
2482 ShouldNotReachHere();
2483 }
2484
2485 // Some accesses require access atomicity for all types, notably longs and doubles.
2486 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2487 bool requires_atomic_access = false;
2488 switch (kind) {
2489 case Relaxed:
2490 requires_atomic_access = AlwaysAtomicAccesses;
2491 break;
2492 case Opaque:
2493 // Opaque accesses are atomic.
2494 case Acquire:
2495 case Release:
2496 case Volatile:
2497 requires_atomic_access = true;
2498 break;
2499 default:
2500 ShouldNotReachHere();
2501 }
2502
2503 // Figure out the memory ordering.
2504 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2505 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2506
2507 // If we are reading the value of the referent field of a Reference
2508 // object (either by using Unsafe directly or through reflection)
2509 // then, if G1 is enabled, we need to record the referent in an
2510 // SATB log buffer using the pre-barrier mechanism.
2511 // Also we need to add memory barrier to prevent commoning reads
2512 // from this field across safepoint since GC can change its value.
2513 bool need_read_barrier = !is_store &&
2514 offset != top() && heap_base_oop != top();
2515
2516 if (!is_store && type == T_OBJECT) {
2517 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2518 if (tjp != NULL) {
2519 value_type = tjp;
2520 }
2521 }
2522
2523 receiver = null_check(receiver);
2524 if (stopped()) {
2525 return true;
2526 }
2527 // Heap pointers get a null-check from the interpreter,
2528 // as a courtesy. However, this is not guaranteed by Unsafe,
2529 // and it is not possible to fully distinguish unintended nulls
2530 // from intended ones in this API.
2531
2532 // We need to emit leading and trailing CPU membars (see below) in
2533 // addition to memory membars for special access modes. This is a little
2534 // too strong, but avoids the need to insert per-alias-type
2535 // volatile membars (for stores; compare Parse::do_put_xxx), which
2536 // we cannot do effectively here because we probably only have a
2537 // rough approximation of type.
2538
2539 switch(kind) {
2540 case Relaxed:
2541 case Opaque:
2542 case Acquire:
2543 break;
2544 case Release:
2545 case Volatile:
2546 if (is_store) {
2547 insert_mem_bar(Op_MemBarRelease);
2548 } else {
2549 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2550 insert_mem_bar(Op_MemBarVolatile);
2551 }
2552 }
2553 break;
2554 default:
2555 ShouldNotReachHere();
2556 }
2557
2558 // Memory barrier to prevent normal and 'unsafe' accesses from
2559 // bypassing each other. Happens after null checks, so the
2560 // exception paths do not take memory state from the memory barrier,
2561 // so there's no problems making a strong assert about mixing users
2562 // of safe & unsafe memory.
2563 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2564
2565 if (!is_store) {
2566 Node* p = NULL;
2567 // Try to constant fold a load from a constant field
2568 ciField* field = alias_type->field();
2569 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2570 // final or stable field
2571 p = make_constant_from_field(field, heap_base_oop);
2572 }
2573 if (p == NULL) {
2574 // To be valid, unsafe loads may depend on other conditions than
2575 // the one that guards them: pin the Load node
2576 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, requires_atomic_access, unaligned, mismatched);
2577 // load value
2578 switch (type) {
2579 case T_BOOLEAN:
2580 {
2581 // Normalize the value returned by getBoolean in the following cases
2582 if (mismatched ||
2583 heap_base_oop == top() || // - heap_base_oop is NULL or
2584 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2585 // and the unsafe access is made to large offset
2586 // (i.e., larger than the maximum offset necessary for any
2587 // field access)
2588 ) {
2589 IdealKit ideal = IdealKit(this);
2590 #define __ ideal.
2591 IdealVariable normalized_result(ideal);
2592 __ declarations_done();
2593 __ set(normalized_result, p);
2594 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2595 __ set(normalized_result, ideal.ConI(1));
2596 ideal.end_if();
2597 final_sync(ideal);
2598 p = __ value(normalized_result);
2599 #undef __
2600 }
2601 }
2602 case T_CHAR:
2603 case T_BYTE:
2604 case T_SHORT:
2605 case T_INT:
2606 case T_LONG:
2607 case T_FLOAT:
2608 case T_DOUBLE:
2609 break;
2610 case T_OBJECT:
2611 if (need_read_barrier) {
2612 // We do not require a mem bar inside pre_barrier if need_mem_bar
2613 // is set: the barriers would be emitted by us.
2614 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2615 }
2616 break;
2617 case T_ADDRESS:
2618 // Cast to an int type.
2619 p = _gvn.transform(new CastP2XNode(NULL, p));
2620 p = ConvX2UL(p);
2621 break;
2622 default:
2623 fatal("unexpected type %d: %s", type, type2name(type));
2624 break;
2625 }
2626 }
2627 // The load node has the control of the preceding MemBarCPUOrder. All
2628 // following nodes will have the control of the MemBarCPUOrder inserted at
2629 // the end of this method. So, pushing the load onto the stack at a later
2630 // point is fine.
2631 set_result(p);
2632 } else {
2633 // place effect of store into memory
2634 switch (type) {
2635 case T_DOUBLE:
2636 val = dstore_rounding(val);
2637 break;
2638 case T_ADDRESS:
2639 // Repackage the long as a pointer.
2640 val = ConvL2X(val);
2641 val = _gvn.transform(new CastX2PNode(val));
2642 break;
2643 }
2644
2645 if (type == T_OBJECT) {
2646 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2647 } else {
2648 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2649 }
2650 }
2651
2652 switch(kind) {
2653 case Relaxed:
2654 case Opaque:
2655 case Release:
2656 break;
2657 case Acquire:
2658 case Volatile:
2659 if (!is_store) {
2660 insert_mem_bar(Op_MemBarAcquire);
2661 } else {
2662 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2663 insert_mem_bar(Op_MemBarVolatile);
2664 }
2665 }
2666 break;
2667 default:
2668 ShouldNotReachHere();
2669 }
2670
2671 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2672
2673 return true;
2674 }
2675
2676 //----------------------------inline_unsafe_load_store----------------------------
2677 // This method serves a couple of different customers (depending on LoadStoreKind):
2678 //
2679 // LS_cmp_swap:
2680 //
2681 // boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2682 // boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2683 // boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2684 //
2685 // LS_cmp_swap_weak:
2686 //
2687 // boolean weakCompareAndSwapObject( Object o, long offset, Object expected, Object x);
2688 // boolean weakCompareAndSwapObjectAcquire(Object o, long offset, Object expected, Object x);
2689 // boolean weakCompareAndSwapObjectRelease(Object o, long offset, Object expected, Object x);
2690 //
2691 // boolean weakCompareAndSwapInt( Object o, long offset, int expected, int x);
2692 // boolean weakCompareAndSwapIntAcquire( Object o, long offset, int expected, int x);
2693 // boolean weakCompareAndSwapIntRelease( Object o, long offset, int expected, int x);
2694 //
2695 // boolean weakCompareAndSwapLong( Object o, long offset, long expected, long x);
2696 // boolean weakCompareAndSwapLongAcquire( Object o, long offset, long expected, long x);
2697 // boolean weakCompareAndSwapLongRelease( Object o, long offset, long expected, long x);
2698 //
2699 // LS_cmp_exchange:
2700 //
2701 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2702 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2703 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2704 //
2705 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2706 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2707 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2708 //
2709 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2710 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2711 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2712 //
2713 // LS_get_add:
2714 //
2715 // int getAndAddInt( Object o, long offset, int delta)
2716 // long getAndAddLong(Object o, long offset, long delta)
2717 //
2718 // LS_get_set:
2719 //
2720 // int getAndSet(Object o, long offset, int newValue)
2721 // long getAndSet(Object o, long offset, long newValue)
2722 // Object getAndSet(Object o, long offset, Object newValue)
2723 //
2724 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2725 // This basic scheme here is the same as inline_unsafe_access, but
2726 // differs in enough details that combining them would make the code
2727 // overly confusing. (This is a true fact! I originally combined
2728 // them, but even I was confused by it!) As much code/comments as
2729 // possible are retained from inline_unsafe_access though to make
2730 // the correspondences clearer. - dl
2731
2732 if (callee()->is_static()) return false; // caller must have the capability!
2733
2734 #ifndef PRODUCT
2735 BasicType rtype;
2736 {
2737 ResourceMark rm;
2738 // Check the signatures.
2739 ciSignature* sig = callee()->signature();
2740 rtype = sig->return_type()->basic_type();
2741 switch(kind) {
2742 case LS_get_add:
2743 case LS_get_set: {
2744 // Check the signatures.
2745 #ifdef ASSERT
2746 assert(rtype == type, "get and set must return the expected type");
2747 assert(sig->count() == 3, "get and set has 3 arguments");
2748 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2749 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2750 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2751 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2752 #endif // ASSERT
2753 break;
2754 }
2755 case LS_cmp_swap:
2756 case LS_cmp_swap_weak: {
2757 // Check the signatures.
2758 #ifdef ASSERT
2759 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2760 assert(sig->count() == 4, "CAS has 4 arguments");
2761 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2762 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2763 #endif // ASSERT
2764 break;
2765 }
2766 case LS_cmp_exchange: {
2767 // Check the signatures.
2768 #ifdef ASSERT
2769 assert(rtype == type, "CAS must return the expected type");
2770 assert(sig->count() == 4, "CAS has 4 arguments");
2771 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2772 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2773 #endif // ASSERT
2774 break;
2775 }
2776 default:
2777 ShouldNotReachHere();
2778 }
2779 }
2780 #endif //PRODUCT
2781
2782 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2783
2784 // Get arguments:
2785 Node* receiver = NULL;
2786 Node* base = NULL;
2787 Node* offset = NULL;
2788 Node* oldval = NULL;
2789 Node* newval = NULL;
2790 switch(kind) {
2791 case LS_cmp_swap:
2792 case LS_cmp_swap_weak:
2793 case LS_cmp_exchange: {
2794 const bool two_slot_type = type2size[type] == 2;
2795 receiver = argument(0); // type: oop
2796 base = argument(1); // type: oop
2797 offset = argument(2); // type: long
2798 oldval = argument(4); // type: oop, int, or long
2799 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2800 break;
2801 }
2802 case LS_get_add:
2803 case LS_get_set: {
2804 receiver = argument(0); // type: oop
2805 base = argument(1); // type: oop
2806 offset = argument(2); // type: long
2807 oldval = NULL;
2808 newval = argument(4); // type: oop, int, or long
2809 break;
2810 }
2811 default:
2812 ShouldNotReachHere();
2813 }
2814
2815 // Build field offset expression.
2816 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2817 // to be plain byte offsets, which are also the same as those accepted
2818 // by oopDesc::field_base.
2819 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2820 // 32-bit machines ignore the high half of long offsets
2821 offset = ConvL2X(offset);
2822 Node* adr = make_unsafe_address(base, offset, true);
2823 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2824
2825 Compile::AliasType* alias_type = C->alias_type(adr_type);
2826 BasicType bt = alias_type->basic_type();
2827 if (bt != T_ILLEGAL &&
2828 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2829 // Don't intrinsify mismatched object accesses.
2830 return false;
2831 }
2832
2833 // For CAS, unlike inline_unsafe_access, there seems no point in
2834 // trying to refine types. Just use the coarse types here.
2835 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2836 const Type *value_type = Type::get_const_basic_type(type);
2837
2838 switch (kind) {
2839 case LS_get_set:
2840 case LS_cmp_exchange: {
2841 if (type == T_OBJECT) {
2842 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2843 if (tjp != NULL) {
2844 value_type = tjp;
2845 }
2846 }
2847 break;
2848 }
2849 case LS_cmp_swap:
2850 case LS_cmp_swap_weak:
2851 case LS_get_add:
2852 break;
2853 default:
2854 ShouldNotReachHere();
2855 }
2856
2857 // Null check receiver.
2858 receiver = null_check(receiver);
2859 if (stopped()) {
2860 return true;
2861 }
2862
2863 int alias_idx = C->get_alias_index(adr_type);
2864
2865 // Memory-model-wise, a LoadStore acts like a little synchronized
2866 // block, so needs barriers on each side. These don't translate
2867 // into actual barriers on most machines, but we still need rest of
2868 // compiler to respect ordering.
2869
2870 switch (access_kind) {
2871 case Relaxed:
2872 case Acquire:
2873 break;
2874 case Release:
2875 insert_mem_bar(Op_MemBarRelease);
2876 break;
2877 case Volatile:
2878 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2879 insert_mem_bar(Op_MemBarVolatile);
2880 } else {
2881 insert_mem_bar(Op_MemBarRelease);
2882 }
2883 break;
2884 default:
2885 ShouldNotReachHere();
2886 }
2887 insert_mem_bar(Op_MemBarCPUOrder);
2888
2889 // Figure out the memory ordering.
2890 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2891
2892 // 4984716: MemBars must be inserted before this
2893 // memory node in order to avoid a false
2894 // dependency which will confuse the scheduler.
2895 Node *mem = memory(alias_idx);
2896
2897 // For now, we handle only those cases that actually exist: ints,
2898 // longs, and Object. Adding others should be straightforward.
2899 Node* load_store = NULL;
2900 switch(type) {
2901 case T_BYTE:
2902 switch(kind) {
2903 case LS_get_add:
2904 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2905 break;
2906 case LS_get_set:
2907 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2908 break;
2909 case LS_cmp_swap_weak:
2910 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2911 break;
2912 case LS_cmp_swap:
2913 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2914 break;
2915 case LS_cmp_exchange:
2916 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2917 break;
2918 default:
2919 ShouldNotReachHere();
2920 }
2921 break;
2922 case T_SHORT:
2923 switch(kind) {
2924 case LS_get_add:
2925 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2926 break;
2927 case LS_get_set:
2928 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2929 break;
2930 case LS_cmp_swap_weak:
2931 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2932 break;
2933 case LS_cmp_swap:
2934 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2935 break;
2936 case LS_cmp_exchange:
2937 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2938 break;
2939 default:
2940 ShouldNotReachHere();
2941 }
2942 break;
2943 case T_INT:
2944 switch(kind) {
2945 case LS_get_add:
2946 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2947 break;
2948 case LS_get_set:
2949 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2950 break;
2951 case LS_cmp_swap_weak:
2952 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2953 break;
2954 case LS_cmp_swap:
2955 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2956 break;
2957 case LS_cmp_exchange:
2958 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2959 break;
2960 default:
2961 ShouldNotReachHere();
2962 }
2963 break;
2964 case T_LONG:
2965 switch(kind) {
2966 case LS_get_add:
2967 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2968 break;
2969 case LS_get_set:
2970 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2971 break;
2972 case LS_cmp_swap_weak:
2973 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2974 break;
2975 case LS_cmp_swap:
2976 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2977 break;
2978 case LS_cmp_exchange:
2979 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
2980 break;
2981 default:
2982 ShouldNotReachHere();
2983 }
2984 break;
2985 case T_OBJECT:
2986 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2987 // could be delayed during Parse (for example, in adjust_map_after_if()).
2988 // Execute transformation here to avoid barrier generation in such case.
2989 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2990 newval = _gvn.makecon(TypePtr::NULL_PTR);
2991
2992 // Reference stores need a store barrier.
2993 switch(kind) {
2994 case LS_get_set: {
2995 // If pre-barrier must execute before the oop store, old value will require do_load here.
2996 if (!can_move_pre_barrier()) {
2997 newval = pre_barrier(true /* do_load*/,
2998 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2999 NULL /* pre_val*/,
3000 T_OBJECT);
3001 } // Else move pre_barrier to use load_store value, see below.
3002 else {
3003 newval = shenandoah_read_barrier_storeval(newval);
3004 }
3005 break;
3006 }
3007 case LS_cmp_swap_weak:
3008 case LS_cmp_swap:
3009 case LS_cmp_exchange: {
3010 // Same as for newval above:
3011 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3012 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3013 }
3014 // The only known value which might get overwritten is oldval.
3015 newval = pre_barrier(false /* do_load */,
3016 control(), NULL, NULL, max_juint, newval, NULL,
3017 oldval /* pre_val */,
3018 T_OBJECT);
3019 break;
3020 }
3021 default:
3022 ShouldNotReachHere();
3023 }
3024
3025 #ifdef _LP64
3026 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3027 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3028
3029 switch(kind) {
3030 case LS_get_set:
3031 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3032 break;
3033 case LS_cmp_swap_weak: {
3034 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3035 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3036 break;
3037 }
3038 case LS_cmp_swap: {
3039 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3040 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3041 break;
3042 }
3043 case LS_cmp_exchange: {
3044 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3045 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3046 break;
3047 }
3048 default:
3049 ShouldNotReachHere();
3050 }
3051 } else
3052 #endif
3053 switch (kind) {
3054 case LS_get_set:
3055 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3056 break;
3057 case LS_cmp_swap_weak:
3058 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3059 break;
3060 case LS_cmp_swap:
3061 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3062 break;
3063 case LS_cmp_exchange:
3064 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3065 break;
3066 default:
3067 ShouldNotReachHere();
3068 }
3069
3070 // Emit the post barrier only when the actual store happened. This makes sense
3071 // to check only for LS_cmp_* that can fail to set the value.
3072 // LS_cmp_exchange does not produce any branches by default, so there is no
3073 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3074 // CompareAndExchange and move branches here, it would make sense to conditionalize
3075 // post_barriers for LS_cmp_exchange as well.
3076 //
3077 // CAS success path is marked more likely since we anticipate this is a performance
3078 // critical path, while CAS failure path can use the penalty for going through unlikely
3079 // path as backoff. Which is still better than doing a store barrier there.
3080 switch (kind) {
3081 case LS_get_set:
3082 case LS_cmp_exchange: {
3083 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3084 break;
3085 }
3086 case LS_cmp_swap_weak:
3087 case LS_cmp_swap: {
3088 IdealKit ideal(this);
3089 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3090 sync_kit(ideal);
3091 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3092 ideal.sync_kit(this);
3093 } ideal.end_if();
3094 final_sync(ideal);
3095 break;
3096 }
3097 default:
3098 ShouldNotReachHere();
3099 }
3100 break;
3101 default:
3102 fatal("unexpected type %d: %s", type, type2name(type));
3103 break;
3104 }
3105
3106 // SCMemProjNodes represent the memory state of a LoadStore. Their
3107 // main role is to prevent LoadStore nodes from being optimized away
3108 // when their results aren't used.
3109 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3110 set_memory(proj, alias_idx);
3111
3112 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3113 #ifdef _LP64
3114 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3115 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3116 }
3117 #endif
3118 if (can_move_pre_barrier()) {
3119 // Don't need to load pre_val. The old value is returned by load_store.
3120 // The pre_barrier can execute after the xchg as long as no safepoint
3121 // gets inserted between them.
3122 pre_barrier(false /* do_load */,
3123 control(), NULL, NULL, max_juint, NULL, NULL,
3124 load_store /* pre_val */,
3125 T_OBJECT);
3126 }
3127 }
3128
3129 // Add the trailing membar surrounding the access
3130 insert_mem_bar(Op_MemBarCPUOrder);
3131
3132 switch (access_kind) {
3133 case Relaxed:
3134 case Release:
3135 break; // do nothing
3136 case Acquire:
3137 case Volatile:
3138 insert_mem_bar(Op_MemBarAcquire);
3139 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3140 break;
3141 default:
3142 ShouldNotReachHere();
3143 }
3144
3145 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3146 set_result(load_store);
3147 return true;
3148 }
3149
3150 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3151 MemNode::MemOrd mo = MemNode::unset;
3152 switch(kind) {
3153 case Opaque:
3154 case Relaxed: mo = MemNode::unordered; break;
3155 case Acquire: mo = MemNode::acquire; break;
3156 case Release: mo = MemNode::release; break;
3157 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3158 default:
3159 ShouldNotReachHere();
3160 }
3161 guarantee(mo != MemNode::unset, "Should select memory ordering");
3162 return mo;
3163 }
3164
3165 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3166 MemNode::MemOrd mo = MemNode::unset;
3167 switch(kind) {
3168 case Opaque:
3169 case Relaxed: mo = MemNode::unordered; break;
3170 case Acquire: mo = MemNode::acquire; break;
3171 case Release: mo = MemNode::release; break;
3172 case Volatile: mo = MemNode::seqcst; break;
3173 default:
3174 ShouldNotReachHere();
3175 }
3176 guarantee(mo != MemNode::unset, "Should select memory ordering");
3177 return mo;
3178 }
3179
3180 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3181 // Regardless of form, don't allow previous ld/st to move down,
3182 // then issue acquire, release, or volatile mem_bar.
3183 insert_mem_bar(Op_MemBarCPUOrder);
3184 switch(id) {
3185 case vmIntrinsics::_loadFence:
3186 insert_mem_bar(Op_LoadFence);
3187 return true;
3188 case vmIntrinsics::_storeFence:
3189 insert_mem_bar(Op_StoreFence);
3190 return true;
3191 case vmIntrinsics::_fullFence:
3192 insert_mem_bar(Op_MemBarVolatile);
3193 return true;
3194 default:
3195 fatal_unexpected_iid(id);
3196 return false;
3197 }
3198 }
3199
3200 bool LibraryCallKit::inline_onspinwait() {
3201 insert_mem_bar(Op_OnSpinWait);
3202 return true;
3203 }
3204
3205 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3206 if (!kls->is_Con()) {
3207 return true;
3208 }
3209 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3210 if (klsptr == NULL) {
3211 return true;
3212 }
3213 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3214 // don't need a guard for a klass that is already initialized
3215 return !ik->is_initialized();
3216 }
3217
3218 //----------------------------inline_unsafe_allocate---------------------------
3219 // public native Object Unsafe.allocateInstance(Class<?> cls);
3220 bool LibraryCallKit::inline_unsafe_allocate() {
3221 if (callee()->is_static()) return false; // caller must have the capability!
3222
3223 null_check_receiver(); // null-check, then ignore
3224 Node* cls = null_check(argument(1));
3225 if (stopped()) return true;
3226
3227 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3228 kls = null_check(kls);
3229 if (stopped()) return true; // argument was like int.class
3230
3231 Node* test = NULL;
3232 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3233 // Note: The argument might still be an illegal value like
3234 // Serializable.class or Object[].class. The runtime will handle it.
3235 // But we must make an explicit check for initialization.
3236 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3237 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3238 // can generate code to load it as unsigned byte.
3239 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3240 Node* bits = intcon(InstanceKlass::fully_initialized);
3241 test = _gvn.transform(new SubINode(inst, bits));
3242 // The 'test' is non-zero if we need to take a slow path.
3243 }
3244
3245 Node* obj = new_instance(kls, test);
3246 set_result(obj);
3247 return true;
3248 }
3249
3250 //------------------------inline_native_time_funcs--------------
3251 // inline code for System.currentTimeMillis() and System.nanoTime()
3252 // these have the same type and signature
3253 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3254 const TypeFunc* tf = OptoRuntime::void_long_Type();
3255 const TypePtr* no_memory_effects = NULL;
3256 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3257 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3258 #ifdef ASSERT
3259 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3260 assert(value_top == top(), "second value must be top");
3261 #endif
3262 set_result(value);
3263 return true;
3264 }
3265
3266 #ifdef TRACE_HAVE_INTRINSICS
3267
3268 /*
3269 * oop -> myklass
3270 * myklass->trace_id |= USED
3271 * return myklass->trace_id & ~0x3
3272 */
3273 bool LibraryCallKit::inline_native_classID() {
3274 Node* cls = null_check(argument(0), T_OBJECT);
3275 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3276 kls = null_check(kls, T_OBJECT);
3277
3278 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3279 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3280 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3281
3282 Node* clsused = longcon(0x01l); // set the class bit
3283 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3284 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3285 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3286
3287 #ifdef TRACE_ID_META_BITS
3288 Node* mbits = longcon(~TRACE_ID_META_BITS);
3289 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3290 #endif
3291 #ifdef TRACE_ID_CLASS_SHIFT
3292 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3293 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3294 #endif
3295
3296 set_result(tvalue);
3297 return true;
3298
3299 }
3300
3301 bool LibraryCallKit::inline_native_getBufferWriter() {
3302 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3303
3304 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3305 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3306 );
3307
3308 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3309
3310 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3311 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3312
3313 IfNode* iff_jobj_null =
3314 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3315
3316 enum { _normal_path = 1,
3317 _null_path = 2,
3318 PATH_LIMIT };
3319
3320 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3321 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3322
3323 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3324 result_rgn->init_req(_null_path, jobj_is_null);
3325 result_val->init_req(_null_path, null());
3326
3327 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3328 result_rgn->init_req(_normal_path, jobj_is_not_null);
3329
3330 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3331 result_val->init_req(_normal_path, res);
3332
3333 set_result(result_rgn, result_val);
3334
3335 return true;
3336 }
3337
3338 #endif
3339
3340 //------------------------inline_native_currentThread------------------
3341 bool LibraryCallKit::inline_native_currentThread() {
3342 Node* junk = NULL;
3343 set_result(generate_current_thread(junk));
3344 return true;
3345 }
3346
3347 //------------------------inline_native_isInterrupted------------------
3348 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3349 bool LibraryCallKit::inline_native_isInterrupted() {
3350 // Add a fast path to t.isInterrupted(clear_int):
3351 // (t == Thread.current() &&
3352 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3353 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3354 // So, in the common case that the interrupt bit is false,
3355 // we avoid making a call into the VM. Even if the interrupt bit
3356 // is true, if the clear_int argument is false, we avoid the VM call.
3357 // However, if the receiver is not currentThread, we must call the VM,
3358 // because there must be some locking done around the operation.
3359
3360 // We only go to the fast case code if we pass two guards.
3361 // Paths which do not pass are accumulated in the slow_region.
3362
3363 enum {
3364 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3365 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3366 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3367 PATH_LIMIT
3368 };
3369
3370 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3371 // out of the function.
3372 insert_mem_bar(Op_MemBarCPUOrder);
3373
3374 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3375 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3376
3377 RegionNode* slow_region = new RegionNode(1);
3378 record_for_igvn(slow_region);
3379
3380 // (a) Receiving thread must be the current thread.
3381 Node* rec_thr = argument(0);
3382 Node* tls_ptr = NULL;
3383 Node* cur_thr = generate_current_thread(tls_ptr);
3384
3385 Node* cmp_thr = cmp_objects(cur_thr, rec_thr);
3386 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3387
3388 generate_slow_guard(bol_thr, slow_region);
3389
3390 // (b) Interrupt bit on TLS must be false.
3391 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3392 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3393 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3394
3395 // Set the control input on the field _interrupted read to prevent it floating up.
3396 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3397 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3398 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3399
3400 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3401
3402 // First fast path: if (!TLS._interrupted) return false;
3403 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3404 result_rgn->init_req(no_int_result_path, false_bit);
3405 result_val->init_req(no_int_result_path, intcon(0));
3406
3407 // drop through to next case
3408 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3409
3410 #ifndef _WINDOWS
3411 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3412 Node* clr_arg = argument(1);
3413 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3414 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3415 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3416
3417 // Second fast path: ... else if (!clear_int) return true;
3418 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3419 result_rgn->init_req(no_clear_result_path, false_arg);
3420 result_val->init_req(no_clear_result_path, intcon(1));
3421
3422 // drop through to next case
3423 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3424 #else
3425 // To return true on Windows you must read the _interrupted field
3426 // and check the event state i.e. take the slow path.
3427 #endif // _WINDOWS
3428
3429 // (d) Otherwise, go to the slow path.
3430 slow_region->add_req(control());
3431 set_control( _gvn.transform(slow_region));
3432
3433 if (stopped()) {
3434 // There is no slow path.
3435 result_rgn->init_req(slow_result_path, top());
3436 result_val->init_req(slow_result_path, top());
3437 } else {
3438 // non-virtual because it is a private non-static
3439 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3440
3441 Node* slow_val = set_results_for_java_call(slow_call);
3442 // this->control() comes from set_results_for_java_call
3443
3444 Node* fast_io = slow_call->in(TypeFunc::I_O);
3445 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3446
3447 // These two phis are pre-filled with copies of of the fast IO and Memory
3448 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3449 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3450
3451 result_rgn->init_req(slow_result_path, control());
3452 result_io ->init_req(slow_result_path, i_o());
3453 result_mem->init_req(slow_result_path, reset_memory());
3454 result_val->init_req(slow_result_path, slow_val);
3455
3456 set_all_memory(_gvn.transform(result_mem));
3457 set_i_o( _gvn.transform(result_io));
3458 }
3459
3460 C->set_has_split_ifs(true); // Has chance for split-if optimization
3461 set_result(result_rgn, result_val);
3462 return true;
3463 }
3464
3465 //---------------------------load_mirror_from_klass----------------------------
3466 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3467 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3468 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3469 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3470 }
3471
3472 //-----------------------load_klass_from_mirror_common-------------------------
3473 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3474 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3475 // and branch to the given path on the region.
3476 // If never_see_null, take an uncommon trap on null, so we can optimistically
3477 // compile for the non-null case.
3478 // If the region is NULL, force never_see_null = true.
3479 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3480 bool never_see_null,
3481 RegionNode* region,
3482 int null_path,
3483 int offset) {
3484 if (region == NULL) never_see_null = true;
3485 Node* p = basic_plus_adr(mirror, offset);
3486 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3487 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3488 Node* null_ctl = top();
3489 kls = null_check_oop(kls, &null_ctl, never_see_null);
3490 if (region != NULL) {
3491 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3492 region->init_req(null_path, null_ctl);
3493 } else {
3494 assert(null_ctl == top(), "no loose ends");
3495 }
3496 return kls;
3497 }
3498
3499 //--------------------(inline_native_Class_query helpers)---------------------
3500 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3501 // Fall through if (mods & mask) == bits, take the guard otherwise.
3502 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3503 // Branch around if the given klass has the given modifier bit set.
3504 // Like generate_guard, adds a new path onto the region.
3505 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3506 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3507 Node* mask = intcon(modifier_mask);
3508 Node* bits = intcon(modifier_bits);
3509 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3510 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3511 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3512 return generate_fair_guard(bol, region);
3513 }
3514 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3515 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3516 }
3517
3518 //-------------------------inline_native_Class_query-------------------
3519 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3520 const Type* return_type = TypeInt::BOOL;
3521 Node* prim_return_value = top(); // what happens if it's a primitive class?
3522 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3523 bool expect_prim = false; // most of these guys expect to work on refs
3524
3525 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3526
3527 Node* mirror = argument(0);
3528
3529 if (ShenandoahVerifyReadsToFromSpace) {
3530 mirror = shenandoah_read_barrier(mirror);
3531 }
3532
3533 Node* obj = top();
3534
3535 switch (id) {
3536 case vmIntrinsics::_isInstance:
3537 // nothing is an instance of a primitive type
3538 prim_return_value = intcon(0);
3539 obj = argument(1);
3540 if (ShenandoahVerifyReadsToFromSpace) {
3541 obj = shenandoah_read_barrier(obj);
3542 }
3543 break;
3544 case vmIntrinsics::_getModifiers:
3545 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3546 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3547 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3548 break;
3549 case vmIntrinsics::_isInterface:
3550 prim_return_value = intcon(0);
3551 break;
3552 case vmIntrinsics::_isArray:
3553 prim_return_value = intcon(0);
3554 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3555 break;
3556 case vmIntrinsics::_isPrimitive:
3557 prim_return_value = intcon(1);
3558 expect_prim = true; // obviously
3559 break;
3560 case vmIntrinsics::_getSuperclass:
3561 prim_return_value = null();
3562 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3563 break;
3564 case vmIntrinsics::_getClassAccessFlags:
3565 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3566 return_type = TypeInt::INT; // not bool! 6297094
3567 break;
3568 default:
3569 fatal_unexpected_iid(id);
3570 break;
3571 }
3572
3573 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3574 if (mirror_con == NULL) return false; // cannot happen?
3575
3576 #ifndef PRODUCT
3577 if (C->print_intrinsics() || C->print_inlining()) {
3578 ciType* k = mirror_con->java_mirror_type();
3579 if (k) {
3580 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3581 k->print_name();
3582 tty->cr();
3583 }
3584 }
3585 #endif
3586
3587 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3588 RegionNode* region = new RegionNode(PATH_LIMIT);
3589 record_for_igvn(region);
3590 PhiNode* phi = new PhiNode(region, return_type);
3591
3592 // The mirror will never be null of Reflection.getClassAccessFlags, however
3593 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3594 // if it is. See bug 4774291.
3595
3596 // For Reflection.getClassAccessFlags(), the null check occurs in
3597 // the wrong place; see inline_unsafe_access(), above, for a similar
3598 // situation.
3599 mirror = null_check(mirror);
3600 // If mirror or obj is dead, only null-path is taken.
3601 if (stopped()) return true;
3602
3603 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3604
3605 // Now load the mirror's klass metaobject, and null-check it.
3606 // Side-effects region with the control path if the klass is null.
3607 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3608 // If kls is null, we have a primitive mirror.
3609 phi->init_req(_prim_path, prim_return_value);
3610 if (stopped()) { set_result(region, phi); return true; }
3611 bool safe_for_replace = (region->in(_prim_path) == top());
3612
3613 Node* p; // handy temp
3614 Node* null_ctl;
3615
3616 // Now that we have the non-null klass, we can perform the real query.
3617 // For constant classes, the query will constant-fold in LoadNode::Value.
3618 Node* query_value = top();
3619 switch (id) {
3620 case vmIntrinsics::_isInstance:
3621 // nothing is an instance of a primitive type
3622 query_value = gen_instanceof(obj, kls, safe_for_replace);
3623 break;
3624
3625 case vmIntrinsics::_getModifiers:
3626 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3627 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3628 break;
3629
3630 case vmIntrinsics::_isInterface:
3631 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3632 if (generate_interface_guard(kls, region) != NULL)
3633 // A guard was added. If the guard is taken, it was an interface.
3634 phi->add_req(intcon(1));
3635 // If we fall through, it's a plain class.
3636 query_value = intcon(0);
3637 break;
3638
3639 case vmIntrinsics::_isArray:
3640 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3641 if (generate_array_guard(kls, region) != NULL)
3642 // A guard was added. If the guard is taken, it was an array.
3643 phi->add_req(intcon(1));
3644 // If we fall through, it's a plain class.
3645 query_value = intcon(0);
3646 break;
3647
3648 case vmIntrinsics::_isPrimitive:
3649 query_value = intcon(0); // "normal" path produces false
3650 break;
3651
3652 case vmIntrinsics::_getSuperclass:
3653 // The rules here are somewhat unfortunate, but we can still do better
3654 // with random logic than with a JNI call.
3655 // Interfaces store null or Object as _super, but must report null.
3656 // Arrays store an intermediate super as _super, but must report Object.
3657 // Other types can report the actual _super.
3658 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3659 if (generate_interface_guard(kls, region) != NULL)
3660 // A guard was added. If the guard is taken, it was an interface.
3661 phi->add_req(null());
3662 if (generate_array_guard(kls, region) != NULL)
3663 // A guard was added. If the guard is taken, it was an array.
3664 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3665 // If we fall through, it's a plain class. Get its _super.
3666 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3667 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3668 null_ctl = top();
3669 kls = null_check_oop(kls, &null_ctl);
3670 if (null_ctl != top()) {
3671 // If the guard is taken, Object.superClass is null (both klass and mirror).
3672 region->add_req(null_ctl);
3673 phi ->add_req(null());
3674 }
3675 if (!stopped()) {
3676 query_value = load_mirror_from_klass(kls);
3677 }
3678 break;
3679
3680 case vmIntrinsics::_getClassAccessFlags:
3681 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3682 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3683 break;
3684
3685 default:
3686 fatal_unexpected_iid(id);
3687 break;
3688 }
3689
3690 // Fall-through is the normal case of a query to a real class.
3691 phi->init_req(1, query_value);
3692 region->init_req(1, control());
3693
3694 C->set_has_split_ifs(true); // Has chance for split-if optimization
3695 set_result(region, phi);
3696 return true;
3697 }
3698
3699 //-------------------------inline_Class_cast-------------------
3700 bool LibraryCallKit::inline_Class_cast() {
3701 Node* mirror = argument(0); // Class
3702 Node* obj = argument(1);
3703 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3704 if (mirror_con == NULL) {
3705 return false; // dead path (mirror->is_top()).
3706 }
3707 if (obj == NULL || obj->is_top()) {
3708 return false; // dead path
3709 }
3710 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3711
3712 // First, see if Class.cast() can be folded statically.
3713 // java_mirror_type() returns non-null for compile-time Class constants.
3714 ciType* tm = mirror_con->java_mirror_type();
3715 if (tm != NULL && tm->is_klass() &&
3716 tp != NULL && tp->klass() != NULL) {
3717 if (!tp->klass()->is_loaded()) {
3718 // Don't use intrinsic when class is not loaded.
3719 return false;
3720 } else {
3721 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3722 if (static_res == Compile::SSC_always_true) {
3723 // isInstance() is true - fold the code.
3724 set_result(obj);
3725 return true;
3726 } else if (static_res == Compile::SSC_always_false) {
3727 // Don't use intrinsic, have to throw ClassCastException.
3728 // If the reference is null, the non-intrinsic bytecode will
3729 // be optimized appropriately.
3730 return false;
3731 }
3732 }
3733 }
3734
3735 // Bailout intrinsic and do normal inlining if exception path is frequent.
3736 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3737 return false;
3738 }
3739
3740 // Generate dynamic checks.
3741 // Class.cast() is java implementation of _checkcast bytecode.
3742 // Do checkcast (Parse::do_checkcast()) optimizations here.
3743
3744 mirror = null_check(mirror);
3745 // If mirror is dead, only null-path is taken.
3746 if (stopped()) {
3747 return true;
3748 }
3749
3750 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3751 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3752 RegionNode* region = new RegionNode(PATH_LIMIT);
3753 record_for_igvn(region);
3754
3755 // Now load the mirror's klass metaobject, and null-check it.
3756 // If kls is null, we have a primitive mirror and
3757 // nothing is an instance of a primitive type.
3758 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3759
3760 Node* res = top();
3761 if (!stopped()) {
3762 Node* bad_type_ctrl = top();
3763 // Do checkcast optimizations.
3764 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3765 region->init_req(_bad_type_path, bad_type_ctrl);
3766 }
3767 if (region->in(_prim_path) != top() ||
3768 region->in(_bad_type_path) != top()) {
3769 // Let Interpreter throw ClassCastException.
3770 PreserveJVMState pjvms(this);
3771 set_control(_gvn.transform(region));
3772 uncommon_trap(Deoptimization::Reason_intrinsic,
3773 Deoptimization::Action_maybe_recompile);
3774 }
3775 if (!stopped()) {
3776 set_result(res);
3777 }
3778 return true;
3779 }
3780
3781
3782 //--------------------------inline_native_subtype_check------------------------
3783 // This intrinsic takes the JNI calls out of the heart of
3784 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3785 bool LibraryCallKit::inline_native_subtype_check() {
3786 // Pull both arguments off the stack.
3787 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3788 args[0] = argument(0);
3789 args[1] = argument(1);
3790 Node* klasses[2]; // corresponding Klasses: superk, subk
3791 klasses[0] = klasses[1] = top();
3792
3793 enum {
3794 // A full decision tree on {superc is prim, subc is prim}:
3795 _prim_0_path = 1, // {P,N} => false
3796 // {P,P} & superc!=subc => false
3797 _prim_same_path, // {P,P} & superc==subc => true
3798 _prim_1_path, // {N,P} => false
3799 _ref_subtype_path, // {N,N} & subtype check wins => true
3800 _both_ref_path, // {N,N} & subtype check loses => false
3801 PATH_LIMIT
3802 };
3803
3804 RegionNode* region = new RegionNode(PATH_LIMIT);
3805 Node* phi = new PhiNode(region, TypeInt::BOOL);
3806 Node* mem_phi= new PhiNode(region, Type::MEMORY, TypePtr::BOTTOM);
3807 record_for_igvn(region);
3808 Node* init_mem = map()->memory();
3809
3810 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3811 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
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 if (ShenandoahVerifyOptoBarriers) {
3828 args[0] = shenandoah_write_barrier(args[0]);
3829 args[1] = shenandoah_write_barrier(args[1]);
3830 }
3831
3832
3833 // Having loaded both klasses, test each for null.
3834 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3835 for (which_arg = 0; which_arg <= 1; which_arg++) {
3836 Node* kls = klasses[which_arg];
3837 Node* null_ctl = top();
3838 kls = null_check_oop(kls, &null_ctl, never_see_null);
3839 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3840 region->init_req(prim_path, null_ctl);
3841 if (stopped()) break;
3842 klasses[which_arg] = kls;
3843 }
3844
3845 if (!stopped()) {
3846 // now we have two reference types, in klasses[0..1]
3847 Node* subk = klasses[1]; // the argument to isAssignableFrom
3848 Node* superk = klasses[0]; // the receiver
3849 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3850 // now we have a successful reference subtype check
3851 region->set_req(_ref_subtype_path, control());
3852 }
3853
3854 // If both operands are primitive (both klasses null), then
3855 // we must return true when they are identical primitives.
3856 // It is convenient to test this after the first null klass check.
3857 set_control(region->in(_prim_0_path)); // go back to first null check
3858 if (!stopped()) {
3859 // Since superc is primitive, make a guard for the superc==subc case.
3860 Node* cmp_eq = cmp_objects(args[0], args[1]);
3861 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3862 generate_guard(bol_eq, region, PROB_FAIR);
3863 if (region->req() == PATH_LIMIT+1) {
3864 // A guard was added. If the added guard is taken, superc==subc.
3865 region->swap_edges(PATH_LIMIT, _prim_same_path);
3866 region->del_req(PATH_LIMIT);
3867 }
3868 region->set_req(_prim_0_path, control()); // Not equal after all.
3869 }
3870
3871 // these are the only paths that produce 'true':
3872 phi->set_req(_prim_same_path, intcon(1));
3873 phi->set_req(_ref_subtype_path, intcon(1));
3874
3875 // pull together the cases:
3876 assert(region->req() == PATH_LIMIT, "sane region");
3877 Node* cur_mem = reset_memory();
3878 for (uint i = 1; i < region->req(); i++) {
3879 Node* ctl = region->in(i);
3880 if (ctl == NULL || ctl == top()) {
3881 region->set_req(i, top());
3882 phi ->set_req(i, top());
3883 mem_phi->set_req(i, top());
3884 } else {
3885 if (phi->in(i) == NULL) {
3886 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3887 }
3888 mem_phi->set_req(i, (i == _prim_0_path || i == _prim_same_path) ? cur_mem : init_mem);
3889 }
3890 }
3891
3892 set_control(_gvn.transform(region));
3893 set_result(_gvn.transform(phi));
3894 set_all_memory(_gvn.transform(mem_phi));
3895 return true;
3896 }
3897
3898 //---------------------generate_array_guard_common------------------------
3899 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3900 bool obj_array, bool not_array) {
3901
3902 if (stopped()) {
3903 return NULL;
3904 }
3905
3906 // If obj_array/non_array==false/false:
3907 // Branch around if the given klass is in fact an array (either obj or prim).
3908 // If obj_array/non_array==false/true:
3909 // Branch around if the given klass is not an array klass of any kind.
3910 // If obj_array/non_array==true/true:
3911 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3912 // If obj_array/non_array==true/false:
3913 // Branch around if the kls is an oop array (Object[] or subtype)
3914 //
3915 // Like generate_guard, adds a new path onto the region.
3916 jint layout_con = 0;
3917 Node* layout_val = get_layout_helper(kls, layout_con);
3918 if (layout_val == NULL) {
3919 bool query = (obj_array
3920 ? Klass::layout_helper_is_objArray(layout_con)
3921 : Klass::layout_helper_is_array(layout_con));
3922 if (query == not_array) {
3923 return NULL; // never a branch
3924 } else { // always a branch
3925 Node* always_branch = control();
3926 if (region != NULL)
3927 region->add_req(always_branch);
3928 set_control(top());
3929 return always_branch;
3930 }
3931 }
3932 // Now test the correct condition.
3933 jint nval = (obj_array
3934 ? (jint)(Klass::_lh_array_tag_type_value
3935 << Klass::_lh_array_tag_shift)
3936 : Klass::_lh_neutral_value);
3937 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3938 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3939 // invert the test if we are looking for a non-array
3940 if (not_array) btest = BoolTest(btest).negate();
3941 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3942 return generate_fair_guard(bol, region);
3943 }
3944
3945
3946 //-----------------------inline_native_newArray--------------------------
3947 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3948 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3949 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3950 Node* mirror;
3951 Node* count_val;
3952 if (uninitialized) {
3953 mirror = argument(1);
3954 count_val = argument(2);
3955 } else {
3956 mirror = argument(0);
3957 count_val = argument(1);
3958 }
3959
3960 mirror = null_check(mirror);
3961 // If mirror or obj is dead, only null-path is taken.
3962 if (stopped()) return true;
3963
3964 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3965 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3966 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3967 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3968 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3969
3970 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3971 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3972 result_reg, _slow_path);
3973 Node* normal_ctl = control();
3974 Node* no_array_ctl = result_reg->in(_slow_path);
3975
3976 // Generate code for the slow case. We make a call to newArray().
3977 set_control(no_array_ctl);
3978 if (!stopped()) {
3979 // Either the input type is void.class, or else the
3980 // array klass has not yet been cached. Either the
3981 // ensuing call will throw an exception, or else it
3982 // will cache the array klass for next time.
3983 PreserveJVMState pjvms(this);
3984 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3985 Node* slow_result = set_results_for_java_call(slow_call);
3986 // this->control() comes from set_results_for_java_call
3987 result_reg->set_req(_slow_path, control());
3988 result_val->set_req(_slow_path, slow_result);
3989 result_io ->set_req(_slow_path, i_o());
3990 result_mem->set_req(_slow_path, reset_memory());
3991 }
3992
3993 set_control(normal_ctl);
3994 if (!stopped()) {
3995 // Normal case: The array type has been cached in the java.lang.Class.
3996 // The following call works fine even if the array type is polymorphic.
3997 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3998 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3999 result_reg->init_req(_normal_path, control());
4000 result_val->init_req(_normal_path, obj);
4001 result_io ->init_req(_normal_path, i_o());
4002 result_mem->init_req(_normal_path, reset_memory());
4003
4004 if (uninitialized) {
4005 // Mark the allocation so that zeroing is skipped
4006 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4007 alloc->maybe_set_complete(&_gvn);
4008 }
4009 }
4010
4011 // Return the combined state.
4012 set_i_o( _gvn.transform(result_io) );
4013 set_all_memory( _gvn.transform(result_mem));
4014
4015 C->set_has_split_ifs(true); // Has chance for split-if optimization
4016 set_result(result_reg, result_val);
4017 return true;
4018 }
4019
4020 //----------------------inline_native_getLength--------------------------
4021 // public static native int java.lang.reflect.Array.getLength(Object array);
4022 bool LibraryCallKit::inline_native_getLength() {
4023 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4024
4025 Node* array = null_check(argument(0));
4026 // If array is dead, only null-path is taken.
4027 if (stopped()) return true;
4028
4029 // Deoptimize if it is a non-array.
4030 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4031
4032 if (non_array != NULL) {
4033 PreserveJVMState pjvms(this);
4034 set_control(non_array);
4035 uncommon_trap(Deoptimization::Reason_intrinsic,
4036 Deoptimization::Action_maybe_recompile);
4037 }
4038
4039 // If control is dead, only non-array-path is taken.
4040 if (stopped()) return true;
4041
4042 // The works fine even if the array type is polymorphic.
4043 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4044 Node* result = load_array_length(array);
4045
4046 C->set_has_split_ifs(true); // Has chance for split-if optimization
4047 set_result(result);
4048 return true;
4049 }
4050
4051 //------------------------inline_array_copyOf----------------------------
4052 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4053 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4054 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4055 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4056
4057 // Get the arguments.
4058 Node* original = argument(0);
4059 Node* start = is_copyOfRange? argument(1): intcon(0);
4060 Node* end = is_copyOfRange? argument(2): argument(1);
4061 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4062
4063 Node* newcopy = NULL;
4064
4065 // Set the original stack and the reexecute bit for the interpreter to reexecute
4066 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4067 { PreserveReexecuteState preexecs(this);
4068 jvms()->set_should_reexecute(true);
4069
4070 array_type_mirror = null_check(array_type_mirror);
4071 original = null_check(original);
4072
4073 // Check if a null path was taken unconditionally.
4074 if (stopped()) return true;
4075
4076 Node* orig_length = load_array_length(original);
4077
4078 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4079 klass_node = null_check(klass_node);
4080
4081 RegionNode* bailout = new RegionNode(1);
4082 record_for_igvn(bailout);
4083
4084 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4085 // Bail out if that is so.
4086 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4087 if (not_objArray != NULL) {
4088 // Improve the klass node's type from the new optimistic assumption:
4089 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4090 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4091 Node* cast = new CastPPNode(klass_node, akls);
4092 cast->init_req(0, control());
4093 klass_node = _gvn.transform(cast);
4094 }
4095
4096 // Bail out if either start or end is negative.
4097 generate_negative_guard(start, bailout, &start);
4098 generate_negative_guard(end, bailout, &end);
4099
4100 Node* length = end;
4101 if (_gvn.type(start) != TypeInt::ZERO) {
4102 length = _gvn.transform(new SubINode(end, start));
4103 }
4104
4105 // Bail out if length is negative.
4106 // Without this the new_array would throw
4107 // NegativeArraySizeException but IllegalArgumentException is what
4108 // should be thrown
4109 generate_negative_guard(length, bailout, &length);
4110
4111 if (bailout->req() > 1) {
4112 PreserveJVMState pjvms(this);
4113 set_control(_gvn.transform(bailout));
4114 uncommon_trap(Deoptimization::Reason_intrinsic,
4115 Deoptimization::Action_maybe_recompile);
4116 }
4117
4118 if (!stopped()) {
4119 // How many elements will we copy from the original?
4120 // The answer is MinI(orig_length - start, length).
4121 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4122 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4123
4124 original = shenandoah_read_barrier(original);
4125
4126 // Generate a direct call to the right arraycopy function(s).
4127 // We know the copy is disjoint but we might not know if the
4128 // oop stores need checking.
4129 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4130 // This will fail a store-check if x contains any non-nulls.
4131
4132 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4133 // loads/stores but it is legal only if we're sure the
4134 // Arrays.copyOf would succeed. So we need all input arguments
4135 // to the copyOf to be validated, including that the copy to the
4136 // new array won't trigger an ArrayStoreException. That subtype
4137 // check can be optimized if we know something on the type of
4138 // the input array from type speculation.
4139 if (_gvn.type(klass_node)->singleton()) {
4140 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4141 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4142
4143 int test = C->static_subtype_check(superk, subk);
4144 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4145 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4146 if (t_original->speculative_type() != NULL) {
4147 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4148 }
4149 }
4150 }
4151
4152 bool validated = false;
4153 // Reason_class_check rather than Reason_intrinsic because we
4154 // want to intrinsify even if this traps.
4155 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4156 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4157 klass_node);
4158
4159 if (not_subtype_ctrl != top()) {
4160 PreserveJVMState pjvms(this);
4161 set_control(not_subtype_ctrl);
4162 uncommon_trap(Deoptimization::Reason_class_check,
4163 Deoptimization::Action_make_not_entrant);
4164 assert(stopped(), "Should be stopped");
4165 }
4166 validated = true;
4167 }
4168
4169 if (!stopped()) {
4170 newcopy = new_array(klass_node, length, 0); // no arguments to push
4171
4172 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4173 load_object_klass(original), klass_node);
4174 if (!is_copyOfRange) {
4175 ac->set_copyof(validated);
4176 } else {
4177 ac->set_copyofrange(validated);
4178 }
4179 Node* n = _gvn.transform(ac);
4180 if (n == ac) {
4181 ac->_adr_type = TypePtr::BOTTOM;
4182 ac->connect_outputs(this);
4183 } else {
4184 assert(validated, "shouldn't transform if all arguments not validated");
4185 set_all_memory(n);
4186 }
4187 }
4188 }
4189 } // original reexecute is set back here
4190
4191 C->set_has_split_ifs(true); // Has chance for split-if optimization
4192 if (!stopped()) {
4193 set_result(newcopy);
4194 }
4195 return true;
4196 }
4197
4198
4199 //----------------------generate_virtual_guard---------------------------
4200 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4201 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4202 RegionNode* slow_region) {
4203 ciMethod* method = callee();
4204 int vtable_index = method->vtable_index();
4205 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4206 "bad index %d", vtable_index);
4207 // Get the Method* out of the appropriate vtable entry.
4208 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4209 vtable_index*vtableEntry::size_in_bytes() +
4210 vtableEntry::method_offset_in_bytes();
4211 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4212 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4213
4214 // Compare the target method with the expected method (e.g., Object.hashCode).
4215 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4216
4217 Node* native_call = makecon(native_call_addr);
4218 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4219 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4220
4221 return generate_slow_guard(test_native, slow_region);
4222 }
4223
4224 //-----------------------generate_method_call----------------------------
4225 // Use generate_method_call to make a slow-call to the real
4226 // method if the fast path fails. An alternative would be to
4227 // use a stub like OptoRuntime::slow_arraycopy_Java.
4228 // This only works for expanding the current library call,
4229 // not another intrinsic. (E.g., don't use this for making an
4230 // arraycopy call inside of the copyOf intrinsic.)
4231 CallJavaNode*
4232 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4233 // When compiling the intrinsic method itself, do not use this technique.
4234 guarantee(callee() != C->method(), "cannot make slow-call to self");
4235
4236 ciMethod* method = callee();
4237 // ensure the JVMS we have will be correct for this call
4238 guarantee(method_id == method->intrinsic_id(), "must match");
4239
4240 const TypeFunc* tf = TypeFunc::make(method);
4241 CallJavaNode* slow_call;
4242 if (is_static) {
4243 assert(!is_virtual, "");
4244 slow_call = new CallStaticJavaNode(C, tf,
4245 SharedRuntime::get_resolve_static_call_stub(),
4246 method, bci());
4247 } else if (is_virtual) {
4248 null_check_receiver();
4249 int vtable_index = Method::invalid_vtable_index;
4250 if (UseInlineCaches) {
4251 // Suppress the vtable call
4252 } else {
4253 // hashCode and clone are not a miranda methods,
4254 // so the vtable index is fixed.
4255 // No need to use the linkResolver to get it.
4256 vtable_index = method->vtable_index();
4257 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4258 "bad index %d", vtable_index);
4259 }
4260 slow_call = new CallDynamicJavaNode(tf,
4261 SharedRuntime::get_resolve_virtual_call_stub(),
4262 method, vtable_index, bci());
4263 } else { // neither virtual nor static: opt_virtual
4264 null_check_receiver();
4265 slow_call = new CallStaticJavaNode(C, tf,
4266 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4267 method, bci());
4268 slow_call->set_optimized_virtual(true);
4269 }
4270 set_arguments_for_java_call(slow_call);
4271 set_edges_for_java_call(slow_call);
4272 return slow_call;
4273 }
4274
4275
4276 /**
4277 * Build special case code for calls to hashCode on an object. This call may
4278 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4279 * slightly different code.
4280 */
4281 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4282 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4283 assert(!(is_virtual && is_static), "either virtual, special, or static");
4284
4285 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4286
4287 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4288 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4289 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4290 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4291 Node* obj = NULL;
4292 if (!is_static) {
4293 // Check for hashing null object
4294 obj = null_check_receiver();
4295 if (stopped()) return true; // unconditionally null
4296 result_reg->init_req(_null_path, top());
4297 result_val->init_req(_null_path, top());
4298 } else {
4299 // Do a null check, and return zero if null.
4300 // System.identityHashCode(null) == 0
4301 obj = argument(0);
4302 Node* null_ctl = top();
4303 obj = null_check_oop(obj, &null_ctl);
4304 result_reg->init_req(_null_path, null_ctl);
4305 result_val->init_req(_null_path, _gvn.intcon(0));
4306 }
4307
4308 if (ShenandoahVerifyReadsToFromSpace) {
4309 obj = shenandoah_read_barrier(obj);
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, false);
4622 Node* dst = make_unsafe_address(dst_ptr, dst_off, true);
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 obj = shenandoah_read_barrier(obj);
4650
4651 AllocateNode* alloc = NULL;
4652 if (ReduceBulkZeroing) {
4653 // We will be completely responsible for initializing this object -
4654 // mark Initialize node as complete.
4655 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4656 // The object was just allocated - there should be no any stores!
4657 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4658 // Mark as complete_with_arraycopy so that on AllocateNode
4659 // expansion, we know this AllocateNode is initialized by an array
4660 // copy and a StoreStore barrier exists after the array copy.
4661 alloc->initialization()->set_complete_with_arraycopy();
4662 }
4663
4664 // Copy the fastest available way.
4665 // TODO: generate fields copies for small objects instead.
4666 Node* src = obj;
4667 Node* dest = alloc_obj;
4668 Node* size = _gvn.transform(obj_size);
4669
4670 // Exclude the header but include array length to copy by 8 bytes words.
4671 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4672 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4673 instanceOopDesc::base_offset_in_bytes();
4674 // base_off:
4675 // 8 - 32-bit VM
4676 // 12 - 64-bit VM, compressed klass
4677 // 16 - 64-bit VM, normal klass
4678 if (base_off % BytesPerLong != 0) {
4679 assert(UseCompressedClassPointers, "");
4680 if (is_array) {
4681 // Exclude length to copy by 8 bytes words.
4682 base_off += sizeof(int);
4683 } else {
4684 // Include klass to copy by 8 bytes words.
4685 base_off = instanceOopDesc::klass_offset_in_bytes();
4686 }
4687 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4688 }
4689 src = basic_plus_adr(src, base_off);
4690 dest = basic_plus_adr(dest, base_off);
4691
4692 // Compute the length also, if needed:
4693 Node* countx = size;
4694 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4695 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4696
4697 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4698
4699 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4700 ac->set_clonebasic();
4701 Node* n = _gvn.transform(ac);
4702 if (n == ac) {
4703 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4704 } else {
4705 set_all_memory(n);
4706 }
4707
4708 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4709 if (card_mark) {
4710 assert(!is_array, "");
4711 // Put in store barrier for any and all oops we are sticking
4712 // into this object. (We could avoid this if we could prove
4713 // that the object type contains no oop fields at all.)
4714 Node* no_particular_value = NULL;
4715 Node* no_particular_field = NULL;
4716 int raw_adr_idx = Compile::AliasIdxRaw;
4717 post_barrier(control(),
4718 memory(raw_adr_type),
4719 alloc_obj,
4720 no_particular_field,
4721 raw_adr_idx,
4722 no_particular_value,
4723 T_OBJECT,
4724 false);
4725 }
4726
4727 // Do not let reads from the cloned object float above the arraycopy.
4728 if (alloc != NULL) {
4729 // Do not let stores that initialize this object be reordered with
4730 // a subsequent store that would make this object accessible by
4731 // other threads.
4732 // Record what AllocateNode this StoreStore protects so that
4733 // escape analysis can go from the MemBarStoreStoreNode to the
4734 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4735 // based on the escape status of the AllocateNode.
4736 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4737 } else {
4738 insert_mem_bar(Op_MemBarCPUOrder);
4739 }
4740 }
4741
4742 //------------------------inline_native_clone----------------------------
4743 // protected native Object java.lang.Object.clone();
4744 //
4745 // Here are the simple edge cases:
4746 // null receiver => normal trap
4747 // virtual and clone was overridden => slow path to out-of-line clone
4748 // not cloneable or finalizer => slow path to out-of-line Object.clone
4749 //
4750 // The general case has two steps, allocation and copying.
4751 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4752 //
4753 // Copying also has two cases, oop arrays and everything else.
4754 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4755 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4756 //
4757 // These steps fold up nicely if and when the cloned object's klass
4758 // can be sharply typed as an object array, a type array, or an instance.
4759 //
4760 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4761 PhiNode* result_val;
4762
4763 // Set the reexecute bit for the interpreter to reexecute
4764 // the bytecode that invokes Object.clone if deoptimization happens.
4765 { PreserveReexecuteState preexecs(this);
4766 jvms()->set_should_reexecute(true);
4767
4768 Node* obj = null_check_receiver();
4769 if (stopped()) return true;
4770
4771 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4772
4773 // If we are going to clone an instance, we need its exact type to
4774 // know the number and types of fields to convert the clone to
4775 // loads/stores. Maybe a speculative type can help us.
4776 if (!obj_type->klass_is_exact() &&
4777 obj_type->speculative_type() != NULL &&
4778 obj_type->speculative_type()->is_instance_klass()) {
4779 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4780 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4781 !spec_ik->has_injected_fields()) {
4782 ciKlass* k = obj_type->klass();
4783 if (!k->is_instance_klass() ||
4784 k->as_instance_klass()->is_interface() ||
4785 k->as_instance_klass()->has_subklass()) {
4786 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4787 }
4788 }
4789 }
4790
4791 Node* obj_klass = load_object_klass(obj);
4792 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4793 const TypeOopPtr* toop = ((tklass != NULL)
4794 ? tklass->as_instance_type()
4795 : TypeInstPtr::NOTNULL);
4796
4797 // Conservatively insert a memory barrier on all memory slices.
4798 // Do not let writes into the original float below the clone.
4799 insert_mem_bar(Op_MemBarCPUOrder);
4800
4801 // paths into result_reg:
4802 enum {
4803 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4804 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4805 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4806 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4807 PATH_LIMIT
4808 };
4809 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4810 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4811 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4812 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4813 record_for_igvn(result_reg);
4814
4815 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4816 int raw_adr_idx = Compile::AliasIdxRaw;
4817
4818 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4819 if (array_ctl != NULL) {
4820 // It's an array.
4821 PreserveJVMState pjvms(this);
4822 set_control(array_ctl);
4823 Node* obj_length = load_array_length(obj);
4824 Node* obj_size = NULL;
4825 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4826
4827 if (!use_ReduceInitialCardMarks()) {
4828 // If it is an oop array, it requires very special treatment,
4829 // because card marking is required on each card of the array.
4830 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4831 if (is_obja != NULL) {
4832 PreserveJVMState pjvms2(this);
4833 set_control(is_obja);
4834
4835 obj = shenandoah_read_barrier(obj);
4836
4837 // Generate a direct call to the right arraycopy function(s).
4838 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4839 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4840 ac->set_cloneoop();
4841 Node* n = _gvn.transform(ac);
4842 assert(n == ac, "cannot disappear");
4843 ac->_adr_type = TypePtr::BOTTOM;
4844 ac->connect_outputs(this);
4845
4846 result_reg->init_req(_objArray_path, control());
4847 result_val->init_req(_objArray_path, alloc_obj);
4848 result_i_o ->set_req(_objArray_path, i_o());
4849 result_mem ->set_req(_objArray_path, reset_memory());
4850 }
4851 }
4852 // Otherwise, there are no card marks to worry about.
4853 // (We can dispense with card marks if we know the allocation
4854 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4855 // causes the non-eden paths to take compensating steps to
4856 // simulate a fresh allocation, so that no further
4857 // card marks are required in compiled code to initialize
4858 // the object.)
4859
4860 if (!stopped()) {
4861 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4862
4863 // Present the results of the copy.
4864 result_reg->init_req(_array_path, control());
4865 result_val->init_req(_array_path, alloc_obj);
4866 result_i_o ->set_req(_array_path, i_o());
4867 result_mem ->set_req(_array_path, reset_memory());
4868 }
4869 }
4870
4871 // We only go to the instance fast case code if we pass a number of guards.
4872 // The paths which do not pass are accumulated in the slow_region.
4873 RegionNode* slow_region = new RegionNode(1);
4874 record_for_igvn(slow_region);
4875 if (!stopped()) {
4876 // It's an instance (we did array above). Make the slow-path tests.
4877 // If this is a virtual call, we generate a funny guard. We grab
4878 // the vtable entry corresponding to clone() from the target object.
4879 // If the target method which we are calling happens to be the
4880 // Object clone() method, we pass the guard. We do not need this
4881 // guard for non-virtual calls; the caller is known to be the native
4882 // Object clone().
4883 if (is_virtual) {
4884 generate_virtual_guard(obj_klass, slow_region);
4885 }
4886
4887 // The object must be easily cloneable and must not have a finalizer.
4888 // Both of these conditions may be checked in a single test.
4889 // We could optimize the test further, but we don't care.
4890 generate_access_flags_guard(obj_klass,
4891 // Test both conditions:
4892 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4893 // Must be cloneable but not finalizer:
4894 JVM_ACC_IS_CLONEABLE_FAST,
4895 slow_region);
4896 }
4897
4898 if (!stopped()) {
4899 // It's an instance, and it passed the slow-path tests.
4900 PreserveJVMState pjvms(this);
4901 Node* obj_size = NULL;
4902 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4903 // is reexecuted if deoptimization occurs and there could be problems when merging
4904 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4905 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4906
4907 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4908
4909 // Present the results of the slow call.
4910 result_reg->init_req(_instance_path, control());
4911 result_val->init_req(_instance_path, alloc_obj);
4912 result_i_o ->set_req(_instance_path, i_o());
4913 result_mem ->set_req(_instance_path, reset_memory());
4914 }
4915
4916 // Generate code for the slow case. We make a call to clone().
4917 set_control(_gvn.transform(slow_region));
4918 if (!stopped()) {
4919 PreserveJVMState pjvms(this);
4920 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4921 Node* slow_result = set_results_for_java_call(slow_call);
4922 // this->control() comes from set_results_for_java_call
4923 result_reg->init_req(_slow_path, control());
4924 result_val->init_req(_slow_path, slow_result);
4925 result_i_o ->set_req(_slow_path, i_o());
4926 result_mem ->set_req(_slow_path, reset_memory());
4927 }
4928
4929 // Return the combined state.
4930 set_control( _gvn.transform(result_reg));
4931 set_i_o( _gvn.transform(result_i_o));
4932 set_all_memory( _gvn.transform(result_mem));
4933 } // original reexecute is set back here
4934
4935 set_result(_gvn.transform(result_val));
4936 return true;
4937 }
4938
4939 // If we have a tighly coupled allocation, the arraycopy may take care
4940 // of the array initialization. If one of the guards we insert between
4941 // the allocation and the arraycopy causes a deoptimization, an
4942 // unitialized array will escape the compiled method. To prevent that
4943 // we set the JVM state for uncommon traps between the allocation and
4944 // the arraycopy to the state before the allocation so, in case of
4945 // deoptimization, we'll reexecute the allocation and the
4946 // initialization.
4947 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4948 if (alloc != NULL) {
4949 ciMethod* trap_method = alloc->jvms()->method();
4950 int trap_bci = alloc->jvms()->bci();
4951
4952 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4953 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4954 // Make sure there's no store between the allocation and the
4955 // arraycopy otherwise visible side effects could be rexecuted
4956 // in case of deoptimization and cause incorrect execution.
4957 bool no_interfering_store = true;
4958 Node* mem = alloc->in(TypeFunc::Memory);
4959 if (mem->is_MergeMem()) {
4960 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4961 Node* n = mms.memory();
4962 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4963 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4964 no_interfering_store = false;
4965 break;
4966 }
4967 }
4968 } else {
4969 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4970 Node* n = mms.memory();
4971 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4972 assert(n->is_Store() || n->Opcode() == Op_ShenandoahWBMemProj, "what else?");
4973 no_interfering_store = false;
4974 break;
4975 }
4976 }
4977 }
4978
4979 if (no_interfering_store) {
4980 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4981 uint size = alloc->req();
4982 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4983 old_jvms->set_map(sfpt);
4984 for (uint i = 0; i < size; i++) {
4985 sfpt->init_req(i, alloc->in(i));
4986 }
4987 // re-push array length for deoptimization
4988 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4989 old_jvms->set_sp(old_jvms->sp()+1);
4990 old_jvms->set_monoff(old_jvms->monoff()+1);
4991 old_jvms->set_scloff(old_jvms->scloff()+1);
4992 old_jvms->set_endoff(old_jvms->endoff()+1);
4993 old_jvms->set_should_reexecute(true);
4994
4995 sfpt->set_i_o(map()->i_o());
4996 sfpt->set_memory(map()->memory());
4997 sfpt->set_control(map()->control());
4998
4999 JVMState* saved_jvms = jvms();
5000 saved_reexecute_sp = _reexecute_sp;
5001
5002 set_jvms(sfpt->jvms());
5003 _reexecute_sp = jvms()->sp();
5004
5005 return saved_jvms;
5006 }
5007 }
5008 }
5009 return NULL;
5010 }
5011
5012 // In case of a deoptimization, we restart execution at the
5013 // allocation, allocating a new array. We would leave an uninitialized
5014 // array in the heap that GCs wouldn't expect. Move the allocation
5015 // after the traps so we don't allocate the array if we
5016 // deoptimize. This is possible because tightly_coupled_allocation()
5017 // guarantees there's no observer of the allocated array at this point
5018 // and the control flow is simple enough.
5019 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
5020 if (saved_jvms != NULL && !stopped()) {
5021 assert(alloc != NULL, "only with a tightly coupled allocation");
5022 // restore JVM state to the state at the arraycopy
5023 saved_jvms->map()->set_control(map()->control());
5024 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5025 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5026 // If we've improved the types of some nodes (null check) while
5027 // emitting the guards, propagate them to the current state
5028 map()->replaced_nodes().apply(saved_jvms->map());
5029 set_jvms(saved_jvms);
5030 _reexecute_sp = saved_reexecute_sp;
5031
5032 // Remove the allocation from above the guards
5033 CallProjections callprojs;
5034 alloc->extract_projections(&callprojs, true);
5035 InitializeNode* init = alloc->initialization();
5036 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5037 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5038 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5039 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5040
5041 // move the allocation here (after the guards)
5042 _gvn.hash_delete(alloc);
5043 alloc->set_req(TypeFunc::Control, control());
5044 alloc->set_req(TypeFunc::I_O, i_o());
5045 Node *mem = reset_memory();
5046 set_all_memory(mem);
5047 alloc->set_req(TypeFunc::Memory, mem);
5048 set_control(init->proj_out(TypeFunc::Control));
5049 set_i_o(callprojs.fallthrough_ioproj);
5050
5051 // Update memory as done in GraphKit::set_output_for_allocation()
5052 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5053 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5054 if (ary_type->isa_aryptr() && length_type != NULL) {
5055 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5056 }
5057 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5058 int elemidx = C->get_alias_index(telemref);
5059 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
5060 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
5061
5062 Node* allocx = _gvn.transform(alloc);
5063 assert(allocx == alloc, "where has the allocation gone?");
5064 assert(dest->is_CheckCastPP(), "not an allocation result?");
5065
5066 _gvn.hash_delete(dest);
5067 dest->set_req(0, control());
5068 Node* destx = _gvn.transform(dest);
5069 assert(destx == dest, "where has the allocation result gone?");
5070 }
5071 }
5072
5073
5074 //------------------------------inline_arraycopy-----------------------
5075 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5076 // Object dest, int destPos,
5077 // int length);
5078 bool LibraryCallKit::inline_arraycopy() {
5079 // Get the arguments.
5080 Node* src = argument(0); // type: oop
5081 Node* src_offset = argument(1); // type: int
5082 Node* dest = argument(2); // type: oop
5083 Node* dest_offset = argument(3); // type: int
5084 Node* length = argument(4); // type: int
5085
5086
5087 // Check for allocation before we add nodes that would confuse
5088 // tightly_coupled_allocation()
5089 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5090
5091 int saved_reexecute_sp = -1;
5092 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5093 // See arraycopy_restore_alloc_state() comment
5094 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5095 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5096 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
5097 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5098
5099 // The following tests must be performed
5100 // (1) src and dest are arrays.
5101 // (2) src and dest arrays must have elements of the same BasicType
5102 // (3) src and dest must not be null.
5103 // (4) src_offset must not be negative.
5104 // (5) dest_offset must not be negative.
5105 // (6) length must not be negative.
5106 // (7) src_offset + length must not exceed length of src.
5107 // (8) dest_offset + length must not exceed length of dest.
5108 // (9) each element of an oop array must be assignable
5109
5110 // (3) src and dest must not be null.
5111 // always do this here because we need the JVM state for uncommon traps
5112 Node* null_ctl = top();
5113 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5114 assert(null_ctl->is_top(), "no null control here");
5115 dest = null_check(dest, T_ARRAY);
5116
5117 if (!can_emit_guards) {
5118 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5119 // guards but the arraycopy node could still take advantage of a
5120 // tightly allocated allocation. tightly_coupled_allocation() is
5121 // called again to make sure it takes the null check above into
5122 // account: the null check is mandatory and if it caused an
5123 // uncommon trap to be emitted then the allocation can't be
5124 // considered tightly coupled in this context.
5125 alloc = tightly_coupled_allocation(dest, NULL);
5126 }
5127
5128 bool validated = false;
5129
5130 const Type* src_type = _gvn.type(src);
5131 const Type* dest_type = _gvn.type(dest);
5132 const TypeAryPtr* top_src = src_type->isa_aryptr();
5133 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5134
5135 // Do we have the type of src?
5136 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5137 // Do we have the type of dest?
5138 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5139 // Is the type for src from speculation?
5140 bool src_spec = false;
5141 // Is the type for dest from speculation?
5142 bool dest_spec = false;
5143
5144 if ((!has_src || !has_dest) && can_emit_guards) {
5145 // We don't have sufficient type information, let's see if
5146 // speculative types can help. We need to have types for both src
5147 // and dest so that it pays off.
5148
5149 // Do we already have or could we have type information for src
5150 bool could_have_src = has_src;
5151 // Do we already have or could we have type information for dest
5152 bool could_have_dest = has_dest;
5153
5154 ciKlass* src_k = NULL;
5155 if (!has_src) {
5156 src_k = src_type->speculative_type_not_null();
5157 if (src_k != NULL && src_k->is_array_klass()) {
5158 could_have_src = true;
5159 }
5160 }
5161
5162 ciKlass* dest_k = NULL;
5163 if (!has_dest) {
5164 dest_k = dest_type->speculative_type_not_null();
5165 if (dest_k != NULL && dest_k->is_array_klass()) {
5166 could_have_dest = true;
5167 }
5168 }
5169
5170 if (could_have_src && could_have_dest) {
5171 // This is going to pay off so emit the required guards
5172 if (!has_src) {
5173 src = maybe_cast_profiled_obj(src, src_k, true);
5174 src_type = _gvn.type(src);
5175 top_src = src_type->isa_aryptr();
5176 has_src = (top_src != NULL && top_src->klass() != NULL);
5177 src_spec = true;
5178 }
5179 if (!has_dest) {
5180 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5181 dest_type = _gvn.type(dest);
5182 top_dest = dest_type->isa_aryptr();
5183 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5184 dest_spec = true;
5185 }
5186 }
5187 }
5188
5189 if (has_src && has_dest && can_emit_guards) {
5190 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5191 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5192 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5193 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5194
5195 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5196 // If both arrays are object arrays then having the exact types
5197 // for both will remove the need for a subtype check at runtime
5198 // before the call and may make it possible to pick a faster copy
5199 // routine (without a subtype check on every element)
5200 // Do we have the exact type of src?
5201 bool could_have_src = src_spec;
5202 // Do we have the exact type of dest?
5203 bool could_have_dest = dest_spec;
5204 ciKlass* src_k = top_src->klass();
5205 ciKlass* dest_k = top_dest->klass();
5206 if (!src_spec) {
5207 src_k = src_type->speculative_type_not_null();
5208 if (src_k != NULL && src_k->is_array_klass()) {
5209 could_have_src = true;
5210 }
5211 }
5212 if (!dest_spec) {
5213 dest_k = dest_type->speculative_type_not_null();
5214 if (dest_k != NULL && dest_k->is_array_klass()) {
5215 could_have_dest = true;
5216 }
5217 }
5218 if (could_have_src && could_have_dest) {
5219 // If we can have both exact types, emit the missing guards
5220 if (could_have_src && !src_spec) {
5221 src = maybe_cast_profiled_obj(src, src_k, true);
5222 }
5223 if (could_have_dest && !dest_spec) {
5224 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5225 }
5226 }
5227 }
5228 }
5229
5230 ciMethod* trap_method = method();
5231 int trap_bci = bci();
5232 if (saved_jvms != NULL) {
5233 trap_method = alloc->jvms()->method();
5234 trap_bci = alloc->jvms()->bci();
5235 }
5236
5237 bool negative_length_guard_generated = false;
5238
5239 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5240 can_emit_guards &&
5241 !src->is_top() && !dest->is_top()) {
5242 // validate arguments: enables transformation the ArrayCopyNode
5243 validated = true;
5244
5245 RegionNode* slow_region = new RegionNode(1);
5246 record_for_igvn(slow_region);
5247
5248 // (1) src and dest are arrays.
5249 generate_non_array_guard(load_object_klass(src), slow_region);
5250 generate_non_array_guard(load_object_klass(dest), slow_region);
5251
5252 // (2) src and dest arrays must have elements of the same BasicType
5253 // done at macro expansion or at Ideal transformation time
5254
5255 // (4) src_offset must not be negative.
5256 generate_negative_guard(src_offset, slow_region);
5257
5258 // (5) dest_offset must not be negative.
5259 generate_negative_guard(dest_offset, slow_region);
5260
5261 // (7) src_offset + length must not exceed length of src.
5262 generate_limit_guard(src_offset, length,
5263 load_array_length(src),
5264 slow_region);
5265
5266 // (8) dest_offset + length must not exceed length of dest.
5267 generate_limit_guard(dest_offset, length,
5268 load_array_length(dest),
5269 slow_region);
5270
5271 // (6) length must not be negative.
5272 // This is also checked in generate_arraycopy() during macro expansion, but
5273 // we also have to check it here for the case where the ArrayCopyNode will
5274 // be eliminated by Escape Analysis.
5275 if (EliminateAllocations) {
5276 generate_negative_guard(length, slow_region);
5277 negative_length_guard_generated = true;
5278 }
5279
5280 // (9) each element of an oop array must be assignable
5281 Node* src_klass = load_object_klass(src);
5282 Node* dest_klass = load_object_klass(dest);
5283 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5284
5285 if (not_subtype_ctrl != top()) {
5286 PreserveJVMState pjvms(this);
5287 set_control(not_subtype_ctrl);
5288 uncommon_trap(Deoptimization::Reason_intrinsic,
5289 Deoptimization::Action_make_not_entrant);
5290 assert(stopped(), "Should be stopped");
5291 }
5292 {
5293 PreserveJVMState pjvms(this);
5294 set_control(_gvn.transform(slow_region));
5295 uncommon_trap(Deoptimization::Reason_intrinsic,
5296 Deoptimization::Action_make_not_entrant);
5297 assert(stopped(), "Should be stopped");
5298 }
5299 }
5300
5301 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5302
5303 if (stopped()) {
5304 return true;
5305 }
5306
5307 Node* new_src = shenandoah_read_barrier(src);
5308 Node* new_dest = shenandoah_write_barrier(dest);
5309
5310 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, new_src, src_offset, new_dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5311 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5312 // so the compiler has a chance to eliminate them: during macro expansion,
5313 // we have to set their control (CastPP nodes are eliminated).
5314 load_object_klass(src), load_object_klass(dest),
5315 load_array_length(src), load_array_length(dest));
5316
5317 ac->set_arraycopy(validated);
5318
5319 Node* n = _gvn.transform(ac);
5320 if (n == ac) {
5321 ac->_adr_type = TypePtr::BOTTOM;
5322 ac->connect_outputs(this);
5323 } else {
5324 assert(validated, "shouldn't transform if all arguments not validated");
5325 set_all_memory(n);
5326 }
5327
5328 return true;
5329 }
5330
5331
5332 // Helper function which determines if an arraycopy immediately follows
5333 // an allocation, with no intervening tests or other escapes for the object.
5334 AllocateArrayNode*
5335 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5336 RegionNode* slow_region) {
5337 if (stopped()) return NULL; // no fast path
5338 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5339
5340 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr);
5341
5342 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5343 if (alloc == NULL) return NULL;
5344
5345 Node* rawmem = memory(Compile::AliasIdxRaw);
5346 // Is the allocation's memory state untouched?
5347 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5348 // Bail out if there have been raw-memory effects since the allocation.
5349 // (Example: There might have been a call or safepoint.)
5350 return NULL;
5351 }
5352 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5353 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5354 return NULL;
5355 }
5356
5357 // There must be no unexpected observers of this allocation.
5358 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5359 Node* obs = ptr->fast_out(i);
5360 if (obs != this->map()) {
5361 return NULL;
5362 }
5363 }
5364
5365 // This arraycopy must unconditionally follow the allocation of the ptr.
5366 Node* alloc_ctl = ptr->in(0);
5367 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5368
5369 Node* ctl = control();
5370 while (ctl != alloc_ctl) {
5371 // There may be guards which feed into the slow_region.
5372 // Any other control flow means that we might not get a chance
5373 // to finish initializing the allocated object.
5374 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5375 IfNode* iff = ctl->in(0)->as_If();
5376 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5377 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5378 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5379 ctl = iff->in(0); // This test feeds the known slow_region.
5380 continue;
5381 }
5382 // One more try: Various low-level checks bottom out in
5383 // uncommon traps. If the debug-info of the trap omits
5384 // any reference to the allocation, as we've already
5385 // observed, then there can be no objection to the trap.
5386 bool found_trap = false;
5387 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5388 Node* obs = not_ctl->fast_out(j);
5389 if (obs->in(0) == not_ctl && obs->is_Call() &&
5390 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5391 found_trap = true; break;
5392 }
5393 }
5394 if (found_trap) {
5395 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5396 continue;
5397 }
5398 }
5399 return NULL;
5400 }
5401
5402 // If we get this far, we have an allocation which immediately
5403 // precedes the arraycopy, and we can take over zeroing the new object.
5404 // The arraycopy will finish the initialization, and provide
5405 // a new control state to which we will anchor the destination pointer.
5406
5407 return alloc;
5408 }
5409
5410 //-------------inline_encodeISOArray-----------------------------------
5411 // encode char[] to byte[] in ISO_8859_1
5412 bool LibraryCallKit::inline_encodeISOArray() {
5413 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5414 // no receiver since it is static method
5415 Node *src = argument(0);
5416 Node *src_offset = argument(1);
5417 Node *dst = argument(2);
5418 Node *dst_offset = argument(3);
5419 Node *length = argument(4);
5420
5421 src = cast_not_null(src, false);
5422 dst = cast_not_null(dst, false);
5423
5424 src = shenandoah_read_barrier(src);
5425 dst = shenandoah_write_barrier(dst);
5426
5427 const Type* src_type = src->Value(&_gvn);
5428 const Type* dst_type = dst->Value(&_gvn);
5429 const TypeAryPtr* top_src = src_type->isa_aryptr();
5430 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5431 if (top_src == NULL || top_src->klass() == NULL ||
5432 top_dest == NULL || top_dest->klass() == NULL) {
5433 // failed array check
5434 return false;
5435 }
5436
5437 // Figure out the size and type of the elements we will be copying.
5438 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5439 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5440 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5441 return false;
5442 }
5443
5444 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5445 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5446 // 'src_start' points to src array + scaled offset
5447 // 'dst_start' points to dst array + scaled offset
5448
5449 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5450 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5451 enc = _gvn.transform(enc);
5452 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5453 set_memory(res_mem, mtype);
5454 set_result(enc);
5455 return true;
5456 }
5457
5458 //-------------inline_multiplyToLen-----------------------------------
5459 bool LibraryCallKit::inline_multiplyToLen() {
5460 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5461
5462 address stubAddr = StubRoutines::multiplyToLen();
5463 if (stubAddr == NULL) {
5464 return false; // Intrinsic's stub is not implemented on this platform
5465 }
5466 const char* stubName = "multiplyToLen";
5467
5468 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5469
5470 // no receiver because it is a static method
5471 Node* x = argument(0);
5472 Node* xlen = argument(1);
5473 Node* y = argument(2);
5474 Node* ylen = argument(3);
5475 Node* z = argument(4);
5476
5477 x = cast_not_null(x, false);
5478 x = shenandoah_read_barrier(x);
5479 y = cast_not_null(y, false);
5480 y = shenandoah_read_barrier(y);
5481 z = shenandoah_write_barrier(z);
5482
5483 const Type* x_type = x->Value(&_gvn);
5484 const Type* y_type = y->Value(&_gvn);
5485 const TypeAryPtr* top_x = x_type->isa_aryptr();
5486 const TypeAryPtr* top_y = y_type->isa_aryptr();
5487 if (top_x == NULL || top_x->klass() == NULL ||
5488 top_y == NULL || top_y->klass() == NULL) {
5489 // failed array check
5490 return false;
5491 }
5492
5493 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5494 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5495 if (x_elem != T_INT || y_elem != T_INT) {
5496 return false;
5497 }
5498
5499 // Set the original stack and the reexecute bit for the interpreter to reexecute
5500 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5501 // on the return from z array allocation in runtime.
5502 { PreserveReexecuteState preexecs(this);
5503 jvms()->set_should_reexecute(true);
5504
5505 Node* x_start = array_element_address(x, intcon(0), x_elem);
5506 Node* y_start = array_element_address(y, intcon(0), y_elem);
5507 // 'x_start' points to x array + scaled xlen
5508 // 'y_start' points to y array + scaled ylen
5509
5510 // Allocate the result array
5511 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5512 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5513 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5514
5515 IdealKit ideal(this);
5516
5517 #define __ ideal.
5518 Node* one = __ ConI(1);
5519 Node* zero = __ ConI(0);
5520 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5521 __ set(need_alloc, zero);
5522 __ set(z_alloc, z);
5523 __ if_then(z, BoolTest::eq, null()); {
5524 __ increment (need_alloc, one);
5525 } __ else_(); {
5526 // Update graphKit memory and control from IdealKit.
5527 sync_kit(ideal);
5528 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5529 cast->init_req(0, control());
5530 _gvn.set_type(cast, cast->bottom_type());
5531 C->record_for_igvn(cast);
5532
5533 Node* zlen_arg = load_array_length(cast);
5534 // Update IdealKit memory and control from graphKit.
5535 __ sync_kit(this);
5536 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5537 __ increment (need_alloc, one);
5538 } __ end_if();
5539 } __ end_if();
5540
5541 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5542 // Update graphKit memory and control from IdealKit.
5543 sync_kit(ideal);
5544 Node * narr = new_array(klass_node, zlen, 1);
5545 // Update IdealKit memory and control from graphKit.
5546 __ sync_kit(this);
5547 __ set(z_alloc, narr);
5548 } __ end_if();
5549
5550 sync_kit(ideal);
5551 z = __ value(z_alloc);
5552 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5553 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5554 // Final sync IdealKit and GraphKit.
5555 final_sync(ideal);
5556 #undef __
5557
5558 Node* z_start = array_element_address(z, intcon(0), T_INT);
5559
5560 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5561 OptoRuntime::multiplyToLen_Type(),
5562 stubAddr, stubName, TypePtr::BOTTOM,
5563 x_start, xlen, y_start, ylen, z_start, zlen);
5564 } // original reexecute is set back here
5565
5566 C->set_has_split_ifs(true); // Has chance for split-if optimization
5567 set_result(z);
5568 return true;
5569 }
5570
5571 //-------------inline_squareToLen------------------------------------
5572 bool LibraryCallKit::inline_squareToLen() {
5573 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5574
5575 address stubAddr = StubRoutines::squareToLen();
5576 if (stubAddr == NULL) {
5577 return false; // Intrinsic's stub is not implemented on this platform
5578 }
5579 const char* stubName = "squareToLen";
5580
5581 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5582
5583 Node* x = argument(0);
5584 Node* len = argument(1);
5585 Node* z = argument(2);
5586 Node* zlen = argument(3);
5587
5588 x = cast_not_null(x, false);
5589 x = shenandoah_read_barrier(x);
5590 z = cast_not_null(z, false);
5591 z = shenandoah_write_barrier(z);
5592
5593 const Type* x_type = x->Value(&_gvn);
5594 const Type* z_type = z->Value(&_gvn);
5595 const TypeAryPtr* top_x = x_type->isa_aryptr();
5596 const TypeAryPtr* top_z = z_type->isa_aryptr();
5597 if (top_x == NULL || top_x->klass() == NULL ||
5598 top_z == NULL || top_z->klass() == NULL) {
5599 // failed array check
5600 return false;
5601 }
5602
5603 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5604 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5605 if (x_elem != T_INT || z_elem != T_INT) {
5606 return false;
5607 }
5608
5609
5610 Node* x_start = array_element_address(x, intcon(0), x_elem);
5611 Node* z_start = array_element_address(z, intcon(0), z_elem);
5612
5613 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5614 OptoRuntime::squareToLen_Type(),
5615 stubAddr, stubName, TypePtr::BOTTOM,
5616 x_start, len, z_start, zlen);
5617
5618 set_result(z);
5619 return true;
5620 }
5621
5622 //-------------inline_mulAdd------------------------------------------
5623 bool LibraryCallKit::inline_mulAdd() {
5624 assert(UseMulAddIntrinsic, "not implemented on this platform");
5625
5626 address stubAddr = StubRoutines::mulAdd();
5627 if (stubAddr == NULL) {
5628 return false; // Intrinsic's stub is not implemented on this platform
5629 }
5630 const char* stubName = "mulAdd";
5631
5632 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5633
5634 Node* out = argument(0);
5635 Node* in = argument(1);
5636 Node* offset = argument(2);
5637 Node* len = argument(3);
5638 Node* k = argument(4);
5639
5640 in = shenandoah_read_barrier(in);
5641 out = cast_not_null(out, false);
5642 out = shenandoah_write_barrier(out);
5643
5644 const Type* out_type = out->Value(&_gvn);
5645 const Type* in_type = in->Value(&_gvn);
5646 const TypeAryPtr* top_out = out_type->isa_aryptr();
5647 const TypeAryPtr* top_in = in_type->isa_aryptr();
5648 if (top_out == NULL || top_out->klass() == NULL ||
5649 top_in == NULL || top_in->klass() == NULL) {
5650 // failed array check
5651 return false;
5652 }
5653
5654 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5655 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5656 if (out_elem != T_INT || in_elem != T_INT) {
5657 return false;
5658 }
5659
5660 Node* outlen = load_array_length(out);
5661 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5662 Node* out_start = array_element_address(out, intcon(0), out_elem);
5663 Node* in_start = array_element_address(in, intcon(0), in_elem);
5664
5665 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5666 OptoRuntime::mulAdd_Type(),
5667 stubAddr, stubName, TypePtr::BOTTOM,
5668 out_start,in_start, new_offset, len, k);
5669 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5670 set_result(result);
5671 return true;
5672 }
5673
5674 //-------------inline_montgomeryMultiply-----------------------------------
5675 bool LibraryCallKit::inline_montgomeryMultiply() {
5676 address stubAddr = StubRoutines::montgomeryMultiply();
5677 if (stubAddr == NULL) {
5678 return false; // Intrinsic's stub is not implemented on this platform
5679 }
5680
5681 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5682 const char* stubName = "montgomery_multiply";
5683
5684 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5685
5686 Node* a = argument(0);
5687 Node* b = argument(1);
5688 Node* n = argument(2);
5689 Node* len = argument(3);
5690 Node* inv = argument(4);
5691 Node* m = argument(6);
5692
5693 a = shenandoah_read_barrier(a);
5694 b = shenandoah_read_barrier(b);
5695 n = shenandoah_read_barrier(n);
5696 m = shenandoah_write_barrier(m);
5697
5698 const Type* a_type = a->Value(&_gvn);
5699 const TypeAryPtr* top_a = a_type->isa_aryptr();
5700 const Type* b_type = b->Value(&_gvn);
5701 const TypeAryPtr* top_b = b_type->isa_aryptr();
5702 const Type* n_type = a->Value(&_gvn);
5703 const TypeAryPtr* top_n = n_type->isa_aryptr();
5704 const Type* m_type = a->Value(&_gvn);
5705 const TypeAryPtr* top_m = m_type->isa_aryptr();
5706 if (top_a == NULL || top_a->klass() == NULL ||
5707 top_b == NULL || top_b->klass() == NULL ||
5708 top_n == NULL || top_n->klass() == NULL ||
5709 top_m == NULL || top_m->klass() == NULL) {
5710 // failed array check
5711 return false;
5712 }
5713
5714 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5715 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5716 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5717 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5718 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5719 return false;
5720 }
5721
5722 // Make the call
5723 {
5724 Node* a_start = array_element_address(a, intcon(0), a_elem);
5725 Node* b_start = array_element_address(b, intcon(0), b_elem);
5726 Node* n_start = array_element_address(n, intcon(0), n_elem);
5727 Node* m_start = array_element_address(m, intcon(0), m_elem);
5728
5729 Node* call = make_runtime_call(RC_LEAF,
5730 OptoRuntime::montgomeryMultiply_Type(),
5731 stubAddr, stubName, TypePtr::BOTTOM,
5732 a_start, b_start, n_start, len, inv, top(),
5733 m_start);
5734 set_result(m);
5735 }
5736
5737 return true;
5738 }
5739
5740 bool LibraryCallKit::inline_montgomerySquare() {
5741 address stubAddr = StubRoutines::montgomerySquare();
5742 if (stubAddr == NULL) {
5743 return false; // Intrinsic's stub is not implemented on this platform
5744 }
5745
5746 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5747 const char* stubName = "montgomery_square";
5748
5749 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5750
5751 Node* a = argument(0);
5752 Node* n = argument(1);
5753 Node* len = argument(2);
5754 Node* inv = argument(3);
5755 Node* m = argument(5);
5756
5757 a = shenandoah_read_barrier(a);
5758 n = shenandoah_read_barrier(n);
5759 m = shenandoah_write_barrier(m);
5760
5761 const Type* a_type = a->Value(&_gvn);
5762 const TypeAryPtr* top_a = a_type->isa_aryptr();
5763 const Type* n_type = a->Value(&_gvn);
5764 const TypeAryPtr* top_n = n_type->isa_aryptr();
5765 const Type* m_type = a->Value(&_gvn);
5766 const TypeAryPtr* top_m = m_type->isa_aryptr();
5767 if (top_a == NULL || top_a->klass() == NULL ||
5768 top_n == NULL || top_n->klass() == NULL ||
5769 top_m == NULL || top_m->klass() == NULL) {
5770 // failed array check
5771 return false;
5772 }
5773
5774 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5775 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5776 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5777 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5778 return false;
5779 }
5780
5781 // Make the call
5782 {
5783 Node* a_start = array_element_address(a, intcon(0), a_elem);
5784 Node* n_start = array_element_address(n, intcon(0), n_elem);
5785 Node* m_start = array_element_address(m, intcon(0), m_elem);
5786
5787 Node* call = make_runtime_call(RC_LEAF,
5788 OptoRuntime::montgomerySquare_Type(),
5789 stubAddr, stubName, TypePtr::BOTTOM,
5790 a_start, n_start, len, inv, top(),
5791 m_start);
5792 set_result(m);
5793 }
5794
5795 return true;
5796 }
5797
5798 //-------------inline_vectorizedMismatch------------------------------
5799 bool LibraryCallKit::inline_vectorizedMismatch() {
5800 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5801
5802 address stubAddr = StubRoutines::vectorizedMismatch();
5803 if (stubAddr == NULL) {
5804 return false; // Intrinsic's stub is not implemented on this platform
5805 }
5806 const char* stubName = "vectorizedMismatch";
5807 int size_l = callee()->signature()->size();
5808 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5809
5810 Node* obja = argument(0);
5811 Node* aoffset = argument(1);
5812 Node* objb = argument(3);
5813 Node* boffset = argument(4);
5814 Node* length = argument(6);
5815 Node* scale = argument(7);
5816
5817 const Type* a_type = obja->Value(&_gvn);
5818 const Type* b_type = objb->Value(&_gvn);
5819 const TypeAryPtr* top_a = a_type->isa_aryptr();
5820 const TypeAryPtr* top_b = b_type->isa_aryptr();
5821 if (top_a == NULL || top_a->klass() == NULL ||
5822 top_b == NULL || top_b->klass() == NULL) {
5823 // failed array check
5824 return false;
5825 }
5826
5827 Node* call;
5828 jvms()->set_should_reexecute(true);
5829
5830 Node* obja_adr = make_unsafe_address(obja, aoffset, false);
5831 Node* objb_adr = make_unsafe_address(objb, boffset, false);
5832
5833 call = make_runtime_call(RC_LEAF,
5834 OptoRuntime::vectorizedMismatch_Type(),
5835 stubAddr, stubName, TypePtr::BOTTOM,
5836 obja_adr, objb_adr, length, scale);
5837
5838 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5839 set_result(result);
5840 return true;
5841 }
5842
5843 /**
5844 * Calculate CRC32 for byte.
5845 * int java.util.zip.CRC32.update(int crc, int b)
5846 */
5847 bool LibraryCallKit::inline_updateCRC32() {
5848 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5849 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5850 // no receiver since it is static method
5851 Node* crc = argument(0); // type: int
5852 Node* b = argument(1); // type: int
5853
5854 /*
5855 * int c = ~ crc;
5856 * b = timesXtoThe32[(b ^ c) & 0xFF];
5857 * b = b ^ (c >>> 8);
5858 * crc = ~b;
5859 */
5860
5861 Node* M1 = intcon(-1);
5862 crc = _gvn.transform(new XorINode(crc, M1));
5863 Node* result = _gvn.transform(new XorINode(crc, b));
5864 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5865
5866 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5867 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5868 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5869 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5870
5871 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5872 result = _gvn.transform(new XorINode(crc, result));
5873 result = _gvn.transform(new XorINode(result, M1));
5874 set_result(result);
5875 return true;
5876 }
5877
5878 /**
5879 * Calculate CRC32 for byte[] array.
5880 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5881 */
5882 bool LibraryCallKit::inline_updateBytesCRC32() {
5883 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5884 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5885 // no receiver since it is static method
5886 Node* crc = argument(0); // type: int
5887 Node* src = argument(1); // type: oop
5888 Node* offset = argument(2); // type: int
5889 Node* length = argument(3); // type: int
5890
5891 const Type* src_type = src->Value(&_gvn);
5892 const TypeAryPtr* top_src = src_type->isa_aryptr();
5893 if (top_src == NULL || top_src->klass() == NULL) {
5894 // failed array check
5895 return false;
5896 }
5897
5898 // Figure out the size and type of the elements we will be copying.
5899 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5900 if (src_elem != T_BYTE) {
5901 return false;
5902 }
5903
5904 // 'src_start' points to src array + scaled offset
5905 src = cast_not_null(src, false);
5906 src = shenandoah_read_barrier(src);
5907 Node* src_start = array_element_address(src, offset, src_elem);
5908
5909 // We assume that range check is done by caller.
5910 // TODO: generate range check (offset+length < src.length) in debug VM.
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 /**
5925 * Calculate CRC32 for ByteBuffer.
5926 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5927 */
5928 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5929 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5930 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5931 // no receiver since it is static method
5932 Node* crc = argument(0); // type: int
5933 Node* src = argument(1); // type: long
5934 Node* offset = argument(3); // type: int
5935 Node* length = argument(4); // type: int
5936
5937 src = ConvL2X(src); // adjust Java long to machine word
5938 Node* base = _gvn.transform(new CastX2PNode(src));
5939 offset = ConvI2X(offset);
5940
5941 // 'src_start' points to src array + scaled offset
5942 Node* src_start = basic_plus_adr(top(), base, offset);
5943
5944 // Call the stub.
5945 address stubAddr = StubRoutines::updateBytesCRC32();
5946 const char *stubName = "updateBytesCRC32";
5947
5948 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5949 stubAddr, stubName, TypePtr::BOTTOM,
5950 crc, src_start, length);
5951 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5952 set_result(result);
5953 return true;
5954 }
5955
5956 //------------------------------get_table_from_crc32c_class-----------------------
5957 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5958 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5959 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5960
5961 return table;
5962 }
5963
5964 //------------------------------inline_updateBytesCRC32C-----------------------
5965 //
5966 // Calculate CRC32C for byte[] array.
5967 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5968 //
5969 bool LibraryCallKit::inline_updateBytesCRC32C() {
5970 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5971 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5972 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5973 // no receiver since it is a static method
5974 Node* crc = argument(0); // type: int
5975 Node* src = argument(1); // type: oop
5976 Node* offset = argument(2); // type: int
5977 Node* end = argument(3); // type: int
5978
5979 Node* length = _gvn.transform(new SubINode(end, offset));
5980
5981 const Type* src_type = src->Value(&_gvn);
5982 const TypeAryPtr* top_src = src_type->isa_aryptr();
5983 if (top_src == NULL || top_src->klass() == NULL) {
5984 // failed array check
5985 return false;
5986 }
5987
5988 // Figure out the size and type of the elements we will be copying.
5989 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5990 if (src_elem != T_BYTE) {
5991 return false;
5992 }
5993
5994 // 'src_start' points to src array + scaled offset
5995 src = shenandoah_read_barrier(src);
5996 src = cast_not_null(src, false);
5997 Node* src_start = array_element_address(src, offset, src_elem);
5998
5999 // static final int[] byteTable in class CRC32C
6000 Node* table = get_table_from_crc32c_class(callee()->holder());
6001 table = cast_not_null(table, false);
6002 table = shenandoah_read_barrier(table);
6003 Node* table_start = array_element_address(table, intcon(0), T_INT);
6004
6005 // We assume that range check is done by caller.
6006 // TODO: generate range check (offset+length < src.length) in debug VM.
6007
6008 // Call the stub.
6009 address stubAddr = StubRoutines::updateBytesCRC32C();
6010 const char *stubName = "updateBytesCRC32C";
6011
6012 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6013 stubAddr, stubName, TypePtr::BOTTOM,
6014 crc, src_start, length, table_start);
6015 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6016 set_result(result);
6017 return true;
6018 }
6019
6020 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
6021 //
6022 // Calculate CRC32C for DirectByteBuffer.
6023 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
6024 //
6025 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
6026 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
6027 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
6028 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
6029 // no receiver since it is a static method
6030 Node* crc = argument(0); // type: int
6031 Node* src = argument(1); // type: long
6032 Node* offset = argument(3); // type: int
6033 Node* end = argument(4); // type: int
6034
6035 Node* length = _gvn.transform(new SubINode(end, offset));
6036
6037 src = ConvL2X(src); // adjust Java long to machine word
6038 Node* base = _gvn.transform(new CastX2PNode(src));
6039 offset = ConvI2X(offset);
6040
6041 // 'src_start' points to src array + scaled offset
6042 Node* src_start = basic_plus_adr(top(), base, offset);
6043
6044 // static final int[] byteTable in class CRC32C
6045 Node* table = get_table_from_crc32c_class(callee()->holder());
6046 table = cast_not_null(table, false);
6047 table = shenandoah_read_barrier(table);
6048 Node* table_start = array_element_address(table, intcon(0), T_INT);
6049
6050 // Call the stub.
6051 address stubAddr = StubRoutines::updateBytesCRC32C();
6052 const char *stubName = "updateBytesCRC32C";
6053
6054 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6055 stubAddr, stubName, TypePtr::BOTTOM,
6056 crc, src_start, length, table_start);
6057 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6058 set_result(result);
6059 return true;
6060 }
6061
6062 //------------------------------inline_updateBytesAdler32----------------------
6063 //
6064 // Calculate Adler32 checksum for byte[] array.
6065 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6066 //
6067 bool LibraryCallKit::inline_updateBytesAdler32() {
6068 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6069 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6070 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6071 // no receiver since it is static method
6072 Node* crc = argument(0); // type: int
6073 Node* src = argument(1); // type: oop
6074 Node* offset = argument(2); // type: int
6075 Node* length = argument(3); // type: int
6076
6077 const Type* src_type = src->Value(&_gvn);
6078 const TypeAryPtr* top_src = src_type->isa_aryptr();
6079 if (top_src == NULL || top_src->klass() == NULL) {
6080 // failed array check
6081 return false;
6082 }
6083
6084 // Figure out the size and type of the elements we will be copying.
6085 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6086 if (src_elem != T_BYTE) {
6087 return false;
6088 }
6089
6090 // 'src_start' points to src array + scaled offset
6091 src = shenandoah_read_barrier(src);
6092 Node* src_start = array_element_address(src, offset, src_elem);
6093
6094 // We assume that range check is done by caller.
6095 // TODO: generate range check (offset+length < src.length) in debug VM.
6096
6097 // Call the stub.
6098 address stubAddr = StubRoutines::updateBytesAdler32();
6099 const char *stubName = "updateBytesAdler32";
6100
6101 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6102 stubAddr, stubName, TypePtr::BOTTOM,
6103 crc, src_start, length);
6104 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6105 set_result(result);
6106 return true;
6107 }
6108
6109 //------------------------------inline_updateByteBufferAdler32---------------
6110 //
6111 // Calculate Adler32 checksum for DirectByteBuffer.
6112 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6113 //
6114 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6115 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6116 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6117 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6118 // no receiver since it is static method
6119 Node* crc = argument(0); // type: int
6120 Node* src = argument(1); // type: long
6121 Node* offset = argument(3); // type: int
6122 Node* length = argument(4); // type: int
6123
6124 src = ConvL2X(src); // adjust Java long to machine word
6125 Node* base = _gvn.transform(new CastX2PNode(src));
6126 offset = ConvI2X(offset);
6127
6128 // 'src_start' points to src array + scaled offset
6129 Node* src_start = basic_plus_adr(top(), base, offset);
6130
6131 // Call the stub.
6132 address stubAddr = StubRoutines::updateBytesAdler32();
6133 const char *stubName = "updateBytesAdler32";
6134
6135 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6136 stubAddr, stubName, TypePtr::BOTTOM,
6137 crc, src_start, length);
6138
6139 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6140 set_result(result);
6141 return true;
6142 }
6143
6144 //----------------------------inline_reference_get----------------------------
6145 // public T java.lang.ref.Reference.get();
6146 bool LibraryCallKit::inline_reference_get() {
6147 const int referent_offset = java_lang_ref_Reference::referent_offset;
6148 guarantee(referent_offset > 0, "should have already been set");
6149
6150 // Get the argument:
6151 Node* reference_obj = null_check_receiver();
6152 if (stopped()) return true;
6153
6154 if (ShenandoahVerifyReadsToFromSpace) {
6155 reference_obj = shenandoah_read_barrier(reference_obj);
6156 }
6157
6158 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6159
6160 ciInstanceKlass* klass = env()->Object_klass();
6161 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6162
6163 Node* no_ctrl = NULL;
6164 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6165
6166 // Use the pre-barrier to record the value in the referent field
6167 pre_barrier(false /* do_load */,
6168 control(),
6169 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6170 result /* pre_val */,
6171 T_OBJECT);
6172
6173 // Add memory barrier to prevent commoning reads from this field
6174 // across safepoint since GC can change its value.
6175 insert_mem_bar(Op_MemBarCPUOrder);
6176
6177 set_result(result);
6178 return true;
6179 }
6180
6181
6182 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6183 bool is_exact=true, bool is_static=false,
6184 ciInstanceKlass * fromKls=NULL) {
6185 if (fromKls == NULL) {
6186 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6187 assert(tinst != NULL, "obj is null");
6188 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6189 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6190 fromKls = tinst->klass()->as_instance_klass();
6191 } else {
6192 assert(is_static, "only for static field access");
6193 }
6194 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6195 ciSymbol::make(fieldTypeString),
6196 is_static);
6197
6198 assert (field != NULL, "undefined field");
6199 if (field == NULL) return (Node *) NULL;
6200
6201 if (is_static) {
6202 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6203 fromObj = makecon(tip);
6204 }
6205
6206 if (! ShenandoahOptimizeFinals || (! field->is_final() && ! field->is_stable())) {
6207 fromObj = shenandoah_read_barrier(fromObj);
6208 }
6209
6210 // Next code copied from Parse::do_get_xxx():
6211
6212 // Compute address and memory type.
6213 int offset = field->offset_in_bytes();
6214 bool is_vol = field->is_volatile();
6215 ciType* field_klass = field->type();
6216 assert(field_klass->is_loaded(), "should be loaded");
6217 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6218 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6219 BasicType bt = field->layout_type();
6220
6221 // Build the resultant type of the load
6222 const Type *type;
6223 if (bt == T_OBJECT) {
6224 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6225 } else {
6226 type = Type::get_const_basic_type(bt);
6227 }
6228
6229 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6230 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6231 }
6232 // Build the load.
6233 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6234 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6235 // If reference is volatile, prevent following memory ops from
6236 // floating up past the volatile read. Also prevents commoning
6237 // another volatile read.
6238 if (is_vol) {
6239 // Memory barrier includes bogus read of value to force load BEFORE membar
6240 insert_mem_bar(Op_MemBarAcquire, loadedField);
6241 }
6242 return loadedField;
6243 }
6244
6245 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6246 bool is_exact = true, bool is_static = false,
6247 ciInstanceKlass * fromKls = NULL) {
6248 if (fromKls == NULL) {
6249 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6250 assert(tinst != NULL, "obj is null");
6251 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6252 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6253 fromKls = tinst->klass()->as_instance_klass();
6254 }
6255 else {
6256 assert(is_static, "only for static field access");
6257 }
6258 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6259 ciSymbol::make(fieldTypeString),
6260 is_static);
6261
6262 assert(field != NULL, "undefined field");
6263 assert(!field->is_volatile(), "not defined for volatile fields");
6264
6265 if (is_static) {
6266 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6267 fromObj = makecon(tip);
6268 }
6269
6270 // Next code copied from Parse::do_get_xxx():
6271
6272 // Compute address and memory type.
6273 int offset = field->offset_in_bytes();
6274 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6275
6276 return adr;
6277 }
6278
6279 //------------------------------inline_aescrypt_Block-----------------------
6280 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6281 address stubAddr = NULL;
6282 const char *stubName;
6283 assert(UseAES, "need AES instruction support");
6284
6285 switch(id) {
6286 case vmIntrinsics::_aescrypt_encryptBlock:
6287 stubAddr = StubRoutines::aescrypt_encryptBlock();
6288 stubName = "aescrypt_encryptBlock";
6289 break;
6290 case vmIntrinsics::_aescrypt_decryptBlock:
6291 stubAddr = StubRoutines::aescrypt_decryptBlock();
6292 stubName = "aescrypt_decryptBlock";
6293 break;
6294 }
6295 if (stubAddr == NULL) return false;
6296
6297 Node* aescrypt_object = argument(0);
6298 Node* src = argument(1);
6299 Node* src_offset = argument(2);
6300 Node* dest = argument(3);
6301 Node* dest_offset = argument(4);
6302
6303 // Resolve src and dest arrays for ShenandoahGC.
6304 src = shenandoah_read_barrier(src);
6305 dest = shenandoah_write_barrier(dest);
6306
6307 // (1) src and dest are arrays.
6308 const Type* src_type = src->Value(&_gvn);
6309 const Type* dest_type = dest->Value(&_gvn);
6310 const TypeAryPtr* top_src = src_type->isa_aryptr();
6311 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6312 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6313
6314 // for the quick and dirty code we will skip all the checks.
6315 // we are just trying to get the call to be generated.
6316 Node* src_start = src;
6317 Node* dest_start = dest;
6318 if (src_offset != NULL || dest_offset != NULL) {
6319 assert(src_offset != NULL && dest_offset != NULL, "");
6320 src_start = array_element_address(src, src_offset, T_BYTE);
6321 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6322 }
6323
6324 // now need to get the start of its expanded key array
6325 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6326 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6327 if (k_start == NULL) return false;
6328
6329 if (Matcher::pass_original_key_for_aes()) {
6330 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6331 // compatibility issues between Java key expansion and SPARC crypto instructions
6332 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6333 if (original_k_start == NULL) return false;
6334
6335 // Call the stub.
6336 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6337 stubAddr, stubName, TypePtr::BOTTOM,
6338 src_start, dest_start, k_start, original_k_start);
6339 } else {
6340 // Call the stub.
6341 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6342 stubAddr, stubName, TypePtr::BOTTOM,
6343 src_start, dest_start, k_start);
6344 }
6345
6346 return true;
6347 }
6348
6349 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6350 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6351 address stubAddr = NULL;
6352 const char *stubName = NULL;
6353
6354 assert(UseAES, "need AES instruction support");
6355
6356 switch(id) {
6357 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6358 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6359 stubName = "cipherBlockChaining_encryptAESCrypt";
6360 break;
6361 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6362 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6363 stubName = "cipherBlockChaining_decryptAESCrypt";
6364 break;
6365 }
6366 if (stubAddr == NULL) return false;
6367
6368 Node* cipherBlockChaining_object = argument(0);
6369 Node* src = argument(1);
6370 Node* src_offset = argument(2);
6371 Node* len = argument(3);
6372 Node* dest = argument(4);
6373 Node* dest_offset = argument(5);
6374
6375
6376 // inline_cipherBlockChaining_AESCrypt_predicate() has its own
6377 // barrier. This one should optimize away.
6378 src = cast_not_null(src, false);
6379 dest = cast_not_null(dest, false);
6380 src = shenandoah_read_barrier(src);
6381 dest = shenandoah_write_barrier(dest);
6382
6383 // (1) src and dest are arrays.
6384 const Type* src_type = src->Value(&_gvn);
6385 const Type* dest_type = dest->Value(&_gvn);
6386 const TypeAryPtr* top_src = src_type->isa_aryptr();
6387 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6388 assert (top_src != NULL && top_src->klass() != NULL
6389 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6390
6391 // checks are the responsibility of the caller
6392 Node* src_start = src;
6393 Node* dest_start = dest;
6394 if (src_offset != NULL || dest_offset != NULL) {
6395 assert(src_offset != NULL && dest_offset != NULL, "");
6396 src_start = array_element_address(src, src_offset, T_BYTE);
6397 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6398 }
6399
6400 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6401 // (because of the predicated logic executed earlier).
6402 // so we cast it here safely.
6403 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6404
6405 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6406 if (embeddedCipherObj == NULL) return false;
6407
6408 // cast it to what we know it will be at runtime
6409 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6410 assert(tinst != NULL, "CBC obj is null");
6411 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6412 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6413 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6414
6415 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6416 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6417 const TypeOopPtr* xtype = aklass->as_instance_type();
6418 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6419 aescrypt_object = _gvn.transform(aescrypt_object);
6420
6421 // we need to get the start of the aescrypt_object's expanded key array
6422 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6423 if (k_start == NULL) return false;
6424
6425 // similarly, get the start address of the r vector
6426 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6427
6428 objRvec = shenandoah_write_barrier(objRvec);
6429
6430 if (objRvec == NULL) return false;
6431 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6432
6433 Node* cbcCrypt;
6434 if (Matcher::pass_original_key_for_aes()) {
6435 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6436 // compatibility issues between Java key expansion and SPARC crypto instructions
6437 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6438 if (original_k_start == NULL) return false;
6439
6440 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6441 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6442 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6443 stubAddr, stubName, TypePtr::BOTTOM,
6444 src_start, dest_start, k_start, r_start, len, original_k_start);
6445 } else {
6446 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6447 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6448 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6449 stubAddr, stubName, TypePtr::BOTTOM,
6450 src_start, dest_start, k_start, r_start, len);
6451 }
6452
6453 // return cipher length (int)
6454 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6455 set_result(retvalue);
6456 return true;
6457 }
6458
6459 //------------------------------inline_counterMode_AESCrypt-----------------------
6460 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6461 assert(UseAES, "need AES instruction support");
6462 if (!UseAESCTRIntrinsics) return false;
6463
6464 address stubAddr = NULL;
6465 const char *stubName = NULL;
6466 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6467 stubAddr = StubRoutines::counterMode_AESCrypt();
6468 stubName = "counterMode_AESCrypt";
6469 }
6470 if (stubAddr == NULL) return false;
6471
6472 Node* counterMode_object = argument(0);
6473 Node* src = argument(1);
6474 Node* src_offset = argument(2);
6475 Node* len = argument(3);
6476 Node* dest = argument(4);
6477 Node* dest_offset = argument(5);
6478
6479 // (1) src and dest are arrays.
6480 const Type* src_type = src->Value(&_gvn);
6481 const Type* dest_type = dest->Value(&_gvn);
6482 const TypeAryPtr* top_src = src_type->isa_aryptr();
6483 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6484 assert(top_src != NULL && top_src->klass() != NULL &&
6485 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6486
6487 src = shenandoah_read_barrier(src);
6488 dest = shenandoah_write_barrier(dest);
6489 counterMode_object = shenandoah_write_barrier(counterMode_object);
6490
6491 // checks are the responsibility of the caller
6492 Node* src_start = src;
6493 Node* dest_start = dest;
6494 if (src_offset != NULL || dest_offset != NULL) {
6495 assert(src_offset != NULL && dest_offset != NULL, "");
6496 src_start = array_element_address(src, src_offset, T_BYTE);
6497 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6498 }
6499
6500 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6501 // (because of the predicated logic executed earlier).
6502 // so we cast it here safely.
6503 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6504 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6505 if (embeddedCipherObj == NULL) return false;
6506 // cast it to what we know it will be at runtime
6507 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6508 assert(tinst != NULL, "CTR obj is null");
6509 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6510 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6511 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6512 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6513 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6514 const TypeOopPtr* xtype = aklass->as_instance_type();
6515 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6516 aescrypt_object = _gvn.transform(aescrypt_object);
6517 // we need to get the start of the aescrypt_object's expanded key array
6518 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6519 if (k_start == NULL) return false;
6520 // similarly, get the start address of the r vector
6521 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6522 if (obj_counter == NULL) return false;
6523 obj_counter = shenandoah_write_barrier(obj_counter);
6524 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6525
6526 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6527 if (saved_encCounter == NULL) return false;
6528 saved_encCounter = shenandoah_write_barrier(saved_encCounter);
6529 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6530 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6531
6532 Node* ctrCrypt;
6533 if (Matcher::pass_original_key_for_aes()) {
6534 // no SPARC version for AES/CTR intrinsics now.
6535 return false;
6536 }
6537 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6538 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6539 OptoRuntime::counterMode_aescrypt_Type(),
6540 stubAddr, stubName, TypePtr::BOTTOM,
6541 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6542
6543 // return cipher length (int)
6544 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6545 set_result(retvalue);
6546 return true;
6547 }
6548
6549 //------------------------------get_key_start_from_aescrypt_object-----------------------
6550 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6551 #if defined(PPC64) || defined(S390)
6552 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6553 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6554 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6555 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6556 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6557 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6558 if (objSessionK == NULL) {
6559 return (Node *) NULL;
6560 }
6561 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6562 #else
6563 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6564 #endif // PPC64
6565 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6566 if (objAESCryptKey == NULL) return (Node *) NULL;
6567
6568 objAESCryptKey = shenandoah_read_barrier(objAESCryptKey);
6569
6570 // now have the array, need to get the start address of the K array
6571 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6572 return k_start;
6573 }
6574
6575 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6576 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6577 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6578 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6579 if (objAESCryptKey == NULL) return (Node *) NULL;
6580
6581 objAESCryptKey = shenandoah_read_barrier(objAESCryptKey);
6582
6583 // now have the array, need to get the start address of the lastKey array
6584 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6585 return original_k_start;
6586 }
6587
6588 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6589 // Return node representing slow path of predicate check.
6590 // the pseudo code we want to emulate with this predicate is:
6591 // for encryption:
6592 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6593 // for decryption:
6594 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6595 // note cipher==plain is more conservative than the original java code but that's OK
6596 //
6597 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6598 // The receiver was checked for NULL already.
6599 Node* objCBC = argument(0);
6600
6601 Node* src = argument(1);
6602 Node* dest = argument(4);
6603
6604 // Load embeddedCipher field of CipherBlockChaining object.
6605 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6606
6607 // get AESCrypt klass for instanceOf check
6608 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6609 // will have same classloader as CipherBlockChaining object
6610 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6611 assert(tinst != NULL, "CBCobj is null");
6612 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6613
6614 // we want to do an instanceof comparison against the AESCrypt class
6615 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6616 if (!klass_AESCrypt->is_loaded()) {
6617 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6618 Node* ctrl = control();
6619 set_control(top()); // no regular fast path
6620 return ctrl;
6621 }
6622
6623 // Resolve src and dest arrays for ShenandoahGC. Here because new
6624 // memory state is not handled by predicate logic in
6625 // inline_cipherBlockChaining_AESCrypt itself
6626 src = cast_not_null(src, false);
6627 dest = cast_not_null(dest, false);
6628 src = shenandoah_write_barrier(src);
6629 dest = shenandoah_write_barrier(dest);
6630
6631 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6632
6633 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6634 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6635 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6636
6637 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6638
6639 // for encryption, we are done
6640 if (!decrypting)
6641 return instof_false; // even if it is NULL
6642
6643 // for decryption, we need to add a further check to avoid
6644 // taking the intrinsic path when cipher and plain are the same
6645 // see the original java code for why.
6646 RegionNode* region = new RegionNode(3);
6647 region->init_req(1, instof_false);
6648
6649 Node* cmp_src_dest = cmp_objects(src, dest);
6650 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6651 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6652 region->init_req(2, src_dest_conjoint);
6653
6654 record_for_igvn(region);
6655 return _gvn.transform(region);
6656 }
6657
6658 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6659 // Return node representing slow path of predicate check.
6660 // the pseudo code we want to emulate with this predicate is:
6661 // for encryption:
6662 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6663 // for decryption:
6664 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6665 // note cipher==plain is more conservative than the original java code but that's OK
6666 //
6667
6668 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6669 // The receiver was checked for NULL already.
6670 Node* objCTR = argument(0);
6671
6672 // Load embeddedCipher field of CipherBlockChaining object.
6673 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6674
6675 // get AESCrypt klass for instanceOf check
6676 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6677 // will have same classloader as CipherBlockChaining object
6678 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6679 assert(tinst != NULL, "CTRobj is null");
6680 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6681
6682 // we want to do an instanceof comparison against the AESCrypt class
6683 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6684 if (!klass_AESCrypt->is_loaded()) {
6685 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6686 Node* ctrl = control();
6687 set_control(top()); // no regular fast path
6688 return ctrl;
6689 }
6690
6691 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6692 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6693 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6694 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6695 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6696
6697 return instof_false; // even if it is NULL
6698 }
6699
6700 //------------------------------inline_ghash_processBlocks
6701 bool LibraryCallKit::inline_ghash_processBlocks() {
6702 address stubAddr;
6703 const char *stubName;
6704 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6705
6706 stubAddr = StubRoutines::ghash_processBlocks();
6707 stubName = "ghash_processBlocks";
6708
6709 Node* data = argument(0);
6710 Node* offset = argument(1);
6711 Node* len = argument(2);
6712 Node* state = argument(3);
6713 Node* subkeyH = argument(4);
6714
6715 state = cast_not_null(state, false);
6716 subkeyH = cast_not_null(subkeyH, false);
6717 data = cast_not_null(data, false);
6718
6719 state = shenandoah_write_barrier(state);
6720 subkeyH = shenandoah_read_barrier(subkeyH);
6721 data = shenandoah_read_barrier(data);
6722
6723 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6724 assert(state_start, "state is NULL");
6725 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6726 assert(subkeyH_start, "subkeyH is NULL");
6727 Node* data_start = array_element_address(data, offset, T_BYTE);
6728 assert(data_start, "data is NULL");
6729
6730 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6731 OptoRuntime::ghash_processBlocks_Type(),
6732 stubAddr, stubName, TypePtr::BOTTOM,
6733 state_start, subkeyH_start, data_start, len);
6734 return true;
6735 }
6736
6737 //------------------------------inline_sha_implCompress-----------------------
6738 //
6739 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6740 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6741 //
6742 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6743 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6744 //
6745 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6746 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6747 //
6748 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6749 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6750
6751 Node* sha_obj = argument(0);
6752 Node* src = argument(1); // type oop
6753 Node* ofs = argument(2); // type int
6754
6755 const Type* src_type = src->Value(&_gvn);
6756 const TypeAryPtr* top_src = src_type->isa_aryptr();
6757 if (top_src == NULL || top_src->klass() == NULL) {
6758 // failed array check
6759 return false;
6760 }
6761 // Figure out the size and type of the elements we will be copying.
6762 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6763 if (src_elem != T_BYTE) {
6764 return false;
6765 }
6766 // 'src_start' points to src array + offset
6767 src = cast_not_null(src, false);
6768 src = shenandoah_read_barrier(src);
6769 Node* src_start = array_element_address(src, ofs, src_elem);
6770 Node* state = NULL;
6771 address stubAddr;
6772 const char *stubName;
6773
6774 switch(id) {
6775 case vmIntrinsics::_sha_implCompress:
6776 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6777 state = get_state_from_sha_object(sha_obj);
6778 stubAddr = StubRoutines::sha1_implCompress();
6779 stubName = "sha1_implCompress";
6780 break;
6781 case vmIntrinsics::_sha2_implCompress:
6782 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6783 state = get_state_from_sha_object(sha_obj);
6784 stubAddr = StubRoutines::sha256_implCompress();
6785 stubName = "sha256_implCompress";
6786 break;
6787 case vmIntrinsics::_sha5_implCompress:
6788 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6789 state = get_state_from_sha5_object(sha_obj);
6790 stubAddr = StubRoutines::sha512_implCompress();
6791 stubName = "sha512_implCompress";
6792 break;
6793 default:
6794 fatal_unexpected_iid(id);
6795 return false;
6796 }
6797 if (state == NULL) return false;
6798
6799 // Call the stub.
6800 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6801 stubAddr, stubName, TypePtr::BOTTOM,
6802 src_start, state);
6803
6804 return true;
6805 }
6806
6807 //------------------------------inline_digestBase_implCompressMB-----------------------
6808 //
6809 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6810 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6811 //
6812 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6813 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6814 "need SHA1/SHA256/SHA512 instruction support");
6815 assert((uint)predicate < 3, "sanity");
6816 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6817
6818 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6819 Node* src = argument(1); // byte[] array
6820 Node* ofs = argument(2); // type int
6821 Node* limit = argument(3); // type int
6822
6823 const Type* src_type = src->Value(&_gvn);
6824 const TypeAryPtr* top_src = src_type->isa_aryptr();
6825 if (top_src == NULL || top_src->klass() == NULL) {
6826 // failed array check
6827 return false;
6828 }
6829 // Figure out the size and type of the elements we will be copying.
6830 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6831 if (src_elem != T_BYTE) {
6832 return false;
6833 }
6834 // 'src_start' points to src array + offset
6835 src = cast_not_null(src, false);
6836 src = shenandoah_read_barrier(src);
6837 Node* src_start = array_element_address(src, ofs, src_elem);
6838
6839 const char* klass_SHA_name = NULL;
6840 const char* stub_name = NULL;
6841 address stub_addr = NULL;
6842 bool long_state = false;
6843
6844 switch (predicate) {
6845 case 0:
6846 if (UseSHA1Intrinsics) {
6847 klass_SHA_name = "sun/security/provider/SHA";
6848 stub_name = "sha1_implCompressMB";
6849 stub_addr = StubRoutines::sha1_implCompressMB();
6850 }
6851 break;
6852 case 1:
6853 if (UseSHA256Intrinsics) {
6854 klass_SHA_name = "sun/security/provider/SHA2";
6855 stub_name = "sha256_implCompressMB";
6856 stub_addr = StubRoutines::sha256_implCompressMB();
6857 }
6858 break;
6859 case 2:
6860 if (UseSHA512Intrinsics) {
6861 klass_SHA_name = "sun/security/provider/SHA5";
6862 stub_name = "sha512_implCompressMB";
6863 stub_addr = StubRoutines::sha512_implCompressMB();
6864 long_state = true;
6865 }
6866 break;
6867 default:
6868 fatal("unknown SHA intrinsic predicate: %d", predicate);
6869 }
6870 if (klass_SHA_name != NULL) {
6871 // get DigestBase klass to lookup for SHA klass
6872 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6873 assert(tinst != NULL, "digestBase_obj is not instance???");
6874 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6875
6876 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6877 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6878 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6879 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6880 }
6881 return false;
6882 }
6883 //------------------------------inline_sha_implCompressMB-----------------------
6884 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6885 bool long_state, address stubAddr, const char *stubName,
6886 Node* src_start, Node* ofs, Node* limit) {
6887 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6888 const TypeOopPtr* xtype = aklass->as_instance_type();
6889 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6890 sha_obj = _gvn.transform(sha_obj);
6891
6892 Node* state;
6893 if (long_state) {
6894 state = get_state_from_sha5_object(sha_obj);
6895 } else {
6896 state = get_state_from_sha_object(sha_obj);
6897 }
6898 if (state == NULL) return false;
6899
6900 // Call the stub.
6901 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6902 OptoRuntime::digestBase_implCompressMB_Type(),
6903 stubAddr, stubName, TypePtr::BOTTOM,
6904 src_start, state, ofs, limit);
6905 // return ofs (int)
6906 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6907 set_result(result);
6908
6909 return true;
6910 }
6911
6912 //------------------------------get_state_from_sha_object-----------------------
6913 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6914 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6915 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6916 if (sha_state == NULL) return (Node *) NULL;
6917
6918 sha_state = shenandoah_write_barrier(sha_state);
6919
6920 // now have the array, need to get the start address of the state array
6921 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6922 return state;
6923 }
6924
6925 //------------------------------get_state_from_sha5_object-----------------------
6926 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6927 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6928 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6929 if (sha_state == NULL) return (Node *) NULL;
6930
6931 sha_state = shenandoah_write_barrier(sha_state);
6932
6933 // now have the array, need to get the start address of the state array
6934 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6935 return state;
6936 }
6937
6938 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6939 // Return node representing slow path of predicate check.
6940 // the pseudo code we want to emulate with this predicate is:
6941 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6942 //
6943 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6944 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6945 "need SHA1/SHA256/SHA512 instruction support");
6946 assert((uint)predicate < 3, "sanity");
6947
6948 // The receiver was checked for NULL already.
6949 Node* digestBaseObj = argument(0);
6950
6951 // get DigestBase klass for instanceOf check
6952 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6953 assert(tinst != NULL, "digestBaseObj is null");
6954 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6955
6956 const char* klass_SHA_name = NULL;
6957 switch (predicate) {
6958 case 0:
6959 if (UseSHA1Intrinsics) {
6960 // we want to do an instanceof comparison against the SHA class
6961 klass_SHA_name = "sun/security/provider/SHA";
6962 }
6963 break;
6964 case 1:
6965 if (UseSHA256Intrinsics) {
6966 // we want to do an instanceof comparison against the SHA2 class
6967 klass_SHA_name = "sun/security/provider/SHA2";
6968 }
6969 break;
6970 case 2:
6971 if (UseSHA512Intrinsics) {
6972 // we want to do an instanceof comparison against the SHA5 class
6973 klass_SHA_name = "sun/security/provider/SHA5";
6974 }
6975 break;
6976 default:
6977 fatal("unknown SHA intrinsic predicate: %d", predicate);
6978 }
6979
6980 ciKlass* klass_SHA = NULL;
6981 if (klass_SHA_name != NULL) {
6982 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6983 }
6984 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6985 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6986 Node* ctrl = control();
6987 set_control(top()); // no intrinsic path
6988 return ctrl;
6989 }
6990 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6991
6992 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6993 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6994 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6995 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6996
6997 return instof_false; // even if it is NULL
6998 }
6999
7000 //-------------inline_fma-----------------------------------
7001 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
7002 Node *a = NULL;
7003 Node *b = NULL;
7004 Node *c = NULL;
7005 Node* result = NULL;
7006 switch (id) {
7007 case vmIntrinsics::_fmaD:
7008 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
7009 // no receiver since it is static method
7010 a = round_double_node(argument(0));
7011 b = round_double_node(argument(2));
7012 c = round_double_node(argument(4));
7013 result = _gvn.transform(new FmaDNode(control(), a, b, c));
7014 break;
7015 case vmIntrinsics::_fmaF:
7016 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
7017 a = argument(0);
7018 b = argument(1);
7019 c = argument(2);
7020 result = _gvn.transform(new FmaFNode(control(), a, b, c));
7021 break;
7022 default:
7023 fatal_unexpected_iid(id); break;
7024 }
7025 set_result(result);
7026 return true;
7027 }
7028
7029 bool LibraryCallKit::inline_profileBoolean() {
7030 Node* counts = argument(1);
7031 const TypeAryPtr* ary = NULL;
7032 ciArray* aobj = NULL;
7033 assert(!(ShenandoahBarrierNode::skip_through_barrier(counts)->is_Con() && !counts->is_Con()), "barrier prevents optimization");
7034 if (counts->is_Con()
7035 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
7036 && (aobj = ary->const_oop()->as_array()) != NULL
7037 && (aobj->length() == 2)) {
7038 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
7039 jint false_cnt = aobj->element_value(0).as_int();
7040 jint true_cnt = aobj->element_value(1).as_int();
7041
7042 if (C->log() != NULL) {
7043 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
7044 false_cnt, true_cnt);
7045 }
7046
7047 if (false_cnt + true_cnt == 0) {
7048 // According to profile, never executed.
7049 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7050 Deoptimization::Action_reinterpret);
7051 return true;
7052 }
7053
7054 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
7055 // is a number of each value occurrences.
7056 Node* result = argument(0);
7057 if (false_cnt == 0 || true_cnt == 0) {
7058 // According to profile, one value has been never seen.
7059 int expected_val = (false_cnt == 0) ? 1 : 0;
7060
7061 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
7062 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
7063
7064 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
7065 Node* fast_path = _gvn.transform(new IfTrueNode(check));
7066 Node* slow_path = _gvn.transform(new IfFalseNode(check));
7067
7068 { // Slow path: uncommon trap for never seen value and then reexecute
7069 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
7070 // the value has been seen at least once.
7071 PreserveJVMState pjvms(this);
7072 PreserveReexecuteState preexecs(this);
7073 jvms()->set_should_reexecute(true);
7074
7075 set_control(slow_path);
7076 set_i_o(i_o());
7077
7078 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
7079 Deoptimization::Action_reinterpret);
7080 }
7081 // The guard for never seen value enables sharpening of the result and
7082 // returning a constant. It allows to eliminate branches on the same value
7083 // later on.
7084 set_control(fast_path);
7085 result = intcon(expected_val);
7086 }
7087 // Stop profiling.
7088 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
7089 // By replacing method body with profile data (represented as ProfileBooleanNode
7090 // on IR level) we effectively disable profiling.
7091 // It enables full speed execution once optimized code is generated.
7092 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
7093 C->record_for_igvn(profile);
7094 set_result(profile);
7095 return true;
7096 } else {
7097 // Continue profiling.
7098 // Profile data isn't available at the moment. So, execute method's bytecode version.
7099 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
7100 // is compiled and counters aren't available since corresponding MethodHandle
7101 // isn't a compile-time constant.
7102 return false;
7103 }
7104 }
7105
7106 bool LibraryCallKit::inline_isCompileConstant() {
7107 Node* n = argument(0);
7108 set_result(n->is_Con() ? intcon(1) : intcon(0));
7109 return true;
7110 }