1 /*
2 * Copyright (c) 1999, 2018, 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 "ci/ciUtilities.inline.hpp"
28 #include "classfile/systemDictionary.hpp"
29 #include "classfile/vmSymbols.hpp"
30 #include "compiler/compileBroker.hpp"
31 #include "compiler/compileLog.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "jfr/support/jfrIntrinsics.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/objArrayKlass.hpp"
36 #include "opto/addnode.hpp"
37 #include "opto/arraycopynode.hpp"
38 #include "opto/c2compiler.hpp"
39 #include "opto/callGenerator.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/cfgnode.hpp"
42 #include "opto/convertnode.hpp"
43 #include "opto/countbitsnode.hpp"
44 #include "opto/intrinsicnode.hpp"
45 #include "opto/idealKit.hpp"
46 #include "opto/mathexactnode.hpp"
47 #include "opto/movenode.hpp"
48 #include "opto/mulnode.hpp"
49 #include "opto/narrowptrnode.hpp"
50 #include "opto/opaquenode.hpp"
51 #include "opto/parse.hpp"
52 #include "opto/runtime.hpp"
53 #include "opto/rootnode.hpp"
54 #include "opto/subnode.hpp"
55 #include "opto/valuetypenode.hpp"
56 #include "prims/nativeLookup.hpp"
57 #include "prims/unsafe.hpp"
58 #include "runtime/objectMonitor.hpp"
59 #include "runtime/sharedRuntime.hpp"
60 #include "utilities/macros.hpp"
61
62
63 class LibraryIntrinsic : public InlineCallGenerator {
64 // Extend the set of intrinsics known to the runtime:
65 public:
66 private:
67 bool _is_virtual;
68 bool _does_virtual_dispatch;
69 int8_t _predicates_count; // Intrinsic is predicated by several conditions
70 int8_t _last_predicate; // Last generated predicate
71 vmIntrinsics::ID _intrinsic_id;
72
73 public:
74 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
75 : InlineCallGenerator(m),
76 _is_virtual(is_virtual),
77 _does_virtual_dispatch(does_virtual_dispatch),
78 _predicates_count((int8_t)predicates_count),
79 _last_predicate((int8_t)-1),
80 _intrinsic_id(id)
81 {
82 }
83 virtual bool is_intrinsic() const { return true; }
84 virtual bool is_virtual() const { return _is_virtual; }
85 virtual bool is_predicated() const { return _predicates_count > 0; }
86 virtual int predicates_count() const { return _predicates_count; }
87 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
88 virtual JVMState* generate(JVMState* jvms);
89 virtual Node* generate_predicate(JVMState* jvms, int predicate);
90 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
91 };
92
93
94 // Local helper class for LibraryIntrinsic:
95 class LibraryCallKit : public GraphKit {
96 private:
97 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
98 Node* _result; // the result node, if any
99 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
100
101 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
102
103 public:
104 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
105 : GraphKit(jvms),
106 _intrinsic(intrinsic),
107 _result(NULL)
108 {
109 // Check if this is a root compile. In that case we don't have a caller.
110 if (!jvms->has_method()) {
111 _reexecute_sp = sp();
112 } else {
113 // Find out how many arguments the interpreter needs when deoptimizing
114 // and save the stack pointer value so it can used by uncommon_trap.
115 // We find the argument count by looking at the declared signature.
116 bool ignored_will_link;
117 ciSignature* declared_signature = NULL;
118 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
119 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
120 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
121 }
122 }
123
124 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
125
126 ciMethod* caller() const { return jvms()->method(); }
127 int bci() const { return jvms()->bci(); }
128 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
129 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
130 ciMethod* callee() const { return _intrinsic->method(); }
131
132 bool try_to_inline(int predicate);
133 Node* try_to_predicate(int predicate);
134
135 void push_result() {
136 // Push the result onto the stack.
137 if (!stopped() && result() != NULL) {
138 BasicType bt = result()->bottom_type()->basic_type();
139 push_node(bt, result());
140 }
141 }
142
143 private:
144 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
145 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
146 }
147
148 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
149 void set_result(RegionNode* region, PhiNode* value);
150 Node* result() { return _result; }
151
152 virtual int reexecute_sp() { return _reexecute_sp; }
153
154 // Helper functions to inline natives
155 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
156 Node* generate_slow_guard(Node* test, RegionNode* region);
157 Node* generate_fair_guard(Node* test, RegionNode* region);
158 Node* generate_negative_guard(Node* index, RegionNode* region,
159 // resulting CastII of index:
160 Node* *pos_index = NULL);
161 Node* generate_limit_guard(Node* offset, Node* subseq_length,
162 Node* array_length,
163 RegionNode* region);
164 void generate_string_range_check(Node* array, Node* offset,
165 Node* length, bool char_count);
166 Node* generate_current_thread(Node* &tls_output);
167 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path,
169 int offset);
170 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
171 RegionNode* region, int null_path) {
172 int offset = java_lang_Class::klass_offset_in_bytes();
173 return load_klass_from_mirror_common(mirror, never_see_null,
174 region, null_path,
175 offset);
176 }
177 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
178 RegionNode* region, int null_path) {
179 int offset = java_lang_Class::array_klass_offset_in_bytes();
180 return load_klass_from_mirror_common(mirror, never_see_null,
181 region, null_path,
182 offset);
183 }
184 Node* generate_access_flags_guard(Node* kls,
185 int modifier_mask, int modifier_bits,
186 RegionNode* region);
187 Node* generate_interface_guard(Node* kls, RegionNode* region);
188 Node* generate_value_guard(Node* kls, RegionNode* region);
189
190 enum ArrayKind {
191 AnyArray,
192 NonArray,
193 ObjectArray,
194 NonObjectArray,
195 TypeArray,
196 ValueArray
197 };
198
199 Node* generate_array_guard(Node* kls, RegionNode* region) {
200 return generate_array_guard_common(kls, region, AnyArray);
201 }
202 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
203 return generate_array_guard_common(kls, region, NonArray);
204 }
205 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
206 return generate_array_guard_common(kls, region, ObjectArray);
207 }
208 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
209 return generate_array_guard_common(kls, region, NonObjectArray);
210 }
211 Node* generate_typeArray_guard(Node* kls, RegionNode* region) {
212 return generate_array_guard_common(kls, region, TypeArray);
213 }
214 Node* generate_valueArray_guard(Node* kls, RegionNode* region) {
215 return generate_array_guard_common(kls, region, ValueArray);
216 }
217 Node* generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind);
218 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
219 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
220 bool is_virtual = false, bool is_static = false);
221 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
222 return generate_method_call(method_id, false, true);
223 }
224 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
225 return generate_method_call(method_id, true, false);
226 }
227 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
228 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
229
230 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
231 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
232 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
233 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
234 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
235 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
236 bool inline_string_indexOfChar();
237 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
238 bool inline_string_toBytesU();
239 bool inline_string_getCharsU();
240 bool inline_string_copy(bool compress);
241 bool inline_string_char_access(bool is_store);
242 Node* round_double_node(Node* n);
243 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
244 bool inline_math_native(vmIntrinsics::ID id);
245 bool inline_math(vmIntrinsics::ID id);
246 template <typename OverflowOp>
247 bool inline_math_overflow(Node* arg1, Node* arg2);
248 void inline_math_mathExact(Node* math, Node* test);
249 bool inline_math_addExactI(bool is_increment);
250 bool inline_math_addExactL(bool is_increment);
251 bool inline_math_multiplyExactI();
252 bool inline_math_multiplyExactL();
253 bool inline_math_multiplyHigh();
254 bool inline_math_negateExactI();
255 bool inline_math_negateExactL();
256 bool inline_math_subtractExactI(bool is_decrement);
257 bool inline_math_subtractExactL(bool is_decrement);
258 bool inline_min_max(vmIntrinsics::ID id);
259 bool inline_notify(vmIntrinsics::ID id);
260 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
261 // This returns Type::AnyPtr, RawPtr, or OopPtr.
262 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
263 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
264
265 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
266 DecoratorSet mo_decorator_for_access_kind(AccessKind kind);
267 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
268 static bool klass_needs_init_guard(Node* kls);
269 bool inline_unsafe_allocate();
270 bool inline_unsafe_newArray(bool uninitialized);
271 bool inline_unsafe_copyMemory();
272 bool inline_native_currentThread();
273
274 bool inline_native_time_funcs(address method, const char* funcName);
275 #ifdef JFR_HAVE_INTRINSICS
276 bool inline_native_classID();
277 bool inline_native_getEventWriter();
278 #endif
279 bool inline_native_isInterrupted();
280 bool inline_native_Class_query(vmIntrinsics::ID id);
281 bool inline_native_subtype_check();
282 bool inline_native_getLength();
283 bool inline_array_copyOf(bool is_copyOfRange);
284 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
285 bool inline_preconditions_checkIndex();
286 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array);
287 bool inline_native_clone(bool is_virtual);
288 bool inline_native_Reflection_getCallerClass();
289 // Helper function for inlining native object hash method
290 bool inline_native_hashcode(bool is_virtual, bool is_static);
291 bool inline_native_getClass();
292
293 // Helper functions for inlining arraycopy
294 bool inline_arraycopy();
295 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
296 RegionNode* slow_region);
297 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
298 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
299 uint new_idx);
300
301 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
302 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
303 bool inline_unsafe_fence(vmIntrinsics::ID id);
304 bool inline_onspinwait();
305 bool inline_fp_conversions(vmIntrinsics::ID id);
306 bool inline_number_methods(vmIntrinsics::ID id);
307 bool inline_reference_get();
308 bool inline_Class_cast();
309 bool inline_aescrypt_Block(vmIntrinsics::ID id);
310 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
311 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
312 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
313 Node* inline_counterMode_AESCrypt_predicate();
314 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
315 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
316 bool inline_ghash_processBlocks();
317 bool inline_base64_encodeBlock();
318 bool inline_sha_implCompress(vmIntrinsics::ID id);
319 bool inline_digestBase_implCompressMB(int predicate);
320 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
321 bool long_state, address stubAddr, const char *stubName,
322 Node* src_start, Node* ofs, Node* limit);
323 Node* get_state_from_sha_object(Node *sha_object);
324 Node* get_state_from_sha5_object(Node *sha_object);
325 Node* inline_digestBase_implCompressMB_predicate(int predicate);
326 bool inline_encodeISOArray();
327 bool inline_updateCRC32();
328 bool inline_updateBytesCRC32();
329 bool inline_updateByteBufferCRC32();
330 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
331 bool inline_updateBytesCRC32C();
332 bool inline_updateDirectByteBufferCRC32C();
333 bool inline_updateBytesAdler32();
334 bool inline_updateByteBufferAdler32();
335 bool inline_multiplyToLen();
336 bool inline_hasNegatives();
337 bool inline_squareToLen();
338 bool inline_mulAdd();
339 bool inline_montgomeryMultiply();
340 bool inline_montgomerySquare();
341 bool inline_vectorizedMismatch();
342 bool inline_fma(vmIntrinsics::ID id);
343
344 bool inline_profileBoolean();
345 bool inline_isCompileConstant();
346 void clear_upper_avx() {
347 #ifdef X86
348 if (UseAVX >= 2) {
349 C->set_clear_upper_avx(true);
350 }
351 #endif
352 }
353 };
354
355 //---------------------------make_vm_intrinsic----------------------------
356 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
357 vmIntrinsics::ID id = m->intrinsic_id();
358 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
359
360 if (!m->is_loaded()) {
361 // Do not attempt to inline unloaded methods.
362 return NULL;
363 }
364
365 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
366 bool is_available = false;
367
368 {
369 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
370 // the compiler must transition to '_thread_in_vm' state because both
371 // methods access VM-internal data.
372 VM_ENTRY_MARK;
373 methodHandle mh(THREAD, m->get_Method());
374 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
375 !C->directive()->is_intrinsic_disabled(mh) &&
376 !vmIntrinsics::is_disabled_by_flags(mh);
377
378 }
379
380 if (is_available) {
381 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
382 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
383 return new LibraryIntrinsic(m, is_virtual,
384 vmIntrinsics::predicates_needed(id),
385 vmIntrinsics::does_virtual_dispatch(id),
386 (vmIntrinsics::ID) id);
387 } else {
388 return NULL;
389 }
390 }
391
392 //----------------------register_library_intrinsics-----------------------
393 // Initialize this file's data structures, for each Compile instance.
394 void Compile::register_library_intrinsics() {
395 // Nothing to do here.
396 }
397
398 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
399 LibraryCallKit kit(jvms, this);
400 Compile* C = kit.C;
401 int nodes = C->unique();
402 #ifndef PRODUCT
403 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
404 char buf[1000];
405 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
406 tty->print_cr("Intrinsic %s", str);
407 }
408 #endif
409 ciMethod* callee = kit.callee();
410 const int bci = kit.bci();
411
412 // Try to inline the intrinsic.
413 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
414 kit.try_to_inline(_last_predicate)) {
415 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
416 : "(intrinsic)";
417 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
418 if (C->print_intrinsics() || C->print_inlining()) {
419 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
420 }
421 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
422 if (C->log()) {
423 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
424 vmIntrinsics::name_at(intrinsic_id()),
425 (is_virtual() ? " virtual='1'" : ""),
426 C->unique() - nodes);
427 }
428 // Push the result from the inlined method onto the stack.
429 kit.push_result();
430 C->print_inlining_update(this);
431 return kit.transfer_exceptions_into_jvms();
432 }
433
434 // The intrinsic bailed out
435 if (jvms->has_method()) {
436 // Not a root compile.
437 const char* msg;
438 if (callee->intrinsic_candidate()) {
439 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
440 } else {
441 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
442 : "failed to inline (intrinsic), method not annotated";
443 }
444 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
445 if (C->print_intrinsics() || C->print_inlining()) {
446 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
447 }
448 } else {
449 // Root compile
450 ResourceMark rm;
451 stringStream msg_stream;
452 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
453 vmIntrinsics::name_at(intrinsic_id()),
454 is_virtual() ? " (virtual)" : "", bci);
455 const char *msg = msg_stream.as_string();
456 log_debug(jit, inlining)("%s", msg);
457 if (C->print_intrinsics() || C->print_inlining()) {
458 tty->print("%s", msg);
459 }
460 }
461 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
462 C->print_inlining_update(this);
463 return NULL;
464 }
465
466 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
467 LibraryCallKit kit(jvms, this);
468 Compile* C = kit.C;
469 int nodes = C->unique();
470 _last_predicate = predicate;
471 #ifndef PRODUCT
472 assert(is_predicated() && predicate < predicates_count(), "sanity");
473 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
474 char buf[1000];
475 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
476 tty->print_cr("Predicate for intrinsic %s", str);
477 }
478 #endif
479 ciMethod* callee = kit.callee();
480 const int bci = kit.bci();
481
482 Node* slow_ctl = kit.try_to_predicate(predicate);
483 if (!kit.failing()) {
484 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
485 : "(intrinsic, predicate)";
486 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
487 if (C->print_intrinsics() || C->print_inlining()) {
488 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
489 }
490 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
491 if (C->log()) {
492 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
493 vmIntrinsics::name_at(intrinsic_id()),
494 (is_virtual() ? " virtual='1'" : ""),
495 C->unique() - nodes);
496 }
497 return slow_ctl; // Could be NULL if the check folds.
498 }
499
500 // The intrinsic bailed out
501 if (jvms->has_method()) {
502 // Not a root compile.
503 const char* msg = "failed to generate predicate for intrinsic";
504 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
505 if (C->print_intrinsics() || C->print_inlining()) {
506 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
507 }
508 } else {
509 // Root compile
510 ResourceMark rm;
511 stringStream msg_stream;
512 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
513 vmIntrinsics::name_at(intrinsic_id()),
514 is_virtual() ? " (virtual)" : "", bci);
515 const char *msg = msg_stream.as_string();
516 log_debug(jit, inlining)("%s", msg);
517 if (C->print_intrinsics() || C->print_inlining()) {
518 C->print_inlining_stream()->print("%s", msg);
519 }
520 }
521 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
522 return NULL;
523 }
524
525 bool LibraryCallKit::try_to_inline(int predicate) {
526 // Handle symbolic names for otherwise undistinguished boolean switches:
527 const bool is_store = true;
528 const bool is_compress = true;
529 const bool is_static = true;
530 const bool is_volatile = true;
531
532 if (!jvms()->has_method()) {
533 // Root JVMState has a null method.
534 assert(map()->memory()->Opcode() == Op_Parm, "");
535 // Insert the memory aliasing node
536 set_all_memory(reset_memory());
537 }
538 assert(merged_memory(), "");
539
540
541 switch (intrinsic_id()) {
542 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
543 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
544 case vmIntrinsics::_getClass: return inline_native_getClass();
545
546 case vmIntrinsics::_dsin:
547 case vmIntrinsics::_dcos:
548 case vmIntrinsics::_dtan:
549 case vmIntrinsics::_dabs:
550 case vmIntrinsics::_datan2:
551 case vmIntrinsics::_dsqrt:
552 case vmIntrinsics::_dexp:
553 case vmIntrinsics::_dlog:
554 case vmIntrinsics::_dlog10:
555 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
556
557 case vmIntrinsics::_min:
558 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
559
560 case vmIntrinsics::_notify:
561 case vmIntrinsics::_notifyAll:
562 if (ObjectMonitor::Knob_InlineNotify) {
563 return inline_notify(intrinsic_id());
564 }
565 return false;
566
567 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
568 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
569 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
570 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
571 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
572 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
573 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
574 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
575 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
576 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
577 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
578 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
579 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
580
581 case vmIntrinsics::_arraycopy: return inline_arraycopy();
582
583 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
584 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
585 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
586 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
587
588 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
589 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
590 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
591 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
592 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
593 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
594 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
595
596 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
597 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
598
599 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
600 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
601 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
602 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
603
604 case vmIntrinsics::_compressStringC:
605 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
606 case vmIntrinsics::_inflateStringC:
607 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
608
609 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
610 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
611 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
612 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
613 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
614 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
615 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
616 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
617 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
618
619 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
620 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
621 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
622 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
623 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
624 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
625 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
626 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
627 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
628
629 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
630 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
631 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
632 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
633 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
634 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
635 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
636 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
637 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
638
639 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
640 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
641 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
642 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
643 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
644 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
645 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
646 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
647 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
648
649 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
650 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
651 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
652 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
653
654 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
655 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
656 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
657 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
658
659 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
660 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
661 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
662 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
663 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
664 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
665 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
666 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
667 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
668
669 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
670 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
671 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
672 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
673 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
674 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
675 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
676 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
677 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
678
679 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
680 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
681 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
682 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
683 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
684 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
685 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
686 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
687 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
688
689 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
690 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
691 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
692 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
693 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
694 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
695 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
696 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
697 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
698
699 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
700 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
701 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
702 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
703 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
704
705 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
706 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
707 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
708 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
709 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
710 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
711 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
712 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
713 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
714 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
715 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
716 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
717 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
718 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
719 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
720 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
721 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
722 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
723 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
724 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
725
726 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
727 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
728 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
729 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
730 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
731 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
732 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
733 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
734 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
735 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
736 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
737 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
738 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
739 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
740 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
741
742 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
743 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
744 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
745 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
746
747 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
748 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
749 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
750 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
751 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
752
753 case vmIntrinsics::_loadFence:
754 case vmIntrinsics::_storeFence:
755 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
756
757 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
758
759 case vmIntrinsics::_currentThread: return inline_native_currentThread();
760 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
761
762 #ifdef JFR_HAVE_INTRINSICS
763 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, JFR_TIME_FUNCTION), "counterTime");
764 case vmIntrinsics::_getClassId: return inline_native_classID();
765 case vmIntrinsics::_getEventWriter: return inline_native_getEventWriter();
766 #endif
767 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
768 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
769 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
770 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
771 case vmIntrinsics::_getLength: return inline_native_getLength();
772 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
773 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
774 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
775 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
776 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
777 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
778
779 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
780 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
781
782 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
783
784 case vmIntrinsics::_isInstance:
785 case vmIntrinsics::_getModifiers:
786 case vmIntrinsics::_isInterface:
787 case vmIntrinsics::_isArray:
788 case vmIntrinsics::_isPrimitive:
789 case vmIntrinsics::_getSuperclass:
790 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
791
792 case vmIntrinsics::_floatToRawIntBits:
793 case vmIntrinsics::_floatToIntBits:
794 case vmIntrinsics::_intBitsToFloat:
795 case vmIntrinsics::_doubleToRawLongBits:
796 case vmIntrinsics::_doubleToLongBits:
797 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
798
799 case vmIntrinsics::_numberOfLeadingZeros_i:
800 case vmIntrinsics::_numberOfLeadingZeros_l:
801 case vmIntrinsics::_numberOfTrailingZeros_i:
802 case vmIntrinsics::_numberOfTrailingZeros_l:
803 case vmIntrinsics::_bitCount_i:
804 case vmIntrinsics::_bitCount_l:
805 case vmIntrinsics::_reverseBytes_i:
806 case vmIntrinsics::_reverseBytes_l:
807 case vmIntrinsics::_reverseBytes_s:
808 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
809
810 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
811
812 case vmIntrinsics::_Reference_get: return inline_reference_get();
813
814 case vmIntrinsics::_Class_cast: return inline_Class_cast();
815
816 case vmIntrinsics::_aescrypt_encryptBlock:
817 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
818
819 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
820 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
821 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
822
823 case vmIntrinsics::_counterMode_AESCrypt:
824 return inline_counterMode_AESCrypt(intrinsic_id());
825
826 case vmIntrinsics::_sha_implCompress:
827 case vmIntrinsics::_sha2_implCompress:
828 case vmIntrinsics::_sha5_implCompress:
829 return inline_sha_implCompress(intrinsic_id());
830
831 case vmIntrinsics::_digestBase_implCompressMB:
832 return inline_digestBase_implCompressMB(predicate);
833
834 case vmIntrinsics::_multiplyToLen:
835 return inline_multiplyToLen();
836
837 case vmIntrinsics::_squareToLen:
838 return inline_squareToLen();
839
840 case vmIntrinsics::_mulAdd:
841 return inline_mulAdd();
842
843 case vmIntrinsics::_montgomeryMultiply:
844 return inline_montgomeryMultiply();
845 case vmIntrinsics::_montgomerySquare:
846 return inline_montgomerySquare();
847
848 case vmIntrinsics::_vectorizedMismatch:
849 return inline_vectorizedMismatch();
850
851 case vmIntrinsics::_ghash_processBlocks:
852 return inline_ghash_processBlocks();
853 case vmIntrinsics::_base64_encodeBlock:
854 return inline_base64_encodeBlock();
855
856 case vmIntrinsics::_encodeISOArray:
857 case vmIntrinsics::_encodeByteISOArray:
858 return inline_encodeISOArray();
859
860 case vmIntrinsics::_updateCRC32:
861 return inline_updateCRC32();
862 case vmIntrinsics::_updateBytesCRC32:
863 return inline_updateBytesCRC32();
864 case vmIntrinsics::_updateByteBufferCRC32:
865 return inline_updateByteBufferCRC32();
866
867 case vmIntrinsics::_updateBytesCRC32C:
868 return inline_updateBytesCRC32C();
869 case vmIntrinsics::_updateDirectByteBufferCRC32C:
870 return inline_updateDirectByteBufferCRC32C();
871
872 case vmIntrinsics::_updateBytesAdler32:
873 return inline_updateBytesAdler32();
874 case vmIntrinsics::_updateByteBufferAdler32:
875 return inline_updateByteBufferAdler32();
876
877 case vmIntrinsics::_profileBoolean:
878 return inline_profileBoolean();
879 case vmIntrinsics::_isCompileConstant:
880 return inline_isCompileConstant();
881
882 case vmIntrinsics::_hasNegatives:
883 return inline_hasNegatives();
884
885 case vmIntrinsics::_fmaD:
886 case vmIntrinsics::_fmaF:
887 return inline_fma(intrinsic_id());
888
889 default:
890 // If you get here, it may be that someone has added a new intrinsic
891 // to the list in vmSymbols.hpp without implementing it here.
892 #ifndef PRODUCT
893 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
894 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
895 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
896 }
897 #endif
898 return false;
899 }
900 }
901
902 Node* LibraryCallKit::try_to_predicate(int predicate) {
903 if (!jvms()->has_method()) {
904 // Root JVMState has a null method.
905 assert(map()->memory()->Opcode() == Op_Parm, "");
906 // Insert the memory aliasing node
907 set_all_memory(reset_memory());
908 }
909 assert(merged_memory(), "");
910
911 switch (intrinsic_id()) {
912 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
913 return inline_cipherBlockChaining_AESCrypt_predicate(false);
914 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
915 return inline_cipherBlockChaining_AESCrypt_predicate(true);
916 case vmIntrinsics::_counterMode_AESCrypt:
917 return inline_counterMode_AESCrypt_predicate();
918 case vmIntrinsics::_digestBase_implCompressMB:
919 return inline_digestBase_implCompressMB_predicate(predicate);
920
921 default:
922 // If you get here, it may be that someone has added a new intrinsic
923 // to the list in vmSymbols.hpp without implementing it here.
924 #ifndef PRODUCT
925 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
926 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
927 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
928 }
929 #endif
930 Node* slow_ctl = control();
931 set_control(top()); // No fast path instrinsic
932 return slow_ctl;
933 }
934 }
935
936 //------------------------------set_result-------------------------------
937 // Helper function for finishing intrinsics.
938 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
939 record_for_igvn(region);
940 set_control(_gvn.transform(region));
941 set_result( _gvn.transform(value));
942 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
943 }
944
945 //------------------------------generate_guard---------------------------
946 // Helper function for generating guarded fast-slow graph structures.
947 // The given 'test', if true, guards a slow path. If the test fails
948 // then a fast path can be taken. (We generally hope it fails.)
949 // In all cases, GraphKit::control() is updated to the fast path.
950 // The returned value represents the control for the slow path.
951 // The return value is never 'top'; it is either a valid control
952 // or NULL if it is obvious that the slow path can never be taken.
953 // Also, if region and the slow control are not NULL, the slow edge
954 // is appended to the region.
955 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
956 if (stopped()) {
957 // Already short circuited.
958 return NULL;
959 }
960
961 // Build an if node and its projections.
962 // If test is true we take the slow path, which we assume is uncommon.
963 if (_gvn.type(test) == TypeInt::ZERO) {
964 // The slow branch is never taken. No need to build this guard.
965 return NULL;
966 }
967
968 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
969
970 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
971 if (if_slow == top()) {
972 // The slow branch is never taken. No need to build this guard.
973 return NULL;
974 }
975
976 if (region != NULL)
977 region->add_req(if_slow);
978
979 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
980 set_control(if_fast);
981
982 return if_slow;
983 }
984
985 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
986 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
987 }
988 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
989 return generate_guard(test, region, PROB_FAIR);
990 }
991
992 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
993 Node* *pos_index) {
994 if (stopped())
995 return NULL; // already stopped
996 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
997 return NULL; // index is already adequately typed
998 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
999 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1000 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
1001 if (is_neg != NULL && pos_index != NULL) {
1002 // Emulate effect of Parse::adjust_map_after_if.
1003 Node* ccast = new CastIINode(index, TypeInt::POS);
1004 ccast->set_req(0, control());
1005 (*pos_index) = _gvn.transform(ccast);
1006 }
1007 return is_neg;
1008 }
1009
1010 // Make sure that 'position' is a valid limit index, in [0..length].
1011 // There are two equivalent plans for checking this:
1012 // A. (offset + copyLength) unsigned<= arrayLength
1013 // B. offset <= (arrayLength - copyLength)
1014 // We require that all of the values above, except for the sum and
1015 // difference, are already known to be non-negative.
1016 // Plan A is robust in the face of overflow, if offset and copyLength
1017 // are both hugely positive.
1018 //
1019 // Plan B is less direct and intuitive, but it does not overflow at
1020 // all, since the difference of two non-negatives is always
1021 // representable. Whenever Java methods must perform the equivalent
1022 // check they generally use Plan B instead of Plan A.
1023 // For the moment we use Plan A.
1024 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1025 Node* subseq_length,
1026 Node* array_length,
1027 RegionNode* region) {
1028 if (stopped())
1029 return NULL; // already stopped
1030 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1031 if (zero_offset && subseq_length->eqv_uncast(array_length))
1032 return NULL; // common case of whole-array copy
1033 Node* last = subseq_length;
1034 if (!zero_offset) // last += offset
1035 last = _gvn.transform(new AddINode(last, offset));
1036 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1037 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1038 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1039 return is_over;
1040 }
1041
1042 // Emit range checks for the given String.value byte array
1043 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1044 if (stopped()) {
1045 return; // already stopped
1046 }
1047 RegionNode* bailout = new RegionNode(1);
1048 record_for_igvn(bailout);
1049 if (char_count) {
1050 // Convert char count to byte count
1051 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1052 }
1053
1054 // Offset and count must not be negative
1055 generate_negative_guard(offset, bailout);
1056 generate_negative_guard(count, bailout);
1057 // Offset + count must not exceed length of array
1058 generate_limit_guard(offset, count, load_array_length(array), bailout);
1059
1060 if (bailout->req() > 1) {
1061 PreserveJVMState pjvms(this);
1062 set_control(_gvn.transform(bailout));
1063 uncommon_trap(Deoptimization::Reason_intrinsic,
1064 Deoptimization::Action_maybe_recompile);
1065 }
1066 }
1067
1068 //--------------------------generate_current_thread--------------------
1069 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1070 ciKlass* thread_klass = env()->Thread_klass();
1071 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1072 Node* thread = _gvn.transform(new ThreadLocalNode());
1073 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1074 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1075 tls_output = thread;
1076 return threadObj;
1077 }
1078
1079
1080 //------------------------------make_string_method_node------------------------
1081 // Helper method for String intrinsic functions. This version is called with
1082 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1083 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1084 // containing the lengths of str1 and str2.
1085 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1086 Node* result = NULL;
1087 switch (opcode) {
1088 case Op_StrIndexOf:
1089 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1090 str1_start, cnt1, str2_start, cnt2, ae);
1091 break;
1092 case Op_StrComp:
1093 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1094 str1_start, cnt1, str2_start, cnt2, ae);
1095 break;
1096 case Op_StrEquals:
1097 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1098 // Use the constant length if there is one because optimized match rule may exist.
1099 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1100 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1101 break;
1102 default:
1103 ShouldNotReachHere();
1104 return NULL;
1105 }
1106
1107 // All these intrinsics have checks.
1108 C->set_has_split_ifs(true); // Has chance for split-if optimization
1109 clear_upper_avx();
1110
1111 return _gvn.transform(result);
1112 }
1113
1114 //------------------------------inline_string_compareTo------------------------
1115 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1116 Node* arg1 = argument(0);
1117 Node* arg2 = argument(1);
1118
1119 arg1 = must_be_not_null(arg1, true);
1120 arg2 = must_be_not_null(arg2, true);
1121
1122 // Get start addr and length of first argument
1123 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1124 Node* arg1_cnt = load_array_length(arg1);
1125
1126 // Get start addr and length of second argument
1127 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1128 Node* arg2_cnt = load_array_length(arg2);
1129
1130 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1131 set_result(result);
1132 return true;
1133 }
1134
1135 //------------------------------inline_string_equals------------------------
1136 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1137 Node* arg1 = argument(0);
1138 Node* arg2 = argument(1);
1139
1140 // paths (plus control) merge
1141 RegionNode* region = new RegionNode(3);
1142 Node* phi = new PhiNode(region, TypeInt::BOOL);
1143
1144 if (!stopped()) {
1145
1146 arg1 = must_be_not_null(arg1, true);
1147 arg2 = must_be_not_null(arg2, true);
1148
1149 // Get start addr and length of first argument
1150 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1151 Node* arg1_cnt = load_array_length(arg1);
1152
1153 // Get start addr and length of second argument
1154 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1155 Node* arg2_cnt = load_array_length(arg2);
1156
1157 // Check for arg1_cnt != arg2_cnt
1158 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1159 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1160 Node* if_ne = generate_slow_guard(bol, NULL);
1161 if (if_ne != NULL) {
1162 phi->init_req(2, intcon(0));
1163 region->init_req(2, if_ne);
1164 }
1165
1166 // Check for count == 0 is done by assembler code for StrEquals.
1167
1168 if (!stopped()) {
1169 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1170 phi->init_req(1, equals);
1171 region->init_req(1, control());
1172 }
1173 }
1174
1175 // post merge
1176 set_control(_gvn.transform(region));
1177 record_for_igvn(region);
1178
1179 set_result(_gvn.transform(phi));
1180 return true;
1181 }
1182
1183 //------------------------------inline_array_equals----------------------------
1184 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1185 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1186 Node* arg1 = argument(0);
1187 Node* arg2 = argument(1);
1188
1189 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1190 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1191 clear_upper_avx();
1192
1193 return true;
1194 }
1195
1196 //------------------------------inline_hasNegatives------------------------------
1197 bool LibraryCallKit::inline_hasNegatives() {
1198 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1199 return false;
1200 }
1201
1202 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1203 // no receiver since it is static method
1204 Node* ba = argument(0);
1205 Node* offset = argument(1);
1206 Node* len = argument(2);
1207
1208 ba = must_be_not_null(ba, true);
1209
1210 // Range checks
1211 generate_string_range_check(ba, offset, len, false);
1212 if (stopped()) {
1213 return true;
1214 }
1215 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1216 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1217 set_result(_gvn.transform(result));
1218 return true;
1219 }
1220
1221 bool LibraryCallKit::inline_preconditions_checkIndex() {
1222 Node* index = argument(0);
1223 Node* length = argument(1);
1224 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1225 return false;
1226 }
1227
1228 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1229 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1230
1231 {
1232 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1233 uncommon_trap(Deoptimization::Reason_intrinsic,
1234 Deoptimization::Action_make_not_entrant);
1235 }
1236
1237 if (stopped()) {
1238 return false;
1239 }
1240
1241 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1242 BoolTest::mask btest = BoolTest::lt;
1243 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1244 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1245 _gvn.set_type(rc, rc->Value(&_gvn));
1246 if (!rc_bool->is_Con()) {
1247 record_for_igvn(rc);
1248 }
1249 set_control(_gvn.transform(new IfTrueNode(rc)));
1250 {
1251 PreserveJVMState pjvms(this);
1252 set_control(_gvn.transform(new IfFalseNode(rc)));
1253 uncommon_trap(Deoptimization::Reason_range_check,
1254 Deoptimization::Action_make_not_entrant);
1255 }
1256
1257 if (stopped()) {
1258 return false;
1259 }
1260
1261 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1262 result->set_req(0, control());
1263 result = _gvn.transform(result);
1264 set_result(result);
1265 replace_in_map(index, result);
1266 clear_upper_avx();
1267 return true;
1268 }
1269
1270 //------------------------------inline_string_indexOf------------------------
1271 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1272 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1273 return false;
1274 }
1275 Node* src = argument(0);
1276 Node* tgt = argument(1);
1277
1278 // Make the merge point
1279 RegionNode* result_rgn = new RegionNode(4);
1280 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1281
1282 src = must_be_not_null(src, true);
1283 tgt = must_be_not_null(tgt, true);
1284
1285 // Get start addr and length of source string
1286 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1287 Node* src_count = load_array_length(src);
1288
1289 // Get start addr and length of substring
1290 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1291 Node* tgt_count = load_array_length(tgt);
1292
1293 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1294 // Divide src size by 2 if String is UTF16 encoded
1295 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1296 }
1297 if (ae == StrIntrinsicNode::UU) {
1298 // Divide substring size by 2 if String is UTF16 encoded
1299 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1300 }
1301
1302 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1303 if (result != NULL) {
1304 result_phi->init_req(3, result);
1305 result_rgn->init_req(3, control());
1306 }
1307 set_control(_gvn.transform(result_rgn));
1308 record_for_igvn(result_rgn);
1309 set_result(_gvn.transform(result_phi));
1310
1311 return true;
1312 }
1313
1314 //-----------------------------inline_string_indexOf-----------------------
1315 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1316 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1317 return false;
1318 }
1319 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1320 return false;
1321 }
1322 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1323 Node* src = argument(0); // byte[]
1324 Node* src_count = argument(1); // char count
1325 Node* tgt = argument(2); // byte[]
1326 Node* tgt_count = argument(3); // char count
1327 Node* from_index = argument(4); // char index
1328
1329 src = must_be_not_null(src, true);
1330 tgt = must_be_not_null(tgt, true);
1331
1332 // Multiply byte array index by 2 if String is UTF16 encoded
1333 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1334 src_count = _gvn.transform(new SubINode(src_count, from_index));
1335 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1336 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1337
1338 // Range checks
1339 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1340 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1341 if (stopped()) {
1342 return true;
1343 }
1344
1345 RegionNode* region = new RegionNode(5);
1346 Node* phi = new PhiNode(region, TypeInt::INT);
1347
1348 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1349 if (result != NULL) {
1350 // The result is index relative to from_index if substring was found, -1 otherwise.
1351 // Generate code which will fold into cmove.
1352 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1353 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1354
1355 Node* if_lt = generate_slow_guard(bol, NULL);
1356 if (if_lt != NULL) {
1357 // result == -1
1358 phi->init_req(3, result);
1359 region->init_req(3, if_lt);
1360 }
1361 if (!stopped()) {
1362 result = _gvn.transform(new AddINode(result, from_index));
1363 phi->init_req(4, result);
1364 region->init_req(4, control());
1365 }
1366 }
1367
1368 set_control(_gvn.transform(region));
1369 record_for_igvn(region);
1370 set_result(_gvn.transform(phi));
1371 clear_upper_avx();
1372
1373 return true;
1374 }
1375
1376 // Create StrIndexOfNode with fast path checks
1377 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1378 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1379 // Check for substr count > string count
1380 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1381 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1382 Node* if_gt = generate_slow_guard(bol, NULL);
1383 if (if_gt != NULL) {
1384 phi->init_req(1, intcon(-1));
1385 region->init_req(1, if_gt);
1386 }
1387 if (!stopped()) {
1388 // Check for substr count == 0
1389 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1390 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1391 Node* if_zero = generate_slow_guard(bol, NULL);
1392 if (if_zero != NULL) {
1393 phi->init_req(2, intcon(0));
1394 region->init_req(2, if_zero);
1395 }
1396 }
1397 if (!stopped()) {
1398 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1399 }
1400 return NULL;
1401 }
1402
1403 //-----------------------------inline_string_indexOfChar-----------------------
1404 bool LibraryCallKit::inline_string_indexOfChar() {
1405 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1406 return false;
1407 }
1408 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1409 return false;
1410 }
1411 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1412 Node* src = argument(0); // byte[]
1413 Node* tgt = argument(1); // tgt is int ch
1414 Node* from_index = argument(2);
1415 Node* max = argument(3);
1416
1417 src = must_be_not_null(src, true);
1418
1419 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1420 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1421 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1422
1423 // Range checks
1424 generate_string_range_check(src, src_offset, src_count, true);
1425 if (stopped()) {
1426 return true;
1427 }
1428
1429 RegionNode* region = new RegionNode(3);
1430 Node* phi = new PhiNode(region, TypeInt::INT);
1431
1432 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1433 C->set_has_split_ifs(true); // Has chance for split-if optimization
1434 _gvn.transform(result);
1435
1436 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1437 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1438
1439 Node* if_lt = generate_slow_guard(bol, NULL);
1440 if (if_lt != NULL) {
1441 // result == -1
1442 phi->init_req(2, result);
1443 region->init_req(2, if_lt);
1444 }
1445 if (!stopped()) {
1446 result = _gvn.transform(new AddINode(result, from_index));
1447 phi->init_req(1, result);
1448 region->init_req(1, control());
1449 }
1450 set_control(_gvn.transform(region));
1451 record_for_igvn(region);
1452 set_result(_gvn.transform(phi));
1453
1454 return true;
1455 }
1456 //---------------------------inline_string_copy---------------------
1457 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1458 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1459 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1460 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1461 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1462 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1463 bool LibraryCallKit::inline_string_copy(bool compress) {
1464 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1465 return false;
1466 }
1467 int nargs = 5; // 2 oops, 3 ints
1468 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1469
1470 Node* src = argument(0);
1471 Node* src_offset = argument(1);
1472 Node* dst = argument(2);
1473 Node* dst_offset = argument(3);
1474 Node* length = argument(4);
1475
1476 // Check for allocation before we add nodes that would confuse
1477 // tightly_coupled_allocation()
1478 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1479
1480 // Figure out the size and type of the elements we will be copying.
1481 const Type* src_type = src->Value(&_gvn);
1482 const Type* dst_type = dst->Value(&_gvn);
1483 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1484 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1485 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1486 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1487 "Unsupported array types for inline_string_copy");
1488
1489 src = must_be_not_null(src, true);
1490 dst = must_be_not_null(dst, true);
1491
1492 // Convert char[] offsets to byte[] offsets
1493 bool convert_src = (compress && src_elem == T_BYTE);
1494 bool convert_dst = (!compress && dst_elem == T_BYTE);
1495 if (convert_src) {
1496 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1497 } else if (convert_dst) {
1498 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1499 }
1500
1501 // Range checks
1502 generate_string_range_check(src, src_offset, length, convert_src);
1503 generate_string_range_check(dst, dst_offset, length, convert_dst);
1504 if (stopped()) {
1505 return true;
1506 }
1507
1508 Node* src_start = array_element_address(src, src_offset, src_elem);
1509 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1510 // 'src_start' points to src array + scaled offset
1511 // 'dst_start' points to dst array + scaled offset
1512 Node* count = NULL;
1513 if (compress) {
1514 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1515 } else {
1516 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1517 }
1518
1519 if (alloc != NULL) {
1520 if (alloc->maybe_set_complete(&_gvn)) {
1521 // "You break it, you buy it."
1522 InitializeNode* init = alloc->initialization();
1523 assert(init->is_complete(), "we just did this");
1524 init->set_complete_with_arraycopy();
1525 assert(dst->is_CheckCastPP(), "sanity");
1526 assert(dst->in(0)->in(0) == init, "dest pinned");
1527 }
1528 // Do not let stores that initialize this object be reordered with
1529 // a subsequent store that would make this object accessible by
1530 // other threads.
1531 // Record what AllocateNode this StoreStore protects so that
1532 // escape analysis can go from the MemBarStoreStoreNode to the
1533 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1534 // based on the escape status of the AllocateNode.
1535 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1536 }
1537 if (compress) {
1538 set_result(_gvn.transform(count));
1539 }
1540 clear_upper_avx();
1541
1542 return true;
1543 }
1544
1545 #ifdef _LP64
1546 #define XTOP ,top() /*additional argument*/
1547 #else //_LP64
1548 #define XTOP /*no additional argument*/
1549 #endif //_LP64
1550
1551 //------------------------inline_string_toBytesU--------------------------
1552 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1553 bool LibraryCallKit::inline_string_toBytesU() {
1554 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1555 return false;
1556 }
1557 // Get the arguments.
1558 Node* value = argument(0);
1559 Node* offset = argument(1);
1560 Node* length = argument(2);
1561
1562 Node* newcopy = NULL;
1563
1564 // Set the original stack and the reexecute bit for the interpreter to reexecute
1565 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1566 { PreserveReexecuteState preexecs(this);
1567 jvms()->set_should_reexecute(true);
1568
1569 // Check if a null path was taken unconditionally.
1570 value = null_check(value);
1571
1572 RegionNode* bailout = new RegionNode(1);
1573 record_for_igvn(bailout);
1574
1575 // Range checks
1576 generate_negative_guard(offset, bailout);
1577 generate_negative_guard(length, bailout);
1578 generate_limit_guard(offset, length, load_array_length(value), bailout);
1579 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1580 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1581
1582 if (bailout->req() > 1) {
1583 PreserveJVMState pjvms(this);
1584 set_control(_gvn.transform(bailout));
1585 uncommon_trap(Deoptimization::Reason_intrinsic,
1586 Deoptimization::Action_maybe_recompile);
1587 }
1588 if (stopped()) {
1589 return true;
1590 }
1591
1592 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1593 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1594 newcopy = new_array(klass_node, size, 0); // no arguments to push
1595 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1596
1597 // Calculate starting addresses.
1598 Node* src_start = array_element_address(value, offset, T_CHAR);
1599 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1600
1601 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1602 const TypeInt* toffset = gvn().type(offset)->is_int();
1603 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1604
1605 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1606 const char* copyfunc_name = "arraycopy";
1607 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1608 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1609 OptoRuntime::fast_arraycopy_Type(),
1610 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1611 src_start, dst_start, ConvI2X(length) XTOP);
1612 // Do not let reads from the cloned object float above the arraycopy.
1613 if (alloc != NULL) {
1614 if (alloc->maybe_set_complete(&_gvn)) {
1615 // "You break it, you buy it."
1616 InitializeNode* init = alloc->initialization();
1617 assert(init->is_complete(), "we just did this");
1618 init->set_complete_with_arraycopy();
1619 assert(newcopy->is_CheckCastPP(), "sanity");
1620 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1621 }
1622 // Do not let stores that initialize this object be reordered with
1623 // a subsequent store that would make this object accessible by
1624 // other threads.
1625 // Record what AllocateNode this StoreStore protects so that
1626 // escape analysis can go from the MemBarStoreStoreNode to the
1627 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1628 // based on the escape status of the AllocateNode.
1629 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1630 } else {
1631 insert_mem_bar(Op_MemBarCPUOrder);
1632 }
1633 } // original reexecute is set back here
1634
1635 C->set_has_split_ifs(true); // Has chance for split-if optimization
1636 if (!stopped()) {
1637 set_result(newcopy);
1638 }
1639 clear_upper_avx();
1640
1641 return true;
1642 }
1643
1644 //------------------------inline_string_getCharsU--------------------------
1645 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1646 bool LibraryCallKit::inline_string_getCharsU() {
1647 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1648 return false;
1649 }
1650
1651 // Get the arguments.
1652 Node* src = argument(0);
1653 Node* src_begin = argument(1);
1654 Node* src_end = argument(2); // exclusive offset (i < src_end)
1655 Node* dst = argument(3);
1656 Node* dst_begin = argument(4);
1657
1658 // Check for allocation before we add nodes that would confuse
1659 // tightly_coupled_allocation()
1660 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1661
1662 // Check if a null path was taken unconditionally.
1663 src = null_check(src);
1664 dst = null_check(dst);
1665 if (stopped()) {
1666 return true;
1667 }
1668
1669 // Get length and convert char[] offset to byte[] offset
1670 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1671 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1672
1673 // Range checks
1674 generate_string_range_check(src, src_begin, length, true);
1675 generate_string_range_check(dst, dst_begin, length, false);
1676 if (stopped()) {
1677 return true;
1678 }
1679
1680 if (!stopped()) {
1681 // Calculate starting addresses.
1682 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1683 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1684
1685 // Check if array addresses are aligned to HeapWordSize
1686 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1687 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1688 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1689 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1690
1691 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1692 const char* copyfunc_name = "arraycopy";
1693 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1694 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1695 OptoRuntime::fast_arraycopy_Type(),
1696 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1697 src_start, dst_start, ConvI2X(length) XTOP);
1698 // Do not let reads from the cloned object float above the arraycopy.
1699 if (alloc != NULL) {
1700 if (alloc->maybe_set_complete(&_gvn)) {
1701 // "You break it, you buy it."
1702 InitializeNode* init = alloc->initialization();
1703 assert(init->is_complete(), "we just did this");
1704 init->set_complete_with_arraycopy();
1705 assert(dst->is_CheckCastPP(), "sanity");
1706 assert(dst->in(0)->in(0) == init, "dest pinned");
1707 }
1708 // Do not let stores that initialize this object be reordered with
1709 // a subsequent store that would make this object accessible by
1710 // other threads.
1711 // Record what AllocateNode this StoreStore protects so that
1712 // escape analysis can go from the MemBarStoreStoreNode to the
1713 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1714 // based on the escape status of the AllocateNode.
1715 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
1716 } else {
1717 insert_mem_bar(Op_MemBarCPUOrder);
1718 }
1719 }
1720
1721 C->set_has_split_ifs(true); // Has chance for split-if optimization
1722 return true;
1723 }
1724
1725 //----------------------inline_string_char_access----------------------------
1726 // Store/Load char to/from byte[] array.
1727 // static void StringUTF16.putChar(byte[] val, int index, int c)
1728 // static char StringUTF16.getChar(byte[] val, int index)
1729 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1730 Node* value = argument(0);
1731 Node* index = argument(1);
1732 Node* ch = is_store ? argument(2) : NULL;
1733
1734 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1735 // correctly requires matched array shapes.
1736 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1737 "sanity: byte[] and char[] bases agree");
1738 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1739 "sanity: byte[] and char[] scales agree");
1740
1741 // Bail when getChar over constants is requested: constant folding would
1742 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1743 // Java method would constant fold nicely instead.
1744 if (!is_store && value->is_Con() && index->is_Con()) {
1745 return false;
1746 }
1747
1748 value = must_be_not_null(value, true);
1749
1750 Node* adr = array_element_address(value, index, T_CHAR);
1751 if (adr->is_top()) {
1752 return false;
1753 }
1754 if (is_store) {
1755 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1756 false, false, true /* mismatched */);
1757 } else {
1758 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1759 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1760 set_result(ch);
1761 }
1762 return true;
1763 }
1764
1765 //--------------------------round_double_node--------------------------------
1766 // Round a double node if necessary.
1767 Node* LibraryCallKit::round_double_node(Node* n) {
1768 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1769 n = _gvn.transform(new RoundDoubleNode(0, n));
1770 return n;
1771 }
1772
1773 //------------------------------inline_math-----------------------------------
1774 // public static double Math.abs(double)
1775 // public static double Math.sqrt(double)
1776 // public static double Math.log(double)
1777 // public static double Math.log10(double)
1778 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1779 Node* arg = round_double_node(argument(0));
1780 Node* n = NULL;
1781 switch (id) {
1782 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1783 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1784 default: fatal_unexpected_iid(id); break;
1785 }
1786 set_result(_gvn.transform(n));
1787 return true;
1788 }
1789
1790 //------------------------------runtime_math-----------------------------
1791 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1792 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1793 "must be (DD)D or (D)D type");
1794
1795 // Inputs
1796 Node* a = round_double_node(argument(0));
1797 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1798
1799 const TypePtr* no_memory_effects = NULL;
1800 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1801 no_memory_effects,
1802 a, top(), b, b ? top() : NULL);
1803 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1804 #ifdef ASSERT
1805 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1806 assert(value_top == top(), "second value must be top");
1807 #endif
1808
1809 set_result(value);
1810 return true;
1811 }
1812
1813 //------------------------------inline_math_native-----------------------------
1814 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1815 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1816 switch (id) {
1817 // These intrinsics are not properly supported on all hardware
1818 case vmIntrinsics::_dsin:
1819 return StubRoutines::dsin() != NULL ?
1820 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1821 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1822 case vmIntrinsics::_dcos:
1823 return StubRoutines::dcos() != NULL ?
1824 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1825 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1826 case vmIntrinsics::_dtan:
1827 return StubRoutines::dtan() != NULL ?
1828 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1829 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1830 case vmIntrinsics::_dlog:
1831 return StubRoutines::dlog() != NULL ?
1832 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1833 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1834 case vmIntrinsics::_dlog10:
1835 return StubRoutines::dlog10() != NULL ?
1836 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1837 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1838
1839 // These intrinsics are supported on all hardware
1840 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1841 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1842
1843 case vmIntrinsics::_dexp:
1844 return StubRoutines::dexp() != NULL ?
1845 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1846 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1847 case vmIntrinsics::_dpow: {
1848 Node* exp = round_double_node(argument(2));
1849 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1850 if (d != NULL && d->getd() == 2.0) {
1851 // Special case: pow(x, 2.0) => x * x
1852 Node* base = round_double_node(argument(0));
1853 set_result(_gvn.transform(new MulDNode(base, base)));
1854 return true;
1855 }
1856 return StubRoutines::dpow() != NULL ?
1857 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1858 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1859 }
1860 #undef FN_PTR
1861
1862 // These intrinsics are not yet correctly implemented
1863 case vmIntrinsics::_datan2:
1864 return false;
1865
1866 default:
1867 fatal_unexpected_iid(id);
1868 return false;
1869 }
1870 }
1871
1872 static bool is_simple_name(Node* n) {
1873 return (n->req() == 1 // constant
1874 || (n->is_Type() && n->as_Type()->type()->singleton())
1875 || n->is_Proj() // parameter or return value
1876 || n->is_Phi() // local of some sort
1877 );
1878 }
1879
1880 //----------------------------inline_notify-----------------------------------*
1881 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1882 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1883 address func;
1884 if (id == vmIntrinsics::_notify) {
1885 func = OptoRuntime::monitor_notify_Java();
1886 } else {
1887 func = OptoRuntime::monitor_notifyAll_Java();
1888 }
1889 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1890 make_slow_call_ex(call, env()->Throwable_klass(), false);
1891 return true;
1892 }
1893
1894
1895 //----------------------------inline_min_max-----------------------------------
1896 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1897 set_result(generate_min_max(id, argument(0), argument(1)));
1898 return true;
1899 }
1900
1901 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1902 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1903 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1904 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1905 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1906
1907 {
1908 PreserveJVMState pjvms(this);
1909 PreserveReexecuteState preexecs(this);
1910 jvms()->set_should_reexecute(true);
1911
1912 set_control(slow_path);
1913 set_i_o(i_o());
1914
1915 uncommon_trap(Deoptimization::Reason_intrinsic,
1916 Deoptimization::Action_none);
1917 }
1918
1919 set_control(fast_path);
1920 set_result(math);
1921 }
1922
1923 template <typename OverflowOp>
1924 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1925 typedef typename OverflowOp::MathOp MathOp;
1926
1927 MathOp* mathOp = new MathOp(arg1, arg2);
1928 Node* operation = _gvn.transform( mathOp );
1929 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1930 inline_math_mathExact(operation, ofcheck);
1931 return true;
1932 }
1933
1934 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1935 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1936 }
1937
1938 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1939 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1940 }
1941
1942 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1943 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1944 }
1945
1946 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1947 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1948 }
1949
1950 bool LibraryCallKit::inline_math_negateExactI() {
1951 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1952 }
1953
1954 bool LibraryCallKit::inline_math_negateExactL() {
1955 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1956 }
1957
1958 bool LibraryCallKit::inline_math_multiplyExactI() {
1959 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1960 }
1961
1962 bool LibraryCallKit::inline_math_multiplyExactL() {
1963 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1964 }
1965
1966 bool LibraryCallKit::inline_math_multiplyHigh() {
1967 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1968 return true;
1969 }
1970
1971 Node*
1972 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1973 // These are the candidate return value:
1974 Node* xvalue = x0;
1975 Node* yvalue = y0;
1976
1977 if (xvalue == yvalue) {
1978 return xvalue;
1979 }
1980
1981 bool want_max = (id == vmIntrinsics::_max);
1982
1983 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1984 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1985 if (txvalue == NULL || tyvalue == NULL) return top();
1986 // This is not really necessary, but it is consistent with a
1987 // hypothetical MaxINode::Value method:
1988 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1989
1990 // %%% This folding logic should (ideally) be in a different place.
1991 // Some should be inside IfNode, and there to be a more reliable
1992 // transformation of ?: style patterns into cmoves. We also want
1993 // more powerful optimizations around cmove and min/max.
1994
1995 // Try to find a dominating comparison of these guys.
1996 // It can simplify the index computation for Arrays.copyOf
1997 // and similar uses of System.arraycopy.
1998 // First, compute the normalized version of CmpI(x, y).
1999 int cmp_op = Op_CmpI;
2000 Node* xkey = xvalue;
2001 Node* ykey = yvalue;
2002 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2003 if (ideal_cmpxy->is_Cmp()) {
2004 // E.g., if we have CmpI(length - offset, count),
2005 // it might idealize to CmpI(length, count + offset)
2006 cmp_op = ideal_cmpxy->Opcode();
2007 xkey = ideal_cmpxy->in(1);
2008 ykey = ideal_cmpxy->in(2);
2009 }
2010
2011 // Start by locating any relevant comparisons.
2012 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2013 Node* cmpxy = NULL;
2014 Node* cmpyx = NULL;
2015 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2016 Node* cmp = start_from->fast_out(k);
2017 if (cmp->outcnt() > 0 && // must have prior uses
2018 cmp->in(0) == NULL && // must be context-independent
2019 cmp->Opcode() == cmp_op) { // right kind of compare
2020 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2021 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2022 }
2023 }
2024
2025 const int NCMPS = 2;
2026 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2027 int cmpn;
2028 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2029 if (cmps[cmpn] != NULL) break; // find a result
2030 }
2031 if (cmpn < NCMPS) {
2032 // Look for a dominating test that tells us the min and max.
2033 int depth = 0; // Limit search depth for speed
2034 Node* dom = control();
2035 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2036 if (++depth >= 100) break;
2037 Node* ifproj = dom;
2038 if (!ifproj->is_Proj()) continue;
2039 Node* iff = ifproj->in(0);
2040 if (!iff->is_If()) continue;
2041 Node* bol = iff->in(1);
2042 if (!bol->is_Bool()) continue;
2043 Node* cmp = bol->in(1);
2044 if (cmp == NULL) continue;
2045 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2046 if (cmps[cmpn] == cmp) break;
2047 if (cmpn == NCMPS) continue;
2048 BoolTest::mask btest = bol->as_Bool()->_test._test;
2049 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2050 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2051 // At this point, we know that 'x btest y' is true.
2052 switch (btest) {
2053 case BoolTest::eq:
2054 // They are proven equal, so we can collapse the min/max.
2055 // Either value is the answer. Choose the simpler.
2056 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2057 return yvalue;
2058 return xvalue;
2059 case BoolTest::lt: // x < y
2060 case BoolTest::le: // x <= y
2061 return (want_max ? yvalue : xvalue);
2062 case BoolTest::gt: // x > y
2063 case BoolTest::ge: // x >= y
2064 return (want_max ? xvalue : yvalue);
2065 default:
2066 break;
2067 }
2068 }
2069 }
2070
2071 // We failed to find a dominating test.
2072 // Let's pick a test that might GVN with prior tests.
2073 Node* best_bol = NULL;
2074 BoolTest::mask best_btest = BoolTest::illegal;
2075 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2076 Node* cmp = cmps[cmpn];
2077 if (cmp == NULL) continue;
2078 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2079 Node* bol = cmp->fast_out(j);
2080 if (!bol->is_Bool()) continue;
2081 BoolTest::mask btest = bol->as_Bool()->_test._test;
2082 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2083 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2084 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2085 best_bol = bol->as_Bool();
2086 best_btest = btest;
2087 }
2088 }
2089 }
2090
2091 Node* answer_if_true = NULL;
2092 Node* answer_if_false = NULL;
2093 switch (best_btest) {
2094 default:
2095 if (cmpxy == NULL)
2096 cmpxy = ideal_cmpxy;
2097 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2098 // and fall through:
2099 case BoolTest::lt: // x < y
2100 case BoolTest::le: // x <= y
2101 answer_if_true = (want_max ? yvalue : xvalue);
2102 answer_if_false = (want_max ? xvalue : yvalue);
2103 break;
2104 case BoolTest::gt: // x > y
2105 case BoolTest::ge: // x >= y
2106 answer_if_true = (want_max ? xvalue : yvalue);
2107 answer_if_false = (want_max ? yvalue : xvalue);
2108 break;
2109 }
2110
2111 jint hi, lo;
2112 if (want_max) {
2113 // We can sharpen the minimum.
2114 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2115 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2116 } else {
2117 // We can sharpen the maximum.
2118 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2119 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2120 }
2121
2122 // Use a flow-free graph structure, to avoid creating excess control edges
2123 // which could hinder other optimizations.
2124 // Since Math.min/max is often used with arraycopy, we want
2125 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2126 Node* cmov = CMoveNode::make(NULL, best_bol,
2127 answer_if_false, answer_if_true,
2128 TypeInt::make(lo, hi, widen));
2129
2130 return _gvn.transform(cmov);
2131
2132 /*
2133 // This is not as desirable as it may seem, since Min and Max
2134 // nodes do not have a full set of optimizations.
2135 // And they would interfere, anyway, with 'if' optimizations
2136 // and with CMoveI canonical forms.
2137 switch (id) {
2138 case vmIntrinsics::_min:
2139 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2140 case vmIntrinsics::_max:
2141 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2142 default:
2143 ShouldNotReachHere();
2144 }
2145 */
2146 }
2147
2148 inline int
2149 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2150 const TypePtr* base_type = TypePtr::NULL_PTR;
2151 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2152 if (base_type == NULL) {
2153 // Unknown type.
2154 return Type::AnyPtr;
2155 } else if (base_type == TypePtr::NULL_PTR) {
2156 // Since this is a NULL+long form, we have to switch to a rawptr.
2157 base = _gvn.transform(new CastX2PNode(offset));
2158 offset = MakeConX(0);
2159 return Type::RawPtr;
2160 } else if (base_type->base() == Type::RawPtr) {
2161 return Type::RawPtr;
2162 } else if (base_type->isa_oopptr()) {
2163 // Base is never null => always a heap address.
2164 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2165 return Type::OopPtr;
2166 }
2167 // Offset is small => always a heap address.
2168 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2169 if (offset_type != NULL &&
2170 base_type->offset() == 0 && // (should always be?)
2171 offset_type->_lo >= 0 &&
2172 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2173 return Type::OopPtr;
2174 } else if (type == T_OBJECT) {
2175 // off heap access to an oop doesn't make any sense. Has to be on
2176 // heap.
2177 return Type::OopPtr;
2178 }
2179 // Otherwise, it might either be oop+off or NULL+addr.
2180 return Type::AnyPtr;
2181 } else {
2182 // No information:
2183 return Type::AnyPtr;
2184 }
2185 }
2186
2187 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2188 Node* uncasted_base = base;
2189 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2190 if (kind == Type::RawPtr) {
2191 return basic_plus_adr(top(), uncasted_base, offset);
2192 } else if (kind == Type::AnyPtr) {
2193 assert(base == uncasted_base, "unexpected base change");
2194 if (can_cast) {
2195 if (!_gvn.type(base)->speculative_maybe_null() &&
2196 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2197 // According to profiling, this access is always on
2198 // heap. Casting the base to not null and thus avoiding membars
2199 // around the access should allow better optimizations
2200 Node* null_ctl = top();
2201 base = null_check_oop(base, &null_ctl, true, true, true);
2202 assert(null_ctl->is_top(), "no null control here");
2203 return basic_plus_adr(base, offset);
2204 } else if (_gvn.type(base)->speculative_always_null() &&
2205 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2206 // According to profiling, this access is always off
2207 // heap.
2208 base = null_assert(base);
2209 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2210 offset = MakeConX(0);
2211 return basic_plus_adr(top(), raw_base, offset);
2212 }
2213 }
2214 // We don't know if it's an on heap or off heap access. Fall back
2215 // to raw memory access.
2216 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2217 return basic_plus_adr(top(), raw, offset);
2218 } else {
2219 assert(base == uncasted_base, "unexpected base change");
2220 // We know it's an on heap access so base can't be null
2221 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2222 base = must_be_not_null(base, true);
2223 }
2224 return basic_plus_adr(base, offset);
2225 }
2226 }
2227
2228 //--------------------------inline_number_methods-----------------------------
2229 // inline int Integer.numberOfLeadingZeros(int)
2230 // inline int Long.numberOfLeadingZeros(long)
2231 //
2232 // inline int Integer.numberOfTrailingZeros(int)
2233 // inline int Long.numberOfTrailingZeros(long)
2234 //
2235 // inline int Integer.bitCount(int)
2236 // inline int Long.bitCount(long)
2237 //
2238 // inline char Character.reverseBytes(char)
2239 // inline short Short.reverseBytes(short)
2240 // inline int Integer.reverseBytes(int)
2241 // inline long Long.reverseBytes(long)
2242 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2243 Node* arg = argument(0);
2244 Node* n = NULL;
2245 switch (id) {
2246 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2247 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2248 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2249 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2250 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2251 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2252 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2253 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2254 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2255 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2256 default: fatal_unexpected_iid(id); break;
2257 }
2258 set_result(_gvn.transform(n));
2259 return true;
2260 }
2261
2262 //----------------------------inline_unsafe_access----------------------------
2263
2264 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2265 // Attempt to infer a sharper value type from the offset and base type.
2266 ciKlass* sharpened_klass = NULL;
2267
2268 // See if it is an instance field, with an object type.
2269 if (alias_type->field() != NULL) {
2270 if (alias_type->field()->type()->is_klass()) {
2271 sharpened_klass = alias_type->field()->type()->as_klass();
2272 }
2273 }
2274
2275 // See if it is a narrow oop array.
2276 if (adr_type->isa_aryptr()) {
2277 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2278 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2279 if (elem_type != NULL) {
2280 sharpened_klass = elem_type->klass();
2281 }
2282 }
2283 }
2284
2285 // The sharpened class might be unloaded if there is no class loader
2286 // contraint in place.
2287 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2288 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2289
2290 #ifndef PRODUCT
2291 if (C->print_intrinsics() || C->print_inlining()) {
2292 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2293 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2294 }
2295 #endif
2296 // Sharpen the value type.
2297 return tjp;
2298 }
2299 return NULL;
2300 }
2301
2302 DecoratorSet LibraryCallKit::mo_decorator_for_access_kind(AccessKind kind) {
2303 switch (kind) {
2304 case Relaxed:
2305 return MO_UNORDERED;
2306 case Opaque:
2307 return MO_RELAXED;
2308 case Acquire:
2309 return MO_ACQUIRE;
2310 case Release:
2311 return MO_RELEASE;
2312 case Volatile:
2313 return MO_SEQ_CST;
2314 default:
2315 ShouldNotReachHere();
2316 return 0;
2317 }
2318 }
2319
2320 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2321 if (callee()->is_static()) return false; // caller must have the capability!
2322 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2323 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2324 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2325 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2326
2327 if (type == T_OBJECT || type == T_ARRAY) {
2328 decorators |= ON_UNKNOWN_OOP_REF;
2329 }
2330
2331 if (unaligned) {
2332 decorators |= C2_UNALIGNED;
2333 }
2334
2335 #ifndef PRODUCT
2336 {
2337 ResourceMark rm;
2338 // Check the signatures.
2339 ciSignature* sig = callee()->signature();
2340 #ifdef ASSERT
2341 if (!is_store) {
2342 // Object getObject(Object base, int/long offset), etc.
2343 BasicType rtype = sig->return_type()->basic_type();
2344 assert(rtype == type, "getter must return the expected value");
2345 assert(sig->count() == 2, "oop getter has 2 arguments");
2346 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2347 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2348 } else {
2349 // void putObject(Object base, int/long offset, Object x), etc.
2350 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2351 assert(sig->count() == 3, "oop putter has 3 arguments");
2352 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2353 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2354 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2355 assert(vtype == type, "putter must accept the expected value");
2356 }
2357 #endif // ASSERT
2358 }
2359 #endif //PRODUCT
2360
2361 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2362
2363 Node* receiver = argument(0); // type: oop
2364
2365 // Build address expression.
2366 Node* adr;
2367 Node* heap_base_oop = top();
2368 Node* offset = top();
2369 Node* val;
2370
2371 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2372 Node* base = argument(1); // type: oop
2373 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2374 offset = argument(2); // type: long
2375 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2376 // to be plain byte offsets, which are also the same as those accepted
2377 // by oopDesc::field_addr.
2378 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2379 "fieldOffset must be byte-scaled");
2380
2381 if (base->is_ValueType()) {
2382 if (is_store) {
2383 return false;
2384 }
2385
2386 ValueTypeNode* vt = base->as_ValueType();
2387 if (offset->is_Con()) {
2388 long off = find_long_con(offset, 0);
2389 ciValueKlass* vk = _gvn.type(vt)->is_valuetype()->value_klass();
2390 if ((long)(int)off != off || !vk->contains_field_offset(off)) {
2391 return false;
2392 }
2393
2394 receiver = null_check(receiver);
2395 if (stopped()) {
2396 return true;
2397 }
2398
2399 set_result(vt->field_value_by_offset((int)off, true));
2400 return true;
2401 } else {
2402 receiver = null_check(receiver);
2403 if (stopped()) {
2404 return true;
2405 }
2406 vt = vt->allocate(this)->as_ValueType();
2407 base = vt->get_oop();
2408 }
2409 }
2410
2411 // 32-bit machines ignore the high half!
2412 offset = ConvL2X(offset);
2413 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2414
2415 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2416 heap_base_oop = base;
2417 } else if (type == T_OBJECT) {
2418 return false; // off-heap oop accesses are not supported
2419 }
2420
2421 // Can base be NULL? Otherwise, always on-heap access.
2422 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2423
2424 if (!can_access_non_heap) {
2425 decorators |= IN_HEAP;
2426 }
2427
2428 val = is_store ? argument(4) : NULL;
2429
2430 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2431
2432 // Try to categorize the address.
2433 Compile::AliasType* alias_type = C->alias_type(adr_type);
2434 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2435
2436 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2437 alias_type->adr_type() == TypeAryPtr::RANGE) {
2438 return false; // not supported
2439 }
2440
2441 bool mismatched = false;
2442 BasicType bt = alias_type->basic_type();
2443 if (bt != T_ILLEGAL) {
2444 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2445 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2446 // Alias type doesn't differentiate between byte[] and boolean[]).
2447 // Use address type to get the element type.
2448 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2449 }
2450 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2451 // accessing an array field with getObject is not a mismatch
2452 bt = T_OBJECT;
2453 }
2454 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2455 // Don't intrinsify mismatched object accesses
2456 return false;
2457 }
2458 mismatched = (bt != type);
2459 } else if (alias_type->adr_type()->isa_oopptr()) {
2460 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2461 }
2462
2463 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2464
2465 if (mismatched) {
2466 decorators |= C2_MISMATCHED;
2467 }
2468
2469 // First guess at the value type.
2470 const Type *value_type = Type::get_const_basic_type(type);
2471
2472 // Figure out the memory ordering.
2473 decorators |= mo_decorator_for_access_kind(kind);
2474
2475 if (!is_store && type == T_OBJECT) {
2476 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2477 if (tjp != NULL) {
2478 value_type = tjp;
2479 }
2480 }
2481
2482 receiver = null_check(receiver);
2483 if (stopped()) {
2484 return true;
2485 }
2486 // Heap pointers get a null-check from the interpreter,
2487 // as a courtesy. However, this is not guaranteed by Unsafe,
2488 // and it is not possible to fully distinguish unintended nulls
2489 // from intended ones in this API.
2490
2491 if (!is_store) {
2492 Node* p = NULL;
2493 // Try to constant fold a load from a constant field
2494 ciField* field = alias_type->field();
2495 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2496 // final or stable field
2497 p = make_constant_from_field(field, heap_base_oop);
2498 }
2499
2500 if (p == NULL) { // Could not constant fold the load
2501 p = access_load_at(heap_base_oop, adr, adr_type, value_type, type, decorators);
2502 // Normalize the value returned by getBoolean in the following cases
2503 if (type == T_BOOLEAN &&
2504 (mismatched ||
2505 heap_base_oop == top() || // - heap_base_oop is NULL or
2506 (can_access_non_heap && field == NULL)) // - heap_base_oop is potentially NULL
2507 // and the unsafe access is made to large offset
2508 // (i.e., larger than the maximum offset necessary for any
2509 // field access)
2510 ) {
2511 IdealKit ideal = IdealKit(this);
2512 #define __ ideal.
2513 IdealVariable normalized_result(ideal);
2514 __ declarations_done();
2515 __ set(normalized_result, p);
2516 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2517 __ set(normalized_result, ideal.ConI(1));
2518 ideal.end_if();
2519 final_sync(ideal);
2520 p = __ value(normalized_result);
2521 #undef __
2522 }
2523 }
2524 if (type == T_ADDRESS) {
2525 p = gvn().transform(new CastP2XNode(NULL, p));
2526 p = ConvX2UL(p);
2527 }
2528 // The load node has the control of the preceding MemBarCPUOrder. All
2529 // following nodes will have the control of the MemBarCPUOrder inserted at
2530 // the end of this method. So, pushing the load onto the stack at a later
2531 // point is fine.
2532 set_result(p);
2533 } else {
2534 if (bt == T_ADDRESS) {
2535 // Repackage the long as a pointer.
2536 val = ConvL2X(val);
2537 val = gvn().transform(new CastX2PNode(val));
2538 }
2539 access_store_at(control(), heap_base_oop, adr, adr_type, val, value_type, type, decorators);
2540 }
2541
2542 return true;
2543 }
2544
2545 //----------------------------inline_unsafe_load_store----------------------------
2546 // This method serves a couple of different customers (depending on LoadStoreKind):
2547 //
2548 // LS_cmp_swap:
2549 //
2550 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2551 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2552 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2553 //
2554 // LS_cmp_swap_weak:
2555 //
2556 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2557 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2558 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2559 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2560 //
2561 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2562 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2563 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2564 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2565 //
2566 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2567 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2568 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2569 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2570 //
2571 // LS_cmp_exchange:
2572 //
2573 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2574 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2575 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2576 //
2577 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2578 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2579 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2580 //
2581 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2582 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2583 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2584 //
2585 // LS_get_add:
2586 //
2587 // int getAndAddInt( Object o, long offset, int delta)
2588 // long getAndAddLong(Object o, long offset, long delta)
2589 //
2590 // LS_get_set:
2591 //
2592 // int getAndSet(Object o, long offset, int newValue)
2593 // long getAndSet(Object o, long offset, long newValue)
2594 // Object getAndSet(Object o, long offset, Object newValue)
2595 //
2596 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2597 // This basic scheme here is the same as inline_unsafe_access, but
2598 // differs in enough details that combining them would make the code
2599 // overly confusing. (This is a true fact! I originally combined
2600 // them, but even I was confused by it!) As much code/comments as
2601 // possible are retained from inline_unsafe_access though to make
2602 // the correspondences clearer. - dl
2603
2604 if (callee()->is_static()) return false; // caller must have the capability!
2605
2606 DecoratorSet decorators = C2_UNSAFE_ACCESS;
2607 decorators |= mo_decorator_for_access_kind(access_kind);
2608
2609 #ifndef PRODUCT
2610 BasicType rtype;
2611 {
2612 ResourceMark rm;
2613 // Check the signatures.
2614 ciSignature* sig = callee()->signature();
2615 rtype = sig->return_type()->basic_type();
2616 switch(kind) {
2617 case LS_get_add:
2618 case LS_get_set: {
2619 // Check the signatures.
2620 #ifdef ASSERT
2621 assert(rtype == type, "get and set must return the expected type");
2622 assert(sig->count() == 3, "get and set has 3 arguments");
2623 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2624 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2625 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2626 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2627 #endif // ASSERT
2628 break;
2629 }
2630 case LS_cmp_swap:
2631 case LS_cmp_swap_weak: {
2632 // Check the signatures.
2633 #ifdef ASSERT
2634 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2635 assert(sig->count() == 4, "CAS has 4 arguments");
2636 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2637 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2638 #endif // ASSERT
2639 break;
2640 }
2641 case LS_cmp_exchange: {
2642 // Check the signatures.
2643 #ifdef ASSERT
2644 assert(rtype == type, "CAS must return the expected type");
2645 assert(sig->count() == 4, "CAS has 4 arguments");
2646 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2647 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2648 #endif // ASSERT
2649 break;
2650 }
2651 default:
2652 ShouldNotReachHere();
2653 }
2654 }
2655 #endif //PRODUCT
2656
2657 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2658
2659 // Get arguments:
2660 Node* receiver = NULL;
2661 Node* base = NULL;
2662 Node* offset = NULL;
2663 Node* oldval = NULL;
2664 Node* newval = NULL;
2665 switch(kind) {
2666 case LS_cmp_swap:
2667 case LS_cmp_swap_weak:
2668 case LS_cmp_exchange: {
2669 const bool two_slot_type = type2size[type] == 2;
2670 receiver = argument(0); // type: oop
2671 base = argument(1); // type: oop
2672 offset = argument(2); // type: long
2673 oldval = argument(4); // type: oop, int, or long
2674 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2675 break;
2676 }
2677 case LS_get_add:
2678 case LS_get_set: {
2679 receiver = argument(0); // type: oop
2680 base = argument(1); // type: oop
2681 offset = argument(2); // type: long
2682 oldval = NULL;
2683 newval = argument(4); // type: oop, int, or long
2684 break;
2685 }
2686 default:
2687 ShouldNotReachHere();
2688 }
2689
2690 // Build field offset expression.
2691 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2692 // to be plain byte offsets, which are also the same as those accepted
2693 // by oopDesc::field_addr.
2694 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2695 // 32-bit machines ignore the high half of long offsets
2696 offset = ConvL2X(offset);
2697 Node* adr = make_unsafe_address(base, offset, type, false);
2698 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2699
2700 Compile::AliasType* alias_type = C->alias_type(adr_type);
2701 BasicType bt = alias_type->basic_type();
2702 if (bt != T_ILLEGAL &&
2703 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2704 // Don't intrinsify mismatched object accesses.
2705 return false;
2706 }
2707
2708 // For CAS, unlike inline_unsafe_access, there seems no point in
2709 // trying to refine types. Just use the coarse types here.
2710 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2711 const Type *value_type = Type::get_const_basic_type(type);
2712
2713 switch (kind) {
2714 case LS_get_set:
2715 case LS_cmp_exchange: {
2716 if (type == T_OBJECT) {
2717 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2718 if (tjp != NULL) {
2719 value_type = tjp;
2720 }
2721 }
2722 break;
2723 }
2724 case LS_cmp_swap:
2725 case LS_cmp_swap_weak:
2726 case LS_get_add:
2727 break;
2728 default:
2729 ShouldNotReachHere();
2730 }
2731
2732 // Null check receiver.
2733 receiver = null_check(receiver);
2734 if (stopped()) {
2735 return true;
2736 }
2737
2738 int alias_idx = C->get_alias_index(adr_type);
2739
2740 if (type == T_OBJECT || type == T_ARRAY) {
2741 decorators |= IN_HEAP | ON_UNKNOWN_OOP_REF;
2742
2743 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2744 // could be delayed during Parse (for example, in adjust_map_after_if()).
2745 // Execute transformation here to avoid barrier generation in such case.
2746 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2747 newval = _gvn.makecon(TypePtr::NULL_PTR);
2748
2749 if (oldval != NULL && _gvn.type(oldval) == TypePtr::NULL_PTR) {
2750 // Refine the value to a null constant, when it is known to be null
2751 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2752 }
2753 }
2754
2755 Node* result = NULL;
2756 switch (kind) {
2757 case LS_cmp_exchange: {
2758 result = access_atomic_cmpxchg_val_at(control(), base, adr, adr_type, alias_idx,
2759 oldval, newval, value_type, type, decorators);
2760 break;
2761 }
2762 case LS_cmp_swap_weak:
2763 decorators |= C2_WEAK_CMPXCHG;
2764 case LS_cmp_swap: {
2765 result = access_atomic_cmpxchg_bool_at(control(), base, adr, adr_type, alias_idx,
2766 oldval, newval, value_type, type, decorators);
2767 break;
2768 }
2769 case LS_get_set: {
2770 result = access_atomic_xchg_at(control(), base, adr, adr_type, alias_idx,
2771 newval, value_type, type, decorators);
2772 break;
2773 }
2774 case LS_get_add: {
2775 result = access_atomic_add_at(control(), base, adr, adr_type, alias_idx,
2776 newval, value_type, type, decorators);
2777 break;
2778 }
2779 default:
2780 ShouldNotReachHere();
2781 }
2782
2783 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2784 set_result(result);
2785 return true;
2786 }
2787
2788 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2789 // Regardless of form, don't allow previous ld/st to move down,
2790 // then issue acquire, release, or volatile mem_bar.
2791 insert_mem_bar(Op_MemBarCPUOrder);
2792 switch(id) {
2793 case vmIntrinsics::_loadFence:
2794 insert_mem_bar(Op_LoadFence);
2795 return true;
2796 case vmIntrinsics::_storeFence:
2797 insert_mem_bar(Op_StoreFence);
2798 return true;
2799 case vmIntrinsics::_fullFence:
2800 insert_mem_bar(Op_MemBarVolatile);
2801 return true;
2802 default:
2803 fatal_unexpected_iid(id);
2804 return false;
2805 }
2806 }
2807
2808 bool LibraryCallKit::inline_onspinwait() {
2809 insert_mem_bar(Op_OnSpinWait);
2810 return true;
2811 }
2812
2813 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2814 if (!kls->is_Con()) {
2815 return true;
2816 }
2817 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2818 if (klsptr == NULL) {
2819 return true;
2820 }
2821 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2822 // don't need a guard for a klass that is already initialized
2823 return !ik->is_initialized();
2824 }
2825
2826 //----------------------------inline_unsafe_allocate---------------------------
2827 // public native Object Unsafe.allocateInstance(Class<?> cls);
2828 bool LibraryCallKit::inline_unsafe_allocate() {
2829 if (callee()->is_static()) return false; // caller must have the capability!
2830
2831 null_check_receiver(); // null-check, then ignore
2832 Node* cls = null_check(argument(1));
2833 if (stopped()) return true;
2834
2835 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2836 kls = null_check(kls);
2837 if (stopped()) return true; // argument was like int.class
2838
2839 Node* test = NULL;
2840 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2841 // Note: The argument might still be an illegal value like
2842 // Serializable.class or Object[].class. The runtime will handle it.
2843 // But we must make an explicit check for initialization.
2844 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2845 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2846 // can generate code to load it as unsigned byte.
2847 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2848 Node* bits = intcon(InstanceKlass::fully_initialized);
2849 test = _gvn.transform(new SubINode(inst, bits));
2850 // The 'test' is non-zero if we need to take a slow path.
2851 }
2852
2853 Node* obj = new_instance(kls, test);
2854 set_result(obj);
2855 return true;
2856 }
2857
2858 //------------------------inline_native_time_funcs--------------
2859 // inline code for System.currentTimeMillis() and System.nanoTime()
2860 // these have the same type and signature
2861 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2862 const TypeFunc* tf = OptoRuntime::void_long_Type();
2863 const TypePtr* no_memory_effects = NULL;
2864 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2865 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2866 #ifdef ASSERT
2867 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2868 assert(value_top == top(), "second value must be top");
2869 #endif
2870 set_result(value);
2871 return true;
2872 }
2873
2874 #ifdef JFR_HAVE_INTRINSICS
2875
2876 /*
2877 * oop -> myklass
2878 * myklass->trace_id |= USED
2879 * return myklass->trace_id & ~0x3
2880 */
2881 bool LibraryCallKit::inline_native_classID() {
2882 Node* cls = null_check(argument(0), T_OBJECT);
2883 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2884 kls = null_check(kls, T_OBJECT);
2885
2886 ByteSize offset = KLASS_TRACE_ID_OFFSET;
2887 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2888 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2889
2890 Node* clsused = longcon(0x01l); // set the class bit
2891 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2892 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2893 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2894
2895 #ifdef TRACE_ID_META_BITS
2896 Node* mbits = longcon(~TRACE_ID_META_BITS);
2897 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
2898 #endif
2899 #ifdef TRACE_ID_SHIFT
2900 Node* cbits = intcon(TRACE_ID_SHIFT);
2901 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
2902 #endif
2903
2904 set_result(tvalue);
2905 return true;
2906
2907 }
2908
2909 bool LibraryCallKit::inline_native_getEventWriter() {
2910 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
2911
2912 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
2913 in_bytes(THREAD_LOCAL_WRITER_OFFSET_JFR));
2914
2915 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
2916
2917 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
2918 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
2919
2920 IfNode* iff_jobj_null =
2921 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
2922
2923 enum { _normal_path = 1,
2924 _null_path = 2,
2925 PATH_LIMIT };
2926
2927 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2928 PhiNode* result_val = new PhiNode(result_rgn, TypeInstPtr::BOTTOM);
2929
2930 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
2931 result_rgn->init_req(_null_path, jobj_is_null);
2932 result_val->init_req(_null_path, null());
2933
2934 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
2935 set_control(jobj_is_not_null);
2936 Node* res = access_load(jobj, TypeInstPtr::NOTNULL, T_OBJECT,
2937 IN_NATIVE | C2_CONTROL_DEPENDENT_LOAD);
2938 result_rgn->init_req(_normal_path, control());
2939 result_val->init_req(_normal_path, res);
2940
2941 set_result(result_rgn, result_val);
2942
2943 return true;
2944 }
2945
2946 #endif // JFR_HAVE_INTRINSICS
2947
2948 //------------------------inline_native_currentThread------------------
2949 bool LibraryCallKit::inline_native_currentThread() {
2950 Node* junk = NULL;
2951 set_result(generate_current_thread(junk));
2952 return true;
2953 }
2954
2955 //------------------------inline_native_isInterrupted------------------
2956 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2957 bool LibraryCallKit::inline_native_isInterrupted() {
2958 // Add a fast path to t.isInterrupted(clear_int):
2959 // (t == Thread.current() &&
2960 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2961 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2962 // So, in the common case that the interrupt bit is false,
2963 // we avoid making a call into the VM. Even if the interrupt bit
2964 // is true, if the clear_int argument is false, we avoid the VM call.
2965 // However, if the receiver is not currentThread, we must call the VM,
2966 // because there must be some locking done around the operation.
2967
2968 // We only go to the fast case code if we pass two guards.
2969 // Paths which do not pass are accumulated in the slow_region.
2970
2971 enum {
2972 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2973 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2974 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2975 PATH_LIMIT
2976 };
2977
2978 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
2979 // out of the function.
2980 insert_mem_bar(Op_MemBarCPUOrder);
2981
2982 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
2983 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
2984
2985 RegionNode* slow_region = new RegionNode(1);
2986 record_for_igvn(slow_region);
2987
2988 // (a) Receiving thread must be the current thread.
2989 Node* rec_thr = argument(0);
2990 Node* tls_ptr = NULL;
2991 Node* cur_thr = generate_current_thread(tls_ptr);
2992 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
2993 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
2994
2995 generate_slow_guard(bol_thr, slow_region);
2996
2997 // (b) Interrupt bit on TLS must be false.
2998 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2999 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3000 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3001
3002 // Set the control input on the field _interrupted read to prevent it floating up.
3003 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3004 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3005 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3006
3007 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3008
3009 // First fast path: if (!TLS._interrupted) return false;
3010 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3011 result_rgn->init_req(no_int_result_path, false_bit);
3012 result_val->init_req(no_int_result_path, intcon(0));
3013
3014 // drop through to next case
3015 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3016
3017 #ifndef _WINDOWS
3018 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3019 Node* clr_arg = argument(1);
3020 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3021 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3022 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3023
3024 // Second fast path: ... else if (!clear_int) return true;
3025 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3026 result_rgn->init_req(no_clear_result_path, false_arg);
3027 result_val->init_req(no_clear_result_path, intcon(1));
3028
3029 // drop through to next case
3030 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3031 #else
3032 // To return true on Windows you must read the _interrupted field
3033 // and check the event state i.e. take the slow path.
3034 #endif // _WINDOWS
3035
3036 // (d) Otherwise, go to the slow path.
3037 slow_region->add_req(control());
3038 set_control( _gvn.transform(slow_region));
3039
3040 if (stopped()) {
3041 // There is no slow path.
3042 result_rgn->init_req(slow_result_path, top());
3043 result_val->init_req(slow_result_path, top());
3044 } else {
3045 // non-virtual because it is a private non-static
3046 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3047
3048 Node* slow_val = set_results_for_java_call(slow_call);
3049 // this->control() comes from set_results_for_java_call
3050
3051 Node* fast_io = slow_call->in(TypeFunc::I_O);
3052 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3053
3054 // These two phis are pre-filled with copies of of the fast IO and Memory
3055 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3056 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3057
3058 result_rgn->init_req(slow_result_path, control());
3059 result_io ->init_req(slow_result_path, i_o());
3060 result_mem->init_req(slow_result_path, reset_memory());
3061 result_val->init_req(slow_result_path, slow_val);
3062
3063 set_all_memory(_gvn.transform(result_mem));
3064 set_i_o( _gvn.transform(result_io));
3065 }
3066
3067 C->set_has_split_ifs(true); // Has chance for split-if optimization
3068 set_result(result_rgn, result_val);
3069 return true;
3070 }
3071
3072 //-----------------------load_klass_from_mirror_common-------------------------
3073 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3074 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3075 // and branch to the given path on the region.
3076 // If never_see_null, take an uncommon trap on null, so we can optimistically
3077 // compile for the non-null case.
3078 // If the region is NULL, force never_see_null = true.
3079 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3080 bool never_see_null,
3081 RegionNode* region,
3082 int null_path,
3083 int offset) {
3084 if (region == NULL) never_see_null = true;
3085 Node* p = basic_plus_adr(mirror, offset);
3086 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3087 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3088 Node* null_ctl = top();
3089 kls = null_check_oop(kls, &null_ctl, never_see_null);
3090 if (region != NULL) {
3091 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3092 region->init_req(null_path, null_ctl);
3093 } else {
3094 assert(null_ctl == top(), "no loose ends");
3095 }
3096 return kls;
3097 }
3098
3099 //--------------------(inline_native_Class_query helpers)---------------------
3100 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3101 // Fall through if (mods & mask) == bits, take the guard otherwise.
3102 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3103 // Branch around if the given klass has the given modifier bit set.
3104 // Like generate_guard, adds a new path onto the region.
3105 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3106 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3107 Node* mask = intcon(modifier_mask);
3108 Node* bits = intcon(modifier_bits);
3109 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3110 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3111 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3112 return generate_fair_guard(bol, region);
3113 }
3114 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3115 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3116 }
3117
3118 Node* LibraryCallKit::generate_value_guard(Node* kls, RegionNode* region) {
3119 return generate_access_flags_guard(kls, JVM_ACC_VALUE, 0, region);
3120 }
3121
3122 //-------------------------inline_native_Class_query-------------------
3123 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3124 const Type* return_type = TypeInt::BOOL;
3125 Node* prim_return_value = top(); // what happens if it's a primitive class?
3126 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3127 bool expect_prim = false; // most of these guys expect to work on refs
3128
3129 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3130
3131 Node* mirror = argument(0);
3132 Node* obj = top();
3133
3134 switch (id) {
3135 case vmIntrinsics::_isInstance:
3136 // nothing is an instance of a primitive type
3137 prim_return_value = intcon(0);
3138 obj = argument(1);
3139 break;
3140 case vmIntrinsics::_getModifiers:
3141 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3142 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3143 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3144 break;
3145 case vmIntrinsics::_isInterface:
3146 prim_return_value = intcon(0);
3147 break;
3148 case vmIntrinsics::_isArray:
3149 prim_return_value = intcon(0);
3150 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3151 break;
3152 case vmIntrinsics::_isPrimitive:
3153 prim_return_value = intcon(1);
3154 expect_prim = true; // obviously
3155 break;
3156 case vmIntrinsics::_getSuperclass:
3157 prim_return_value = null();
3158 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3159 break;
3160 case vmIntrinsics::_getClassAccessFlags:
3161 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3162 return_type = TypeInt::INT; // not bool! 6297094
3163 break;
3164 default:
3165 fatal_unexpected_iid(id);
3166 break;
3167 }
3168
3169 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3170 if (mirror_con == NULL) return false; // cannot happen?
3171
3172 #ifndef PRODUCT
3173 if (C->print_intrinsics() || C->print_inlining()) {
3174 ciType* k = mirror_con->java_mirror_type();
3175 if (k) {
3176 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3177 k->print_name();
3178 tty->cr();
3179 }
3180 }
3181 #endif
3182
3183 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3184 RegionNode* region = new RegionNode(PATH_LIMIT);
3185 record_for_igvn(region);
3186 PhiNode* phi = new PhiNode(region, return_type);
3187
3188 // The mirror will never be null of Reflection.getClassAccessFlags, however
3189 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3190 // if it is. See bug 4774291.
3191
3192 // For Reflection.getClassAccessFlags(), the null check occurs in
3193 // the wrong place; see inline_unsafe_access(), above, for a similar
3194 // situation.
3195 mirror = null_check(mirror);
3196 // If mirror or obj is dead, only null-path is taken.
3197 if (stopped()) return true;
3198
3199 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3200
3201 // Now load the mirror's klass metaobject, and null-check it.
3202 // Side-effects region with the control path if the klass is null.
3203 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3204 // If kls is null, we have a primitive mirror.
3205 phi->init_req(_prim_path, prim_return_value);
3206 if (stopped()) { set_result(region, phi); return true; }
3207 bool safe_for_replace = (region->in(_prim_path) == top());
3208
3209 Node* p; // handy temp
3210 Node* null_ctl;
3211
3212 // Now that we have the non-null klass, we can perform the real query.
3213 // For constant classes, the query will constant-fold in LoadNode::Value.
3214 Node* query_value = top();
3215 switch (id) {
3216 case vmIntrinsics::_isInstance:
3217 // nothing is an instance of a primitive type
3218 query_value = gen_instanceof(obj, kls, safe_for_replace);
3219 break;
3220
3221 case vmIntrinsics::_getModifiers:
3222 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3223 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3224 break;
3225
3226 case vmIntrinsics::_isInterface:
3227 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3228 if (generate_interface_guard(kls, region) != NULL)
3229 // A guard was added. If the guard is taken, it was an interface.
3230 phi->add_req(intcon(1));
3231 // If we fall through, it's a plain class.
3232 query_value = intcon(0);
3233 break;
3234
3235 case vmIntrinsics::_isArray:
3236 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3237 if (generate_array_guard(kls, region) != NULL)
3238 // A guard was added. If the guard is taken, it was an array.
3239 phi->add_req(intcon(1));
3240 // If we fall through, it's a plain class.
3241 query_value = intcon(0);
3242 break;
3243
3244 case vmIntrinsics::_isPrimitive:
3245 query_value = intcon(0); // "normal" path produces false
3246 break;
3247
3248 case vmIntrinsics::_getSuperclass:
3249 // The rules here are somewhat unfortunate, but we can still do better
3250 // with random logic than with a JNI call.
3251 // Interfaces store null or Object as _super, but must report null.
3252 // Arrays store an intermediate super as _super, but must report Object.
3253 // Other types can report the actual _super.
3254 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3255 if (generate_interface_guard(kls, region) != NULL)
3256 // A guard was added. If the guard is taken, it was an interface.
3257 phi->add_req(null());
3258 if (generate_array_guard(kls, region) != NULL)
3259 // A guard was added. If the guard is taken, it was an array.
3260 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3261 // If we fall through, it's a plain class. Get its _super.
3262 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3263 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::BOTTOM));
3264 null_ctl = top();
3265 kls = null_check_oop(kls, &null_ctl);
3266 if (null_ctl != top()) {
3267 // If the guard is taken, Object.superClass is null (both klass and mirror).
3268 region->add_req(null_ctl);
3269 phi ->add_req(null());
3270 }
3271 if (!stopped()) {
3272 query_value = load_mirror_from_klass(kls);
3273 }
3274 break;
3275
3276 case vmIntrinsics::_getClassAccessFlags:
3277 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3278 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3279 break;
3280
3281 default:
3282 fatal_unexpected_iid(id);
3283 break;
3284 }
3285
3286 // Fall-through is the normal case of a query to a real class.
3287 phi->init_req(1, query_value);
3288 region->init_req(1, control());
3289
3290 C->set_has_split_ifs(true); // Has chance for split-if optimization
3291 set_result(region, phi);
3292 return true;
3293 }
3294
3295 //-------------------------inline_Class_cast-------------------
3296 bool LibraryCallKit::inline_Class_cast() {
3297 Node* mirror = argument(0); // Class
3298 Node* obj = argument(1);
3299 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3300 if (mirror_con == NULL) {
3301 return false; // dead path (mirror->is_top()).
3302 }
3303 if (obj == NULL || obj->is_top()) {
3304 return false; // dead path
3305 }
3306
3307 ciKlass* obj_klass = NULL;
3308 if (obj->is_ValueType()) {
3309 const TypeValueType* tvt = _gvn.type(obj)->is_valuetype();
3310 obj_klass = tvt->value_klass();
3311 } else {
3312 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3313 if (tp != NULL) {
3314 obj_klass = tp->klass();
3315 }
3316 }
3317
3318 // First, see if Class.cast() can be folded statically.
3319 // java_mirror_type() returns non-null for compile-time Class constants.
3320 ciType* tm = mirror_con->java_mirror_type();
3321 if (tm != NULL && tm->is_klass() &&
3322 obj_klass != NULL) {
3323 if (!obj_klass->is_loaded()) {
3324 // Don't use intrinsic when class is not loaded.
3325 return false;
3326 } else {
3327 int static_res = C->static_subtype_check(tm->as_klass(), obj_klass);
3328 if (static_res == Compile::SSC_always_true) {
3329 // isInstance() is true - fold the code.
3330 set_result(obj);
3331 return true;
3332 } else if (static_res == Compile::SSC_always_false) {
3333 // Don't use intrinsic, have to throw ClassCastException.
3334 // If the reference is null, the non-intrinsic bytecode will
3335 // be optimized appropriately.
3336 return false;
3337 }
3338 }
3339 }
3340
3341 // Bailout intrinsic and do normal inlining if exception path is frequent.
3342 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3343 return false;
3344 }
3345
3346 // Generate dynamic checks.
3347 // Class.cast() is java implementation of _checkcast bytecode.
3348 // Do checkcast (Parse::do_checkcast()) optimizations here.
3349
3350 mirror = null_check(mirror);
3351 // If mirror is dead, only null-path is taken.
3352 if (stopped()) {
3353 return true;
3354 }
3355
3356 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3357 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3358 RegionNode* region = new RegionNode(PATH_LIMIT);
3359 record_for_igvn(region);
3360
3361 // Now load the mirror's klass metaobject, and null-check it.
3362 // If kls is null, we have a primitive mirror and
3363 // nothing is an instance of a primitive type.
3364 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3365
3366 Node* res = top();
3367 if (!stopped()) {
3368 Node* bad_type_ctrl = top();
3369 // Do checkcast optimizations.
3370 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3371 region->init_req(_bad_type_path, bad_type_ctrl);
3372 }
3373 if (region->in(_prim_path) != top() ||
3374 region->in(_bad_type_path) != top()) {
3375 // Let Interpreter throw ClassCastException.
3376 PreserveJVMState pjvms(this);
3377 set_control(_gvn.transform(region));
3378 uncommon_trap(Deoptimization::Reason_intrinsic,
3379 Deoptimization::Action_maybe_recompile);
3380 }
3381 if (!stopped()) {
3382 set_result(res);
3383 }
3384 return true;
3385 }
3386
3387
3388 //--------------------------inline_native_subtype_check------------------------
3389 // This intrinsic takes the JNI calls out of the heart of
3390 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3391 bool LibraryCallKit::inline_native_subtype_check() {
3392 // Pull both arguments off the stack.
3393 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3394 args[0] = argument(0);
3395 args[1] = argument(1);
3396 Node* klasses[2]; // corresponding Klasses: superk, subk
3397 klasses[0] = klasses[1] = top();
3398
3399 enum {
3400 // A full decision tree on {superc is prim, subc is prim}:
3401 _prim_0_path = 1, // {P,N} => false
3402 // {P,P} & superc!=subc => false
3403 _prim_same_path, // {P,P} & superc==subc => true
3404 _prim_1_path, // {N,P} => false
3405 _ref_subtype_path, // {N,N} & subtype check wins => true
3406 _both_ref_path, // {N,N} & subtype check loses => false
3407 PATH_LIMIT
3408 };
3409
3410 RegionNode* region = new RegionNode(PATH_LIMIT);
3411 Node* phi = new PhiNode(region, TypeInt::BOOL);
3412 record_for_igvn(region);
3413
3414 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3415 const TypeKlassPtr* kls_type = TypeKlassPtr::BOTTOM;
3416 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3417
3418 // First null-check both mirrors and load each mirror's klass metaobject.
3419 int which_arg;
3420 for (which_arg = 0; which_arg <= 1; which_arg++) {
3421 Node* arg = args[which_arg];
3422 arg = null_check(arg);
3423 if (stopped()) break;
3424 args[which_arg] = arg;
3425
3426 Node* p = basic_plus_adr(arg, class_klass_offset);
3427 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3428 klasses[which_arg] = _gvn.transform(kls);
3429 }
3430
3431 // Having loaded both klasses, test each for null.
3432 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3433 for (which_arg = 0; which_arg <= 1; which_arg++) {
3434 Node* kls = klasses[which_arg];
3435 Node* null_ctl = top();
3436 kls = null_check_oop(kls, &null_ctl, never_see_null);
3437 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3438 region->init_req(prim_path, null_ctl);
3439 if (stopped()) break;
3440 klasses[which_arg] = kls;
3441 }
3442
3443 if (!stopped()) {
3444 // now we have two reference types, in klasses[0..1]
3445 Node* subk = klasses[1]; // the argument to isAssignableFrom
3446 Node* superk = klasses[0]; // the receiver
3447 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3448 // now we have a successful reference subtype check
3449 region->set_req(_ref_subtype_path, control());
3450 }
3451
3452 // If both operands are primitive (both klasses null), then
3453 // we must return true when they are identical primitives.
3454 // It is convenient to test this after the first null klass check.
3455 set_control(region->in(_prim_0_path)); // go back to first null check
3456 if (!stopped()) {
3457 // Since superc is primitive, make a guard for the superc==subc case.
3458 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3459 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3460 generate_guard(bol_eq, region, PROB_FAIR);
3461 if (region->req() == PATH_LIMIT+1) {
3462 // A guard was added. If the added guard is taken, superc==subc.
3463 region->swap_edges(PATH_LIMIT, _prim_same_path);
3464 region->del_req(PATH_LIMIT);
3465 }
3466 region->set_req(_prim_0_path, control()); // Not equal after all.
3467 }
3468
3469 // these are the only paths that produce 'true':
3470 phi->set_req(_prim_same_path, intcon(1));
3471 phi->set_req(_ref_subtype_path, intcon(1));
3472
3473 // pull together the cases:
3474 assert(region->req() == PATH_LIMIT, "sane region");
3475 for (uint i = 1; i < region->req(); i++) {
3476 Node* ctl = region->in(i);
3477 if (ctl == NULL || ctl == top()) {
3478 region->set_req(i, top());
3479 phi ->set_req(i, top());
3480 } else if (phi->in(i) == NULL) {
3481 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3482 }
3483 }
3484
3485 set_control(_gvn.transform(region));
3486 set_result(_gvn.transform(phi));
3487 return true;
3488 }
3489
3490 //---------------------generate_array_guard_common------------------------
3491 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region, ArrayKind kind) {
3492
3493 if (stopped()) {
3494 return NULL;
3495 }
3496
3497 // Like generate_guard, adds a new path onto the region.
3498 jint layout_con = 0;
3499 Node* layout_val = get_layout_helper(kls, layout_con);
3500 if (layout_val == NULL) {
3501 bool query = 0;
3502 switch(kind) {
3503 case ObjectArray: query = Klass::layout_helper_is_objArray(layout_con); break;
3504 case NonObjectArray: query = !Klass::layout_helper_is_objArray(layout_con); break;
3505 case TypeArray: query = Klass::layout_helper_is_typeArray(layout_con); break;
3506 case ValueArray: query = Klass::layout_helper_is_valueArray(layout_con); break;
3507 case AnyArray: query = Klass::layout_helper_is_array(layout_con); break;
3508 case NonArray: query = !Klass::layout_helper_is_array(layout_con); break;
3509 default:
3510 ShouldNotReachHere();
3511 }
3512 if (!query) {
3513 return NULL; // never a branch
3514 } else { // always a branch
3515 Node* always_branch = control();
3516 if (region != NULL)
3517 region->add_req(always_branch);
3518 set_control(top());
3519 return always_branch;
3520 }
3521 }
3522 unsigned int value = 0;
3523 BoolTest::mask btest = BoolTest::illegal;
3524 switch(kind) {
3525 case ObjectArray:
3526 case NonObjectArray: {
3527 value = Klass::_lh_array_tag_obj_value;
3528 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3529 btest = kind == ObjectArray ? BoolTest::eq : BoolTest::ne;
3530 break;
3531 }
3532 case TypeArray: {
3533 value = Klass::_lh_array_tag_type_value;
3534 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3535 btest = BoolTest::eq;
3536 break;
3537 }
3538 case ValueArray: {
3539 value = Klass::_lh_array_tag_vt_value;
3540 layout_val = _gvn.transform(new RShiftINode(layout_val, intcon(Klass::_lh_array_tag_shift)));
3541 btest = BoolTest::eq;
3542 break;
3543 }
3544 case AnyArray: value = Klass::_lh_neutral_value; btest = BoolTest::lt; break;
3545 case NonArray: value = Klass::_lh_neutral_value; btest = BoolTest::gt; break;
3546 default:
3547 ShouldNotReachHere();
3548 }
3549 // Now test the correct condition.
3550 jint nval = (jint)value;
3551 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3552 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3553 return generate_fair_guard(bol, region);
3554 }
3555
3556
3557 //-----------------------inline_native_newArray--------------------------
3558 // private static native Object java.lang.reflect.Array.newArray(Class<?> componentType, int length);
3559 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3560 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3561 Node* mirror;
3562 Node* count_val;
3563 if (uninitialized) {
3564 mirror = argument(1);
3565 count_val = argument(2);
3566 } else {
3567 mirror = argument(0);
3568 count_val = argument(1);
3569 }
3570
3571 mirror = null_check(mirror);
3572 // If mirror or obj is dead, only null-path is taken.
3573 if (stopped()) return true;
3574
3575 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3576 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3577 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3578 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3579 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3580
3581 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3582 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3583 result_reg, _slow_path);
3584 Node* normal_ctl = control();
3585 Node* no_array_ctl = result_reg->in(_slow_path);
3586
3587 // Generate code for the slow case. We make a call to newArray().
3588 set_control(no_array_ctl);
3589 if (!stopped()) {
3590 // Either the input type is void.class, or else the
3591 // array klass has not yet been cached. Either the
3592 // ensuing call will throw an exception, or else it
3593 // will cache the array klass for next time.
3594 PreserveJVMState pjvms(this);
3595 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3596 Node* slow_result = set_results_for_java_call(slow_call);
3597 // this->control() comes from set_results_for_java_call
3598 result_reg->set_req(_slow_path, control());
3599 result_val->set_req(_slow_path, slow_result);
3600 result_io ->set_req(_slow_path, i_o());
3601 result_mem->set_req(_slow_path, reset_memory());
3602 }
3603
3604 set_control(normal_ctl);
3605 if (!stopped()) {
3606 // Normal case: The array type has been cached in the java.lang.Class.
3607 // The following call works fine even if the array type is polymorphic.
3608 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3609 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3610 result_reg->init_req(_normal_path, control());
3611 result_val->init_req(_normal_path, obj);
3612 result_io ->init_req(_normal_path, i_o());
3613 result_mem->init_req(_normal_path, reset_memory());
3614
3615 if (uninitialized) {
3616 // Mark the allocation so that zeroing is skipped
3617 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
3618 alloc->maybe_set_complete(&_gvn);
3619 }
3620 }
3621
3622 // Return the combined state.
3623 set_i_o( _gvn.transform(result_io) );
3624 set_all_memory( _gvn.transform(result_mem));
3625
3626 C->set_has_split_ifs(true); // Has chance for split-if optimization
3627 set_result(result_reg, result_val);
3628 return true;
3629 }
3630
3631 //----------------------inline_native_getLength--------------------------
3632 // public static native int java.lang.reflect.Array.getLength(Object array);
3633 bool LibraryCallKit::inline_native_getLength() {
3634 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3635
3636 Node* array = null_check(argument(0));
3637 // If array is dead, only null-path is taken.
3638 if (stopped()) return true;
3639
3640 // Deoptimize if it is a non-array.
3641 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3642
3643 if (non_array != NULL) {
3644 PreserveJVMState pjvms(this);
3645 set_control(non_array);
3646 uncommon_trap(Deoptimization::Reason_intrinsic,
3647 Deoptimization::Action_maybe_recompile);
3648 }
3649
3650 // If control is dead, only non-array-path is taken.
3651 if (stopped()) return true;
3652
3653 // The works fine even if the array type is polymorphic.
3654 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3655 Node* result = load_array_length(array);
3656
3657 C->set_has_split_ifs(true); // Has chance for split-if optimization
3658 set_result(result);
3659 return true;
3660 }
3661
3662 //------------------------inline_array_copyOf----------------------------
3663 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3664 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3665 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3666 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3667
3668 // Get the arguments.
3669 Node* original = argument(0);
3670 Node* start = is_copyOfRange? argument(1): intcon(0);
3671 Node* end = is_copyOfRange? argument(2): argument(1);
3672 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3673
3674 const TypeAryPtr* original_t = _gvn.type(original)->isa_aryptr();
3675 const TypeInstPtr* mirror_t = _gvn.type(array_type_mirror)->isa_instptr();
3676 if (EnableValhalla && ValueArrayFlatten &&
3677 (original_t == NULL || mirror_t == NULL ||
3678 (mirror_t->java_mirror_type() == NULL &&
3679 (original_t->elem()->isa_valuetype() ||
3680 (original_t->elem()->make_oopptr() != NULL &&
3681 original_t->elem()->make_oopptr()->can_be_value_type()))))) {
3682 // We need to know statically if the copy is to a flattened array
3683 // or not but can't tell.
3684 return false;
3685 }
3686
3687 Node* newcopy = NULL;
3688
3689 // Set the original stack and the reexecute bit for the interpreter to reexecute
3690 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3691 { PreserveReexecuteState preexecs(this);
3692 jvms()->set_should_reexecute(true);
3693
3694 array_type_mirror = null_check(array_type_mirror);
3695 original = null_check(original);
3696
3697 // Check if a null path was taken unconditionally.
3698 if (stopped()) return true;
3699
3700 Node* orig_length = load_array_length(original);
3701
3702 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3703 klass_node = null_check(klass_node);
3704
3705 RegionNode* bailout = new RegionNode(1);
3706 record_for_igvn(bailout);
3707
3708 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3709 // Bail out if that is so.
3710 // Value type array may have object field that would require a
3711 // write barrier. Conservatively, go to slow path.
3712 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
3713 Node* not_objArray = !bs->array_copy_requires_gc_barriers(T_OBJECT) ? generate_typeArray_guard(klass_node, bailout) : generate_non_objArray_guard(klass_node, bailout);
3714 if (not_objArray != NULL) {
3715 // Improve the klass node's type from the new optimistic assumption:
3716 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3717 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, Type::Offset(0));
3718 Node* cast = new CastPPNode(klass_node, akls);
3719 cast->init_req(0, control());
3720 klass_node = _gvn.transform(cast);
3721 }
3722
3723 Node* original_kls = load_object_klass(original);
3724 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3725 // loads/stores but it is legal only if we're sure the
3726 // Arrays.copyOf would succeed. So we need all input arguments
3727 // to the copyOf to be validated, including that the copy to the
3728 // new array won't trigger an ArrayStoreException. That subtype
3729 // check can be optimized if we know something on the type of
3730 // the input array from type speculation.
3731 if (_gvn.type(klass_node)->singleton() && !stopped()) {
3732 ciKlass* subk = _gvn.type(original_kls)->is_klassptr()->klass();
3733 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3734
3735 int test = C->static_subtype_check(superk, subk);
3736 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3737 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3738 if (t_original->speculative_type() != NULL) {
3739 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3740 original_kls = load_object_klass(original);
3741 }
3742 }
3743 }
3744
3745 if (EnableValhalla) {
3746 // Either both or neither new array klass and original array
3747 // klass must be flattened
3748 Node* flattened_klass = generate_valueArray_guard(klass_node, NULL);
3749 generate_valueArray_guard(original_kls, bailout);
3750 if (flattened_klass != NULL) {
3751 RegionNode* r = new RegionNode(2);
3752 record_for_igvn(r);
3753 r->init_req(1, control());
3754 set_control(flattened_klass);
3755 generate_valueArray_guard(original_kls, r);
3756 bailout->add_req(control());
3757 set_control(_gvn.transform(r));
3758 }
3759 }
3760
3761 // Bail out if either start or end is negative.
3762 generate_negative_guard(start, bailout, &start);
3763 generate_negative_guard(end, bailout, &end);
3764
3765 Node* length = end;
3766 if (_gvn.type(start) != TypeInt::ZERO) {
3767 length = _gvn.transform(new SubINode(end, start));
3768 }
3769
3770 // Bail out if length is negative.
3771 // Without this the new_array would throw
3772 // NegativeArraySizeException but IllegalArgumentException is what
3773 // should be thrown
3774 generate_negative_guard(length, bailout, &length);
3775
3776 if (bailout->req() > 1) {
3777 PreserveJVMState pjvms(this);
3778 set_control(_gvn.transform(bailout));
3779 uncommon_trap(Deoptimization::Reason_intrinsic,
3780 Deoptimization::Action_maybe_recompile);
3781 }
3782
3783 if (!stopped()) {
3784 // How many elements will we copy from the original?
3785 // The answer is MinI(orig_length - start, length).
3786 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3787 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3788
3789 // Generate a direct call to the right arraycopy function(s).
3790 // We know the copy is disjoint but we might not know if the
3791 // oop stores need checking.
3792 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3793 // This will fail a store-check if x contains any non-nulls.
3794
3795 bool validated = false;
3796 // Reason_class_check rather than Reason_intrinsic because we
3797 // want to intrinsify even if this traps.
3798 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3799 Node* not_subtype_ctrl = gen_subtype_check(original_kls,
3800 klass_node);
3801
3802 if (not_subtype_ctrl != top()) {
3803 PreserveJVMState pjvms(this);
3804 set_control(not_subtype_ctrl);
3805 uncommon_trap(Deoptimization::Reason_class_check,
3806 Deoptimization::Action_make_not_entrant);
3807 assert(stopped(), "Should be stopped");
3808 }
3809 validated = true;
3810 }
3811
3812 if (!stopped()) {
3813 newcopy = new_array(klass_node, length, 0); // no arguments to push
3814
3815 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
3816 original_kls, klass_node);
3817 if (!is_copyOfRange) {
3818 ac->set_copyof(validated);
3819 } else {
3820 ac->set_copyofrange(validated);
3821 }
3822 Node* n = _gvn.transform(ac);
3823 if (n == ac) {
3824 ac->connect_outputs(this);
3825 } else {
3826 assert(validated, "shouldn't transform if all arguments not validated");
3827 set_all_memory(n);
3828 }
3829 }
3830 }
3831 } // original reexecute is set back here
3832
3833 C->set_has_split_ifs(true); // Has chance for split-if optimization
3834 if (!stopped()) {
3835 set_result(newcopy);
3836 }
3837 return true;
3838 }
3839
3840
3841 //----------------------generate_virtual_guard---------------------------
3842 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3843 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3844 RegionNode* slow_region) {
3845 ciMethod* method = callee();
3846 int vtable_index = method->vtable_index();
3847 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3848 "bad index %d", vtable_index);
3849 // Get the Method* out of the appropriate vtable entry.
3850 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
3851 vtable_index*vtableEntry::size_in_bytes() +
3852 vtableEntry::method_offset_in_bytes();
3853 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3854 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3855
3856 // Compare the target method with the expected method (e.g., Object.hashCode).
3857 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3858
3859 Node* native_call = makecon(native_call_addr);
3860 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3861 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3862
3863 return generate_slow_guard(test_native, slow_region);
3864 }
3865
3866 //-----------------------generate_method_call----------------------------
3867 // Use generate_method_call to make a slow-call to the real
3868 // method if the fast path fails. An alternative would be to
3869 // use a stub like OptoRuntime::slow_arraycopy_Java.
3870 // This only works for expanding the current library call,
3871 // not another intrinsic. (E.g., don't use this for making an
3872 // arraycopy call inside of the copyOf intrinsic.)
3873 CallJavaNode*
3874 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3875 // When compiling the intrinsic method itself, do not use this technique.
3876 guarantee(callee() != C->method(), "cannot make slow-call to self");
3877
3878 ciMethod* method = callee();
3879 // ensure the JVMS we have will be correct for this call
3880 guarantee(method_id == method->intrinsic_id(), "must match");
3881
3882 const TypeFunc* tf = TypeFunc::make(method);
3883 CallJavaNode* slow_call;
3884 if (is_static) {
3885 assert(!is_virtual, "");
3886 slow_call = new CallStaticJavaNode(C, tf,
3887 SharedRuntime::get_resolve_static_call_stub(),
3888 method, bci());
3889 } else if (is_virtual) {
3890 null_check_receiver();
3891 int vtable_index = Method::invalid_vtable_index;
3892 if (UseInlineCaches) {
3893 // Suppress the vtable call
3894 } else {
3895 // hashCode and clone are not a miranda methods,
3896 // so the vtable index is fixed.
3897 // No need to use the linkResolver to get it.
3898 vtable_index = method->vtable_index();
3899 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3900 "bad index %d", vtable_index);
3901 }
3902 slow_call = new CallDynamicJavaNode(tf,
3903 SharedRuntime::get_resolve_virtual_call_stub(),
3904 method, vtable_index, bci());
3905 } else { // neither virtual nor static: opt_virtual
3906 null_check_receiver();
3907 slow_call = new CallStaticJavaNode(C, tf,
3908 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3909 method, bci());
3910 slow_call->set_optimized_virtual(true);
3911 }
3912 set_arguments_for_java_call(slow_call);
3913 set_edges_for_java_call(slow_call);
3914 return slow_call;
3915 }
3916
3917
3918 /**
3919 * Build special case code for calls to hashCode on an object. This call may
3920 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3921 * slightly different code.
3922 */
3923 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3924 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3925 assert(!(is_virtual && is_static), "either virtual, special, or static");
3926
3927 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3928
3929 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3930 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3931 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3932 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3933 Node* obj = argument(0);
3934
3935 if (obj->is_ValueType()) {
3936 return false;
3937 }
3938
3939 if (!is_static) {
3940 // Check for hashing null object
3941 obj = null_check_receiver();
3942 if (stopped()) return true; // unconditionally null
3943 result_reg->init_req(_null_path, top());
3944 result_val->init_req(_null_path, top());
3945 } else {
3946 // Do a null check, and return zero if null.
3947 // System.identityHashCode(null) == 0
3948 Node* null_ctl = top();
3949 obj = null_check_oop(obj, &null_ctl);
3950 result_reg->init_req(_null_path, null_ctl);
3951 result_val->init_req(_null_path, _gvn.intcon(0));
3952 }
3953
3954 // Unconditionally null? Then return right away.
3955 if (stopped()) {
3956 set_control( result_reg->in(_null_path));
3957 if (!stopped())
3958 set_result(result_val->in(_null_path));
3959 return true;
3960 }
3961
3962 // We only go to the fast case code if we pass a number of guards. The
3963 // paths which do not pass are accumulated in the slow_region.
3964 RegionNode* slow_region = new RegionNode(1);
3965 record_for_igvn(slow_region);
3966
3967 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3968 assert(!obj_type->isa_valuetype() || !obj_type->is_valuetypeptr(), "no value type here");
3969 if (is_static && obj_type->can_be_value_type()) {
3970 Node* obj_klass = load_object_klass(obj);
3971 generate_value_guard(obj_klass, slow_region);
3972 }
3973
3974 // If this is a virtual call, we generate a funny guard. We pull out
3975 // the vtable entry corresponding to hashCode() from the target object.
3976 // If the target method which we are calling happens to be the native
3977 // Object hashCode() method, we pass the guard. We do not need this
3978 // guard for non-virtual calls -- the caller is known to be the native
3979 // Object hashCode().
3980 if (is_virtual) {
3981 // After null check, get the object's klass.
3982 Node* obj_klass = load_object_klass(obj);
3983 generate_virtual_guard(obj_klass, slow_region);
3984 }
3985
3986 // Get the header out of the object, use LoadMarkNode when available
3987 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3988 // The control of the load must be NULL. Otherwise, the load can move before
3989 // the null check after castPP removal.
3990 Node* no_ctrl = NULL;
3991 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3992
3993 // Test the header to see if it is unlocked.
3994 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3995 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3996 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3997 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3998 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3999
4000 generate_slow_guard(test_unlocked, slow_region);
4001
4002 // Get the hash value and check to see that it has been properly assigned.
4003 // We depend on hash_mask being at most 32 bits and avoid the use of
4004 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4005 // vm: see markOop.hpp.
4006 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4007 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4008 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4009 // This hack lets the hash bits live anywhere in the mark object now, as long
4010 // as the shift drops the relevant bits into the low 32 bits. Note that
4011 // Java spec says that HashCode is an int so there's no point in capturing
4012 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4013 hshifted_header = ConvX2I(hshifted_header);
4014 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4015
4016 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4017 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4018 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4019
4020 generate_slow_guard(test_assigned, slow_region);
4021
4022 Node* init_mem = reset_memory();
4023 // fill in the rest of the null path:
4024 result_io ->init_req(_null_path, i_o());
4025 result_mem->init_req(_null_path, init_mem);
4026
4027 result_val->init_req(_fast_path, hash_val);
4028 result_reg->init_req(_fast_path, control());
4029 result_io ->init_req(_fast_path, i_o());
4030 result_mem->init_req(_fast_path, init_mem);
4031
4032 // Generate code for the slow case. We make a call to hashCode().
4033 set_control(_gvn.transform(slow_region));
4034 if (!stopped()) {
4035 // No need for PreserveJVMState, because we're using up the present state.
4036 set_all_memory(init_mem);
4037 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4038 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4039 Node* slow_result = set_results_for_java_call(slow_call);
4040 // this->control() comes from set_results_for_java_call
4041 result_reg->init_req(_slow_path, control());
4042 result_val->init_req(_slow_path, slow_result);
4043 result_io ->set_req(_slow_path, i_o());
4044 result_mem ->set_req(_slow_path, reset_memory());
4045 }
4046
4047 // Return the combined state.
4048 set_i_o( _gvn.transform(result_io) );
4049 set_all_memory( _gvn.transform(result_mem));
4050
4051 set_result(result_reg, result_val);
4052 return true;
4053 }
4054
4055 //---------------------------inline_native_getClass----------------------------
4056 // public final native Class<?> java.lang.Object.getClass();
4057 //
4058 // Build special case code for calls to getClass on an object.
4059 bool LibraryCallKit::inline_native_getClass() {
4060 Node* obj = argument(0);
4061 if (obj->is_ValueType()) {
4062 ciKlass* vk = _gvn.type(obj)->is_valuetype()->value_klass();
4063 set_result(makecon(TypeInstPtr::make(vk->java_mirror())));
4064 return true;
4065 }
4066 obj = null_check_receiver();
4067 if (stopped()) return true;
4068 set_result(load_mirror_from_klass(load_object_klass(obj)));
4069 return true;
4070 }
4071
4072 //-----------------inline_native_Reflection_getCallerClass---------------------
4073 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4074 //
4075 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4076 //
4077 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4078 // in that it must skip particular security frames and checks for
4079 // caller sensitive methods.
4080 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4081 #ifndef PRODUCT
4082 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4083 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4084 }
4085 #endif
4086
4087 if (!jvms()->has_method()) {
4088 #ifndef PRODUCT
4089 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4090 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4091 }
4092 #endif
4093 return false;
4094 }
4095
4096 // Walk back up the JVM state to find the caller at the required
4097 // depth.
4098 JVMState* caller_jvms = jvms();
4099
4100 // Cf. JVM_GetCallerClass
4101 // NOTE: Start the loop at depth 1 because the current JVM state does
4102 // not include the Reflection.getCallerClass() frame.
4103 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4104 ciMethod* m = caller_jvms->method();
4105 switch (n) {
4106 case 0:
4107 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4108 break;
4109 case 1:
4110 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4111 if (!m->caller_sensitive()) {
4112 #ifndef PRODUCT
4113 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4114 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4115 }
4116 #endif
4117 return false; // bail-out; let JVM_GetCallerClass do the work
4118 }
4119 break;
4120 default:
4121 if (!m->is_ignored_by_security_stack_walk()) {
4122 // We have reached the desired frame; return the holder class.
4123 // Acquire method holder as java.lang.Class and push as constant.
4124 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4125 ciInstance* caller_mirror = caller_klass->java_mirror();
4126 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4127
4128 #ifndef PRODUCT
4129 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4130 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());
4131 tty->print_cr(" JVM state at this point:");
4132 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4133 ciMethod* m = jvms()->of_depth(i)->method();
4134 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4135 }
4136 }
4137 #endif
4138 return true;
4139 }
4140 break;
4141 }
4142 }
4143
4144 #ifndef PRODUCT
4145 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4146 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4147 tty->print_cr(" JVM state at this point:");
4148 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4149 ciMethod* m = jvms()->of_depth(i)->method();
4150 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4151 }
4152 }
4153 #endif
4154
4155 return false; // bail-out; let JVM_GetCallerClass do the work
4156 }
4157
4158 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4159 Node* arg = argument(0);
4160 Node* result = NULL;
4161
4162 switch (id) {
4163 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4164 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4165 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4166 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4167
4168 case vmIntrinsics::_doubleToLongBits: {
4169 // two paths (plus control) merge in a wood
4170 RegionNode *r = new RegionNode(3);
4171 Node *phi = new PhiNode(r, TypeLong::LONG);
4172
4173 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4174 // Build the boolean node
4175 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4176
4177 // Branch either way.
4178 // NaN case is less traveled, which makes all the difference.
4179 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4180 Node *opt_isnan = _gvn.transform(ifisnan);
4181 assert( opt_isnan->is_If(), "Expect an IfNode");
4182 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4183 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4184
4185 set_control(iftrue);
4186
4187 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4188 Node *slow_result = longcon(nan_bits); // return NaN
4189 phi->init_req(1, _gvn.transform( slow_result ));
4190 r->init_req(1, iftrue);
4191
4192 // Else fall through
4193 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4194 set_control(iffalse);
4195
4196 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4197 r->init_req(2, iffalse);
4198
4199 // Post merge
4200 set_control(_gvn.transform(r));
4201 record_for_igvn(r);
4202
4203 C->set_has_split_ifs(true); // Has chance for split-if optimization
4204 result = phi;
4205 assert(result->bottom_type()->isa_long(), "must be");
4206 break;
4207 }
4208
4209 case vmIntrinsics::_floatToIntBits: {
4210 // two paths (plus control) merge in a wood
4211 RegionNode *r = new RegionNode(3);
4212 Node *phi = new PhiNode(r, TypeInt::INT);
4213
4214 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4215 // Build the boolean node
4216 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4217
4218 // Branch either way.
4219 // NaN case is less traveled, which makes all the difference.
4220 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4221 Node *opt_isnan = _gvn.transform(ifisnan);
4222 assert( opt_isnan->is_If(), "Expect an IfNode");
4223 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4224 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4225
4226 set_control(iftrue);
4227
4228 static const jint nan_bits = 0x7fc00000;
4229 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4230 phi->init_req(1, _gvn.transform( slow_result ));
4231 r->init_req(1, iftrue);
4232
4233 // Else fall through
4234 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4235 set_control(iffalse);
4236
4237 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4238 r->init_req(2, iffalse);
4239
4240 // Post merge
4241 set_control(_gvn.transform(r));
4242 record_for_igvn(r);
4243
4244 C->set_has_split_ifs(true); // Has chance for split-if optimization
4245 result = phi;
4246 assert(result->bottom_type()->isa_int(), "must be");
4247 break;
4248 }
4249
4250 default:
4251 fatal_unexpected_iid(id);
4252 break;
4253 }
4254 set_result(_gvn.transform(result));
4255 return true;
4256 }
4257
4258 //----------------------inline_unsafe_copyMemory-------------------------
4259 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4260 bool LibraryCallKit::inline_unsafe_copyMemory() {
4261 if (callee()->is_static()) return false; // caller must have the capability!
4262 null_check_receiver(); // null-check receiver
4263 if (stopped()) return true;
4264
4265 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4266
4267 Node* src_ptr = argument(1); // type: oop
4268 Node* src_off = ConvL2X(argument(2)); // type: long
4269 Node* dst_ptr = argument(4); // type: oop
4270 Node* dst_off = ConvL2X(argument(5)); // type: long
4271 Node* size = ConvL2X(argument(7)); // type: long
4272
4273 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4274 "fieldOffset must be byte-scaled");
4275
4276 Node* src = make_unsafe_address(src_ptr, src_off);
4277 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4278
4279 // Conservatively insert a memory barrier on all memory slices.
4280 // Do not let writes of the copy source or destination float below the copy.
4281 insert_mem_bar(Op_MemBarCPUOrder);
4282
4283 // Call it. Note that the length argument is not scaled.
4284 make_runtime_call(RC_LEAF|RC_NO_FP,
4285 OptoRuntime::fast_arraycopy_Type(),
4286 StubRoutines::unsafe_arraycopy(),
4287 "unsafe_arraycopy",
4288 TypeRawPtr::BOTTOM,
4289 src, dst, size XTOP);
4290
4291 // Do not let reads of the copy destination float above the copy.
4292 insert_mem_bar(Op_MemBarCPUOrder);
4293
4294 return true;
4295 }
4296
4297 //------------------------clone_coping-----------------------------------
4298 // Helper function for inline_native_clone.
4299 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array) {
4300 assert(obj_size != NULL, "");
4301 Node* raw_obj = alloc_obj->in(1);
4302 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4303
4304 AllocateNode* alloc = NULL;
4305 if (ReduceBulkZeroing) {
4306 // We will be completely responsible for initializing this object -
4307 // mark Initialize node as complete.
4308 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4309 // The object was just allocated - there should be no any stores!
4310 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4311 // Mark as complete_with_arraycopy so that on AllocateNode
4312 // expansion, we know this AllocateNode is initialized by an array
4313 // copy and a StoreStore barrier exists after the array copy.
4314 alloc->initialization()->set_complete_with_arraycopy();
4315 }
4316
4317 // Copy the fastest available way.
4318 // TODO: generate fields copies for small objects instead.
4319 Node* size = _gvn.transform(obj_size);
4320
4321 // Exclude the header but include array length to copy by 8 bytes words.
4322 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4323 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4324 instanceOopDesc::base_offset_in_bytes();
4325 // base_off:
4326 // 8 - 32-bit VM
4327 // 12 - 64-bit VM, compressed klass
4328 // 16 - 64-bit VM, normal klass
4329 if (base_off % BytesPerLong != 0) {
4330 assert(UseCompressedClassPointers, "");
4331 if (is_array) {
4332 // Exclude length to copy by 8 bytes words.
4333 base_off += sizeof(int);
4334 } else {
4335 // Include klass to copy by 8 bytes words.
4336 base_off = instanceOopDesc::klass_offset_in_bytes();
4337 }
4338 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4339 }
4340 Node* src_base = basic_plus_adr(obj, base_off);
4341 Node* dst_base = basic_plus_adr(alloc_obj, base_off);
4342
4343 // Compute the length also, if needed:
4344 Node* countx = size;
4345 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4346 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong)));
4347
4348 access_clone(control(), src_base, dst_base, countx, is_array);
4349
4350 // Do not let reads from the cloned object float above the arraycopy.
4351 if (alloc != NULL) {
4352 // Do not let stores that initialize this object be reordered with
4353 // a subsequent store that would make this object accessible by
4354 // other threads.
4355 // Record what AllocateNode this StoreStore protects so that
4356 // escape analysis can go from the MemBarStoreStoreNode to the
4357 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4358 // based on the escape status of the AllocateNode.
4359 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out_or_null(AllocateNode::RawAddress));
4360 } else {
4361 insert_mem_bar(Op_MemBarCPUOrder);
4362 }
4363 }
4364
4365 //------------------------inline_native_clone----------------------------
4366 // protected native Object java.lang.Object.clone();
4367 //
4368 // Here are the simple edge cases:
4369 // null receiver => normal trap
4370 // virtual and clone was overridden => slow path to out-of-line clone
4371 // not cloneable or finalizer => slow path to out-of-line Object.clone
4372 //
4373 // The general case has two steps, allocation and copying.
4374 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4375 //
4376 // Copying also has two cases, oop arrays and everything else.
4377 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4378 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4379 //
4380 // These steps fold up nicely if and when the cloned object's klass
4381 // can be sharply typed as an object array, a type array, or an instance.
4382 //
4383 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4384 PhiNode* result_val;
4385
4386 // Set the reexecute bit for the interpreter to reexecute
4387 // the bytecode that invokes Object.clone if deoptimization happens.
4388 { PreserveReexecuteState preexecs(this);
4389 jvms()->set_should_reexecute(true);
4390
4391 Node* obj = argument(0);
4392 if (obj->is_ValueType()) {
4393 return false;
4394 }
4395
4396 obj = null_check_receiver();
4397 if (stopped()) return true;
4398
4399 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4400
4401 // If we are going to clone an instance, we need its exact type to
4402 // know the number and types of fields to convert the clone to
4403 // loads/stores. Maybe a speculative type can help us.
4404 if (!obj_type->klass_is_exact() &&
4405 obj_type->speculative_type() != NULL &&
4406 obj_type->speculative_type()->is_instance_klass() &&
4407 !obj_type->speculative_type()->is_valuetype()) {
4408 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4409 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4410 !spec_ik->has_injected_fields()) {
4411 ciKlass* k = obj_type->klass();
4412 if (!k->is_instance_klass() ||
4413 k->as_instance_klass()->is_interface() ||
4414 k->as_instance_klass()->has_subklass()) {
4415 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4416 }
4417 }
4418 }
4419
4420 Node* obj_klass = load_object_klass(obj);
4421 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4422 const TypeOopPtr* toop = ((tklass != NULL)
4423 ? tklass->as_instance_type()
4424 : TypeInstPtr::NOTNULL);
4425
4426 // Conservatively insert a memory barrier on all memory slices.
4427 // Do not let writes into the original float below the clone.
4428 insert_mem_bar(Op_MemBarCPUOrder);
4429
4430 // paths into result_reg:
4431 enum {
4432 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4433 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4434 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4435 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4436 PATH_LIMIT
4437 };
4438 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4439 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4440 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4441 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4442 record_for_igvn(result_reg);
4443
4444 // We only go to the fast case code if we pass a number of guards.
4445 // The paths which do not pass are accumulated in the slow_region.
4446 RegionNode* slow_region = new RegionNode(1);
4447 record_for_igvn(slow_region);
4448
4449 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4450 if (array_ctl != NULL) {
4451 // It's an array.
4452 PreserveJVMState pjvms(this);
4453 set_control(array_ctl);
4454
4455 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4456 if (bs->array_copy_requires_gc_barriers(T_OBJECT)) {
4457 // Value type array may have object field that would require a
4458 // write barrier. Conservatively, go to slow path.
4459 generate_valueArray_guard(obj_klass, slow_region);
4460 }
4461
4462 if (!stopped()) {
4463 Node* obj_length = load_array_length(obj);
4464 Node* obj_size = NULL;
4465 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4466
4467 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2();
4468 if (bs->array_copy_requires_gc_barriers(T_OBJECT)) {
4469 // If it is an oop array, it requires very special treatment,
4470 // because gc barriers are required when accessing the array.
4471 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4472 if (is_obja != NULL) {
4473 PreserveJVMState pjvms2(this);
4474 set_control(is_obja);
4475 // Generate a direct call to the right arraycopy function(s).
4476 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4477 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4478 ac->set_cloneoop();
4479 Node* n = _gvn.transform(ac);
4480 assert(n == ac, "cannot disappear");
4481 ac->connect_outputs(this);
4482
4483 result_reg->init_req(_objArray_path, control());
4484 result_val->init_req(_objArray_path, alloc_obj);
4485 result_i_o ->set_req(_objArray_path, i_o());
4486 result_mem ->set_req(_objArray_path, reset_memory());
4487 }
4488 }
4489
4490 // Otherwise, there are no barriers to worry about.
4491 // (We can dispense with card marks if we know the allocation
4492 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4493 // causes the non-eden paths to take compensating steps to
4494 // simulate a fresh allocation, so that no further
4495 // card marks are required in compiled code to initialize
4496 // the object.)
4497
4498 if (!stopped()) {
4499 copy_to_clone(obj, alloc_obj, obj_size, true);
4500
4501 // Present the results of the copy.
4502 result_reg->init_req(_array_path, control());
4503 result_val->init_req(_array_path, alloc_obj);
4504 result_i_o ->set_req(_array_path, i_o());
4505 result_mem ->set_req(_array_path, reset_memory());
4506 }
4507 }
4508 }
4509
4510 if (!stopped()) {
4511 // It's an instance (we did array above). Make the slow-path tests.
4512 // If this is a virtual call, we generate a funny guard. We grab
4513 // the vtable entry corresponding to clone() from the target object.
4514 // If the target method which we are calling happens to be the
4515 // Object clone() method, we pass the guard. We do not need this
4516 // guard for non-virtual calls; the caller is known to be the native
4517 // Object clone().
4518 if (is_virtual) {
4519 generate_virtual_guard(obj_klass, slow_region);
4520 }
4521
4522 // The object must be easily cloneable and must not have a finalizer.
4523 // Both of these conditions may be checked in a single test.
4524 // We could optimize the test further, but we don't care.
4525 generate_access_flags_guard(obj_klass,
4526 // Test both conditions:
4527 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4528 // Must be cloneable but not finalizer:
4529 JVM_ACC_IS_CLONEABLE_FAST,
4530 slow_region);
4531 }
4532
4533 if (!stopped()) {
4534 // It's an instance, and it passed the slow-path tests.
4535 PreserveJVMState pjvms(this);
4536 Node* obj_size = NULL;
4537 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4538 // is reexecuted if deoptimization occurs and there could be problems when merging
4539 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4540 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4541
4542 copy_to_clone(obj, alloc_obj, obj_size, false);
4543
4544 // Present the results of the slow call.
4545 result_reg->init_req(_instance_path, control());
4546 result_val->init_req(_instance_path, alloc_obj);
4547 result_i_o ->set_req(_instance_path, i_o());
4548 result_mem ->set_req(_instance_path, reset_memory());
4549 }
4550
4551 // Generate code for the slow case. We make a call to clone().
4552 set_control(_gvn.transform(slow_region));
4553 if (!stopped()) {
4554 PreserveJVMState pjvms(this);
4555 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4556 // We need to deoptimize on exception (see comment above)
4557 Node* slow_result = set_results_for_java_call(slow_call, false, /* deoptimize */ true);
4558 // this->control() comes from set_results_for_java_call
4559 result_reg->init_req(_slow_path, control());
4560 result_val->init_req(_slow_path, slow_result);
4561 result_i_o ->set_req(_slow_path, i_o());
4562 result_mem ->set_req(_slow_path, reset_memory());
4563 }
4564
4565 // Return the combined state.
4566 set_control( _gvn.transform(result_reg));
4567 set_i_o( _gvn.transform(result_i_o));
4568 set_all_memory( _gvn.transform(result_mem));
4569 } // original reexecute is set back here
4570
4571 set_result(_gvn.transform(result_val));
4572 return true;
4573 }
4574
4575 // If we have a tighly coupled allocation, the arraycopy may take care
4576 // of the array initialization. If one of the guards we insert between
4577 // the allocation and the arraycopy causes a deoptimization, an
4578 // unitialized array will escape the compiled method. To prevent that
4579 // we set the JVM state for uncommon traps between the allocation and
4580 // the arraycopy to the state before the allocation so, in case of
4581 // deoptimization, we'll reexecute the allocation and the
4582 // initialization.
4583 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4584 if (alloc != NULL) {
4585 ciMethod* trap_method = alloc->jvms()->method();
4586 int trap_bci = alloc->jvms()->bci();
4587
4588 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4589 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4590 // Make sure there's no store between the allocation and the
4591 // arraycopy otherwise visible side effects could be rexecuted
4592 // in case of deoptimization and cause incorrect execution.
4593 bool no_interfering_store = true;
4594 Node* mem = alloc->in(TypeFunc::Memory);
4595 if (mem->is_MergeMem()) {
4596 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4597 Node* n = mms.memory();
4598 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4599 assert(n->is_Store(), "what else?");
4600 no_interfering_store = false;
4601 break;
4602 }
4603 }
4604 } else {
4605 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4606 Node* n = mms.memory();
4607 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4608 assert(n->is_Store(), "what else?");
4609 no_interfering_store = false;
4610 break;
4611 }
4612 }
4613 }
4614
4615 if (no_interfering_store) {
4616 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4617 uint size = alloc->req();
4618 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4619 old_jvms->set_map(sfpt);
4620 for (uint i = 0; i < size; i++) {
4621 sfpt->init_req(i, alloc->in(i));
4622 }
4623 // re-push array length for deoptimization
4624 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4625 old_jvms->set_sp(old_jvms->sp()+1);
4626 old_jvms->set_monoff(old_jvms->monoff()+1);
4627 old_jvms->set_scloff(old_jvms->scloff()+1);
4628 old_jvms->set_endoff(old_jvms->endoff()+1);
4629 old_jvms->set_should_reexecute(true);
4630
4631 sfpt->set_i_o(map()->i_o());
4632 sfpt->set_memory(map()->memory());
4633 sfpt->set_control(map()->control());
4634
4635 JVMState* saved_jvms = jvms();
4636 saved_reexecute_sp = _reexecute_sp;
4637
4638 set_jvms(sfpt->jvms());
4639 _reexecute_sp = jvms()->sp();
4640
4641 return saved_jvms;
4642 }
4643 }
4644 }
4645 return NULL;
4646 }
4647
4648 // In case of a deoptimization, we restart execution at the
4649 // allocation, allocating a new array. We would leave an uninitialized
4650 // array in the heap that GCs wouldn't expect. Move the allocation
4651 // after the traps so we don't allocate the array if we
4652 // deoptimize. This is possible because tightly_coupled_allocation()
4653 // guarantees there's no observer of the allocated array at this point
4654 // and the control flow is simple enough.
4655 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
4656 int saved_reexecute_sp, uint new_idx) {
4657 if (saved_jvms != NULL && !stopped()) {
4658 assert(alloc != NULL, "only with a tightly coupled allocation");
4659 // restore JVM state to the state at the arraycopy
4660 saved_jvms->map()->set_control(map()->control());
4661 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4662 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4663 // If we've improved the types of some nodes (null check) while
4664 // emitting the guards, propagate them to the current state
4665 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
4666 set_jvms(saved_jvms);
4667 _reexecute_sp = saved_reexecute_sp;
4668
4669 // Remove the allocation from above the guards
4670 CallProjections* callprojs = alloc->extract_projections(true);
4671 InitializeNode* init = alloc->initialization();
4672 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4673 C->gvn_replace_by(callprojs->fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4674 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4675 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4676
4677 // move the allocation here (after the guards)
4678 _gvn.hash_delete(alloc);
4679 alloc->set_req(TypeFunc::Control, control());
4680 alloc->set_req(TypeFunc::I_O, i_o());
4681 Node *mem = reset_memory();
4682 set_all_memory(mem);
4683 alloc->set_req(TypeFunc::Memory, mem);
4684 set_control(init->proj_out_or_null(TypeFunc::Control));
4685 set_i_o(callprojs->fallthrough_ioproj);
4686
4687 // Update memory as done in GraphKit::set_output_for_allocation()
4688 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4689 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4690 if (ary_type->isa_aryptr() && length_type != NULL) {
4691 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4692 }
4693 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4694 int elemidx = C->get_alias_index(telemref);
4695 set_memory(init->proj_out_or_null(TypeFunc::Memory), Compile::AliasIdxRaw);
4696 set_memory(init->proj_out_or_null(TypeFunc::Memory), elemidx);
4697
4698 Node* allocx = _gvn.transform(alloc);
4699 assert(allocx == alloc, "where has the allocation gone?");
4700 assert(dest->is_CheckCastPP(), "not an allocation result?");
4701
4702 _gvn.hash_delete(dest);
4703 dest->set_req(0, control());
4704 Node* destx = _gvn.transform(dest);
4705 assert(destx == dest, "where has the allocation result gone?");
4706 }
4707 }
4708
4709
4710 //------------------------------inline_arraycopy-----------------------
4711 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4712 // Object dest, int destPos,
4713 // int length);
4714 bool LibraryCallKit::inline_arraycopy() {
4715 // Get the arguments.
4716 Node* src = argument(0); // type: oop
4717 Node* src_offset = argument(1); // type: int
4718 Node* dest = argument(2); // type: oop
4719 Node* dest_offset = argument(3); // type: int
4720 Node* length = argument(4); // type: int
4721
4722 uint new_idx = C->unique();
4723
4724 // Check for allocation before we add nodes that would confuse
4725 // tightly_coupled_allocation()
4726 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4727
4728 int saved_reexecute_sp = -1;
4729 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4730 // See arraycopy_restore_alloc_state() comment
4731 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4732 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4733 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
4734 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4735
4736 // The following tests must be performed
4737 // (1) src and dest are arrays.
4738 // (2) src and dest arrays must have elements of the same BasicType
4739 // (3) src and dest must not be null.
4740 // (4) src_offset must not be negative.
4741 // (5) dest_offset must not be negative.
4742 // (6) length must not be negative.
4743 // (7) src_offset + length must not exceed length of src.
4744 // (8) dest_offset + length must not exceed length of dest.
4745 // (9) each element of an oop array must be assignable
4746
4747 // (3) src and dest must not be null.
4748 // always do this here because we need the JVM state for uncommon traps
4749 Node* null_ctl = top();
4750 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4751 assert(null_ctl->is_top(), "no null control here");
4752 dest = null_check(dest, T_ARRAY);
4753
4754 if (!can_emit_guards) {
4755 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4756 // guards but the arraycopy node could still take advantage of a
4757 // tightly allocated allocation. tightly_coupled_allocation() is
4758 // called again to make sure it takes the null check above into
4759 // account: the null check is mandatory and if it caused an
4760 // uncommon trap to be emitted then the allocation can't be
4761 // considered tightly coupled in this context.
4762 alloc = tightly_coupled_allocation(dest, NULL);
4763 }
4764
4765 bool validated = false;
4766
4767 const Type* src_type = _gvn.type(src);
4768 const Type* dest_type = _gvn.type(dest);
4769 const TypeAryPtr* top_src = src_type->isa_aryptr();
4770 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4771
4772 // Do we have the type of src?
4773 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4774 // Do we have the type of dest?
4775 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4776 // Is the type for src from speculation?
4777 bool src_spec = false;
4778 // Is the type for dest from speculation?
4779 bool dest_spec = false;
4780
4781 if ((!has_src || !has_dest) && can_emit_guards) {
4782 // We don't have sufficient type information, let's see if
4783 // speculative types can help. We need to have types for both src
4784 // and dest so that it pays off.
4785
4786 // Do we already have or could we have type information for src
4787 bool could_have_src = has_src;
4788 // Do we already have or could we have type information for dest
4789 bool could_have_dest = has_dest;
4790
4791 ciKlass* src_k = NULL;
4792 if (!has_src) {
4793 src_k = src_type->speculative_type_not_null();
4794 if (src_k != NULL && src_k->is_array_klass()) {
4795 could_have_src = true;
4796 }
4797 }
4798
4799 ciKlass* dest_k = NULL;
4800 if (!has_dest) {
4801 dest_k = dest_type->speculative_type_not_null();
4802 if (dest_k != NULL && dest_k->is_array_klass()) {
4803 could_have_dest = true;
4804 }
4805 }
4806
4807 if (could_have_src && could_have_dest) {
4808 // This is going to pay off so emit the required guards
4809 if (!has_src) {
4810 src = maybe_cast_profiled_obj(src, src_k, true);
4811 src_type = _gvn.type(src);
4812 top_src = src_type->isa_aryptr();
4813 has_src = (top_src != NULL && top_src->klass() != NULL);
4814 src_spec = true;
4815 }
4816 if (!has_dest) {
4817 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4818 dest_type = _gvn.type(dest);
4819 top_dest = dest_type->isa_aryptr();
4820 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4821 dest_spec = true;
4822 }
4823 }
4824 }
4825
4826 if (has_src && has_dest && can_emit_guards) {
4827 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4828 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4829 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4830 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4831
4832 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4833 // If both arrays are object arrays then having the exact types
4834 // for both will remove the need for a subtype check at runtime
4835 // before the call and may make it possible to pick a faster copy
4836 // routine (without a subtype check on every element)
4837 // Do we have the exact type of src?
4838 bool could_have_src = src_spec;
4839 // Do we have the exact type of dest?
4840 bool could_have_dest = dest_spec;
4841 ciKlass* src_k = top_src->klass();
4842 ciKlass* dest_k = top_dest->klass();
4843 if (!src_spec) {
4844 src_k = src_type->speculative_type_not_null();
4845 if (src_k != NULL && src_k->is_array_klass()) {
4846 could_have_src = true;
4847 }
4848 }
4849 if (!dest_spec) {
4850 dest_k = dest_type->speculative_type_not_null();
4851 if (dest_k != NULL && dest_k->is_array_klass()) {
4852 could_have_dest = true;
4853 }
4854 }
4855 if (could_have_src && could_have_dest) {
4856 // If we can have both exact types, emit the missing guards
4857 if (could_have_src && !src_spec) {
4858 src = maybe_cast_profiled_obj(src, src_k, true);
4859 }
4860 if (could_have_dest && !dest_spec) {
4861 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4862 }
4863 }
4864 }
4865 }
4866
4867 ciMethod* trap_method = method();
4868 int trap_bci = bci();
4869 if (saved_jvms != NULL) {
4870 trap_method = alloc->jvms()->method();
4871 trap_bci = alloc->jvms()->bci();
4872 }
4873
4874 bool negative_length_guard_generated = false;
4875
4876 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4877 can_emit_guards &&
4878 !src->is_top() && !dest->is_top()) {
4879 // validate arguments: enables transformation the ArrayCopyNode
4880 validated = true;
4881
4882 RegionNode* slow_region = new RegionNode(1);
4883 record_for_igvn(slow_region);
4884
4885 // (1) src and dest are arrays.
4886 generate_non_array_guard(load_object_klass(src), slow_region);
4887 generate_non_array_guard(load_object_klass(dest), slow_region);
4888
4889 // (2) src and dest arrays must have elements of the same BasicType
4890 // done at macro expansion or at Ideal transformation time
4891
4892 // (4) src_offset must not be negative.
4893 generate_negative_guard(src_offset, slow_region);
4894
4895 // (5) dest_offset must not be negative.
4896 generate_negative_guard(dest_offset, slow_region);
4897
4898 // (7) src_offset + length must not exceed length of src.
4899 generate_limit_guard(src_offset, length,
4900 load_array_length(src),
4901 slow_region);
4902
4903 // (8) dest_offset + length must not exceed length of dest.
4904 generate_limit_guard(dest_offset, length,
4905 load_array_length(dest),
4906 slow_region);
4907
4908 // (6) length must not be negative.
4909 // This is also checked in generate_arraycopy() during macro expansion, but
4910 // we also have to check it here for the case where the ArrayCopyNode will
4911 // be eliminated by Escape Analysis.
4912 if (EliminateAllocations) {
4913 generate_negative_guard(length, slow_region);
4914 negative_length_guard_generated = true;
4915 }
4916
4917 // (9) each element of an oop array must be assignable
4918 Node* src_klass = load_object_klass(src);
4919 Node* dest_klass = load_object_klass(dest);
4920 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4921
4922 if (not_subtype_ctrl != top()) {
4923 PreserveJVMState pjvms(this);
4924 set_control(not_subtype_ctrl);
4925 uncommon_trap(Deoptimization::Reason_intrinsic,
4926 Deoptimization::Action_make_not_entrant);
4927 assert(stopped(), "Should be stopped");
4928 }
4929
4930 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
4931 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
4932 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
4933
4934 src_type = _gvn.type(src);
4935 top_src = src_type->isa_aryptr();
4936
4937 if (top_dest != NULL &&
4938 top_dest->elem()->make_oopptr() != NULL &&
4939 top_dest->elem()->make_oopptr()->can_be_value_type()) {
4940 generate_valueArray_guard(load_object_klass(dest), slow_region);
4941 }
4942
4943 if (top_src != NULL &&
4944 top_src->elem()->make_oopptr() != NULL &&
4945 top_src->elem()->make_oopptr()->can_be_value_type()) {
4946 generate_valueArray_guard(load_object_klass(src), slow_region);
4947 }
4948
4949 {
4950 PreserveJVMState pjvms(this);
4951 set_control(_gvn.transform(slow_region));
4952 uncommon_trap(Deoptimization::Reason_intrinsic,
4953 Deoptimization::Action_make_not_entrant);
4954 assert(stopped(), "Should be stopped");
4955 }
4956 }
4957
4958 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
4959
4960 if (stopped()) {
4961 return true;
4962 }
4963
4964 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
4965 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4966 // so the compiler has a chance to eliminate them: during macro expansion,
4967 // we have to set their control (CastPP nodes are eliminated).
4968 load_object_klass(src), load_object_klass(dest),
4969 load_array_length(src), load_array_length(dest));
4970
4971 ac->set_arraycopy(validated);
4972
4973 Node* n = _gvn.transform(ac);
4974 if (n == ac) {
4975 ac->connect_outputs(this);
4976 } else {
4977 assert(validated, "shouldn't transform if all arguments not validated");
4978 set_all_memory(n);
4979 }
4980 clear_upper_avx();
4981
4982
4983 return true;
4984 }
4985
4986
4987 // Helper function which determines if an arraycopy immediately follows
4988 // an allocation, with no intervening tests or other escapes for the object.
4989 AllocateArrayNode*
4990 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4991 RegionNode* slow_region) {
4992 if (stopped()) return NULL; // no fast path
4993 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4994
4995 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4996 if (alloc == NULL) return NULL;
4997
4998 Node* rawmem = memory(Compile::AliasIdxRaw);
4999 // Is the allocation's memory state untouched?
5000 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5001 // Bail out if there have been raw-memory effects since the allocation.
5002 // (Example: There might have been a call or safepoint.)
5003 return NULL;
5004 }
5005 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5006 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5007 return NULL;
5008 }
5009
5010 // There must be no unexpected observers of this allocation.
5011 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5012 Node* obs = ptr->fast_out(i);
5013 if (obs != this->map()) {
5014 return NULL;
5015 }
5016 }
5017
5018 // This arraycopy must unconditionally follow the allocation of the ptr.
5019 Node* alloc_ctl = ptr->in(0);
5020 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5021
5022 Node* ctl = control();
5023 while (ctl != alloc_ctl) {
5024 // There may be guards which feed into the slow_region.
5025 // Any other control flow means that we might not get a chance
5026 // to finish initializing the allocated object.
5027 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5028 IfNode* iff = ctl->in(0)->as_If();
5029 Node* not_ctl = iff->proj_out_or_null(1 - ctl->as_Proj()->_con);
5030 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5031 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5032 ctl = iff->in(0); // This test feeds the known slow_region.
5033 continue;
5034 }
5035 // One more try: Various low-level checks bottom out in
5036 // uncommon traps. If the debug-info of the trap omits
5037 // any reference to the allocation, as we've already
5038 // observed, then there can be no objection to the trap.
5039 bool found_trap = false;
5040 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5041 Node* obs = not_ctl->fast_out(j);
5042 if (obs->in(0) == not_ctl && obs->is_Call() &&
5043 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5044 found_trap = true; break;
5045 }
5046 }
5047 if (found_trap) {
5048 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5049 continue;
5050 }
5051 }
5052 return NULL;
5053 }
5054
5055 // If we get this far, we have an allocation which immediately
5056 // precedes the arraycopy, and we can take over zeroing the new object.
5057 // The arraycopy will finish the initialization, and provide
5058 // a new control state to which we will anchor the destination pointer.
5059
5060 return alloc;
5061 }
5062
5063 //-------------inline_encodeISOArray-----------------------------------
5064 // encode char[] to byte[] in ISO_8859_1
5065 bool LibraryCallKit::inline_encodeISOArray() {
5066 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5067 // no receiver since it is static method
5068 Node *src = argument(0);
5069 Node *src_offset = argument(1);
5070 Node *dst = argument(2);
5071 Node *dst_offset = argument(3);
5072 Node *length = argument(4);
5073
5074 src = must_be_not_null(src, true);
5075 dst = must_be_not_null(dst, true);
5076
5077 const Type* src_type = src->Value(&_gvn);
5078 const Type* dst_type = dst->Value(&_gvn);
5079 const TypeAryPtr* top_src = src_type->isa_aryptr();
5080 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5081 if (top_src == NULL || top_src->klass() == NULL ||
5082 top_dest == NULL || top_dest->klass() == NULL) {
5083 // failed array check
5084 return false;
5085 }
5086
5087 // Figure out the size and type of the elements we will be copying.
5088 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5089 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5090 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5091 return false;
5092 }
5093
5094 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5095 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5096 // 'src_start' points to src array + scaled offset
5097 // 'dst_start' points to dst array + scaled offset
5098
5099 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5100 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5101 enc = _gvn.transform(enc);
5102 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5103 set_memory(res_mem, mtype);
5104 set_result(enc);
5105 clear_upper_avx();
5106
5107 return true;
5108 }
5109
5110 //-------------inline_multiplyToLen-----------------------------------
5111 bool LibraryCallKit::inline_multiplyToLen() {
5112 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5113
5114 address stubAddr = StubRoutines::multiplyToLen();
5115 if (stubAddr == NULL) {
5116 return false; // Intrinsic's stub is not implemented on this platform
5117 }
5118 const char* stubName = "multiplyToLen";
5119
5120 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5121
5122 // no receiver because it is a static method
5123 Node* x = argument(0);
5124 Node* xlen = argument(1);
5125 Node* y = argument(2);
5126 Node* ylen = argument(3);
5127 Node* z = argument(4);
5128
5129 x = must_be_not_null(x, true);
5130 y = must_be_not_null(y, true);
5131
5132 const Type* x_type = x->Value(&_gvn);
5133 const Type* y_type = y->Value(&_gvn);
5134 const TypeAryPtr* top_x = x_type->isa_aryptr();
5135 const TypeAryPtr* top_y = y_type->isa_aryptr();
5136 if (top_x == NULL || top_x->klass() == NULL ||
5137 top_y == NULL || top_y->klass() == NULL) {
5138 // failed array check
5139 return false;
5140 }
5141
5142 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5143 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5144 if (x_elem != T_INT || y_elem != T_INT) {
5145 return false;
5146 }
5147
5148 // Set the original stack and the reexecute bit for the interpreter to reexecute
5149 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5150 // on the return from z array allocation in runtime.
5151 { PreserveReexecuteState preexecs(this);
5152 jvms()->set_should_reexecute(true);
5153
5154 Node* x_start = array_element_address(x, intcon(0), x_elem);
5155 Node* y_start = array_element_address(y, intcon(0), y_elem);
5156 // 'x_start' points to x array + scaled xlen
5157 // 'y_start' points to y array + scaled ylen
5158
5159 // Allocate the result array
5160 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5161 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5162 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5163
5164 IdealKit ideal(this);
5165
5166 #define __ ideal.
5167 Node* one = __ ConI(1);
5168 Node* zero = __ ConI(0);
5169 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5170 __ set(need_alloc, zero);
5171 __ set(z_alloc, z);
5172 __ if_then(z, BoolTest::eq, null()); {
5173 __ increment (need_alloc, one);
5174 } __ else_(); {
5175 // Update graphKit memory and control from IdealKit.
5176 sync_kit(ideal);
5177 Node *cast = new CastPPNode(z, TypePtr::NOTNULL);
5178 cast->init_req(0, control());
5179 _gvn.set_type(cast, cast->bottom_type());
5180 C->record_for_igvn(cast);
5181
5182 Node* zlen_arg = load_array_length(cast);
5183 // Update IdealKit memory and control from graphKit.
5184 __ sync_kit(this);
5185 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5186 __ increment (need_alloc, one);
5187 } __ end_if();
5188 } __ end_if();
5189
5190 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5191 // Update graphKit memory and control from IdealKit.
5192 sync_kit(ideal);
5193 Node * narr = new_array(klass_node, zlen, 1);
5194 // Update IdealKit memory and control from graphKit.
5195 __ sync_kit(this);
5196 __ set(z_alloc, narr);
5197 } __ end_if();
5198
5199 sync_kit(ideal);
5200 z = __ value(z_alloc);
5201 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5202 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5203 // Final sync IdealKit and GraphKit.
5204 final_sync(ideal);
5205 #undef __
5206
5207 Node* z_start = array_element_address(z, intcon(0), T_INT);
5208
5209 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5210 OptoRuntime::multiplyToLen_Type(),
5211 stubAddr, stubName, TypePtr::BOTTOM,
5212 x_start, xlen, y_start, ylen, z_start, zlen);
5213 } // original reexecute is set back here
5214
5215 C->set_has_split_ifs(true); // Has chance for split-if optimization
5216 set_result(z);
5217 return true;
5218 }
5219
5220 //-------------inline_squareToLen------------------------------------
5221 bool LibraryCallKit::inline_squareToLen() {
5222 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5223
5224 address stubAddr = StubRoutines::squareToLen();
5225 if (stubAddr == NULL) {
5226 return false; // Intrinsic's stub is not implemented on this platform
5227 }
5228 const char* stubName = "squareToLen";
5229
5230 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5231
5232 Node* x = argument(0);
5233 Node* len = argument(1);
5234 Node* z = argument(2);
5235 Node* zlen = argument(3);
5236
5237 x = must_be_not_null(x, true);
5238 z = must_be_not_null(z, true);
5239
5240 const Type* x_type = x->Value(&_gvn);
5241 const Type* z_type = z->Value(&_gvn);
5242 const TypeAryPtr* top_x = x_type->isa_aryptr();
5243 const TypeAryPtr* top_z = z_type->isa_aryptr();
5244 if (top_x == NULL || top_x->klass() == NULL ||
5245 top_z == NULL || top_z->klass() == NULL) {
5246 // failed array check
5247 return false;
5248 }
5249
5250 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5251 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5252 if (x_elem != T_INT || z_elem != T_INT) {
5253 return false;
5254 }
5255
5256
5257 Node* x_start = array_element_address(x, intcon(0), x_elem);
5258 Node* z_start = array_element_address(z, intcon(0), z_elem);
5259
5260 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5261 OptoRuntime::squareToLen_Type(),
5262 stubAddr, stubName, TypePtr::BOTTOM,
5263 x_start, len, z_start, zlen);
5264
5265 set_result(z);
5266 return true;
5267 }
5268
5269 //-------------inline_mulAdd------------------------------------------
5270 bool LibraryCallKit::inline_mulAdd() {
5271 assert(UseMulAddIntrinsic, "not implemented on this platform");
5272
5273 address stubAddr = StubRoutines::mulAdd();
5274 if (stubAddr == NULL) {
5275 return false; // Intrinsic's stub is not implemented on this platform
5276 }
5277 const char* stubName = "mulAdd";
5278
5279 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5280
5281 Node* out = argument(0);
5282 Node* in = argument(1);
5283 Node* offset = argument(2);
5284 Node* len = argument(3);
5285 Node* k = argument(4);
5286
5287 out = must_be_not_null(out, true);
5288
5289 const Type* out_type = out->Value(&_gvn);
5290 const Type* in_type = in->Value(&_gvn);
5291 const TypeAryPtr* top_out = out_type->isa_aryptr();
5292 const TypeAryPtr* top_in = in_type->isa_aryptr();
5293 if (top_out == NULL || top_out->klass() == NULL ||
5294 top_in == NULL || top_in->klass() == NULL) {
5295 // failed array check
5296 return false;
5297 }
5298
5299 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5300 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5301 if (out_elem != T_INT || in_elem != T_INT) {
5302 return false;
5303 }
5304
5305 Node* outlen = load_array_length(out);
5306 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5307 Node* out_start = array_element_address(out, intcon(0), out_elem);
5308 Node* in_start = array_element_address(in, intcon(0), in_elem);
5309
5310 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5311 OptoRuntime::mulAdd_Type(),
5312 stubAddr, stubName, TypePtr::BOTTOM,
5313 out_start,in_start, new_offset, len, k);
5314 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5315 set_result(result);
5316 return true;
5317 }
5318
5319 //-------------inline_montgomeryMultiply-----------------------------------
5320 bool LibraryCallKit::inline_montgomeryMultiply() {
5321 address stubAddr = StubRoutines::montgomeryMultiply();
5322 if (stubAddr == NULL) {
5323 return false; // Intrinsic's stub is not implemented on this platform
5324 }
5325
5326 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5327 const char* stubName = "montgomery_multiply";
5328
5329 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5330
5331 Node* a = argument(0);
5332 Node* b = argument(1);
5333 Node* n = argument(2);
5334 Node* len = argument(3);
5335 Node* inv = argument(4);
5336 Node* m = argument(6);
5337
5338 const Type* a_type = a->Value(&_gvn);
5339 const TypeAryPtr* top_a = a_type->isa_aryptr();
5340 const Type* b_type = b->Value(&_gvn);
5341 const TypeAryPtr* top_b = b_type->isa_aryptr();
5342 const Type* n_type = a->Value(&_gvn);
5343 const TypeAryPtr* top_n = n_type->isa_aryptr();
5344 const Type* m_type = a->Value(&_gvn);
5345 const TypeAryPtr* top_m = m_type->isa_aryptr();
5346 if (top_a == NULL || top_a->klass() == NULL ||
5347 top_b == NULL || top_b->klass() == NULL ||
5348 top_n == NULL || top_n->klass() == NULL ||
5349 top_m == NULL || top_m->klass() == NULL) {
5350 // failed array check
5351 return false;
5352 }
5353
5354 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5355 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5356 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5357 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5358 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5359 return false;
5360 }
5361
5362 // Make the call
5363 {
5364 Node* a_start = array_element_address(a, intcon(0), a_elem);
5365 Node* b_start = array_element_address(b, intcon(0), b_elem);
5366 Node* n_start = array_element_address(n, intcon(0), n_elem);
5367 Node* m_start = array_element_address(m, intcon(0), m_elem);
5368
5369 Node* call = make_runtime_call(RC_LEAF,
5370 OptoRuntime::montgomeryMultiply_Type(),
5371 stubAddr, stubName, TypePtr::BOTTOM,
5372 a_start, b_start, n_start, len, inv, top(),
5373 m_start);
5374 set_result(m);
5375 }
5376
5377 return true;
5378 }
5379
5380 bool LibraryCallKit::inline_montgomerySquare() {
5381 address stubAddr = StubRoutines::montgomerySquare();
5382 if (stubAddr == NULL) {
5383 return false; // Intrinsic's stub is not implemented on this platform
5384 }
5385
5386 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5387 const char* stubName = "montgomery_square";
5388
5389 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5390
5391 Node* a = argument(0);
5392 Node* n = argument(1);
5393 Node* len = argument(2);
5394 Node* inv = argument(3);
5395 Node* m = argument(5);
5396
5397 const Type* a_type = a->Value(&_gvn);
5398 const TypeAryPtr* top_a = a_type->isa_aryptr();
5399 const Type* n_type = a->Value(&_gvn);
5400 const TypeAryPtr* top_n = n_type->isa_aryptr();
5401 const Type* m_type = a->Value(&_gvn);
5402 const TypeAryPtr* top_m = m_type->isa_aryptr();
5403 if (top_a == NULL || top_a->klass() == NULL ||
5404 top_n == NULL || top_n->klass() == NULL ||
5405 top_m == NULL || top_m->klass() == NULL) {
5406 // failed array check
5407 return false;
5408 }
5409
5410 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5411 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5412 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5413 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5414 return false;
5415 }
5416
5417 // Make the call
5418 {
5419 Node* a_start = array_element_address(a, intcon(0), a_elem);
5420 Node* n_start = array_element_address(n, intcon(0), n_elem);
5421 Node* m_start = array_element_address(m, intcon(0), m_elem);
5422
5423 Node* call = make_runtime_call(RC_LEAF,
5424 OptoRuntime::montgomerySquare_Type(),
5425 stubAddr, stubName, TypePtr::BOTTOM,
5426 a_start, n_start, len, inv, top(),
5427 m_start);
5428 set_result(m);
5429 }
5430
5431 return true;
5432 }
5433
5434 //-------------inline_vectorizedMismatch------------------------------
5435 bool LibraryCallKit::inline_vectorizedMismatch() {
5436 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5437
5438 address stubAddr = StubRoutines::vectorizedMismatch();
5439 if (stubAddr == NULL) {
5440 return false; // Intrinsic's stub is not implemented on this platform
5441 }
5442 const char* stubName = "vectorizedMismatch";
5443 int size_l = callee()->signature()->size();
5444 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5445
5446 Node* obja = argument(0);
5447 Node* aoffset = argument(1);
5448 Node* objb = argument(3);
5449 Node* boffset = argument(4);
5450 Node* length = argument(6);
5451 Node* scale = argument(7);
5452
5453 const Type* a_type = obja->Value(&_gvn);
5454 const Type* b_type = objb->Value(&_gvn);
5455 const TypeAryPtr* top_a = a_type->isa_aryptr();
5456 const TypeAryPtr* top_b = b_type->isa_aryptr();
5457 if (top_a == NULL || top_a->klass() == NULL ||
5458 top_b == NULL || top_b->klass() == NULL) {
5459 // failed array check
5460 return false;
5461 }
5462
5463 Node* call;
5464 jvms()->set_should_reexecute(true);
5465
5466 Node* obja_adr = make_unsafe_address(obja, aoffset);
5467 Node* objb_adr = make_unsafe_address(objb, boffset);
5468
5469 call = make_runtime_call(RC_LEAF,
5470 OptoRuntime::vectorizedMismatch_Type(),
5471 stubAddr, stubName, TypePtr::BOTTOM,
5472 obja_adr, objb_adr, length, scale);
5473
5474 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5475 set_result(result);
5476 return true;
5477 }
5478
5479 /**
5480 * Calculate CRC32 for byte.
5481 * int java.util.zip.CRC32.update(int crc, int b)
5482 */
5483 bool LibraryCallKit::inline_updateCRC32() {
5484 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5485 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5486 // no receiver since it is static method
5487 Node* crc = argument(0); // type: int
5488 Node* b = argument(1); // type: int
5489
5490 /*
5491 * int c = ~ crc;
5492 * b = timesXtoThe32[(b ^ c) & 0xFF];
5493 * b = b ^ (c >>> 8);
5494 * crc = ~b;
5495 */
5496
5497 Node* M1 = intcon(-1);
5498 crc = _gvn.transform(new XorINode(crc, M1));
5499 Node* result = _gvn.transform(new XorINode(crc, b));
5500 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5501
5502 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5503 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5504 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5505 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5506
5507 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5508 result = _gvn.transform(new XorINode(crc, result));
5509 result = _gvn.transform(new XorINode(result, M1));
5510 set_result(result);
5511 return true;
5512 }
5513
5514 /**
5515 * Calculate CRC32 for byte[] array.
5516 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5517 */
5518 bool LibraryCallKit::inline_updateBytesCRC32() {
5519 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5520 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5521 // no receiver since it is static method
5522 Node* crc = argument(0); // type: int
5523 Node* src = argument(1); // type: oop
5524 Node* offset = argument(2); // type: int
5525 Node* length = argument(3); // type: int
5526
5527 const Type* src_type = src->Value(&_gvn);
5528 const TypeAryPtr* top_src = src_type->isa_aryptr();
5529 if (top_src == NULL || top_src->klass() == NULL) {
5530 // failed array check
5531 return false;
5532 }
5533
5534 // Figure out the size and type of the elements we will be copying.
5535 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5536 if (src_elem != T_BYTE) {
5537 return false;
5538 }
5539
5540 // 'src_start' points to src array + scaled offset
5541 src = must_be_not_null(src, true);
5542 Node* src_start = array_element_address(src, offset, src_elem);
5543
5544 // We assume that range check is done by caller.
5545 // TODO: generate range check (offset+length < src.length) in debug VM.
5546
5547 // Call the stub.
5548 address stubAddr = StubRoutines::updateBytesCRC32();
5549 const char *stubName = "updateBytesCRC32";
5550
5551 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5552 stubAddr, stubName, TypePtr::BOTTOM,
5553 crc, src_start, length);
5554 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5555 set_result(result);
5556 return true;
5557 }
5558
5559 /**
5560 * Calculate CRC32 for ByteBuffer.
5561 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5562 */
5563 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5564 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5565 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5566 // no receiver since it is static method
5567 Node* crc = argument(0); // type: int
5568 Node* src = argument(1); // type: long
5569 Node* offset = argument(3); // type: int
5570 Node* length = argument(4); // type: int
5571
5572 src = ConvL2X(src); // adjust Java long to machine word
5573 Node* base = _gvn.transform(new CastX2PNode(src));
5574 offset = ConvI2X(offset);
5575
5576 // 'src_start' points to src array + scaled offset
5577 Node* src_start = basic_plus_adr(top(), base, offset);
5578
5579 // Call the stub.
5580 address stubAddr = StubRoutines::updateBytesCRC32();
5581 const char *stubName = "updateBytesCRC32";
5582
5583 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5584 stubAddr, stubName, TypePtr::BOTTOM,
5585 crc, src_start, length);
5586 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5587 set_result(result);
5588 return true;
5589 }
5590
5591 //------------------------------get_table_from_crc32c_class-----------------------
5592 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5593 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5594 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5595
5596 return table;
5597 }
5598
5599 //------------------------------inline_updateBytesCRC32C-----------------------
5600 //
5601 // Calculate CRC32C for byte[] array.
5602 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5603 //
5604 bool LibraryCallKit::inline_updateBytesCRC32C() {
5605 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5606 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5607 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5608 // no receiver since it is a static method
5609 Node* crc = argument(0); // type: int
5610 Node* src = argument(1); // type: oop
5611 Node* offset = argument(2); // type: int
5612 Node* end = argument(3); // type: int
5613
5614 Node* length = _gvn.transform(new SubINode(end, offset));
5615
5616 const Type* src_type = src->Value(&_gvn);
5617 const TypeAryPtr* top_src = src_type->isa_aryptr();
5618 if (top_src == NULL || top_src->klass() == NULL) {
5619 // failed array check
5620 return false;
5621 }
5622
5623 // Figure out the size and type of the elements we will be copying.
5624 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5625 if (src_elem != T_BYTE) {
5626 return false;
5627 }
5628
5629 // 'src_start' points to src array + scaled offset
5630 src = must_be_not_null(src, true);
5631 Node* src_start = array_element_address(src, offset, src_elem);
5632
5633 // static final int[] byteTable in class CRC32C
5634 Node* table = get_table_from_crc32c_class(callee()->holder());
5635 table = must_be_not_null(table, true);
5636 Node* table_start = array_element_address(table, intcon(0), T_INT);
5637
5638 // We assume that range check is done by caller.
5639 // TODO: generate range check (offset+length < src.length) in debug VM.
5640
5641 // Call the stub.
5642 address stubAddr = StubRoutines::updateBytesCRC32C();
5643 const char *stubName = "updateBytesCRC32C";
5644
5645 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5646 stubAddr, stubName, TypePtr::BOTTOM,
5647 crc, src_start, length, table_start);
5648 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5649 set_result(result);
5650 return true;
5651 }
5652
5653 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5654 //
5655 // Calculate CRC32C for DirectByteBuffer.
5656 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5657 //
5658 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5659 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5660 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5661 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5662 // no receiver since it is a static method
5663 Node* crc = argument(0); // type: int
5664 Node* src = argument(1); // type: long
5665 Node* offset = argument(3); // type: int
5666 Node* end = argument(4); // type: int
5667
5668 Node* length = _gvn.transform(new SubINode(end, offset));
5669
5670 src = ConvL2X(src); // adjust Java long to machine word
5671 Node* base = _gvn.transform(new CastX2PNode(src));
5672 offset = ConvI2X(offset);
5673
5674 // 'src_start' points to src array + scaled offset
5675 Node* src_start = basic_plus_adr(top(), base, offset);
5676
5677 // static final int[] byteTable in class CRC32C
5678 Node* table = get_table_from_crc32c_class(callee()->holder());
5679 table = must_be_not_null(table, true);
5680 Node* table_start = array_element_address(table, intcon(0), T_INT);
5681
5682 // Call the stub.
5683 address stubAddr = StubRoutines::updateBytesCRC32C();
5684 const char *stubName = "updateBytesCRC32C";
5685
5686 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5687 stubAddr, stubName, TypePtr::BOTTOM,
5688 crc, src_start, length, table_start);
5689 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5690 set_result(result);
5691 return true;
5692 }
5693
5694 //------------------------------inline_updateBytesAdler32----------------------
5695 //
5696 // Calculate Adler32 checksum for byte[] array.
5697 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5698 //
5699 bool LibraryCallKit::inline_updateBytesAdler32() {
5700 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5701 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5702 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5703 // no receiver since it is static method
5704 Node* crc = argument(0); // type: int
5705 Node* src = argument(1); // type: oop
5706 Node* offset = argument(2); // type: int
5707 Node* length = argument(3); // type: int
5708
5709 const Type* src_type = src->Value(&_gvn);
5710 const TypeAryPtr* top_src = src_type->isa_aryptr();
5711 if (top_src == NULL || top_src->klass() == NULL) {
5712 // failed array check
5713 return false;
5714 }
5715
5716 // Figure out the size and type of the elements we will be copying.
5717 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5718 if (src_elem != T_BYTE) {
5719 return false;
5720 }
5721
5722 // 'src_start' points to src array + scaled offset
5723 Node* src_start = array_element_address(src, offset, src_elem);
5724
5725 // We assume that range check is done by caller.
5726 // TODO: generate range check (offset+length < src.length) in debug VM.
5727
5728 // Call the stub.
5729 address stubAddr = StubRoutines::updateBytesAdler32();
5730 const char *stubName = "updateBytesAdler32";
5731
5732 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5733 stubAddr, stubName, TypePtr::BOTTOM,
5734 crc, src_start, length);
5735 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5736 set_result(result);
5737 return true;
5738 }
5739
5740 //------------------------------inline_updateByteBufferAdler32---------------
5741 //
5742 // Calculate Adler32 checksum for DirectByteBuffer.
5743 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5744 //
5745 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5746 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5747 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5748 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5749 // no receiver since it is static method
5750 Node* crc = argument(0); // type: int
5751 Node* src = argument(1); // type: long
5752 Node* offset = argument(3); // type: int
5753 Node* length = argument(4); // type: int
5754
5755 src = ConvL2X(src); // adjust Java long to machine word
5756 Node* base = _gvn.transform(new CastX2PNode(src));
5757 offset = ConvI2X(offset);
5758
5759 // 'src_start' points to src array + scaled offset
5760 Node* src_start = basic_plus_adr(top(), base, offset);
5761
5762 // Call the stub.
5763 address stubAddr = StubRoutines::updateBytesAdler32();
5764 const char *stubName = "updateBytesAdler32";
5765
5766 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5767 stubAddr, stubName, TypePtr::BOTTOM,
5768 crc, src_start, length);
5769
5770 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5771 set_result(result);
5772 return true;
5773 }
5774
5775 //----------------------------inline_reference_get----------------------------
5776 // public T java.lang.ref.Reference.get();
5777 bool LibraryCallKit::inline_reference_get() {
5778 const int referent_offset = java_lang_ref_Reference::referent_offset;
5779 guarantee(referent_offset > 0, "should have already been set");
5780
5781 // Get the argument:
5782 Node* reference_obj = null_check_receiver();
5783 if (stopped()) return true;
5784
5785 const TypeInstPtr* tinst = _gvn.type(reference_obj)->isa_instptr();
5786 assert(tinst != NULL, "obj is null");
5787 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5788 ciInstanceKlass* referenceKlass = tinst->klass()->as_instance_klass();
5789 ciField* field = referenceKlass->get_field_by_name(ciSymbol::make("referent"),
5790 ciSymbol::make("Ljava/lang/Object;"),
5791 false);
5792 assert (field != NULL, "undefined field");
5793
5794 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5795 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5796
5797 ciInstanceKlass* klass = env()->Object_klass();
5798 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5799
5800 DecoratorSet decorators = IN_HEAP | ON_WEAK_OOP_REF;
5801 Node* result = access_load_at(reference_obj, adr, adr_type, object_type, T_OBJECT, decorators);
5802 // Add memory barrier to prevent commoning reads from this field
5803 // across safepoint since GC can change its value.
5804 insert_mem_bar(Op_MemBarCPUOrder);
5805
5806 set_result(result);
5807 return true;
5808 }
5809
5810
5811 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5812 bool is_exact=true, bool is_static=false,
5813 ciInstanceKlass * fromKls=NULL) {
5814 if (fromKls == NULL) {
5815 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5816 assert(tinst != NULL, "obj is null");
5817 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5818 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5819 fromKls = tinst->klass()->as_instance_klass();
5820 } else {
5821 assert(is_static, "only for static field access");
5822 }
5823 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5824 ciSymbol::make(fieldTypeString),
5825 is_static);
5826
5827 assert (field != NULL, "undefined field");
5828 if (field == NULL) return (Node *) NULL;
5829
5830 if (is_static) {
5831 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5832 fromObj = makecon(tip);
5833 }
5834
5835 // Next code copied from Parse::do_get_xxx():
5836
5837 // Compute address and memory type.
5838 int offset = field->offset_in_bytes();
5839 bool is_vol = field->is_volatile();
5840 ciType* field_klass = field->type();
5841 assert(field_klass->is_loaded(), "should be loaded");
5842 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5843 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5844 BasicType bt = field->layout_type();
5845
5846 // Build the resultant type of the load
5847 const Type *type;
5848 if (bt == T_OBJECT) {
5849 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5850 } else {
5851 type = Type::get_const_basic_type(bt);
5852 }
5853
5854 DecoratorSet decorators = IN_HEAP;
5855
5856 if (is_vol) {
5857 decorators |= MO_SEQ_CST;
5858 }
5859
5860 return access_load_at(fromObj, adr, adr_type, type, bt, decorators);
5861 }
5862
5863 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5864 bool is_exact = true, bool is_static = false,
5865 ciInstanceKlass * fromKls = NULL) {
5866 if (fromKls == NULL) {
5867 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5868 assert(tinst != NULL, "obj is null");
5869 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5870 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5871 fromKls = tinst->klass()->as_instance_klass();
5872 }
5873 else {
5874 assert(is_static, "only for static field access");
5875 }
5876 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5877 ciSymbol::make(fieldTypeString),
5878 is_static);
5879
5880 assert(field != NULL, "undefined field");
5881 assert(!field->is_volatile(), "not defined for volatile fields");
5882
5883 if (is_static) {
5884 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5885 fromObj = makecon(tip);
5886 }
5887
5888 // Next code copied from Parse::do_get_xxx():
5889
5890 // Compute address and memory type.
5891 int offset = field->offset_in_bytes();
5892 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5893
5894 return adr;
5895 }
5896
5897 //------------------------------inline_aescrypt_Block-----------------------
5898 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5899 address stubAddr = NULL;
5900 const char *stubName;
5901 assert(UseAES, "need AES instruction support");
5902
5903 switch(id) {
5904 case vmIntrinsics::_aescrypt_encryptBlock:
5905 stubAddr = StubRoutines::aescrypt_encryptBlock();
5906 stubName = "aescrypt_encryptBlock";
5907 break;
5908 case vmIntrinsics::_aescrypt_decryptBlock:
5909 stubAddr = StubRoutines::aescrypt_decryptBlock();
5910 stubName = "aescrypt_decryptBlock";
5911 break;
5912 default:
5913 break;
5914 }
5915 if (stubAddr == NULL) return false;
5916
5917 Node* aescrypt_object = argument(0);
5918 Node* src = argument(1);
5919 Node* src_offset = argument(2);
5920 Node* dest = argument(3);
5921 Node* dest_offset = argument(4);
5922
5923 src = must_be_not_null(src, true);
5924 dest = must_be_not_null(dest, true);
5925
5926 // (1) src and dest are arrays.
5927 const Type* src_type = src->Value(&_gvn);
5928 const Type* dest_type = dest->Value(&_gvn);
5929 const TypeAryPtr* top_src = src_type->isa_aryptr();
5930 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5931 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5932
5933 // for the quick and dirty code we will skip all the checks.
5934 // we are just trying to get the call to be generated.
5935 Node* src_start = src;
5936 Node* dest_start = dest;
5937 if (src_offset != NULL || dest_offset != NULL) {
5938 assert(src_offset != NULL && dest_offset != NULL, "");
5939 src_start = array_element_address(src, src_offset, T_BYTE);
5940 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5941 }
5942
5943 // now need to get the start of its expanded key array
5944 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5945 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5946 if (k_start == NULL) return false;
5947
5948 if (Matcher::pass_original_key_for_aes()) {
5949 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5950 // compatibility issues between Java key expansion and SPARC crypto instructions
5951 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5952 if (original_k_start == NULL) return false;
5953
5954 // Call the stub.
5955 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5956 stubAddr, stubName, TypePtr::BOTTOM,
5957 src_start, dest_start, k_start, original_k_start);
5958 } else {
5959 // Call the stub.
5960 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5961 stubAddr, stubName, TypePtr::BOTTOM,
5962 src_start, dest_start, k_start);
5963 }
5964
5965 return true;
5966 }
5967
5968 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5969 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5970 address stubAddr = NULL;
5971 const char *stubName = NULL;
5972
5973 assert(UseAES, "need AES instruction support");
5974
5975 switch(id) {
5976 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5977 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5978 stubName = "cipherBlockChaining_encryptAESCrypt";
5979 break;
5980 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5981 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5982 stubName = "cipherBlockChaining_decryptAESCrypt";
5983 break;
5984 default:
5985 break;
5986 }
5987 if (stubAddr == NULL) return false;
5988
5989 Node* cipherBlockChaining_object = argument(0);
5990 Node* src = argument(1);
5991 Node* src_offset = argument(2);
5992 Node* len = argument(3);
5993 Node* dest = argument(4);
5994 Node* dest_offset = argument(5);
5995
5996 src = must_be_not_null(src, false);
5997 dest = must_be_not_null(dest, false);
5998
5999 // (1) src and dest are arrays.
6000 const Type* src_type = src->Value(&_gvn);
6001 const Type* dest_type = dest->Value(&_gvn);
6002 const TypeAryPtr* top_src = src_type->isa_aryptr();
6003 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6004 assert (top_src != NULL && top_src->klass() != NULL
6005 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6006
6007 // checks are the responsibility of the caller
6008 Node* src_start = src;
6009 Node* dest_start = dest;
6010 if (src_offset != NULL || dest_offset != NULL) {
6011 assert(src_offset != NULL && dest_offset != NULL, "");
6012 src_start = array_element_address(src, src_offset, T_BYTE);
6013 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6014 }
6015
6016 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6017 // (because of the predicated logic executed earlier).
6018 // so we cast it here safely.
6019 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6020
6021 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6022 if (embeddedCipherObj == NULL) return false;
6023
6024 // cast it to what we know it will be at runtime
6025 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6026 assert(tinst != NULL, "CBC obj is null");
6027 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6028 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6029 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6030
6031 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6032 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6033 const TypeOopPtr* xtype = aklass->as_instance_type();
6034 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6035 aescrypt_object = _gvn.transform(aescrypt_object);
6036
6037 // we need to get the start of the aescrypt_object's expanded key array
6038 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6039 if (k_start == NULL) return false;
6040
6041 // similarly, get the start address of the r vector
6042 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6043 if (objRvec == NULL) return false;
6044 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6045
6046 Node* cbcCrypt;
6047 if (Matcher::pass_original_key_for_aes()) {
6048 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6049 // compatibility issues between Java key expansion and SPARC crypto instructions
6050 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6051 if (original_k_start == NULL) return false;
6052
6053 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6054 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6055 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6056 stubAddr, stubName, TypePtr::BOTTOM,
6057 src_start, dest_start, k_start, r_start, len, original_k_start);
6058 } else {
6059 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6060 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6061 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6062 stubAddr, stubName, TypePtr::BOTTOM,
6063 src_start, dest_start, k_start, r_start, len);
6064 }
6065
6066 // return cipher length (int)
6067 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6068 set_result(retvalue);
6069 return true;
6070 }
6071
6072 //------------------------------inline_counterMode_AESCrypt-----------------------
6073 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6074 assert(UseAES, "need AES instruction support");
6075 if (!UseAESCTRIntrinsics) return false;
6076
6077 address stubAddr = NULL;
6078 const char *stubName = NULL;
6079 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6080 stubAddr = StubRoutines::counterMode_AESCrypt();
6081 stubName = "counterMode_AESCrypt";
6082 }
6083 if (stubAddr == NULL) return false;
6084
6085 Node* counterMode_object = argument(0);
6086 Node* src = argument(1);
6087 Node* src_offset = argument(2);
6088 Node* len = argument(3);
6089 Node* dest = argument(4);
6090 Node* dest_offset = argument(5);
6091
6092 // (1) src and dest are arrays.
6093 const Type* src_type = src->Value(&_gvn);
6094 const Type* dest_type = dest->Value(&_gvn);
6095 const TypeAryPtr* top_src = src_type->isa_aryptr();
6096 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6097 assert(top_src != NULL && top_src->klass() != NULL &&
6098 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6099
6100 // checks are the responsibility of the caller
6101 Node* src_start = src;
6102 Node* dest_start = dest;
6103 if (src_offset != NULL || dest_offset != NULL) {
6104 assert(src_offset != NULL && dest_offset != NULL, "");
6105 src_start = array_element_address(src, src_offset, T_BYTE);
6106 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6107 }
6108
6109 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6110 // (because of the predicated logic executed earlier).
6111 // so we cast it here safely.
6112 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6113 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6114 if (embeddedCipherObj == NULL) return false;
6115 // cast it to what we know it will be at runtime
6116 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6117 assert(tinst != NULL, "CTR obj is null");
6118 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6119 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6120 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6121 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6122 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6123 const TypeOopPtr* xtype = aklass->as_instance_type();
6124 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6125 aescrypt_object = _gvn.transform(aescrypt_object);
6126 // we need to get the start of the aescrypt_object's expanded key array
6127 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6128 if (k_start == NULL) return false;
6129 // similarly, get the start address of the r vector
6130 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6131 if (obj_counter == NULL) return false;
6132 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6133
6134 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6135 if (saved_encCounter == NULL) return false;
6136 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6137 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6138
6139 Node* ctrCrypt;
6140 if (Matcher::pass_original_key_for_aes()) {
6141 // no SPARC version for AES/CTR intrinsics now.
6142 return false;
6143 }
6144 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6145 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6146 OptoRuntime::counterMode_aescrypt_Type(),
6147 stubAddr, stubName, TypePtr::BOTTOM,
6148 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6149
6150 // return cipher length (int)
6151 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6152 set_result(retvalue);
6153 return true;
6154 }
6155
6156 //------------------------------get_key_start_from_aescrypt_object-----------------------
6157 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6158 #if defined(PPC64) || defined(S390)
6159 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6160 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6161 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6162 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6163 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6164 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6165 if (objSessionK == NULL) {
6166 return (Node *) NULL;
6167 }
6168 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6169 #else
6170 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6171 #endif // PPC64
6172 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6173 if (objAESCryptKey == NULL) return (Node *) NULL;
6174
6175 // now have the array, need to get the start address of the K array
6176 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6177 return k_start;
6178 }
6179
6180 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6181 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6182 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6183 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6184 if (objAESCryptKey == NULL) return (Node *) NULL;
6185
6186 // now have the array, need to get the start address of the lastKey array
6187 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6188 return original_k_start;
6189 }
6190
6191 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6192 // Return node representing slow path of predicate check.
6193 // the pseudo code we want to emulate with this predicate is:
6194 // for encryption:
6195 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6196 // for decryption:
6197 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6198 // note cipher==plain is more conservative than the original java code but that's OK
6199 //
6200 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6201 // The receiver was checked for NULL already.
6202 Node* objCBC = argument(0);
6203
6204 Node* src = argument(1);
6205 Node* dest = argument(4);
6206
6207 // Load embeddedCipher field of CipherBlockChaining object.
6208 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6209
6210 // get AESCrypt klass for instanceOf check
6211 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6212 // will have same classloader as CipherBlockChaining object
6213 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6214 assert(tinst != NULL, "CBCobj is null");
6215 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6216
6217 // we want to do an instanceof comparison against the AESCrypt class
6218 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6219 if (!klass_AESCrypt->is_loaded()) {
6220 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6221 Node* ctrl = control();
6222 set_control(top()); // no regular fast path
6223 return ctrl;
6224 }
6225
6226 src = must_be_not_null(src, true);
6227 dest = must_be_not_null(dest, true);
6228
6229 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6230
6231 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6232 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6233 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6234
6235 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6236
6237 // for encryption, we are done
6238 if (!decrypting)
6239 return instof_false; // even if it is NULL
6240
6241 // for decryption, we need to add a further check to avoid
6242 // taking the intrinsic path when cipher and plain are the same
6243 // see the original java code for why.
6244 RegionNode* region = new RegionNode(3);
6245 region->init_req(1, instof_false);
6246
6247 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6248 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6249 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6250 region->init_req(2, src_dest_conjoint);
6251
6252 record_for_igvn(region);
6253 return _gvn.transform(region);
6254 }
6255
6256 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6257 // Return node representing slow path of predicate check.
6258 // the pseudo code we want to emulate with this predicate is:
6259 // for encryption:
6260 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6261 // for decryption:
6262 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6263 // note cipher==plain is more conservative than the original java code but that's OK
6264 //
6265
6266 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6267 // The receiver was checked for NULL already.
6268 Node* objCTR = argument(0);
6269
6270 // Load embeddedCipher field of CipherBlockChaining object.
6271 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6272
6273 // get AESCrypt klass for instanceOf check
6274 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6275 // will have same classloader as CipherBlockChaining object
6276 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6277 assert(tinst != NULL, "CTRobj is null");
6278 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6279
6280 // we want to do an instanceof comparison against the AESCrypt class
6281 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6282 if (!klass_AESCrypt->is_loaded()) {
6283 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6284 Node* ctrl = control();
6285 set_control(top()); // no regular fast path
6286 return ctrl;
6287 }
6288
6289 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6290 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6291 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6292 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6293 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6294
6295 return instof_false; // even if it is NULL
6296 }
6297
6298 //------------------------------inline_ghash_processBlocks
6299 bool LibraryCallKit::inline_ghash_processBlocks() {
6300 address stubAddr;
6301 const char *stubName;
6302 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6303
6304 stubAddr = StubRoutines::ghash_processBlocks();
6305 stubName = "ghash_processBlocks";
6306
6307 Node* data = argument(0);
6308 Node* offset = argument(1);
6309 Node* len = argument(2);
6310 Node* state = argument(3);
6311 Node* subkeyH = argument(4);
6312
6313 state = must_be_not_null(state, true);
6314 subkeyH = must_be_not_null(subkeyH, true);
6315 data = must_be_not_null(data, true);
6316
6317 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6318 assert(state_start, "state is NULL");
6319 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6320 assert(subkeyH_start, "subkeyH is NULL");
6321 Node* data_start = array_element_address(data, offset, T_BYTE);
6322 assert(data_start, "data is NULL");
6323
6324 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6325 OptoRuntime::ghash_processBlocks_Type(),
6326 stubAddr, stubName, TypePtr::BOTTOM,
6327 state_start, subkeyH_start, data_start, len);
6328 return true;
6329 }
6330
6331 bool LibraryCallKit::inline_base64_encodeBlock() {
6332 address stubAddr;
6333 const char *stubName;
6334 assert(UseBASE64Intrinsics, "need Base64 intrinsics support");
6335 assert(callee()->signature()->size() == 6, "base64_encodeBlock has 6 parameters");
6336 stubAddr = StubRoutines::base64_encodeBlock();
6337 stubName = "encodeBlock";
6338
6339 if (!stubAddr) return false;
6340 Node* base64obj = argument(0);
6341 Node* src = argument(1);
6342 Node* offset = argument(2);
6343 Node* len = argument(3);
6344 Node* dest = argument(4);
6345 Node* dp = argument(5);
6346 Node* isURL = argument(6);
6347
6348 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
6349 assert(src_start, "source array is NULL");
6350 Node* dest_start = array_element_address(dest, intcon(0), T_BYTE);
6351 assert(dest_start, "destination array is NULL");
6352
6353 Node* base64 = make_runtime_call(RC_LEAF,
6354 OptoRuntime::base64_encodeBlock_Type(),
6355 stubAddr, stubName, TypePtr::BOTTOM,
6356 src_start, offset, len, dest_start, dp, isURL);
6357 return true;
6358 }
6359
6360 //------------------------------inline_sha_implCompress-----------------------
6361 //
6362 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6363 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6364 //
6365 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6366 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6367 //
6368 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6369 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6370 //
6371 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6372 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6373
6374 Node* sha_obj = argument(0);
6375 Node* src = argument(1); // type oop
6376 Node* ofs = argument(2); // type int
6377
6378 const Type* src_type = src->Value(&_gvn);
6379 const TypeAryPtr* top_src = src_type->isa_aryptr();
6380 if (top_src == NULL || top_src->klass() == NULL) {
6381 // failed array check
6382 return false;
6383 }
6384 // Figure out the size and type of the elements we will be copying.
6385 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6386 if (src_elem != T_BYTE) {
6387 return false;
6388 }
6389 // 'src_start' points to src array + offset
6390 src = must_be_not_null(src, true);
6391 Node* src_start = array_element_address(src, ofs, src_elem);
6392 Node* state = NULL;
6393 address stubAddr;
6394 const char *stubName;
6395
6396 switch(id) {
6397 case vmIntrinsics::_sha_implCompress:
6398 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6399 state = get_state_from_sha_object(sha_obj);
6400 stubAddr = StubRoutines::sha1_implCompress();
6401 stubName = "sha1_implCompress";
6402 break;
6403 case vmIntrinsics::_sha2_implCompress:
6404 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6405 state = get_state_from_sha_object(sha_obj);
6406 stubAddr = StubRoutines::sha256_implCompress();
6407 stubName = "sha256_implCompress";
6408 break;
6409 case vmIntrinsics::_sha5_implCompress:
6410 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6411 state = get_state_from_sha5_object(sha_obj);
6412 stubAddr = StubRoutines::sha512_implCompress();
6413 stubName = "sha512_implCompress";
6414 break;
6415 default:
6416 fatal_unexpected_iid(id);
6417 return false;
6418 }
6419 if (state == NULL) return false;
6420
6421 // Call the stub.
6422 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6423 stubAddr, stubName, TypePtr::BOTTOM,
6424 src_start, state);
6425
6426 return true;
6427 }
6428
6429 //------------------------------inline_digestBase_implCompressMB-----------------------
6430 //
6431 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6432 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6433 //
6434 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6435 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6436 "need SHA1/SHA256/SHA512 instruction support");
6437 assert((uint)predicate < 3, "sanity");
6438 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6439
6440 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6441 Node* src = argument(1); // byte[] array
6442 Node* ofs = argument(2); // type int
6443 Node* limit = argument(3); // type int
6444
6445 const Type* src_type = src->Value(&_gvn);
6446 const TypeAryPtr* top_src = src_type->isa_aryptr();
6447 if (top_src == NULL || top_src->klass() == NULL) {
6448 // failed array check
6449 return false;
6450 }
6451 // Figure out the size and type of the elements we will be copying.
6452 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6453 if (src_elem != T_BYTE) {
6454 return false;
6455 }
6456 // 'src_start' points to src array + offset
6457 src = must_be_not_null(src, false);
6458 Node* src_start = array_element_address(src, ofs, src_elem);
6459
6460 const char* klass_SHA_name = NULL;
6461 const char* stub_name = NULL;
6462 address stub_addr = NULL;
6463 bool long_state = false;
6464
6465 switch (predicate) {
6466 case 0:
6467 if (UseSHA1Intrinsics) {
6468 klass_SHA_name = "sun/security/provider/SHA";
6469 stub_name = "sha1_implCompressMB";
6470 stub_addr = StubRoutines::sha1_implCompressMB();
6471 }
6472 break;
6473 case 1:
6474 if (UseSHA256Intrinsics) {
6475 klass_SHA_name = "sun/security/provider/SHA2";
6476 stub_name = "sha256_implCompressMB";
6477 stub_addr = StubRoutines::sha256_implCompressMB();
6478 }
6479 break;
6480 case 2:
6481 if (UseSHA512Intrinsics) {
6482 klass_SHA_name = "sun/security/provider/SHA5";
6483 stub_name = "sha512_implCompressMB";
6484 stub_addr = StubRoutines::sha512_implCompressMB();
6485 long_state = true;
6486 }
6487 break;
6488 default:
6489 fatal("unknown SHA intrinsic predicate: %d", predicate);
6490 }
6491 if (klass_SHA_name != NULL) {
6492 // get DigestBase klass to lookup for SHA klass
6493 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6494 assert(tinst != NULL, "digestBase_obj is not instance???");
6495 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6496
6497 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6498 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6499 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6500 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6501 }
6502 return false;
6503 }
6504 //------------------------------inline_sha_implCompressMB-----------------------
6505 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6506 bool long_state, address stubAddr, const char *stubName,
6507 Node* src_start, Node* ofs, Node* limit) {
6508 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6509 const TypeOopPtr* xtype = aklass->as_instance_type();
6510 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6511 sha_obj = _gvn.transform(sha_obj);
6512
6513 Node* state;
6514 if (long_state) {
6515 state = get_state_from_sha5_object(sha_obj);
6516 } else {
6517 state = get_state_from_sha_object(sha_obj);
6518 }
6519 if (state == NULL) return false;
6520
6521 // Call the stub.
6522 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6523 OptoRuntime::digestBase_implCompressMB_Type(),
6524 stubAddr, stubName, TypePtr::BOTTOM,
6525 src_start, state, ofs, limit);
6526 // return ofs (int)
6527 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6528 set_result(result);
6529
6530 return true;
6531 }
6532
6533 //------------------------------get_state_from_sha_object-----------------------
6534 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6535 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6536 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6537 if (sha_state == NULL) return (Node *) NULL;
6538
6539 // now have the array, need to get the start address of the state array
6540 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6541 return state;
6542 }
6543
6544 //------------------------------get_state_from_sha5_object-----------------------
6545 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6546 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6547 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6548 if (sha_state == NULL) return (Node *) NULL;
6549
6550 // now have the array, need to get the start address of the state array
6551 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6552 return state;
6553 }
6554
6555 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6556 // Return node representing slow path of predicate check.
6557 // the pseudo code we want to emulate with this predicate is:
6558 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6559 //
6560 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6561 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6562 "need SHA1/SHA256/SHA512 instruction support");
6563 assert((uint)predicate < 3, "sanity");
6564
6565 // The receiver was checked for NULL already.
6566 Node* digestBaseObj = argument(0);
6567
6568 // get DigestBase klass for instanceOf check
6569 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6570 assert(tinst != NULL, "digestBaseObj is null");
6571 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6572
6573 const char* klass_SHA_name = NULL;
6574 switch (predicate) {
6575 case 0:
6576 if (UseSHA1Intrinsics) {
6577 // we want to do an instanceof comparison against the SHA class
6578 klass_SHA_name = "sun/security/provider/SHA";
6579 }
6580 break;
6581 case 1:
6582 if (UseSHA256Intrinsics) {
6583 // we want to do an instanceof comparison against the SHA2 class
6584 klass_SHA_name = "sun/security/provider/SHA2";
6585 }
6586 break;
6587 case 2:
6588 if (UseSHA512Intrinsics) {
6589 // we want to do an instanceof comparison against the SHA5 class
6590 klass_SHA_name = "sun/security/provider/SHA5";
6591 }
6592 break;
6593 default:
6594 fatal("unknown SHA intrinsic predicate: %d", predicate);
6595 }
6596
6597 ciKlass* klass_SHA = NULL;
6598 if (klass_SHA_name != NULL) {
6599 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6600 }
6601 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6602 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6603 Node* ctrl = control();
6604 set_control(top()); // no intrinsic path
6605 return ctrl;
6606 }
6607 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6608
6609 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6610 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6611 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6612 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6613
6614 return instof_false; // even if it is NULL
6615 }
6616
6617 //-------------inline_fma-----------------------------------
6618 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6619 Node *a = NULL;
6620 Node *b = NULL;
6621 Node *c = NULL;
6622 Node* result = NULL;
6623 switch (id) {
6624 case vmIntrinsics::_fmaD:
6625 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6626 // no receiver since it is static method
6627 a = round_double_node(argument(0));
6628 b = round_double_node(argument(2));
6629 c = round_double_node(argument(4));
6630 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6631 break;
6632 case vmIntrinsics::_fmaF:
6633 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6634 a = argument(0);
6635 b = argument(1);
6636 c = argument(2);
6637 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6638 break;
6639 default:
6640 fatal_unexpected_iid(id); break;
6641 }
6642 set_result(result);
6643 return true;
6644 }
6645
6646 bool LibraryCallKit::inline_profileBoolean() {
6647 Node* counts = argument(1);
6648 const TypeAryPtr* ary = NULL;
6649 ciArray* aobj = NULL;
6650 if (counts->is_Con()
6651 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6652 && (aobj = ary->const_oop()->as_array()) != NULL
6653 && (aobj->length() == 2)) {
6654 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6655 jint false_cnt = aobj->element_value(0).as_int();
6656 jint true_cnt = aobj->element_value(1).as_int();
6657
6658 if (C->log() != NULL) {
6659 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6660 false_cnt, true_cnt);
6661 }
6662
6663 if (false_cnt + true_cnt == 0) {
6664 // According to profile, never executed.
6665 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6666 Deoptimization::Action_reinterpret);
6667 return true;
6668 }
6669
6670 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6671 // is a number of each value occurrences.
6672 Node* result = argument(0);
6673 if (false_cnt == 0 || true_cnt == 0) {
6674 // According to profile, one value has been never seen.
6675 int expected_val = (false_cnt == 0) ? 1 : 0;
6676
6677 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6678 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6679
6680 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6681 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6682 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6683
6684 { // Slow path: uncommon trap for never seen value and then reexecute
6685 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6686 // the value has been seen at least once.
6687 PreserveJVMState pjvms(this);
6688 PreserveReexecuteState preexecs(this);
6689 jvms()->set_should_reexecute(true);
6690
6691 set_control(slow_path);
6692 set_i_o(i_o());
6693
6694 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6695 Deoptimization::Action_reinterpret);
6696 }
6697 // The guard for never seen value enables sharpening of the result and
6698 // returning a constant. It allows to eliminate branches on the same value
6699 // later on.
6700 set_control(fast_path);
6701 result = intcon(expected_val);
6702 }
6703 // Stop profiling.
6704 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6705 // By replacing method body with profile data (represented as ProfileBooleanNode
6706 // on IR level) we effectively disable profiling.
6707 // It enables full speed execution once optimized code is generated.
6708 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6709 C->record_for_igvn(profile);
6710 set_result(profile);
6711 return true;
6712 } else {
6713 // Continue profiling.
6714 // Profile data isn't available at the moment. So, execute method's bytecode version.
6715 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6716 // is compiled and counters aren't available since corresponding MethodHandle
6717 // isn't a compile-time constant.
6718 return false;
6719 }
6720 }
6721
6722 bool LibraryCallKit::inline_isCompileConstant() {
6723 Node* n = argument(0);
6724 set_result(n->is_Con() ? intcon(1) : intcon(0));
6725 return true;
6726 }