1 /*
2 * Copyright (c) 1999, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "classfile/systemDictionary.hpp"
28 #include "classfile/vmSymbols.hpp"
29 #include "compiler/compileBroker.hpp"
30 #include "compiler/compileLog.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/rootnode.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "prims/unsafe.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #ifdef TRACE_HAVE_INTRINSICS
56 #include "trace/traceMacros.hpp"
57 #endif
58
59 class LibraryIntrinsic : public InlineCallGenerator {
60 // Extend the set of intrinsics known to the runtime:
61 public:
62 private:
63 bool _is_virtual;
64 bool _does_virtual_dispatch;
65 int8_t _predicates_count; // Intrinsic is predicated by several conditions
66 int8_t _last_predicate; // Last generated predicate
67 vmIntrinsics::ID _intrinsic_id;
68
69 public:
70 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
71 : InlineCallGenerator(m),
72 _is_virtual(is_virtual),
73 _does_virtual_dispatch(does_virtual_dispatch),
74 _predicates_count((int8_t)predicates_count),
75 _last_predicate((int8_t)-1),
76 _intrinsic_id(id)
77 {
78 }
79 virtual bool is_intrinsic() const { return true; }
80 virtual bool is_virtual() const { return _is_virtual; }
81 virtual bool is_predicated() const { return _predicates_count > 0; }
82 virtual int predicates_count() const { return _predicates_count; }
83 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
84 virtual JVMState* generate(JVMState* jvms);
85 virtual Node* generate_predicate(JVMState* jvms, int predicate);
86 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
87 };
88
89
90 // Local helper class for LibraryIntrinsic:
91 class LibraryCallKit : public GraphKit {
92 private:
93 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
94 Node* _result; // the result node, if any
95 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
96
97 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type);
98
99 public:
100 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
101 : GraphKit(jvms),
102 _intrinsic(intrinsic),
103 _result(NULL)
104 {
105 // Check if this is a root compile. In that case we don't have a caller.
106 if (!jvms->has_method()) {
107 _reexecute_sp = sp();
108 } else {
109 // Find out how many arguments the interpreter needs when deoptimizing
110 // and save the stack pointer value so it can used by uncommon_trap.
111 // We find the argument count by looking at the declared signature.
112 bool ignored_will_link;
113 ciSignature* declared_signature = NULL;
114 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
115 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
116 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
117 }
118 }
119
120 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
121
122 ciMethod* caller() const { return jvms()->method(); }
123 int bci() const { return jvms()->bci(); }
124 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
125 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
126 ciMethod* callee() const { return _intrinsic->method(); }
127
128 bool try_to_inline(int predicate);
129 Node* try_to_predicate(int predicate);
130
131 void push_result() {
132 // Push the result onto the stack.
133 if (!stopped() && result() != NULL) {
134 BasicType bt = result()->bottom_type()->basic_type();
135 push_node(bt, result());
136 }
137 }
138
139 private:
140 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
141 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
142 }
143
144 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
145 void set_result(RegionNode* region, PhiNode* value);
146 Node* result() { return _result; }
147
148 virtual int reexecute_sp() { return _reexecute_sp; }
149
150 // Helper functions to inline natives
151 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
152 Node* generate_slow_guard(Node* test, RegionNode* region);
153 Node* generate_fair_guard(Node* test, RegionNode* region);
154 Node* generate_negative_guard(Node* index, RegionNode* region,
155 // resulting CastII of index:
156 Node* *pos_index = NULL);
157 Node* generate_limit_guard(Node* offset, Node* subseq_length,
158 Node* array_length,
159 RegionNode* region);
160 void generate_string_range_check(Node* array, Node* offset,
161 Node* length, bool char_count);
162 Node* generate_current_thread(Node* &tls_output);
163 Node* load_mirror_from_klass(Node* klass);
164 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
165 RegionNode* region, int null_path,
166 int offset);
167 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
175 RegionNode* region, int null_path) {
176 int offset = java_lang_Class::array_klass_offset_in_bytes();
177 return load_klass_from_mirror_common(mirror, never_see_null,
178 region, null_path,
179 offset);
180 }
181 Node* generate_access_flags_guard(Node* kls,
182 int modifier_mask, int modifier_bits,
183 RegionNode* region);
184 Node* generate_interface_guard(Node* kls, RegionNode* region);
185 Node* generate_array_guard(Node* kls, RegionNode* region) {
186 return generate_array_guard_common(kls, region, false, false);
187 }
188 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
189 return generate_array_guard_common(kls, region, false, true);
190 }
191 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
192 return generate_array_guard_common(kls, region, true, false);
193 }
194 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
195 return generate_array_guard_common(kls, region, true, true);
196 }
197 Node* generate_array_guard_common(Node* kls, RegionNode* region,
198 bool obj_array, bool not_array);
199 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
200 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
201 bool is_virtual = false, bool is_static = false);
202 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
203 return generate_method_call(method_id, false, true);
204 }
205 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
206 return generate_method_call(method_id, true, false);
207 }
208 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
209 Node * field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
210
211 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
212 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
213 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
214 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
215 Node* make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
216 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae);
217 bool inline_string_indexOfChar();
218 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
219 bool inline_string_toBytesU();
220 bool inline_string_getCharsU();
221 bool inline_string_copy(bool compress);
222 bool inline_string_char_access(bool is_store);
223 Node* round_double_node(Node* n);
224 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
225 bool inline_math_native(vmIntrinsics::ID id);
226 bool inline_math(vmIntrinsics::ID id);
227 template <typename OverflowOp>
228 bool inline_math_overflow(Node* arg1, Node* arg2);
229 void inline_math_mathExact(Node* math, Node* test);
230 bool inline_math_addExactI(bool is_increment);
231 bool inline_math_addExactL(bool is_increment);
232 bool inline_math_multiplyExactI();
233 bool inline_math_multiplyExactL();
234 bool inline_math_multiplyHigh();
235 bool inline_math_negateExactI();
236 bool inline_math_negateExactL();
237 bool inline_math_subtractExactI(bool is_decrement);
238 bool inline_math_subtractExactL(bool is_decrement);
239 bool inline_min_max(vmIntrinsics::ID id);
240 bool inline_notify(vmIntrinsics::ID id);
241 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
242 // This returns Type::AnyPtr, RawPtr, or OopPtr.
243 int classify_unsafe_addr(Node* &base, Node* &offset, BasicType type);
244 Node* make_unsafe_address(Node*& base, Node* offset, BasicType type = T_ILLEGAL, bool can_cast = false);
245 // Helper for inline_unsafe_access.
246 // Generates the guards that check whether the result of
247 // Unsafe.getObject should be recorded in an SATB log buffer.
248 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
249
250 typedef enum { Relaxed, Opaque, Volatile, Acquire, Release } AccessKind;
251 bool inline_unsafe_access(bool is_store, BasicType type, AccessKind kind, bool is_unaligned);
252 static bool klass_needs_init_guard(Node* kls);
253 bool inline_unsafe_allocate();
254 bool inline_unsafe_newArray(bool uninitialized);
255 bool inline_unsafe_copyMemory();
256 bool inline_native_currentThread();
257
258 bool inline_native_time_funcs(address method, const char* funcName);
259 #ifdef TRACE_HAVE_INTRINSICS
260 bool inline_native_classID();
261 bool inline_native_getBufferWriter();
262 #endif
263 bool inline_native_isInterrupted();
264 bool inline_native_Class_query(vmIntrinsics::ID id);
265 bool inline_native_subtype_check();
266 bool inline_native_getLength();
267 bool inline_array_copyOf(bool is_copyOfRange);
268 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
269 bool inline_preconditions_checkIndex();
270 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
271 bool inline_native_clone(bool is_virtual);
272 bool inline_native_Reflection_getCallerClass();
273 // Helper function for inlining native object hash method
274 bool inline_native_hashcode(bool is_virtual, bool is_static);
275 bool inline_native_getClass();
276
277 // Helper functions for inlining arraycopy
278 bool inline_arraycopy();
279 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
280 RegionNode* slow_region);
281 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
282 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp,
283 uint new_idx);
284
285 typedef enum { LS_get_add, LS_get_set, LS_cmp_swap, LS_cmp_swap_weak, LS_cmp_exchange } LoadStoreKind;
286 MemNode::MemOrd access_kind_to_memord_LS(AccessKind access_kind, bool is_store);
287 MemNode::MemOrd access_kind_to_memord(AccessKind access_kind);
288 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind, AccessKind access_kind);
289 bool inline_unsafe_fence(vmIntrinsics::ID id);
290 bool inline_onspinwait();
291 bool inline_fp_conversions(vmIntrinsics::ID id);
292 bool inline_number_methods(vmIntrinsics::ID id);
293 bool inline_reference_get();
294 bool inline_Class_cast();
295 bool inline_aescrypt_Block(vmIntrinsics::ID id);
296 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
297 bool inline_counterMode_AESCrypt(vmIntrinsics::ID id);
298 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
299 Node* inline_counterMode_AESCrypt_predicate();
300 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
301 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
302 bool inline_ghash_processBlocks();
303 bool inline_sha_implCompress(vmIntrinsics::ID id);
304 bool inline_digestBase_implCompressMB(int predicate);
305 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
306 bool long_state, address stubAddr, const char *stubName,
307 Node* src_start, Node* ofs, Node* limit);
308 Node* get_state_from_sha_object(Node *sha_object);
309 Node* get_state_from_sha5_object(Node *sha_object);
310 Node* inline_digestBase_implCompressMB_predicate(int predicate);
311 bool inline_encodeISOArray();
312 bool inline_updateCRC32();
313 bool inline_updateBytesCRC32();
314 bool inline_updateByteBufferCRC32();
315 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
316 bool inline_updateBytesCRC32C();
317 bool inline_updateDirectByteBufferCRC32C();
318 bool inline_updateBytesAdler32();
319 bool inline_updateByteBufferAdler32();
320 bool inline_multiplyToLen();
321 bool inline_hasNegatives();
322 bool inline_squareToLen();
323 bool inline_mulAdd();
324 bool inline_montgomeryMultiply();
325 bool inline_montgomerySquare();
326 bool inline_vectorizedMismatch();
327 bool inline_fma(vmIntrinsics::ID id);
328
329 bool inline_profileBoolean();
330 bool inline_isCompileConstant();
331 void clear_upper_avx() {
332 #ifdef X86
333 if (UseAVX >= 2) {
334 C->set_clear_upper_avx(true);
335 }
336 #endif
337 }
338 };
339
340 //---------------------------make_vm_intrinsic----------------------------
341 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
342 vmIntrinsics::ID id = m->intrinsic_id();
343 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
344
345 if (!m->is_loaded()) {
346 // Do not attempt to inline unloaded methods.
347 return NULL;
348 }
349
350 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
351 bool is_available = false;
352
353 {
354 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
355 // the compiler must transition to '_thread_in_vm' state because both
356 // methods access VM-internal data.
357 VM_ENTRY_MARK;
358 methodHandle mh(THREAD, m->get_Method());
359 is_available = compiler != NULL && compiler->is_intrinsic_supported(mh, is_virtual) &&
360 !C->directive()->is_intrinsic_disabled(mh) &&
361 !vmIntrinsics::is_disabled_by_flags(mh);
362
363 }
364
365 if (is_available) {
366 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
367 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
368 return new LibraryIntrinsic(m, is_virtual,
369 vmIntrinsics::predicates_needed(id),
370 vmIntrinsics::does_virtual_dispatch(id),
371 (vmIntrinsics::ID) id);
372 } else {
373 return NULL;
374 }
375 }
376
377 //----------------------register_library_intrinsics-----------------------
378 // Initialize this file's data structures, for each Compile instance.
379 void Compile::register_library_intrinsics() {
380 // Nothing to do here.
381 }
382
383 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
384 LibraryCallKit kit(jvms, this);
385 Compile* C = kit.C;
386 int nodes = C->unique();
387 #ifndef PRODUCT
388 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
389 char buf[1000];
390 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
391 tty->print_cr("Intrinsic %s", str);
392 }
393 #endif
394 ciMethod* callee = kit.callee();
395 const int bci = kit.bci();
396
397 // Try to inline the intrinsic.
398 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
399 kit.try_to_inline(_last_predicate)) {
400 const char *inline_msg = is_virtual() ? "(intrinsic, virtual)"
401 : "(intrinsic)";
402 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
403 if (C->print_intrinsics() || C->print_inlining()) {
404 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
405 }
406 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
407 if (C->log()) {
408 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
409 vmIntrinsics::name_at(intrinsic_id()),
410 (is_virtual() ? " virtual='1'" : ""),
411 C->unique() - nodes);
412 }
413 // Push the result from the inlined method onto the stack.
414 kit.push_result();
415 C->print_inlining_update(this);
416 return kit.transfer_exceptions_into_jvms();
417 }
418
419 // The intrinsic bailed out
420 if (jvms->has_method()) {
421 // Not a root compile.
422 const char* msg;
423 if (callee->intrinsic_candidate()) {
424 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
425 } else {
426 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
427 : "failed to inline (intrinsic), method not annotated";
428 }
429 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, msg);
430 if (C->print_intrinsics() || C->print_inlining()) {
431 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
432 }
433 } else {
434 // Root compile
435 ResourceMark rm;
436 stringStream msg_stream;
437 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
438 vmIntrinsics::name_at(intrinsic_id()),
439 is_virtual() ? " (virtual)" : "", bci);
440 const char *msg = msg_stream.as_string();
441 log_debug(jit, inlining)("%s", msg);
442 if (C->print_intrinsics() || C->print_inlining()) {
443 tty->print("%s", msg);
444 }
445 }
446 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
447 C->print_inlining_update(this);
448 return NULL;
449 }
450
451 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
452 LibraryCallKit kit(jvms, this);
453 Compile* C = kit.C;
454 int nodes = C->unique();
455 _last_predicate = predicate;
456 #ifndef PRODUCT
457 assert(is_predicated() && predicate < predicates_count(), "sanity");
458 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
459 char buf[1000];
460 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
461 tty->print_cr("Predicate for intrinsic %s", str);
462 }
463 #endif
464 ciMethod* callee = kit.callee();
465 const int bci = kit.bci();
466
467 Node* slow_ctl = kit.try_to_predicate(predicate);
468 if (!kit.failing()) {
469 const char *inline_msg = is_virtual() ? "(intrinsic, virtual, predicate)"
470 : "(intrinsic, predicate)";
471 CompileTask::print_inlining_ul(callee, jvms->depth() - 1, bci, inline_msg);
472 if (C->print_intrinsics() || C->print_inlining()) {
473 C->print_inlining(callee, jvms->depth() - 1, bci, inline_msg);
474 }
475 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
476 if (C->log()) {
477 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
478 vmIntrinsics::name_at(intrinsic_id()),
479 (is_virtual() ? " virtual='1'" : ""),
480 C->unique() - nodes);
481 }
482 return slow_ctl; // Could be NULL if the check folds.
483 }
484
485 // The intrinsic bailed out
486 if (jvms->has_method()) {
487 // Not a root compile.
488 const char* msg = "failed to generate predicate for intrinsic";
489 CompileTask::print_inlining_ul(kit.callee(), jvms->depth() - 1, bci, msg);
490 if (C->print_intrinsics() || C->print_inlining()) {
491 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
492 }
493 } else {
494 // Root compile
495 ResourceMark rm;
496 stringStream msg_stream;
497 msg_stream.print("Did not generate intrinsic %s%s at bci:%d in",
498 vmIntrinsics::name_at(intrinsic_id()),
499 is_virtual() ? " (virtual)" : "", bci);
500 const char *msg = msg_stream.as_string();
501 log_debug(jit, inlining)("%s", msg);
502 if (C->print_intrinsics() || C->print_inlining()) {
503 C->print_inlining_stream()->print("%s", msg);
504 }
505 }
506 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
507 return NULL;
508 }
509
510 bool LibraryCallKit::try_to_inline(int predicate) {
511 // Handle symbolic names for otherwise undistinguished boolean switches:
512 const bool is_store = true;
513 const bool is_compress = true;
514 const bool is_static = true;
515 const bool is_volatile = true;
516
517 if (!jvms()->has_method()) {
518 // Root JVMState has a null method.
519 assert(map()->memory()->Opcode() == Op_Parm, "");
520 // Insert the memory aliasing node
521 set_all_memory(reset_memory());
522 }
523 assert(merged_memory(), "");
524
525
526 switch (intrinsic_id()) {
527 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
528 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
529 case vmIntrinsics::_getClass: return inline_native_getClass();
530
531 case vmIntrinsics::_dsin:
532 case vmIntrinsics::_dcos:
533 case vmIntrinsics::_dtan:
534 case vmIntrinsics::_dabs:
535 case vmIntrinsics::_datan2:
536 case vmIntrinsics::_dsqrt:
537 case vmIntrinsics::_dexp:
538 case vmIntrinsics::_dlog:
539 case vmIntrinsics::_dlog10:
540 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
541
542 case vmIntrinsics::_min:
543 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
544
545 case vmIntrinsics::_notify:
546 case vmIntrinsics::_notifyAll:
547 if (InlineNotify) {
548 return inline_notify(intrinsic_id());
549 }
550 return false;
551
552 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
553 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
554 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
555 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
556 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
557 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
558 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
559 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
560 case vmIntrinsics::_multiplyHigh: return inline_math_multiplyHigh();
561 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
562 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
563 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
564 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
565
566 case vmIntrinsics::_arraycopy: return inline_arraycopy();
567
568 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
569 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
570 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
571 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
572
573 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
574 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
575 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
576 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
577 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
578 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
579 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
580
581 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
582 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
583
584 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
585 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
586 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
587 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
588
589 case vmIntrinsics::_compressStringC:
590 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
591 case vmIntrinsics::_inflateStringC:
592 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
593
594 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_store, T_OBJECT, Relaxed, false);
595 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_store, T_BOOLEAN, Relaxed, false);
596 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_store, T_BYTE, Relaxed, false);
597 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, false);
598 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, false);
599 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_store, T_INT, Relaxed, false);
600 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_store, T_LONG, Relaxed, false);
601 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_store, T_FLOAT, Relaxed, false);
602 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_store, T_DOUBLE, Relaxed, false);
603
604 case vmIntrinsics::_putObject: return inline_unsafe_access( is_store, T_OBJECT, Relaxed, false);
605 case vmIntrinsics::_putBoolean: return inline_unsafe_access( is_store, T_BOOLEAN, Relaxed, false);
606 case vmIntrinsics::_putByte: return inline_unsafe_access( is_store, T_BYTE, Relaxed, false);
607 case vmIntrinsics::_putShort: return inline_unsafe_access( is_store, T_SHORT, Relaxed, false);
608 case vmIntrinsics::_putChar: return inline_unsafe_access( is_store, T_CHAR, Relaxed, false);
609 case vmIntrinsics::_putInt: return inline_unsafe_access( is_store, T_INT, Relaxed, false);
610 case vmIntrinsics::_putLong: return inline_unsafe_access( is_store, T_LONG, Relaxed, false);
611 case vmIntrinsics::_putFloat: return inline_unsafe_access( is_store, T_FLOAT, Relaxed, false);
612 case vmIntrinsics::_putDouble: return inline_unsafe_access( is_store, T_DOUBLE, Relaxed, false);
613
614 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_store, T_OBJECT, Volatile, false);
615 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_store, T_BOOLEAN, Volatile, false);
616 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_store, T_BYTE, Volatile, false);
617 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_store, T_SHORT, Volatile, false);
618 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_store, T_CHAR, Volatile, false);
619 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_store, T_INT, Volatile, false);
620 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_store, T_LONG, Volatile, false);
621 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_store, T_FLOAT, Volatile, false);
622 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_store, T_DOUBLE, Volatile, false);
623
624 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access( is_store, T_OBJECT, Volatile, false);
625 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access( is_store, T_BOOLEAN, Volatile, false);
626 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access( is_store, T_BYTE, Volatile, false);
627 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access( is_store, T_SHORT, Volatile, false);
628 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access( is_store, T_CHAR, Volatile, false);
629 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access( is_store, T_INT, Volatile, false);
630 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access( is_store, T_LONG, Volatile, false);
631 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access( is_store, T_FLOAT, Volatile, false);
632 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access( is_store, T_DOUBLE, Volatile, false);
633
634 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_store, T_SHORT, Relaxed, true);
635 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_store, T_CHAR, Relaxed, true);
636 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_store, T_INT, Relaxed, true);
637 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_store, T_LONG, Relaxed, true);
638
639 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access( is_store, T_SHORT, Relaxed, true);
640 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access( is_store, T_CHAR, Relaxed, true);
641 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access( is_store, T_INT, Relaxed, true);
642 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access( is_store, T_LONG, Relaxed, true);
643
644 case vmIntrinsics::_getObjectAcquire: return inline_unsafe_access(!is_store, T_OBJECT, Acquire, false);
645 case vmIntrinsics::_getBooleanAcquire: return inline_unsafe_access(!is_store, T_BOOLEAN, Acquire, false);
646 case vmIntrinsics::_getByteAcquire: return inline_unsafe_access(!is_store, T_BYTE, Acquire, false);
647 case vmIntrinsics::_getShortAcquire: return inline_unsafe_access(!is_store, T_SHORT, Acquire, false);
648 case vmIntrinsics::_getCharAcquire: return inline_unsafe_access(!is_store, T_CHAR, Acquire, false);
649 case vmIntrinsics::_getIntAcquire: return inline_unsafe_access(!is_store, T_INT, Acquire, false);
650 case vmIntrinsics::_getLongAcquire: return inline_unsafe_access(!is_store, T_LONG, Acquire, false);
651 case vmIntrinsics::_getFloatAcquire: return inline_unsafe_access(!is_store, T_FLOAT, Acquire, false);
652 case vmIntrinsics::_getDoubleAcquire: return inline_unsafe_access(!is_store, T_DOUBLE, Acquire, false);
653
654 case vmIntrinsics::_putObjectRelease: return inline_unsafe_access( is_store, T_OBJECT, Release, false);
655 case vmIntrinsics::_putBooleanRelease: return inline_unsafe_access( is_store, T_BOOLEAN, Release, false);
656 case vmIntrinsics::_putByteRelease: return inline_unsafe_access( is_store, T_BYTE, Release, false);
657 case vmIntrinsics::_putShortRelease: return inline_unsafe_access( is_store, T_SHORT, Release, false);
658 case vmIntrinsics::_putCharRelease: return inline_unsafe_access( is_store, T_CHAR, Release, false);
659 case vmIntrinsics::_putIntRelease: return inline_unsafe_access( is_store, T_INT, Release, false);
660 case vmIntrinsics::_putLongRelease: return inline_unsafe_access( is_store, T_LONG, Release, false);
661 case vmIntrinsics::_putFloatRelease: return inline_unsafe_access( is_store, T_FLOAT, Release, false);
662 case vmIntrinsics::_putDoubleRelease: return inline_unsafe_access( is_store, T_DOUBLE, Release, false);
663
664 case vmIntrinsics::_getObjectOpaque: return inline_unsafe_access(!is_store, T_OBJECT, Opaque, false);
665 case vmIntrinsics::_getBooleanOpaque: return inline_unsafe_access(!is_store, T_BOOLEAN, Opaque, false);
666 case vmIntrinsics::_getByteOpaque: return inline_unsafe_access(!is_store, T_BYTE, Opaque, false);
667 case vmIntrinsics::_getShortOpaque: return inline_unsafe_access(!is_store, T_SHORT, Opaque, false);
668 case vmIntrinsics::_getCharOpaque: return inline_unsafe_access(!is_store, T_CHAR, Opaque, false);
669 case vmIntrinsics::_getIntOpaque: return inline_unsafe_access(!is_store, T_INT, Opaque, false);
670 case vmIntrinsics::_getLongOpaque: return inline_unsafe_access(!is_store, T_LONG, Opaque, false);
671 case vmIntrinsics::_getFloatOpaque: return inline_unsafe_access(!is_store, T_FLOAT, Opaque, false);
672 case vmIntrinsics::_getDoubleOpaque: return inline_unsafe_access(!is_store, T_DOUBLE, Opaque, false);
673
674 case vmIntrinsics::_putObjectOpaque: return inline_unsafe_access( is_store, T_OBJECT, Opaque, false);
675 case vmIntrinsics::_putBooleanOpaque: return inline_unsafe_access( is_store, T_BOOLEAN, Opaque, false);
676 case vmIntrinsics::_putByteOpaque: return inline_unsafe_access( is_store, T_BYTE, Opaque, false);
677 case vmIntrinsics::_putShortOpaque: return inline_unsafe_access( is_store, T_SHORT, Opaque, false);
678 case vmIntrinsics::_putCharOpaque: return inline_unsafe_access( is_store, T_CHAR, Opaque, false);
679 case vmIntrinsics::_putIntOpaque: return inline_unsafe_access( is_store, T_INT, Opaque, false);
680 case vmIntrinsics::_putLongOpaque: return inline_unsafe_access( is_store, T_LONG, Opaque, false);
681 case vmIntrinsics::_putFloatOpaque: return inline_unsafe_access( is_store, T_FLOAT, Opaque, false);
682 case vmIntrinsics::_putDoubleOpaque: return inline_unsafe_access( is_store, T_DOUBLE, Opaque, false);
683
684 case vmIntrinsics::_compareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap, Volatile);
685 case vmIntrinsics::_compareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap, Volatile);
686 case vmIntrinsics::_compareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap, Volatile);
687 case vmIntrinsics::_compareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap, Volatile);
688 case vmIntrinsics::_compareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap, Volatile);
689
690 case vmIntrinsics::_weakCompareAndSetObjectPlain: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Relaxed);
691 case vmIntrinsics::_weakCompareAndSetObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Acquire);
692 case vmIntrinsics::_weakCompareAndSetObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Release);
693 case vmIntrinsics::_weakCompareAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_swap_weak, Volatile);
694 case vmIntrinsics::_weakCompareAndSetBytePlain: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Relaxed);
695 case vmIntrinsics::_weakCompareAndSetByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Acquire);
696 case vmIntrinsics::_weakCompareAndSetByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Release);
697 case vmIntrinsics::_weakCompareAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_swap_weak, Volatile);
698 case vmIntrinsics::_weakCompareAndSetShortPlain: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Relaxed);
699 case vmIntrinsics::_weakCompareAndSetShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Acquire);
700 case vmIntrinsics::_weakCompareAndSetShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Release);
701 case vmIntrinsics::_weakCompareAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_swap_weak, Volatile);
702 case vmIntrinsics::_weakCompareAndSetIntPlain: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Relaxed);
703 case vmIntrinsics::_weakCompareAndSetIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Acquire);
704 case vmIntrinsics::_weakCompareAndSetIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Release);
705 case vmIntrinsics::_weakCompareAndSetInt: return inline_unsafe_load_store(T_INT, LS_cmp_swap_weak, Volatile);
706 case vmIntrinsics::_weakCompareAndSetLongPlain: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Relaxed);
707 case vmIntrinsics::_weakCompareAndSetLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Acquire);
708 case vmIntrinsics::_weakCompareAndSetLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Release);
709 case vmIntrinsics::_weakCompareAndSetLong: return inline_unsafe_load_store(T_LONG, LS_cmp_swap_weak, Volatile);
710
711 case vmIntrinsics::_compareAndExchangeObject: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Volatile);
712 case vmIntrinsics::_compareAndExchangeObjectAcquire: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Acquire);
713 case vmIntrinsics::_compareAndExchangeObjectRelease: return inline_unsafe_load_store(T_OBJECT, LS_cmp_exchange, Release);
714 case vmIntrinsics::_compareAndExchangeByte: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Volatile);
715 case vmIntrinsics::_compareAndExchangeByteAcquire: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Acquire);
716 case vmIntrinsics::_compareAndExchangeByteRelease: return inline_unsafe_load_store(T_BYTE, LS_cmp_exchange, Release);
717 case vmIntrinsics::_compareAndExchangeShort: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Volatile);
718 case vmIntrinsics::_compareAndExchangeShortAcquire: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Acquire);
719 case vmIntrinsics::_compareAndExchangeShortRelease: return inline_unsafe_load_store(T_SHORT, LS_cmp_exchange, Release);
720 case vmIntrinsics::_compareAndExchangeInt: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Volatile);
721 case vmIntrinsics::_compareAndExchangeIntAcquire: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Acquire);
722 case vmIntrinsics::_compareAndExchangeIntRelease: return inline_unsafe_load_store(T_INT, LS_cmp_exchange, Release);
723 case vmIntrinsics::_compareAndExchangeLong: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Volatile);
724 case vmIntrinsics::_compareAndExchangeLongAcquire: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Acquire);
725 case vmIntrinsics::_compareAndExchangeLongRelease: return inline_unsafe_load_store(T_LONG, LS_cmp_exchange, Release);
726
727 case vmIntrinsics::_getAndAddByte: return inline_unsafe_load_store(T_BYTE, LS_get_add, Volatile);
728 case vmIntrinsics::_getAndAddShort: return inline_unsafe_load_store(T_SHORT, LS_get_add, Volatile);
729 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_get_add, Volatile);
730 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_get_add, Volatile);
731
732 case vmIntrinsics::_getAndSetByte: return inline_unsafe_load_store(T_BYTE, LS_get_set, Volatile);
733 case vmIntrinsics::_getAndSetShort: return inline_unsafe_load_store(T_SHORT, LS_get_set, Volatile);
734 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_get_set, Volatile);
735 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_get_set, Volatile);
736 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_get_set, Volatile);
737
738 case vmIntrinsics::_loadFence:
739 case vmIntrinsics::_storeFence:
740 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
741
742 case vmIntrinsics::_onSpinWait: return inline_onspinwait();
743
744 case vmIntrinsics::_currentThread: return inline_native_currentThread();
745 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
746
747 #ifdef TRACE_HAVE_INTRINSICS
748 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
749 case vmIntrinsics::_getClassId: return inline_native_classID();
750 case vmIntrinsics::_getBufferWriter: return inline_native_getBufferWriter();
751 #endif
752 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
753 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
754 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
755 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
756 case vmIntrinsics::_getLength: return inline_native_getLength();
757 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
758 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
759 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
760 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
761 case vmIntrinsics::_Preconditions_checkIndex: return inline_preconditions_checkIndex();
762 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
763
764 case vmIntrinsics::_allocateUninitializedArray: return inline_unsafe_newArray(true);
765 case vmIntrinsics::_newArray: return inline_unsafe_newArray(false);
766
767 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
768
769 case vmIntrinsics::_isInstance:
770 case vmIntrinsics::_getModifiers:
771 case vmIntrinsics::_isInterface:
772 case vmIntrinsics::_isArray:
773 case vmIntrinsics::_isPrimitive:
774 case vmIntrinsics::_getSuperclass:
775 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
776
777 case vmIntrinsics::_floatToRawIntBits:
778 case vmIntrinsics::_floatToIntBits:
779 case vmIntrinsics::_intBitsToFloat:
780 case vmIntrinsics::_doubleToRawLongBits:
781 case vmIntrinsics::_doubleToLongBits:
782 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
783
784 case vmIntrinsics::_numberOfLeadingZeros_i:
785 case vmIntrinsics::_numberOfLeadingZeros_l:
786 case vmIntrinsics::_numberOfTrailingZeros_i:
787 case vmIntrinsics::_numberOfTrailingZeros_l:
788 case vmIntrinsics::_bitCount_i:
789 case vmIntrinsics::_bitCount_l:
790 case vmIntrinsics::_reverseBytes_i:
791 case vmIntrinsics::_reverseBytes_l:
792 case vmIntrinsics::_reverseBytes_s:
793 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
794
795 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
796
797 case vmIntrinsics::_Reference_get: return inline_reference_get();
798
799 case vmIntrinsics::_Class_cast: return inline_Class_cast();
800
801 case vmIntrinsics::_aescrypt_encryptBlock:
802 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
803
804 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
805 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
806 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
807
808 case vmIntrinsics::_counterMode_AESCrypt:
809 return inline_counterMode_AESCrypt(intrinsic_id());
810
811 case vmIntrinsics::_sha_implCompress:
812 case vmIntrinsics::_sha2_implCompress:
813 case vmIntrinsics::_sha5_implCompress:
814 return inline_sha_implCompress(intrinsic_id());
815
816 case vmIntrinsics::_digestBase_implCompressMB:
817 return inline_digestBase_implCompressMB(predicate);
818
819 case vmIntrinsics::_multiplyToLen:
820 return inline_multiplyToLen();
821
822 case vmIntrinsics::_squareToLen:
823 return inline_squareToLen();
824
825 case vmIntrinsics::_mulAdd:
826 return inline_mulAdd();
827
828 case vmIntrinsics::_montgomeryMultiply:
829 return inline_montgomeryMultiply();
830 case vmIntrinsics::_montgomerySquare:
831 return inline_montgomerySquare();
832
833 case vmIntrinsics::_vectorizedMismatch:
834 return inline_vectorizedMismatch();
835
836 case vmIntrinsics::_ghash_processBlocks:
837 return inline_ghash_processBlocks();
838
839 case vmIntrinsics::_encodeISOArray:
840 case vmIntrinsics::_encodeByteISOArray:
841 return inline_encodeISOArray();
842
843 case vmIntrinsics::_updateCRC32:
844 return inline_updateCRC32();
845 case vmIntrinsics::_updateBytesCRC32:
846 return inline_updateBytesCRC32();
847 case vmIntrinsics::_updateByteBufferCRC32:
848 return inline_updateByteBufferCRC32();
849
850 case vmIntrinsics::_updateBytesCRC32C:
851 return inline_updateBytesCRC32C();
852 case vmIntrinsics::_updateDirectByteBufferCRC32C:
853 return inline_updateDirectByteBufferCRC32C();
854
855 case vmIntrinsics::_updateBytesAdler32:
856 return inline_updateBytesAdler32();
857 case vmIntrinsics::_updateByteBufferAdler32:
858 return inline_updateByteBufferAdler32();
859
860 case vmIntrinsics::_profileBoolean:
861 return inline_profileBoolean();
862 case vmIntrinsics::_isCompileConstant:
863 return inline_isCompileConstant();
864
865 case vmIntrinsics::_hasNegatives:
866 return inline_hasNegatives();
867
868 case vmIntrinsics::_fmaD:
869 case vmIntrinsics::_fmaF:
870 return inline_fma(intrinsic_id());
871
872 default:
873 // If you get here, it may be that someone has added a new intrinsic
874 // to the list in vmSymbols.hpp without implementing it here.
875 #ifndef PRODUCT
876 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
877 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
878 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
879 }
880 #endif
881 return false;
882 }
883 }
884
885 Node* LibraryCallKit::try_to_predicate(int predicate) {
886 if (!jvms()->has_method()) {
887 // Root JVMState has a null method.
888 assert(map()->memory()->Opcode() == Op_Parm, "");
889 // Insert the memory aliasing node
890 set_all_memory(reset_memory());
891 }
892 assert(merged_memory(), "");
893
894 switch (intrinsic_id()) {
895 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
896 return inline_cipherBlockChaining_AESCrypt_predicate(false);
897 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
898 return inline_cipherBlockChaining_AESCrypt_predicate(true);
899 case vmIntrinsics::_counterMode_AESCrypt:
900 return inline_counterMode_AESCrypt_predicate();
901 case vmIntrinsics::_digestBase_implCompressMB:
902 return inline_digestBase_implCompressMB_predicate(predicate);
903
904 default:
905 // If you get here, it may be that someone has added a new intrinsic
906 // to the list in vmSymbols.hpp without implementing it here.
907 #ifndef PRODUCT
908 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
909 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
910 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
911 }
912 #endif
913 Node* slow_ctl = control();
914 set_control(top()); // No fast path instrinsic
915 return slow_ctl;
916 }
917 }
918
919 //------------------------------set_result-------------------------------
920 // Helper function for finishing intrinsics.
921 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
922 record_for_igvn(region);
923 set_control(_gvn.transform(region));
924 set_result( _gvn.transform(value));
925 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
926 }
927
928 //------------------------------generate_guard---------------------------
929 // Helper function for generating guarded fast-slow graph structures.
930 // The given 'test', if true, guards a slow path. If the test fails
931 // then a fast path can be taken. (We generally hope it fails.)
932 // In all cases, GraphKit::control() is updated to the fast path.
933 // The returned value represents the control for the slow path.
934 // The return value is never 'top'; it is either a valid control
935 // or NULL if it is obvious that the slow path can never be taken.
936 // Also, if region and the slow control are not NULL, the slow edge
937 // is appended to the region.
938 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
939 if (stopped()) {
940 // Already short circuited.
941 return NULL;
942 }
943
944 // Build an if node and its projections.
945 // If test is true we take the slow path, which we assume is uncommon.
946 if (_gvn.type(test) == TypeInt::ZERO) {
947 // The slow branch is never taken. No need to build this guard.
948 return NULL;
949 }
950
951 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
952
953 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
954 if (if_slow == top()) {
955 // The slow branch is never taken. No need to build this guard.
956 return NULL;
957 }
958
959 if (region != NULL)
960 region->add_req(if_slow);
961
962 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
963 set_control(if_fast);
964
965 return if_slow;
966 }
967
968 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
969 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
970 }
971 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
972 return generate_guard(test, region, PROB_FAIR);
973 }
974
975 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
976 Node* *pos_index) {
977 if (stopped())
978 return NULL; // already stopped
979 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
980 return NULL; // index is already adequately typed
981 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
982 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
983 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
984 if (is_neg != NULL && pos_index != NULL) {
985 // Emulate effect of Parse::adjust_map_after_if.
986 Node* ccast = new CastIINode(index, TypeInt::POS);
987 ccast->set_req(0, control());
988 (*pos_index) = _gvn.transform(ccast);
989 }
990 return is_neg;
991 }
992
993 // Make sure that 'position' is a valid limit index, in [0..length].
994 // There are two equivalent plans for checking this:
995 // A. (offset + copyLength) unsigned<= arrayLength
996 // B. offset <= (arrayLength - copyLength)
997 // We require that all of the values above, except for the sum and
998 // difference, are already known to be non-negative.
999 // Plan A is robust in the face of overflow, if offset and copyLength
1000 // are both hugely positive.
1001 //
1002 // Plan B is less direct and intuitive, but it does not overflow at
1003 // all, since the difference of two non-negatives is always
1004 // representable. Whenever Java methods must perform the equivalent
1005 // check they generally use Plan B instead of Plan A.
1006 // For the moment we use Plan A.
1007 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1008 Node* subseq_length,
1009 Node* array_length,
1010 RegionNode* region) {
1011 if (stopped())
1012 return NULL; // already stopped
1013 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1014 if (zero_offset && subseq_length->eqv_uncast(array_length))
1015 return NULL; // common case of whole-array copy
1016 Node* last = subseq_length;
1017 if (!zero_offset) // last += offset
1018 last = _gvn.transform(new AddINode(last, offset));
1019 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
1020 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
1021 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1022 return is_over;
1023 }
1024
1025 // Emit range checks for the given String.value byte array
1026 void LibraryCallKit::generate_string_range_check(Node* array, Node* offset, Node* count, bool char_count) {
1027 if (stopped()) {
1028 return; // already stopped
1029 }
1030 RegionNode* bailout = new RegionNode(1);
1031 record_for_igvn(bailout);
1032 if (char_count) {
1033 // Convert char count to byte count
1034 count = _gvn.transform(new LShiftINode(count, intcon(1)));
1035 }
1036
1037 // Offset and count must not be negative
1038 generate_negative_guard(offset, bailout);
1039 generate_negative_guard(count, bailout);
1040 // Offset + count must not exceed length of array
1041 generate_limit_guard(offset, count, load_array_length(array), bailout);
1042
1043 if (bailout->req() > 1) {
1044 PreserveJVMState pjvms(this);
1045 set_control(_gvn.transform(bailout));
1046 uncommon_trap(Deoptimization::Reason_intrinsic,
1047 Deoptimization::Action_maybe_recompile);
1048 }
1049 }
1050
1051 //--------------------------generate_current_thread--------------------
1052 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1053 ciKlass* thread_klass = env()->Thread_klass();
1054 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1055 Node* thread = _gvn.transform(new ThreadLocalNode());
1056 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1057 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1058 tls_output = thread;
1059 return threadObj;
1060 }
1061
1062
1063 //------------------------------make_string_method_node------------------------
1064 // Helper method for String intrinsic functions. This version is called with
1065 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
1066 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
1067 // containing the lengths of str1 and str2.
1068 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
1069 Node* result = NULL;
1070 switch (opcode) {
1071 case Op_StrIndexOf:
1072 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
1073 str1_start, cnt1, str2_start, cnt2, ae);
1074 break;
1075 case Op_StrComp:
1076 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
1077 str1_start, cnt1, str2_start, cnt2, ae);
1078 break;
1079 case Op_StrEquals:
1080 // We already know that cnt1 == cnt2 here (checked in 'inline_string_equals').
1081 // Use the constant length if there is one because optimized match rule may exist.
1082 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
1083 str1_start, str2_start, cnt2->is_Con() ? cnt2 : cnt1, ae);
1084 break;
1085 default:
1086 ShouldNotReachHere();
1087 return NULL;
1088 }
1089
1090 // All these intrinsics have checks.
1091 C->set_has_split_ifs(true); // Has chance for split-if optimization
1092 clear_upper_avx();
1093
1094 return _gvn.transform(result);
1095 }
1096
1097 //------------------------------inline_string_compareTo------------------------
1098 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
1099 Node* arg1 = argument(0);
1100 Node* arg2 = argument(1);
1101
1102 // Get start addr and length of first argument
1103 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1104 Node* arg1_cnt = load_array_length(arg1);
1105
1106 // Get start addr and length of second argument
1107 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1108 Node* arg2_cnt = load_array_length(arg2);
1109
1110 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1111 set_result(result);
1112 return true;
1113 }
1114
1115 //------------------------------inline_string_equals------------------------
1116 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
1117 Node* arg1 = argument(0);
1118 Node* arg2 = argument(1);
1119
1120 // paths (plus control) merge
1121 RegionNode* region = new RegionNode(3);
1122 Node* phi = new PhiNode(region, TypeInt::BOOL);
1123
1124 if (!stopped()) {
1125 // Get start addr and length of first argument
1126 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
1127 Node* arg1_cnt = load_array_length(arg1);
1128
1129 // Get start addr and length of second argument
1130 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
1131 Node* arg2_cnt = load_array_length(arg2);
1132
1133 // Check for arg1_cnt != arg2_cnt
1134 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
1135 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1136 Node* if_ne = generate_slow_guard(bol, NULL);
1137 if (if_ne != NULL) {
1138 phi->init_req(2, intcon(0));
1139 region->init_req(2, if_ne);
1140 }
1141
1142 // Check for count == 0 is done by assembler code for StrEquals.
1143
1144 if (!stopped()) {
1145 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
1146 phi->init_req(1, equals);
1147 region->init_req(1, control());
1148 }
1149 }
1150
1151 // post merge
1152 set_control(_gvn.transform(region));
1153 record_for_igvn(region);
1154
1155 set_result(_gvn.transform(phi));
1156 return true;
1157 }
1158
1159 //------------------------------inline_array_equals----------------------------
1160 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1161 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1162 Node* arg1 = argument(0);
1163 Node* arg2 = argument(1);
1164
1165 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1166 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1167 clear_upper_avx();
1168
1169 return true;
1170 }
1171
1172 //------------------------------inline_hasNegatives------------------------------
1173 bool LibraryCallKit::inline_hasNegatives() {
1174 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1175 return false;
1176 }
1177
1178 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1179 // no receiver since it is static method
1180 Node* ba = argument(0);
1181 Node* offset = argument(1);
1182 Node* len = argument(2);
1183
1184 // Range checks
1185 generate_string_range_check(ba, offset, len, false);
1186 if (stopped()) {
1187 return true;
1188 }
1189 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1190 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1191 set_result(_gvn.transform(result));
1192 return true;
1193 }
1194
1195 bool LibraryCallKit::inline_preconditions_checkIndex() {
1196 Node* index = argument(0);
1197 Node* length = argument(1);
1198 if (too_many_traps(Deoptimization::Reason_intrinsic) || too_many_traps(Deoptimization::Reason_range_check)) {
1199 return false;
1200 }
1201
1202 Node* len_pos_cmp = _gvn.transform(new CmpINode(length, intcon(0)));
1203 Node* len_pos_bol = _gvn.transform(new BoolNode(len_pos_cmp, BoolTest::ge));
1204
1205 {
1206 BuildCutout unless(this, len_pos_bol, PROB_MAX);
1207 uncommon_trap(Deoptimization::Reason_intrinsic,
1208 Deoptimization::Action_make_not_entrant);
1209 }
1210
1211 if (stopped()) {
1212 return false;
1213 }
1214
1215 Node* rc_cmp = _gvn.transform(new CmpUNode(index, length));
1216 BoolTest::mask btest = BoolTest::lt;
1217 Node* rc_bool = _gvn.transform(new BoolNode(rc_cmp, btest));
1218 RangeCheckNode* rc = new RangeCheckNode(control(), rc_bool, PROB_MAX, COUNT_UNKNOWN);
1219 _gvn.set_type(rc, rc->Value(&_gvn));
1220 if (!rc_bool->is_Con()) {
1221 record_for_igvn(rc);
1222 }
1223 set_control(_gvn.transform(new IfTrueNode(rc)));
1224 {
1225 PreserveJVMState pjvms(this);
1226 set_control(_gvn.transform(new IfFalseNode(rc)));
1227 uncommon_trap(Deoptimization::Reason_range_check,
1228 Deoptimization::Action_make_not_entrant);
1229 }
1230
1231 if (stopped()) {
1232 return false;
1233 }
1234
1235 Node* result = new CastIINode(index, TypeInt::make(0, _gvn.type(length)->is_int()->_hi, Type::WidenMax));
1236 result->set_req(0, control());
1237 result = _gvn.transform(result);
1238 set_result(result);
1239 replace_in_map(index, result);
1240 clear_upper_avx();
1241 return true;
1242 }
1243
1244 //------------------------------inline_string_indexOf------------------------
1245 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1246 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1247 return false;
1248 }
1249 Node* src = argument(0);
1250 Node* tgt = argument(1);
1251
1252 // Make the merge point
1253 RegionNode* result_rgn = new RegionNode(4);
1254 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1255
1256 // Get start addr and length of source string
1257 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1258 Node* src_count = load_array_length(src);
1259
1260 // Get start addr and length of substring
1261 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1262 Node* tgt_count = load_array_length(tgt);
1263
1264 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1265 // Divide src size by 2 if String is UTF16 encoded
1266 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1267 }
1268 if (ae == StrIntrinsicNode::UU) {
1269 // Divide substring size by 2 if String is UTF16 encoded
1270 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1271 }
1272
1273 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, result_rgn, result_phi, ae);
1274 if (result != NULL) {
1275 result_phi->init_req(3, result);
1276 result_rgn->init_req(3, control());
1277 }
1278 set_control(_gvn.transform(result_rgn));
1279 record_for_igvn(result_rgn);
1280 set_result(_gvn.transform(result_phi));
1281
1282 return true;
1283 }
1284
1285 //-----------------------------inline_string_indexOf-----------------------
1286 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1287 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1288 return false;
1289 }
1290 if (!Matcher::match_rule_supported(Op_StrIndexOf)) {
1291 return false;
1292 }
1293 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1294 Node* src = argument(0); // byte[]
1295 Node* src_count = argument(1); // char count
1296 Node* tgt = argument(2); // byte[]
1297 Node* tgt_count = argument(3); // char count
1298 Node* from_index = argument(4); // char index
1299
1300 // Multiply byte array index by 2 if String is UTF16 encoded
1301 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1302 src_count = _gvn.transform(new SubINode(src_count, from_index));
1303 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1304 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1305
1306 // Range checks
1307 generate_string_range_check(src, src_offset, src_count, ae != StrIntrinsicNode::LL);
1308 generate_string_range_check(tgt, intcon(0), tgt_count, ae == StrIntrinsicNode::UU);
1309 if (stopped()) {
1310 return true;
1311 }
1312
1313 RegionNode* region = new RegionNode(5);
1314 Node* phi = new PhiNode(region, TypeInt::INT);
1315
1316 Node* result = make_indexOf_node(src_start, src_count, tgt_start, tgt_count, region, phi, ae);
1317 if (result != NULL) {
1318 // The result is index relative to from_index if substring was found, -1 otherwise.
1319 // Generate code which will fold into cmove.
1320 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1321 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1322
1323 Node* if_lt = generate_slow_guard(bol, NULL);
1324 if (if_lt != NULL) {
1325 // result == -1
1326 phi->init_req(3, result);
1327 region->init_req(3, if_lt);
1328 }
1329 if (!stopped()) {
1330 result = _gvn.transform(new AddINode(result, from_index));
1331 phi->init_req(4, result);
1332 region->init_req(4, control());
1333 }
1334 }
1335
1336 set_control(_gvn.transform(region));
1337 record_for_igvn(region);
1338 set_result(_gvn.transform(phi));
1339 clear_upper_avx();
1340
1341 return true;
1342 }
1343
1344 // Create StrIndexOfNode with fast path checks
1345 Node* LibraryCallKit::make_indexOf_node(Node* src_start, Node* src_count, Node* tgt_start, Node* tgt_count,
1346 RegionNode* region, Node* phi, StrIntrinsicNode::ArgEnc ae) {
1347 // Check for substr count > string count
1348 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1349 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1350 Node* if_gt = generate_slow_guard(bol, NULL);
1351 if (if_gt != NULL) {
1352 phi->init_req(1, intcon(-1));
1353 region->init_req(1, if_gt);
1354 }
1355 if (!stopped()) {
1356 // Check for substr count == 0
1357 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1358 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1359 Node* if_zero = generate_slow_guard(bol, NULL);
1360 if (if_zero != NULL) {
1361 phi->init_req(2, intcon(0));
1362 region->init_req(2, if_zero);
1363 }
1364 }
1365 if (!stopped()) {
1366 return make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1367 }
1368 return NULL;
1369 }
1370
1371 //-----------------------------inline_string_indexOfChar-----------------------
1372 bool LibraryCallKit::inline_string_indexOfChar() {
1373 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1374 return false;
1375 }
1376 if (!Matcher::match_rule_supported(Op_StrIndexOfChar)) {
1377 return false;
1378 }
1379 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1380 Node* src = argument(0); // byte[]
1381 Node* tgt = argument(1); // tgt is int ch
1382 Node* from_index = argument(2);
1383 Node* max = argument(3);
1384
1385 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1386 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1387 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1388
1389 // Range checks
1390 generate_string_range_check(src, src_offset, src_count, true);
1391 if (stopped()) {
1392 return true;
1393 }
1394
1395 RegionNode* region = new RegionNode(3);
1396 Node* phi = new PhiNode(region, TypeInt::INT);
1397
1398 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1399 C->set_has_split_ifs(true); // Has chance for split-if optimization
1400 _gvn.transform(result);
1401
1402 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1403 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1404
1405 Node* if_lt = generate_slow_guard(bol, NULL);
1406 if (if_lt != NULL) {
1407 // result == -1
1408 phi->init_req(2, result);
1409 region->init_req(2, if_lt);
1410 }
1411 if (!stopped()) {
1412 result = _gvn.transform(new AddINode(result, from_index));
1413 phi->init_req(1, result);
1414 region->init_req(1, control());
1415 }
1416 set_control(_gvn.transform(region));
1417 record_for_igvn(region);
1418 set_result(_gvn.transform(phi));
1419
1420 return true;
1421 }
1422 //---------------------------inline_string_copy---------------------
1423 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1424 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1425 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1426 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1427 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1428 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1429 bool LibraryCallKit::inline_string_copy(bool compress) {
1430 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1431 return false;
1432 }
1433 int nargs = 5; // 2 oops, 3 ints
1434 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1435
1436 Node* src = argument(0);
1437 Node* src_offset = argument(1);
1438 Node* dst = argument(2);
1439 Node* dst_offset = argument(3);
1440 Node* length = argument(4);
1441
1442 // Check for allocation before we add nodes that would confuse
1443 // tightly_coupled_allocation()
1444 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1445
1446 // Figure out the size and type of the elements we will be copying.
1447 const Type* src_type = src->Value(&_gvn);
1448 const Type* dst_type = dst->Value(&_gvn);
1449 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1450 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1451 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1452 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1453 "Unsupported array types for inline_string_copy");
1454
1455 // Convert char[] offsets to byte[] offsets
1456 bool convert_src = (compress && src_elem == T_BYTE);
1457 bool convert_dst = (!compress && dst_elem == T_BYTE);
1458 if (convert_src) {
1459 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1460 } else if (convert_dst) {
1461 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1462 }
1463
1464 // Range checks
1465 generate_string_range_check(src, src_offset, length, convert_src);
1466 generate_string_range_check(dst, dst_offset, length, convert_dst);
1467 if (stopped()) {
1468 return true;
1469 }
1470
1471 Node* src_start = array_element_address(src, src_offset, src_elem);
1472 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1473 // 'src_start' points to src array + scaled offset
1474 // 'dst_start' points to dst array + scaled offset
1475 Node* count = NULL;
1476 if (compress) {
1477 count = compress_string(src_start, TypeAryPtr::get_array_body_type(src_elem), dst_start, length);
1478 } else {
1479 inflate_string(src_start, dst_start, TypeAryPtr::get_array_body_type(dst_elem), length);
1480 }
1481
1482 if (alloc != NULL) {
1483 if (alloc->maybe_set_complete(&_gvn)) {
1484 // "You break it, you buy it."
1485 InitializeNode* init = alloc->initialization();
1486 assert(init->is_complete(), "we just did this");
1487 init->set_complete_with_arraycopy();
1488 assert(dst->is_CheckCastPP(), "sanity");
1489 assert(dst->in(0)->in(0) == init, "dest pinned");
1490 }
1491 // Do not let stores that initialize this object be reordered with
1492 // a subsequent store that would make this object accessible by
1493 // other threads.
1494 // Record what AllocateNode this StoreStore protects so that
1495 // escape analysis can go from the MemBarStoreStoreNode to the
1496 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1497 // based on the escape status of the AllocateNode.
1498 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1499 }
1500 if (compress) {
1501 set_result(_gvn.transform(count));
1502 }
1503 clear_upper_avx();
1504
1505 return true;
1506 }
1507
1508 #ifdef _LP64
1509 #define XTOP ,top() /*additional argument*/
1510 #else //_LP64
1511 #define XTOP /*no additional argument*/
1512 #endif //_LP64
1513
1514 //------------------------inline_string_toBytesU--------------------------
1515 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1516 bool LibraryCallKit::inline_string_toBytesU() {
1517 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1518 return false;
1519 }
1520 // Get the arguments.
1521 Node* value = argument(0);
1522 Node* offset = argument(1);
1523 Node* length = argument(2);
1524
1525 Node* newcopy = NULL;
1526
1527 // Set the original stack and the reexecute bit for the interpreter to reexecute
1528 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1529 { PreserveReexecuteState preexecs(this);
1530 jvms()->set_should_reexecute(true);
1531
1532 // Check if a null path was taken unconditionally.
1533 value = null_check(value);
1534
1535 RegionNode* bailout = new RegionNode(1);
1536 record_for_igvn(bailout);
1537
1538 // Range checks
1539 generate_negative_guard(offset, bailout);
1540 generate_negative_guard(length, bailout);
1541 generate_limit_guard(offset, length, load_array_length(value), bailout);
1542 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1543 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1544
1545 if (bailout->req() > 1) {
1546 PreserveJVMState pjvms(this);
1547 set_control(_gvn.transform(bailout));
1548 uncommon_trap(Deoptimization::Reason_intrinsic,
1549 Deoptimization::Action_maybe_recompile);
1550 }
1551 if (stopped()) {
1552 return true;
1553 }
1554
1555 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1556 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1557 newcopy = new_array(klass_node, size, 0); // no arguments to push
1558 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1559
1560 // Calculate starting addresses.
1561 Node* src_start = array_element_address(value, offset, T_CHAR);
1562 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1563
1564 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1565 const TypeInt* toffset = gvn().type(offset)->is_int();
1566 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1567
1568 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1569 const char* copyfunc_name = "arraycopy";
1570 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1571 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1572 OptoRuntime::fast_arraycopy_Type(),
1573 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1574 src_start, dst_start, ConvI2X(length) XTOP);
1575 // Do not let reads from the cloned object float above the arraycopy.
1576 if (alloc != NULL) {
1577 if (alloc->maybe_set_complete(&_gvn)) {
1578 // "You break it, you buy it."
1579 InitializeNode* init = alloc->initialization();
1580 assert(init->is_complete(), "we just did this");
1581 init->set_complete_with_arraycopy();
1582 assert(newcopy->is_CheckCastPP(), "sanity");
1583 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1584 }
1585 // Do not let stores that initialize this object be reordered with
1586 // a subsequent store that would make this object accessible by
1587 // other threads.
1588 // Record what AllocateNode this StoreStore protects so that
1589 // escape analysis can go from the MemBarStoreStoreNode to the
1590 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1591 // based on the escape status of the AllocateNode.
1592 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1593 } else {
1594 insert_mem_bar(Op_MemBarCPUOrder);
1595 }
1596 } // original reexecute is set back here
1597
1598 C->set_has_split_ifs(true); // Has chance for split-if optimization
1599 if (!stopped()) {
1600 set_result(newcopy);
1601 }
1602 clear_upper_avx();
1603
1604 return true;
1605 }
1606
1607 //------------------------inline_string_getCharsU--------------------------
1608 // public void StringUTF16.getChars(byte[] src, int srcBegin, int srcEnd, char dst[], int dstBegin)
1609 bool LibraryCallKit::inline_string_getCharsU() {
1610 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
1611 return false;
1612 }
1613
1614 // Get the arguments.
1615 Node* src = argument(0);
1616 Node* src_begin = argument(1);
1617 Node* src_end = argument(2); // exclusive offset (i < src_end)
1618 Node* dst = argument(3);
1619 Node* dst_begin = argument(4);
1620
1621 // Check for allocation before we add nodes that would confuse
1622 // tightly_coupled_allocation()
1623 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1624
1625 // Check if a null path was taken unconditionally.
1626 src = null_check(src);
1627 dst = null_check(dst);
1628 if (stopped()) {
1629 return true;
1630 }
1631
1632 // Get length and convert char[] offset to byte[] offset
1633 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1634 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1635
1636 // Range checks
1637 generate_string_range_check(src, src_begin, length, true);
1638 generate_string_range_check(dst, dst_begin, length, false);
1639 if (stopped()) {
1640 return true;
1641 }
1642
1643 if (!stopped()) {
1644 // Calculate starting addresses.
1645 Node* src_start = array_element_address(src, src_begin, T_BYTE);
1646 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1647
1648 // Check if array addresses are aligned to HeapWordSize
1649 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1650 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1651 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1652 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1653
1654 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1655 const char* copyfunc_name = "arraycopy";
1656 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1657 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1658 OptoRuntime::fast_arraycopy_Type(),
1659 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1660 src_start, dst_start, ConvI2X(length) XTOP);
1661 // Do not let reads from the cloned object float above the arraycopy.
1662 if (alloc != NULL) {
1663 if (alloc->maybe_set_complete(&_gvn)) {
1664 // "You break it, you buy it."
1665 InitializeNode* init = alloc->initialization();
1666 assert(init->is_complete(), "we just did this");
1667 init->set_complete_with_arraycopy();
1668 assert(dst->is_CheckCastPP(), "sanity");
1669 assert(dst->in(0)->in(0) == init, "dest pinned");
1670 }
1671 // Do not let stores that initialize this object be reordered with
1672 // a subsequent store that would make this object accessible by
1673 // other threads.
1674 // Record what AllocateNode this StoreStore protects so that
1675 // escape analysis can go from the MemBarStoreStoreNode to the
1676 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1677 // based on the escape status of the AllocateNode.
1678 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1679 } else {
1680 insert_mem_bar(Op_MemBarCPUOrder);
1681 }
1682 }
1683
1684 C->set_has_split_ifs(true); // Has chance for split-if optimization
1685 return true;
1686 }
1687
1688 //----------------------inline_string_char_access----------------------------
1689 // Store/Load char to/from byte[] array.
1690 // static void StringUTF16.putChar(byte[] val, int index, int c)
1691 // static char StringUTF16.getChar(byte[] val, int index)
1692 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1693 Node* value = argument(0);
1694 Node* index = argument(1);
1695 Node* ch = is_store ? argument(2) : NULL;
1696
1697 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1698 // correctly requires matched array shapes.
1699 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1700 "sanity: byte[] and char[] bases agree");
1701 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1702 "sanity: byte[] and char[] scales agree");
1703
1704 // Bail when getChar over constants is requested: constant folding would
1705 // reject folding mismatched char access over byte[]. A normal inlining for getChar
1706 // Java method would constant fold nicely instead.
1707 if (!is_store && value->is_Con() && index->is_Con()) {
1708 return false;
1709 }
1710
1711 Node* adr = array_element_address(value, index, T_CHAR);
1712 if (adr->is_top()) {
1713 return false;
1714 }
1715 if (is_store) {
1716 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1717 false, false, true /* mismatched */);
1718 } else {
1719 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered,
1720 LoadNode::DependsOnlyOnTest, false, false, true /* mismatched */);
1721 set_result(ch);
1722 }
1723 return true;
1724 }
1725
1726 //--------------------------round_double_node--------------------------------
1727 // Round a double node if necessary.
1728 Node* LibraryCallKit::round_double_node(Node* n) {
1729 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1730 n = _gvn.transform(new RoundDoubleNode(0, n));
1731 return n;
1732 }
1733
1734 //------------------------------inline_math-----------------------------------
1735 // public static double Math.abs(double)
1736 // public static double Math.sqrt(double)
1737 // public static double Math.log(double)
1738 // public static double Math.log10(double)
1739 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1740 Node* arg = round_double_node(argument(0));
1741 Node* n = NULL;
1742 switch (id) {
1743 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1744 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1745 default: fatal_unexpected_iid(id); break;
1746 }
1747 set_result(_gvn.transform(n));
1748 return true;
1749 }
1750
1751 //------------------------------runtime_math-----------------------------
1752 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1753 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1754 "must be (DD)D or (D)D type");
1755
1756 // Inputs
1757 Node* a = round_double_node(argument(0));
1758 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1759
1760 const TypePtr* no_memory_effects = NULL;
1761 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1762 no_memory_effects,
1763 a, top(), b, b ? top() : NULL);
1764 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1765 #ifdef ASSERT
1766 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1767 assert(value_top == top(), "second value must be top");
1768 #endif
1769
1770 set_result(value);
1771 return true;
1772 }
1773
1774 //------------------------------inline_math_native-----------------------------
1775 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1776 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1777 switch (id) {
1778 // These intrinsics are not properly supported on all hardware
1779 case vmIntrinsics::_dsin:
1780 return StubRoutines::dsin() != NULL ?
1781 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dsin(), "dsin") :
1782 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1783 case vmIntrinsics::_dcos:
1784 return StubRoutines::dcos() != NULL ?
1785 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dcos(), "dcos") :
1786 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1787 case vmIntrinsics::_dtan:
1788 return StubRoutines::dtan() != NULL ?
1789 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dtan(), "dtan") :
1790 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1791 case vmIntrinsics::_dlog:
1792 return StubRoutines::dlog() != NULL ?
1793 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1794 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1795 case vmIntrinsics::_dlog10:
1796 return StubRoutines::dlog10() != NULL ?
1797 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog10(), "dlog10") :
1798 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1799
1800 // These intrinsics are supported on all hardware
1801 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1802 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1803
1804 case vmIntrinsics::_dexp:
1805 return StubRoutines::dexp() != NULL ?
1806 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1807 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1808 case vmIntrinsics::_dpow: {
1809 Node* exp = round_double_node(argument(2));
1810 const TypeD* d = _gvn.type(exp)->isa_double_constant();
1811 if (d != NULL && d->getd() == 2.0) {
1812 // Special case: pow(x, 2.0) => x * x
1813 Node* base = round_double_node(argument(0));
1814 set_result(_gvn.transform(new MulDNode(base, base)));
1815 return true;
1816 }
1817 return StubRoutines::dexp() != NULL ?
1818 runtime_math(OptoRuntime::Math_DD_D_Type(), StubRoutines::dpow(), "dpow") :
1819 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1820 }
1821 #undef FN_PTR
1822
1823 // These intrinsics are not yet correctly implemented
1824 case vmIntrinsics::_datan2:
1825 return false;
1826
1827 default:
1828 fatal_unexpected_iid(id);
1829 return false;
1830 }
1831 }
1832
1833 static bool is_simple_name(Node* n) {
1834 return (n->req() == 1 // constant
1835 || (n->is_Type() && n->as_Type()->type()->singleton())
1836 || n->is_Proj() // parameter or return value
1837 || n->is_Phi() // local of some sort
1838 );
1839 }
1840
1841 //----------------------------inline_notify-----------------------------------*
1842 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1843 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1844 address func;
1845 if (id == vmIntrinsics::_notify) {
1846 func = OptoRuntime::monitor_notify_Java();
1847 } else {
1848 func = OptoRuntime::monitor_notifyAll_Java();
1849 }
1850 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1851 make_slow_call_ex(call, env()->Throwable_klass(), false);
1852 return true;
1853 }
1854
1855
1856 //----------------------------inline_min_max-----------------------------------
1857 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1858 set_result(generate_min_max(id, argument(0), argument(1)));
1859 return true;
1860 }
1861
1862 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1863 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1864 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1865 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1866 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1867
1868 {
1869 PreserveJVMState pjvms(this);
1870 PreserveReexecuteState preexecs(this);
1871 jvms()->set_should_reexecute(true);
1872
1873 set_control(slow_path);
1874 set_i_o(i_o());
1875
1876 uncommon_trap(Deoptimization::Reason_intrinsic,
1877 Deoptimization::Action_none);
1878 }
1879
1880 set_control(fast_path);
1881 set_result(math);
1882 }
1883
1884 template <typename OverflowOp>
1885 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1886 typedef typename OverflowOp::MathOp MathOp;
1887
1888 MathOp* mathOp = new MathOp(arg1, arg2);
1889 Node* operation = _gvn.transform( mathOp );
1890 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1891 inline_math_mathExact(operation, ofcheck);
1892 return true;
1893 }
1894
1895 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1896 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1897 }
1898
1899 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1900 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1901 }
1902
1903 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1904 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1905 }
1906
1907 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1908 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1909 }
1910
1911 bool LibraryCallKit::inline_math_negateExactI() {
1912 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1913 }
1914
1915 bool LibraryCallKit::inline_math_negateExactL() {
1916 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1917 }
1918
1919 bool LibraryCallKit::inline_math_multiplyExactI() {
1920 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1921 }
1922
1923 bool LibraryCallKit::inline_math_multiplyExactL() {
1924 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1925 }
1926
1927 bool LibraryCallKit::inline_math_multiplyHigh() {
1928 set_result(_gvn.transform(new MulHiLNode(argument(0), argument(2))));
1929 return true;
1930 }
1931
1932 Node*
1933 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1934 // These are the candidate return value:
1935 Node* xvalue = x0;
1936 Node* yvalue = y0;
1937
1938 if (xvalue == yvalue) {
1939 return xvalue;
1940 }
1941
1942 bool want_max = (id == vmIntrinsics::_max);
1943
1944 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1945 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1946 if (txvalue == NULL || tyvalue == NULL) return top();
1947 // This is not really necessary, but it is consistent with a
1948 // hypothetical MaxINode::Value method:
1949 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1950
1951 // %%% This folding logic should (ideally) be in a different place.
1952 // Some should be inside IfNode, and there to be a more reliable
1953 // transformation of ?: style patterns into cmoves. We also want
1954 // more powerful optimizations around cmove and min/max.
1955
1956 // Try to find a dominating comparison of these guys.
1957 // It can simplify the index computation for Arrays.copyOf
1958 // and similar uses of System.arraycopy.
1959 // First, compute the normalized version of CmpI(x, y).
1960 int cmp_op = Op_CmpI;
1961 Node* xkey = xvalue;
1962 Node* ykey = yvalue;
1963 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1964 if (ideal_cmpxy->is_Cmp()) {
1965 // E.g., if we have CmpI(length - offset, count),
1966 // it might idealize to CmpI(length, count + offset)
1967 cmp_op = ideal_cmpxy->Opcode();
1968 xkey = ideal_cmpxy->in(1);
1969 ykey = ideal_cmpxy->in(2);
1970 }
1971
1972 // Start by locating any relevant comparisons.
1973 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1974 Node* cmpxy = NULL;
1975 Node* cmpyx = NULL;
1976 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1977 Node* cmp = start_from->fast_out(k);
1978 if (cmp->outcnt() > 0 && // must have prior uses
1979 cmp->in(0) == NULL && // must be context-independent
1980 cmp->Opcode() == cmp_op) { // right kind of compare
1981 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1982 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1983 }
1984 }
1985
1986 const int NCMPS = 2;
1987 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1988 int cmpn;
1989 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1990 if (cmps[cmpn] != NULL) break; // find a result
1991 }
1992 if (cmpn < NCMPS) {
1993 // Look for a dominating test that tells us the min and max.
1994 int depth = 0; // Limit search depth for speed
1995 Node* dom = control();
1996 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1997 if (++depth >= 100) break;
1998 Node* ifproj = dom;
1999 if (!ifproj->is_Proj()) continue;
2000 Node* iff = ifproj->in(0);
2001 if (!iff->is_If()) continue;
2002 Node* bol = iff->in(1);
2003 if (!bol->is_Bool()) continue;
2004 Node* cmp = bol->in(1);
2005 if (cmp == NULL) continue;
2006 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2007 if (cmps[cmpn] == cmp) break;
2008 if (cmpn == NCMPS) continue;
2009 BoolTest::mask btest = bol->as_Bool()->_test._test;
2010 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2011 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2012 // At this point, we know that 'x btest y' is true.
2013 switch (btest) {
2014 case BoolTest::eq:
2015 // They are proven equal, so we can collapse the min/max.
2016 // Either value is the answer. Choose the simpler.
2017 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2018 return yvalue;
2019 return xvalue;
2020 case BoolTest::lt: // x < y
2021 case BoolTest::le: // x <= y
2022 return (want_max ? yvalue : xvalue);
2023 case BoolTest::gt: // x > y
2024 case BoolTest::ge: // x >= y
2025 return (want_max ? xvalue : yvalue);
2026 default:
2027 break;
2028 }
2029 }
2030 }
2031
2032 // We failed to find a dominating test.
2033 // Let's pick a test that might GVN with prior tests.
2034 Node* best_bol = NULL;
2035 BoolTest::mask best_btest = BoolTest::illegal;
2036 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2037 Node* cmp = cmps[cmpn];
2038 if (cmp == NULL) continue;
2039 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2040 Node* bol = cmp->fast_out(j);
2041 if (!bol->is_Bool()) continue;
2042 BoolTest::mask btest = bol->as_Bool()->_test._test;
2043 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2044 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2045 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2046 best_bol = bol->as_Bool();
2047 best_btest = btest;
2048 }
2049 }
2050 }
2051
2052 Node* answer_if_true = NULL;
2053 Node* answer_if_false = NULL;
2054 switch (best_btest) {
2055 default:
2056 if (cmpxy == NULL)
2057 cmpxy = ideal_cmpxy;
2058 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2059 // and fall through:
2060 case BoolTest::lt: // x < y
2061 case BoolTest::le: // x <= y
2062 answer_if_true = (want_max ? yvalue : xvalue);
2063 answer_if_false = (want_max ? xvalue : yvalue);
2064 break;
2065 case BoolTest::gt: // x > y
2066 case BoolTest::ge: // x >= y
2067 answer_if_true = (want_max ? xvalue : yvalue);
2068 answer_if_false = (want_max ? yvalue : xvalue);
2069 break;
2070 }
2071
2072 jint hi, lo;
2073 if (want_max) {
2074 // We can sharpen the minimum.
2075 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2076 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2077 } else {
2078 // We can sharpen the maximum.
2079 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2080 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2081 }
2082
2083 // Use a flow-free graph structure, to avoid creating excess control edges
2084 // which could hinder other optimizations.
2085 // Since Math.min/max is often used with arraycopy, we want
2086 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2087 Node* cmov = CMoveNode::make(NULL, best_bol,
2088 answer_if_false, answer_if_true,
2089 TypeInt::make(lo, hi, widen));
2090
2091 return _gvn.transform(cmov);
2092
2093 /*
2094 // This is not as desirable as it may seem, since Min and Max
2095 // nodes do not have a full set of optimizations.
2096 // And they would interfere, anyway, with 'if' optimizations
2097 // and with CMoveI canonical forms.
2098 switch (id) {
2099 case vmIntrinsics::_min:
2100 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2101 case vmIntrinsics::_max:
2102 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2103 default:
2104 ShouldNotReachHere();
2105 }
2106 */
2107 }
2108
2109 inline int
2110 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset, BasicType type) {
2111 const TypePtr* base_type = TypePtr::NULL_PTR;
2112 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2113 if (base_type == NULL) {
2114 // Unknown type.
2115 return Type::AnyPtr;
2116 } else if (base_type == TypePtr::NULL_PTR) {
2117 // Since this is a NULL+long form, we have to switch to a rawptr.
2118 base = _gvn.transform(new CastX2PNode(offset));
2119 offset = MakeConX(0);
2120 return Type::RawPtr;
2121 } else if (base_type->base() == Type::RawPtr) {
2122 return Type::RawPtr;
2123 } else if (base_type->isa_oopptr()) {
2124 // Base is never null => always a heap address.
2125 if (!TypePtr::NULL_PTR->higher_equal(base_type)) {
2126 return Type::OopPtr;
2127 }
2128 // Offset is small => always a heap address.
2129 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2130 if (offset_type != NULL &&
2131 base_type->offset() == 0 && // (should always be?)
2132 offset_type->_lo >= 0 &&
2133 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2134 return Type::OopPtr;
2135 } else if (type == T_OBJECT) {
2136 // off heap access to an oop doesn't make any sense. Has to be on
2137 // heap.
2138 return Type::OopPtr;
2139 }
2140 // Otherwise, it might either be oop+off or NULL+addr.
2141 return Type::AnyPtr;
2142 } else {
2143 // No information:
2144 return Type::AnyPtr;
2145 }
2146 }
2147
2148 inline Node* LibraryCallKit::make_unsafe_address(Node*& base, Node* offset, BasicType type, bool can_cast) {
2149 Node* uncasted_base = base;
2150 int kind = classify_unsafe_addr(uncasted_base, offset, type);
2151 if (kind == Type::RawPtr) {
2152 return basic_plus_adr(top(), uncasted_base, offset);
2153 } else if (kind == Type::AnyPtr) {
2154 assert(base == uncasted_base, "unexpected base change");
2155 if (can_cast) {
2156 if (!_gvn.type(base)->speculative_maybe_null() &&
2157 !too_many_traps(Deoptimization::Reason_speculate_null_check)) {
2158 // According to profiling, this access is always on
2159 // heap. Casting the base to not null and thus avoiding membars
2160 // around the access should allow better optimizations
2161 Node* null_ctl = top();
2162 base = null_check_oop(base, &null_ctl, true, true, true);
2163 assert(null_ctl->is_top(), "no null control here");
2164 return basic_plus_adr(base, offset);
2165 } else if (_gvn.type(base)->speculative_always_null() &&
2166 !too_many_traps(Deoptimization::Reason_speculate_null_assert)) {
2167 // According to profiling, this access is always off
2168 // heap.
2169 base = null_assert(base);
2170 Node* raw_base = _gvn.transform(new CastX2PNode(offset));
2171 offset = MakeConX(0);
2172 return basic_plus_adr(top(), raw_base, offset);
2173 }
2174 }
2175 // We don't know if it's an on heap or off heap access. Fall back
2176 // to raw memory access.
2177 Node* raw = _gvn.transform(new CheckCastPPNode(control(), base, TypeRawPtr::BOTTOM));
2178 return basic_plus_adr(top(), raw, offset);
2179 } else {
2180 assert(base == uncasted_base, "unexpected base change");
2181 // We know it's an on heap access so base can't be null
2182 if (TypePtr::NULL_PTR->higher_equal(_gvn.type(base))) {
2183 base = must_be_not_null(base, true);
2184 }
2185 return basic_plus_adr(base, offset);
2186 }
2187 }
2188
2189 //--------------------------inline_number_methods-----------------------------
2190 // inline int Integer.numberOfLeadingZeros(int)
2191 // inline int Long.numberOfLeadingZeros(long)
2192 //
2193 // inline int Integer.numberOfTrailingZeros(int)
2194 // inline int Long.numberOfTrailingZeros(long)
2195 //
2196 // inline int Integer.bitCount(int)
2197 // inline int Long.bitCount(long)
2198 //
2199 // inline char Character.reverseBytes(char)
2200 // inline short Short.reverseBytes(short)
2201 // inline int Integer.reverseBytes(int)
2202 // inline long Long.reverseBytes(long)
2203 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2204 Node* arg = argument(0);
2205 Node* n = NULL;
2206 switch (id) {
2207 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2208 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2209 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2210 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2211 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2212 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2213 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2214 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2215 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2216 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2217 default: fatal_unexpected_iid(id); break;
2218 }
2219 set_result(_gvn.transform(n));
2220 return true;
2221 }
2222
2223 //----------------------------inline_unsafe_access----------------------------
2224
2225 // Helper that guards and inserts a pre-barrier.
2226 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2227 Node* pre_val, bool need_mem_bar) {
2228 // We could be accessing the referent field of a reference object. If so, when G1
2229 // is enabled, we need to log the value in the referent field in an SATB buffer.
2230 // This routine performs some compile time filters and generates suitable
2231 // runtime filters that guard the pre-barrier code.
2232 // Also add memory barrier for non volatile load from the referent field
2233 // to prevent commoning of loads across safepoint.
2234 if (!UseG1GC && !need_mem_bar)
2235 return;
2236
2237 // Some compile time checks.
2238
2239 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2240 const TypeX* otype = offset->find_intptr_t_type();
2241 if (otype != NULL && otype->is_con() &&
2242 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2243 // Constant offset but not the reference_offset so just return
2244 return;
2245 }
2246
2247 // We only need to generate the runtime guards for instances.
2248 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2249 if (btype != NULL) {
2250 if (btype->isa_aryptr()) {
2251 // Array type so nothing to do
2252 return;
2253 }
2254
2255 const TypeInstPtr* itype = btype->isa_instptr();
2256 if (itype != NULL) {
2257 // Can the klass of base_oop be statically determined to be
2258 // _not_ a sub-class of Reference and _not_ Object?
2259 ciKlass* klass = itype->klass();
2260 if ( klass->is_loaded() &&
2261 !klass->is_subtype_of(env()->Reference_klass()) &&
2262 !env()->Object_klass()->is_subtype_of(klass)) {
2263 return;
2264 }
2265 }
2266 }
2267
2268 // The compile time filters did not reject base_oop/offset so
2269 // we need to generate the following runtime filters
2270 //
2271 // if (offset == java_lang_ref_Reference::_reference_offset) {
2272 // if (instance_of(base, java.lang.ref.Reference)) {
2273 // pre_barrier(_, pre_val, ...);
2274 // }
2275 // }
2276
2277 float likely = PROB_LIKELY( 0.999);
2278 float unlikely = PROB_UNLIKELY(0.999);
2279
2280 IdealKit ideal(this);
2281 #define __ ideal.
2282
2283 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2284
2285 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2286 // Update graphKit memory and control from IdealKit.
2287 sync_kit(ideal);
2288
2289 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2290 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2291
2292 // Update IdealKit memory and control from graphKit.
2293 __ sync_kit(this);
2294
2295 Node* one = __ ConI(1);
2296 // is_instof == 0 if base_oop == NULL
2297 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2298
2299 // Update graphKit from IdeakKit.
2300 sync_kit(ideal);
2301
2302 // Use the pre-barrier to record the value in the referent field
2303 pre_barrier(false /* do_load */,
2304 __ ctrl(),
2305 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2306 pre_val /* pre_val */,
2307 T_OBJECT);
2308 if (need_mem_bar) {
2309 // Add memory barrier to prevent commoning reads from this field
2310 // across safepoint since GC can change its value.
2311 insert_mem_bar(Op_MemBarCPUOrder);
2312 }
2313 // Update IdealKit from graphKit.
2314 __ sync_kit(this);
2315
2316 } __ end_if(); // _ref_type != ref_none
2317 } __ end_if(); // offset == referent_offset
2318
2319 // Final sync IdealKit and GraphKit.
2320 final_sync(ideal);
2321 #undef __
2322 }
2323
2324
2325 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type) {
2326 // Attempt to infer a sharper value type from the offset and base type.
2327 ciKlass* sharpened_klass = NULL;
2328
2329 // See if it is an instance field, with an object type.
2330 if (alias_type->field() != NULL) {
2331 if (alias_type->field()->type()->is_klass()) {
2332 sharpened_klass = alias_type->field()->type()->as_klass();
2333 }
2334 }
2335
2336 // See if it is a narrow oop array.
2337 if (adr_type->isa_aryptr()) {
2338 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2339 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2340 if (elem_type != NULL) {
2341 sharpened_klass = elem_type->klass();
2342 }
2343 }
2344 }
2345
2346 // The sharpened class might be unloaded if there is no class loader
2347 // contraint in place.
2348 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2349 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2350
2351 #ifndef PRODUCT
2352 if (C->print_intrinsics() || C->print_inlining()) {
2353 tty->print(" from base type: "); adr_type->dump(); tty->cr();
2354 tty->print(" sharpened value: "); tjp->dump(); tty->cr();
2355 }
2356 #endif
2357 // Sharpen the value type.
2358 return tjp;
2359 }
2360 return NULL;
2361 }
2362
2363 bool LibraryCallKit::inline_unsafe_access(bool is_store, const BasicType type, const AccessKind kind, const bool unaligned) {
2364 if (callee()->is_static()) return false; // caller must have the capability!
2365 guarantee(!is_store || kind != Acquire, "Acquire accesses can be produced only for loads");
2366 guarantee( is_store || kind != Release, "Release accesses can be produced only for stores");
2367 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2368
2369 #ifndef PRODUCT
2370 {
2371 ResourceMark rm;
2372 // Check the signatures.
2373 ciSignature* sig = callee()->signature();
2374 #ifdef ASSERT
2375 if (!is_store) {
2376 // Object getObject(Object base, int/long offset), etc.
2377 BasicType rtype = sig->return_type()->basic_type();
2378 assert(rtype == type, "getter must return the expected value");
2379 assert(sig->count() == 2, "oop getter has 2 arguments");
2380 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2381 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2382 } else {
2383 // void putObject(Object base, int/long offset, Object x), etc.
2384 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2385 assert(sig->count() == 3, "oop putter has 3 arguments");
2386 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2387 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2388 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2389 assert(vtype == type, "putter must accept the expected value");
2390 }
2391 #endif // ASSERT
2392 }
2393 #endif //PRODUCT
2394
2395 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2396
2397 Node* receiver = argument(0); // type: oop
2398
2399 // Build address expression.
2400 Node* adr;
2401 Node* heap_base_oop = top();
2402 Node* offset = top();
2403 Node* val;
2404
2405 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2406 Node* base = argument(1); // type: oop
2407 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2408 offset = argument(2); // type: long
2409 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2410 // to be plain byte offsets, which are also the same as those accepted
2411 // by oopDesc::field_base.
2412 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2413 "fieldOffset must be byte-scaled");
2414 // 32-bit machines ignore the high half!
2415 offset = ConvL2X(offset);
2416 adr = make_unsafe_address(base, offset, type, kind == Relaxed);
2417
2418 if (_gvn.type(base)->isa_ptr() != TypePtr::NULL_PTR) {
2419 heap_base_oop = base;
2420 } else if (type == T_OBJECT) {
2421 return false; // off-heap oop accesses are not supported
2422 }
2423
2424 // Can base be NULL? Otherwise, always on-heap access.
2425 bool can_access_non_heap = TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop));
2426
2427 val = is_store ? argument(4) : NULL;
2428
2429 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2430
2431 // Try to categorize the address.
2432 Compile::AliasType* alias_type = C->alias_type(adr_type);
2433 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2434
2435 if (alias_type->adr_type() == TypeInstPtr::KLASS ||
2436 alias_type->adr_type() == TypeAryPtr::RANGE) {
2437 return false; // not supported
2438 }
2439
2440 bool mismatched = false;
2441 BasicType bt = alias_type->basic_type();
2442 if (bt != T_ILLEGAL) {
2443 assert(alias_type->adr_type()->is_oopptr(), "should be on-heap access");
2444 if (bt == T_BYTE && adr_type->isa_aryptr()) {
2445 // Alias type doesn't differentiate between byte[] and boolean[]).
2446 // Use address type to get the element type.
2447 bt = adr_type->is_aryptr()->elem()->array_element_basic_type();
2448 }
2449 if (bt == T_ARRAY || bt == T_NARROWOOP) {
2450 // accessing an array field with getObject is not a mismatch
2451 bt = T_OBJECT;
2452 }
2453 if ((bt == T_OBJECT) != (type == T_OBJECT)) {
2454 // Don't intrinsify mismatched object accesses
2455 return false;
2456 }
2457 mismatched = (bt != type);
2458 } else if (alias_type->adr_type()->isa_oopptr()) {
2459 mismatched = true; // conservatively mark all "wide" on-heap accesses as mismatched
2460 }
2461
2462 assert(!mismatched || alias_type->adr_type()->is_oopptr(), "off-heap access can't be mismatched");
2463
2464 // First guess at the value type.
2465 const Type *value_type = Type::get_const_basic_type(type);
2466
2467 // We will need memory barriers unless we can determine a unique
2468 // alias category for this reference. (Note: If for some reason
2469 // the barriers get omitted and the unsafe reference begins to "pollute"
2470 // the alias analysis of the rest of the graph, either Compile::can_alias
2471 // or Compile::must_alias will throw a diagnostic assert.)
2472 bool need_mem_bar = false;
2473 switch (kind) {
2474 case Relaxed:
2475 need_mem_bar = (mismatched && !adr_type->isa_aryptr()) || can_access_non_heap;
2476 break;
2477 case Opaque:
2478 // Opaque uses CPUOrder membars for protection against code movement.
2479 case Acquire:
2480 case Release:
2481 case Volatile:
2482 need_mem_bar = true;
2483 break;
2484 default:
2485 ShouldNotReachHere();
2486 }
2487
2488 // Some accesses require access atomicity for all types, notably longs and doubles.
2489 // When AlwaysAtomicAccesses is enabled, all accesses are atomic.
2490 bool requires_atomic_access = false;
2491 switch (kind) {
2492 case Relaxed:
2493 requires_atomic_access = AlwaysAtomicAccesses;
2494 break;
2495 case Opaque:
2496 // Opaque accesses are atomic.
2497 case Acquire:
2498 case Release:
2499 case Volatile:
2500 requires_atomic_access = true;
2501 break;
2502 default:
2503 ShouldNotReachHere();
2504 }
2505
2506 // Figure out the memory ordering.
2507 // Acquire/Release/Volatile accesses require marking the loads/stores with MemOrd
2508 MemNode::MemOrd mo = access_kind_to_memord_LS(kind, is_store);
2509
2510 // If we are reading the value of the referent field of a Reference
2511 // object (either by using Unsafe directly or through reflection)
2512 // then, if G1 is enabled, we need to record the referent in an
2513 // SATB log buffer using the pre-barrier mechanism.
2514 // Also we need to add memory barrier to prevent commoning reads
2515 // from this field across safepoint since GC can change its value.
2516 bool need_read_barrier = !is_store &&
2517 offset != top() && heap_base_oop != top();
2518
2519 if (!is_store && type == T_OBJECT) {
2520 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2521 if (tjp != NULL) {
2522 value_type = tjp;
2523 }
2524 }
2525
2526 receiver = null_check(receiver);
2527 if (stopped()) {
2528 return true;
2529 }
2530 // Heap pointers get a null-check from the interpreter,
2531 // as a courtesy. However, this is not guaranteed by Unsafe,
2532 // and it is not possible to fully distinguish unintended nulls
2533 // from intended ones in this API.
2534
2535 // We need to emit leading and trailing CPU membars (see below) in
2536 // addition to memory membars for special access modes. This is a little
2537 // too strong, but avoids the need to insert per-alias-type
2538 // volatile membars (for stores; compare Parse::do_put_xxx), which
2539 // we cannot do effectively here because we probably only have a
2540 // rough approximation of type.
2541
2542 switch(kind) {
2543 case Relaxed:
2544 case Opaque:
2545 case Acquire:
2546 break;
2547 case Release:
2548 case Volatile:
2549 if (is_store) {
2550 insert_mem_bar(Op_MemBarRelease);
2551 } else {
2552 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2553 insert_mem_bar(Op_MemBarVolatile);
2554 }
2555 }
2556 break;
2557 default:
2558 ShouldNotReachHere();
2559 }
2560
2561 // Memory barrier to prevent normal and 'unsafe' accesses from
2562 // bypassing each other. Happens after null checks, so the
2563 // exception paths do not take memory state from the memory barrier,
2564 // so there's no problems making a strong assert about mixing users
2565 // of safe & unsafe memory.
2566 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2567
2568 if (!is_store) {
2569 Node* p = NULL;
2570 // Try to constant fold a load from a constant field
2571 ciField* field = alias_type->field();
2572 if (heap_base_oop != top() && field != NULL && field->is_constant() && !mismatched) {
2573 // final or stable field
2574 p = make_constant_from_field(field, heap_base_oop);
2575 }
2576 if (p == NULL) {
2577 // To be valid, unsafe loads may depend on other conditions than
2578 // the one that guards them: pin the Load node
2579 LoadNode::ControlDependency dep = LoadNode::Pinned;
2580 Node* ctrl = control();
2581 if (adr_type->isa_instptr()) {
2582 assert(adr_type->meet(TypePtr::NULL_PTR) != adr_type->remove_speculative(), "should be not null");
2583 intptr_t offset = Type::OffsetBot;
2584 AddPNode::Ideal_base_and_offset(adr, &_gvn, offset);
2585 if (offset >= 0) {
2586 int s = Klass::layout_helper_size_in_bytes(adr_type->isa_instptr()->klass()->layout_helper());
2587 if (offset < s) {
2588 // Guaranteed to be a valid access, no need to pin it
2589 dep = LoadNode::DependsOnlyOnTest;
2590 ctrl = NULL;
2591 }
2592 }
2593 }
2594 p = make_load(ctrl, adr, value_type, type, adr_type, mo, dep, requires_atomic_access, unaligned, mismatched);
2595 // load value
2596 switch (type) {
2597 case T_BOOLEAN:
2598 {
2599 // Normalize the value returned by getBoolean in the following cases
2600 if (mismatched ||
2601 heap_base_oop == top() || // - heap_base_oop is NULL or
2602 (can_access_non_heap && alias_type->field() == NULL) // - heap_base_oop is potentially NULL
2603 // and the unsafe access is made to large offset
2604 // (i.e., larger than the maximum offset necessary for any
2605 // field access)
2606 ) {
2607 IdealKit ideal = IdealKit(this);
2608 #define __ ideal.
2609 IdealVariable normalized_result(ideal);
2610 __ declarations_done();
2611 __ set(normalized_result, p);
2612 __ if_then(p, BoolTest::ne, ideal.ConI(0));
2613 __ set(normalized_result, ideal.ConI(1));
2614 ideal.end_if();
2615 final_sync(ideal);
2616 p = __ value(normalized_result);
2617 #undef __
2618 }
2619 }
2620 case T_CHAR:
2621 case T_BYTE:
2622 case T_SHORT:
2623 case T_INT:
2624 case T_LONG:
2625 case T_FLOAT:
2626 case T_DOUBLE:
2627 break;
2628 case T_OBJECT:
2629 if (need_read_barrier) {
2630 // We do not require a mem bar inside pre_barrier if need_mem_bar
2631 // is set: the barriers would be emitted by us.
2632 insert_pre_barrier(heap_base_oop, offset, p, !need_mem_bar);
2633 }
2634 break;
2635 case T_ADDRESS:
2636 // Cast to an int type.
2637 p = _gvn.transform(new CastP2XNode(NULL, p));
2638 p = ConvX2UL(p);
2639 break;
2640 default:
2641 fatal("unexpected type %d: %s", type, type2name(type));
2642 break;
2643 }
2644 }
2645 // The load node has the control of the preceding MemBarCPUOrder. All
2646 // following nodes will have the control of the MemBarCPUOrder inserted at
2647 // the end of this method. So, pushing the load onto the stack at a later
2648 // point is fine.
2649 set_result(p);
2650 } else {
2651 // place effect of store into memory
2652 switch (type) {
2653 case T_DOUBLE:
2654 val = dstore_rounding(val);
2655 break;
2656 case T_ADDRESS:
2657 // Repackage the long as a pointer.
2658 val = ConvL2X(val);
2659 val = _gvn.transform(new CastX2PNode(val));
2660 break;
2661 default:
2662 break;
2663 }
2664
2665 if (type == T_OBJECT) {
2666 store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2667 } else {
2668 store_to_memory(control(), adr, val, type, adr_type, mo, requires_atomic_access, unaligned, mismatched);
2669 }
2670 }
2671
2672 switch(kind) {
2673 case Relaxed:
2674 case Opaque:
2675 case Release:
2676 break;
2677 case Acquire:
2678 case Volatile:
2679 if (!is_store) {
2680 insert_mem_bar(Op_MemBarAcquire);
2681 } else {
2682 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2683 insert_mem_bar(Op_MemBarVolatile);
2684 }
2685 }
2686 break;
2687 default:
2688 ShouldNotReachHere();
2689 }
2690
2691 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2692
2693 return true;
2694 }
2695
2696 //----------------------------inline_unsafe_load_store----------------------------
2697 // This method serves a couple of different customers (depending on LoadStoreKind):
2698 //
2699 // LS_cmp_swap:
2700 //
2701 // boolean compareAndSetObject(Object o, long offset, Object expected, Object x);
2702 // boolean compareAndSetInt( Object o, long offset, int expected, int x);
2703 // boolean compareAndSetLong( Object o, long offset, long expected, long x);
2704 //
2705 // LS_cmp_swap_weak:
2706 //
2707 // boolean weakCompareAndSetObject( Object o, long offset, Object expected, Object x);
2708 // boolean weakCompareAndSetObjectPlain( Object o, long offset, Object expected, Object x);
2709 // boolean weakCompareAndSetObjectAcquire(Object o, long offset, Object expected, Object x);
2710 // boolean weakCompareAndSetObjectRelease(Object o, long offset, Object expected, Object x);
2711 //
2712 // boolean weakCompareAndSetInt( Object o, long offset, int expected, int x);
2713 // boolean weakCompareAndSetIntPlain( Object o, long offset, int expected, int x);
2714 // boolean weakCompareAndSetIntAcquire( Object o, long offset, int expected, int x);
2715 // boolean weakCompareAndSetIntRelease( Object o, long offset, int expected, int x);
2716 //
2717 // boolean weakCompareAndSetLong( Object o, long offset, long expected, long x);
2718 // boolean weakCompareAndSetLongPlain( Object o, long offset, long expected, long x);
2719 // boolean weakCompareAndSetLongAcquire( Object o, long offset, long expected, long x);
2720 // boolean weakCompareAndSetLongRelease( Object o, long offset, long expected, long x);
2721 //
2722 // LS_cmp_exchange:
2723 //
2724 // Object compareAndExchangeObjectVolatile(Object o, long offset, Object expected, Object x);
2725 // Object compareAndExchangeObjectAcquire( Object o, long offset, Object expected, Object x);
2726 // Object compareAndExchangeObjectRelease( Object o, long offset, Object expected, Object x);
2727 //
2728 // Object compareAndExchangeIntVolatile( Object o, long offset, Object expected, Object x);
2729 // Object compareAndExchangeIntAcquire( Object o, long offset, Object expected, Object x);
2730 // Object compareAndExchangeIntRelease( Object o, long offset, Object expected, Object x);
2731 //
2732 // Object compareAndExchangeLongVolatile( Object o, long offset, Object expected, Object x);
2733 // Object compareAndExchangeLongAcquire( Object o, long offset, Object expected, Object x);
2734 // Object compareAndExchangeLongRelease( Object o, long offset, Object expected, Object x);
2735 //
2736 // LS_get_add:
2737 //
2738 // int getAndAddInt( Object o, long offset, int delta)
2739 // long getAndAddLong(Object o, long offset, long delta)
2740 //
2741 // LS_get_set:
2742 //
2743 // int getAndSet(Object o, long offset, int newValue)
2744 // long getAndSet(Object o, long offset, long newValue)
2745 // Object getAndSet(Object o, long offset, Object newValue)
2746 //
2747 bool LibraryCallKit::inline_unsafe_load_store(const BasicType type, const LoadStoreKind kind, const AccessKind access_kind) {
2748 // This basic scheme here is the same as inline_unsafe_access, but
2749 // differs in enough details that combining them would make the code
2750 // overly confusing. (This is a true fact! I originally combined
2751 // them, but even I was confused by it!) As much code/comments as
2752 // possible are retained from inline_unsafe_access though to make
2753 // the correspondences clearer. - dl
2754
2755 if (callee()->is_static()) return false; // caller must have the capability!
2756
2757 #ifndef PRODUCT
2758 BasicType rtype;
2759 {
2760 ResourceMark rm;
2761 // Check the signatures.
2762 ciSignature* sig = callee()->signature();
2763 rtype = sig->return_type()->basic_type();
2764 switch(kind) {
2765 case LS_get_add:
2766 case LS_get_set: {
2767 // Check the signatures.
2768 #ifdef ASSERT
2769 assert(rtype == type, "get and set must return the expected type");
2770 assert(sig->count() == 3, "get and set has 3 arguments");
2771 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2772 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2773 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2774 assert(access_kind == Volatile, "mo is not passed to intrinsic nodes in current implementation");
2775 #endif // ASSERT
2776 break;
2777 }
2778 case LS_cmp_swap:
2779 case LS_cmp_swap_weak: {
2780 // Check the signatures.
2781 #ifdef ASSERT
2782 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2783 assert(sig->count() == 4, "CAS has 4 arguments");
2784 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2785 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2786 #endif // ASSERT
2787 break;
2788 }
2789 case LS_cmp_exchange: {
2790 // Check the signatures.
2791 #ifdef ASSERT
2792 assert(rtype == type, "CAS must return the expected type");
2793 assert(sig->count() == 4, "CAS has 4 arguments");
2794 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2795 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2796 #endif // ASSERT
2797 break;
2798 }
2799 default:
2800 ShouldNotReachHere();
2801 }
2802 }
2803 #endif //PRODUCT
2804
2805 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2806
2807 // Get arguments:
2808 Node* receiver = NULL;
2809 Node* base = NULL;
2810 Node* offset = NULL;
2811 Node* oldval = NULL;
2812 Node* newval = NULL;
2813 switch(kind) {
2814 case LS_cmp_swap:
2815 case LS_cmp_swap_weak:
2816 case LS_cmp_exchange: {
2817 const bool two_slot_type = type2size[type] == 2;
2818 receiver = argument(0); // type: oop
2819 base = argument(1); // type: oop
2820 offset = argument(2); // type: long
2821 oldval = argument(4); // type: oop, int, or long
2822 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2823 break;
2824 }
2825 case LS_get_add:
2826 case LS_get_set: {
2827 receiver = argument(0); // type: oop
2828 base = argument(1); // type: oop
2829 offset = argument(2); // type: long
2830 oldval = NULL;
2831 newval = argument(4); // type: oop, int, or long
2832 break;
2833 }
2834 default:
2835 ShouldNotReachHere();
2836 }
2837
2838 // Build field offset expression.
2839 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2840 // to be plain byte offsets, which are also the same as those accepted
2841 // by oopDesc::field_base.
2842 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2843 // 32-bit machines ignore the high half of long offsets
2844 offset = ConvL2X(offset);
2845 Node* adr = make_unsafe_address(base, offset, type, false);
2846 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2847
2848 Compile::AliasType* alias_type = C->alias_type(adr_type);
2849 BasicType bt = alias_type->basic_type();
2850 if (bt != T_ILLEGAL &&
2851 ((bt == T_OBJECT || bt == T_ARRAY) != (type == T_OBJECT))) {
2852 // Don't intrinsify mismatched object accesses.
2853 return false;
2854 }
2855
2856 // For CAS, unlike inline_unsafe_access, there seems no point in
2857 // trying to refine types. Just use the coarse types here.
2858 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2859 const Type *value_type = Type::get_const_basic_type(type);
2860
2861 switch (kind) {
2862 case LS_get_set:
2863 case LS_cmp_exchange: {
2864 if (type == T_OBJECT) {
2865 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2866 if (tjp != NULL) {
2867 value_type = tjp;
2868 }
2869 }
2870 break;
2871 }
2872 case LS_cmp_swap:
2873 case LS_cmp_swap_weak:
2874 case LS_get_add:
2875 break;
2876 default:
2877 ShouldNotReachHere();
2878 }
2879
2880 // Null check receiver.
2881 receiver = null_check(receiver);
2882 if (stopped()) {
2883 return true;
2884 }
2885
2886 int alias_idx = C->get_alias_index(adr_type);
2887
2888 // Memory-model-wise, a LoadStore acts like a little synchronized
2889 // block, so needs barriers on each side. These don't translate
2890 // into actual barriers on most machines, but we still need rest of
2891 // compiler to respect ordering.
2892
2893 switch (access_kind) {
2894 case Relaxed:
2895 case Acquire:
2896 break;
2897 case Release:
2898 insert_mem_bar(Op_MemBarRelease);
2899 break;
2900 case Volatile:
2901 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2902 insert_mem_bar(Op_MemBarVolatile);
2903 } else {
2904 insert_mem_bar(Op_MemBarRelease);
2905 }
2906 break;
2907 default:
2908 ShouldNotReachHere();
2909 }
2910 insert_mem_bar(Op_MemBarCPUOrder);
2911
2912 // Figure out the memory ordering.
2913 MemNode::MemOrd mo = access_kind_to_memord(access_kind);
2914
2915 // 4984716: MemBars must be inserted before this
2916 // memory node in order to avoid a false
2917 // dependency which will confuse the scheduler.
2918 Node *mem = memory(alias_idx);
2919
2920 // For now, we handle only those cases that actually exist: ints,
2921 // longs, and Object. Adding others should be straightforward.
2922 Node* load_store = NULL;
2923 switch(type) {
2924 case T_BYTE:
2925 switch(kind) {
2926 case LS_get_add:
2927 load_store = _gvn.transform(new GetAndAddBNode(control(), mem, adr, newval, adr_type));
2928 break;
2929 case LS_get_set:
2930 load_store = _gvn.transform(new GetAndSetBNode(control(), mem, adr, newval, adr_type));
2931 break;
2932 case LS_cmp_swap_weak:
2933 load_store = _gvn.transform(new WeakCompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2934 break;
2935 case LS_cmp_swap:
2936 load_store = _gvn.transform(new CompareAndSwapBNode(control(), mem, adr, newval, oldval, mo));
2937 break;
2938 case LS_cmp_exchange:
2939 load_store = _gvn.transform(new CompareAndExchangeBNode(control(), mem, adr, newval, oldval, adr_type, mo));
2940 break;
2941 default:
2942 ShouldNotReachHere();
2943 }
2944 break;
2945 case T_SHORT:
2946 switch(kind) {
2947 case LS_get_add:
2948 load_store = _gvn.transform(new GetAndAddSNode(control(), mem, adr, newval, adr_type));
2949 break;
2950 case LS_get_set:
2951 load_store = _gvn.transform(new GetAndSetSNode(control(), mem, adr, newval, adr_type));
2952 break;
2953 case LS_cmp_swap_weak:
2954 load_store = _gvn.transform(new WeakCompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2955 break;
2956 case LS_cmp_swap:
2957 load_store = _gvn.transform(new CompareAndSwapSNode(control(), mem, adr, newval, oldval, mo));
2958 break;
2959 case LS_cmp_exchange:
2960 load_store = _gvn.transform(new CompareAndExchangeSNode(control(), mem, adr, newval, oldval, adr_type, mo));
2961 break;
2962 default:
2963 ShouldNotReachHere();
2964 }
2965 break;
2966 case T_INT:
2967 switch(kind) {
2968 case LS_get_add:
2969 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2970 break;
2971 case LS_get_set:
2972 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2973 break;
2974 case LS_cmp_swap_weak:
2975 load_store = _gvn.transform(new WeakCompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2976 break;
2977 case LS_cmp_swap:
2978 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval, mo));
2979 break;
2980 case LS_cmp_exchange:
2981 load_store = _gvn.transform(new CompareAndExchangeINode(control(), mem, adr, newval, oldval, adr_type, mo));
2982 break;
2983 default:
2984 ShouldNotReachHere();
2985 }
2986 break;
2987 case T_LONG:
2988 switch(kind) {
2989 case LS_get_add:
2990 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2991 break;
2992 case LS_get_set:
2993 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2994 break;
2995 case LS_cmp_swap_weak:
2996 load_store = _gvn.transform(new WeakCompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
2997 break;
2998 case LS_cmp_swap:
2999 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval, mo));
3000 break;
3001 case LS_cmp_exchange:
3002 load_store = _gvn.transform(new CompareAndExchangeLNode(control(), mem, adr, newval, oldval, adr_type, mo));
3003 break;
3004 default:
3005 ShouldNotReachHere();
3006 }
3007 break;
3008 case T_OBJECT:
3009 // Transformation of a value which could be NULL pointer (CastPP #NULL)
3010 // could be delayed during Parse (for example, in adjust_map_after_if()).
3011 // Execute transformation here to avoid barrier generation in such case.
3012 if (_gvn.type(newval) == TypePtr::NULL_PTR)
3013 newval = _gvn.makecon(TypePtr::NULL_PTR);
3014
3015 // Reference stores need a store barrier.
3016 switch(kind) {
3017 case LS_get_set: {
3018 // If pre-barrier must execute before the oop store, old value will require do_load here.
3019 if (!can_move_pre_barrier()) {
3020 pre_barrier(true /* do_load*/,
3021 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
3022 NULL /* pre_val*/,
3023 T_OBJECT);
3024 } // Else move pre_barrier to use load_store value, see below.
3025 break;
3026 }
3027 case LS_cmp_swap_weak:
3028 case LS_cmp_swap:
3029 case LS_cmp_exchange: {
3030 // Same as for newval above:
3031 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
3032 oldval = _gvn.makecon(TypePtr::NULL_PTR);
3033 }
3034 // The only known value which might get overwritten is oldval.
3035 pre_barrier(false /* do_load */,
3036 control(), NULL, NULL, max_juint, NULL, NULL,
3037 oldval /* pre_val */,
3038 T_OBJECT);
3039 break;
3040 }
3041 default:
3042 ShouldNotReachHere();
3043 }
3044
3045 #ifdef _LP64
3046 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3047 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
3048
3049 switch(kind) {
3050 case LS_get_set:
3051 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr, newval_enc, adr_type, value_type->make_narrowoop()));
3052 break;
3053 case LS_cmp_swap_weak: {
3054 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3055 load_store = _gvn.transform(new WeakCompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3056 break;
3057 }
3058 case LS_cmp_swap: {
3059 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3060 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr, newval_enc, oldval_enc, mo));
3061 break;
3062 }
3063 case LS_cmp_exchange: {
3064 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
3065 load_store = _gvn.transform(new CompareAndExchangeNNode(control(), mem, adr, newval_enc, oldval_enc, adr_type, value_type->make_narrowoop(), mo));
3066 break;
3067 }
3068 default:
3069 ShouldNotReachHere();
3070 }
3071 } else
3072 #endif
3073 switch (kind) {
3074 case LS_get_set:
3075 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3076 break;
3077 case LS_cmp_swap_weak:
3078 load_store = _gvn.transform(new WeakCompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3079 break;
3080 case LS_cmp_swap:
3081 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval, mo));
3082 break;
3083 case LS_cmp_exchange:
3084 load_store = _gvn.transform(new CompareAndExchangePNode(control(), mem, adr, newval, oldval, adr_type, value_type->is_oopptr(), mo));
3085 break;
3086 default:
3087 ShouldNotReachHere();
3088 }
3089
3090 // Emit the post barrier only when the actual store happened. This makes sense
3091 // to check only for LS_cmp_* that can fail to set the value.
3092 // LS_cmp_exchange does not produce any branches by default, so there is no
3093 // boolean result to piggyback on. TODO: When we merge CompareAndSwap with
3094 // CompareAndExchange and move branches here, it would make sense to conditionalize
3095 // post_barriers for LS_cmp_exchange as well.
3096 //
3097 // CAS success path is marked more likely since we anticipate this is a performance
3098 // critical path, while CAS failure path can use the penalty for going through unlikely
3099 // path as backoff. Which is still better than doing a store barrier there.
3100 switch (kind) {
3101 case LS_get_set:
3102 case LS_cmp_exchange: {
3103 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3104 break;
3105 }
3106 case LS_cmp_swap_weak:
3107 case LS_cmp_swap: {
3108 IdealKit ideal(this);
3109 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
3110 sync_kit(ideal);
3111 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3112 ideal.sync_kit(this);
3113 } ideal.end_if();
3114 final_sync(ideal);
3115 break;
3116 }
3117 default:
3118 ShouldNotReachHere();
3119 }
3120 break;
3121 default:
3122 fatal("unexpected type %d: %s", type, type2name(type));
3123 break;
3124 }
3125
3126 // SCMemProjNodes represent the memory state of a LoadStore. Their
3127 // main role is to prevent LoadStore nodes from being optimized away
3128 // when their results aren't used.
3129 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
3130 set_memory(proj, alias_idx);
3131
3132 if (type == T_OBJECT && (kind == LS_get_set || kind == LS_cmp_exchange)) {
3133 #ifdef _LP64
3134 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3135 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
3136 }
3137 #endif
3138 if (can_move_pre_barrier() && kind == LS_get_set) {
3139 // Don't need to load pre_val. The old value is returned by load_store.
3140 // The pre_barrier can execute after the xchg as long as no safepoint
3141 // gets inserted between them.
3142 pre_barrier(false /* do_load */,
3143 control(), NULL, NULL, max_juint, NULL, NULL,
3144 load_store /* pre_val */,
3145 T_OBJECT);
3146 }
3147 }
3148
3149 // Add the trailing membar surrounding the access
3150 insert_mem_bar(Op_MemBarCPUOrder);
3151
3152 switch (access_kind) {
3153 case Relaxed:
3154 case Release:
3155 break; // do nothing
3156 case Acquire:
3157 case Volatile:
3158 insert_mem_bar(Op_MemBarAcquire);
3159 // !support_IRIW_for_not_multiple_copy_atomic_cpu handled in platform code
3160 break;
3161 default:
3162 ShouldNotReachHere();
3163 }
3164
3165 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3166 set_result(load_store);
3167 return true;
3168 }
3169
3170 MemNode::MemOrd LibraryCallKit::access_kind_to_memord_LS(AccessKind kind, bool is_store) {
3171 MemNode::MemOrd mo = MemNode::unset;
3172 switch(kind) {
3173 case Opaque:
3174 case Relaxed: mo = MemNode::unordered; break;
3175 case Acquire: mo = MemNode::acquire; break;
3176 case Release: mo = MemNode::release; break;
3177 case Volatile: mo = is_store ? MemNode::release : MemNode::acquire; break;
3178 default:
3179 ShouldNotReachHere();
3180 }
3181 guarantee(mo != MemNode::unset, "Should select memory ordering");
3182 return mo;
3183 }
3184
3185 MemNode::MemOrd LibraryCallKit::access_kind_to_memord(AccessKind kind) {
3186 MemNode::MemOrd mo = MemNode::unset;
3187 switch(kind) {
3188 case Opaque:
3189 case Relaxed: mo = MemNode::unordered; break;
3190 case Acquire: mo = MemNode::acquire; break;
3191 case Release: mo = MemNode::release; break;
3192 case Volatile: mo = MemNode::seqcst; break;
3193 default:
3194 ShouldNotReachHere();
3195 }
3196 guarantee(mo != MemNode::unset, "Should select memory ordering");
3197 return mo;
3198 }
3199
3200 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3201 // Regardless of form, don't allow previous ld/st to move down,
3202 // then issue acquire, release, or volatile mem_bar.
3203 insert_mem_bar(Op_MemBarCPUOrder);
3204 switch(id) {
3205 case vmIntrinsics::_loadFence:
3206 insert_mem_bar(Op_LoadFence);
3207 return true;
3208 case vmIntrinsics::_storeFence:
3209 insert_mem_bar(Op_StoreFence);
3210 return true;
3211 case vmIntrinsics::_fullFence:
3212 insert_mem_bar(Op_MemBarVolatile);
3213 return true;
3214 default:
3215 fatal_unexpected_iid(id);
3216 return false;
3217 }
3218 }
3219
3220 bool LibraryCallKit::inline_onspinwait() {
3221 insert_mem_bar(Op_OnSpinWait);
3222 return true;
3223 }
3224
3225 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3226 if (!kls->is_Con()) {
3227 return true;
3228 }
3229 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3230 if (klsptr == NULL) {
3231 return true;
3232 }
3233 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3234 // don't need a guard for a klass that is already initialized
3235 return !ik->is_initialized();
3236 }
3237
3238 //----------------------------inline_unsafe_allocate---------------------------
3239 // public native Object Unsafe.allocateInstance(Class<?> cls);
3240 bool LibraryCallKit::inline_unsafe_allocate() {
3241 if (callee()->is_static()) return false; // caller must have the capability!
3242
3243 null_check_receiver(); // null-check, then ignore
3244 Node* cls = null_check(argument(1));
3245 if (stopped()) return true;
3246
3247 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3248 kls = null_check(kls);
3249 if (stopped()) return true; // argument was like int.class
3250
3251 Node* test = NULL;
3252 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3253 // Note: The argument might still be an illegal value like
3254 // Serializable.class or Object[].class. The runtime will handle it.
3255 // But we must make an explicit check for initialization.
3256 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3257 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3258 // can generate code to load it as unsigned byte.
3259 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3260 Node* bits = intcon(InstanceKlass::fully_initialized);
3261 test = _gvn.transform(new SubINode(inst, bits));
3262 // The 'test' is non-zero if we need to take a slow path.
3263 }
3264
3265 Node* obj = new_instance(kls, test);
3266 set_result(obj);
3267 return true;
3268 }
3269
3270 //------------------------inline_native_time_funcs--------------
3271 // inline code for System.currentTimeMillis() and System.nanoTime()
3272 // these have the same type and signature
3273 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3274 const TypeFunc* tf = OptoRuntime::void_long_Type();
3275 const TypePtr* no_memory_effects = NULL;
3276 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3277 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3278 #ifdef ASSERT
3279 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3280 assert(value_top == top(), "second value must be top");
3281 #endif
3282 set_result(value);
3283 return true;
3284 }
3285
3286 #ifdef TRACE_HAVE_INTRINSICS
3287
3288 /*
3289 * oop -> myklass
3290 * myklass->trace_id |= USED
3291 * return myklass->trace_id & ~0x3
3292 */
3293 bool LibraryCallKit::inline_native_classID() {
3294 Node* cls = null_check(argument(0), T_OBJECT);
3295 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3296 kls = null_check(kls, T_OBJECT);
3297
3298 ByteSize offset = TRACE_KLASS_TRACE_ID_OFFSET;
3299 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3300 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3301
3302 Node* clsused = longcon(0x01l); // set the class bit
3303 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3304 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3305 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3306
3307 #ifdef TRACE_ID_META_BITS
3308 Node* mbits = longcon(~TRACE_ID_META_BITS);
3309 tvalue = _gvn.transform(new AndLNode(tvalue, mbits));
3310 #endif
3311 #ifdef TRACE_ID_CLASS_SHIFT
3312 Node* cbits = intcon(TRACE_ID_CLASS_SHIFT);
3313 tvalue = _gvn.transform(new URShiftLNode(tvalue, cbits));
3314 #endif
3315
3316 set_result(tvalue);
3317 return true;
3318
3319 }
3320
3321 bool LibraryCallKit::inline_native_getBufferWriter() {
3322 Node* tls_ptr = _gvn.transform(new ThreadLocalNode());
3323
3324 Node* jobj_ptr = basic_plus_adr(top(), tls_ptr,
3325 in_bytes(TRACE_THREAD_DATA_WRITER_OFFSET)
3326 );
3327
3328 Node* jobj = make_load(control(), jobj_ptr, TypeRawPtr::BOTTOM, T_ADDRESS, MemNode::unordered);
3329
3330 Node* jobj_cmp_null = _gvn.transform( new CmpPNode(jobj, null()) );
3331 Node* test_jobj_eq_null = _gvn.transform( new BoolNode(jobj_cmp_null, BoolTest::eq) );
3332
3333 IfNode* iff_jobj_null =
3334 create_and_map_if(control(), test_jobj_eq_null, PROB_MIN, COUNT_UNKNOWN);
3335
3336 enum { _normal_path = 1,
3337 _null_path = 2,
3338 PATH_LIMIT };
3339
3340 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3341 PhiNode* result_val = new PhiNode(result_rgn, TypePtr::BOTTOM);
3342
3343 Node* jobj_is_null = _gvn.transform(new IfTrueNode(iff_jobj_null));
3344 result_rgn->init_req(_null_path, jobj_is_null);
3345 result_val->init_req(_null_path, null());
3346
3347 Node* jobj_is_not_null = _gvn.transform(new IfFalseNode(iff_jobj_null));
3348 result_rgn->init_req(_normal_path, jobj_is_not_null);
3349
3350 Node* res = make_load(jobj_is_not_null, jobj, TypeInstPtr::NOTNULL, T_OBJECT, MemNode::unordered);
3351 result_val->init_req(_normal_path, res);
3352
3353 set_result(result_rgn, result_val);
3354
3355 return true;
3356 }
3357
3358 #endif
3359
3360 //------------------------inline_native_currentThread------------------
3361 bool LibraryCallKit::inline_native_currentThread() {
3362 Node* junk = NULL;
3363 set_result(generate_current_thread(junk));
3364 return true;
3365 }
3366
3367 //------------------------inline_native_isInterrupted------------------
3368 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3369 bool LibraryCallKit::inline_native_isInterrupted() {
3370 // Add a fast path to t.isInterrupted(clear_int):
3371 // (t == Thread.current() &&
3372 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3373 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3374 // So, in the common case that the interrupt bit is false,
3375 // we avoid making a call into the VM. Even if the interrupt bit
3376 // is true, if the clear_int argument is false, we avoid the VM call.
3377 // However, if the receiver is not currentThread, we must call the VM,
3378 // because there must be some locking done around the operation.
3379
3380 // We only go to the fast case code if we pass two guards.
3381 // Paths which do not pass are accumulated in the slow_region.
3382
3383 enum {
3384 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3385 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3386 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3387 PATH_LIMIT
3388 };
3389
3390 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3391 // out of the function.
3392 insert_mem_bar(Op_MemBarCPUOrder);
3393
3394 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3395 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3396
3397 RegionNode* slow_region = new RegionNode(1);
3398 record_for_igvn(slow_region);
3399
3400 // (a) Receiving thread must be the current thread.
3401 Node* rec_thr = argument(0);
3402 Node* tls_ptr = NULL;
3403 Node* cur_thr = generate_current_thread(tls_ptr);
3404 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3405 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3406
3407 generate_slow_guard(bol_thr, slow_region);
3408
3409 // (b) Interrupt bit on TLS must be false.
3410 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3411 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3412 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3413
3414 // Set the control input on the field _interrupted read to prevent it floating up.
3415 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3416 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3417 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3418
3419 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3420
3421 // First fast path: if (!TLS._interrupted) return false;
3422 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3423 result_rgn->init_req(no_int_result_path, false_bit);
3424 result_val->init_req(no_int_result_path, intcon(0));
3425
3426 // drop through to next case
3427 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3428
3429 #ifndef _WINDOWS
3430 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3431 Node* clr_arg = argument(1);
3432 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3433 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3434 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3435
3436 // Second fast path: ... else if (!clear_int) return true;
3437 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3438 result_rgn->init_req(no_clear_result_path, false_arg);
3439 result_val->init_req(no_clear_result_path, intcon(1));
3440
3441 // drop through to next case
3442 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3443 #else
3444 // To return true on Windows you must read the _interrupted field
3445 // and check the event state i.e. take the slow path.
3446 #endif // _WINDOWS
3447
3448 // (d) Otherwise, go to the slow path.
3449 slow_region->add_req(control());
3450 set_control( _gvn.transform(slow_region));
3451
3452 if (stopped()) {
3453 // There is no slow path.
3454 result_rgn->init_req(slow_result_path, top());
3455 result_val->init_req(slow_result_path, top());
3456 } else {
3457 // non-virtual because it is a private non-static
3458 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3459
3460 Node* slow_val = set_results_for_java_call(slow_call);
3461 // this->control() comes from set_results_for_java_call
3462
3463 Node* fast_io = slow_call->in(TypeFunc::I_O);
3464 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3465
3466 // These two phis are pre-filled with copies of of the fast IO and Memory
3467 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3468 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3469
3470 result_rgn->init_req(slow_result_path, control());
3471 result_io ->init_req(slow_result_path, i_o());
3472 result_mem->init_req(slow_result_path, reset_memory());
3473 result_val->init_req(slow_result_path, slow_val);
3474
3475 set_all_memory(_gvn.transform(result_mem));
3476 set_i_o( _gvn.transform(result_io));
3477 }
3478
3479 C->set_has_split_ifs(true); // Has chance for split-if optimization
3480 set_result(result_rgn, result_val);
3481 return true;
3482 }
3483
3484 //---------------------------load_mirror_from_klass----------------------------
3485 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3486 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3487 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3488 Node* load = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3489 return make_load(NULL, load, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3490 }
3491
3492 //-----------------------load_klass_from_mirror_common-------------------------
3493 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3494 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3495 // and branch to the given path on the region.
3496 // If never_see_null, take an uncommon trap on null, so we can optimistically
3497 // compile for the non-null case.
3498 // If the region is NULL, force never_see_null = true.
3499 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3500 bool never_see_null,
3501 RegionNode* region,
3502 int null_path,
3503 int offset) {
3504 if (region == NULL) never_see_null = true;
3505 Node* p = basic_plus_adr(mirror, offset);
3506 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3507 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3508 Node* null_ctl = top();
3509 kls = null_check_oop(kls, &null_ctl, never_see_null);
3510 if (region != NULL) {
3511 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3512 region->init_req(null_path, null_ctl);
3513 } else {
3514 assert(null_ctl == top(), "no loose ends");
3515 }
3516 return kls;
3517 }
3518
3519 //--------------------(inline_native_Class_query helpers)---------------------
3520 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE_FAST, JVM_ACC_HAS_FINALIZER.
3521 // Fall through if (mods & mask) == bits, take the guard otherwise.
3522 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3523 // Branch around if the given klass has the given modifier bit set.
3524 // Like generate_guard, adds a new path onto the region.
3525 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3526 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3527 Node* mask = intcon(modifier_mask);
3528 Node* bits = intcon(modifier_bits);
3529 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3530 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3531 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3532 return generate_fair_guard(bol, region);
3533 }
3534 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3535 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3536 }
3537
3538 //-------------------------inline_native_Class_query-------------------
3539 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3540 const Type* return_type = TypeInt::BOOL;
3541 Node* prim_return_value = top(); // what happens if it's a primitive class?
3542 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3543 bool expect_prim = false; // most of these guys expect to work on refs
3544
3545 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3546
3547 Node* mirror = argument(0);
3548 Node* obj = top();
3549
3550 switch (id) {
3551 case vmIntrinsics::_isInstance:
3552 // nothing is an instance of a primitive type
3553 prim_return_value = intcon(0);
3554 obj = argument(1);
3555 break;
3556 case vmIntrinsics::_getModifiers:
3557 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3558 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3559 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3560 break;
3561 case vmIntrinsics::_isInterface:
3562 prim_return_value = intcon(0);
3563 break;
3564 case vmIntrinsics::_isArray:
3565 prim_return_value = intcon(0);
3566 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3567 break;
3568 case vmIntrinsics::_isPrimitive:
3569 prim_return_value = intcon(1);
3570 expect_prim = true; // obviously
3571 break;
3572 case vmIntrinsics::_getSuperclass:
3573 prim_return_value = null();
3574 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3575 break;
3576 case vmIntrinsics::_getClassAccessFlags:
3577 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3578 return_type = TypeInt::INT; // not bool! 6297094
3579 break;
3580 default:
3581 fatal_unexpected_iid(id);
3582 break;
3583 }
3584
3585 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3586 if (mirror_con == NULL) return false; // cannot happen?
3587
3588 #ifndef PRODUCT
3589 if (C->print_intrinsics() || C->print_inlining()) {
3590 ciType* k = mirror_con->java_mirror_type();
3591 if (k) {
3592 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3593 k->print_name();
3594 tty->cr();
3595 }
3596 }
3597 #endif
3598
3599 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3600 RegionNode* region = new RegionNode(PATH_LIMIT);
3601 record_for_igvn(region);
3602 PhiNode* phi = new PhiNode(region, return_type);
3603
3604 // The mirror will never be null of Reflection.getClassAccessFlags, however
3605 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3606 // if it is. See bug 4774291.
3607
3608 // For Reflection.getClassAccessFlags(), the null check occurs in
3609 // the wrong place; see inline_unsafe_access(), above, for a similar
3610 // situation.
3611 mirror = null_check(mirror);
3612 // If mirror or obj is dead, only null-path is taken.
3613 if (stopped()) return true;
3614
3615 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3616
3617 // Now load the mirror's klass metaobject, and null-check it.
3618 // Side-effects region with the control path if the klass is null.
3619 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3620 // If kls is null, we have a primitive mirror.
3621 phi->init_req(_prim_path, prim_return_value);
3622 if (stopped()) { set_result(region, phi); return true; }
3623 bool safe_for_replace = (region->in(_prim_path) == top());
3624
3625 Node* p; // handy temp
3626 Node* null_ctl;
3627
3628 // Now that we have the non-null klass, we can perform the real query.
3629 // For constant classes, the query will constant-fold in LoadNode::Value.
3630 Node* query_value = top();
3631 switch (id) {
3632 case vmIntrinsics::_isInstance:
3633 // nothing is an instance of a primitive type
3634 query_value = gen_instanceof(obj, kls, safe_for_replace);
3635 break;
3636
3637 case vmIntrinsics::_getModifiers:
3638 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3639 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3640 break;
3641
3642 case vmIntrinsics::_isInterface:
3643 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3644 if (generate_interface_guard(kls, region) != NULL)
3645 // A guard was added. If the guard is taken, it was an interface.
3646 phi->add_req(intcon(1));
3647 // If we fall through, it's a plain class.
3648 query_value = intcon(0);
3649 break;
3650
3651 case vmIntrinsics::_isArray:
3652 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3653 if (generate_array_guard(kls, region) != NULL)
3654 // A guard was added. If the guard is taken, it was an array.
3655 phi->add_req(intcon(1));
3656 // If we fall through, it's a plain class.
3657 query_value = intcon(0);
3658 break;
3659
3660 case vmIntrinsics::_isPrimitive:
3661 query_value = intcon(0); // "normal" path produces false
3662 break;
3663
3664 case vmIntrinsics::_getSuperclass:
3665 // The rules here are somewhat unfortunate, but we can still do better
3666 // with random logic than with a JNI call.
3667 // Interfaces store null or Object as _super, but must report null.
3668 // Arrays store an intermediate super as _super, but must report Object.
3669 // Other types can report the actual _super.
3670 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3671 if (generate_interface_guard(kls, region) != NULL)
3672 // A guard was added. If the guard is taken, it was an interface.
3673 phi->add_req(null());
3674 if (generate_array_guard(kls, region) != NULL)
3675 // A guard was added. If the guard is taken, it was an array.
3676 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3677 // If we fall through, it's a plain class. Get its _super.
3678 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3679 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3680 null_ctl = top();
3681 kls = null_check_oop(kls, &null_ctl);
3682 if (null_ctl != top()) {
3683 // If the guard is taken, Object.superClass is null (both klass and mirror).
3684 region->add_req(null_ctl);
3685 phi ->add_req(null());
3686 }
3687 if (!stopped()) {
3688 query_value = load_mirror_from_klass(kls);
3689 }
3690 break;
3691
3692 case vmIntrinsics::_getClassAccessFlags:
3693 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3694 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3695 break;
3696
3697 default:
3698 fatal_unexpected_iid(id);
3699 break;
3700 }
3701
3702 // Fall-through is the normal case of a query to a real class.
3703 phi->init_req(1, query_value);
3704 region->init_req(1, control());
3705
3706 C->set_has_split_ifs(true); // Has chance for split-if optimization
3707 set_result(region, phi);
3708 return true;
3709 }
3710
3711 //-------------------------inline_Class_cast-------------------
3712 bool LibraryCallKit::inline_Class_cast() {
3713 Node* mirror = argument(0); // Class
3714 Node* obj = argument(1);
3715 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3716 if (mirror_con == NULL) {
3717 return false; // dead path (mirror->is_top()).
3718 }
3719 if (obj == NULL || obj->is_top()) {
3720 return false; // dead path
3721 }
3722 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3723
3724 // First, see if Class.cast() can be folded statically.
3725 // java_mirror_type() returns non-null for compile-time Class constants.
3726 ciType* tm = mirror_con->java_mirror_type();
3727 if (tm != NULL && tm->is_klass() &&
3728 tp != NULL && tp->klass() != NULL) {
3729 if (!tp->klass()->is_loaded()) {
3730 // Don't use intrinsic when class is not loaded.
3731 return false;
3732 } else {
3733 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3734 if (static_res == Compile::SSC_always_true) {
3735 // isInstance() is true - fold the code.
3736 set_result(obj);
3737 return true;
3738 } else if (static_res == Compile::SSC_always_false) {
3739 // Don't use intrinsic, have to throw ClassCastException.
3740 // If the reference is null, the non-intrinsic bytecode will
3741 // be optimized appropriately.
3742 return false;
3743 }
3744 }
3745 }
3746
3747 // Bailout intrinsic and do normal inlining if exception path is frequent.
3748 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3749 return false;
3750 }
3751
3752 // Generate dynamic checks.
3753 // Class.cast() is java implementation of _checkcast bytecode.
3754 // Do checkcast (Parse::do_checkcast()) optimizations here.
3755
3756 mirror = null_check(mirror);
3757 // If mirror is dead, only null-path is taken.
3758 if (stopped()) {
3759 return true;
3760 }
3761
3762 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3763 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3764 RegionNode* region = new RegionNode(PATH_LIMIT);
3765 record_for_igvn(region);
3766
3767 // Now load the mirror's klass metaobject, and null-check it.
3768 // If kls is null, we have a primitive mirror and
3769 // nothing is an instance of a primitive type.
3770 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3771
3772 Node* res = top();
3773 if (!stopped()) {
3774 Node* bad_type_ctrl = top();
3775 // Do checkcast optimizations.
3776 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3777 region->init_req(_bad_type_path, bad_type_ctrl);
3778 }
3779 if (region->in(_prim_path) != top() ||
3780 region->in(_bad_type_path) != top()) {
3781 // Let Interpreter throw ClassCastException.
3782 PreserveJVMState pjvms(this);
3783 set_control(_gvn.transform(region));
3784 uncommon_trap(Deoptimization::Reason_intrinsic,
3785 Deoptimization::Action_maybe_recompile);
3786 }
3787 if (!stopped()) {
3788 set_result(res);
3789 }
3790 return true;
3791 }
3792
3793
3794 //--------------------------inline_native_subtype_check------------------------
3795 // This intrinsic takes the JNI calls out of the heart of
3796 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3797 bool LibraryCallKit::inline_native_subtype_check() {
3798 // Pull both arguments off the stack.
3799 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3800 args[0] = argument(0);
3801 args[1] = argument(1);
3802 Node* klasses[2]; // corresponding Klasses: superk, subk
3803 klasses[0] = klasses[1] = top();
3804
3805 enum {
3806 // A full decision tree on {superc is prim, subc is prim}:
3807 _prim_0_path = 1, // {P,N} => false
3808 // {P,P} & superc!=subc => false
3809 _prim_same_path, // {P,P} & superc==subc => true
3810 _prim_1_path, // {N,P} => false
3811 _ref_subtype_path, // {N,N} & subtype check wins => true
3812 _both_ref_path, // {N,N} & subtype check loses => false
3813 PATH_LIMIT
3814 };
3815
3816 RegionNode* region = new RegionNode(PATH_LIMIT);
3817 Node* phi = new PhiNode(region, TypeInt::BOOL);
3818 record_for_igvn(region);
3819
3820 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3821 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3822 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3823
3824 // First null-check both mirrors and load each mirror's klass metaobject.
3825 int which_arg;
3826 for (which_arg = 0; which_arg <= 1; which_arg++) {
3827 Node* arg = args[which_arg];
3828 arg = null_check(arg);
3829 if (stopped()) break;
3830 args[which_arg] = arg;
3831
3832 Node* p = basic_plus_adr(arg, class_klass_offset);
3833 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3834 klasses[which_arg] = _gvn.transform(kls);
3835 }
3836
3837 // Having loaded both klasses, test each for null.
3838 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3839 for (which_arg = 0; which_arg <= 1; which_arg++) {
3840 Node* kls = klasses[which_arg];
3841 Node* null_ctl = top();
3842 kls = null_check_oop(kls, &null_ctl, never_see_null);
3843 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3844 region->init_req(prim_path, null_ctl);
3845 if (stopped()) break;
3846 klasses[which_arg] = kls;
3847 }
3848
3849 if (!stopped()) {
3850 // now we have two reference types, in klasses[0..1]
3851 Node* subk = klasses[1]; // the argument to isAssignableFrom
3852 Node* superk = klasses[0]; // the receiver
3853 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3854 // now we have a successful reference subtype check
3855 region->set_req(_ref_subtype_path, control());
3856 }
3857
3858 // If both operands are primitive (both klasses null), then
3859 // we must return true when they are identical primitives.
3860 // It is convenient to test this after the first null klass check.
3861 set_control(region->in(_prim_0_path)); // go back to first null check
3862 if (!stopped()) {
3863 // Since superc is primitive, make a guard for the superc==subc case.
3864 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3865 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3866 generate_guard(bol_eq, region, PROB_FAIR);
3867 if (region->req() == PATH_LIMIT+1) {
3868 // A guard was added. If the added guard is taken, superc==subc.
3869 region->swap_edges(PATH_LIMIT, _prim_same_path);
3870 region->del_req(PATH_LIMIT);
3871 }
3872 region->set_req(_prim_0_path, control()); // Not equal after all.
3873 }
3874
3875 // these are the only paths that produce 'true':
3876 phi->set_req(_prim_same_path, intcon(1));
3877 phi->set_req(_ref_subtype_path, intcon(1));
3878
3879 // pull together the cases:
3880 assert(region->req() == PATH_LIMIT, "sane region");
3881 for (uint i = 1; i < region->req(); i++) {
3882 Node* ctl = region->in(i);
3883 if (ctl == NULL || ctl == top()) {
3884 region->set_req(i, top());
3885 phi ->set_req(i, top());
3886 } else if (phi->in(i) == NULL) {
3887 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3888 }
3889 }
3890
3891 set_control(_gvn.transform(region));
3892 set_result(_gvn.transform(phi));
3893 return true;
3894 }
3895
3896 //---------------------generate_array_guard_common------------------------
3897 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3898 bool obj_array, bool not_array) {
3899
3900 if (stopped()) {
3901 return NULL;
3902 }
3903
3904 // If obj_array/non_array==false/false:
3905 // Branch around if the given klass is in fact an array (either obj or prim).
3906 // If obj_array/non_array==false/true:
3907 // Branch around if the given klass is not an array klass of any kind.
3908 // If obj_array/non_array==true/true:
3909 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3910 // If obj_array/non_array==true/false:
3911 // Branch around if the kls is an oop array (Object[] or subtype)
3912 //
3913 // Like generate_guard, adds a new path onto the region.
3914 jint layout_con = 0;
3915 Node* layout_val = get_layout_helper(kls, layout_con);
3916 if (layout_val == NULL) {
3917 bool query = (obj_array
3918 ? Klass::layout_helper_is_objArray(layout_con)
3919 : Klass::layout_helper_is_array(layout_con));
3920 if (query == not_array) {
3921 return NULL; // never a branch
3922 } else { // always a branch
3923 Node* always_branch = control();
3924 if (region != NULL)
3925 region->add_req(always_branch);
3926 set_control(top());
3927 return always_branch;
3928 }
3929 }
3930 // Now test the correct condition.
3931 jint nval = (obj_array
3932 ? (jint)(Klass::_lh_array_tag_type_value
3933 << Klass::_lh_array_tag_shift)
3934 : Klass::_lh_neutral_value);
3935 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3936 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3937 // invert the test if we are looking for a non-array
3938 if (not_array) btest = BoolTest(btest).negate();
3939 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3940 return generate_fair_guard(bol, region);
3941 }
3942
3943
3944 //-----------------------inline_native_newArray--------------------------
3945 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3946 // private native Object Unsafe.allocateUninitializedArray0(Class<?> cls, int size);
3947 bool LibraryCallKit::inline_unsafe_newArray(bool uninitialized) {
3948 Node* mirror;
3949 Node* count_val;
3950 if (uninitialized) {
3951 mirror = argument(1);
3952 count_val = argument(2);
3953 } else {
3954 mirror = argument(0);
3955 count_val = argument(1);
3956 }
3957
3958 mirror = null_check(mirror);
3959 // If mirror or obj is dead, only null-path is taken.
3960 if (stopped()) return true;
3961
3962 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3963 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3964 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3965 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3966 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3967
3968 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3969 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3970 result_reg, _slow_path);
3971 Node* normal_ctl = control();
3972 Node* no_array_ctl = result_reg->in(_slow_path);
3973
3974 // Generate code for the slow case. We make a call to newArray().
3975 set_control(no_array_ctl);
3976 if (!stopped()) {
3977 // Either the input type is void.class, or else the
3978 // array klass has not yet been cached. Either the
3979 // ensuing call will throw an exception, or else it
3980 // will cache the array klass for next time.
3981 PreserveJVMState pjvms(this);
3982 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3983 Node* slow_result = set_results_for_java_call(slow_call);
3984 // this->control() comes from set_results_for_java_call
3985 result_reg->set_req(_slow_path, control());
3986 result_val->set_req(_slow_path, slow_result);
3987 result_io ->set_req(_slow_path, i_o());
3988 result_mem->set_req(_slow_path, reset_memory());
3989 }
3990
3991 set_control(normal_ctl);
3992 if (!stopped()) {
3993 // Normal case: The array type has been cached in the java.lang.Class.
3994 // The following call works fine even if the array type is polymorphic.
3995 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3996 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3997 result_reg->init_req(_normal_path, control());
3998 result_val->init_req(_normal_path, obj);
3999 result_io ->init_req(_normal_path, i_o());
4000 result_mem->init_req(_normal_path, reset_memory());
4001
4002 if (uninitialized) {
4003 // Mark the allocation so that zeroing is skipped
4004 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(obj, &_gvn);
4005 alloc->maybe_set_complete(&_gvn);
4006 }
4007 }
4008
4009 // Return the combined state.
4010 set_i_o( _gvn.transform(result_io) );
4011 set_all_memory( _gvn.transform(result_mem));
4012
4013 C->set_has_split_ifs(true); // Has chance for split-if optimization
4014 set_result(result_reg, result_val);
4015 return true;
4016 }
4017
4018 //----------------------inline_native_getLength--------------------------
4019 // public static native int java.lang.reflect.Array.getLength(Object array);
4020 bool LibraryCallKit::inline_native_getLength() {
4021 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4022
4023 Node* array = null_check(argument(0));
4024 // If array is dead, only null-path is taken.
4025 if (stopped()) return true;
4026
4027 // Deoptimize if it is a non-array.
4028 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
4029
4030 if (non_array != NULL) {
4031 PreserveJVMState pjvms(this);
4032 set_control(non_array);
4033 uncommon_trap(Deoptimization::Reason_intrinsic,
4034 Deoptimization::Action_maybe_recompile);
4035 }
4036
4037 // If control is dead, only non-array-path is taken.
4038 if (stopped()) return true;
4039
4040 // The works fine even if the array type is polymorphic.
4041 // It could be a dynamic mix of int[], boolean[], Object[], etc.
4042 Node* result = load_array_length(array);
4043
4044 C->set_has_split_ifs(true); // Has chance for split-if optimization
4045 set_result(result);
4046 return true;
4047 }
4048
4049 //------------------------inline_array_copyOf----------------------------
4050 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
4051 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
4052 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
4053 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
4054
4055 // Get the arguments.
4056 Node* original = argument(0);
4057 Node* start = is_copyOfRange? argument(1): intcon(0);
4058 Node* end = is_copyOfRange? argument(2): argument(1);
4059 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
4060
4061 Node* newcopy = NULL;
4062
4063 // Set the original stack and the reexecute bit for the interpreter to reexecute
4064 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
4065 { PreserveReexecuteState preexecs(this);
4066 jvms()->set_should_reexecute(true);
4067
4068 array_type_mirror = null_check(array_type_mirror);
4069 original = null_check(original);
4070
4071 // Check if a null path was taken unconditionally.
4072 if (stopped()) return true;
4073
4074 Node* orig_length = load_array_length(original);
4075
4076 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
4077 klass_node = null_check(klass_node);
4078
4079 RegionNode* bailout = new RegionNode(1);
4080 record_for_igvn(bailout);
4081
4082 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
4083 // Bail out if that is so.
4084 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
4085 if (not_objArray != NULL) {
4086 // Improve the klass node's type from the new optimistic assumption:
4087 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
4088 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
4089 Node* cast = new CastPPNode(klass_node, akls);
4090 cast->init_req(0, control());
4091 klass_node = _gvn.transform(cast);
4092 }
4093
4094 // Bail out if either start or end is negative.
4095 generate_negative_guard(start, bailout, &start);
4096 generate_negative_guard(end, bailout, &end);
4097
4098 Node* length = end;
4099 if (_gvn.type(start) != TypeInt::ZERO) {
4100 length = _gvn.transform(new SubINode(end, start));
4101 }
4102
4103 // Bail out if length is negative.
4104 // Without this the new_array would throw
4105 // NegativeArraySizeException but IllegalArgumentException is what
4106 // should be thrown
4107 generate_negative_guard(length, bailout, &length);
4108
4109 if (bailout->req() > 1) {
4110 PreserveJVMState pjvms(this);
4111 set_control(_gvn.transform(bailout));
4112 uncommon_trap(Deoptimization::Reason_intrinsic,
4113 Deoptimization::Action_maybe_recompile);
4114 }
4115
4116 if (!stopped()) {
4117 // How many elements will we copy from the original?
4118 // The answer is MinI(orig_length - start, length).
4119 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
4120 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
4121
4122 // Generate a direct call to the right arraycopy function(s).
4123 // We know the copy is disjoint but we might not know if the
4124 // oop stores need checking.
4125 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
4126 // This will fail a store-check if x contains any non-nulls.
4127
4128 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
4129 // loads/stores but it is legal only if we're sure the
4130 // Arrays.copyOf would succeed. So we need all input arguments
4131 // to the copyOf to be validated, including that the copy to the
4132 // new array won't trigger an ArrayStoreException. That subtype
4133 // check can be optimized if we know something on the type of
4134 // the input array from type speculation.
4135 if (_gvn.type(klass_node)->singleton()) {
4136 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
4137 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
4138
4139 int test = C->static_subtype_check(superk, subk);
4140 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
4141 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
4142 if (t_original->speculative_type() != NULL) {
4143 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
4144 }
4145 }
4146 }
4147
4148 bool validated = false;
4149 // Reason_class_check rather than Reason_intrinsic because we
4150 // want to intrinsify even if this traps.
4151 if (!too_many_traps(Deoptimization::Reason_class_check)) {
4152 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
4153 klass_node);
4154
4155 if (not_subtype_ctrl != top()) {
4156 PreserveJVMState pjvms(this);
4157 set_control(not_subtype_ctrl);
4158 uncommon_trap(Deoptimization::Reason_class_check,
4159 Deoptimization::Action_make_not_entrant);
4160 assert(stopped(), "Should be stopped");
4161 }
4162 validated = true;
4163 }
4164
4165 if (!stopped()) {
4166 newcopy = new_array(klass_node, length, 0); // no arguments to push
4167
4168 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true, false,
4169 load_object_klass(original), klass_node);
4170 if (!is_copyOfRange) {
4171 ac->set_copyof(validated);
4172 } else {
4173 ac->set_copyofrange(validated);
4174 }
4175 Node* n = _gvn.transform(ac);
4176 if (n == ac) {
4177 ac->connect_outputs(this);
4178 } else {
4179 assert(validated, "shouldn't transform if all arguments not validated");
4180 set_all_memory(n);
4181 }
4182 }
4183 }
4184 } // original reexecute is set back here
4185
4186 C->set_has_split_ifs(true); // Has chance for split-if optimization
4187 if (!stopped()) {
4188 set_result(newcopy);
4189 }
4190 return true;
4191 }
4192
4193
4194 //----------------------generate_virtual_guard---------------------------
4195 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
4196 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
4197 RegionNode* slow_region) {
4198 ciMethod* method = callee();
4199 int vtable_index = method->vtable_index();
4200 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4201 "bad index %d", vtable_index);
4202 // Get the Method* out of the appropriate vtable entry.
4203 int entry_offset = in_bytes(Klass::vtable_start_offset()) +
4204 vtable_index*vtableEntry::size_in_bytes() +
4205 vtableEntry::method_offset_in_bytes();
4206 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
4207 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
4208
4209 // Compare the target method with the expected method (e.g., Object.hashCode).
4210 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
4211
4212 Node* native_call = makecon(native_call_addr);
4213 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
4214 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
4215
4216 return generate_slow_guard(test_native, slow_region);
4217 }
4218
4219 //-----------------------generate_method_call----------------------------
4220 // Use generate_method_call to make a slow-call to the real
4221 // method if the fast path fails. An alternative would be to
4222 // use a stub like OptoRuntime::slow_arraycopy_Java.
4223 // This only works for expanding the current library call,
4224 // not another intrinsic. (E.g., don't use this for making an
4225 // arraycopy call inside of the copyOf intrinsic.)
4226 CallJavaNode*
4227 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
4228 // When compiling the intrinsic method itself, do not use this technique.
4229 guarantee(callee() != C->method(), "cannot make slow-call to self");
4230
4231 ciMethod* method = callee();
4232 // ensure the JVMS we have will be correct for this call
4233 guarantee(method_id == method->intrinsic_id(), "must match");
4234
4235 const TypeFunc* tf = TypeFunc::make(method);
4236 CallJavaNode* slow_call;
4237 if (is_static) {
4238 assert(!is_virtual, "");
4239 slow_call = new CallStaticJavaNode(C, tf,
4240 SharedRuntime::get_resolve_static_call_stub(),
4241 method, bci());
4242 } else if (is_virtual) {
4243 null_check_receiver();
4244 int vtable_index = Method::invalid_vtable_index;
4245 if (UseInlineCaches) {
4246 // Suppress the vtable call
4247 } else {
4248 // hashCode and clone are not a miranda methods,
4249 // so the vtable index is fixed.
4250 // No need to use the linkResolver to get it.
4251 vtable_index = method->vtable_index();
4252 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
4253 "bad index %d", vtable_index);
4254 }
4255 slow_call = new CallDynamicJavaNode(tf,
4256 SharedRuntime::get_resolve_virtual_call_stub(),
4257 method, vtable_index, bci());
4258 } else { // neither virtual nor static: opt_virtual
4259 null_check_receiver();
4260 slow_call = new CallStaticJavaNode(C, tf,
4261 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4262 method, bci());
4263 slow_call->set_optimized_virtual(true);
4264 }
4265 set_arguments_for_java_call(slow_call);
4266 set_edges_for_java_call(slow_call);
4267 return slow_call;
4268 }
4269
4270
4271 /**
4272 * Build special case code for calls to hashCode on an object. This call may
4273 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4274 * slightly different code.
4275 */
4276 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4277 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4278 assert(!(is_virtual && is_static), "either virtual, special, or static");
4279
4280 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4281
4282 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4283 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4284 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4285 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4286 Node* obj = NULL;
4287 if (!is_static) {
4288 // Check for hashing null object
4289 obj = null_check_receiver();
4290 if (stopped()) return true; // unconditionally null
4291 result_reg->init_req(_null_path, top());
4292 result_val->init_req(_null_path, top());
4293 } else {
4294 // Do a null check, and return zero if null.
4295 // System.identityHashCode(null) == 0
4296 obj = argument(0);
4297 Node* null_ctl = top();
4298 obj = null_check_oop(obj, &null_ctl);
4299 result_reg->init_req(_null_path, null_ctl);
4300 result_val->init_req(_null_path, _gvn.intcon(0));
4301 }
4302
4303 // Unconditionally null? Then return right away.
4304 if (stopped()) {
4305 set_control( result_reg->in(_null_path));
4306 if (!stopped())
4307 set_result(result_val->in(_null_path));
4308 return true;
4309 }
4310
4311 // We only go to the fast case code if we pass a number of guards. The
4312 // paths which do not pass are accumulated in the slow_region.
4313 RegionNode* slow_region = new RegionNode(1);
4314 record_for_igvn(slow_region);
4315
4316 // If this is a virtual call, we generate a funny guard. We pull out
4317 // the vtable entry corresponding to hashCode() from the target object.
4318 // If the target method which we are calling happens to be the native
4319 // Object hashCode() method, we pass the guard. We do not need this
4320 // guard for non-virtual calls -- the caller is known to be the native
4321 // Object hashCode().
4322 if (is_virtual) {
4323 // After null check, get the object's klass.
4324 Node* obj_klass = load_object_klass(obj);
4325 generate_virtual_guard(obj_klass, slow_region);
4326 }
4327
4328 // Get the header out of the object, use LoadMarkNode when available
4329 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4330 // The control of the load must be NULL. Otherwise, the load can move before
4331 // the null check after castPP removal.
4332 Node* no_ctrl = NULL;
4333 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4334
4335 // Test the header to see if it is unlocked.
4336 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4337 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4338 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4339 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4340 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4341
4342 generate_slow_guard(test_unlocked, slow_region);
4343
4344 // Get the hash value and check to see that it has been properly assigned.
4345 // We depend on hash_mask being at most 32 bits and avoid the use of
4346 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4347 // vm: see markOop.hpp.
4348 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4349 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4350 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4351 // This hack lets the hash bits live anywhere in the mark object now, as long
4352 // as the shift drops the relevant bits into the low 32 bits. Note that
4353 // Java spec says that HashCode is an int so there's no point in capturing
4354 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4355 hshifted_header = ConvX2I(hshifted_header);
4356 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4357
4358 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4359 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4360 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4361
4362 generate_slow_guard(test_assigned, slow_region);
4363
4364 Node* init_mem = reset_memory();
4365 // fill in the rest of the null path:
4366 result_io ->init_req(_null_path, i_o());
4367 result_mem->init_req(_null_path, init_mem);
4368
4369 result_val->init_req(_fast_path, hash_val);
4370 result_reg->init_req(_fast_path, control());
4371 result_io ->init_req(_fast_path, i_o());
4372 result_mem->init_req(_fast_path, init_mem);
4373
4374 // Generate code for the slow case. We make a call to hashCode().
4375 set_control(_gvn.transform(slow_region));
4376 if (!stopped()) {
4377 // No need for PreserveJVMState, because we're using up the present state.
4378 set_all_memory(init_mem);
4379 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4380 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4381 Node* slow_result = set_results_for_java_call(slow_call);
4382 // this->control() comes from set_results_for_java_call
4383 result_reg->init_req(_slow_path, control());
4384 result_val->init_req(_slow_path, slow_result);
4385 result_io ->set_req(_slow_path, i_o());
4386 result_mem ->set_req(_slow_path, reset_memory());
4387 }
4388
4389 // Return the combined state.
4390 set_i_o( _gvn.transform(result_io) );
4391 set_all_memory( _gvn.transform(result_mem));
4392
4393 set_result(result_reg, result_val);
4394 return true;
4395 }
4396
4397 //---------------------------inline_native_getClass----------------------------
4398 // public final native Class<?> java.lang.Object.getClass();
4399 //
4400 // Build special case code for calls to getClass on an object.
4401 bool LibraryCallKit::inline_native_getClass() {
4402 Node* obj = null_check_receiver();
4403 if (stopped()) return true;
4404 set_result(load_mirror_from_klass(load_object_klass(obj)));
4405 return true;
4406 }
4407
4408 //-----------------inline_native_Reflection_getCallerClass---------------------
4409 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4410 //
4411 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4412 //
4413 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4414 // in that it must skip particular security frames and checks for
4415 // caller sensitive methods.
4416 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4417 #ifndef PRODUCT
4418 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4419 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4420 }
4421 #endif
4422
4423 if (!jvms()->has_method()) {
4424 #ifndef PRODUCT
4425 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4426 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4427 }
4428 #endif
4429 return false;
4430 }
4431
4432 // Walk back up the JVM state to find the caller at the required
4433 // depth.
4434 JVMState* caller_jvms = jvms();
4435
4436 // Cf. JVM_GetCallerClass
4437 // NOTE: Start the loop at depth 1 because the current JVM state does
4438 // not include the Reflection.getCallerClass() frame.
4439 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4440 ciMethod* m = caller_jvms->method();
4441 switch (n) {
4442 case 0:
4443 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4444 break;
4445 case 1:
4446 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4447 if (!m->caller_sensitive()) {
4448 #ifndef PRODUCT
4449 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4450 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4451 }
4452 #endif
4453 return false; // bail-out; let JVM_GetCallerClass do the work
4454 }
4455 break;
4456 default:
4457 if (!m->is_ignored_by_security_stack_walk()) {
4458 // We have reached the desired frame; return the holder class.
4459 // Acquire method holder as java.lang.Class and push as constant.
4460 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4461 ciInstance* caller_mirror = caller_klass->java_mirror();
4462 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4463
4464 #ifndef PRODUCT
4465 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4466 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());
4467 tty->print_cr(" JVM state at this point:");
4468 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4469 ciMethod* m = jvms()->of_depth(i)->method();
4470 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4471 }
4472 }
4473 #endif
4474 return true;
4475 }
4476 break;
4477 }
4478 }
4479
4480 #ifndef PRODUCT
4481 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4482 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4483 tty->print_cr(" JVM state at this point:");
4484 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4485 ciMethod* m = jvms()->of_depth(i)->method();
4486 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4487 }
4488 }
4489 #endif
4490
4491 return false; // bail-out; let JVM_GetCallerClass do the work
4492 }
4493
4494 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4495 Node* arg = argument(0);
4496 Node* result = NULL;
4497
4498 switch (id) {
4499 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4500 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4501 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4502 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4503
4504 case vmIntrinsics::_doubleToLongBits: {
4505 // two paths (plus control) merge in a wood
4506 RegionNode *r = new RegionNode(3);
4507 Node *phi = new PhiNode(r, TypeLong::LONG);
4508
4509 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4510 // Build the boolean node
4511 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4512
4513 // Branch either way.
4514 // NaN case is less traveled, which makes all the difference.
4515 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4516 Node *opt_isnan = _gvn.transform(ifisnan);
4517 assert( opt_isnan->is_If(), "Expect an IfNode");
4518 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4519 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4520
4521 set_control(iftrue);
4522
4523 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4524 Node *slow_result = longcon(nan_bits); // return NaN
4525 phi->init_req(1, _gvn.transform( slow_result ));
4526 r->init_req(1, iftrue);
4527
4528 // Else fall through
4529 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4530 set_control(iffalse);
4531
4532 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4533 r->init_req(2, iffalse);
4534
4535 // Post merge
4536 set_control(_gvn.transform(r));
4537 record_for_igvn(r);
4538
4539 C->set_has_split_ifs(true); // Has chance for split-if optimization
4540 result = phi;
4541 assert(result->bottom_type()->isa_long(), "must be");
4542 break;
4543 }
4544
4545 case vmIntrinsics::_floatToIntBits: {
4546 // two paths (plus control) merge in a wood
4547 RegionNode *r = new RegionNode(3);
4548 Node *phi = new PhiNode(r, TypeInt::INT);
4549
4550 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4551 // Build the boolean node
4552 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4553
4554 // Branch either way.
4555 // NaN case is less traveled, which makes all the difference.
4556 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4557 Node *opt_isnan = _gvn.transform(ifisnan);
4558 assert( opt_isnan->is_If(), "Expect an IfNode");
4559 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4560 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4561
4562 set_control(iftrue);
4563
4564 static const jint nan_bits = 0x7fc00000;
4565 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4566 phi->init_req(1, _gvn.transform( slow_result ));
4567 r->init_req(1, iftrue);
4568
4569 // Else fall through
4570 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4571 set_control(iffalse);
4572
4573 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4574 r->init_req(2, iffalse);
4575
4576 // Post merge
4577 set_control(_gvn.transform(r));
4578 record_for_igvn(r);
4579
4580 C->set_has_split_ifs(true); // Has chance for split-if optimization
4581 result = phi;
4582 assert(result->bottom_type()->isa_int(), "must be");
4583 break;
4584 }
4585
4586 default:
4587 fatal_unexpected_iid(id);
4588 break;
4589 }
4590 set_result(_gvn.transform(result));
4591 return true;
4592 }
4593
4594 //----------------------inline_unsafe_copyMemory-------------------------
4595 // public native void Unsafe.copyMemory0(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4596 bool LibraryCallKit::inline_unsafe_copyMemory() {
4597 if (callee()->is_static()) return false; // caller must have the capability!
4598 null_check_receiver(); // null-check receiver
4599 if (stopped()) return true;
4600
4601 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4602
4603 Node* src_ptr = argument(1); // type: oop
4604 Node* src_off = ConvL2X(argument(2)); // type: long
4605 Node* dst_ptr = argument(4); // type: oop
4606 Node* dst_off = ConvL2X(argument(5)); // type: long
4607 Node* size = ConvL2X(argument(7)); // type: long
4608
4609 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4610 "fieldOffset must be byte-scaled");
4611
4612 Node* src = make_unsafe_address(src_ptr, src_off);
4613 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4614
4615 // Conservatively insert a memory barrier on all memory slices.
4616 // Do not let writes of the copy source or destination float below the copy.
4617 insert_mem_bar(Op_MemBarCPUOrder);
4618
4619 // Call it. Note that the length argument is not scaled.
4620 make_runtime_call(RC_LEAF|RC_NO_FP,
4621 OptoRuntime::fast_arraycopy_Type(),
4622 StubRoutines::unsafe_arraycopy(),
4623 "unsafe_arraycopy",
4624 TypeRawPtr::BOTTOM,
4625 src, dst, size XTOP);
4626
4627 // Do not let reads of the copy destination float above the copy.
4628 insert_mem_bar(Op_MemBarCPUOrder);
4629
4630 return true;
4631 }
4632
4633 //------------------------clone_coping-----------------------------------
4634 // Helper function for inline_native_clone.
4635 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4636 assert(obj_size != NULL, "");
4637 Node* raw_obj = alloc_obj->in(1);
4638 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4639
4640 AllocateNode* alloc = NULL;
4641 if (ReduceBulkZeroing) {
4642 // We will be completely responsible for initializing this object -
4643 // mark Initialize node as complete.
4644 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4645 // The object was just allocated - there should be no any stores!
4646 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4647 // Mark as complete_with_arraycopy so that on AllocateNode
4648 // expansion, we know this AllocateNode is initialized by an array
4649 // copy and a StoreStore barrier exists after the array copy.
4650 alloc->initialization()->set_complete_with_arraycopy();
4651 }
4652
4653 // Copy the fastest available way.
4654 // TODO: generate fields copies for small objects instead.
4655 Node* src = obj;
4656 Node* dest = alloc_obj;
4657 Node* size = _gvn.transform(obj_size);
4658
4659 // Exclude the header but include array length to copy by 8 bytes words.
4660 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4661 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4662 instanceOopDesc::base_offset_in_bytes();
4663 // base_off:
4664 // 8 - 32-bit VM
4665 // 12 - 64-bit VM, compressed klass
4666 // 16 - 64-bit VM, normal klass
4667 if (base_off % BytesPerLong != 0) {
4668 assert(UseCompressedClassPointers, "");
4669 if (is_array) {
4670 // Exclude length to copy by 8 bytes words.
4671 base_off += sizeof(int);
4672 } else {
4673 // Include klass to copy by 8 bytes words.
4674 base_off = instanceOopDesc::klass_offset_in_bytes();
4675 }
4676 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4677 }
4678 src = basic_plus_adr(src, base_off);
4679 dest = basic_plus_adr(dest, base_off);
4680
4681 // Compute the length also, if needed:
4682 Node* countx = size;
4683 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4684 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4685
4686 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4687
4688 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false, false);
4689 ac->set_clonebasic();
4690 Node* n = _gvn.transform(ac);
4691 if (n == ac) {
4692 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4693 } else {
4694 set_all_memory(n);
4695 }
4696
4697 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4698 if (card_mark) {
4699 assert(!is_array, "");
4700 // Put in store barrier for any and all oops we are sticking
4701 // into this object. (We could avoid this if we could prove
4702 // that the object type contains no oop fields at all.)
4703 Node* no_particular_value = NULL;
4704 Node* no_particular_field = NULL;
4705 int raw_adr_idx = Compile::AliasIdxRaw;
4706 post_barrier(control(),
4707 memory(raw_adr_type),
4708 alloc_obj,
4709 no_particular_field,
4710 raw_adr_idx,
4711 no_particular_value,
4712 T_OBJECT,
4713 false);
4714 }
4715
4716 // Do not let reads from the cloned object float above the arraycopy.
4717 if (alloc != NULL) {
4718 // Do not let stores that initialize this object be reordered with
4719 // a subsequent store that would make this object accessible by
4720 // other threads.
4721 // Record what AllocateNode this StoreStore protects so that
4722 // escape analysis can go from the MemBarStoreStoreNode to the
4723 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4724 // based on the escape status of the AllocateNode.
4725 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4726 } else {
4727 insert_mem_bar(Op_MemBarCPUOrder);
4728 }
4729 }
4730
4731 //------------------------inline_native_clone----------------------------
4732 // protected native Object java.lang.Object.clone();
4733 //
4734 // Here are the simple edge cases:
4735 // null receiver => normal trap
4736 // virtual and clone was overridden => slow path to out-of-line clone
4737 // not cloneable or finalizer => slow path to out-of-line Object.clone
4738 //
4739 // The general case has two steps, allocation and copying.
4740 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4741 //
4742 // Copying also has two cases, oop arrays and everything else.
4743 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4744 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4745 //
4746 // These steps fold up nicely if and when the cloned object's klass
4747 // can be sharply typed as an object array, a type array, or an instance.
4748 //
4749 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4750 PhiNode* result_val;
4751
4752 // Set the reexecute bit for the interpreter to reexecute
4753 // the bytecode that invokes Object.clone if deoptimization happens.
4754 { PreserveReexecuteState preexecs(this);
4755 jvms()->set_should_reexecute(true);
4756
4757 Node* obj = null_check_receiver();
4758 if (stopped()) return true;
4759
4760 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4761
4762 // If we are going to clone an instance, we need its exact type to
4763 // know the number and types of fields to convert the clone to
4764 // loads/stores. Maybe a speculative type can help us.
4765 if (!obj_type->klass_is_exact() &&
4766 obj_type->speculative_type() != NULL &&
4767 obj_type->speculative_type()->is_instance_klass()) {
4768 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4769 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4770 !spec_ik->has_injected_fields()) {
4771 ciKlass* k = obj_type->klass();
4772 if (!k->is_instance_klass() ||
4773 k->as_instance_klass()->is_interface() ||
4774 k->as_instance_klass()->has_subklass()) {
4775 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4776 }
4777 }
4778 }
4779
4780 Node* obj_klass = load_object_klass(obj);
4781 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4782 const TypeOopPtr* toop = ((tklass != NULL)
4783 ? tklass->as_instance_type()
4784 : TypeInstPtr::NOTNULL);
4785
4786 // Conservatively insert a memory barrier on all memory slices.
4787 // Do not let writes into the original float below the clone.
4788 insert_mem_bar(Op_MemBarCPUOrder);
4789
4790 // paths into result_reg:
4791 enum {
4792 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4793 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4794 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4795 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4796 PATH_LIMIT
4797 };
4798 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4799 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4800 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4801 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4802 record_for_igvn(result_reg);
4803
4804 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4805 int raw_adr_idx = Compile::AliasIdxRaw;
4806
4807 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4808 if (array_ctl != NULL) {
4809 // It's an array.
4810 PreserveJVMState pjvms(this);
4811 set_control(array_ctl);
4812 Node* obj_length = load_array_length(obj);
4813 Node* obj_size = NULL;
4814 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4815
4816 if (!use_ReduceInitialCardMarks()) {
4817 // If it is an oop array, it requires very special treatment,
4818 // because card marking is required on each card of the array.
4819 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4820 if (is_obja != NULL) {
4821 PreserveJVMState pjvms2(this);
4822 set_control(is_obja);
4823 // Generate a direct call to the right arraycopy function(s).
4824 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4825 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL, false);
4826 ac->set_cloneoop();
4827 Node* n = _gvn.transform(ac);
4828 assert(n == ac, "cannot disappear");
4829 ac->connect_outputs(this);
4830
4831 result_reg->init_req(_objArray_path, control());
4832 result_val->init_req(_objArray_path, alloc_obj);
4833 result_i_o ->set_req(_objArray_path, i_o());
4834 result_mem ->set_req(_objArray_path, reset_memory());
4835 }
4836 }
4837 // Otherwise, there are no card marks to worry about.
4838 // (We can dispense with card marks if we know the allocation
4839 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4840 // causes the non-eden paths to take compensating steps to
4841 // simulate a fresh allocation, so that no further
4842 // card marks are required in compiled code to initialize
4843 // the object.)
4844
4845 if (!stopped()) {
4846 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4847
4848 // Present the results of the copy.
4849 result_reg->init_req(_array_path, control());
4850 result_val->init_req(_array_path, alloc_obj);
4851 result_i_o ->set_req(_array_path, i_o());
4852 result_mem ->set_req(_array_path, reset_memory());
4853 }
4854 }
4855
4856 // We only go to the instance fast case code if we pass a number of guards.
4857 // The paths which do not pass are accumulated in the slow_region.
4858 RegionNode* slow_region = new RegionNode(1);
4859 record_for_igvn(slow_region);
4860 if (!stopped()) {
4861 // It's an instance (we did array above). Make the slow-path tests.
4862 // If this is a virtual call, we generate a funny guard. We grab
4863 // the vtable entry corresponding to clone() from the target object.
4864 // If the target method which we are calling happens to be the
4865 // Object clone() method, we pass the guard. We do not need this
4866 // guard for non-virtual calls; the caller is known to be the native
4867 // Object clone().
4868 if (is_virtual) {
4869 generate_virtual_guard(obj_klass, slow_region);
4870 }
4871
4872 // The object must be easily cloneable and must not have a finalizer.
4873 // Both of these conditions may be checked in a single test.
4874 // We could optimize the test further, but we don't care.
4875 generate_access_flags_guard(obj_klass,
4876 // Test both conditions:
4877 JVM_ACC_IS_CLONEABLE_FAST | JVM_ACC_HAS_FINALIZER,
4878 // Must be cloneable but not finalizer:
4879 JVM_ACC_IS_CLONEABLE_FAST,
4880 slow_region);
4881 }
4882
4883 if (!stopped()) {
4884 // It's an instance, and it passed the slow-path tests.
4885 PreserveJVMState pjvms(this);
4886 Node* obj_size = NULL;
4887 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4888 // is reexecuted if deoptimization occurs and there could be problems when merging
4889 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4890 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4891
4892 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4893
4894 // Present the results of the slow call.
4895 result_reg->init_req(_instance_path, control());
4896 result_val->init_req(_instance_path, alloc_obj);
4897 result_i_o ->set_req(_instance_path, i_o());
4898 result_mem ->set_req(_instance_path, reset_memory());
4899 }
4900
4901 // Generate code for the slow case. We make a call to clone().
4902 set_control(_gvn.transform(slow_region));
4903 if (!stopped()) {
4904 PreserveJVMState pjvms(this);
4905 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4906 Node* slow_result = set_results_for_java_call(slow_call);
4907 // this->control() comes from set_results_for_java_call
4908 result_reg->init_req(_slow_path, control());
4909 result_val->init_req(_slow_path, slow_result);
4910 result_i_o ->set_req(_slow_path, i_o());
4911 result_mem ->set_req(_slow_path, reset_memory());
4912 }
4913
4914 // Return the combined state.
4915 set_control( _gvn.transform(result_reg));
4916 set_i_o( _gvn.transform(result_i_o));
4917 set_all_memory( _gvn.transform(result_mem));
4918 } // original reexecute is set back here
4919
4920 set_result(_gvn.transform(result_val));
4921 return true;
4922 }
4923
4924 // If we have a tighly coupled allocation, the arraycopy may take care
4925 // of the array initialization. If one of the guards we insert between
4926 // the allocation and the arraycopy causes a deoptimization, an
4927 // unitialized array will escape the compiled method. To prevent that
4928 // we set the JVM state for uncommon traps between the allocation and
4929 // the arraycopy to the state before the allocation so, in case of
4930 // deoptimization, we'll reexecute the allocation and the
4931 // initialization.
4932 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4933 if (alloc != NULL) {
4934 ciMethod* trap_method = alloc->jvms()->method();
4935 int trap_bci = alloc->jvms()->bci();
4936
4937 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4938 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4939 // Make sure there's no store between the allocation and the
4940 // arraycopy otherwise visible side effects could be rexecuted
4941 // in case of deoptimization and cause incorrect execution.
4942 bool no_interfering_store = true;
4943 Node* mem = alloc->in(TypeFunc::Memory);
4944 if (mem->is_MergeMem()) {
4945 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4946 Node* n = mms.memory();
4947 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4948 assert(n->is_Store(), "what else?");
4949 no_interfering_store = false;
4950 break;
4951 }
4952 }
4953 } else {
4954 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4955 Node* n = mms.memory();
4956 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4957 assert(n->is_Store(), "what else?");
4958 no_interfering_store = false;
4959 break;
4960 }
4961 }
4962 }
4963
4964 if (no_interfering_store) {
4965 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4966 uint size = alloc->req();
4967 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4968 old_jvms->set_map(sfpt);
4969 for (uint i = 0; i < size; i++) {
4970 sfpt->init_req(i, alloc->in(i));
4971 }
4972 // re-push array length for deoptimization
4973 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4974 old_jvms->set_sp(old_jvms->sp()+1);
4975 old_jvms->set_monoff(old_jvms->monoff()+1);
4976 old_jvms->set_scloff(old_jvms->scloff()+1);
4977 old_jvms->set_endoff(old_jvms->endoff()+1);
4978 old_jvms->set_should_reexecute(true);
4979
4980 sfpt->set_i_o(map()->i_o());
4981 sfpt->set_memory(map()->memory());
4982 sfpt->set_control(map()->control());
4983
4984 JVMState* saved_jvms = jvms();
4985 saved_reexecute_sp = _reexecute_sp;
4986
4987 set_jvms(sfpt->jvms());
4988 _reexecute_sp = jvms()->sp();
4989
4990 return saved_jvms;
4991 }
4992 }
4993 }
4994 return NULL;
4995 }
4996
4997 // In case of a deoptimization, we restart execution at the
4998 // allocation, allocating a new array. We would leave an uninitialized
4999 // array in the heap that GCs wouldn't expect. Move the allocation
5000 // after the traps so we don't allocate the array if we
5001 // deoptimize. This is possible because tightly_coupled_allocation()
5002 // guarantees there's no observer of the allocated array at this point
5003 // and the control flow is simple enough.
5004 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms,
5005 int saved_reexecute_sp, uint new_idx) {
5006 if (saved_jvms != NULL && !stopped()) {
5007 assert(alloc != NULL, "only with a tightly coupled allocation");
5008 // restore JVM state to the state at the arraycopy
5009 saved_jvms->map()->set_control(map()->control());
5010 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
5011 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
5012 // If we've improved the types of some nodes (null check) while
5013 // emitting the guards, propagate them to the current state
5014 map()->replaced_nodes().apply(saved_jvms->map(), new_idx);
5015 set_jvms(saved_jvms);
5016 _reexecute_sp = saved_reexecute_sp;
5017
5018 // Remove the allocation from above the guards
5019 CallProjections callprojs;
5020 alloc->extract_projections(&callprojs, true);
5021 InitializeNode* init = alloc->initialization();
5022 Node* alloc_mem = alloc->in(TypeFunc::Memory);
5023 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
5024 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
5025 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
5026
5027 // move the allocation here (after the guards)
5028 _gvn.hash_delete(alloc);
5029 alloc->set_req(TypeFunc::Control, control());
5030 alloc->set_req(TypeFunc::I_O, i_o());
5031 Node *mem = reset_memory();
5032 set_all_memory(mem);
5033 alloc->set_req(TypeFunc::Memory, mem);
5034 set_control(init->proj_out(TypeFunc::Control));
5035 set_i_o(callprojs.fallthrough_ioproj);
5036
5037 // Update memory as done in GraphKit::set_output_for_allocation()
5038 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
5039 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
5040 if (ary_type->isa_aryptr() && length_type != NULL) {
5041 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
5042 }
5043 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
5044 int elemidx = C->get_alias_index(telemref);
5045 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
5046 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
5047
5048 Node* allocx = _gvn.transform(alloc);
5049 assert(allocx == alloc, "where has the allocation gone?");
5050 assert(dest->is_CheckCastPP(), "not an allocation result?");
5051
5052 _gvn.hash_delete(dest);
5053 dest->set_req(0, control());
5054 Node* destx = _gvn.transform(dest);
5055 assert(destx == dest, "where has the allocation result gone?");
5056 }
5057 }
5058
5059
5060 //------------------------------inline_arraycopy-----------------------
5061 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
5062 // Object dest, int destPos,
5063 // int length);
5064 bool LibraryCallKit::inline_arraycopy() {
5065 // Get the arguments.
5066 Node* src = argument(0); // type: oop
5067 Node* src_offset = argument(1); // type: int
5068 Node* dest = argument(2); // type: oop
5069 Node* dest_offset = argument(3); // type: int
5070 Node* length = argument(4); // type: int
5071
5072 uint new_idx = C->unique();
5073
5074 // Check for allocation before we add nodes that would confuse
5075 // tightly_coupled_allocation()
5076 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
5077
5078 int saved_reexecute_sp = -1;
5079 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
5080 // See arraycopy_restore_alloc_state() comment
5081 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
5082 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
5083 // if saved_jvms == NULL and alloc != NULL, we can't emit any guards
5084 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
5085
5086 // The following tests must be performed
5087 // (1) src and dest are arrays.
5088 // (2) src and dest arrays must have elements of the same BasicType
5089 // (3) src and dest must not be null.
5090 // (4) src_offset must not be negative.
5091 // (5) dest_offset must not be negative.
5092 // (6) length must not be negative.
5093 // (7) src_offset + length must not exceed length of src.
5094 // (8) dest_offset + length must not exceed length of dest.
5095 // (9) each element of an oop array must be assignable
5096
5097 // (3) src and dest must not be null.
5098 // always do this here because we need the JVM state for uncommon traps
5099 Node* null_ctl = top();
5100 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
5101 assert(null_ctl->is_top(), "no null control here");
5102 dest = null_check(dest, T_ARRAY);
5103
5104 if (!can_emit_guards) {
5105 // if saved_jvms == NULL and alloc != NULL, we don't emit any
5106 // guards but the arraycopy node could still take advantage of a
5107 // tightly allocated allocation. tightly_coupled_allocation() is
5108 // called again to make sure it takes the null check above into
5109 // account: the null check is mandatory and if it caused an
5110 // uncommon trap to be emitted then the allocation can't be
5111 // considered tightly coupled in this context.
5112 alloc = tightly_coupled_allocation(dest, NULL);
5113 }
5114
5115 bool validated = false;
5116
5117 const Type* src_type = _gvn.type(src);
5118 const Type* dest_type = _gvn.type(dest);
5119 const TypeAryPtr* top_src = src_type->isa_aryptr();
5120 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5121
5122 // Do we have the type of src?
5123 bool has_src = (top_src != NULL && top_src->klass() != NULL);
5124 // Do we have the type of dest?
5125 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5126 // Is the type for src from speculation?
5127 bool src_spec = false;
5128 // Is the type for dest from speculation?
5129 bool dest_spec = false;
5130
5131 if ((!has_src || !has_dest) && can_emit_guards) {
5132 // We don't have sufficient type information, let's see if
5133 // speculative types can help. We need to have types for both src
5134 // and dest so that it pays off.
5135
5136 // Do we already have or could we have type information for src
5137 bool could_have_src = has_src;
5138 // Do we already have or could we have type information for dest
5139 bool could_have_dest = has_dest;
5140
5141 ciKlass* src_k = NULL;
5142 if (!has_src) {
5143 src_k = src_type->speculative_type_not_null();
5144 if (src_k != NULL && src_k->is_array_klass()) {
5145 could_have_src = true;
5146 }
5147 }
5148
5149 ciKlass* dest_k = NULL;
5150 if (!has_dest) {
5151 dest_k = dest_type->speculative_type_not_null();
5152 if (dest_k != NULL && dest_k->is_array_klass()) {
5153 could_have_dest = true;
5154 }
5155 }
5156
5157 if (could_have_src && could_have_dest) {
5158 // This is going to pay off so emit the required guards
5159 if (!has_src) {
5160 src = maybe_cast_profiled_obj(src, src_k, true);
5161 src_type = _gvn.type(src);
5162 top_src = src_type->isa_aryptr();
5163 has_src = (top_src != NULL && top_src->klass() != NULL);
5164 src_spec = true;
5165 }
5166 if (!has_dest) {
5167 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5168 dest_type = _gvn.type(dest);
5169 top_dest = dest_type->isa_aryptr();
5170 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
5171 dest_spec = true;
5172 }
5173 }
5174 }
5175
5176 if (has_src && has_dest && can_emit_guards) {
5177 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
5178 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
5179 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
5180 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
5181
5182 if (src_elem == dest_elem && src_elem == T_OBJECT) {
5183 // If both arrays are object arrays then having the exact types
5184 // for both will remove the need for a subtype check at runtime
5185 // before the call and may make it possible to pick a faster copy
5186 // routine (without a subtype check on every element)
5187 // Do we have the exact type of src?
5188 bool could_have_src = src_spec;
5189 // Do we have the exact type of dest?
5190 bool could_have_dest = dest_spec;
5191 ciKlass* src_k = top_src->klass();
5192 ciKlass* dest_k = top_dest->klass();
5193 if (!src_spec) {
5194 src_k = src_type->speculative_type_not_null();
5195 if (src_k != NULL && src_k->is_array_klass()) {
5196 could_have_src = true;
5197 }
5198 }
5199 if (!dest_spec) {
5200 dest_k = dest_type->speculative_type_not_null();
5201 if (dest_k != NULL && dest_k->is_array_klass()) {
5202 could_have_dest = true;
5203 }
5204 }
5205 if (could_have_src && could_have_dest) {
5206 // If we can have both exact types, emit the missing guards
5207 if (could_have_src && !src_spec) {
5208 src = maybe_cast_profiled_obj(src, src_k, true);
5209 }
5210 if (could_have_dest && !dest_spec) {
5211 dest = maybe_cast_profiled_obj(dest, dest_k, true);
5212 }
5213 }
5214 }
5215 }
5216
5217 ciMethod* trap_method = method();
5218 int trap_bci = bci();
5219 if (saved_jvms != NULL) {
5220 trap_method = alloc->jvms()->method();
5221 trap_bci = alloc->jvms()->bci();
5222 }
5223
5224 bool negative_length_guard_generated = false;
5225
5226 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
5227 can_emit_guards &&
5228 !src->is_top() && !dest->is_top()) {
5229 // validate arguments: enables transformation the ArrayCopyNode
5230 validated = true;
5231
5232 RegionNode* slow_region = new RegionNode(1);
5233 record_for_igvn(slow_region);
5234
5235 // (1) src and dest are arrays.
5236 generate_non_array_guard(load_object_klass(src), slow_region);
5237 generate_non_array_guard(load_object_klass(dest), slow_region);
5238
5239 // (2) src and dest arrays must have elements of the same BasicType
5240 // done at macro expansion or at Ideal transformation time
5241
5242 // (4) src_offset must not be negative.
5243 generate_negative_guard(src_offset, slow_region);
5244
5245 // (5) dest_offset must not be negative.
5246 generate_negative_guard(dest_offset, slow_region);
5247
5248 // (7) src_offset + length must not exceed length of src.
5249 generate_limit_guard(src_offset, length,
5250 load_array_length(src),
5251 slow_region);
5252
5253 // (8) dest_offset + length must not exceed length of dest.
5254 generate_limit_guard(dest_offset, length,
5255 load_array_length(dest),
5256 slow_region);
5257
5258 // (6) length must not be negative.
5259 // This is also checked in generate_arraycopy() during macro expansion, but
5260 // we also have to check it here for the case where the ArrayCopyNode will
5261 // be eliminated by Escape Analysis.
5262 if (EliminateAllocations) {
5263 generate_negative_guard(length, slow_region);
5264 negative_length_guard_generated = true;
5265 }
5266
5267 // (9) each element of an oop array must be assignable
5268 Node* src_klass = load_object_klass(src);
5269 Node* dest_klass = load_object_klass(dest);
5270 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5271
5272 if (not_subtype_ctrl != top()) {
5273 PreserveJVMState pjvms(this);
5274 set_control(not_subtype_ctrl);
5275 uncommon_trap(Deoptimization::Reason_intrinsic,
5276 Deoptimization::Action_make_not_entrant);
5277 assert(stopped(), "Should be stopped");
5278 }
5279 {
5280 PreserveJVMState pjvms(this);
5281 set_control(_gvn.transform(slow_region));
5282 uncommon_trap(Deoptimization::Reason_intrinsic,
5283 Deoptimization::Action_make_not_entrant);
5284 assert(stopped(), "Should be stopped");
5285 }
5286
5287 const TypeKlassPtr* dest_klass_t = _gvn.type(dest_klass)->is_klassptr();
5288 const Type *toop = TypeOopPtr::make_from_klass(dest_klass_t->klass());
5289 src = _gvn.transform(new CheckCastPPNode(control(), src, toop));
5290 }
5291
5292 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp, new_idx);
5293
5294 if (stopped()) {
5295 return true;
5296 }
5297
5298 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL, negative_length_guard_generated,
5299 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5300 // so the compiler has a chance to eliminate them: during macro expansion,
5301 // we have to set their control (CastPP nodes are eliminated).
5302 load_object_klass(src), load_object_klass(dest),
5303 load_array_length(src), load_array_length(dest));
5304
5305 ac->set_arraycopy(validated);
5306
5307 Node* n = _gvn.transform(ac);
5308 if (n == ac) {
5309 ac->connect_outputs(this);
5310 } else {
5311 assert(validated, "shouldn't transform if all arguments not validated");
5312 set_all_memory(n);
5313 }
5314 clear_upper_avx();
5315
5316
5317 return true;
5318 }
5319
5320
5321 // Helper function which determines if an arraycopy immediately follows
5322 // an allocation, with no intervening tests or other escapes for the object.
5323 AllocateArrayNode*
5324 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5325 RegionNode* slow_region) {
5326 if (stopped()) return NULL; // no fast path
5327 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5328
5329 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5330 if (alloc == NULL) return NULL;
5331
5332 Node* rawmem = memory(Compile::AliasIdxRaw);
5333 // Is the allocation's memory state untouched?
5334 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5335 // Bail out if there have been raw-memory effects since the allocation.
5336 // (Example: There might have been a call or safepoint.)
5337 return NULL;
5338 }
5339 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5340 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5341 return NULL;
5342 }
5343
5344 // There must be no unexpected observers of this allocation.
5345 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5346 Node* obs = ptr->fast_out(i);
5347 if (obs != this->map()) {
5348 return NULL;
5349 }
5350 }
5351
5352 // This arraycopy must unconditionally follow the allocation of the ptr.
5353 Node* alloc_ctl = ptr->in(0);
5354 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5355
5356 Node* ctl = control();
5357 while (ctl != alloc_ctl) {
5358 // There may be guards which feed into the slow_region.
5359 // Any other control flow means that we might not get a chance
5360 // to finish initializing the allocated object.
5361 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5362 IfNode* iff = ctl->in(0)->as_If();
5363 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5364 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5365 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5366 ctl = iff->in(0); // This test feeds the known slow_region.
5367 continue;
5368 }
5369 // One more try: Various low-level checks bottom out in
5370 // uncommon traps. If the debug-info of the trap omits
5371 // any reference to the allocation, as we've already
5372 // observed, then there can be no objection to the trap.
5373 bool found_trap = false;
5374 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5375 Node* obs = not_ctl->fast_out(j);
5376 if (obs->in(0) == not_ctl && obs->is_Call() &&
5377 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5378 found_trap = true; break;
5379 }
5380 }
5381 if (found_trap) {
5382 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5383 continue;
5384 }
5385 }
5386 return NULL;
5387 }
5388
5389 // If we get this far, we have an allocation which immediately
5390 // precedes the arraycopy, and we can take over zeroing the new object.
5391 // The arraycopy will finish the initialization, and provide
5392 // a new control state to which we will anchor the destination pointer.
5393
5394 return alloc;
5395 }
5396
5397 //-------------inline_encodeISOArray-----------------------------------
5398 // encode char[] to byte[] in ISO_8859_1
5399 bool LibraryCallKit::inline_encodeISOArray() {
5400 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5401 // no receiver since it is static method
5402 Node *src = argument(0);
5403 Node *src_offset = argument(1);
5404 Node *dst = argument(2);
5405 Node *dst_offset = argument(3);
5406 Node *length = argument(4);
5407
5408 const Type* src_type = src->Value(&_gvn);
5409 const Type* dst_type = dst->Value(&_gvn);
5410 const TypeAryPtr* top_src = src_type->isa_aryptr();
5411 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5412 if (top_src == NULL || top_src->klass() == NULL ||
5413 top_dest == NULL || top_dest->klass() == NULL) {
5414 // failed array check
5415 return false;
5416 }
5417
5418 // Figure out the size and type of the elements we will be copying.
5419 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5420 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5421 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5422 return false;
5423 }
5424
5425 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5426 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5427 // 'src_start' points to src array + scaled offset
5428 // 'dst_start' points to dst array + scaled offset
5429
5430 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5431 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5432 enc = _gvn.transform(enc);
5433 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5434 set_memory(res_mem, mtype);
5435 set_result(enc);
5436 clear_upper_avx();
5437
5438 return true;
5439 }
5440
5441 //-------------inline_multiplyToLen-----------------------------------
5442 bool LibraryCallKit::inline_multiplyToLen() {
5443 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5444
5445 address stubAddr = StubRoutines::multiplyToLen();
5446 if (stubAddr == NULL) {
5447 return false; // Intrinsic's stub is not implemented on this platform
5448 }
5449 const char* stubName = "multiplyToLen";
5450
5451 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5452
5453 // no receiver because it is a static method
5454 Node* x = argument(0);
5455 Node* xlen = argument(1);
5456 Node* y = argument(2);
5457 Node* ylen = argument(3);
5458 Node* z = argument(4);
5459
5460 const Type* x_type = x->Value(&_gvn);
5461 const Type* y_type = y->Value(&_gvn);
5462 const TypeAryPtr* top_x = x_type->isa_aryptr();
5463 const TypeAryPtr* top_y = y_type->isa_aryptr();
5464 if (top_x == NULL || top_x->klass() == NULL ||
5465 top_y == NULL || top_y->klass() == NULL) {
5466 // failed array check
5467 return false;
5468 }
5469
5470 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5471 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5472 if (x_elem != T_INT || y_elem != T_INT) {
5473 return false;
5474 }
5475
5476 // Set the original stack and the reexecute bit for the interpreter to reexecute
5477 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5478 // on the return from z array allocation in runtime.
5479 { PreserveReexecuteState preexecs(this);
5480 jvms()->set_should_reexecute(true);
5481
5482 Node* x_start = array_element_address(x, intcon(0), x_elem);
5483 Node* y_start = array_element_address(y, intcon(0), y_elem);
5484 // 'x_start' points to x array + scaled xlen
5485 // 'y_start' points to y array + scaled ylen
5486
5487 // Allocate the result array
5488 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5489 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5490 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5491
5492 IdealKit ideal(this);
5493
5494 #define __ ideal.
5495 Node* one = __ ConI(1);
5496 Node* zero = __ ConI(0);
5497 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5498 __ set(need_alloc, zero);
5499 __ set(z_alloc, z);
5500 __ if_then(z, BoolTest::eq, null()); {
5501 __ increment (need_alloc, one);
5502 } __ else_(); {
5503 // Update graphKit memory and control from IdealKit.
5504 sync_kit(ideal);
5505 Node* zlen_arg = load_array_length(z);
5506 // Update IdealKit memory and control from graphKit.
5507 __ sync_kit(this);
5508 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5509 __ increment (need_alloc, one);
5510 } __ end_if();
5511 } __ end_if();
5512
5513 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5514 // Update graphKit memory and control from IdealKit.
5515 sync_kit(ideal);
5516 Node * narr = new_array(klass_node, zlen, 1);
5517 // Update IdealKit memory and control from graphKit.
5518 __ sync_kit(this);
5519 __ set(z_alloc, narr);
5520 } __ end_if();
5521
5522 sync_kit(ideal);
5523 z = __ value(z_alloc);
5524 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5525 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5526 // Final sync IdealKit and GraphKit.
5527 final_sync(ideal);
5528 #undef __
5529
5530 Node* z_start = array_element_address(z, intcon(0), T_INT);
5531
5532 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5533 OptoRuntime::multiplyToLen_Type(),
5534 stubAddr, stubName, TypePtr::BOTTOM,
5535 x_start, xlen, y_start, ylen, z_start, zlen);
5536 } // original reexecute is set back here
5537
5538 C->set_has_split_ifs(true); // Has chance for split-if optimization
5539 set_result(z);
5540 return true;
5541 }
5542
5543 //-------------inline_squareToLen------------------------------------
5544 bool LibraryCallKit::inline_squareToLen() {
5545 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5546
5547 address stubAddr = StubRoutines::squareToLen();
5548 if (stubAddr == NULL) {
5549 return false; // Intrinsic's stub is not implemented on this platform
5550 }
5551 const char* stubName = "squareToLen";
5552
5553 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5554
5555 Node* x = argument(0);
5556 Node* len = argument(1);
5557 Node* z = argument(2);
5558 Node* zlen = argument(3);
5559
5560 const Type* x_type = x->Value(&_gvn);
5561 const Type* z_type = z->Value(&_gvn);
5562 const TypeAryPtr* top_x = x_type->isa_aryptr();
5563 const TypeAryPtr* top_z = z_type->isa_aryptr();
5564 if (top_x == NULL || top_x->klass() == NULL ||
5565 top_z == NULL || top_z->klass() == NULL) {
5566 // failed array check
5567 return false;
5568 }
5569
5570 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5571 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5572 if (x_elem != T_INT || z_elem != T_INT) {
5573 return false;
5574 }
5575
5576
5577 Node* x_start = array_element_address(x, intcon(0), x_elem);
5578 Node* z_start = array_element_address(z, intcon(0), z_elem);
5579
5580 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5581 OptoRuntime::squareToLen_Type(),
5582 stubAddr, stubName, TypePtr::BOTTOM,
5583 x_start, len, z_start, zlen);
5584
5585 set_result(z);
5586 return true;
5587 }
5588
5589 //-------------inline_mulAdd------------------------------------------
5590 bool LibraryCallKit::inline_mulAdd() {
5591 assert(UseMulAddIntrinsic, "not implemented on this platform");
5592
5593 address stubAddr = StubRoutines::mulAdd();
5594 if (stubAddr == NULL) {
5595 return false; // Intrinsic's stub is not implemented on this platform
5596 }
5597 const char* stubName = "mulAdd";
5598
5599 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5600
5601 Node* out = argument(0);
5602 Node* in = argument(1);
5603 Node* offset = argument(2);
5604 Node* len = argument(3);
5605 Node* k = argument(4);
5606
5607 const Type* out_type = out->Value(&_gvn);
5608 const Type* in_type = in->Value(&_gvn);
5609 const TypeAryPtr* top_out = out_type->isa_aryptr();
5610 const TypeAryPtr* top_in = in_type->isa_aryptr();
5611 if (top_out == NULL || top_out->klass() == NULL ||
5612 top_in == NULL || top_in->klass() == NULL) {
5613 // failed array check
5614 return false;
5615 }
5616
5617 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5618 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5619 if (out_elem != T_INT || in_elem != T_INT) {
5620 return false;
5621 }
5622
5623 Node* outlen = load_array_length(out);
5624 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5625 Node* out_start = array_element_address(out, intcon(0), out_elem);
5626 Node* in_start = array_element_address(in, intcon(0), in_elem);
5627
5628 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5629 OptoRuntime::mulAdd_Type(),
5630 stubAddr, stubName, TypePtr::BOTTOM,
5631 out_start,in_start, new_offset, len, k);
5632 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5633 set_result(result);
5634 return true;
5635 }
5636
5637 //-------------inline_montgomeryMultiply-----------------------------------
5638 bool LibraryCallKit::inline_montgomeryMultiply() {
5639 address stubAddr = StubRoutines::montgomeryMultiply();
5640 if (stubAddr == NULL) {
5641 return false; // Intrinsic's stub is not implemented on this platform
5642 }
5643
5644 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5645 const char* stubName = "montgomery_multiply";
5646
5647 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5648
5649 Node* a = argument(0);
5650 Node* b = argument(1);
5651 Node* n = argument(2);
5652 Node* len = argument(3);
5653 Node* inv = argument(4);
5654 Node* m = argument(6);
5655
5656 const Type* a_type = a->Value(&_gvn);
5657 const TypeAryPtr* top_a = a_type->isa_aryptr();
5658 const Type* b_type = b->Value(&_gvn);
5659 const TypeAryPtr* top_b = b_type->isa_aryptr();
5660 const Type* n_type = a->Value(&_gvn);
5661 const TypeAryPtr* top_n = n_type->isa_aryptr();
5662 const Type* m_type = a->Value(&_gvn);
5663 const TypeAryPtr* top_m = m_type->isa_aryptr();
5664 if (top_a == NULL || top_a->klass() == NULL ||
5665 top_b == NULL || top_b->klass() == NULL ||
5666 top_n == NULL || top_n->klass() == NULL ||
5667 top_m == NULL || top_m->klass() == NULL) {
5668 // failed array check
5669 return false;
5670 }
5671
5672 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5673 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5674 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5675 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5676 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5677 return false;
5678 }
5679
5680 // Make the call
5681 {
5682 Node* a_start = array_element_address(a, intcon(0), a_elem);
5683 Node* b_start = array_element_address(b, intcon(0), b_elem);
5684 Node* n_start = array_element_address(n, intcon(0), n_elem);
5685 Node* m_start = array_element_address(m, intcon(0), m_elem);
5686
5687 Node* call = make_runtime_call(RC_LEAF,
5688 OptoRuntime::montgomeryMultiply_Type(),
5689 stubAddr, stubName, TypePtr::BOTTOM,
5690 a_start, b_start, n_start, len, inv, top(),
5691 m_start);
5692 set_result(m);
5693 }
5694
5695 return true;
5696 }
5697
5698 bool LibraryCallKit::inline_montgomerySquare() {
5699 address stubAddr = StubRoutines::montgomerySquare();
5700 if (stubAddr == NULL) {
5701 return false; // Intrinsic's stub is not implemented on this platform
5702 }
5703
5704 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5705 const char* stubName = "montgomery_square";
5706
5707 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5708
5709 Node* a = argument(0);
5710 Node* n = argument(1);
5711 Node* len = argument(2);
5712 Node* inv = argument(3);
5713 Node* m = argument(5);
5714
5715 const Type* a_type = a->Value(&_gvn);
5716 const TypeAryPtr* top_a = a_type->isa_aryptr();
5717 const Type* n_type = a->Value(&_gvn);
5718 const TypeAryPtr* top_n = n_type->isa_aryptr();
5719 const Type* m_type = a->Value(&_gvn);
5720 const TypeAryPtr* top_m = m_type->isa_aryptr();
5721 if (top_a == NULL || top_a->klass() == NULL ||
5722 top_n == NULL || top_n->klass() == NULL ||
5723 top_m == NULL || top_m->klass() == NULL) {
5724 // failed array check
5725 return false;
5726 }
5727
5728 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5729 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5730 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5731 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5732 return false;
5733 }
5734
5735 // Make the call
5736 {
5737 Node* a_start = array_element_address(a, intcon(0), a_elem);
5738 Node* n_start = array_element_address(n, intcon(0), n_elem);
5739 Node* m_start = array_element_address(m, intcon(0), m_elem);
5740
5741 Node* call = make_runtime_call(RC_LEAF,
5742 OptoRuntime::montgomerySquare_Type(),
5743 stubAddr, stubName, TypePtr::BOTTOM,
5744 a_start, n_start, len, inv, top(),
5745 m_start);
5746 set_result(m);
5747 }
5748
5749 return true;
5750 }
5751
5752 //-------------inline_vectorizedMismatch------------------------------
5753 bool LibraryCallKit::inline_vectorizedMismatch() {
5754 assert(UseVectorizedMismatchIntrinsic, "not implementated on this platform");
5755
5756 address stubAddr = StubRoutines::vectorizedMismatch();
5757 if (stubAddr == NULL) {
5758 return false; // Intrinsic's stub is not implemented on this platform
5759 }
5760 const char* stubName = "vectorizedMismatch";
5761 int size_l = callee()->signature()->size();
5762 assert(callee()->signature()->size() == 8, "vectorizedMismatch has 6 parameters");
5763
5764 Node* obja = argument(0);
5765 Node* aoffset = argument(1);
5766 Node* objb = argument(3);
5767 Node* boffset = argument(4);
5768 Node* length = argument(6);
5769 Node* scale = argument(7);
5770
5771 const Type* a_type = obja->Value(&_gvn);
5772 const Type* b_type = objb->Value(&_gvn);
5773 const TypeAryPtr* top_a = a_type->isa_aryptr();
5774 const TypeAryPtr* top_b = b_type->isa_aryptr();
5775 if (top_a == NULL || top_a->klass() == NULL ||
5776 top_b == NULL || top_b->klass() == NULL) {
5777 // failed array check
5778 return false;
5779 }
5780
5781 Node* call;
5782 jvms()->set_should_reexecute(true);
5783
5784 Node* obja_adr = make_unsafe_address(obja, aoffset);
5785 Node* objb_adr = make_unsafe_address(objb, boffset);
5786
5787 call = make_runtime_call(RC_LEAF,
5788 OptoRuntime::vectorizedMismatch_Type(),
5789 stubAddr, stubName, TypePtr::BOTTOM,
5790 obja_adr, objb_adr, length, scale);
5791
5792 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5793 set_result(result);
5794 return true;
5795 }
5796
5797 /**
5798 * Calculate CRC32 for byte.
5799 * int java.util.zip.CRC32.update(int crc, int b)
5800 */
5801 bool LibraryCallKit::inline_updateCRC32() {
5802 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5803 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5804 // no receiver since it is static method
5805 Node* crc = argument(0); // type: int
5806 Node* b = argument(1); // type: int
5807
5808 /*
5809 * int c = ~ crc;
5810 * b = timesXtoThe32[(b ^ c) & 0xFF];
5811 * b = b ^ (c >>> 8);
5812 * crc = ~b;
5813 */
5814
5815 Node* M1 = intcon(-1);
5816 crc = _gvn.transform(new XorINode(crc, M1));
5817 Node* result = _gvn.transform(new XorINode(crc, b));
5818 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5819
5820 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5821 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5822 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5823 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5824
5825 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5826 result = _gvn.transform(new XorINode(crc, result));
5827 result = _gvn.transform(new XorINode(result, M1));
5828 set_result(result);
5829 return true;
5830 }
5831
5832 /**
5833 * Calculate CRC32 for byte[] array.
5834 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5835 */
5836 bool LibraryCallKit::inline_updateBytesCRC32() {
5837 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5838 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5839 // no receiver since it is static method
5840 Node* crc = argument(0); // type: int
5841 Node* src = argument(1); // type: oop
5842 Node* offset = argument(2); // type: int
5843 Node* length = argument(3); // type: int
5844
5845 const Type* src_type = src->Value(&_gvn);
5846 const TypeAryPtr* top_src = src_type->isa_aryptr();
5847 if (top_src == NULL || top_src->klass() == NULL) {
5848 // failed array check
5849 return false;
5850 }
5851
5852 // Figure out the size and type of the elements we will be copying.
5853 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5854 if (src_elem != T_BYTE) {
5855 return false;
5856 }
5857
5858 // 'src_start' points to src array + scaled offset
5859 Node* src_start = array_element_address(src, offset, src_elem);
5860
5861 // We assume that range check is done by caller.
5862 // TODO: generate range check (offset+length < src.length) in debug VM.
5863
5864 // Call the stub.
5865 address stubAddr = StubRoutines::updateBytesCRC32();
5866 const char *stubName = "updateBytesCRC32";
5867
5868 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5869 stubAddr, stubName, TypePtr::BOTTOM,
5870 crc, src_start, length);
5871 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5872 set_result(result);
5873 return true;
5874 }
5875
5876 /**
5877 * Calculate CRC32 for ByteBuffer.
5878 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5879 */
5880 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5881 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5882 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5883 // no receiver since it is static method
5884 Node* crc = argument(0); // type: int
5885 Node* src = argument(1); // type: long
5886 Node* offset = argument(3); // type: int
5887 Node* length = argument(4); // type: int
5888
5889 src = ConvL2X(src); // adjust Java long to machine word
5890 Node* base = _gvn.transform(new CastX2PNode(src));
5891 offset = ConvI2X(offset);
5892
5893 // 'src_start' points to src array + scaled offset
5894 Node* src_start = basic_plus_adr(top(), base, offset);
5895
5896 // Call the stub.
5897 address stubAddr = StubRoutines::updateBytesCRC32();
5898 const char *stubName = "updateBytesCRC32";
5899
5900 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5901 stubAddr, stubName, TypePtr::BOTTOM,
5902 crc, src_start, length);
5903 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5904 set_result(result);
5905 return true;
5906 }
5907
5908 //------------------------------get_table_from_crc32c_class-----------------------
5909 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5910 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5911 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5912
5913 return table;
5914 }
5915
5916 //------------------------------inline_updateBytesCRC32C-----------------------
5917 //
5918 // Calculate CRC32C for byte[] array.
5919 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5920 //
5921 bool LibraryCallKit::inline_updateBytesCRC32C() {
5922 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5923 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5924 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5925 // no receiver since it is a static method
5926 Node* crc = argument(0); // type: int
5927 Node* src = argument(1); // type: oop
5928 Node* offset = argument(2); // type: int
5929 Node* end = argument(3); // type: int
5930
5931 Node* length = _gvn.transform(new SubINode(end, offset));
5932
5933 const Type* src_type = src->Value(&_gvn);
5934 const TypeAryPtr* top_src = src_type->isa_aryptr();
5935 if (top_src == NULL || top_src->klass() == NULL) {
5936 // failed array check
5937 return false;
5938 }
5939
5940 // Figure out the size and type of the elements we will be copying.
5941 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5942 if (src_elem != T_BYTE) {
5943 return false;
5944 }
5945
5946 // 'src_start' points to src array + scaled offset
5947 Node* src_start = array_element_address(src, offset, src_elem);
5948
5949 // static final int[] byteTable in class CRC32C
5950 Node* table = get_table_from_crc32c_class(callee()->holder());
5951 Node* table_start = array_element_address(table, intcon(0), T_INT);
5952
5953 // We assume that range check is done by caller.
5954 // TODO: generate range check (offset+length < src.length) in debug VM.
5955
5956 // Call the stub.
5957 address stubAddr = StubRoutines::updateBytesCRC32C();
5958 const char *stubName = "updateBytesCRC32C";
5959
5960 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5961 stubAddr, stubName, TypePtr::BOTTOM,
5962 crc, src_start, length, table_start);
5963 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5964 set_result(result);
5965 return true;
5966 }
5967
5968 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5969 //
5970 // Calculate CRC32C for DirectByteBuffer.
5971 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5972 //
5973 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5974 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5975 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5976 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5977 // no receiver since it is a static method
5978 Node* crc = argument(0); // type: int
5979 Node* src = argument(1); // type: long
5980 Node* offset = argument(3); // type: int
5981 Node* end = argument(4); // type: int
5982
5983 Node* length = _gvn.transform(new SubINode(end, offset));
5984
5985 src = ConvL2X(src); // adjust Java long to machine word
5986 Node* base = _gvn.transform(new CastX2PNode(src));
5987 offset = ConvI2X(offset);
5988
5989 // 'src_start' points to src array + scaled offset
5990 Node* src_start = basic_plus_adr(top(), base, offset);
5991
5992 // static final int[] byteTable in class CRC32C
5993 Node* table = get_table_from_crc32c_class(callee()->holder());
5994 Node* table_start = array_element_address(table, intcon(0), T_INT);
5995
5996 // Call the stub.
5997 address stubAddr = StubRoutines::updateBytesCRC32C();
5998 const char *stubName = "updateBytesCRC32C";
5999
6000 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
6001 stubAddr, stubName, TypePtr::BOTTOM,
6002 crc, src_start, length, table_start);
6003 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6004 set_result(result);
6005 return true;
6006 }
6007
6008 //------------------------------inline_updateBytesAdler32----------------------
6009 //
6010 // Calculate Adler32 checksum for byte[] array.
6011 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
6012 //
6013 bool LibraryCallKit::inline_updateBytesAdler32() {
6014 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6015 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
6016 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6017 // no receiver since it is static method
6018 Node* crc = argument(0); // type: int
6019 Node* src = argument(1); // type: oop
6020 Node* offset = argument(2); // type: int
6021 Node* length = argument(3); // type: int
6022
6023 const Type* src_type = src->Value(&_gvn);
6024 const TypeAryPtr* top_src = src_type->isa_aryptr();
6025 if (top_src == NULL || top_src->klass() == NULL) {
6026 // failed array check
6027 return false;
6028 }
6029
6030 // Figure out the size and type of the elements we will be copying.
6031 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6032 if (src_elem != T_BYTE) {
6033 return false;
6034 }
6035
6036 // 'src_start' points to src array + scaled offset
6037 Node* src_start = array_element_address(src, offset, src_elem);
6038
6039 // We assume that range check is done by caller.
6040 // TODO: generate range check (offset+length < src.length) in debug VM.
6041
6042 // Call the stub.
6043 address stubAddr = StubRoutines::updateBytesAdler32();
6044 const char *stubName = "updateBytesAdler32";
6045
6046 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6047 stubAddr, stubName, TypePtr::BOTTOM,
6048 crc, src_start, length);
6049 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6050 set_result(result);
6051 return true;
6052 }
6053
6054 //------------------------------inline_updateByteBufferAdler32---------------
6055 //
6056 // Calculate Adler32 checksum for DirectByteBuffer.
6057 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
6058 //
6059 bool LibraryCallKit::inline_updateByteBufferAdler32() {
6060 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
6061 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
6062 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
6063 // no receiver since it is static method
6064 Node* crc = argument(0); // type: int
6065 Node* src = argument(1); // type: long
6066 Node* offset = argument(3); // type: int
6067 Node* length = argument(4); // type: int
6068
6069 src = ConvL2X(src); // adjust Java long to machine word
6070 Node* base = _gvn.transform(new CastX2PNode(src));
6071 offset = ConvI2X(offset);
6072
6073 // 'src_start' points to src array + scaled offset
6074 Node* src_start = basic_plus_adr(top(), base, offset);
6075
6076 // Call the stub.
6077 address stubAddr = StubRoutines::updateBytesAdler32();
6078 const char *stubName = "updateBytesAdler32";
6079
6080 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
6081 stubAddr, stubName, TypePtr::BOTTOM,
6082 crc, src_start, length);
6083
6084 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6085 set_result(result);
6086 return true;
6087 }
6088
6089 //----------------------------inline_reference_get----------------------------
6090 // public T java.lang.ref.Reference.get();
6091 bool LibraryCallKit::inline_reference_get() {
6092 const int referent_offset = java_lang_ref_Reference::referent_offset;
6093 guarantee(referent_offset > 0, "should have already been set");
6094
6095 // Get the argument:
6096 Node* reference_obj = null_check_receiver();
6097 if (stopped()) return true;
6098
6099 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
6100
6101 ciInstanceKlass* klass = env()->Object_klass();
6102 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
6103
6104 Node* no_ctrl = NULL;
6105 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
6106
6107 // Use the pre-barrier to record the value in the referent field
6108 pre_barrier(false /* do_load */,
6109 control(),
6110 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
6111 result /* pre_val */,
6112 T_OBJECT);
6113
6114 // Add memory barrier to prevent commoning reads from this field
6115 // across safepoint since GC can change its value.
6116 insert_mem_bar(Op_MemBarCPUOrder);
6117
6118 set_result(result);
6119 return true;
6120 }
6121
6122
6123 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6124 bool is_exact=true, bool is_static=false,
6125 ciInstanceKlass * fromKls=NULL) {
6126 if (fromKls == NULL) {
6127 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6128 assert(tinst != NULL, "obj is null");
6129 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6130 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6131 fromKls = tinst->klass()->as_instance_klass();
6132 } else {
6133 assert(is_static, "only for static field access");
6134 }
6135 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6136 ciSymbol::make(fieldTypeString),
6137 is_static);
6138
6139 assert (field != NULL, "undefined field");
6140 if (field == NULL) return (Node *) NULL;
6141
6142 if (is_static) {
6143 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6144 fromObj = makecon(tip);
6145 }
6146
6147 // Next code copied from Parse::do_get_xxx():
6148
6149 // Compute address and memory type.
6150 int offset = field->offset_in_bytes();
6151 bool is_vol = field->is_volatile();
6152 ciType* field_klass = field->type();
6153 assert(field_klass->is_loaded(), "should be loaded");
6154 const TypePtr* adr_type = C->alias_type(field)->adr_type();
6155 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6156 BasicType bt = field->layout_type();
6157
6158 // Build the resultant type of the load
6159 const Type *type;
6160 if (bt == T_OBJECT) {
6161 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
6162 } else {
6163 type = Type::get_const_basic_type(bt);
6164 }
6165
6166 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
6167 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
6168 }
6169 // Build the load.
6170 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
6171 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
6172 // If reference is volatile, prevent following memory ops from
6173 // floating up past the volatile read. Also prevents commoning
6174 // another volatile read.
6175 if (is_vol) {
6176 // Memory barrier includes bogus read of value to force load BEFORE membar
6177 insert_mem_bar(Op_MemBarAcquire, loadedField);
6178 }
6179 return loadedField;
6180 }
6181
6182 Node * LibraryCallKit::field_address_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
6183 bool is_exact = true, bool is_static = false,
6184 ciInstanceKlass * fromKls = NULL) {
6185 if (fromKls == NULL) {
6186 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
6187 assert(tinst != NULL, "obj is null");
6188 assert(tinst->klass()->is_loaded(), "obj is not loaded");
6189 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
6190 fromKls = tinst->klass()->as_instance_klass();
6191 }
6192 else {
6193 assert(is_static, "only for static field access");
6194 }
6195 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
6196 ciSymbol::make(fieldTypeString),
6197 is_static);
6198
6199 assert(field != NULL, "undefined field");
6200 assert(!field->is_volatile(), "not defined for volatile fields");
6201
6202 if (is_static) {
6203 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
6204 fromObj = makecon(tip);
6205 }
6206
6207 // Next code copied from Parse::do_get_xxx():
6208
6209 // Compute address and memory type.
6210 int offset = field->offset_in_bytes();
6211 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
6212
6213 return adr;
6214 }
6215
6216 //------------------------------inline_aescrypt_Block-----------------------
6217 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
6218 address stubAddr = NULL;
6219 const char *stubName;
6220 assert(UseAES, "need AES instruction support");
6221
6222 switch(id) {
6223 case vmIntrinsics::_aescrypt_encryptBlock:
6224 stubAddr = StubRoutines::aescrypt_encryptBlock();
6225 stubName = "aescrypt_encryptBlock";
6226 break;
6227 case vmIntrinsics::_aescrypt_decryptBlock:
6228 stubAddr = StubRoutines::aescrypt_decryptBlock();
6229 stubName = "aescrypt_decryptBlock";
6230 break;
6231 default:
6232 break;
6233 }
6234 if (stubAddr == NULL) return false;
6235
6236 Node* aescrypt_object = argument(0);
6237 Node* src = argument(1);
6238 Node* src_offset = argument(2);
6239 Node* dest = argument(3);
6240 Node* dest_offset = argument(4);
6241
6242 // (1) src and dest are arrays.
6243 const Type* src_type = src->Value(&_gvn);
6244 const Type* dest_type = dest->Value(&_gvn);
6245 const TypeAryPtr* top_src = src_type->isa_aryptr();
6246 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6247 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6248
6249 // for the quick and dirty code we will skip all the checks.
6250 // we are just trying to get the call to be generated.
6251 Node* src_start = src;
6252 Node* dest_start = dest;
6253 if (src_offset != NULL || dest_offset != NULL) {
6254 assert(src_offset != NULL && dest_offset != NULL, "");
6255 src_start = array_element_address(src, src_offset, T_BYTE);
6256 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6257 }
6258
6259 // now need to get the start of its expanded key array
6260 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6261 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6262 if (k_start == NULL) return false;
6263
6264 if (Matcher::pass_original_key_for_aes()) {
6265 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6266 // compatibility issues between Java key expansion and SPARC crypto instructions
6267 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6268 if (original_k_start == NULL) return false;
6269
6270 // Call the stub.
6271 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6272 stubAddr, stubName, TypePtr::BOTTOM,
6273 src_start, dest_start, k_start, original_k_start);
6274 } else {
6275 // Call the stub.
6276 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
6277 stubAddr, stubName, TypePtr::BOTTOM,
6278 src_start, dest_start, k_start);
6279 }
6280
6281 return true;
6282 }
6283
6284 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
6285 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
6286 address stubAddr = NULL;
6287 const char *stubName = NULL;
6288
6289 assert(UseAES, "need AES instruction support");
6290
6291 switch(id) {
6292 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
6293 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
6294 stubName = "cipherBlockChaining_encryptAESCrypt";
6295 break;
6296 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
6297 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
6298 stubName = "cipherBlockChaining_decryptAESCrypt";
6299 break;
6300 default:
6301 break;
6302 }
6303 if (stubAddr == NULL) return false;
6304
6305 Node* cipherBlockChaining_object = argument(0);
6306 Node* src = argument(1);
6307 Node* src_offset = argument(2);
6308 Node* len = argument(3);
6309 Node* dest = argument(4);
6310 Node* dest_offset = argument(5);
6311
6312 // (1) src and dest are arrays.
6313 const Type* src_type = src->Value(&_gvn);
6314 const Type* dest_type = dest->Value(&_gvn);
6315 const TypeAryPtr* top_src = src_type->isa_aryptr();
6316 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6317 assert (top_src != NULL && top_src->klass() != NULL
6318 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6319
6320 // checks are the responsibility of the caller
6321 Node* src_start = src;
6322 Node* dest_start = dest;
6323 if (src_offset != NULL || dest_offset != NULL) {
6324 assert(src_offset != NULL && dest_offset != NULL, "");
6325 src_start = array_element_address(src, src_offset, T_BYTE);
6326 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6327 }
6328
6329 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6330 // (because of the predicated logic executed earlier).
6331 // so we cast it here safely.
6332 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6333
6334 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6335 if (embeddedCipherObj == NULL) return false;
6336
6337 // cast it to what we know it will be at runtime
6338 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6339 assert(tinst != NULL, "CBC obj is null");
6340 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6341 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6342 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6343
6344 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6345 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6346 const TypeOopPtr* xtype = aklass->as_instance_type();
6347 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6348 aescrypt_object = _gvn.transform(aescrypt_object);
6349
6350 // we need to get the start of the aescrypt_object's expanded key array
6351 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6352 if (k_start == NULL) return false;
6353
6354 // similarly, get the start address of the r vector
6355 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6356 if (objRvec == NULL) return false;
6357 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6358
6359 Node* cbcCrypt;
6360 if (Matcher::pass_original_key_for_aes()) {
6361 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6362 // compatibility issues between Java key expansion and SPARC crypto instructions
6363 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6364 if (original_k_start == NULL) return false;
6365
6366 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6367 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6368 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6369 stubAddr, stubName, TypePtr::BOTTOM,
6370 src_start, dest_start, k_start, r_start, len, original_k_start);
6371 } else {
6372 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6373 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6374 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6375 stubAddr, stubName, TypePtr::BOTTOM,
6376 src_start, dest_start, k_start, r_start, len);
6377 }
6378
6379 // return cipher length (int)
6380 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6381 set_result(retvalue);
6382 return true;
6383 }
6384
6385 //------------------------------inline_counterMode_AESCrypt-----------------------
6386 bool LibraryCallKit::inline_counterMode_AESCrypt(vmIntrinsics::ID id) {
6387 assert(UseAES, "need AES instruction support");
6388 if (!UseAESCTRIntrinsics) return false;
6389
6390 address stubAddr = NULL;
6391 const char *stubName = NULL;
6392 if (id == vmIntrinsics::_counterMode_AESCrypt) {
6393 stubAddr = StubRoutines::counterMode_AESCrypt();
6394 stubName = "counterMode_AESCrypt";
6395 }
6396 if (stubAddr == NULL) return false;
6397
6398 Node* counterMode_object = argument(0);
6399 Node* src = argument(1);
6400 Node* src_offset = argument(2);
6401 Node* len = argument(3);
6402 Node* dest = argument(4);
6403 Node* dest_offset = argument(5);
6404
6405 // (1) src and dest are arrays.
6406 const Type* src_type = src->Value(&_gvn);
6407 const Type* dest_type = dest->Value(&_gvn);
6408 const TypeAryPtr* top_src = src_type->isa_aryptr();
6409 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6410 assert(top_src != NULL && top_src->klass() != NULL &&
6411 top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6412
6413 // checks are the responsibility of the caller
6414 Node* src_start = src;
6415 Node* dest_start = dest;
6416 if (src_offset != NULL || dest_offset != NULL) {
6417 assert(src_offset != NULL && dest_offset != NULL, "");
6418 src_start = array_element_address(src, src_offset, T_BYTE);
6419 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6420 }
6421
6422 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6423 // (because of the predicated logic executed earlier).
6424 // so we cast it here safely.
6425 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6426 Node* embeddedCipherObj = load_field_from_object(counterMode_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6427 if (embeddedCipherObj == NULL) return false;
6428 // cast it to what we know it will be at runtime
6429 const TypeInstPtr* tinst = _gvn.type(counterMode_object)->isa_instptr();
6430 assert(tinst != NULL, "CTR obj is null");
6431 assert(tinst->klass()->is_loaded(), "CTR obj is not loaded");
6432 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6433 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6434 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6435 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6436 const TypeOopPtr* xtype = aklass->as_instance_type();
6437 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6438 aescrypt_object = _gvn.transform(aescrypt_object);
6439 // we need to get the start of the aescrypt_object's expanded key array
6440 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6441 if (k_start == NULL) return false;
6442 // similarly, get the start address of the r vector
6443 Node* obj_counter = load_field_from_object(counterMode_object, "counter", "[B", /*is_exact*/ false);
6444 if (obj_counter == NULL) return false;
6445 Node* cnt_start = array_element_address(obj_counter, intcon(0), T_BYTE);
6446
6447 Node* saved_encCounter = load_field_from_object(counterMode_object, "encryptedCounter", "[B", /*is_exact*/ false);
6448 if (saved_encCounter == NULL) return false;
6449 Node* saved_encCounter_start = array_element_address(saved_encCounter, intcon(0), T_BYTE);
6450 Node* used = field_address_from_object(counterMode_object, "used", "I", /*is_exact*/ false);
6451
6452 Node* ctrCrypt;
6453 if (Matcher::pass_original_key_for_aes()) {
6454 // no SPARC version for AES/CTR intrinsics now.
6455 return false;
6456 }
6457 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6458 ctrCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6459 OptoRuntime::counterMode_aescrypt_Type(),
6460 stubAddr, stubName, TypePtr::BOTTOM,
6461 src_start, dest_start, k_start, cnt_start, len, saved_encCounter_start, used);
6462
6463 // return cipher length (int)
6464 Node* retvalue = _gvn.transform(new ProjNode(ctrCrypt, TypeFunc::Parms));
6465 set_result(retvalue);
6466 return true;
6467 }
6468
6469 //------------------------------get_key_start_from_aescrypt_object-----------------------
6470 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6471 #if defined(PPC64) || defined(S390)
6472 // MixColumns for decryption can be reduced by preprocessing MixColumns with round keys.
6473 // Intel's extention is based on this optimization and AESCrypt generates round keys by preprocessing MixColumns.
6474 // However, ppc64 vncipher processes MixColumns and requires the same round keys with encryption.
6475 // The ppc64 stubs of encryption and decryption use the same round keys (sessionK[0]).
6476 Node* objSessionK = load_field_from_object(aescrypt_object, "sessionK", "[[I", /*is_exact*/ false);
6477 assert (objSessionK != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6478 if (objSessionK == NULL) {
6479 return (Node *) NULL;
6480 }
6481 Node* objAESCryptKey = load_array_element(control(), objSessionK, intcon(0), TypeAryPtr::OOPS);
6482 #else
6483 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6484 #endif // PPC64
6485 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6486 if (objAESCryptKey == NULL) return (Node *) NULL;
6487
6488 // now have the array, need to get the start address of the K array
6489 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6490 return k_start;
6491 }
6492
6493 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6494 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6495 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6496 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6497 if (objAESCryptKey == NULL) return (Node *) NULL;
6498
6499 // now have the array, need to get the start address of the lastKey array
6500 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6501 return original_k_start;
6502 }
6503
6504 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6505 // Return node representing slow path of predicate check.
6506 // the pseudo code we want to emulate with this predicate is:
6507 // for encryption:
6508 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6509 // for decryption:
6510 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6511 // note cipher==plain is more conservative than the original java code but that's OK
6512 //
6513 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6514 // The receiver was checked for NULL already.
6515 Node* objCBC = argument(0);
6516
6517 // Load embeddedCipher field of CipherBlockChaining object.
6518 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6519
6520 // get AESCrypt klass for instanceOf check
6521 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6522 // will have same classloader as CipherBlockChaining object
6523 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6524 assert(tinst != NULL, "CBCobj is null");
6525 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6526
6527 // we want to do an instanceof comparison against the AESCrypt class
6528 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6529 if (!klass_AESCrypt->is_loaded()) {
6530 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6531 Node* ctrl = control();
6532 set_control(top()); // no regular fast path
6533 return ctrl;
6534 }
6535 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6536
6537 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6538 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6539 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6540
6541 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6542
6543 // for encryption, we are done
6544 if (!decrypting)
6545 return instof_false; // even if it is NULL
6546
6547 // for decryption, we need to add a further check to avoid
6548 // taking the intrinsic path when cipher and plain are the same
6549 // see the original java code for why.
6550 RegionNode* region = new RegionNode(3);
6551 region->init_req(1, instof_false);
6552 Node* src = argument(1);
6553 Node* dest = argument(4);
6554 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6555 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6556 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6557 region->init_req(2, src_dest_conjoint);
6558
6559 record_for_igvn(region);
6560 return _gvn.transform(region);
6561 }
6562
6563 //----------------------------inline_counterMode_AESCrypt_predicate----------------------------
6564 // Return node representing slow path of predicate check.
6565 // the pseudo code we want to emulate with this predicate is:
6566 // for encryption:
6567 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6568 // for decryption:
6569 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6570 // note cipher==plain is more conservative than the original java code but that's OK
6571 //
6572
6573 Node* LibraryCallKit::inline_counterMode_AESCrypt_predicate() {
6574 // The receiver was checked for NULL already.
6575 Node* objCTR = argument(0);
6576
6577 // Load embeddedCipher field of CipherBlockChaining object.
6578 Node* embeddedCipherObj = load_field_from_object(objCTR, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6579
6580 // get AESCrypt klass for instanceOf check
6581 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6582 // will have same classloader as CipherBlockChaining object
6583 const TypeInstPtr* tinst = _gvn.type(objCTR)->isa_instptr();
6584 assert(tinst != NULL, "CTRobj is null");
6585 assert(tinst->klass()->is_loaded(), "CTRobj is not loaded");
6586
6587 // we want to do an instanceof comparison against the AESCrypt class
6588 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6589 if (!klass_AESCrypt->is_loaded()) {
6590 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6591 Node* ctrl = control();
6592 set_control(top()); // no regular fast path
6593 return ctrl;
6594 }
6595
6596 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6597 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6598 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6599 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6600 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6601
6602 return instof_false; // even if it is NULL
6603 }
6604
6605 //------------------------------inline_ghash_processBlocks
6606 bool LibraryCallKit::inline_ghash_processBlocks() {
6607 address stubAddr;
6608 const char *stubName;
6609 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6610
6611 stubAddr = StubRoutines::ghash_processBlocks();
6612 stubName = "ghash_processBlocks";
6613
6614 Node* data = argument(0);
6615 Node* offset = argument(1);
6616 Node* len = argument(2);
6617 Node* state = argument(3);
6618 Node* subkeyH = argument(4);
6619
6620 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6621 assert(state_start, "state is NULL");
6622 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6623 assert(subkeyH_start, "subkeyH is NULL");
6624 Node* data_start = array_element_address(data, offset, T_BYTE);
6625 assert(data_start, "data is NULL");
6626
6627 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6628 OptoRuntime::ghash_processBlocks_Type(),
6629 stubAddr, stubName, TypePtr::BOTTOM,
6630 state_start, subkeyH_start, data_start, len);
6631 return true;
6632 }
6633
6634 //------------------------------inline_sha_implCompress-----------------------
6635 //
6636 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6637 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6638 //
6639 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6640 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6641 //
6642 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6643 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6644 //
6645 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6646 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6647
6648 Node* sha_obj = argument(0);
6649 Node* src = argument(1); // type oop
6650 Node* ofs = argument(2); // type int
6651
6652 const Type* src_type = src->Value(&_gvn);
6653 const TypeAryPtr* top_src = src_type->isa_aryptr();
6654 if (top_src == NULL || top_src->klass() == NULL) {
6655 // failed array check
6656 return false;
6657 }
6658 // Figure out the size and type of the elements we will be copying.
6659 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6660 if (src_elem != T_BYTE) {
6661 return false;
6662 }
6663 // 'src_start' points to src array + offset
6664 Node* src_start = array_element_address(src, ofs, src_elem);
6665 Node* state = NULL;
6666 address stubAddr;
6667 const char *stubName;
6668
6669 switch(id) {
6670 case vmIntrinsics::_sha_implCompress:
6671 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6672 state = get_state_from_sha_object(sha_obj);
6673 stubAddr = StubRoutines::sha1_implCompress();
6674 stubName = "sha1_implCompress";
6675 break;
6676 case vmIntrinsics::_sha2_implCompress:
6677 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6678 state = get_state_from_sha_object(sha_obj);
6679 stubAddr = StubRoutines::sha256_implCompress();
6680 stubName = "sha256_implCompress";
6681 break;
6682 case vmIntrinsics::_sha5_implCompress:
6683 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6684 state = get_state_from_sha5_object(sha_obj);
6685 stubAddr = StubRoutines::sha512_implCompress();
6686 stubName = "sha512_implCompress";
6687 break;
6688 default:
6689 fatal_unexpected_iid(id);
6690 return false;
6691 }
6692 if (state == NULL) return false;
6693
6694 // Call the stub.
6695 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6696 stubAddr, stubName, TypePtr::BOTTOM,
6697 src_start, state);
6698
6699 return true;
6700 }
6701
6702 //------------------------------inline_digestBase_implCompressMB-----------------------
6703 //
6704 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6705 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6706 //
6707 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6708 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6709 "need SHA1/SHA256/SHA512 instruction support");
6710 assert((uint)predicate < 3, "sanity");
6711 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6712
6713 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6714 Node* src = argument(1); // byte[] array
6715 Node* ofs = argument(2); // type int
6716 Node* limit = argument(3); // type int
6717
6718 const Type* src_type = src->Value(&_gvn);
6719 const TypeAryPtr* top_src = src_type->isa_aryptr();
6720 if (top_src == NULL || top_src->klass() == NULL) {
6721 // failed array check
6722 return false;
6723 }
6724 // Figure out the size and type of the elements we will be copying.
6725 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6726 if (src_elem != T_BYTE) {
6727 return false;
6728 }
6729 // 'src_start' points to src array + offset
6730 Node* src_start = array_element_address(src, ofs, src_elem);
6731
6732 const char* klass_SHA_name = NULL;
6733 const char* stub_name = NULL;
6734 address stub_addr = NULL;
6735 bool long_state = false;
6736
6737 switch (predicate) {
6738 case 0:
6739 if (UseSHA1Intrinsics) {
6740 klass_SHA_name = "sun/security/provider/SHA";
6741 stub_name = "sha1_implCompressMB";
6742 stub_addr = StubRoutines::sha1_implCompressMB();
6743 }
6744 break;
6745 case 1:
6746 if (UseSHA256Intrinsics) {
6747 klass_SHA_name = "sun/security/provider/SHA2";
6748 stub_name = "sha256_implCompressMB";
6749 stub_addr = StubRoutines::sha256_implCompressMB();
6750 }
6751 break;
6752 case 2:
6753 if (UseSHA512Intrinsics) {
6754 klass_SHA_name = "sun/security/provider/SHA5";
6755 stub_name = "sha512_implCompressMB";
6756 stub_addr = StubRoutines::sha512_implCompressMB();
6757 long_state = true;
6758 }
6759 break;
6760 default:
6761 fatal("unknown SHA intrinsic predicate: %d", predicate);
6762 }
6763 if (klass_SHA_name != NULL) {
6764 // get DigestBase klass to lookup for SHA klass
6765 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6766 assert(tinst != NULL, "digestBase_obj is not instance???");
6767 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6768
6769 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6770 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6771 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6772 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6773 }
6774 return false;
6775 }
6776 //------------------------------inline_sha_implCompressMB-----------------------
6777 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6778 bool long_state, address stubAddr, const char *stubName,
6779 Node* src_start, Node* ofs, Node* limit) {
6780 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6781 const TypeOopPtr* xtype = aklass->as_instance_type();
6782 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6783 sha_obj = _gvn.transform(sha_obj);
6784
6785 Node* state;
6786 if (long_state) {
6787 state = get_state_from_sha5_object(sha_obj);
6788 } else {
6789 state = get_state_from_sha_object(sha_obj);
6790 }
6791 if (state == NULL) return false;
6792
6793 // Call the stub.
6794 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6795 OptoRuntime::digestBase_implCompressMB_Type(),
6796 stubAddr, stubName, TypePtr::BOTTOM,
6797 src_start, state, ofs, limit);
6798 // return ofs (int)
6799 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6800 set_result(result);
6801
6802 return true;
6803 }
6804
6805 //------------------------------get_state_from_sha_object-----------------------
6806 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6807 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6808 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6809 if (sha_state == NULL) return (Node *) NULL;
6810
6811 // now have the array, need to get the start address of the state array
6812 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6813 return state;
6814 }
6815
6816 //------------------------------get_state_from_sha5_object-----------------------
6817 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6818 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6819 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6820 if (sha_state == NULL) return (Node *) NULL;
6821
6822 // now have the array, need to get the start address of the state array
6823 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6824 return state;
6825 }
6826
6827 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6828 // Return node representing slow path of predicate check.
6829 // the pseudo code we want to emulate with this predicate is:
6830 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6831 //
6832 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6833 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6834 "need SHA1/SHA256/SHA512 instruction support");
6835 assert((uint)predicate < 3, "sanity");
6836
6837 // The receiver was checked for NULL already.
6838 Node* digestBaseObj = argument(0);
6839
6840 // get DigestBase klass for instanceOf check
6841 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6842 assert(tinst != NULL, "digestBaseObj is null");
6843 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6844
6845 const char* klass_SHA_name = NULL;
6846 switch (predicate) {
6847 case 0:
6848 if (UseSHA1Intrinsics) {
6849 // we want to do an instanceof comparison against the SHA class
6850 klass_SHA_name = "sun/security/provider/SHA";
6851 }
6852 break;
6853 case 1:
6854 if (UseSHA256Intrinsics) {
6855 // we want to do an instanceof comparison against the SHA2 class
6856 klass_SHA_name = "sun/security/provider/SHA2";
6857 }
6858 break;
6859 case 2:
6860 if (UseSHA512Intrinsics) {
6861 // we want to do an instanceof comparison against the SHA5 class
6862 klass_SHA_name = "sun/security/provider/SHA5";
6863 }
6864 break;
6865 default:
6866 fatal("unknown SHA intrinsic predicate: %d", predicate);
6867 }
6868
6869 ciKlass* klass_SHA = NULL;
6870 if (klass_SHA_name != NULL) {
6871 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6872 }
6873 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6874 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6875 Node* ctrl = control();
6876 set_control(top()); // no intrinsic path
6877 return ctrl;
6878 }
6879 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6880
6881 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6882 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6883 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6884 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6885
6886 return instof_false; // even if it is NULL
6887 }
6888
6889 //-------------inline_fma-----------------------------------
6890 bool LibraryCallKit::inline_fma(vmIntrinsics::ID id) {
6891 Node *a = NULL;
6892 Node *b = NULL;
6893 Node *c = NULL;
6894 Node* result = NULL;
6895 switch (id) {
6896 case vmIntrinsics::_fmaD:
6897 assert(callee()->signature()->size() == 6, "fma has 3 parameters of size 2 each.");
6898 // no receiver since it is static method
6899 a = round_double_node(argument(0));
6900 b = round_double_node(argument(2));
6901 c = round_double_node(argument(4));
6902 result = _gvn.transform(new FmaDNode(control(), a, b, c));
6903 break;
6904 case vmIntrinsics::_fmaF:
6905 assert(callee()->signature()->size() == 3, "fma has 3 parameters of size 1 each.");
6906 a = argument(0);
6907 b = argument(1);
6908 c = argument(2);
6909 result = _gvn.transform(new FmaFNode(control(), a, b, c));
6910 break;
6911 default:
6912 fatal_unexpected_iid(id); break;
6913 }
6914 set_result(result);
6915 return true;
6916 }
6917
6918 bool LibraryCallKit::inline_profileBoolean() {
6919 Node* counts = argument(1);
6920 const TypeAryPtr* ary = NULL;
6921 ciArray* aobj = NULL;
6922 if (counts->is_Con()
6923 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6924 && (aobj = ary->const_oop()->as_array()) != NULL
6925 && (aobj->length() == 2)) {
6926 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6927 jint false_cnt = aobj->element_value(0).as_int();
6928 jint true_cnt = aobj->element_value(1).as_int();
6929
6930 if (C->log() != NULL) {
6931 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6932 false_cnt, true_cnt);
6933 }
6934
6935 if (false_cnt + true_cnt == 0) {
6936 // According to profile, never executed.
6937 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6938 Deoptimization::Action_reinterpret);
6939 return true;
6940 }
6941
6942 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6943 // is a number of each value occurrences.
6944 Node* result = argument(0);
6945 if (false_cnt == 0 || true_cnt == 0) {
6946 // According to profile, one value has been never seen.
6947 int expected_val = (false_cnt == 0) ? 1 : 0;
6948
6949 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6950 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6951
6952 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6953 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6954 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6955
6956 { // Slow path: uncommon trap for never seen value and then reexecute
6957 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6958 // the value has been seen at least once.
6959 PreserveJVMState pjvms(this);
6960 PreserveReexecuteState preexecs(this);
6961 jvms()->set_should_reexecute(true);
6962
6963 set_control(slow_path);
6964 set_i_o(i_o());
6965
6966 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6967 Deoptimization::Action_reinterpret);
6968 }
6969 // The guard for never seen value enables sharpening of the result and
6970 // returning a constant. It allows to eliminate branches on the same value
6971 // later on.
6972 set_control(fast_path);
6973 result = intcon(expected_val);
6974 }
6975 // Stop profiling.
6976 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6977 // By replacing method body with profile data (represented as ProfileBooleanNode
6978 // on IR level) we effectively disable profiling.
6979 // It enables full speed execution once optimized code is generated.
6980 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6981 C->record_for_igvn(profile);
6982 set_result(profile);
6983 return true;
6984 } else {
6985 // Continue profiling.
6986 // Profile data isn't available at the moment. So, execute method's bytecode version.
6987 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6988 // is compiled and counters aren't available since corresponding MethodHandle
6989 // isn't a compile-time constant.
6990 return false;
6991 }
6992 }
6993
6994 bool LibraryCallKit::inline_isCompileConstant() {
6995 Node* n = argument(0);
6996 set_result(n->is_Con() ? intcon(1) : intcon(0));
6997 return true;
6998 }