1 /*
2 * Copyright (c) 1999, 2015, 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 "oops/objArrayKlass.hpp"
32 #include "opto/addnode.hpp"
33 #include "opto/arraycopynode.hpp"
34 #include "opto/c2compiler.hpp"
35 #include "opto/callGenerator.hpp"
36 #include "opto/castnode.hpp"
37 #include "opto/cfgnode.hpp"
38 #include "opto/convertnode.hpp"
39 #include "opto/countbitsnode.hpp"
40 #include "opto/intrinsicnode.hpp"
41 #include "opto/idealKit.hpp"
42 #include "opto/mathexactnode.hpp"
43 #include "opto/movenode.hpp"
44 #include "opto/mulnode.hpp"
45 #include "opto/narrowptrnode.hpp"
46 #include "opto/opaquenode.hpp"
47 #include "opto/parse.hpp"
48 #include "opto/runtime.hpp"
49 #include "opto/subnode.hpp"
50 #include "prims/nativeLookup.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "trace/traceMacros.hpp"
53
54 class LibraryIntrinsic : public InlineCallGenerator {
55 // Extend the set of intrinsics known to the runtime:
56 public:
57 private:
58 bool _is_virtual;
59 bool _does_virtual_dispatch;
60 int8_t _predicates_count; // Intrinsic is predicated by several conditions
61 int8_t _last_predicate; // Last generated predicate
62 vmIntrinsics::ID _intrinsic_id;
63
64 public:
65 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
66 : InlineCallGenerator(m),
67 _is_virtual(is_virtual),
68 _does_virtual_dispatch(does_virtual_dispatch),
69 _predicates_count((int8_t)predicates_count),
70 _last_predicate((int8_t)-1),
71 _intrinsic_id(id)
72 {
73 }
74 virtual bool is_intrinsic() const { return true; }
75 virtual bool is_virtual() const { return _is_virtual; }
76 virtual bool is_predicated() const { return _predicates_count > 0; }
77 virtual int predicates_count() const { return _predicates_count; }
78 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
79 virtual JVMState* generate(JVMState* jvms);
80 virtual Node* generate_predicate(JVMState* jvms, int predicate);
81 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
82 };
83
84
85 // Local helper class for LibraryIntrinsic:
86 class LibraryCallKit : public GraphKit {
87 private:
88 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
89 Node* _result; // the result node, if any
90 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
91
92 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
93
94 public:
95 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
96 : GraphKit(jvms),
97 _intrinsic(intrinsic),
98 _result(NULL)
99 {
100 // Check if this is a root compile. In that case we don't have a caller.
101 if (!jvms->has_method()) {
102 _reexecute_sp = sp();
103 } else {
104 // Find out how many arguments the interpreter needs when deoptimizing
105 // and save the stack pointer value so it can used by uncommon_trap.
106 // We find the argument count by looking at the declared signature.
107 bool ignored_will_link;
108 ciSignature* declared_signature = NULL;
109 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
110 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
111 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
112 }
113 }
114
115 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
116
117 ciMethod* caller() const { return jvms()->method(); }
118 int bci() const { return jvms()->bci(); }
119 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
120 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
121 ciMethod* callee() const { return _intrinsic->method(); }
122
123 bool try_to_inline(int predicate);
124 Node* try_to_predicate(int predicate);
125
126 void push_result() {
127 // Push the result onto the stack.
128 if (!stopped() && result() != NULL) {
129 BasicType bt = result()->bottom_type()->basic_type();
130 push_node(bt, result());
131 }
132 }
133
134 private:
135 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
136 fatal("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid));
137 }
138
139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
140 void set_result(RegionNode* region, PhiNode* value);
141 Node* result() { return _result; }
142
143 virtual int reexecute_sp() { return _reexecute_sp; }
144
145 // Helper functions to inline natives
146 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
147 Node* generate_slow_guard(Node* test, RegionNode* region);
148 Node* generate_fair_guard(Node* test, RegionNode* region);
149 Node* generate_negative_guard(Node* index, RegionNode* region,
150 // resulting CastII of index:
151 Node* *pos_index = NULL);
152 Node* generate_limit_guard(Node* offset, Node* subseq_length,
153 Node* array_length,
154 RegionNode* region);
155 Node* generate_current_thread(Node* &tls_output);
156 Node* load_mirror_from_klass(Node* klass);
157 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path,
159 int offset);
160 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
161 RegionNode* region, int null_path) {
162 int offset = java_lang_Class::klass_offset_in_bytes();
163 return load_klass_from_mirror_common(mirror, never_see_null,
164 region, null_path,
165 offset);
166 }
167 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::array_klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* generate_access_flags_guard(Node* kls,
175 int modifier_mask, int modifier_bits,
176 RegionNode* region);
177 Node* generate_interface_guard(Node* kls, RegionNode* region);
178 Node* generate_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, false);
180 }
181 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, false, true);
183 }
184 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, false);
186 }
187 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, true, true);
189 }
190 Node* generate_array_guard_common(Node* kls, RegionNode* region,
191 bool obj_array, bool not_array);
192 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
193 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
194 bool is_virtual = false, bool is_static = false);
195 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, false, true);
197 }
198 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
199 return generate_method_call(method_id, true, false);
200 }
201 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
202
203 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae);
204 bool inline_string_compareTo(StrIntrinsicNode::ArgEnc ae);
205 bool inline_string_indexOf(StrIntrinsicNode::ArgEnc ae);
206 bool inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae);
207 bool inline_string_indexOfChar();
208 bool inline_string_equals(StrIntrinsicNode::ArgEnc ae);
209 bool inline_string_toBytesU();
210 bool inline_string_getCharsU();
211 bool inline_string_copy(bool compress);
212 bool inline_string_char_access(bool is_store);
213 Node* round_double_node(Node* n);
214 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
215 bool inline_math_native(vmIntrinsics::ID id);
216 bool inline_trig(vmIntrinsics::ID id);
217 bool inline_math(vmIntrinsics::ID id);
218 template <typename OverflowOp>
219 bool inline_math_overflow(Node* arg1, Node* arg2);
220 void inline_math_mathExact(Node* math, Node* test);
221 bool inline_math_addExactI(bool is_increment);
222 bool inline_math_addExactL(bool is_increment);
223 bool inline_math_multiplyExactI();
224 bool inline_math_multiplyExactL();
225 bool inline_math_negateExactI();
226 bool inline_math_negateExactL();
227 bool inline_math_subtractExactI(bool is_decrement);
228 bool inline_math_subtractExactL(bool is_decrement);
229 bool inline_pow();
230 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
231 bool inline_min_max(vmIntrinsics::ID id);
232 bool inline_notify(vmIntrinsics::ID id);
233 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
234 // This returns Type::AnyPtr, RawPtr, or OopPtr.
235 int classify_unsafe_addr(Node* &base, Node* &offset);
236 Node* make_unsafe_address(Node* base, Node* offset);
237 // Helper for inline_unsafe_access.
238 // Generates the guards that check whether the result of
239 // Unsafe.getObject should be recorded in an SATB log buffer.
240 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
241 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool is_unaligned);
242 static bool klass_needs_init_guard(Node* kls);
243 bool inline_unsafe_allocate();
244 bool inline_unsafe_copyMemory();
245 bool inline_native_currentThread();
246 #ifdef TRACE_HAVE_INTRINSICS
247 bool inline_native_classID();
248 bool inline_native_threadID();
249 #endif
250 bool inline_native_time_funcs(address method, const char* funcName);
251 bool inline_native_isInterrupted();
252 bool inline_native_Class_query(vmIntrinsics::ID id);
253 bool inline_native_subtype_check();
254
255 bool inline_native_newArray();
256 bool inline_native_getLength();
257 bool inline_array_copyOf(bool is_copyOfRange);
258 bool inline_array_equals(StrIntrinsicNode::ArgEnc ae);
259 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
260 bool inline_native_clone(bool is_virtual);
261 bool inline_native_Reflection_getCallerClass();
262 // Helper function for inlining native object hash method
263 bool inline_native_hashcode(bool is_virtual, bool is_static);
264 bool inline_native_getClass();
265
266 // Helper functions for inlining arraycopy
267 bool inline_arraycopy();
268 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
269 RegionNode* slow_region);
270 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
271 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
272
273 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
274 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
275 bool inline_unsafe_ordered_store(BasicType type);
276 bool inline_unsafe_fence(vmIntrinsics::ID id);
277 bool inline_fp_conversions(vmIntrinsics::ID id);
278 bool inline_number_methods(vmIntrinsics::ID id);
279 bool inline_reference_get();
280 bool inline_Class_cast();
281 bool inline_aescrypt_Block(vmIntrinsics::ID id);
282 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
283 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
284 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
285 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
286 bool inline_ghash_processBlocks();
287 bool inline_sha_implCompress(vmIntrinsics::ID id);
288 bool inline_digestBase_implCompressMB(int predicate);
289 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
290 bool long_state, address stubAddr, const char *stubName,
291 Node* src_start, Node* ofs, Node* limit);
292 Node* get_state_from_sha_object(Node *sha_object);
293 Node* get_state_from_sha5_object(Node *sha_object);
294 Node* inline_digestBase_implCompressMB_predicate(int predicate);
295 bool inline_encodeISOArray();
296 bool inline_updateCRC32();
297 bool inline_updateBytesCRC32();
298 bool inline_updateByteBufferCRC32();
299 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
300 bool inline_updateBytesCRC32C();
301 bool inline_updateDirectByteBufferCRC32C();
302 bool inline_updateBytesAdler32();
303 bool inline_updateByteBufferAdler32();
304 bool inline_multiplyToLen();
305 bool inline_hasNegatives();
306 bool inline_squareToLen();
307 bool inline_mulAdd();
308 bool inline_montgomeryMultiply();
309 bool inline_montgomerySquare();
310
311 bool inline_profileBoolean();
312 bool inline_isCompileConstant();
313 };
314
315 //---------------------------make_vm_intrinsic----------------------------
316 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
317 vmIntrinsics::ID id = m->intrinsic_id();
318 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
319
320 if (!m->is_loaded()) {
321 // Do not attempt to inline unloaded methods.
322 return NULL;
323 }
324
325 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
326 bool is_available = false;
327
328 {
329 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
330 // the compiler must transition to '_thread_in_vm' state because both
331 // methods access VM-internal data.
332 VM_ENTRY_MARK;
333 methodHandle mh(THREAD, m->get_Method());
334 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
335 !C->directive()->is_intrinsic_disabled(mh) &&
336 !vmIntrinsics::is_disabled_by_flags(mh);
337
338 }
339
340 if (is_available) {
341 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
342 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
343 return new LibraryIntrinsic(m, is_virtual,
344 vmIntrinsics::predicates_needed(id),
345 vmIntrinsics::does_virtual_dispatch(id),
346 (vmIntrinsics::ID) id);
347 } else {
348 return NULL;
349 }
350 }
351
352 //----------------------register_library_intrinsics-----------------------
353 // Initialize this file's data structures, for each Compile instance.
354 void Compile::register_library_intrinsics() {
355 // Nothing to do here.
356 }
357
358 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
359 LibraryCallKit kit(jvms, this);
360 Compile* C = kit.C;
361 int nodes = C->unique();
362 #ifndef PRODUCT
363 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
364 char buf[1000];
365 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
366 tty->print_cr("Intrinsic %s", str);
367 }
368 #endif
369 ciMethod* callee = kit.callee();
370 const int bci = kit.bci();
371
372 // Try to inline the intrinsic.
373 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
374 kit.try_to_inline(_last_predicate)) {
375 if (C->print_intrinsics() || C->print_inlining()) {
376 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
377 }
378 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
379 if (C->log()) {
380 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
381 vmIntrinsics::name_at(intrinsic_id()),
382 (is_virtual() ? " virtual='1'" : ""),
383 C->unique() - nodes);
384 }
385 // Push the result from the inlined method onto the stack.
386 kit.push_result();
387 C->print_inlining_update(this);
388 return kit.transfer_exceptions_into_jvms();
389 }
390
391 // The intrinsic bailed out
392 if (C->print_intrinsics() || C->print_inlining()) {
393 if (jvms->has_method()) {
394 // Not a root compile.
395 const char* msg;
396 if (callee->intrinsic_candidate()) {
397 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
398 } else {
399 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
400 : "failed to inline (intrinsic), method not annotated";
401 }
402 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
403 } else {
404 // Root compile
405 tty->print("Did not generate intrinsic %s%s at bci:%d in",
406 vmIntrinsics::name_at(intrinsic_id()),
407 (is_virtual() ? " (virtual)" : ""), bci);
408 }
409 }
410 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
411 C->print_inlining_update(this);
412 return NULL;
413 }
414
415 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
416 LibraryCallKit kit(jvms, this);
417 Compile* C = kit.C;
418 int nodes = C->unique();
419 _last_predicate = predicate;
420 #ifndef PRODUCT
421 assert(is_predicated() && predicate < predicates_count(), "sanity");
422 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
423 char buf[1000];
424 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
425 tty->print_cr("Predicate for intrinsic %s", str);
426 }
427 #endif
428 ciMethod* callee = kit.callee();
429 const int bci = kit.bci();
430
431 Node* slow_ctl = kit.try_to_predicate(predicate);
432 if (!kit.failing()) {
433 if (C->print_intrinsics() || C->print_inlining()) {
434 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
435 }
436 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
437 if (C->log()) {
438 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
439 vmIntrinsics::name_at(intrinsic_id()),
440 (is_virtual() ? " virtual='1'" : ""),
441 C->unique() - nodes);
442 }
443 return slow_ctl; // Could be NULL if the check folds.
444 }
445
446 // The intrinsic bailed out
447 if (C->print_intrinsics() || C->print_inlining()) {
448 if (jvms->has_method()) {
449 // Not a root compile.
450 const char* msg = "failed to generate predicate for intrinsic";
451 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
452 } else {
453 // Root compile
454 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
455 vmIntrinsics::name_at(intrinsic_id()),
456 (is_virtual() ? " (virtual)" : ""), bci);
457 }
458 }
459 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
460 return NULL;
461 }
462
463 bool LibraryCallKit::try_to_inline(int predicate) {
464 // Handle symbolic names for otherwise undistinguished boolean switches:
465 const bool is_store = true;
466 const bool is_compress = true;
467 const bool is_native_ptr = true;
468 const bool is_static = true;
469 const bool is_volatile = true;
470
471 if (!jvms()->has_method()) {
472 // Root JVMState has a null method.
473 assert(map()->memory()->Opcode() == Op_Parm, "");
474 // Insert the memory aliasing node
475 set_all_memory(reset_memory());
476 }
477 assert(merged_memory(), "");
478
479
480 switch (intrinsic_id()) {
481 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
482 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
483 case vmIntrinsics::_getClass: return inline_native_getClass();
484
485 case vmIntrinsics::_dsin:
486 case vmIntrinsics::_dcos:
487 case vmIntrinsics::_dtan:
488 case vmIntrinsics::_dabs:
489 case vmIntrinsics::_datan2:
490 case vmIntrinsics::_dsqrt:
491 case vmIntrinsics::_dexp:
492 case vmIntrinsics::_dlog:
493 case vmIntrinsics::_dlog10:
494 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
495
496 case vmIntrinsics::_min:
497 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
498
499 case vmIntrinsics::_notify:
500 case vmIntrinsics::_notifyAll:
501 if (InlineNotify) {
502 return inline_notify(intrinsic_id());
503 }
504 return false;
505
506 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
507 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
508 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
509 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
510 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
511 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
512 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
513 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
514 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
515 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
516 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
517 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
518
519 case vmIntrinsics::_arraycopy: return inline_arraycopy();
520
521 case vmIntrinsics::_compareToL: return inline_string_compareTo(StrIntrinsicNode::LL);
522 case vmIntrinsics::_compareToU: return inline_string_compareTo(StrIntrinsicNode::UU);
523 case vmIntrinsics::_compareToLU: return inline_string_compareTo(StrIntrinsicNode::LU);
524 case vmIntrinsics::_compareToUL: return inline_string_compareTo(StrIntrinsicNode::UL);
525
526 case vmIntrinsics::_indexOfL: return inline_string_indexOf(StrIntrinsicNode::LL);
527 case vmIntrinsics::_indexOfU: return inline_string_indexOf(StrIntrinsicNode::UU);
528 case vmIntrinsics::_indexOfUL: return inline_string_indexOf(StrIntrinsicNode::UL);
529 case vmIntrinsics::_indexOfIL: return inline_string_indexOfI(StrIntrinsicNode::LL);
530 case vmIntrinsics::_indexOfIU: return inline_string_indexOfI(StrIntrinsicNode::UU);
531 case vmIntrinsics::_indexOfIUL: return inline_string_indexOfI(StrIntrinsicNode::UL);
532 case vmIntrinsics::_indexOfU_char: return inline_string_indexOfChar();
533
534 case vmIntrinsics::_equalsL: return inline_string_equals(StrIntrinsicNode::LL);
535 case vmIntrinsics::_equalsU: return inline_string_equals(StrIntrinsicNode::UU);
536
537 case vmIntrinsics::_toBytesStringU: return inline_string_toBytesU();
538 case vmIntrinsics::_getCharsStringU: return inline_string_getCharsU();
539 case vmIntrinsics::_getCharStringU: return inline_string_char_access(!is_store);
540 case vmIntrinsics::_putCharStringU: return inline_string_char_access( is_store);
541
542 case vmIntrinsics::_compressStringC:
543 case vmIntrinsics::_compressStringB: return inline_string_copy( is_compress);
544 case vmIntrinsics::_inflateStringC:
545 case vmIntrinsics::_inflateStringB: return inline_string_copy(!is_compress);
546
547 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile, false);
548 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile, false);
549 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
550 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
551 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
552 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, false);
553 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, false);
554 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
555 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
556 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile, false);
557 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile, false);
558 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile, false);
559 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, false);
560 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, false);
561 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, false);
562 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, false);
563 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
564 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
565
566 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile, false);
567 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile, false);
568 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile, false);
569 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile, false);
570 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile, false);
571 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile, false);
572 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile, false);
573 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile, false);
574
575 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile, false);
576 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile, false);
577 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile, false);
578 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile, false);
579 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile, false);
580 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile, false);
581 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile, false);
582 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile, false);
583
584 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile, false);
585 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile, false);
586 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile, false);
587 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile, false);
588 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile, false);
589 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile, false);
590 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile, false);
591 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile, false);
592 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile, false);
593
594 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile, false);
595 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile, false);
596 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile, false);
597 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile, false);
598 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile, false);
599 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile, false);
600 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile, false);
601 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile, false);
602 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile, false);
603
604 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile, true);
605 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile, true);
606 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile, true);
607 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile, true);
608
609 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile, true);
610 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile, true);
611 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile, true);
612 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile, true);
613
614 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
615 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
616 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
617
618 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
619 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
620 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
621
622 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
623 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
624 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
625 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
626 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
627
628 case vmIntrinsics::_loadFence:
629 case vmIntrinsics::_storeFence:
630 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
631
632 case vmIntrinsics::_currentThread: return inline_native_currentThread();
633 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
634
635 #ifdef TRACE_HAVE_INTRINSICS
636 case vmIntrinsics::_classID: return inline_native_classID();
637 case vmIntrinsics::_threadID: return inline_native_threadID();
638 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
639 #endif
640 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
641 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
642 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
643 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
644 case vmIntrinsics::_newArray: return inline_native_newArray();
645 case vmIntrinsics::_getLength: return inline_native_getLength();
646 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
647 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
648 case vmIntrinsics::_equalsB: return inline_array_equals(StrIntrinsicNode::LL);
649 case vmIntrinsics::_equalsC: return inline_array_equals(StrIntrinsicNode::UU);
650 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
651
652 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
653
654 case vmIntrinsics::_isInstance:
655 case vmIntrinsics::_getModifiers:
656 case vmIntrinsics::_isInterface:
657 case vmIntrinsics::_isArray:
658 case vmIntrinsics::_isPrimitive:
659 case vmIntrinsics::_getSuperclass:
660 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
661
662 case vmIntrinsics::_floatToRawIntBits:
663 case vmIntrinsics::_floatToIntBits:
664 case vmIntrinsics::_intBitsToFloat:
665 case vmIntrinsics::_doubleToRawLongBits:
666 case vmIntrinsics::_doubleToLongBits:
667 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
668
669 case vmIntrinsics::_numberOfLeadingZeros_i:
670 case vmIntrinsics::_numberOfLeadingZeros_l:
671 case vmIntrinsics::_numberOfTrailingZeros_i:
672 case vmIntrinsics::_numberOfTrailingZeros_l:
673 case vmIntrinsics::_bitCount_i:
674 case vmIntrinsics::_bitCount_l:
675 case vmIntrinsics::_reverseBytes_i:
676 case vmIntrinsics::_reverseBytes_l:
677 case vmIntrinsics::_reverseBytes_s:
678 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
679
680 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
681
682 case vmIntrinsics::_Reference_get: return inline_reference_get();
683
684 case vmIntrinsics::_Class_cast: return inline_Class_cast();
685
686 case vmIntrinsics::_aescrypt_encryptBlock:
687 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
688
689 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
690 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
691 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
692
693 case vmIntrinsics::_sha_implCompress:
694 case vmIntrinsics::_sha2_implCompress:
695 case vmIntrinsics::_sha5_implCompress:
696 return inline_sha_implCompress(intrinsic_id());
697
698 case vmIntrinsics::_digestBase_implCompressMB:
699 return inline_digestBase_implCompressMB(predicate);
700
701 case vmIntrinsics::_multiplyToLen:
702 return inline_multiplyToLen();
703
704 case vmIntrinsics::_squareToLen:
705 return inline_squareToLen();
706
707 case vmIntrinsics::_mulAdd:
708 return inline_mulAdd();
709
710 case vmIntrinsics::_montgomeryMultiply:
711 return inline_montgomeryMultiply();
712 case vmIntrinsics::_montgomerySquare:
713 return inline_montgomerySquare();
714
715 case vmIntrinsics::_ghash_processBlocks:
716 return inline_ghash_processBlocks();
717
718 case vmIntrinsics::_encodeISOArray:
719 case vmIntrinsics::_encodeByteISOArray:
720 return inline_encodeISOArray();
721
722 case vmIntrinsics::_updateCRC32:
723 return inline_updateCRC32();
724 case vmIntrinsics::_updateBytesCRC32:
725 return inline_updateBytesCRC32();
726 case vmIntrinsics::_updateByteBufferCRC32:
727 return inline_updateByteBufferCRC32();
728
729 case vmIntrinsics::_updateBytesCRC32C:
730 return inline_updateBytesCRC32C();
731 case vmIntrinsics::_updateDirectByteBufferCRC32C:
732 return inline_updateDirectByteBufferCRC32C();
733
734 case vmIntrinsics::_updateBytesAdler32:
735 return inline_updateBytesAdler32();
736 case vmIntrinsics::_updateByteBufferAdler32:
737 return inline_updateByteBufferAdler32();
738
739 case vmIntrinsics::_profileBoolean:
740 return inline_profileBoolean();
741 case vmIntrinsics::_isCompileConstant:
742 return inline_isCompileConstant();
743
744 case vmIntrinsics::_hasNegatives:
745 return inline_hasNegatives();
746
747 default:
748 // If you get here, it may be that someone has added a new intrinsic
749 // to the list in vmSymbols.hpp without implementing it here.
750 #ifndef PRODUCT
751 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
752 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
753 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
754 }
755 #endif
756 return false;
757 }
758 }
759
760 Node* LibraryCallKit::try_to_predicate(int predicate) {
761 if (!jvms()->has_method()) {
762 // Root JVMState has a null method.
763 assert(map()->memory()->Opcode() == Op_Parm, "");
764 // Insert the memory aliasing node
765 set_all_memory(reset_memory());
766 }
767 assert(merged_memory(), "");
768
769 switch (intrinsic_id()) {
770 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
771 return inline_cipherBlockChaining_AESCrypt_predicate(false);
772 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
773 return inline_cipherBlockChaining_AESCrypt_predicate(true);
774 case vmIntrinsics::_digestBase_implCompressMB:
775 return inline_digestBase_implCompressMB_predicate(predicate);
776
777 default:
778 // If you get here, it may be that someone has added a new intrinsic
779 // to the list in vmSymbols.hpp without implementing it here.
780 #ifndef PRODUCT
781 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
782 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
783 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
784 }
785 #endif
786 Node* slow_ctl = control();
787 set_control(top()); // No fast path instrinsic
788 return slow_ctl;
789 }
790 }
791
792 //------------------------------set_result-------------------------------
793 // Helper function for finishing intrinsics.
794 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
795 record_for_igvn(region);
796 set_control(_gvn.transform(region));
797 set_result( _gvn.transform(value));
798 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
799 }
800
801 //------------------------------generate_guard---------------------------
802 // Helper function for generating guarded fast-slow graph structures.
803 // The given 'test', if true, guards a slow path. If the test fails
804 // then a fast path can be taken. (We generally hope it fails.)
805 // In all cases, GraphKit::control() is updated to the fast path.
806 // The returned value represents the control for the slow path.
807 // The return value is never 'top'; it is either a valid control
808 // or NULL if it is obvious that the slow path can never be taken.
809 // Also, if region and the slow control are not NULL, the slow edge
810 // is appended to the region.
811 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
812 if (stopped()) {
813 // Already short circuited.
814 return NULL;
815 }
816
817 // Build an if node and its projections.
818 // If test is true we take the slow path, which we assume is uncommon.
819 if (_gvn.type(test) == TypeInt::ZERO) {
820 // The slow branch is never taken. No need to build this guard.
821 return NULL;
822 }
823
824 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
825
826 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
827 if (if_slow == top()) {
828 // The slow branch is never taken. No need to build this guard.
829 return NULL;
830 }
831
832 if (region != NULL)
833 region->add_req(if_slow);
834
835 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
836 set_control(if_fast);
837
838 return if_slow;
839 }
840
841 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
842 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
843 }
844 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
845 return generate_guard(test, region, PROB_FAIR);
846 }
847
848 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
849 Node* *pos_index) {
850 if (stopped())
851 return NULL; // already stopped
852 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
853 return NULL; // index is already adequately typed
854 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
855 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
856 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
857 if (is_neg != NULL && pos_index != NULL) {
858 // Emulate effect of Parse::adjust_map_after_if.
859 Node* ccast = new CastIINode(index, TypeInt::POS);
860 ccast->set_req(0, control());
861 (*pos_index) = _gvn.transform(ccast);
862 }
863 return is_neg;
864 }
865
866 // Make sure that 'position' is a valid limit index, in [0..length].
867 // There are two equivalent plans for checking this:
868 // A. (offset + copyLength) unsigned<= arrayLength
869 // B. offset <= (arrayLength - copyLength)
870 // We require that all of the values above, except for the sum and
871 // difference, are already known to be non-negative.
872 // Plan A is robust in the face of overflow, if offset and copyLength
873 // are both hugely positive.
874 //
875 // Plan B is less direct and intuitive, but it does not overflow at
876 // all, since the difference of two non-negatives is always
877 // representable. Whenever Java methods must perform the equivalent
878 // check they generally use Plan B instead of Plan A.
879 // For the moment we use Plan A.
880 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
881 Node* subseq_length,
882 Node* array_length,
883 RegionNode* region) {
884 if (stopped())
885 return NULL; // already stopped
886 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
887 if (zero_offset && subseq_length->eqv_uncast(array_length))
888 return NULL; // common case of whole-array copy
889 Node* last = subseq_length;
890 if (!zero_offset) // last += offset
891 last = _gvn.transform(new AddINode(last, offset));
892 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
893 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
894 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
895 return is_over;
896 }
897
898
899 //--------------------------generate_current_thread--------------------
900 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
901 ciKlass* thread_klass = env()->Thread_klass();
902 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
903 Node* thread = _gvn.transform(new ThreadLocalNode());
904 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
905 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
906 tls_output = thread;
907 return threadObj;
908 }
909
910
911 //------------------------------make_string_method_node------------------------
912 // Helper method for String intrinsic functions. This version is called with
913 // str1 and str2 pointing to byte[] nodes containing Latin1 or UTF16 encoded
914 // characters (depending on 'is_byte'). cnt1 and cnt2 are pointing to Int nodes
915 // containing the lengths of str1 and str2.
916 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2, StrIntrinsicNode::ArgEnc ae) {
917 Node* result = NULL;
918 switch (opcode) {
919 case Op_StrIndexOf:
920 result = new StrIndexOfNode(control(), memory(TypeAryPtr::BYTES),
921 str1_start, cnt1, str2_start, cnt2, ae);
922 break;
923 case Op_StrComp:
924 result = new StrCompNode(control(), memory(TypeAryPtr::BYTES),
925 str1_start, cnt1, str2_start, cnt2, ae);
926 break;
927 case Op_StrEquals:
928 result = new StrEqualsNode(control(), memory(TypeAryPtr::BYTES),
929 str1_start, str2_start, cnt1, ae);
930 break;
931 default:
932 ShouldNotReachHere();
933 return NULL;
934 }
935
936 // All these intrinsics have checks.
937 C->set_has_split_ifs(true); // Has chance for split-if optimization
938
939 return _gvn.transform(result);
940 }
941
942 //------------------------------inline_string_compareTo------------------------
943 bool LibraryCallKit::inline_string_compareTo(StrIntrinsicNode::ArgEnc ae) {
944 Node* arg1 = argument(0);
945 Node* arg2 = argument(1);
946
947 // Get start addr and length of first argument
948 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
949 Node* arg1_cnt = load_array_length(arg1);
950
951 // Get start addr and length of second argument
952 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
953 Node* arg2_cnt = load_array_length(arg2);
954
955 Node* result = make_string_method_node(Op_StrComp, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
956 set_result(result);
957 return true;
958 }
959
960 //------------------------------inline_string_equals------------------------
961 bool LibraryCallKit::inline_string_equals(StrIntrinsicNode::ArgEnc ae) {
962 Node* arg1 = argument(0);
963 Node* arg2 = argument(1);
964
965 // paths (plus control) merge
966 RegionNode* region = new RegionNode(3);
967 Node* phi = new PhiNode(region, TypeInt::BOOL);
968
969 if (!stopped()) {
970 // Get start addr and length of first argument
971 Node* arg1_start = array_element_address(arg1, intcon(0), T_BYTE);
972 Node* arg1_cnt = load_array_length(arg1);
973
974 // Get start addr and length of second argument
975 Node* arg2_start = array_element_address(arg2, intcon(0), T_BYTE);
976 Node* arg2_cnt = load_array_length(arg2);
977
978 // Check for arg1_cnt != arg2_cnt
979 Node* cmp = _gvn.transform(new CmpINode(arg1_cnt, arg2_cnt));
980 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
981 Node* if_ne = generate_slow_guard(bol, NULL);
982 if (if_ne != NULL) {
983 phi->init_req(2, intcon(0));
984 region->init_req(2, if_ne);
985 }
986
987 // Check for count == 0 is done by assembler code for StrEquals.
988
989 if (!stopped()) {
990 Node* equals = make_string_method_node(Op_StrEquals, arg1_start, arg1_cnt, arg2_start, arg2_cnt, ae);
991 phi->init_req(1, equals);
992 region->init_req(1, control());
993 }
994 }
995
996 // post merge
997 set_control(_gvn.transform(region));
998 record_for_igvn(region);
999
1000 set_result(_gvn.transform(phi));
1001 return true;
1002 }
1003
1004 //------------------------------inline_array_equals----------------------------
1005 bool LibraryCallKit::inline_array_equals(StrIntrinsicNode::ArgEnc ae) {
1006 assert(ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::LL, "unsupported array types");
1007 Node* arg1 = argument(0);
1008 Node* arg2 = argument(1);
1009
1010 const TypeAryPtr* mtype = (ae == StrIntrinsicNode::UU) ? TypeAryPtr::CHARS : TypeAryPtr::BYTES;
1011 set_result(_gvn.transform(new AryEqNode(control(), memory(mtype), arg1, arg2, ae)));
1012 return true;
1013 }
1014
1015 //------------------------------inline_hasNegatives------------------------------
1016 bool LibraryCallKit::inline_hasNegatives() {
1017 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1018
1019 assert(callee()->signature()->size() == 3, "hasNegatives has 3 parameters");
1020 // no receiver since it is static method
1021 Node* ba = argument(0);
1022 Node* offset = argument(1);
1023 Node* len = argument(2);
1024
1025 RegionNode* bailout = new RegionNode(1);
1026 record_for_igvn(bailout);
1027
1028 // offset must not be negative.
1029 generate_negative_guard(offset, bailout);
1030
1031 // offset + length must not exceed length of ba.
1032 generate_limit_guard(offset, len, load_array_length(ba), bailout);
1033
1034 if (bailout->req() > 1) {
1035 PreserveJVMState pjvms(this);
1036 set_control(_gvn.transform(bailout));
1037 uncommon_trap(Deoptimization::Reason_intrinsic,
1038 Deoptimization::Action_maybe_recompile);
1039 }
1040 if (!stopped()) {
1041 Node* ba_start = array_element_address(ba, offset, T_BYTE);
1042 Node* result = new HasNegativesNode(control(), memory(TypeAryPtr::BYTES), ba_start, len);
1043 set_result(_gvn.transform(result));
1044 }
1045 return true;
1046 }
1047
1048 //------------------------------inline_string_indexOf------------------------
1049 bool LibraryCallKit::inline_string_indexOf(StrIntrinsicNode::ArgEnc ae) {
1050 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1051 return false;
1052 }
1053 Node* src = argument(0);
1054 Node* tgt = argument(1);
1055
1056 // Make the merge point
1057 RegionNode* result_rgn = new RegionNode(4);
1058 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1059
1060 // Get start addr and length of source string
1061 Node* src_start = array_element_address(src, intcon(0), T_BYTE);
1062 Node* src_count = load_array_length(src);
1063
1064 // Get start addr and length of substring
1065 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1066 Node* tgt_count = load_array_length(tgt);
1067
1068 if (ae == StrIntrinsicNode::UU || ae == StrIntrinsicNode::UL) {
1069 // Divide src size by 2 if String is UTF16 encoded
1070 src_count = _gvn.transform(new RShiftINode(src_count, intcon(1)));
1071 }
1072 if (ae == StrIntrinsicNode::UU) {
1073 // Divide substring size by 2 if String is UTF16 encoded
1074 tgt_count = _gvn.transform(new RShiftINode(tgt_count, intcon(1)));
1075 }
1076
1077 // Check for substr count > string count
1078 Node* cmp = _gvn.transform(new CmpINode(tgt_count, src_count));
1079 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1080 Node* if_gt = generate_slow_guard(bol, NULL);
1081 if (if_gt != NULL) {
1082 result_phi->init_req(2, intcon(-1));
1083 result_rgn->init_req(2, if_gt);
1084 }
1085
1086 if (!stopped()) {
1087 // Check for substr count == 0
1088 cmp = _gvn.transform(new CmpINode(tgt_count, intcon(0)));
1089 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1090 Node* if_zero = generate_slow_guard(bol, NULL);
1091 if (if_zero != NULL) {
1092 result_phi->init_req(3, intcon(0));
1093 result_rgn->init_req(3, if_zero);
1094 }
1095 }
1096
1097 if (!stopped()) {
1098 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1099 result_phi->init_req(1, result);
1100 result_rgn->init_req(1, control());
1101 }
1102 set_control(_gvn.transform(result_rgn));
1103 record_for_igvn(result_rgn);
1104 set_result(_gvn.transform(result_phi));
1105
1106 return true;
1107 }
1108
1109 //-----------------------------inline_string_indexOf-----------------------
1110 bool LibraryCallKit::inline_string_indexOfI(StrIntrinsicNode::ArgEnc ae) {
1111 if (!Matcher::has_match_rule(Op_StrIndexOf) || !UseSSE42Intrinsics) {
1112 return false;
1113 }
1114 assert(callee()->signature()->size() == 5, "String.indexOf() has 5 arguments");
1115 Node* src = argument(0); // byte[]
1116 Node* src_count = argument(1);
1117 Node* tgt = argument(2); // byte[]
1118 Node* tgt_count = argument(3);
1119 Node* from_index = argument(4);
1120
1121 // Java code which calls this method has range checks for from_index value.
1122 src_count = _gvn.transform(new SubINode(src_count, from_index));
1123
1124 // Multiply byte array index by 2 if String is UTF16 encoded
1125 Node* src_offset = (ae == StrIntrinsicNode::LL) ? from_index : _gvn.transform(new LShiftINode(from_index, intcon(1)));
1126 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1127 Node* tgt_start = array_element_address(tgt, intcon(0), T_BYTE);
1128
1129 Node* result = make_string_method_node(Op_StrIndexOf, src_start, src_count, tgt_start, tgt_count, ae);
1130
1131 // The result is index relative to from_index if substring was found, -1 otherwise.
1132 // Generate code which will fold into cmove.
1133 RegionNode* region = new RegionNode(3);
1134 Node* phi = new PhiNode(region, TypeInt::INT);
1135
1136 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1137 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1138
1139 Node* if_lt = generate_slow_guard(bol, NULL);
1140 if (if_lt != NULL) {
1141 // result == -1
1142 phi->init_req(2, result);
1143 region->init_req(2, if_lt);
1144 }
1145 if (!stopped()) {
1146 result = _gvn.transform(new AddINode(result, from_index));
1147 phi->init_req(1, result);
1148 region->init_req(1, control());
1149 }
1150
1151 set_control(_gvn.transform(region));
1152 record_for_igvn(region);
1153 set_result(_gvn.transform(phi));
1154
1155 return true;
1156 }
1157
1158 //-----------------------------inline_string_indexOfChar-----------------------
1159 bool LibraryCallKit::inline_string_indexOfChar() {
1160 if (!Matcher::has_match_rule(Op_StrIndexOfChar) || !(UseSSE > 4)) {
1161 return false;
1162 }
1163 assert(callee()->signature()->size() == 4, "String.indexOfChar() has 4 arguments");
1164 Node* src = argument(0); // byte[]
1165 Node* tgt = argument(1); // tgt is int ch
1166 Node* from_index = argument(2);
1167 Node* max = argument(3);
1168
1169 Node* src_offset = _gvn.transform(new LShiftINode(from_index, intcon(1)));
1170 Node* src_start = array_element_address(src, src_offset, T_BYTE);
1171
1172 Node* src_count = _gvn.transform(new SubINode(max, from_index));
1173
1174 RegionNode* region = new RegionNode(3);
1175 Node* phi = new PhiNode(region, TypeInt::INT);
1176
1177 Node* result = new StrIndexOfCharNode(control(), memory(TypeAryPtr::BYTES), src_start, src_count, tgt, StrIntrinsicNode::none);
1178 C->set_has_split_ifs(true); // Has chance for split-if optimization
1179 _gvn.transform(result);
1180
1181 Node* cmp = _gvn.transform(new CmpINode(result, intcon(0)));
1182 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::lt));
1183
1184 Node* if_lt = generate_slow_guard(bol, NULL);
1185 if (if_lt != NULL) {
1186 // result == -1
1187 phi->init_req(2, result);
1188 region->init_req(2, if_lt);
1189 }
1190 if (!stopped()) {
1191 result = _gvn.transform(new AddINode(result, from_index));
1192 phi->init_req(1, result);
1193 region->init_req(1, control());
1194 }
1195 set_control(_gvn.transform(region));
1196 record_for_igvn(region);
1197 set_result(_gvn.transform(phi));
1198
1199 return true;
1200 }
1201 //---------------------------inline_string_copy---------------------
1202 // compressIt == true --> generate a compressed copy operation (compress char[]/byte[] to byte[])
1203 // int StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
1204 // int StringUTF16.compress(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1205 // compressIt == false --> generate an inflated copy operation (inflate byte[] to char[]/byte[])
1206 // void StringLatin1.inflate(byte[] src, int srcOff, char[] dst, int dstOff, int len)
1207 // void StringLatin1.inflate(byte[] src, int srcOff, byte[] dst, int dstOff, int len)
1208 bool LibraryCallKit::inline_string_copy(bool compress) {
1209 int nargs = 5; // 2 oops, 3 ints
1210 assert(callee()->signature()->size() == nargs, "string copy has 5 arguments");
1211
1212 Node* src = argument(0);
1213 Node* src_offset = argument(1);
1214 Node* dst = argument(2);
1215 Node* dst_offset = argument(3);
1216 Node* length = argument(4);
1217
1218 // Check for allocation before we add nodes that would confuse
1219 // tightly_coupled_allocation()
1220 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1221
1222 // Figure out the size and type of the elements we will be copying.
1223 const Type* src_type = src->Value(&_gvn);
1224 const Type* dst_type = dst->Value(&_gvn);
1225 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1226 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
1227 assert((compress && dst_elem == T_BYTE && (src_elem == T_BYTE || src_elem == T_CHAR)) ||
1228 (!compress && src_elem == T_BYTE && (dst_elem == T_BYTE || dst_elem == T_CHAR)),
1229 "Unsupported array types for inline_string_copy");
1230
1231 // Convert char[] offsets to byte[] offsets
1232 if (compress && src_elem == T_BYTE) {
1233 src_offset = _gvn.transform(new LShiftINode(src_offset, intcon(1)));
1234 } else if (!compress && dst_elem == T_BYTE) {
1235 dst_offset = _gvn.transform(new LShiftINode(dst_offset, intcon(1)));
1236 }
1237
1238 Node* src_start = array_element_address(src, src_offset, src_elem);
1239 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
1240 // 'src_start' points to src array + scaled offset
1241 // 'dst_start' points to dst array + scaled offset
1242 Node* count;
1243 if (compress) {
1244 count = compress_string(src_start, dst_start, length);
1245 } else {
1246 inflate_string(src_start, dst_start, length);
1247 }
1248
1249 if (alloc != NULL) {
1250 if (alloc->maybe_set_complete(&_gvn)) {
1251 // "You break it, you buy it."
1252 InitializeNode* init = alloc->initialization();
1253 assert(init->is_complete(), "we just did this");
1254 init->set_complete_with_arraycopy();
1255 assert(dst->is_CheckCastPP(), "sanity");
1256 assert(dst->in(0)->in(0) == init, "dest pinned");
1257 }
1258 // Do not let stores that initialize this object be reordered with
1259 // a subsequent store that would make this object accessible by
1260 // other threads.
1261 // Record what AllocateNode this StoreStore protects so that
1262 // escape analysis can go from the MemBarStoreStoreNode to the
1263 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1264 // based on the escape status of the AllocateNode.
1265 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1266 }
1267 if (compress) {
1268 set_result(_gvn.transform(count));
1269 }
1270 return true;
1271 }
1272
1273 #ifdef _LP64
1274 #define XTOP ,top() /*additional argument*/
1275 #else //_LP64
1276 #define XTOP /*no additional argument*/
1277 #endif //_LP64
1278
1279 //------------------------inline_string_toBytesU--------------------------
1280 // public static byte[] StringUTF16.toBytes(char[] value, int off, int len)
1281 bool LibraryCallKit::inline_string_toBytesU() {
1282 // Get the arguments.
1283 Node* value = argument(0);
1284 Node* offset = argument(1);
1285 Node* length = argument(2);
1286
1287 Node* newcopy = NULL;
1288
1289 // Set the original stack and the reexecute bit for the interpreter to reexecute
1290 // the bytecode that invokes StringUTF16.toBytes() if deoptimization happens.
1291 { PreserveReexecuteState preexecs(this);
1292 jvms()->set_should_reexecute(true);
1293
1294 // Check if a null path was taken unconditionally.
1295 value = null_check(value);
1296
1297 RegionNode* bailout = new RegionNode(1);
1298 record_for_igvn(bailout);
1299
1300 // Make sure that resulting byte[] length does not overflow Integer.MAX_VALUE
1301 generate_negative_guard(length, bailout);
1302 generate_limit_guard(length, intcon(0), intcon(max_jint/2), bailout);
1303
1304 if (bailout->req() > 1) {
1305 PreserveJVMState pjvms(this);
1306 set_control(_gvn.transform(bailout));
1307 uncommon_trap(Deoptimization::Reason_intrinsic,
1308 Deoptimization::Action_maybe_recompile);
1309 }
1310 if (stopped()) return true;
1311
1312 // Range checks are done by caller.
1313
1314 Node* size = _gvn.transform(new LShiftINode(length, intcon(1)));
1315 Node* klass_node = makecon(TypeKlassPtr::make(ciTypeArrayKlass::make(T_BYTE)));
1316 newcopy = new_array(klass_node, size, 0); // no arguments to push
1317 AllocateArrayNode* alloc = tightly_coupled_allocation(newcopy, NULL);
1318
1319 // Calculate starting addresses.
1320 Node* src_start = array_element_address(value, offset, T_CHAR);
1321 Node* dst_start = basic_plus_adr(newcopy, arrayOopDesc::base_offset_in_bytes(T_BYTE));
1322
1323 // Check if src array address is aligned to HeapWordSize (dst is always aligned)
1324 const TypeInt* toffset = gvn().type(offset)->is_int();
1325 bool aligned = toffset->is_con() && ((toffset->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1326
1327 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1328 const char* copyfunc_name = "arraycopy";
1329 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1330 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1331 OptoRuntime::fast_arraycopy_Type(),
1332 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1333 src_start, dst_start, ConvI2X(length) XTOP);
1334 // Do not let reads from the cloned object float above the arraycopy.
1335 if (alloc != NULL) {
1336 if (alloc->maybe_set_complete(&_gvn)) {
1337 // "You break it, you buy it."
1338 InitializeNode* init = alloc->initialization();
1339 assert(init->is_complete(), "we just did this");
1340 init->set_complete_with_arraycopy();
1341 assert(newcopy->is_CheckCastPP(), "sanity");
1342 assert(newcopy->in(0)->in(0) == init, "dest pinned");
1343 }
1344 // Do not let stores that initialize this object be reordered with
1345 // a subsequent store that would make this object accessible by
1346 // other threads.
1347 // Record what AllocateNode this StoreStore protects so that
1348 // escape analysis can go from the MemBarStoreStoreNode to the
1349 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1350 // based on the escape status of the AllocateNode.
1351 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1352 } else {
1353 insert_mem_bar(Op_MemBarCPUOrder);
1354 }
1355 } // original reexecute is set back here
1356
1357 C->set_has_split_ifs(true); // Has chance for split-if optimization
1358 if (!stopped()) {
1359 set_result(newcopy);
1360 }
1361 return true;
1362 }
1363
1364 //------------------------inline_string_getCharsU--------------------------
1365 // public void StringUTF16.getChars(byte[] value, int srcBegin, int srcEnd, char dst[], int dstBegin)
1366 bool LibraryCallKit::inline_string_getCharsU() {
1367 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1368
1369 // Get the arguments.
1370 Node* value = argument(0);
1371 Node* src_begin = argument(1);
1372 Node* src_end = argument(2); // exclusive offset (i < src_end)
1373 Node* dst = argument(3);
1374 Node* dst_begin = argument(4);
1375
1376 // Check for allocation before we add nodes that would confuse
1377 // tightly_coupled_allocation()
1378 AllocateArrayNode* alloc = tightly_coupled_allocation(dst, NULL);
1379
1380 // Check if a null path was taken unconditionally.
1381 value = null_check(value);
1382 dst = null_check(dst);
1383 if (stopped()) {
1384 return true;
1385 }
1386
1387 // Range checks are done by caller.
1388
1389 // Get length and convert char[] offset to byte[] offset
1390 Node* length = _gvn.transform(new SubINode(src_end, src_begin));
1391 src_begin = _gvn.transform(new LShiftINode(src_begin, intcon(1)));
1392
1393 if (!stopped()) {
1394 // Calculate starting addresses.
1395 Node* src_start = array_element_address(value, src_begin, T_BYTE);
1396 Node* dst_start = array_element_address(dst, dst_begin, T_CHAR);
1397
1398 // Check if array addresses are aligned to HeapWordSize
1399 const TypeInt* tsrc = gvn().type(src_begin)->is_int();
1400 const TypeInt* tdst = gvn().type(dst_begin)->is_int();
1401 bool aligned = tsrc->is_con() && ((tsrc->get_con() * type2aelembytes(T_BYTE)) % HeapWordSize == 0) &&
1402 tdst->is_con() && ((tdst->get_con() * type2aelembytes(T_CHAR)) % HeapWordSize == 0);
1403
1404 // Figure out which arraycopy runtime method to call (disjoint, uninitialized).
1405 const char* copyfunc_name = "arraycopy";
1406 address copyfunc_addr = StubRoutines::select_arraycopy_function(T_CHAR, aligned, true, copyfunc_name, true);
1407 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
1408 OptoRuntime::fast_arraycopy_Type(),
1409 copyfunc_addr, copyfunc_name, TypeRawPtr::BOTTOM,
1410 src_start, dst_start, ConvI2X(length) XTOP);
1411 // Do not let reads from the cloned object float above the arraycopy.
1412 if (alloc != NULL) {
1413 if (alloc->maybe_set_complete(&_gvn)) {
1414 // "You break it, you buy it."
1415 InitializeNode* init = alloc->initialization();
1416 assert(init->is_complete(), "we just did this");
1417 init->set_complete_with_arraycopy();
1418 assert(dst->is_CheckCastPP(), "sanity");
1419 assert(dst->in(0)->in(0) == init, "dest pinned");
1420 }
1421 // Do not let stores that initialize this object be reordered with
1422 // a subsequent store that would make this object accessible by
1423 // other threads.
1424 // Record what AllocateNode this StoreStore protects so that
1425 // escape analysis can go from the MemBarStoreStoreNode to the
1426 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
1427 // based on the escape status of the AllocateNode.
1428 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
1429 } else {
1430 insert_mem_bar(Op_MemBarCPUOrder);
1431 }
1432 }
1433
1434 C->set_has_split_ifs(true); // Has chance for split-if optimization
1435 return true;
1436 }
1437
1438 //----------------------inline_string_char_access----------------------------
1439 // Store/Load char to/from byte[] array.
1440 // static void StringUTF16.putChar(byte[] val, int index, int c)
1441 // static char StringUTF16.getChar(byte[] val, int index)
1442 bool LibraryCallKit::inline_string_char_access(bool is_store) {
1443 Node* value = argument(0);
1444 Node* index = argument(1);
1445 Node* ch = is_store ? argument(2) : NULL;
1446
1447 // This intrinsic accesses byte[] array as char[] array. Computing the offsets
1448 // correctly requires matched array shapes.
1449 assert (arrayOopDesc::base_offset_in_bytes(T_CHAR) == arrayOopDesc::base_offset_in_bytes(T_BYTE),
1450 "sanity: byte[] and char[] bases agree");
1451 assert (type2aelembytes(T_CHAR) == type2aelembytes(T_BYTE)*2,
1452 "sanity: byte[] and char[] scales agree");
1453
1454 Node* adr = array_element_address(value, index, T_CHAR);
1455 if (is_store) {
1456 (void) store_to_memory(control(), adr, ch, T_CHAR, TypeAryPtr::BYTES, MemNode::unordered);
1457 } else {
1458 ch = make_load(control(), adr, TypeInt::CHAR, T_CHAR, MemNode::unordered);
1459 set_result(ch);
1460 }
1461 return true;
1462 }
1463
1464 //--------------------------round_double_node--------------------------------
1465 // Round a double node if necessary.
1466 Node* LibraryCallKit::round_double_node(Node* n) {
1467 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1468 n = _gvn.transform(new RoundDoubleNode(0, n));
1469 return n;
1470 }
1471
1472 //------------------------------inline_math-----------------------------------
1473 // public static double Math.abs(double)
1474 // public static double Math.sqrt(double)
1475 // public static double Math.log(double)
1476 // public static double Math.log10(double)
1477 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1478 Node* arg = round_double_node(argument(0));
1479 Node* n;
1480 switch (id) {
1481 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1482 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1483 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1484 default: fatal_unexpected_iid(id); break;
1485 }
1486 set_result(_gvn.transform(n));
1487 return true;
1488 }
1489
1490 //------------------------------inline_trig----------------------------------
1491 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1492 // argument reduction which will turn into a fast/slow diamond.
1493 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1494 Node* arg = round_double_node(argument(0));
1495 Node* n = NULL;
1496
1497 switch (id) {
1498 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1499 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1500 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1501 default: fatal_unexpected_iid(id); break;
1502 }
1503 n = _gvn.transform(n);
1504
1505 // Rounding required? Check for argument reduction!
1506 if (Matcher::strict_fp_requires_explicit_rounding) {
1507 static const double pi_4 = 0.7853981633974483;
1508 static const double neg_pi_4 = -0.7853981633974483;
1509 // pi/2 in 80-bit extended precision
1510 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1511 // -pi/2 in 80-bit extended precision
1512 // static const unsigned char neg_pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0xbf,0x00,0x00,0x00,0x00,0x00,0x00};
1513 // Cutoff value for using this argument reduction technique
1514 //static const double pi_2_minus_epsilon = 1.564660403643354;
1515 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1516
1517 // Pseudocode for sin:
1518 // if (x <= Math.PI / 4.0) {
1519 // if (x >= -Math.PI / 4.0) return fsin(x);
1520 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1521 // } else {
1522 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1523 // }
1524 // return StrictMath.sin(x);
1525
1526 // Pseudocode for cos:
1527 // if (x <= Math.PI / 4.0) {
1528 // if (x >= -Math.PI / 4.0) return fcos(x);
1529 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1530 // } else {
1531 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1532 // }
1533 // return StrictMath.cos(x);
1534
1535 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1536 // requires a special machine instruction to load it. Instead we'll try
1537 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1538 // probably do the math inside the SIN encoding.
1539
1540 // Make the merge point
1541 RegionNode* r = new RegionNode(3);
1542 Node* phi = new PhiNode(r, Type::DOUBLE);
1543
1544 // Flatten arg so we need only 1 test
1545 Node *abs = _gvn.transform(new AbsDNode(arg));
1546 // Node for PI/4 constant
1547 Node *pi4 = makecon(TypeD::make(pi_4));
1548 // Check PI/4 : abs(arg)
1549 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1550 // Check: If PI/4 < abs(arg) then go slow
1551 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1552 // Branch either way
1553 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1554 set_control(opt_iff(r,iff));
1555
1556 // Set fast path result
1557 phi->init_req(2, n);
1558
1559 // Slow path - non-blocking leaf call
1560 Node* call = NULL;
1561 switch (id) {
1562 case vmIntrinsics::_dsin:
1563 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1564 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1565 "Sin", NULL, arg, top());
1566 break;
1567 case vmIntrinsics::_dcos:
1568 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1569 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1570 "Cos", NULL, arg, top());
1571 break;
1572 case vmIntrinsics::_dtan:
1573 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1574 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1575 "Tan", NULL, arg, top());
1576 break;
1577 }
1578 assert(control()->in(0) == call, "");
1579 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1580 r->init_req(1, control());
1581 phi->init_req(1, slow_result);
1582
1583 // Post-merge
1584 set_control(_gvn.transform(r));
1585 record_for_igvn(r);
1586 n = _gvn.transform(phi);
1587
1588 C->set_has_split_ifs(true); // Has chance for split-if optimization
1589 }
1590 set_result(n);
1591 return true;
1592 }
1593
1594 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1595 //-------------------
1596 //result=(result.isNaN())? funcAddr():result;
1597 // Check: If isNaN() by checking result!=result? then either trap
1598 // or go to runtime
1599 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1600 // Build the boolean node
1601 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1602
1603 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1604 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1605 // The pow or exp intrinsic returned a NaN, which requires a call
1606 // to the runtime. Recompile with the runtime call.
1607 uncommon_trap(Deoptimization::Reason_intrinsic,
1608 Deoptimization::Action_make_not_entrant);
1609 }
1610 return result;
1611 } else {
1612 // If this inlining ever returned NaN in the past, we compile a call
1613 // to the runtime to properly handle corner cases
1614
1615 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1616 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1617 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1618
1619 if (!if_slow->is_top()) {
1620 RegionNode* result_region = new RegionNode(3);
1621 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1622
1623 result_region->init_req(1, if_fast);
1624 result_val->init_req(1, result);
1625
1626 set_control(if_slow);
1627
1628 const TypePtr* no_memory_effects = NULL;
1629 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1630 no_memory_effects,
1631 x, top(), y, y ? top() : NULL);
1632 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1633 #ifdef ASSERT
1634 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1635 assert(value_top == top(), "second value must be top");
1636 #endif
1637
1638 result_region->init_req(2, control());
1639 result_val->init_req(2, value);
1640 set_control(_gvn.transform(result_region));
1641 return _gvn.transform(result_val);
1642 } else {
1643 return result;
1644 }
1645 }
1646 }
1647
1648 //------------------------------inline_pow-------------------------------------
1649 // Inline power instructions, if possible.
1650 bool LibraryCallKit::inline_pow() {
1651 // Pseudocode for pow
1652 // if (y == 2) {
1653 // return x * x;
1654 // } else {
1655 // if (x <= 0.0) {
1656 // long longy = (long)y;
1657 // if ((double)longy == y) { // if y is long
1658 // if (y + 1 == y) longy = 0; // huge number: even
1659 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1660 // } else {
1661 // result = NaN;
1662 // }
1663 // } else {
1664 // result = DPow(x,y);
1665 // }
1666 // if (result != result)? {
1667 // result = uncommon_trap() or runtime_call();
1668 // }
1669 // return result;
1670 // }
1671
1672 Node* x = round_double_node(argument(0));
1673 Node* y = round_double_node(argument(2));
1674
1675 Node* result = NULL;
1676
1677 Node* const_two_node = makecon(TypeD::make(2.0));
1678 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1679 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1680 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1681 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1682 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1683
1684 RegionNode* region_node = new RegionNode(3);
1685 region_node->init_req(1, if_true);
1686
1687 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1688 // special case for x^y where y == 2, we can convert it to x * x
1689 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1690
1691 // set control to if_false since we will now process the false branch
1692 set_control(if_false);
1693
1694 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1695 // Short form: skip the fancy tests and just check for NaN result.
1696 result = _gvn.transform(new PowDNode(C, control(), x, y));
1697 } else {
1698 // If this inlining ever returned NaN in the past, include all
1699 // checks + call to the runtime.
1700
1701 // Set the merge point for If node with condition of (x <= 0.0)
1702 // There are four possible paths to region node and phi node
1703 RegionNode *r = new RegionNode(4);
1704 Node *phi = new PhiNode(r, Type::DOUBLE);
1705
1706 // Build the first if node: if (x <= 0.0)
1707 // Node for 0 constant
1708 Node *zeronode = makecon(TypeD::ZERO);
1709 // Check x:0
1710 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1711 // Check: If (x<=0) then go complex path
1712 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1713 // Branch either way
1714 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1715 // Fast path taken; set region slot 3
1716 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1717 r->init_req(3,fast_taken); // Capture fast-control
1718
1719 // Fast path not-taken, i.e. slow path
1720 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1721
1722 // Set fast path result
1723 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1724 phi->init_req(3, fast_result);
1725
1726 // Complex path
1727 // Build the second if node (if y is long)
1728 // Node for (long)y
1729 Node *longy = _gvn.transform(new ConvD2LNode(y));
1730 // Node for (double)((long) y)
1731 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1732 // Check (double)((long) y) : y
1733 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1734 // Check if (y isn't long) then go to slow path
1735
1736 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1737 // Branch either way
1738 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1739 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1740
1741 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1742
1743 // Calculate DPow(abs(x), y)*(1 & (long)y)
1744 // Node for constant 1
1745 Node *conone = longcon(1);
1746 // 1& (long)y
1747 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1748
1749 // A huge number is always even. Detect a huge number by checking
1750 // if y + 1 == y and set integer to be tested for parity to 0.
1751 // Required for corner case:
1752 // (long)9.223372036854776E18 = max_jlong
1753 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1754 // max_jlong is odd but 9.223372036854776E18 is even
1755 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1756 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1757 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1758 Node* correctedsign = NULL;
1759 if (ConditionalMoveLimit != 0) {
1760 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1761 } else {
1762 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1763 RegionNode *r = new RegionNode(3);
1764 Node *phi = new PhiNode(r, TypeLong::LONG);
1765 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1766 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1767 phi->init_req(1, signnode);
1768 phi->init_req(2, longcon(0));
1769 correctedsign = _gvn.transform(phi);
1770 ylong_path = _gvn.transform(r);
1771 record_for_igvn(r);
1772 }
1773
1774 // zero node
1775 Node *conzero = longcon(0);
1776 // Check (1&(long)y)==0?
1777 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1778 // Check if (1&(long)y)!=0?, if so the result is negative
1779 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1780 // abs(x)
1781 Node *absx=_gvn.transform(new AbsDNode(x));
1782 // abs(x)^y
1783 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1784 // -abs(x)^y
1785 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1786 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1787 Node *signresult = NULL;
1788 if (ConditionalMoveLimit != 0) {
1789 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1790 } else {
1791 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1792 RegionNode *r = new RegionNode(3);
1793 Node *phi = new PhiNode(r, Type::DOUBLE);
1794 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1795 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1796 phi->init_req(1, absxpowy);
1797 phi->init_req(2, negabsxpowy);
1798 signresult = _gvn.transform(phi);
1799 ylong_path = _gvn.transform(r);
1800 record_for_igvn(r);
1801 }
1802 // Set complex path fast result
1803 r->init_req(2, ylong_path);
1804 phi->init_req(2, signresult);
1805
1806 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1807 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1808 r->init_req(1,slow_path);
1809 phi->init_req(1,slow_result);
1810
1811 // Post merge
1812 set_control(_gvn.transform(r));
1813 record_for_igvn(r);
1814 result = _gvn.transform(phi);
1815 }
1816
1817 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1818
1819 // control from finish_pow_exp is now input to the region node
1820 region_node->set_req(2, control());
1821 // the result from finish_pow_exp is now input to the phi node
1822 phi_node->init_req(2, result);
1823 set_control(_gvn.transform(region_node));
1824 record_for_igvn(region_node);
1825 set_result(_gvn.transform(phi_node));
1826
1827 C->set_has_split_ifs(true); // Has chance for split-if optimization
1828 return true;
1829 }
1830
1831 //------------------------------runtime_math-----------------------------
1832 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1833 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1834 "must be (DD)D or (D)D type");
1835
1836 // Inputs
1837 Node* a = round_double_node(argument(0));
1838 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1839
1840 const TypePtr* no_memory_effects = NULL;
1841 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1842 no_memory_effects,
1843 a, top(), b, b ? top() : NULL);
1844 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1845 #ifdef ASSERT
1846 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1847 assert(value_top == top(), "second value must be top");
1848 #endif
1849
1850 set_result(value);
1851 return true;
1852 }
1853
1854 //------------------------------inline_math_native-----------------------------
1855 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1856 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1857 switch (id) {
1858 // These intrinsics are not properly supported on all hardware
1859 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1860 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1861 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1862 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1863 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1864 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1865
1866 case vmIntrinsics::_dlog:
1867 return StubRoutines::dlog() != NULL ?
1868 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1869 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1870 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1871 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1872
1873 // These intrinsics are supported on all hardware
1874 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1875 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1876
1877 case vmIntrinsics::_dexp:
1878 return StubRoutines::dexp() != NULL ?
1879 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1880 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1881 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1882 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1883 #undef FN_PTR
1884
1885 // These intrinsics are not yet correctly implemented
1886 case vmIntrinsics::_datan2:
1887 return false;
1888
1889 default:
1890 fatal_unexpected_iid(id);
1891 return false;
1892 }
1893 }
1894
1895 static bool is_simple_name(Node* n) {
1896 return (n->req() == 1 // constant
1897 || (n->is_Type() && n->as_Type()->type()->singleton())
1898 || n->is_Proj() // parameter or return value
1899 || n->is_Phi() // local of some sort
1900 );
1901 }
1902
1903 //----------------------------inline_notify-----------------------------------*
1904 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1905 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1906 address func;
1907 if (id == vmIntrinsics::_notify) {
1908 func = OptoRuntime::monitor_notify_Java();
1909 } else {
1910 func = OptoRuntime::monitor_notifyAll_Java();
1911 }
1912 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1913 make_slow_call_ex(call, env()->Throwable_klass(), false);
1914 return true;
1915 }
1916
1917
1918 //----------------------------inline_min_max-----------------------------------
1919 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1920 set_result(generate_min_max(id, argument(0), argument(1)));
1921 return true;
1922 }
1923
1924 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1925 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1926 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1927 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1928 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1929
1930 {
1931 PreserveJVMState pjvms(this);
1932 PreserveReexecuteState preexecs(this);
1933 jvms()->set_should_reexecute(true);
1934
1935 set_control(slow_path);
1936 set_i_o(i_o());
1937
1938 uncommon_trap(Deoptimization::Reason_intrinsic,
1939 Deoptimization::Action_none);
1940 }
1941
1942 set_control(fast_path);
1943 set_result(math);
1944 }
1945
1946 template <typename OverflowOp>
1947 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1948 typedef typename OverflowOp::MathOp MathOp;
1949
1950 MathOp* mathOp = new MathOp(arg1, arg2);
1951 Node* operation = _gvn.transform( mathOp );
1952 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1953 inline_math_mathExact(operation, ofcheck);
1954 return true;
1955 }
1956
1957 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1958 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1959 }
1960
1961 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1962 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1963 }
1964
1965 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1966 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1967 }
1968
1969 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1970 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1971 }
1972
1973 bool LibraryCallKit::inline_math_negateExactI() {
1974 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1975 }
1976
1977 bool LibraryCallKit::inline_math_negateExactL() {
1978 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1979 }
1980
1981 bool LibraryCallKit::inline_math_multiplyExactI() {
1982 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1983 }
1984
1985 bool LibraryCallKit::inline_math_multiplyExactL() {
1986 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1987 }
1988
1989 Node*
1990 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1991 // These are the candidate return value:
1992 Node* xvalue = x0;
1993 Node* yvalue = y0;
1994
1995 if (xvalue == yvalue) {
1996 return xvalue;
1997 }
1998
1999 bool want_max = (id == vmIntrinsics::_max);
2000
2001 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2002 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2003 if (txvalue == NULL || tyvalue == NULL) return top();
2004 // This is not really necessary, but it is consistent with a
2005 // hypothetical MaxINode::Value method:
2006 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2007
2008 // %%% This folding logic should (ideally) be in a different place.
2009 // Some should be inside IfNode, and there to be a more reliable
2010 // transformation of ?: style patterns into cmoves. We also want
2011 // more powerful optimizations around cmove and min/max.
2012
2013 // Try to find a dominating comparison of these guys.
2014 // It can simplify the index computation for Arrays.copyOf
2015 // and similar uses of System.arraycopy.
2016 // First, compute the normalized version of CmpI(x, y).
2017 int cmp_op = Op_CmpI;
2018 Node* xkey = xvalue;
2019 Node* ykey = yvalue;
2020 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
2021 if (ideal_cmpxy->is_Cmp()) {
2022 // E.g., if we have CmpI(length - offset, count),
2023 // it might idealize to CmpI(length, count + offset)
2024 cmp_op = ideal_cmpxy->Opcode();
2025 xkey = ideal_cmpxy->in(1);
2026 ykey = ideal_cmpxy->in(2);
2027 }
2028
2029 // Start by locating any relevant comparisons.
2030 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2031 Node* cmpxy = NULL;
2032 Node* cmpyx = NULL;
2033 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2034 Node* cmp = start_from->fast_out(k);
2035 if (cmp->outcnt() > 0 && // must have prior uses
2036 cmp->in(0) == NULL && // must be context-independent
2037 cmp->Opcode() == cmp_op) { // right kind of compare
2038 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2039 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2040 }
2041 }
2042
2043 const int NCMPS = 2;
2044 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2045 int cmpn;
2046 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2047 if (cmps[cmpn] != NULL) break; // find a result
2048 }
2049 if (cmpn < NCMPS) {
2050 // Look for a dominating test that tells us the min and max.
2051 int depth = 0; // Limit search depth for speed
2052 Node* dom = control();
2053 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2054 if (++depth >= 100) break;
2055 Node* ifproj = dom;
2056 if (!ifproj->is_Proj()) continue;
2057 Node* iff = ifproj->in(0);
2058 if (!iff->is_If()) continue;
2059 Node* bol = iff->in(1);
2060 if (!bol->is_Bool()) continue;
2061 Node* cmp = bol->in(1);
2062 if (cmp == NULL) continue;
2063 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2064 if (cmps[cmpn] == cmp) break;
2065 if (cmpn == NCMPS) continue;
2066 BoolTest::mask btest = bol->as_Bool()->_test._test;
2067 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2068 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2069 // At this point, we know that 'x btest y' is true.
2070 switch (btest) {
2071 case BoolTest::eq:
2072 // They are proven equal, so we can collapse the min/max.
2073 // Either value is the answer. Choose the simpler.
2074 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2075 return yvalue;
2076 return xvalue;
2077 case BoolTest::lt: // x < y
2078 case BoolTest::le: // x <= y
2079 return (want_max ? yvalue : xvalue);
2080 case BoolTest::gt: // x > y
2081 case BoolTest::ge: // x >= y
2082 return (want_max ? xvalue : yvalue);
2083 }
2084 }
2085 }
2086
2087 // We failed to find a dominating test.
2088 // Let's pick a test that might GVN with prior tests.
2089 Node* best_bol = NULL;
2090 BoolTest::mask best_btest = BoolTest::illegal;
2091 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2092 Node* cmp = cmps[cmpn];
2093 if (cmp == NULL) continue;
2094 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2095 Node* bol = cmp->fast_out(j);
2096 if (!bol->is_Bool()) continue;
2097 BoolTest::mask btest = bol->as_Bool()->_test._test;
2098 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2099 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2100 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2101 best_bol = bol->as_Bool();
2102 best_btest = btest;
2103 }
2104 }
2105 }
2106
2107 Node* answer_if_true = NULL;
2108 Node* answer_if_false = NULL;
2109 switch (best_btest) {
2110 default:
2111 if (cmpxy == NULL)
2112 cmpxy = ideal_cmpxy;
2113 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2114 // and fall through:
2115 case BoolTest::lt: // x < y
2116 case BoolTest::le: // x <= y
2117 answer_if_true = (want_max ? yvalue : xvalue);
2118 answer_if_false = (want_max ? xvalue : yvalue);
2119 break;
2120 case BoolTest::gt: // x > y
2121 case BoolTest::ge: // x >= y
2122 answer_if_true = (want_max ? xvalue : yvalue);
2123 answer_if_false = (want_max ? yvalue : xvalue);
2124 break;
2125 }
2126
2127 jint hi, lo;
2128 if (want_max) {
2129 // We can sharpen the minimum.
2130 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2131 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2132 } else {
2133 // We can sharpen the maximum.
2134 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2135 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2136 }
2137
2138 // Use a flow-free graph structure, to avoid creating excess control edges
2139 // which could hinder other optimizations.
2140 // Since Math.min/max is often used with arraycopy, we want
2141 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2142 Node* cmov = CMoveNode::make(NULL, best_bol,
2143 answer_if_false, answer_if_true,
2144 TypeInt::make(lo, hi, widen));
2145
2146 return _gvn.transform(cmov);
2147
2148 /*
2149 // This is not as desirable as it may seem, since Min and Max
2150 // nodes do not have a full set of optimizations.
2151 // And they would interfere, anyway, with 'if' optimizations
2152 // and with CMoveI canonical forms.
2153 switch (id) {
2154 case vmIntrinsics::_min:
2155 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2156 case vmIntrinsics::_max:
2157 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2158 default:
2159 ShouldNotReachHere();
2160 }
2161 */
2162 }
2163
2164 inline int
2165 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2166 const TypePtr* base_type = TypePtr::NULL_PTR;
2167 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2168 if (base_type == NULL) {
2169 // Unknown type.
2170 return Type::AnyPtr;
2171 } else if (base_type == TypePtr::NULL_PTR) {
2172 // Since this is a NULL+long form, we have to switch to a rawptr.
2173 base = _gvn.transform(new CastX2PNode(offset));
2174 offset = MakeConX(0);
2175 return Type::RawPtr;
2176 } else if (base_type->base() == Type::RawPtr) {
2177 return Type::RawPtr;
2178 } else if (base_type->isa_oopptr()) {
2179 // Base is never null => always a heap address.
2180 if (base_type->ptr() == TypePtr::NotNull) {
2181 return Type::OopPtr;
2182 }
2183 // Offset is small => always a heap address.
2184 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2185 if (offset_type != NULL &&
2186 base_type->offset() == 0 && // (should always be?)
2187 offset_type->_lo >= 0 &&
2188 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2189 return Type::OopPtr;
2190 }
2191 // Otherwise, it might either be oop+off or NULL+addr.
2192 return Type::AnyPtr;
2193 } else {
2194 // No information:
2195 return Type::AnyPtr;
2196 }
2197 }
2198
2199 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2200 int kind = classify_unsafe_addr(base, offset);
2201 if (kind == Type::RawPtr) {
2202 return basic_plus_adr(top(), base, offset);
2203 } else {
2204 return basic_plus_adr(base, offset);
2205 }
2206 }
2207
2208 //--------------------------inline_number_methods-----------------------------
2209 // inline int Integer.numberOfLeadingZeros(int)
2210 // inline int Long.numberOfLeadingZeros(long)
2211 //
2212 // inline int Integer.numberOfTrailingZeros(int)
2213 // inline int Long.numberOfTrailingZeros(long)
2214 //
2215 // inline int Integer.bitCount(int)
2216 // inline int Long.bitCount(long)
2217 //
2218 // inline char Character.reverseBytes(char)
2219 // inline short Short.reverseBytes(short)
2220 // inline int Integer.reverseBytes(int)
2221 // inline long Long.reverseBytes(long)
2222 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2223 Node* arg = argument(0);
2224 Node* n;
2225 switch (id) {
2226 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2227 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2228 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2229 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2230 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2231 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2232 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2233 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2234 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2235 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2236 default: fatal_unexpected_iid(id); break;
2237 }
2238 set_result(_gvn.transform(n));
2239 return true;
2240 }
2241
2242 //----------------------------inline_unsafe_access----------------------------
2243
2244 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2245
2246 // Helper that guards and inserts a pre-barrier.
2247 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2248 Node* pre_val, bool need_mem_bar) {
2249 // We could be accessing the referent field of a reference object. If so, when G1
2250 // is enabled, we need to log the value in the referent field in an SATB buffer.
2251 // This routine performs some compile time filters and generates suitable
2252 // runtime filters that guard the pre-barrier code.
2253 // Also add memory barrier for non volatile load from the referent field
2254 // to prevent commoning of loads across safepoint.
2255 if (!UseG1GC && !need_mem_bar)
2256 return;
2257
2258 // Some compile time checks.
2259
2260 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2261 const TypeX* otype = offset->find_intptr_t_type();
2262 if (otype != NULL && otype->is_con() &&
2263 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2264 // Constant offset but not the reference_offset so just return
2265 return;
2266 }
2267
2268 // We only need to generate the runtime guards for instances.
2269 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2270 if (btype != NULL) {
2271 if (btype->isa_aryptr()) {
2272 // Array type so nothing to do
2273 return;
2274 }
2275
2276 const TypeInstPtr* itype = btype->isa_instptr();
2277 if (itype != NULL) {
2278 // Can the klass of base_oop be statically determined to be
2279 // _not_ a sub-class of Reference and _not_ Object?
2280 ciKlass* klass = itype->klass();
2281 if ( klass->is_loaded() &&
2282 !klass->is_subtype_of(env()->Reference_klass()) &&
2283 !env()->Object_klass()->is_subtype_of(klass)) {
2284 return;
2285 }
2286 }
2287 }
2288
2289 // The compile time filters did not reject base_oop/offset so
2290 // we need to generate the following runtime filters
2291 //
2292 // if (offset == java_lang_ref_Reference::_reference_offset) {
2293 // if (instance_of(base, java.lang.ref.Reference)) {
2294 // pre_barrier(_, pre_val, ...);
2295 // }
2296 // }
2297
2298 float likely = PROB_LIKELY( 0.999);
2299 float unlikely = PROB_UNLIKELY(0.999);
2300
2301 IdealKit ideal(this);
2302 #define __ ideal.
2303
2304 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2305
2306 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2307 // Update graphKit memory and control from IdealKit.
2308 sync_kit(ideal);
2309
2310 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2311 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2312
2313 // Update IdealKit memory and control from graphKit.
2314 __ sync_kit(this);
2315
2316 Node* one = __ ConI(1);
2317 // is_instof == 0 if base_oop == NULL
2318 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2319
2320 // Update graphKit from IdeakKit.
2321 sync_kit(ideal);
2322
2323 // Use the pre-barrier to record the value in the referent field
2324 pre_barrier(false /* do_load */,
2325 __ ctrl(),
2326 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2327 pre_val /* pre_val */,
2328 T_OBJECT);
2329 if (need_mem_bar) {
2330 // Add memory barrier to prevent commoning reads from this field
2331 // across safepoint since GC can change its value.
2332 insert_mem_bar(Op_MemBarCPUOrder);
2333 }
2334 // Update IdealKit from graphKit.
2335 __ sync_kit(this);
2336
2337 } __ end_if(); // _ref_type != ref_none
2338 } __ end_if(); // offset == referent_offset
2339
2340 // Final sync IdealKit and GraphKit.
2341 final_sync(ideal);
2342 #undef __
2343 }
2344
2345
2346 // Interpret Unsafe.fieldOffset cookies correctly:
2347 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2348
2349 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2350 // Attempt to infer a sharper value type from the offset and base type.
2351 ciKlass* sharpened_klass = NULL;
2352
2353 // See if it is an instance field, with an object type.
2354 if (alias_type->field() != NULL) {
2355 assert(!is_native_ptr, "native pointer op cannot use a java address");
2356 if (alias_type->field()->type()->is_klass()) {
2357 sharpened_klass = alias_type->field()->type()->as_klass();
2358 }
2359 }
2360
2361 // See if it is a narrow oop array.
2362 if (adr_type->isa_aryptr()) {
2363 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2364 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2365 if (elem_type != NULL) {
2366 sharpened_klass = elem_type->klass();
2367 }
2368 }
2369 }
2370
2371 // The sharpened class might be unloaded if there is no class loader
2372 // contraint in place.
2373 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2374 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2375
2376 #ifndef PRODUCT
2377 if (C->print_intrinsics() || C->print_inlining()) {
2378 tty->print(" from base type: "); adr_type->dump();
2379 tty->print(" sharpened value: "); tjp->dump();
2380 }
2381 #endif
2382 // Sharpen the value type.
2383 return tjp;
2384 }
2385 return NULL;
2386 }
2387
2388 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile, bool unaligned) {
2389 if (callee()->is_static()) return false; // caller must have the capability!
2390
2391 #ifndef PRODUCT
2392 {
2393 ResourceMark rm;
2394 // Check the signatures.
2395 ciSignature* sig = callee()->signature();
2396 #ifdef ASSERT
2397 if (!is_store) {
2398 // Object getObject(Object base, int/long offset), etc.
2399 BasicType rtype = sig->return_type()->basic_type();
2400 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2401 rtype = T_ADDRESS; // it is really a C void*
2402 assert(rtype == type, "getter must return the expected value");
2403 if (!is_native_ptr) {
2404 assert(sig->count() == 2, "oop getter has 2 arguments");
2405 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2406 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2407 } else {
2408 assert(sig->count() == 1, "native getter has 1 argument");
2409 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2410 }
2411 } else {
2412 // void putObject(Object base, int/long offset, Object x), etc.
2413 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2414 if (!is_native_ptr) {
2415 assert(sig->count() == 3, "oop putter has 3 arguments");
2416 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2417 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2418 } else {
2419 assert(sig->count() == 2, "native putter has 2 arguments");
2420 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2421 }
2422 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2423 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2424 vtype = T_ADDRESS; // it is really a C void*
2425 assert(vtype == type, "putter must accept the expected value");
2426 }
2427 #endif // ASSERT
2428 }
2429 #endif //PRODUCT
2430
2431 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2432
2433 Node* receiver = argument(0); // type: oop
2434
2435 // Build address expression.
2436 Node* adr;
2437 Node* heap_base_oop = top();
2438 Node* offset = top();
2439 Node* val;
2440
2441 if (!is_native_ptr) {
2442 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2443 Node* base = argument(1); // type: oop
2444 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2445 offset = argument(2); // type: long
2446 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2447 // to be plain byte offsets, which are also the same as those accepted
2448 // by oopDesc::field_base.
2449 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2450 "fieldOffset must be byte-scaled");
2451 // 32-bit machines ignore the high half!
2452 offset = ConvL2X(offset);
2453 adr = make_unsafe_address(base, offset);
2454 heap_base_oop = base;
2455 val = is_store ? argument(4) : NULL;
2456 } else {
2457 Node* ptr = argument(1); // type: long
2458 ptr = ConvL2X(ptr); // adjust Java long to machine word
2459 adr = make_unsafe_address(NULL, ptr);
2460 val = is_store ? argument(3) : NULL;
2461 }
2462
2463 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2464
2465 // First guess at the value type.
2466 const Type *value_type = Type::get_const_basic_type(type);
2467
2468 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2469 // there was not enough information to nail it down.
2470 Compile::AliasType* alias_type = C->alias_type(adr_type);
2471 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2472
2473 // We will need memory barriers unless we can determine a unique
2474 // alias category for this reference. (Note: If for some reason
2475 // the barriers get omitted and the unsafe reference begins to "pollute"
2476 // the alias analysis of the rest of the graph, either Compile::can_alias
2477 // or Compile::must_alias will throw a diagnostic assert.)
2478 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2479
2480 // If we are reading the value of the referent field of a Reference
2481 // object (either by using Unsafe directly or through reflection)
2482 // then, if G1 is enabled, we need to record the referent in an
2483 // SATB log buffer using the pre-barrier mechanism.
2484 // Also we need to add memory barrier to prevent commoning reads
2485 // from this field across safepoint since GC can change its value.
2486 bool need_read_barrier = !is_native_ptr && !is_store &&
2487 offset != top() && heap_base_oop != top();
2488
2489 if (!is_store && type == T_OBJECT) {
2490 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2491 if (tjp != NULL) {
2492 value_type = tjp;
2493 }
2494 }
2495
2496 receiver = null_check(receiver);
2497 if (stopped()) {
2498 return true;
2499 }
2500 // Heap pointers get a null-check from the interpreter,
2501 // as a courtesy. However, this is not guaranteed by Unsafe,
2502 // and it is not possible to fully distinguish unintended nulls
2503 // from intended ones in this API.
2504
2505 if (is_volatile) {
2506 // We need to emit leading and trailing CPU membars (see below) in
2507 // addition to memory membars when is_volatile. This is a little
2508 // too strong, but avoids the need to insert per-alias-type
2509 // volatile membars (for stores; compare Parse::do_put_xxx), which
2510 // we cannot do effectively here because we probably only have a
2511 // rough approximation of type.
2512 need_mem_bar = true;
2513 // For Stores, place a memory ordering barrier now.
2514 if (is_store) {
2515 insert_mem_bar(Op_MemBarRelease);
2516 } else {
2517 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2518 insert_mem_bar(Op_MemBarVolatile);
2519 }
2520 }
2521 }
2522
2523 // Memory barrier to prevent normal and 'unsafe' accesses from
2524 // bypassing each other. Happens after null checks, so the
2525 // exception paths do not take memory state from the memory barrier,
2526 // so there's no problems making a strong assert about mixing users
2527 // of safe & unsafe memory.
2528 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2529
2530 assert(alias_type->adr_type() == TypeRawPtr::BOTTOM || alias_type->adr_type() == TypeOopPtr::BOTTOM ||
2531 alias_type->field() != NULL || alias_type->element() != NULL, "field, array element or unknown");
2532 bool mismatched = false;
2533 if (alias_type->element() != NULL || alias_type->field() != NULL) {
2534 BasicType bt;
2535 if (alias_type->element() != NULL) {
2536 const Type* element = alias_type->element();
2537 bt = element->isa_narrowoop() ? T_OBJECT : element->array_element_basic_type();
2538 } else {
2539 bt = alias_type->field()->type()->basic_type();
2540 }
2541 if (bt != type) {
2542 mismatched = true;
2543 }
2544 }
2545 assert(type != T_OBJECT || !unaligned, "unaligned access not supported with object type");
2546
2547 if (!is_store) {
2548 Node* p = NULL;
2549 // Try to constant fold a load from a constant field
2550 ciField* field = alias_type->field();
2551 if (heap_base_oop != top() &&
2552 field != NULL && field->is_constant() && field->layout_type() == type) {
2553 // final or stable field
2554 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2555 if (con_type != NULL) {
2556 p = makecon(con_type);
2557 }
2558 }
2559 if (p == NULL) {
2560 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2561 // To be valid, unsafe loads may depend on other conditions than
2562 // the one that guards them: pin the Load node
2563 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile, unaligned, mismatched);
2564 // load value
2565 switch (type) {
2566 case T_BOOLEAN:
2567 case T_CHAR:
2568 case T_BYTE:
2569 case T_SHORT:
2570 case T_INT:
2571 case T_LONG:
2572 case T_FLOAT:
2573 case T_DOUBLE:
2574 break;
2575 case T_OBJECT:
2576 if (need_read_barrier) {
2577 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2578 }
2579 break;
2580 case T_ADDRESS:
2581 // Cast to an int type.
2582 p = _gvn.transform(new CastP2XNode(NULL, p));
2583 p = ConvX2UL(p);
2584 break;
2585 default:
2586 fatal("unexpected type %d: %s", type, type2name(type));
2587 break;
2588 }
2589 }
2590 // The load node has the control of the preceding MemBarCPUOrder. All
2591 // following nodes will have the control of the MemBarCPUOrder inserted at
2592 // the end of this method. So, pushing the load onto the stack at a later
2593 // point is fine.
2594 set_result(p);
2595 } else {
2596 // place effect of store into memory
2597 switch (type) {
2598 case T_DOUBLE:
2599 val = dstore_rounding(val);
2600 break;
2601 case T_ADDRESS:
2602 // Repackage the long as a pointer.
2603 val = ConvL2X(val);
2604 val = _gvn.transform(new CastX2PNode(val));
2605 break;
2606 }
2607
2608 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2609 if (type != T_OBJECT ) {
2610 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile, unaligned, mismatched);
2611 } else {
2612 // Possibly an oop being stored to Java heap or native memory
2613 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2614 // oop to Java heap.
2615 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2616 } else {
2617 // We can't tell at compile time if we are storing in the Java heap or outside
2618 // of it. So we need to emit code to conditionally do the proper type of
2619 // store.
2620
2621 IdealKit ideal(this);
2622 #define __ ideal.
2623 // QQQ who knows what probability is here??
2624 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2625 // Sync IdealKit and graphKit.
2626 sync_kit(ideal);
2627 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo, mismatched);
2628 // Update IdealKit memory.
2629 __ sync_kit(this);
2630 } __ else_(); {
2631 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile, mismatched);
2632 } __ end_if();
2633 // Final sync IdealKit and GraphKit.
2634 final_sync(ideal);
2635 #undef __
2636 }
2637 }
2638 }
2639
2640 if (is_volatile) {
2641 if (!is_store) {
2642 insert_mem_bar(Op_MemBarAcquire);
2643 } else {
2644 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2645 insert_mem_bar(Op_MemBarVolatile);
2646 }
2647 }
2648 }
2649
2650 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2651
2652 return true;
2653 }
2654
2655 //----------------------------inline_unsafe_load_store----------------------------
2656 // This method serves a couple of different customers (depending on LoadStoreKind):
2657 //
2658 // LS_cmpxchg:
2659 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2660 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2661 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2662 //
2663 // LS_xadd:
2664 // public int getAndAddInt( Object o, long offset, int delta)
2665 // public long getAndAddLong(Object o, long offset, long delta)
2666 //
2667 // LS_xchg:
2668 // int getAndSet(Object o, long offset, int newValue)
2669 // long getAndSet(Object o, long offset, long newValue)
2670 // Object getAndSet(Object o, long offset, Object newValue)
2671 //
2672 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2673 // This basic scheme here is the same as inline_unsafe_access, but
2674 // differs in enough details that combining them would make the code
2675 // overly confusing. (This is a true fact! I originally combined
2676 // them, but even I was confused by it!) As much code/comments as
2677 // possible are retained from inline_unsafe_access though to make
2678 // the correspondences clearer. - dl
2679
2680 if (callee()->is_static()) return false; // caller must have the capability!
2681
2682 #ifndef PRODUCT
2683 BasicType rtype;
2684 {
2685 ResourceMark rm;
2686 // Check the signatures.
2687 ciSignature* sig = callee()->signature();
2688 rtype = sig->return_type()->basic_type();
2689 if (kind == LS_xadd || kind == LS_xchg) {
2690 // Check the signatures.
2691 #ifdef ASSERT
2692 assert(rtype == type, "get and set must return the expected type");
2693 assert(sig->count() == 3, "get and set has 3 arguments");
2694 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2695 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2696 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2697 #endif // ASSERT
2698 } else if (kind == LS_cmpxchg) {
2699 // Check the signatures.
2700 #ifdef ASSERT
2701 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2702 assert(sig->count() == 4, "CAS has 4 arguments");
2703 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2704 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2705 #endif // ASSERT
2706 } else {
2707 ShouldNotReachHere();
2708 }
2709 }
2710 #endif //PRODUCT
2711
2712 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2713
2714 // Get arguments:
2715 Node* receiver = NULL;
2716 Node* base = NULL;
2717 Node* offset = NULL;
2718 Node* oldval = NULL;
2719 Node* newval = NULL;
2720 if (kind == LS_cmpxchg) {
2721 const bool two_slot_type = type2size[type] == 2;
2722 receiver = argument(0); // type: oop
2723 base = argument(1); // type: oop
2724 offset = argument(2); // type: long
2725 oldval = argument(4); // type: oop, int, or long
2726 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2727 } else if (kind == LS_xadd || kind == LS_xchg){
2728 receiver = argument(0); // type: oop
2729 base = argument(1); // type: oop
2730 offset = argument(2); // type: long
2731 oldval = NULL;
2732 newval = argument(4); // type: oop, int, or long
2733 }
2734
2735 // Null check receiver.
2736 receiver = null_check(receiver);
2737 if (stopped()) {
2738 return true;
2739 }
2740
2741 // Build field offset expression.
2742 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2743 // to be plain byte offsets, which are also the same as those accepted
2744 // by oopDesc::field_base.
2745 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2746 // 32-bit machines ignore the high half of long offsets
2747 offset = ConvL2X(offset);
2748 Node* adr = make_unsafe_address(base, offset);
2749 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2750
2751 // For CAS, unlike inline_unsafe_access, there seems no point in
2752 // trying to refine types. Just use the coarse types here.
2753 const Type *value_type = Type::get_const_basic_type(type);
2754 Compile::AliasType* alias_type = C->alias_type(adr_type);
2755 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2756
2757 if (kind == LS_xchg && type == T_OBJECT) {
2758 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2759 if (tjp != NULL) {
2760 value_type = tjp;
2761 }
2762 }
2763
2764 int alias_idx = C->get_alias_index(adr_type);
2765
2766 // Memory-model-wise, a LoadStore acts like a little synchronized
2767 // block, so needs barriers on each side. These don't translate
2768 // into actual barriers on most machines, but we still need rest of
2769 // compiler to respect ordering.
2770
2771 insert_mem_bar(Op_MemBarRelease);
2772 insert_mem_bar(Op_MemBarCPUOrder);
2773
2774 // 4984716: MemBars must be inserted before this
2775 // memory node in order to avoid a false
2776 // dependency which will confuse the scheduler.
2777 Node *mem = memory(alias_idx);
2778
2779 // For now, we handle only those cases that actually exist: ints,
2780 // longs, and Object. Adding others should be straightforward.
2781 Node* load_store;
2782 switch(type) {
2783 case T_INT:
2784 if (kind == LS_xadd) {
2785 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2786 } else if (kind == LS_xchg) {
2787 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2788 } else if (kind == LS_cmpxchg) {
2789 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2790 } else {
2791 ShouldNotReachHere();
2792 }
2793 break;
2794 case T_LONG:
2795 if (kind == LS_xadd) {
2796 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2797 } else if (kind == LS_xchg) {
2798 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2799 } else if (kind == LS_cmpxchg) {
2800 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2801 } else {
2802 ShouldNotReachHere();
2803 }
2804 break;
2805 case T_OBJECT:
2806 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2807 // could be delayed during Parse (for example, in adjust_map_after_if()).
2808 // Execute transformation here to avoid barrier generation in such case.
2809 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2810 newval = _gvn.makecon(TypePtr::NULL_PTR);
2811
2812 // Reference stores need a store barrier.
2813 if (kind == LS_xchg) {
2814 // If pre-barrier must execute before the oop store, old value will require do_load here.
2815 if (!can_move_pre_barrier()) {
2816 pre_barrier(true /* do_load*/,
2817 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2818 NULL /* pre_val*/,
2819 T_OBJECT);
2820 } // Else move pre_barrier to use load_store value, see below.
2821 } else if (kind == LS_cmpxchg) {
2822 // Same as for newval above:
2823 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2824 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2825 }
2826 // The only known value which might get overwritten is oldval.
2827 pre_barrier(false /* do_load */,
2828 control(), NULL, NULL, max_juint, NULL, NULL,
2829 oldval /* pre_val */,
2830 T_OBJECT);
2831 } else {
2832 ShouldNotReachHere();
2833 }
2834
2835 #ifdef _LP64
2836 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2837 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2838 if (kind == LS_xchg) {
2839 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2840 newval_enc, adr_type, value_type->make_narrowoop()));
2841 } else {
2842 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2843 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2844 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2845 newval_enc, oldval_enc));
2846 }
2847 } else
2848 #endif
2849 {
2850 if (kind == LS_xchg) {
2851 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2852 } else {
2853 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2854 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2855 }
2856 }
2857 if (kind == LS_cmpxchg) {
2858 // Emit the post barrier only when the actual store happened.
2859 // This makes sense to check only for compareAndSet that can fail to set the value.
2860 // CAS success path is marked more likely since we anticipate this is a performance
2861 // critical path, while CAS failure path can use the penalty for going through unlikely
2862 // path as backoff. Which is still better than doing a store barrier there.
2863 IdealKit ideal(this);
2864 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2865 sync_kit(ideal);
2866 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2867 ideal.sync_kit(this);
2868 } ideal.end_if();
2869 final_sync(ideal);
2870 } else {
2871 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2872 }
2873 break;
2874 default:
2875 fatal("unexpected type %d: %s", type, type2name(type));
2876 break;
2877 }
2878
2879 // SCMemProjNodes represent the memory state of a LoadStore. Their
2880 // main role is to prevent LoadStore nodes from being optimized away
2881 // when their results aren't used.
2882 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2883 set_memory(proj, alias_idx);
2884
2885 if (type == T_OBJECT && kind == LS_xchg) {
2886 #ifdef _LP64
2887 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2888 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2889 }
2890 #endif
2891 if (can_move_pre_barrier()) {
2892 // Don't need to load pre_val. The old value is returned by load_store.
2893 // The pre_barrier can execute after the xchg as long as no safepoint
2894 // gets inserted between them.
2895 pre_barrier(false /* do_load */,
2896 control(), NULL, NULL, max_juint, NULL, NULL,
2897 load_store /* pre_val */,
2898 T_OBJECT);
2899 }
2900 }
2901
2902 // Add the trailing membar surrounding the access
2903 insert_mem_bar(Op_MemBarCPUOrder);
2904 insert_mem_bar(Op_MemBarAcquire);
2905
2906 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2907 set_result(load_store);
2908 return true;
2909 }
2910
2911 //----------------------------inline_unsafe_ordered_store----------------------
2912 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2913 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2914 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2915 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2916 // This is another variant of inline_unsafe_access, differing in
2917 // that it always issues store-store ("release") barrier and ensures
2918 // store-atomicity (which only matters for "long").
2919
2920 if (callee()->is_static()) return false; // caller must have the capability!
2921
2922 #ifndef PRODUCT
2923 {
2924 ResourceMark rm;
2925 // Check the signatures.
2926 ciSignature* sig = callee()->signature();
2927 #ifdef ASSERT
2928 BasicType rtype = sig->return_type()->basic_type();
2929 assert(rtype == T_VOID, "must return void");
2930 assert(sig->count() == 3, "has 3 arguments");
2931 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2932 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2933 #endif // ASSERT
2934 }
2935 #endif //PRODUCT
2936
2937 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2938
2939 // Get arguments:
2940 Node* receiver = argument(0); // type: oop
2941 Node* base = argument(1); // type: oop
2942 Node* offset = argument(2); // type: long
2943 Node* val = argument(4); // type: oop, int, or long
2944
2945 // Null check receiver.
2946 receiver = null_check(receiver);
2947 if (stopped()) {
2948 return true;
2949 }
2950
2951 // Build field offset expression.
2952 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2953 // 32-bit machines ignore the high half of long offsets
2954 offset = ConvL2X(offset);
2955 Node* adr = make_unsafe_address(base, offset);
2956 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2957 const Type *value_type = Type::get_const_basic_type(type);
2958 Compile::AliasType* alias_type = C->alias_type(adr_type);
2959
2960 insert_mem_bar(Op_MemBarRelease);
2961 insert_mem_bar(Op_MemBarCPUOrder);
2962 // Ensure that the store is atomic for longs:
2963 const bool require_atomic_access = true;
2964 Node* store;
2965 if (type == T_OBJECT) // reference stores need a store barrier.
2966 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2967 else {
2968 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2969 }
2970 insert_mem_bar(Op_MemBarCPUOrder);
2971 return true;
2972 }
2973
2974 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2975 // Regardless of form, don't allow previous ld/st to move down,
2976 // then issue acquire, release, or volatile mem_bar.
2977 insert_mem_bar(Op_MemBarCPUOrder);
2978 switch(id) {
2979 case vmIntrinsics::_loadFence:
2980 insert_mem_bar(Op_LoadFence);
2981 return true;
2982 case vmIntrinsics::_storeFence:
2983 insert_mem_bar(Op_StoreFence);
2984 return true;
2985 case vmIntrinsics::_fullFence:
2986 insert_mem_bar(Op_MemBarVolatile);
2987 return true;
2988 default:
2989 fatal_unexpected_iid(id);
2990 return false;
2991 }
2992 }
2993
2994 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2995 if (!kls->is_Con()) {
2996 return true;
2997 }
2998 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2999 if (klsptr == NULL) {
3000 return true;
3001 }
3002 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3003 // don't need a guard for a klass that is already initialized
3004 return !ik->is_initialized();
3005 }
3006
3007 //----------------------------inline_unsafe_allocate---------------------------
3008 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3009 bool LibraryCallKit::inline_unsafe_allocate() {
3010 if (callee()->is_static()) return false; // caller must have the capability!
3011
3012 null_check_receiver(); // null-check, then ignore
3013 Node* cls = null_check(argument(1));
3014 if (stopped()) return true;
3015
3016 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3017 kls = null_check(kls);
3018 if (stopped()) return true; // argument was like int.class
3019
3020 Node* test = NULL;
3021 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3022 // Note: The argument might still be an illegal value like
3023 // Serializable.class or Object[].class. The runtime will handle it.
3024 // But we must make an explicit check for initialization.
3025 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3026 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3027 // can generate code to load it as unsigned byte.
3028 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3029 Node* bits = intcon(InstanceKlass::fully_initialized);
3030 test = _gvn.transform(new SubINode(inst, bits));
3031 // The 'test' is non-zero if we need to take a slow path.
3032 }
3033
3034 Node* obj = new_instance(kls, test);
3035 set_result(obj);
3036 return true;
3037 }
3038
3039 #ifdef TRACE_HAVE_INTRINSICS
3040 /*
3041 * oop -> myklass
3042 * myklass->trace_id |= USED
3043 * return myklass->trace_id & ~0x3
3044 */
3045 bool LibraryCallKit::inline_native_classID() {
3046 null_check_receiver(); // null-check, then ignore
3047 Node* cls = null_check(argument(1), T_OBJECT);
3048 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3049 kls = null_check(kls, T_OBJECT);
3050 ByteSize offset = TRACE_ID_OFFSET;
3051 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3052 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3053 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3054 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3055 Node* clsused = longcon(0x01l); // set the class bit
3056 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3057
3058 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3059 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3060 set_result(andl);
3061 return true;
3062 }
3063
3064 bool LibraryCallKit::inline_native_threadID() {
3065 Node* tls_ptr = NULL;
3066 Node* cur_thr = generate_current_thread(tls_ptr);
3067 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3068 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3069 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3070
3071 Node* threadid = NULL;
3072 size_t thread_id_size = OSThread::thread_id_size();
3073 if (thread_id_size == (size_t) BytesPerLong) {
3074 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3075 } else if (thread_id_size == (size_t) BytesPerInt) {
3076 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3077 } else {
3078 ShouldNotReachHere();
3079 }
3080 set_result(threadid);
3081 return true;
3082 }
3083 #endif
3084
3085 //------------------------inline_native_time_funcs--------------
3086 // inline code for System.currentTimeMillis() and System.nanoTime()
3087 // these have the same type and signature
3088 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3089 const TypeFunc* tf = OptoRuntime::void_long_Type();
3090 const TypePtr* no_memory_effects = NULL;
3091 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3092 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3093 #ifdef ASSERT
3094 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3095 assert(value_top == top(), "second value must be top");
3096 #endif
3097 set_result(value);
3098 return true;
3099 }
3100
3101 //------------------------inline_native_currentThread------------------
3102 bool LibraryCallKit::inline_native_currentThread() {
3103 Node* junk = NULL;
3104 set_result(generate_current_thread(junk));
3105 return true;
3106 }
3107
3108 //------------------------inline_native_isInterrupted------------------
3109 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3110 bool LibraryCallKit::inline_native_isInterrupted() {
3111 // Add a fast path to t.isInterrupted(clear_int):
3112 // (t == Thread.current() &&
3113 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3114 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3115 // So, in the common case that the interrupt bit is false,
3116 // we avoid making a call into the VM. Even if the interrupt bit
3117 // is true, if the clear_int argument is false, we avoid the VM call.
3118 // However, if the receiver is not currentThread, we must call the VM,
3119 // because there must be some locking done around the operation.
3120
3121 // We only go to the fast case code if we pass two guards.
3122 // Paths which do not pass are accumulated in the slow_region.
3123
3124 enum {
3125 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3126 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3127 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3128 PATH_LIMIT
3129 };
3130
3131 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3132 // out of the function.
3133 insert_mem_bar(Op_MemBarCPUOrder);
3134
3135 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3136 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3137
3138 RegionNode* slow_region = new RegionNode(1);
3139 record_for_igvn(slow_region);
3140
3141 // (a) Receiving thread must be the current thread.
3142 Node* rec_thr = argument(0);
3143 Node* tls_ptr = NULL;
3144 Node* cur_thr = generate_current_thread(tls_ptr);
3145 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3146 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3147
3148 generate_slow_guard(bol_thr, slow_region);
3149
3150 // (b) Interrupt bit on TLS must be false.
3151 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3152 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3153 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3154
3155 // Set the control input on the field _interrupted read to prevent it floating up.
3156 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3157 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3158 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3159
3160 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3161
3162 // First fast path: if (!TLS._interrupted) return false;
3163 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3164 result_rgn->init_req(no_int_result_path, false_bit);
3165 result_val->init_req(no_int_result_path, intcon(0));
3166
3167 // drop through to next case
3168 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3169
3170 #ifndef TARGET_OS_FAMILY_windows
3171 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3172 Node* clr_arg = argument(1);
3173 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3174 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3175 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3176
3177 // Second fast path: ... else if (!clear_int) return true;
3178 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3179 result_rgn->init_req(no_clear_result_path, false_arg);
3180 result_val->init_req(no_clear_result_path, intcon(1));
3181
3182 // drop through to next case
3183 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3184 #else
3185 // To return true on Windows you must read the _interrupted field
3186 // and check the the event state i.e. take the slow path.
3187 #endif // TARGET_OS_FAMILY_windows
3188
3189 // (d) Otherwise, go to the slow path.
3190 slow_region->add_req(control());
3191 set_control( _gvn.transform(slow_region));
3192
3193 if (stopped()) {
3194 // There is no slow path.
3195 result_rgn->init_req(slow_result_path, top());
3196 result_val->init_req(slow_result_path, top());
3197 } else {
3198 // non-virtual because it is a private non-static
3199 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3200
3201 Node* slow_val = set_results_for_java_call(slow_call);
3202 // this->control() comes from set_results_for_java_call
3203
3204 Node* fast_io = slow_call->in(TypeFunc::I_O);
3205 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3206
3207 // These two phis are pre-filled with copies of of the fast IO and Memory
3208 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3209 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3210
3211 result_rgn->init_req(slow_result_path, control());
3212 result_io ->init_req(slow_result_path, i_o());
3213 result_mem->init_req(slow_result_path, reset_memory());
3214 result_val->init_req(slow_result_path, slow_val);
3215
3216 set_all_memory(_gvn.transform(result_mem));
3217 set_i_o( _gvn.transform(result_io));
3218 }
3219
3220 C->set_has_split_ifs(true); // Has chance for split-if optimization
3221 set_result(result_rgn, result_val);
3222 return true;
3223 }
3224
3225 //---------------------------load_mirror_from_klass----------------------------
3226 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3227 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3228 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3229 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3230 }
3231
3232 //-----------------------load_klass_from_mirror_common-------------------------
3233 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3234 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3235 // and branch to the given path on the region.
3236 // If never_see_null, take an uncommon trap on null, so we can optimistically
3237 // compile for the non-null case.
3238 // If the region is NULL, force never_see_null = true.
3239 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3240 bool never_see_null,
3241 RegionNode* region,
3242 int null_path,
3243 int offset) {
3244 if (region == NULL) never_see_null = true;
3245 Node* p = basic_plus_adr(mirror, offset);
3246 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3247 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3248 Node* null_ctl = top();
3249 kls = null_check_oop(kls, &null_ctl, never_see_null);
3250 if (region != NULL) {
3251 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3252 region->init_req(null_path, null_ctl);
3253 } else {
3254 assert(null_ctl == top(), "no loose ends");
3255 }
3256 return kls;
3257 }
3258
3259 //--------------------(inline_native_Class_query helpers)---------------------
3260 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3261 // Fall through if (mods & mask) == bits, take the guard otherwise.
3262 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3263 // Branch around if the given klass has the given modifier bit set.
3264 // Like generate_guard, adds a new path onto the region.
3265 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3266 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3267 Node* mask = intcon(modifier_mask);
3268 Node* bits = intcon(modifier_bits);
3269 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3270 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3271 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3272 return generate_fair_guard(bol, region);
3273 }
3274 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3275 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3276 }
3277
3278 //-------------------------inline_native_Class_query-------------------
3279 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3280 const Type* return_type = TypeInt::BOOL;
3281 Node* prim_return_value = top(); // what happens if it's a primitive class?
3282 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3283 bool expect_prim = false; // most of these guys expect to work on refs
3284
3285 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3286
3287 Node* mirror = argument(0);
3288 Node* obj = top();
3289
3290 switch (id) {
3291 case vmIntrinsics::_isInstance:
3292 // nothing is an instance of a primitive type
3293 prim_return_value = intcon(0);
3294 obj = argument(1);
3295 break;
3296 case vmIntrinsics::_getModifiers:
3297 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3298 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3299 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3300 break;
3301 case vmIntrinsics::_isInterface:
3302 prim_return_value = intcon(0);
3303 break;
3304 case vmIntrinsics::_isArray:
3305 prim_return_value = intcon(0);
3306 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3307 break;
3308 case vmIntrinsics::_isPrimitive:
3309 prim_return_value = intcon(1);
3310 expect_prim = true; // obviously
3311 break;
3312 case vmIntrinsics::_getSuperclass:
3313 prim_return_value = null();
3314 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3315 break;
3316 case vmIntrinsics::_getClassAccessFlags:
3317 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3318 return_type = TypeInt::INT; // not bool! 6297094
3319 break;
3320 default:
3321 fatal_unexpected_iid(id);
3322 break;
3323 }
3324
3325 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3326 if (mirror_con == NULL) return false; // cannot happen?
3327
3328 #ifndef PRODUCT
3329 if (C->print_intrinsics() || C->print_inlining()) {
3330 ciType* k = mirror_con->java_mirror_type();
3331 if (k) {
3332 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3333 k->print_name();
3334 tty->cr();
3335 }
3336 }
3337 #endif
3338
3339 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3340 RegionNode* region = new RegionNode(PATH_LIMIT);
3341 record_for_igvn(region);
3342 PhiNode* phi = new PhiNode(region, return_type);
3343
3344 // The mirror will never be null of Reflection.getClassAccessFlags, however
3345 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3346 // if it is. See bug 4774291.
3347
3348 // For Reflection.getClassAccessFlags(), the null check occurs in
3349 // the wrong place; see inline_unsafe_access(), above, for a similar
3350 // situation.
3351 mirror = null_check(mirror);
3352 // If mirror or obj is dead, only null-path is taken.
3353 if (stopped()) return true;
3354
3355 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3356
3357 // Now load the mirror's klass metaobject, and null-check it.
3358 // Side-effects region with the control path if the klass is null.
3359 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3360 // If kls is null, we have a primitive mirror.
3361 phi->init_req(_prim_path, prim_return_value);
3362 if (stopped()) { set_result(region, phi); return true; }
3363 bool safe_for_replace = (region->in(_prim_path) == top());
3364
3365 Node* p; // handy temp
3366 Node* null_ctl;
3367
3368 // Now that we have the non-null klass, we can perform the real query.
3369 // For constant classes, the query will constant-fold in LoadNode::Value.
3370 Node* query_value = top();
3371 switch (id) {
3372 case vmIntrinsics::_isInstance:
3373 // nothing is an instance of a primitive type
3374 query_value = gen_instanceof(obj, kls, safe_for_replace);
3375 break;
3376
3377 case vmIntrinsics::_getModifiers:
3378 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3379 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3380 break;
3381
3382 case vmIntrinsics::_isInterface:
3383 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3384 if (generate_interface_guard(kls, region) != NULL)
3385 // A guard was added. If the guard is taken, it was an interface.
3386 phi->add_req(intcon(1));
3387 // If we fall through, it's a plain class.
3388 query_value = intcon(0);
3389 break;
3390
3391 case vmIntrinsics::_isArray:
3392 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3393 if (generate_array_guard(kls, region) != NULL)
3394 // A guard was added. If the guard is taken, it was an array.
3395 phi->add_req(intcon(1));
3396 // If we fall through, it's a plain class.
3397 query_value = intcon(0);
3398 break;
3399
3400 case vmIntrinsics::_isPrimitive:
3401 query_value = intcon(0); // "normal" path produces false
3402 break;
3403
3404 case vmIntrinsics::_getSuperclass:
3405 // The rules here are somewhat unfortunate, but we can still do better
3406 // with random logic than with a JNI call.
3407 // Interfaces store null or Object as _super, but must report null.
3408 // Arrays store an intermediate super as _super, but must report Object.
3409 // Other types can report the actual _super.
3410 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3411 if (generate_interface_guard(kls, region) != NULL)
3412 // A guard was added. If the guard is taken, it was an interface.
3413 phi->add_req(null());
3414 if (generate_array_guard(kls, region) != NULL)
3415 // A guard was added. If the guard is taken, it was an array.
3416 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3417 // If we fall through, it's a plain class. Get its _super.
3418 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3419 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3420 null_ctl = top();
3421 kls = null_check_oop(kls, &null_ctl);
3422 if (null_ctl != top()) {
3423 // If the guard is taken, Object.superClass is null (both klass and mirror).
3424 region->add_req(null_ctl);
3425 phi ->add_req(null());
3426 }
3427 if (!stopped()) {
3428 query_value = load_mirror_from_klass(kls);
3429 }
3430 break;
3431
3432 case vmIntrinsics::_getClassAccessFlags:
3433 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3434 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3435 break;
3436
3437 default:
3438 fatal_unexpected_iid(id);
3439 break;
3440 }
3441
3442 // Fall-through is the normal case of a query to a real class.
3443 phi->init_req(1, query_value);
3444 region->init_req(1, control());
3445
3446 C->set_has_split_ifs(true); // Has chance for split-if optimization
3447 set_result(region, phi);
3448 return true;
3449 }
3450
3451 //-------------------------inline_Class_cast-------------------
3452 bool LibraryCallKit::inline_Class_cast() {
3453 Node* mirror = argument(0); // Class
3454 Node* obj = argument(1);
3455 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3456 if (mirror_con == NULL) {
3457 return false; // dead path (mirror->is_top()).
3458 }
3459 if (obj == NULL || obj->is_top()) {
3460 return false; // dead path
3461 }
3462 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3463
3464 // First, see if Class.cast() can be folded statically.
3465 // java_mirror_type() returns non-null for compile-time Class constants.
3466 ciType* tm = mirror_con->java_mirror_type();
3467 if (tm != NULL && tm->is_klass() &&
3468 tp != NULL && tp->klass() != NULL) {
3469 if (!tp->klass()->is_loaded()) {
3470 // Don't use intrinsic when class is not loaded.
3471 return false;
3472 } else {
3473 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3474 if (static_res == Compile::SSC_always_true) {
3475 // isInstance() is true - fold the code.
3476 set_result(obj);
3477 return true;
3478 } else if (static_res == Compile::SSC_always_false) {
3479 // Don't use intrinsic, have to throw ClassCastException.
3480 // If the reference is null, the non-intrinsic bytecode will
3481 // be optimized appropriately.
3482 return false;
3483 }
3484 }
3485 }
3486
3487 // Bailout intrinsic and do normal inlining if exception path is frequent.
3488 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3489 return false;
3490 }
3491
3492 // Generate dynamic checks.
3493 // Class.cast() is java implementation of _checkcast bytecode.
3494 // Do checkcast (Parse::do_checkcast()) optimizations here.
3495
3496 mirror = null_check(mirror);
3497 // If mirror is dead, only null-path is taken.
3498 if (stopped()) {
3499 return true;
3500 }
3501
3502 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3503 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3504 RegionNode* region = new RegionNode(PATH_LIMIT);
3505 record_for_igvn(region);
3506
3507 // Now load the mirror's klass metaobject, and null-check it.
3508 // If kls is null, we have a primitive mirror and
3509 // nothing is an instance of a primitive type.
3510 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3511
3512 Node* res = top();
3513 if (!stopped()) {
3514 Node* bad_type_ctrl = top();
3515 // Do checkcast optimizations.
3516 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3517 region->init_req(_bad_type_path, bad_type_ctrl);
3518 }
3519 if (region->in(_prim_path) != top() ||
3520 region->in(_bad_type_path) != top()) {
3521 // Let Interpreter throw ClassCastException.
3522 PreserveJVMState pjvms(this);
3523 set_control(_gvn.transform(region));
3524 uncommon_trap(Deoptimization::Reason_intrinsic,
3525 Deoptimization::Action_maybe_recompile);
3526 }
3527 if (!stopped()) {
3528 set_result(res);
3529 }
3530 return true;
3531 }
3532
3533
3534 //--------------------------inline_native_subtype_check------------------------
3535 // This intrinsic takes the JNI calls out of the heart of
3536 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3537 bool LibraryCallKit::inline_native_subtype_check() {
3538 // Pull both arguments off the stack.
3539 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3540 args[0] = argument(0);
3541 args[1] = argument(1);
3542 Node* klasses[2]; // corresponding Klasses: superk, subk
3543 klasses[0] = klasses[1] = top();
3544
3545 enum {
3546 // A full decision tree on {superc is prim, subc is prim}:
3547 _prim_0_path = 1, // {P,N} => false
3548 // {P,P} & superc!=subc => false
3549 _prim_same_path, // {P,P} & superc==subc => true
3550 _prim_1_path, // {N,P} => false
3551 _ref_subtype_path, // {N,N} & subtype check wins => true
3552 _both_ref_path, // {N,N} & subtype check loses => false
3553 PATH_LIMIT
3554 };
3555
3556 RegionNode* region = new RegionNode(PATH_LIMIT);
3557 Node* phi = new PhiNode(region, TypeInt::BOOL);
3558 record_for_igvn(region);
3559
3560 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3561 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3562 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3563
3564 // First null-check both mirrors and load each mirror's klass metaobject.
3565 int which_arg;
3566 for (which_arg = 0; which_arg <= 1; which_arg++) {
3567 Node* arg = args[which_arg];
3568 arg = null_check(arg);
3569 if (stopped()) break;
3570 args[which_arg] = arg;
3571
3572 Node* p = basic_plus_adr(arg, class_klass_offset);
3573 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3574 klasses[which_arg] = _gvn.transform(kls);
3575 }
3576
3577 // Having loaded both klasses, test each for null.
3578 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3579 for (which_arg = 0; which_arg <= 1; which_arg++) {
3580 Node* kls = klasses[which_arg];
3581 Node* null_ctl = top();
3582 kls = null_check_oop(kls, &null_ctl, never_see_null);
3583 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3584 region->init_req(prim_path, null_ctl);
3585 if (stopped()) break;
3586 klasses[which_arg] = kls;
3587 }
3588
3589 if (!stopped()) {
3590 // now we have two reference types, in klasses[0..1]
3591 Node* subk = klasses[1]; // the argument to isAssignableFrom
3592 Node* superk = klasses[0]; // the receiver
3593 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3594 // now we have a successful reference subtype check
3595 region->set_req(_ref_subtype_path, control());
3596 }
3597
3598 // If both operands are primitive (both klasses null), then
3599 // we must return true when they are identical primitives.
3600 // It is convenient to test this after the first null klass check.
3601 set_control(region->in(_prim_0_path)); // go back to first null check
3602 if (!stopped()) {
3603 // Since superc is primitive, make a guard for the superc==subc case.
3604 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3605 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3606 generate_guard(bol_eq, region, PROB_FAIR);
3607 if (region->req() == PATH_LIMIT+1) {
3608 // A guard was added. If the added guard is taken, superc==subc.
3609 region->swap_edges(PATH_LIMIT, _prim_same_path);
3610 region->del_req(PATH_LIMIT);
3611 }
3612 region->set_req(_prim_0_path, control()); // Not equal after all.
3613 }
3614
3615 // these are the only paths that produce 'true':
3616 phi->set_req(_prim_same_path, intcon(1));
3617 phi->set_req(_ref_subtype_path, intcon(1));
3618
3619 // pull together the cases:
3620 assert(region->req() == PATH_LIMIT, "sane region");
3621 for (uint i = 1; i < region->req(); i++) {
3622 Node* ctl = region->in(i);
3623 if (ctl == NULL || ctl == top()) {
3624 region->set_req(i, top());
3625 phi ->set_req(i, top());
3626 } else if (phi->in(i) == NULL) {
3627 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3628 }
3629 }
3630
3631 set_control(_gvn.transform(region));
3632 set_result(_gvn.transform(phi));
3633 return true;
3634 }
3635
3636 //---------------------generate_array_guard_common------------------------
3637 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3638 bool obj_array, bool not_array) {
3639
3640 if (stopped()) {
3641 return NULL;
3642 }
3643
3644 // If obj_array/non_array==false/false:
3645 // Branch around if the given klass is in fact an array (either obj or prim).
3646 // If obj_array/non_array==false/true:
3647 // Branch around if the given klass is not an array klass of any kind.
3648 // If obj_array/non_array==true/true:
3649 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3650 // If obj_array/non_array==true/false:
3651 // Branch around if the kls is an oop array (Object[] or subtype)
3652 //
3653 // Like generate_guard, adds a new path onto the region.
3654 jint layout_con = 0;
3655 Node* layout_val = get_layout_helper(kls, layout_con);
3656 if (layout_val == NULL) {
3657 bool query = (obj_array
3658 ? Klass::layout_helper_is_objArray(layout_con)
3659 : Klass::layout_helper_is_array(layout_con));
3660 if (query == not_array) {
3661 return NULL; // never a branch
3662 } else { // always a branch
3663 Node* always_branch = control();
3664 if (region != NULL)
3665 region->add_req(always_branch);
3666 set_control(top());
3667 return always_branch;
3668 }
3669 }
3670 // Now test the correct condition.
3671 jint nval = (obj_array
3672 ? ((jint)Klass::_lh_array_tag_type_value
3673 << Klass::_lh_array_tag_shift)
3674 : Klass::_lh_neutral_value);
3675 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3676 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3677 // invert the test if we are looking for a non-array
3678 if (not_array) btest = BoolTest(btest).negate();
3679 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3680 return generate_fair_guard(bol, region);
3681 }
3682
3683
3684 //-----------------------inline_native_newArray--------------------------
3685 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3686 bool LibraryCallKit::inline_native_newArray() {
3687 Node* mirror = argument(0);
3688 Node* count_val = argument(1);
3689
3690 mirror = null_check(mirror);
3691 // If mirror or obj is dead, only null-path is taken.
3692 if (stopped()) return true;
3693
3694 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3695 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3696 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3697 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3698 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3699
3700 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3701 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3702 result_reg, _slow_path);
3703 Node* normal_ctl = control();
3704 Node* no_array_ctl = result_reg->in(_slow_path);
3705
3706 // Generate code for the slow case. We make a call to newArray().
3707 set_control(no_array_ctl);
3708 if (!stopped()) {
3709 // Either the input type is void.class, or else the
3710 // array klass has not yet been cached. Either the
3711 // ensuing call will throw an exception, or else it
3712 // will cache the array klass for next time.
3713 PreserveJVMState pjvms(this);
3714 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3715 Node* slow_result = set_results_for_java_call(slow_call);
3716 // this->control() comes from set_results_for_java_call
3717 result_reg->set_req(_slow_path, control());
3718 result_val->set_req(_slow_path, slow_result);
3719 result_io ->set_req(_slow_path, i_o());
3720 result_mem->set_req(_slow_path, reset_memory());
3721 }
3722
3723 set_control(normal_ctl);
3724 if (!stopped()) {
3725 // Normal case: The array type has been cached in the java.lang.Class.
3726 // The following call works fine even if the array type is polymorphic.
3727 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3728 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3729 result_reg->init_req(_normal_path, control());
3730 result_val->init_req(_normal_path, obj);
3731 result_io ->init_req(_normal_path, i_o());
3732 result_mem->init_req(_normal_path, reset_memory());
3733 }
3734
3735 // Return the combined state.
3736 set_i_o( _gvn.transform(result_io) );
3737 set_all_memory( _gvn.transform(result_mem));
3738
3739 C->set_has_split_ifs(true); // Has chance for split-if optimization
3740 set_result(result_reg, result_val);
3741 return true;
3742 }
3743
3744 //----------------------inline_native_getLength--------------------------
3745 // public static native int java.lang.reflect.Array.getLength(Object array);
3746 bool LibraryCallKit::inline_native_getLength() {
3747 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3748
3749 Node* array = null_check(argument(0));
3750 // If array is dead, only null-path is taken.
3751 if (stopped()) return true;
3752
3753 // Deoptimize if it is a non-array.
3754 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3755
3756 if (non_array != NULL) {
3757 PreserveJVMState pjvms(this);
3758 set_control(non_array);
3759 uncommon_trap(Deoptimization::Reason_intrinsic,
3760 Deoptimization::Action_maybe_recompile);
3761 }
3762
3763 // If control is dead, only non-array-path is taken.
3764 if (stopped()) return true;
3765
3766 // The works fine even if the array type is polymorphic.
3767 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3768 Node* result = load_array_length(array);
3769
3770 C->set_has_split_ifs(true); // Has chance for split-if optimization
3771 set_result(result);
3772 return true;
3773 }
3774
3775 //------------------------inline_array_copyOf----------------------------
3776 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3777 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3778 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3779 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3780
3781 // Get the arguments.
3782 Node* original = argument(0);
3783 Node* start = is_copyOfRange? argument(1): intcon(0);
3784 Node* end = is_copyOfRange? argument(2): argument(1);
3785 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3786
3787 Node* newcopy;
3788
3789 // Set the original stack and the reexecute bit for the interpreter to reexecute
3790 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3791 { PreserveReexecuteState preexecs(this);
3792 jvms()->set_should_reexecute(true);
3793
3794 array_type_mirror = null_check(array_type_mirror);
3795 original = null_check(original);
3796
3797 // Check if a null path was taken unconditionally.
3798 if (stopped()) return true;
3799
3800 Node* orig_length = load_array_length(original);
3801
3802 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3803 klass_node = null_check(klass_node);
3804
3805 RegionNode* bailout = new RegionNode(1);
3806 record_for_igvn(bailout);
3807
3808 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3809 // Bail out if that is so.
3810 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3811 if (not_objArray != NULL) {
3812 // Improve the klass node's type from the new optimistic assumption:
3813 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3814 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3815 Node* cast = new CastPPNode(klass_node, akls);
3816 cast->init_req(0, control());
3817 klass_node = _gvn.transform(cast);
3818 }
3819
3820 // Bail out if either start or end is negative.
3821 generate_negative_guard(start, bailout, &start);
3822 generate_negative_guard(end, bailout, &end);
3823
3824 Node* length = end;
3825 if (_gvn.type(start) != TypeInt::ZERO) {
3826 length = _gvn.transform(new SubINode(end, start));
3827 }
3828
3829 // Bail out if length is negative.
3830 // Without this the new_array would throw
3831 // NegativeArraySizeException but IllegalArgumentException is what
3832 // should be thrown
3833 generate_negative_guard(length, bailout, &length);
3834
3835 if (bailout->req() > 1) {
3836 PreserveJVMState pjvms(this);
3837 set_control(_gvn.transform(bailout));
3838 uncommon_trap(Deoptimization::Reason_intrinsic,
3839 Deoptimization::Action_maybe_recompile);
3840 }
3841
3842 if (!stopped()) {
3843 // How many elements will we copy from the original?
3844 // The answer is MinI(orig_length - start, length).
3845 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3846 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3847
3848 // Generate a direct call to the right arraycopy function(s).
3849 // We know the copy is disjoint but we might not know if the
3850 // oop stores need checking.
3851 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3852 // This will fail a store-check if x contains any non-nulls.
3853
3854 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3855 // loads/stores but it is legal only if we're sure the
3856 // Arrays.copyOf would succeed. So we need all input arguments
3857 // to the copyOf to be validated, including that the copy to the
3858 // new array won't trigger an ArrayStoreException. That subtype
3859 // check can be optimized if we know something on the type of
3860 // the input array from type speculation.
3861 if (_gvn.type(klass_node)->singleton()) {
3862 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3863 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3864
3865 int test = C->static_subtype_check(superk, subk);
3866 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3867 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3868 if (t_original->speculative_type() != NULL) {
3869 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3870 }
3871 }
3872 }
3873
3874 bool validated = false;
3875 // Reason_class_check rather than Reason_intrinsic because we
3876 // want to intrinsify even if this traps.
3877 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3878 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3879 klass_node);
3880
3881 if (not_subtype_ctrl != top()) {
3882 PreserveJVMState pjvms(this);
3883 set_control(not_subtype_ctrl);
3884 uncommon_trap(Deoptimization::Reason_class_check,
3885 Deoptimization::Action_make_not_entrant);
3886 assert(stopped(), "Should be stopped");
3887 }
3888 validated = true;
3889 }
3890
3891 if (!stopped()) {
3892 newcopy = new_array(klass_node, length, 0); // no arguments to push
3893
3894 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3895 load_object_klass(original), klass_node);
3896 if (!is_copyOfRange) {
3897 ac->set_copyof(validated);
3898 } else {
3899 ac->set_copyofrange(validated);
3900 }
3901 Node* n = _gvn.transform(ac);
3902 if (n == ac) {
3903 ac->connect_outputs(this);
3904 } else {
3905 assert(validated, "shouldn't transform if all arguments not validated");
3906 set_all_memory(n);
3907 }
3908 }
3909 }
3910 } // original reexecute is set back here
3911
3912 C->set_has_split_ifs(true); // Has chance for split-if optimization
3913 if (!stopped()) {
3914 set_result(newcopy);
3915 }
3916 return true;
3917 }
3918
3919
3920 //----------------------generate_virtual_guard---------------------------
3921 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3922 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3923 RegionNode* slow_region) {
3924 ciMethod* method = callee();
3925 int vtable_index = method->vtable_index();
3926 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3927 "bad index %d", vtable_index);
3928 // Get the Method* out of the appropriate vtable entry.
3929 int entry_offset = (InstanceKlass::vtable_start_offset() +
3930 vtable_index*vtableEntry::size()) * wordSize +
3931 vtableEntry::method_offset_in_bytes();
3932 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3933 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3934
3935 // Compare the target method with the expected method (e.g., Object.hashCode).
3936 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3937
3938 Node* native_call = makecon(native_call_addr);
3939 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3940 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3941
3942 return generate_slow_guard(test_native, slow_region);
3943 }
3944
3945 //-----------------------generate_method_call----------------------------
3946 // Use generate_method_call to make a slow-call to the real
3947 // method if the fast path fails. An alternative would be to
3948 // use a stub like OptoRuntime::slow_arraycopy_Java.
3949 // This only works for expanding the current library call,
3950 // not another intrinsic. (E.g., don't use this for making an
3951 // arraycopy call inside of the copyOf intrinsic.)
3952 CallJavaNode*
3953 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3954 // When compiling the intrinsic method itself, do not use this technique.
3955 guarantee(callee() != C->method(), "cannot make slow-call to self");
3956
3957 ciMethod* method = callee();
3958 // ensure the JVMS we have will be correct for this call
3959 guarantee(method_id == method->intrinsic_id(), "must match");
3960
3961 const TypeFunc* tf = TypeFunc::make(method);
3962 CallJavaNode* slow_call;
3963 if (is_static) {
3964 assert(!is_virtual, "");
3965 slow_call = new CallStaticJavaNode(C, tf,
3966 SharedRuntime::get_resolve_static_call_stub(),
3967 method, bci());
3968 } else if (is_virtual) {
3969 null_check_receiver();
3970 int vtable_index = Method::invalid_vtable_index;
3971 if (UseInlineCaches) {
3972 // Suppress the vtable call
3973 } else {
3974 // hashCode and clone are not a miranda methods,
3975 // so the vtable index is fixed.
3976 // No need to use the linkResolver to get it.
3977 vtable_index = method->vtable_index();
3978 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3979 "bad index %d", vtable_index);
3980 }
3981 slow_call = new CallDynamicJavaNode(tf,
3982 SharedRuntime::get_resolve_virtual_call_stub(),
3983 method, vtable_index, bci());
3984 } else { // neither virtual nor static: opt_virtual
3985 null_check_receiver();
3986 slow_call = new CallStaticJavaNode(C, tf,
3987 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3988 method, bci());
3989 slow_call->set_optimized_virtual(true);
3990 }
3991 set_arguments_for_java_call(slow_call);
3992 set_edges_for_java_call(slow_call);
3993 return slow_call;
3994 }
3995
3996
3997 /**
3998 * Build special case code for calls to hashCode on an object. This call may
3999 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4000 * slightly different code.
4001 */
4002 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4003 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4004 assert(!(is_virtual && is_static), "either virtual, special, or static");
4005
4006 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4007
4008 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4009 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4010 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4011 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4012 Node* obj = NULL;
4013 if (!is_static) {
4014 // Check for hashing null object
4015 obj = null_check_receiver();
4016 if (stopped()) return true; // unconditionally null
4017 result_reg->init_req(_null_path, top());
4018 result_val->init_req(_null_path, top());
4019 } else {
4020 // Do a null check, and return zero if null.
4021 // System.identityHashCode(null) == 0
4022 obj = argument(0);
4023 Node* null_ctl = top();
4024 obj = null_check_oop(obj, &null_ctl);
4025 result_reg->init_req(_null_path, null_ctl);
4026 result_val->init_req(_null_path, _gvn.intcon(0));
4027 }
4028
4029 // Unconditionally null? Then return right away.
4030 if (stopped()) {
4031 set_control( result_reg->in(_null_path));
4032 if (!stopped())
4033 set_result(result_val->in(_null_path));
4034 return true;
4035 }
4036
4037 // We only go to the fast case code if we pass a number of guards. The
4038 // paths which do not pass are accumulated in the slow_region.
4039 RegionNode* slow_region = new RegionNode(1);
4040 record_for_igvn(slow_region);
4041
4042 // If this is a virtual call, we generate a funny guard. We pull out
4043 // the vtable entry corresponding to hashCode() from the target object.
4044 // If the target method which we are calling happens to be the native
4045 // Object hashCode() method, we pass the guard. We do not need this
4046 // guard for non-virtual calls -- the caller is known to be the native
4047 // Object hashCode().
4048 if (is_virtual) {
4049 // After null check, get the object's klass.
4050 Node* obj_klass = load_object_klass(obj);
4051 generate_virtual_guard(obj_klass, slow_region);
4052 }
4053
4054 // Get the header out of the object, use LoadMarkNode when available
4055 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4056 // The control of the load must be NULL. Otherwise, the load can move before
4057 // the null check after castPP removal.
4058 Node* no_ctrl = NULL;
4059 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4060
4061 // Test the header to see if it is unlocked.
4062 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4063 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4064 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4065 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4066 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4067
4068 generate_slow_guard(test_unlocked, slow_region);
4069
4070 // Get the hash value and check to see that it has been properly assigned.
4071 // We depend on hash_mask being at most 32 bits and avoid the use of
4072 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4073 // vm: see markOop.hpp.
4074 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4075 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4076 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4077 // This hack lets the hash bits live anywhere in the mark object now, as long
4078 // as the shift drops the relevant bits into the low 32 bits. Note that
4079 // Java spec says that HashCode is an int so there's no point in capturing
4080 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4081 hshifted_header = ConvX2I(hshifted_header);
4082 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4083
4084 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4085 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4086 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4087
4088 generate_slow_guard(test_assigned, slow_region);
4089
4090 Node* init_mem = reset_memory();
4091 // fill in the rest of the null path:
4092 result_io ->init_req(_null_path, i_o());
4093 result_mem->init_req(_null_path, init_mem);
4094
4095 result_val->init_req(_fast_path, hash_val);
4096 result_reg->init_req(_fast_path, control());
4097 result_io ->init_req(_fast_path, i_o());
4098 result_mem->init_req(_fast_path, init_mem);
4099
4100 // Generate code for the slow case. We make a call to hashCode().
4101 set_control(_gvn.transform(slow_region));
4102 if (!stopped()) {
4103 // No need for PreserveJVMState, because we're using up the present state.
4104 set_all_memory(init_mem);
4105 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4106 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4107 Node* slow_result = set_results_for_java_call(slow_call);
4108 // this->control() comes from set_results_for_java_call
4109 result_reg->init_req(_slow_path, control());
4110 result_val->init_req(_slow_path, slow_result);
4111 result_io ->set_req(_slow_path, i_o());
4112 result_mem ->set_req(_slow_path, reset_memory());
4113 }
4114
4115 // Return the combined state.
4116 set_i_o( _gvn.transform(result_io) );
4117 set_all_memory( _gvn.transform(result_mem));
4118
4119 set_result(result_reg, result_val);
4120 return true;
4121 }
4122
4123 //---------------------------inline_native_getClass----------------------------
4124 // public final native Class<?> java.lang.Object.getClass();
4125 //
4126 // Build special case code for calls to getClass on an object.
4127 bool LibraryCallKit::inline_native_getClass() {
4128 Node* obj = null_check_receiver();
4129 if (stopped()) return true;
4130 set_result(load_mirror_from_klass(load_object_klass(obj)));
4131 return true;
4132 }
4133
4134 //-----------------inline_native_Reflection_getCallerClass---------------------
4135 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4136 //
4137 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4138 //
4139 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4140 // in that it must skip particular security frames and checks for
4141 // caller sensitive methods.
4142 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4143 #ifndef PRODUCT
4144 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4145 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4146 }
4147 #endif
4148
4149 if (!jvms()->has_method()) {
4150 #ifndef PRODUCT
4151 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4152 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4153 }
4154 #endif
4155 return false;
4156 }
4157
4158 // Walk back up the JVM state to find the caller at the required
4159 // depth.
4160 JVMState* caller_jvms = jvms();
4161
4162 // Cf. JVM_GetCallerClass
4163 // NOTE: Start the loop at depth 1 because the current JVM state does
4164 // not include the Reflection.getCallerClass() frame.
4165 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4166 ciMethod* m = caller_jvms->method();
4167 switch (n) {
4168 case 0:
4169 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4170 break;
4171 case 1:
4172 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4173 if (!m->caller_sensitive()) {
4174 #ifndef PRODUCT
4175 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4176 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4177 }
4178 #endif
4179 return false; // bail-out; let JVM_GetCallerClass do the work
4180 }
4181 break;
4182 default:
4183 if (!m->is_ignored_by_security_stack_walk()) {
4184 // We have reached the desired frame; return the holder class.
4185 // Acquire method holder as java.lang.Class and push as constant.
4186 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4187 ciInstance* caller_mirror = caller_klass->java_mirror();
4188 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4189
4190 #ifndef PRODUCT
4191 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4192 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());
4193 tty->print_cr(" JVM state at this point:");
4194 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4195 ciMethod* m = jvms()->of_depth(i)->method();
4196 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4197 }
4198 }
4199 #endif
4200 return true;
4201 }
4202 break;
4203 }
4204 }
4205
4206 #ifndef PRODUCT
4207 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4208 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4209 tty->print_cr(" JVM state at this point:");
4210 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4211 ciMethod* m = jvms()->of_depth(i)->method();
4212 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4213 }
4214 }
4215 #endif
4216
4217 return false; // bail-out; let JVM_GetCallerClass do the work
4218 }
4219
4220 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4221 Node* arg = argument(0);
4222 Node* result;
4223
4224 switch (id) {
4225 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4226 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4227 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4228 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4229
4230 case vmIntrinsics::_doubleToLongBits: {
4231 // two paths (plus control) merge in a wood
4232 RegionNode *r = new RegionNode(3);
4233 Node *phi = new PhiNode(r, TypeLong::LONG);
4234
4235 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4236 // Build the boolean node
4237 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4238
4239 // Branch either way.
4240 // NaN case is less traveled, which makes all the difference.
4241 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4242 Node *opt_isnan = _gvn.transform(ifisnan);
4243 assert( opt_isnan->is_If(), "Expect an IfNode");
4244 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4245 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4246
4247 set_control(iftrue);
4248
4249 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4250 Node *slow_result = longcon(nan_bits); // return NaN
4251 phi->init_req(1, _gvn.transform( slow_result ));
4252 r->init_req(1, iftrue);
4253
4254 // Else fall through
4255 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4256 set_control(iffalse);
4257
4258 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4259 r->init_req(2, iffalse);
4260
4261 // Post merge
4262 set_control(_gvn.transform(r));
4263 record_for_igvn(r);
4264
4265 C->set_has_split_ifs(true); // Has chance for split-if optimization
4266 result = phi;
4267 assert(result->bottom_type()->isa_long(), "must be");
4268 break;
4269 }
4270
4271 case vmIntrinsics::_floatToIntBits: {
4272 // two paths (plus control) merge in a wood
4273 RegionNode *r = new RegionNode(3);
4274 Node *phi = new PhiNode(r, TypeInt::INT);
4275
4276 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4277 // Build the boolean node
4278 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4279
4280 // Branch either way.
4281 // NaN case is less traveled, which makes all the difference.
4282 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4283 Node *opt_isnan = _gvn.transform(ifisnan);
4284 assert( opt_isnan->is_If(), "Expect an IfNode");
4285 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4286 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4287
4288 set_control(iftrue);
4289
4290 static const jint nan_bits = 0x7fc00000;
4291 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4292 phi->init_req(1, _gvn.transform( slow_result ));
4293 r->init_req(1, iftrue);
4294
4295 // Else fall through
4296 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4297 set_control(iffalse);
4298
4299 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4300 r->init_req(2, iffalse);
4301
4302 // Post merge
4303 set_control(_gvn.transform(r));
4304 record_for_igvn(r);
4305
4306 C->set_has_split_ifs(true); // Has chance for split-if optimization
4307 result = phi;
4308 assert(result->bottom_type()->isa_int(), "must be");
4309 break;
4310 }
4311
4312 default:
4313 fatal_unexpected_iid(id);
4314 break;
4315 }
4316 set_result(_gvn.transform(result));
4317 return true;
4318 }
4319
4320 //----------------------inline_unsafe_copyMemory-------------------------
4321 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4322 bool LibraryCallKit::inline_unsafe_copyMemory() {
4323 if (callee()->is_static()) return false; // caller must have the capability!
4324 null_check_receiver(); // null-check receiver
4325 if (stopped()) return true;
4326
4327 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4328
4329 Node* src_ptr = argument(1); // type: oop
4330 Node* src_off = ConvL2X(argument(2)); // type: long
4331 Node* dst_ptr = argument(4); // type: oop
4332 Node* dst_off = ConvL2X(argument(5)); // type: long
4333 Node* size = ConvL2X(argument(7)); // type: long
4334
4335 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4336 "fieldOffset must be byte-scaled");
4337
4338 Node* src = make_unsafe_address(src_ptr, src_off);
4339 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4340
4341 // Conservatively insert a memory barrier on all memory slices.
4342 // Do not let writes of the copy source or destination float below the copy.
4343 insert_mem_bar(Op_MemBarCPUOrder);
4344
4345 // Call it. Note that the length argument is not scaled.
4346 make_runtime_call(RC_LEAF|RC_NO_FP,
4347 OptoRuntime::fast_arraycopy_Type(),
4348 StubRoutines::unsafe_arraycopy(),
4349 "unsafe_arraycopy",
4350 TypeRawPtr::BOTTOM,
4351 src, dst, size XTOP);
4352
4353 // Do not let reads of the copy destination float above the copy.
4354 insert_mem_bar(Op_MemBarCPUOrder);
4355
4356 return true;
4357 }
4358
4359 //------------------------clone_coping-----------------------------------
4360 // Helper function for inline_native_clone.
4361 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4362 assert(obj_size != NULL, "");
4363 Node* raw_obj = alloc_obj->in(1);
4364 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4365
4366 AllocateNode* alloc = NULL;
4367 if (ReduceBulkZeroing) {
4368 // We will be completely responsible for initializing this object -
4369 // mark Initialize node as complete.
4370 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4371 // The object was just allocated - there should be no any stores!
4372 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4373 // Mark as complete_with_arraycopy so that on AllocateNode
4374 // expansion, we know this AllocateNode is initialized by an array
4375 // copy and a StoreStore barrier exists after the array copy.
4376 alloc->initialization()->set_complete_with_arraycopy();
4377 }
4378
4379 // Copy the fastest available way.
4380 // TODO: generate fields copies for small objects instead.
4381 Node* src = obj;
4382 Node* dest = alloc_obj;
4383 Node* size = _gvn.transform(obj_size);
4384
4385 // Exclude the header but include array length to copy by 8 bytes words.
4386 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4387 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4388 instanceOopDesc::base_offset_in_bytes();
4389 // base_off:
4390 // 8 - 32-bit VM
4391 // 12 - 64-bit VM, compressed klass
4392 // 16 - 64-bit VM, normal klass
4393 if (base_off % BytesPerLong != 0) {
4394 assert(UseCompressedClassPointers, "");
4395 if (is_array) {
4396 // Exclude length to copy by 8 bytes words.
4397 base_off += sizeof(int);
4398 } else {
4399 // Include klass to copy by 8 bytes words.
4400 base_off = instanceOopDesc::klass_offset_in_bytes();
4401 }
4402 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4403 }
4404 src = basic_plus_adr(src, base_off);
4405 dest = basic_plus_adr(dest, base_off);
4406
4407 // Compute the length also, if needed:
4408 Node* countx = size;
4409 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4410 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4411
4412 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4413
4414 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4415 ac->set_clonebasic();
4416 Node* n = _gvn.transform(ac);
4417 if (n == ac) {
4418 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4419 } else {
4420 set_all_memory(n);
4421 }
4422
4423 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4424 if (card_mark) {
4425 assert(!is_array, "");
4426 // Put in store barrier for any and all oops we are sticking
4427 // into this object. (We could avoid this if we could prove
4428 // that the object type contains no oop fields at all.)
4429 Node* no_particular_value = NULL;
4430 Node* no_particular_field = NULL;
4431 int raw_adr_idx = Compile::AliasIdxRaw;
4432 post_barrier(control(),
4433 memory(raw_adr_type),
4434 alloc_obj,
4435 no_particular_field,
4436 raw_adr_idx,
4437 no_particular_value,
4438 T_OBJECT,
4439 false);
4440 }
4441
4442 // Do not let reads from the cloned object float above the arraycopy.
4443 if (alloc != NULL) {
4444 // Do not let stores that initialize this object be reordered with
4445 // a subsequent store that would make this object accessible by
4446 // other threads.
4447 // Record what AllocateNode this StoreStore protects so that
4448 // escape analysis can go from the MemBarStoreStoreNode to the
4449 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4450 // based on the escape status of the AllocateNode.
4451 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4452 } else {
4453 insert_mem_bar(Op_MemBarCPUOrder);
4454 }
4455 }
4456
4457 //------------------------inline_native_clone----------------------------
4458 // protected native Object java.lang.Object.clone();
4459 //
4460 // Here are the simple edge cases:
4461 // null receiver => normal trap
4462 // virtual and clone was overridden => slow path to out-of-line clone
4463 // not cloneable or finalizer => slow path to out-of-line Object.clone
4464 //
4465 // The general case has two steps, allocation and copying.
4466 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4467 //
4468 // Copying also has two cases, oop arrays and everything else.
4469 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4470 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4471 //
4472 // These steps fold up nicely if and when the cloned object's klass
4473 // can be sharply typed as an object array, a type array, or an instance.
4474 //
4475 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4476 PhiNode* result_val;
4477
4478 // Set the reexecute bit for the interpreter to reexecute
4479 // the bytecode that invokes Object.clone if deoptimization happens.
4480 { PreserveReexecuteState preexecs(this);
4481 jvms()->set_should_reexecute(true);
4482
4483 Node* obj = null_check_receiver();
4484 if (stopped()) return true;
4485
4486 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4487
4488 // If we are going to clone an instance, we need its exact type to
4489 // know the number and types of fields to convert the clone to
4490 // loads/stores. Maybe a speculative type can help us.
4491 if (!obj_type->klass_is_exact() &&
4492 obj_type->speculative_type() != NULL &&
4493 obj_type->speculative_type()->is_instance_klass()) {
4494 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4495 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4496 !spec_ik->has_injected_fields()) {
4497 ciKlass* k = obj_type->klass();
4498 if (!k->is_instance_klass() ||
4499 k->as_instance_klass()->is_interface() ||
4500 k->as_instance_klass()->has_subklass()) {
4501 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4502 }
4503 }
4504 }
4505
4506 Node* obj_klass = load_object_klass(obj);
4507 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4508 const TypeOopPtr* toop = ((tklass != NULL)
4509 ? tklass->as_instance_type()
4510 : TypeInstPtr::NOTNULL);
4511
4512 // Conservatively insert a memory barrier on all memory slices.
4513 // Do not let writes into the original float below the clone.
4514 insert_mem_bar(Op_MemBarCPUOrder);
4515
4516 // paths into result_reg:
4517 enum {
4518 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4519 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4520 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4521 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4522 PATH_LIMIT
4523 };
4524 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4525 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4526 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4527 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4528 record_for_igvn(result_reg);
4529
4530 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4531 int raw_adr_idx = Compile::AliasIdxRaw;
4532
4533 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4534 if (array_ctl != NULL) {
4535 // It's an array.
4536 PreserveJVMState pjvms(this);
4537 set_control(array_ctl);
4538 Node* obj_length = load_array_length(obj);
4539 Node* obj_size = NULL;
4540 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4541
4542 if (!use_ReduceInitialCardMarks()) {
4543 // If it is an oop array, it requires very special treatment,
4544 // because card marking is required on each card of the array.
4545 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4546 if (is_obja != NULL) {
4547 PreserveJVMState pjvms2(this);
4548 set_control(is_obja);
4549 // Generate a direct call to the right arraycopy function(s).
4550 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4551 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4552 ac->set_cloneoop();
4553 Node* n = _gvn.transform(ac);
4554 assert(n == ac, "cannot disappear");
4555 ac->connect_outputs(this);
4556
4557 result_reg->init_req(_objArray_path, control());
4558 result_val->init_req(_objArray_path, alloc_obj);
4559 result_i_o ->set_req(_objArray_path, i_o());
4560 result_mem ->set_req(_objArray_path, reset_memory());
4561 }
4562 }
4563 // Otherwise, there are no card marks to worry about.
4564 // (We can dispense with card marks if we know the allocation
4565 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4566 // causes the non-eden paths to take compensating steps to
4567 // simulate a fresh allocation, so that no further
4568 // card marks are required in compiled code to initialize
4569 // the object.)
4570
4571 if (!stopped()) {
4572 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4573
4574 // Present the results of the copy.
4575 result_reg->init_req(_array_path, control());
4576 result_val->init_req(_array_path, alloc_obj);
4577 result_i_o ->set_req(_array_path, i_o());
4578 result_mem ->set_req(_array_path, reset_memory());
4579 }
4580 }
4581
4582 // We only go to the instance fast case code if we pass a number of guards.
4583 // The paths which do not pass are accumulated in the slow_region.
4584 RegionNode* slow_region = new RegionNode(1);
4585 record_for_igvn(slow_region);
4586 if (!stopped()) {
4587 // It's an instance (we did array above). Make the slow-path tests.
4588 // If this is a virtual call, we generate a funny guard. We grab
4589 // the vtable entry corresponding to clone() from the target object.
4590 // If the target method which we are calling happens to be the
4591 // Object clone() method, we pass the guard. We do not need this
4592 // guard for non-virtual calls; the caller is known to be the native
4593 // Object clone().
4594 if (is_virtual) {
4595 generate_virtual_guard(obj_klass, slow_region);
4596 }
4597
4598 // The object must be cloneable and must not have a finalizer.
4599 // Both of these conditions may be checked in a single test.
4600 // We could optimize the cloneable test further, but we don't care.
4601 generate_access_flags_guard(obj_klass,
4602 // Test both conditions:
4603 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4604 // Must be cloneable but not finalizer:
4605 JVM_ACC_IS_CLONEABLE,
4606 slow_region);
4607 }
4608
4609 if (!stopped()) {
4610 // It's an instance, and it passed the slow-path tests.
4611 PreserveJVMState pjvms(this);
4612 Node* obj_size = NULL;
4613 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4614 // is reexecuted if deoptimization occurs and there could be problems when merging
4615 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4616 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4617
4618 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4619
4620 // Present the results of the slow call.
4621 result_reg->init_req(_instance_path, control());
4622 result_val->init_req(_instance_path, alloc_obj);
4623 result_i_o ->set_req(_instance_path, i_o());
4624 result_mem ->set_req(_instance_path, reset_memory());
4625 }
4626
4627 // Generate code for the slow case. We make a call to clone().
4628 set_control(_gvn.transform(slow_region));
4629 if (!stopped()) {
4630 PreserveJVMState pjvms(this);
4631 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4632 Node* slow_result = set_results_for_java_call(slow_call);
4633 // this->control() comes from set_results_for_java_call
4634 result_reg->init_req(_slow_path, control());
4635 result_val->init_req(_slow_path, slow_result);
4636 result_i_o ->set_req(_slow_path, i_o());
4637 result_mem ->set_req(_slow_path, reset_memory());
4638 }
4639
4640 // Return the combined state.
4641 set_control( _gvn.transform(result_reg));
4642 set_i_o( _gvn.transform(result_i_o));
4643 set_all_memory( _gvn.transform(result_mem));
4644 } // original reexecute is set back here
4645
4646 set_result(_gvn.transform(result_val));
4647 return true;
4648 }
4649
4650 // If we have a tighly coupled allocation, the arraycopy may take care
4651 // of the array initialization. If one of the guards we insert between
4652 // the allocation and the arraycopy causes a deoptimization, an
4653 // unitialized array will escape the compiled method. To prevent that
4654 // we set the JVM state for uncommon traps between the allocation and
4655 // the arraycopy to the state before the allocation so, in case of
4656 // deoptimization, we'll reexecute the allocation and the
4657 // initialization.
4658 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4659 if (alloc != NULL) {
4660 ciMethod* trap_method = alloc->jvms()->method();
4661 int trap_bci = alloc->jvms()->bci();
4662
4663 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4664 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4665 // Make sure there's no store between the allocation and the
4666 // arraycopy otherwise visible side effects could be rexecuted
4667 // in case of deoptimization and cause incorrect execution.
4668 bool no_interfering_store = true;
4669 Node* mem = alloc->in(TypeFunc::Memory);
4670 if (mem->is_MergeMem()) {
4671 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4672 Node* n = mms.memory();
4673 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4674 assert(n->is_Store(), "what else?");
4675 no_interfering_store = false;
4676 break;
4677 }
4678 }
4679 } else {
4680 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4681 Node* n = mms.memory();
4682 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4683 assert(n->is_Store(), "what else?");
4684 no_interfering_store = false;
4685 break;
4686 }
4687 }
4688 }
4689
4690 if (no_interfering_store) {
4691 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4692 uint size = alloc->req();
4693 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4694 old_jvms->set_map(sfpt);
4695 for (uint i = 0; i < size; i++) {
4696 sfpt->init_req(i, alloc->in(i));
4697 }
4698 // re-push array length for deoptimization
4699 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4700 old_jvms->set_sp(old_jvms->sp()+1);
4701 old_jvms->set_monoff(old_jvms->monoff()+1);
4702 old_jvms->set_scloff(old_jvms->scloff()+1);
4703 old_jvms->set_endoff(old_jvms->endoff()+1);
4704 old_jvms->set_should_reexecute(true);
4705
4706 sfpt->set_i_o(map()->i_o());
4707 sfpt->set_memory(map()->memory());
4708 sfpt->set_control(map()->control());
4709
4710 JVMState* saved_jvms = jvms();
4711 saved_reexecute_sp = _reexecute_sp;
4712
4713 set_jvms(sfpt->jvms());
4714 _reexecute_sp = jvms()->sp();
4715
4716 return saved_jvms;
4717 }
4718 }
4719 }
4720 return NULL;
4721 }
4722
4723 // In case of a deoptimization, we restart execution at the
4724 // allocation, allocating a new array. We would leave an uninitialized
4725 // array in the heap that GCs wouldn't expect. Move the allocation
4726 // after the traps so we don't allocate the array if we
4727 // deoptimize. This is possible because tightly_coupled_allocation()
4728 // guarantees there's no observer of the allocated array at this point
4729 // and the control flow is simple enough.
4730 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4731 if (saved_jvms != NULL && !stopped()) {
4732 assert(alloc != NULL, "only with a tightly coupled allocation");
4733 // restore JVM state to the state at the arraycopy
4734 saved_jvms->map()->set_control(map()->control());
4735 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4736 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4737 // If we've improved the types of some nodes (null check) while
4738 // emitting the guards, propagate them to the current state
4739 map()->replaced_nodes().apply(saved_jvms->map());
4740 set_jvms(saved_jvms);
4741 _reexecute_sp = saved_reexecute_sp;
4742
4743 // Remove the allocation from above the guards
4744 CallProjections callprojs;
4745 alloc->extract_projections(&callprojs, true);
4746 InitializeNode* init = alloc->initialization();
4747 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4748 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4749 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4750 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4751
4752 // move the allocation here (after the guards)
4753 _gvn.hash_delete(alloc);
4754 alloc->set_req(TypeFunc::Control, control());
4755 alloc->set_req(TypeFunc::I_O, i_o());
4756 Node *mem = reset_memory();
4757 set_all_memory(mem);
4758 alloc->set_req(TypeFunc::Memory, mem);
4759 set_control(init->proj_out(TypeFunc::Control));
4760 set_i_o(callprojs.fallthrough_ioproj);
4761
4762 // Update memory as done in GraphKit::set_output_for_allocation()
4763 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4764 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4765 if (ary_type->isa_aryptr() && length_type != NULL) {
4766 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4767 }
4768 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4769 int elemidx = C->get_alias_index(telemref);
4770 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4771 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4772
4773 Node* allocx = _gvn.transform(alloc);
4774 assert(allocx == alloc, "where has the allocation gone?");
4775 assert(dest->is_CheckCastPP(), "not an allocation result?");
4776
4777 _gvn.hash_delete(dest);
4778 dest->set_req(0, control());
4779 Node* destx = _gvn.transform(dest);
4780 assert(destx == dest, "where has the allocation result gone?");
4781 }
4782 }
4783
4784
4785 //------------------------------inline_arraycopy-----------------------
4786 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4787 // Object dest, int destPos,
4788 // int length);
4789 bool LibraryCallKit::inline_arraycopy() {
4790 // Get the arguments.
4791 Node* src = argument(0); // type: oop
4792 Node* src_offset = argument(1); // type: int
4793 Node* dest = argument(2); // type: oop
4794 Node* dest_offset = argument(3); // type: int
4795 Node* length = argument(4); // type: int
4796
4797
4798 // Check for allocation before we add nodes that would confuse
4799 // tightly_coupled_allocation()
4800 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4801
4802 int saved_reexecute_sp = -1;
4803 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4804 // See arraycopy_restore_alloc_state() comment
4805 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4806 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4807 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4808 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4809
4810 // The following tests must be performed
4811 // (1) src and dest are arrays.
4812 // (2) src and dest arrays must have elements of the same BasicType
4813 // (3) src and dest must not be null.
4814 // (4) src_offset must not be negative.
4815 // (5) dest_offset must not be negative.
4816 // (6) length must not be negative.
4817 // (7) src_offset + length must not exceed length of src.
4818 // (8) dest_offset + length must not exceed length of dest.
4819 // (9) each element of an oop array must be assignable
4820
4821 // (3) src and dest must not be null.
4822 // always do this here because we need the JVM state for uncommon traps
4823 Node* null_ctl = top();
4824 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4825 assert(null_ctl->is_top(), "no null control here");
4826 dest = null_check(dest, T_ARRAY);
4827
4828 if (!can_emit_guards) {
4829 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4830 // guards but the arraycopy node could still take advantage of a
4831 // tightly allocated allocation. tightly_coupled_allocation() is
4832 // called again to make sure it takes the null check above into
4833 // account: the null check is mandatory and if it caused an
4834 // uncommon trap to be emitted then the allocation can't be
4835 // considered tightly coupled in this context.
4836 alloc = tightly_coupled_allocation(dest, NULL);
4837 }
4838
4839 bool validated = false;
4840
4841 const Type* src_type = _gvn.type(src);
4842 const Type* dest_type = _gvn.type(dest);
4843 const TypeAryPtr* top_src = src_type->isa_aryptr();
4844 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4845
4846 // Do we have the type of src?
4847 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4848 // Do we have the type of dest?
4849 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4850 // Is the type for src from speculation?
4851 bool src_spec = false;
4852 // Is the type for dest from speculation?
4853 bool dest_spec = false;
4854
4855 if ((!has_src || !has_dest) && can_emit_guards) {
4856 // We don't have sufficient type information, let's see if
4857 // speculative types can help. We need to have types for both src
4858 // and dest so that it pays off.
4859
4860 // Do we already have or could we have type information for src
4861 bool could_have_src = has_src;
4862 // Do we already have or could we have type information for dest
4863 bool could_have_dest = has_dest;
4864
4865 ciKlass* src_k = NULL;
4866 if (!has_src) {
4867 src_k = src_type->speculative_type_not_null();
4868 if (src_k != NULL && src_k->is_array_klass()) {
4869 could_have_src = true;
4870 }
4871 }
4872
4873 ciKlass* dest_k = NULL;
4874 if (!has_dest) {
4875 dest_k = dest_type->speculative_type_not_null();
4876 if (dest_k != NULL && dest_k->is_array_klass()) {
4877 could_have_dest = true;
4878 }
4879 }
4880
4881 if (could_have_src && could_have_dest) {
4882 // This is going to pay off so emit the required guards
4883 if (!has_src) {
4884 src = maybe_cast_profiled_obj(src, src_k, true);
4885 src_type = _gvn.type(src);
4886 top_src = src_type->isa_aryptr();
4887 has_src = (top_src != NULL && top_src->klass() != NULL);
4888 src_spec = true;
4889 }
4890 if (!has_dest) {
4891 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4892 dest_type = _gvn.type(dest);
4893 top_dest = dest_type->isa_aryptr();
4894 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4895 dest_spec = true;
4896 }
4897 }
4898 }
4899
4900 if (has_src && has_dest && can_emit_guards) {
4901 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4902 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4903 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4904 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4905
4906 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4907 // If both arrays are object arrays then having the exact types
4908 // for both will remove the need for a subtype check at runtime
4909 // before the call and may make it possible to pick a faster copy
4910 // routine (without a subtype check on every element)
4911 // Do we have the exact type of src?
4912 bool could_have_src = src_spec;
4913 // Do we have the exact type of dest?
4914 bool could_have_dest = dest_spec;
4915 ciKlass* src_k = top_src->klass();
4916 ciKlass* dest_k = top_dest->klass();
4917 if (!src_spec) {
4918 src_k = src_type->speculative_type_not_null();
4919 if (src_k != NULL && src_k->is_array_klass()) {
4920 could_have_src = true;
4921 }
4922 }
4923 if (!dest_spec) {
4924 dest_k = dest_type->speculative_type_not_null();
4925 if (dest_k != NULL && dest_k->is_array_klass()) {
4926 could_have_dest = true;
4927 }
4928 }
4929 if (could_have_src && could_have_dest) {
4930 // If we can have both exact types, emit the missing guards
4931 if (could_have_src && !src_spec) {
4932 src = maybe_cast_profiled_obj(src, src_k, true);
4933 }
4934 if (could_have_dest && !dest_spec) {
4935 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4936 }
4937 }
4938 }
4939 }
4940
4941 ciMethod* trap_method = method();
4942 int trap_bci = bci();
4943 if (saved_jvms != NULL) {
4944 trap_method = alloc->jvms()->method();
4945 trap_bci = alloc->jvms()->bci();
4946 }
4947
4948 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4949 can_emit_guards &&
4950 !src->is_top() && !dest->is_top()) {
4951 // validate arguments: enables transformation the ArrayCopyNode
4952 validated = true;
4953
4954 RegionNode* slow_region = new RegionNode(1);
4955 record_for_igvn(slow_region);
4956
4957 // (1) src and dest are arrays.
4958 generate_non_array_guard(load_object_klass(src), slow_region);
4959 generate_non_array_guard(load_object_klass(dest), slow_region);
4960
4961 // (2) src and dest arrays must have elements of the same BasicType
4962 // done at macro expansion or at Ideal transformation time
4963
4964 // (4) src_offset must not be negative.
4965 generate_negative_guard(src_offset, slow_region);
4966
4967 // (5) dest_offset must not be negative.
4968 generate_negative_guard(dest_offset, slow_region);
4969
4970 // (7) src_offset + length must not exceed length of src.
4971 generate_limit_guard(src_offset, length,
4972 load_array_length(src),
4973 slow_region);
4974
4975 // (8) dest_offset + length must not exceed length of dest.
4976 generate_limit_guard(dest_offset, length,
4977 load_array_length(dest),
4978 slow_region);
4979
4980 // (9) each element of an oop array must be assignable
4981 Node* src_klass = load_object_klass(src);
4982 Node* dest_klass = load_object_klass(dest);
4983 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4984
4985 if (not_subtype_ctrl != top()) {
4986 PreserveJVMState pjvms(this);
4987 set_control(not_subtype_ctrl);
4988 uncommon_trap(Deoptimization::Reason_intrinsic,
4989 Deoptimization::Action_make_not_entrant);
4990 assert(stopped(), "Should be stopped");
4991 }
4992 {
4993 PreserveJVMState pjvms(this);
4994 set_control(_gvn.transform(slow_region));
4995 uncommon_trap(Deoptimization::Reason_intrinsic,
4996 Deoptimization::Action_make_not_entrant);
4997 assert(stopped(), "Should be stopped");
4998 }
4999 }
5000
5001 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5002
5003 if (stopped()) {
5004 return true;
5005 }
5006
5007 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5008 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5009 // so the compiler has a chance to eliminate them: during macro expansion,
5010 // we have to set their control (CastPP nodes are eliminated).
5011 load_object_klass(src), load_object_klass(dest),
5012 load_array_length(src), load_array_length(dest));
5013
5014 ac->set_arraycopy(validated);
5015
5016 Node* n = _gvn.transform(ac);
5017 if (n == ac) {
5018 ac->connect_outputs(this);
5019 } else {
5020 assert(validated, "shouldn't transform if all arguments not validated");
5021 set_all_memory(n);
5022 }
5023
5024 return true;
5025 }
5026
5027
5028 // Helper function which determines if an arraycopy immediately follows
5029 // an allocation, with no intervening tests or other escapes for the object.
5030 AllocateArrayNode*
5031 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5032 RegionNode* slow_region) {
5033 if (stopped()) return NULL; // no fast path
5034 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5035
5036 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5037 if (alloc == NULL) return NULL;
5038
5039 Node* rawmem = memory(Compile::AliasIdxRaw);
5040 // Is the allocation's memory state untouched?
5041 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5042 // Bail out if there have been raw-memory effects since the allocation.
5043 // (Example: There might have been a call or safepoint.)
5044 return NULL;
5045 }
5046 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5047 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5048 return NULL;
5049 }
5050
5051 // There must be no unexpected observers of this allocation.
5052 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5053 Node* obs = ptr->fast_out(i);
5054 if (obs != this->map()) {
5055 return NULL;
5056 }
5057 }
5058
5059 // This arraycopy must unconditionally follow the allocation of the ptr.
5060 Node* alloc_ctl = ptr->in(0);
5061 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5062
5063 Node* ctl = control();
5064 while (ctl != alloc_ctl) {
5065 // There may be guards which feed into the slow_region.
5066 // Any other control flow means that we might not get a chance
5067 // to finish initializing the allocated object.
5068 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5069 IfNode* iff = ctl->in(0)->as_If();
5070 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5071 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5072 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5073 ctl = iff->in(0); // This test feeds the known slow_region.
5074 continue;
5075 }
5076 // One more try: Various low-level checks bottom out in
5077 // uncommon traps. If the debug-info of the trap omits
5078 // any reference to the allocation, as we've already
5079 // observed, then there can be no objection to the trap.
5080 bool found_trap = false;
5081 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5082 Node* obs = not_ctl->fast_out(j);
5083 if (obs->in(0) == not_ctl && obs->is_Call() &&
5084 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5085 found_trap = true; break;
5086 }
5087 }
5088 if (found_trap) {
5089 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5090 continue;
5091 }
5092 }
5093 return NULL;
5094 }
5095
5096 // If we get this far, we have an allocation which immediately
5097 // precedes the arraycopy, and we can take over zeroing the new object.
5098 // The arraycopy will finish the initialization, and provide
5099 // a new control state to which we will anchor the destination pointer.
5100
5101 return alloc;
5102 }
5103
5104 //-------------inline_encodeISOArray-----------------------------------
5105 // encode char[] to byte[] in ISO_8859_1
5106 bool LibraryCallKit::inline_encodeISOArray() {
5107 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5108 // no receiver since it is static method
5109 Node *src = argument(0);
5110 Node *src_offset = argument(1);
5111 Node *dst = argument(2);
5112 Node *dst_offset = argument(3);
5113 Node *length = argument(4);
5114
5115 const Type* src_type = src->Value(&_gvn);
5116 const Type* dst_type = dst->Value(&_gvn);
5117 const TypeAryPtr* top_src = src_type->isa_aryptr();
5118 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5119 if (top_src == NULL || top_src->klass() == NULL ||
5120 top_dest == NULL || top_dest->klass() == NULL) {
5121 // failed array check
5122 return false;
5123 }
5124
5125 // Figure out the size and type of the elements we will be copying.
5126 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5127 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5128 if (!((src_elem == T_CHAR) || (src_elem== T_BYTE)) || dst_elem != T_BYTE) {
5129 return false;
5130 }
5131
5132 Node* src_start = array_element_address(src, src_offset, T_CHAR);
5133 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5134 // 'src_start' points to src array + scaled offset
5135 // 'dst_start' points to dst array + scaled offset
5136
5137 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5138 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5139 enc = _gvn.transform(enc);
5140 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5141 set_memory(res_mem, mtype);
5142 set_result(enc);
5143 return true;
5144 }
5145
5146 //-------------inline_multiplyToLen-----------------------------------
5147 bool LibraryCallKit::inline_multiplyToLen() {
5148 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5149
5150 address stubAddr = StubRoutines::multiplyToLen();
5151 if (stubAddr == NULL) {
5152 return false; // Intrinsic's stub is not implemented on this platform
5153 }
5154 const char* stubName = "multiplyToLen";
5155
5156 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5157
5158 // no receiver because it is a static method
5159 Node* x = argument(0);
5160 Node* xlen = argument(1);
5161 Node* y = argument(2);
5162 Node* ylen = argument(3);
5163 Node* z = argument(4);
5164
5165 const Type* x_type = x->Value(&_gvn);
5166 const Type* y_type = y->Value(&_gvn);
5167 const TypeAryPtr* top_x = x_type->isa_aryptr();
5168 const TypeAryPtr* top_y = y_type->isa_aryptr();
5169 if (top_x == NULL || top_x->klass() == NULL ||
5170 top_y == NULL || top_y->klass() == NULL) {
5171 // failed array check
5172 return false;
5173 }
5174
5175 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5176 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5177 if (x_elem != T_INT || y_elem != T_INT) {
5178 return false;
5179 }
5180
5181 // Set the original stack and the reexecute bit for the interpreter to reexecute
5182 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5183 // on the return from z array allocation in runtime.
5184 { PreserveReexecuteState preexecs(this);
5185 jvms()->set_should_reexecute(true);
5186
5187 Node* x_start = array_element_address(x, intcon(0), x_elem);
5188 Node* y_start = array_element_address(y, intcon(0), y_elem);
5189 // 'x_start' points to x array + scaled xlen
5190 // 'y_start' points to y array + scaled ylen
5191
5192 // Allocate the result array
5193 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5194 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5195 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5196
5197 IdealKit ideal(this);
5198
5199 #define __ ideal.
5200 Node* one = __ ConI(1);
5201 Node* zero = __ ConI(0);
5202 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5203 __ set(need_alloc, zero);
5204 __ set(z_alloc, z);
5205 __ if_then(z, BoolTest::eq, null()); {
5206 __ increment (need_alloc, one);
5207 } __ else_(); {
5208 // Update graphKit memory and control from IdealKit.
5209 sync_kit(ideal);
5210 Node* zlen_arg = load_array_length(z);
5211 // Update IdealKit memory and control from graphKit.
5212 __ sync_kit(this);
5213 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5214 __ increment (need_alloc, one);
5215 } __ end_if();
5216 } __ end_if();
5217
5218 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5219 // Update graphKit memory and control from IdealKit.
5220 sync_kit(ideal);
5221 Node * narr = new_array(klass_node, zlen, 1);
5222 // Update IdealKit memory and control from graphKit.
5223 __ sync_kit(this);
5224 __ set(z_alloc, narr);
5225 } __ end_if();
5226
5227 sync_kit(ideal);
5228 z = __ value(z_alloc);
5229 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5230 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5231 // Final sync IdealKit and GraphKit.
5232 final_sync(ideal);
5233 #undef __
5234
5235 Node* z_start = array_element_address(z, intcon(0), T_INT);
5236
5237 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5238 OptoRuntime::multiplyToLen_Type(),
5239 stubAddr, stubName, TypePtr::BOTTOM,
5240 x_start, xlen, y_start, ylen, z_start, zlen);
5241 } // original reexecute is set back here
5242
5243 C->set_has_split_ifs(true); // Has chance for split-if optimization
5244 set_result(z);
5245 return true;
5246 }
5247
5248 //-------------inline_squareToLen------------------------------------
5249 bool LibraryCallKit::inline_squareToLen() {
5250 assert(UseSquareToLenIntrinsic, "not implemented on this platform");
5251
5252 address stubAddr = StubRoutines::squareToLen();
5253 if (stubAddr == NULL) {
5254 return false; // Intrinsic's stub is not implemented on this platform
5255 }
5256 const char* stubName = "squareToLen";
5257
5258 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5259
5260 Node* x = argument(0);
5261 Node* len = argument(1);
5262 Node* z = argument(2);
5263 Node* zlen = argument(3);
5264
5265 const Type* x_type = x->Value(&_gvn);
5266 const Type* z_type = z->Value(&_gvn);
5267 const TypeAryPtr* top_x = x_type->isa_aryptr();
5268 const TypeAryPtr* top_z = z_type->isa_aryptr();
5269 if (top_x == NULL || top_x->klass() == NULL ||
5270 top_z == NULL || top_z->klass() == NULL) {
5271 // failed array check
5272 return false;
5273 }
5274
5275 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5276 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5277 if (x_elem != T_INT || z_elem != T_INT) {
5278 return false;
5279 }
5280
5281
5282 Node* x_start = array_element_address(x, intcon(0), x_elem);
5283 Node* z_start = array_element_address(z, intcon(0), z_elem);
5284
5285 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5286 OptoRuntime::squareToLen_Type(),
5287 stubAddr, stubName, TypePtr::BOTTOM,
5288 x_start, len, z_start, zlen);
5289
5290 set_result(z);
5291 return true;
5292 }
5293
5294 //-------------inline_mulAdd------------------------------------------
5295 bool LibraryCallKit::inline_mulAdd() {
5296 assert(UseMulAddIntrinsic, "not implemented on this platform");
5297
5298 address stubAddr = StubRoutines::mulAdd();
5299 if (stubAddr == NULL) {
5300 return false; // Intrinsic's stub is not implemented on this platform
5301 }
5302 const char* stubName = "mulAdd";
5303
5304 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5305
5306 Node* out = argument(0);
5307 Node* in = argument(1);
5308 Node* offset = argument(2);
5309 Node* len = argument(3);
5310 Node* k = argument(4);
5311
5312 const Type* out_type = out->Value(&_gvn);
5313 const Type* in_type = in->Value(&_gvn);
5314 const TypeAryPtr* top_out = out_type->isa_aryptr();
5315 const TypeAryPtr* top_in = in_type->isa_aryptr();
5316 if (top_out == NULL || top_out->klass() == NULL ||
5317 top_in == NULL || top_in->klass() == NULL) {
5318 // failed array check
5319 return false;
5320 }
5321
5322 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5323 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5324 if (out_elem != T_INT || in_elem != T_INT) {
5325 return false;
5326 }
5327
5328 Node* outlen = load_array_length(out);
5329 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5330 Node* out_start = array_element_address(out, intcon(0), out_elem);
5331 Node* in_start = array_element_address(in, intcon(0), in_elem);
5332
5333 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5334 OptoRuntime::mulAdd_Type(),
5335 stubAddr, stubName, TypePtr::BOTTOM,
5336 out_start,in_start, new_offset, len, k);
5337 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5338 set_result(result);
5339 return true;
5340 }
5341
5342 //-------------inline_montgomeryMultiply-----------------------------------
5343 bool LibraryCallKit::inline_montgomeryMultiply() {
5344 address stubAddr = StubRoutines::montgomeryMultiply();
5345 if (stubAddr == NULL) {
5346 return false; // Intrinsic's stub is not implemented on this platform
5347 }
5348
5349 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5350 const char* stubName = "montgomery_square";
5351
5352 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5353
5354 Node* a = argument(0);
5355 Node* b = argument(1);
5356 Node* n = argument(2);
5357 Node* len = argument(3);
5358 Node* inv = argument(4);
5359 Node* m = argument(6);
5360
5361 const Type* a_type = a->Value(&_gvn);
5362 const TypeAryPtr* top_a = a_type->isa_aryptr();
5363 const Type* b_type = b->Value(&_gvn);
5364 const TypeAryPtr* top_b = b_type->isa_aryptr();
5365 const Type* n_type = a->Value(&_gvn);
5366 const TypeAryPtr* top_n = n_type->isa_aryptr();
5367 const Type* m_type = a->Value(&_gvn);
5368 const TypeAryPtr* top_m = m_type->isa_aryptr();
5369 if (top_a == NULL || top_a->klass() == NULL ||
5370 top_b == NULL || top_b->klass() == NULL ||
5371 top_n == NULL || top_n->klass() == NULL ||
5372 top_m == NULL || top_m->klass() == NULL) {
5373 // failed array check
5374 return false;
5375 }
5376
5377 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5378 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5379 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5380 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5381 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5382 return false;
5383 }
5384
5385 // Make the call
5386 {
5387 Node* a_start = array_element_address(a, intcon(0), a_elem);
5388 Node* b_start = array_element_address(b, intcon(0), b_elem);
5389 Node* n_start = array_element_address(n, intcon(0), n_elem);
5390 Node* m_start = array_element_address(m, intcon(0), m_elem);
5391
5392 Node* call = make_runtime_call(RC_LEAF,
5393 OptoRuntime::montgomeryMultiply_Type(),
5394 stubAddr, stubName, TypePtr::BOTTOM,
5395 a_start, b_start, n_start, len, inv, top(),
5396 m_start);
5397 set_result(m);
5398 }
5399
5400 return true;
5401 }
5402
5403 bool LibraryCallKit::inline_montgomerySquare() {
5404 address stubAddr = StubRoutines::montgomerySquare();
5405 if (stubAddr == NULL) {
5406 return false; // Intrinsic's stub is not implemented on this platform
5407 }
5408
5409 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5410 const char* stubName = "montgomery_square";
5411
5412 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5413
5414 Node* a = argument(0);
5415 Node* n = argument(1);
5416 Node* len = argument(2);
5417 Node* inv = argument(3);
5418 Node* m = argument(5);
5419
5420 const Type* a_type = a->Value(&_gvn);
5421 const TypeAryPtr* top_a = a_type->isa_aryptr();
5422 const Type* n_type = a->Value(&_gvn);
5423 const TypeAryPtr* top_n = n_type->isa_aryptr();
5424 const Type* m_type = a->Value(&_gvn);
5425 const TypeAryPtr* top_m = m_type->isa_aryptr();
5426 if (top_a == NULL || top_a->klass() == NULL ||
5427 top_n == NULL || top_n->klass() == NULL ||
5428 top_m == NULL || top_m->klass() == NULL) {
5429 // failed array check
5430 return false;
5431 }
5432
5433 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5434 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5435 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5436 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5437 return false;
5438 }
5439
5440 // Make the call
5441 {
5442 Node* a_start = array_element_address(a, intcon(0), a_elem);
5443 Node* n_start = array_element_address(n, intcon(0), n_elem);
5444 Node* m_start = array_element_address(m, intcon(0), m_elem);
5445
5446 Node* call = make_runtime_call(RC_LEAF,
5447 OptoRuntime::montgomerySquare_Type(),
5448 stubAddr, stubName, TypePtr::BOTTOM,
5449 a_start, n_start, len, inv, top(),
5450 m_start);
5451 set_result(m);
5452 }
5453
5454 return true;
5455 }
5456
5457
5458 /**
5459 * Calculate CRC32 for byte.
5460 * int java.util.zip.CRC32.update(int crc, int b)
5461 */
5462 bool LibraryCallKit::inline_updateCRC32() {
5463 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5464 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5465 // no receiver since it is static method
5466 Node* crc = argument(0); // type: int
5467 Node* b = argument(1); // type: int
5468
5469 /*
5470 * int c = ~ crc;
5471 * b = timesXtoThe32[(b ^ c) & 0xFF];
5472 * b = b ^ (c >>> 8);
5473 * crc = ~b;
5474 */
5475
5476 Node* M1 = intcon(-1);
5477 crc = _gvn.transform(new XorINode(crc, M1));
5478 Node* result = _gvn.transform(new XorINode(crc, b));
5479 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5480
5481 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5482 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5483 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5484 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5485
5486 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5487 result = _gvn.transform(new XorINode(crc, result));
5488 result = _gvn.transform(new XorINode(result, M1));
5489 set_result(result);
5490 return true;
5491 }
5492
5493 /**
5494 * Calculate CRC32 for byte[] array.
5495 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5496 */
5497 bool LibraryCallKit::inline_updateBytesCRC32() {
5498 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5499 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5500 // no receiver since it is static method
5501 Node* crc = argument(0); // type: int
5502 Node* src = argument(1); // type: oop
5503 Node* offset = argument(2); // type: int
5504 Node* length = argument(3); // type: int
5505
5506 const Type* src_type = src->Value(&_gvn);
5507 const TypeAryPtr* top_src = src_type->isa_aryptr();
5508 if (top_src == NULL || top_src->klass() == NULL) {
5509 // failed array check
5510 return false;
5511 }
5512
5513 // Figure out the size and type of the elements we will be copying.
5514 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5515 if (src_elem != T_BYTE) {
5516 return false;
5517 }
5518
5519 // 'src_start' points to src array + scaled offset
5520 Node* src_start = array_element_address(src, offset, src_elem);
5521
5522 // We assume that range check is done by caller.
5523 // TODO: generate range check (offset+length < src.length) in debug VM.
5524
5525 // Call the stub.
5526 address stubAddr = StubRoutines::updateBytesCRC32();
5527 const char *stubName = "updateBytesCRC32";
5528
5529 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5530 stubAddr, stubName, TypePtr::BOTTOM,
5531 crc, src_start, length);
5532 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5533 set_result(result);
5534 return true;
5535 }
5536
5537 /**
5538 * Calculate CRC32 for ByteBuffer.
5539 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5540 */
5541 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5542 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5543 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5544 // no receiver since it is static method
5545 Node* crc = argument(0); // type: int
5546 Node* src = argument(1); // type: long
5547 Node* offset = argument(3); // type: int
5548 Node* length = argument(4); // type: int
5549
5550 src = ConvL2X(src); // adjust Java long to machine word
5551 Node* base = _gvn.transform(new CastX2PNode(src));
5552 offset = ConvI2X(offset);
5553
5554 // 'src_start' points to src array + scaled offset
5555 Node* src_start = basic_plus_adr(top(), base, offset);
5556
5557 // Call the stub.
5558 address stubAddr = StubRoutines::updateBytesCRC32();
5559 const char *stubName = "updateBytesCRC32";
5560
5561 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5562 stubAddr, stubName, TypePtr::BOTTOM,
5563 crc, src_start, length);
5564 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5565 set_result(result);
5566 return true;
5567 }
5568
5569 //------------------------------get_table_from_crc32c_class-----------------------
5570 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5571 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5572 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5573
5574 return table;
5575 }
5576
5577 //------------------------------inline_updateBytesCRC32C-----------------------
5578 //
5579 // Calculate CRC32C for byte[] array.
5580 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5581 //
5582 bool LibraryCallKit::inline_updateBytesCRC32C() {
5583 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5584 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5585 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5586 // no receiver since it is a static method
5587 Node* crc = argument(0); // type: int
5588 Node* src = argument(1); // type: oop
5589 Node* offset = argument(2); // type: int
5590 Node* end = argument(3); // type: int
5591
5592 Node* length = _gvn.transform(new SubINode(end, offset));
5593
5594 const Type* src_type = src->Value(&_gvn);
5595 const TypeAryPtr* top_src = src_type->isa_aryptr();
5596 if (top_src == NULL || top_src->klass() == NULL) {
5597 // failed array check
5598 return false;
5599 }
5600
5601 // Figure out the size and type of the elements we will be copying.
5602 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5603 if (src_elem != T_BYTE) {
5604 return false;
5605 }
5606
5607 // 'src_start' points to src array + scaled offset
5608 Node* src_start = array_element_address(src, offset, src_elem);
5609
5610 // static final int[] byteTable in class CRC32C
5611 Node* table = get_table_from_crc32c_class(callee()->holder());
5612 Node* table_start = array_element_address(table, intcon(0), T_INT);
5613
5614 // We assume that range check is done by caller.
5615 // TODO: generate range check (offset+length < src.length) in debug VM.
5616
5617 // Call the stub.
5618 address stubAddr = StubRoutines::updateBytesCRC32C();
5619 const char *stubName = "updateBytesCRC32C";
5620
5621 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5622 stubAddr, stubName, TypePtr::BOTTOM,
5623 crc, src_start, length, table_start);
5624 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5625 set_result(result);
5626 return true;
5627 }
5628
5629 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5630 //
5631 // Calculate CRC32C for DirectByteBuffer.
5632 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5633 //
5634 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5635 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5636 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5637 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5638 // no receiver since it is a static method
5639 Node* crc = argument(0); // type: int
5640 Node* src = argument(1); // type: long
5641 Node* offset = argument(3); // type: int
5642 Node* end = argument(4); // type: int
5643
5644 Node* length = _gvn.transform(new SubINode(end, offset));
5645
5646 src = ConvL2X(src); // adjust Java long to machine word
5647 Node* base = _gvn.transform(new CastX2PNode(src));
5648 offset = ConvI2X(offset);
5649
5650 // 'src_start' points to src array + scaled offset
5651 Node* src_start = basic_plus_adr(top(), base, offset);
5652
5653 // static final int[] byteTable in class CRC32C
5654 Node* table = get_table_from_crc32c_class(callee()->holder());
5655 Node* table_start = array_element_address(table, intcon(0), T_INT);
5656
5657 // Call the stub.
5658 address stubAddr = StubRoutines::updateBytesCRC32C();
5659 const char *stubName = "updateBytesCRC32C";
5660
5661 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5662 stubAddr, stubName, TypePtr::BOTTOM,
5663 crc, src_start, length, table_start);
5664 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5665 set_result(result);
5666 return true;
5667 }
5668
5669 //------------------------------inline_updateBytesAdler32----------------------
5670 //
5671 // Calculate Adler32 checksum for byte[] array.
5672 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5673 //
5674 bool LibraryCallKit::inline_updateBytesAdler32() {
5675 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5676 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5677 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5678 // no receiver since it is static method
5679 Node* crc = argument(0); // type: int
5680 Node* src = argument(1); // type: oop
5681 Node* offset = argument(2); // type: int
5682 Node* length = argument(3); // type: int
5683
5684 const Type* src_type = src->Value(&_gvn);
5685 const TypeAryPtr* top_src = src_type->isa_aryptr();
5686 if (top_src == NULL || top_src->klass() == NULL) {
5687 // failed array check
5688 return false;
5689 }
5690
5691 // Figure out the size and type of the elements we will be copying.
5692 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5693 if (src_elem != T_BYTE) {
5694 return false;
5695 }
5696
5697 // 'src_start' points to src array + scaled offset
5698 Node* src_start = array_element_address(src, offset, src_elem);
5699
5700 // We assume that range check is done by caller.
5701 // TODO: generate range check (offset+length < src.length) in debug VM.
5702
5703 // Call the stub.
5704 address stubAddr = StubRoutines::updateBytesAdler32();
5705 const char *stubName = "updateBytesAdler32";
5706
5707 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5708 stubAddr, stubName, TypePtr::BOTTOM,
5709 crc, src_start, length);
5710 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5711 set_result(result);
5712 return true;
5713 }
5714
5715 //------------------------------inline_updateByteBufferAdler32---------------
5716 //
5717 // Calculate Adler32 checksum for DirectByteBuffer.
5718 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5719 //
5720 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5721 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5722 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5723 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5724 // no receiver since it is static method
5725 Node* crc = argument(0); // type: int
5726 Node* src = argument(1); // type: long
5727 Node* offset = argument(3); // type: int
5728 Node* length = argument(4); // type: int
5729
5730 src = ConvL2X(src); // adjust Java long to machine word
5731 Node* base = _gvn.transform(new CastX2PNode(src));
5732 offset = ConvI2X(offset);
5733
5734 // 'src_start' points to src array + scaled offset
5735 Node* src_start = basic_plus_adr(top(), base, offset);
5736
5737 // Call the stub.
5738 address stubAddr = StubRoutines::updateBytesAdler32();
5739 const char *stubName = "updateBytesAdler32";
5740
5741 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5742 stubAddr, stubName, TypePtr::BOTTOM,
5743 crc, src_start, length);
5744
5745 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5746 set_result(result);
5747 return true;
5748 }
5749
5750 //----------------------------inline_reference_get----------------------------
5751 // public T java.lang.ref.Reference.get();
5752 bool LibraryCallKit::inline_reference_get() {
5753 const int referent_offset = java_lang_ref_Reference::referent_offset;
5754 guarantee(referent_offset > 0, "should have already been set");
5755
5756 // Get the argument:
5757 Node* reference_obj = null_check_receiver();
5758 if (stopped()) return true;
5759
5760 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5761
5762 ciInstanceKlass* klass = env()->Object_klass();
5763 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5764
5765 Node* no_ctrl = NULL;
5766 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5767
5768 // Use the pre-barrier to record the value in the referent field
5769 pre_barrier(false /* do_load */,
5770 control(),
5771 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5772 result /* pre_val */,
5773 T_OBJECT);
5774
5775 // Add memory barrier to prevent commoning reads from this field
5776 // across safepoint since GC can change its value.
5777 insert_mem_bar(Op_MemBarCPUOrder);
5778
5779 set_result(result);
5780 return true;
5781 }
5782
5783
5784 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5785 bool is_exact=true, bool is_static=false,
5786 ciInstanceKlass * fromKls=NULL) {
5787 if (fromKls == NULL) {
5788 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5789 assert(tinst != NULL, "obj is null");
5790 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5791 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5792 fromKls = tinst->klass()->as_instance_klass();
5793 } else {
5794 assert(is_static, "only for static field access");
5795 }
5796 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5797 ciSymbol::make(fieldTypeString),
5798 is_static);
5799
5800 assert (field != NULL, "undefined field");
5801 if (field == NULL) return (Node *) NULL;
5802
5803 if (is_static) {
5804 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5805 fromObj = makecon(tip);
5806 }
5807
5808 // Next code copied from Parse::do_get_xxx():
5809
5810 // Compute address and memory type.
5811 int offset = field->offset_in_bytes();
5812 bool is_vol = field->is_volatile();
5813 ciType* field_klass = field->type();
5814 assert(field_klass->is_loaded(), "should be loaded");
5815 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5816 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5817 BasicType bt = field->layout_type();
5818
5819 // Build the resultant type of the load
5820 const Type *type;
5821 if (bt == T_OBJECT) {
5822 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5823 } else {
5824 type = Type::get_const_basic_type(bt);
5825 }
5826
5827 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5828 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5829 }
5830 // Build the load.
5831 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5832 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5833 // If reference is volatile, prevent following memory ops from
5834 // floating up past the volatile read. Also prevents commoning
5835 // another volatile read.
5836 if (is_vol) {
5837 // Memory barrier includes bogus read of value to force load BEFORE membar
5838 insert_mem_bar(Op_MemBarAcquire, loadedField);
5839 }
5840 return loadedField;
5841 }
5842
5843
5844 //------------------------------inline_aescrypt_Block-----------------------
5845 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5846 address stubAddr;
5847 const char *stubName;
5848 assert(UseAES, "need AES instruction support");
5849
5850 switch(id) {
5851 case vmIntrinsics::_aescrypt_encryptBlock:
5852 stubAddr = StubRoutines::aescrypt_encryptBlock();
5853 stubName = "aescrypt_encryptBlock";
5854 break;
5855 case vmIntrinsics::_aescrypt_decryptBlock:
5856 stubAddr = StubRoutines::aescrypt_decryptBlock();
5857 stubName = "aescrypt_decryptBlock";
5858 break;
5859 }
5860 if (stubAddr == NULL) return false;
5861
5862 Node* aescrypt_object = argument(0);
5863 Node* src = argument(1);
5864 Node* src_offset = argument(2);
5865 Node* dest = argument(3);
5866 Node* dest_offset = argument(4);
5867
5868 // (1) src and dest are arrays.
5869 const Type* src_type = src->Value(&_gvn);
5870 const Type* dest_type = dest->Value(&_gvn);
5871 const TypeAryPtr* top_src = src_type->isa_aryptr();
5872 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5873 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5874
5875 // for the quick and dirty code we will skip all the checks.
5876 // we are just trying to get the call to be generated.
5877 Node* src_start = src;
5878 Node* dest_start = dest;
5879 if (src_offset != NULL || dest_offset != NULL) {
5880 assert(src_offset != NULL && dest_offset != NULL, "");
5881 src_start = array_element_address(src, src_offset, T_BYTE);
5882 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5883 }
5884
5885 // now need to get the start of its expanded key array
5886 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5887 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5888 if (k_start == NULL) return false;
5889
5890 if (Matcher::pass_original_key_for_aes()) {
5891 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5892 // compatibility issues between Java key expansion and SPARC crypto instructions
5893 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5894 if (original_k_start == NULL) return false;
5895
5896 // Call the stub.
5897 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5898 stubAddr, stubName, TypePtr::BOTTOM,
5899 src_start, dest_start, k_start, original_k_start);
5900 } else {
5901 // Call the stub.
5902 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5903 stubAddr, stubName, TypePtr::BOTTOM,
5904 src_start, dest_start, k_start);
5905 }
5906
5907 return true;
5908 }
5909
5910 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5911 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5912 address stubAddr;
5913 const char *stubName;
5914
5915 assert(UseAES, "need AES instruction support");
5916
5917 switch(id) {
5918 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5919 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5920 stubName = "cipherBlockChaining_encryptAESCrypt";
5921 break;
5922 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5923 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5924 stubName = "cipherBlockChaining_decryptAESCrypt";
5925 break;
5926 }
5927 if (stubAddr == NULL) return false;
5928
5929 Node* cipherBlockChaining_object = argument(0);
5930 Node* src = argument(1);
5931 Node* src_offset = argument(2);
5932 Node* len = argument(3);
5933 Node* dest = argument(4);
5934 Node* dest_offset = argument(5);
5935
5936 // (1) src and dest are arrays.
5937 const Type* src_type = src->Value(&_gvn);
5938 const Type* dest_type = dest->Value(&_gvn);
5939 const TypeAryPtr* top_src = src_type->isa_aryptr();
5940 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5941 assert (top_src != NULL && top_src->klass() != NULL
5942 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5943
5944 // checks are the responsibility of the caller
5945 Node* src_start = src;
5946 Node* dest_start = dest;
5947 if (src_offset != NULL || dest_offset != NULL) {
5948 assert(src_offset != NULL && dest_offset != NULL, "");
5949 src_start = array_element_address(src, src_offset, T_BYTE);
5950 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5951 }
5952
5953 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5954 // (because of the predicated logic executed earlier).
5955 // so we cast it here safely.
5956 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5957
5958 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5959 if (embeddedCipherObj == NULL) return false;
5960
5961 // cast it to what we know it will be at runtime
5962 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5963 assert(tinst != NULL, "CBC obj is null");
5964 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5965 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5966 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5967
5968 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5969 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5970 const TypeOopPtr* xtype = aklass->as_instance_type();
5971 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5972 aescrypt_object = _gvn.transform(aescrypt_object);
5973
5974 // we need to get the start of the aescrypt_object's expanded key array
5975 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5976 if (k_start == NULL) return false;
5977
5978 // similarly, get the start address of the r vector
5979 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5980 if (objRvec == NULL) return false;
5981 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5982
5983 Node* cbcCrypt;
5984 if (Matcher::pass_original_key_for_aes()) {
5985 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5986 // compatibility issues between Java key expansion and SPARC crypto instructions
5987 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5988 if (original_k_start == NULL) return false;
5989
5990 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5991 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5992 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5993 stubAddr, stubName, TypePtr::BOTTOM,
5994 src_start, dest_start, k_start, r_start, len, original_k_start);
5995 } else {
5996 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5997 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5998 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5999 stubAddr, stubName, TypePtr::BOTTOM,
6000 src_start, dest_start, k_start, r_start, len);
6001 }
6002
6003 // return cipher length (int)
6004 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6005 set_result(retvalue);
6006 return true;
6007 }
6008
6009 //------------------------------get_key_start_from_aescrypt_object-----------------------
6010 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6011 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6012 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6013 if (objAESCryptKey == NULL) return (Node *) NULL;
6014
6015 // now have the array, need to get the start address of the K array
6016 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6017 return k_start;
6018 }
6019
6020 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6021 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6022 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6023 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6024 if (objAESCryptKey == NULL) return (Node *) NULL;
6025
6026 // now have the array, need to get the start address of the lastKey array
6027 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6028 return original_k_start;
6029 }
6030
6031 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6032 // Return node representing slow path of predicate check.
6033 // the pseudo code we want to emulate with this predicate is:
6034 // for encryption:
6035 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6036 // for decryption:
6037 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6038 // note cipher==plain is more conservative than the original java code but that's OK
6039 //
6040 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6041 // The receiver was checked for NULL already.
6042 Node* objCBC = argument(0);
6043
6044 // Load embeddedCipher field of CipherBlockChaining object.
6045 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6046
6047 // get AESCrypt klass for instanceOf check
6048 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6049 // will have same classloader as CipherBlockChaining object
6050 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6051 assert(tinst != NULL, "CBCobj is null");
6052 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6053
6054 // we want to do an instanceof comparison against the AESCrypt class
6055 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6056 if (!klass_AESCrypt->is_loaded()) {
6057 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6058 Node* ctrl = control();
6059 set_control(top()); // no regular fast path
6060 return ctrl;
6061 }
6062 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6063
6064 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6065 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6066 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6067
6068 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6069
6070 // for encryption, we are done
6071 if (!decrypting)
6072 return instof_false; // even if it is NULL
6073
6074 // for decryption, we need to add a further check to avoid
6075 // taking the intrinsic path when cipher and plain are the same
6076 // see the original java code for why.
6077 RegionNode* region = new RegionNode(3);
6078 region->init_req(1, instof_false);
6079 Node* src = argument(1);
6080 Node* dest = argument(4);
6081 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6082 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6083 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6084 region->init_req(2, src_dest_conjoint);
6085
6086 record_for_igvn(region);
6087 return _gvn.transform(region);
6088 }
6089
6090 //------------------------------inline_ghash_processBlocks
6091 bool LibraryCallKit::inline_ghash_processBlocks() {
6092 address stubAddr;
6093 const char *stubName;
6094 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6095
6096 stubAddr = StubRoutines::ghash_processBlocks();
6097 stubName = "ghash_processBlocks";
6098
6099 Node* data = argument(0);
6100 Node* offset = argument(1);
6101 Node* len = argument(2);
6102 Node* state = argument(3);
6103 Node* subkeyH = argument(4);
6104
6105 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6106 assert(state_start, "state is NULL");
6107 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6108 assert(subkeyH_start, "subkeyH is NULL");
6109 Node* data_start = array_element_address(data, offset, T_BYTE);
6110 assert(data_start, "data is NULL");
6111
6112 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6113 OptoRuntime::ghash_processBlocks_Type(),
6114 stubAddr, stubName, TypePtr::BOTTOM,
6115 state_start, subkeyH_start, data_start, len);
6116 return true;
6117 }
6118
6119 //------------------------------inline_sha_implCompress-----------------------
6120 //
6121 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6122 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6123 //
6124 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6125 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6126 //
6127 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6128 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6129 //
6130 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6131 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6132
6133 Node* sha_obj = argument(0);
6134 Node* src = argument(1); // type oop
6135 Node* ofs = argument(2); // type int
6136
6137 const Type* src_type = src->Value(&_gvn);
6138 const TypeAryPtr* top_src = src_type->isa_aryptr();
6139 if (top_src == NULL || top_src->klass() == NULL) {
6140 // failed array check
6141 return false;
6142 }
6143 // Figure out the size and type of the elements we will be copying.
6144 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6145 if (src_elem != T_BYTE) {
6146 return false;
6147 }
6148 // 'src_start' points to src array + offset
6149 Node* src_start = array_element_address(src, ofs, src_elem);
6150 Node* state = NULL;
6151 address stubAddr;
6152 const char *stubName;
6153
6154 switch(id) {
6155 case vmIntrinsics::_sha_implCompress:
6156 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6157 state = get_state_from_sha_object(sha_obj);
6158 stubAddr = StubRoutines::sha1_implCompress();
6159 stubName = "sha1_implCompress";
6160 break;
6161 case vmIntrinsics::_sha2_implCompress:
6162 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6163 state = get_state_from_sha_object(sha_obj);
6164 stubAddr = StubRoutines::sha256_implCompress();
6165 stubName = "sha256_implCompress";
6166 break;
6167 case vmIntrinsics::_sha5_implCompress:
6168 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6169 state = get_state_from_sha5_object(sha_obj);
6170 stubAddr = StubRoutines::sha512_implCompress();
6171 stubName = "sha512_implCompress";
6172 break;
6173 default:
6174 fatal_unexpected_iid(id);
6175 return false;
6176 }
6177 if (state == NULL) return false;
6178
6179 // Call the stub.
6180 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6181 stubAddr, stubName, TypePtr::BOTTOM,
6182 src_start, state);
6183
6184 return true;
6185 }
6186
6187 //------------------------------inline_digestBase_implCompressMB-----------------------
6188 //
6189 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6190 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6191 //
6192 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6193 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6194 "need SHA1/SHA256/SHA512 instruction support");
6195 assert((uint)predicate < 3, "sanity");
6196 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6197
6198 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6199 Node* src = argument(1); // byte[] array
6200 Node* ofs = argument(2); // type int
6201 Node* limit = argument(3); // type int
6202
6203 const Type* src_type = src->Value(&_gvn);
6204 const TypeAryPtr* top_src = src_type->isa_aryptr();
6205 if (top_src == NULL || top_src->klass() == NULL) {
6206 // failed array check
6207 return false;
6208 }
6209 // Figure out the size and type of the elements we will be copying.
6210 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6211 if (src_elem != T_BYTE) {
6212 return false;
6213 }
6214 // 'src_start' points to src array + offset
6215 Node* src_start = array_element_address(src, ofs, src_elem);
6216
6217 const char* klass_SHA_name = NULL;
6218 const char* stub_name = NULL;
6219 address stub_addr = NULL;
6220 bool long_state = false;
6221
6222 switch (predicate) {
6223 case 0:
6224 if (UseSHA1Intrinsics) {
6225 klass_SHA_name = "sun/security/provider/SHA";
6226 stub_name = "sha1_implCompressMB";
6227 stub_addr = StubRoutines::sha1_implCompressMB();
6228 }
6229 break;
6230 case 1:
6231 if (UseSHA256Intrinsics) {
6232 klass_SHA_name = "sun/security/provider/SHA2";
6233 stub_name = "sha256_implCompressMB";
6234 stub_addr = StubRoutines::sha256_implCompressMB();
6235 }
6236 break;
6237 case 2:
6238 if (UseSHA512Intrinsics) {
6239 klass_SHA_name = "sun/security/provider/SHA5";
6240 stub_name = "sha512_implCompressMB";
6241 stub_addr = StubRoutines::sha512_implCompressMB();
6242 long_state = true;
6243 }
6244 break;
6245 default:
6246 fatal("unknown SHA intrinsic predicate: %d", predicate);
6247 }
6248 if (klass_SHA_name != NULL) {
6249 // get DigestBase klass to lookup for SHA klass
6250 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6251 assert(tinst != NULL, "digestBase_obj is not instance???");
6252 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6253
6254 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6255 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6256 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6257 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6258 }
6259 return false;
6260 }
6261 //------------------------------inline_sha_implCompressMB-----------------------
6262 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6263 bool long_state, address stubAddr, const char *stubName,
6264 Node* src_start, Node* ofs, Node* limit) {
6265 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6266 const TypeOopPtr* xtype = aklass->as_instance_type();
6267 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6268 sha_obj = _gvn.transform(sha_obj);
6269
6270 Node* state;
6271 if (long_state) {
6272 state = get_state_from_sha5_object(sha_obj);
6273 } else {
6274 state = get_state_from_sha_object(sha_obj);
6275 }
6276 if (state == NULL) return false;
6277
6278 // Call the stub.
6279 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6280 OptoRuntime::digestBase_implCompressMB_Type(),
6281 stubAddr, stubName, TypePtr::BOTTOM,
6282 src_start, state, ofs, limit);
6283 // return ofs (int)
6284 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6285 set_result(result);
6286
6287 return true;
6288 }
6289
6290 //------------------------------get_state_from_sha_object-----------------------
6291 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6292 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6293 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6294 if (sha_state == NULL) return (Node *) NULL;
6295
6296 // now have the array, need to get the start address of the state array
6297 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6298 return state;
6299 }
6300
6301 //------------------------------get_state_from_sha5_object-----------------------
6302 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6303 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6304 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6305 if (sha_state == NULL) return (Node *) NULL;
6306
6307 // now have the array, need to get the start address of the state array
6308 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6309 return state;
6310 }
6311
6312 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6313 // Return node representing slow path of predicate check.
6314 // the pseudo code we want to emulate with this predicate is:
6315 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6316 //
6317 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6318 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6319 "need SHA1/SHA256/SHA512 instruction support");
6320 assert((uint)predicate < 3, "sanity");
6321
6322 // The receiver was checked for NULL already.
6323 Node* digestBaseObj = argument(0);
6324
6325 // get DigestBase klass for instanceOf check
6326 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6327 assert(tinst != NULL, "digestBaseObj is null");
6328 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6329
6330 const char* klass_SHA_name = NULL;
6331 switch (predicate) {
6332 case 0:
6333 if (UseSHA1Intrinsics) {
6334 // we want to do an instanceof comparison against the SHA class
6335 klass_SHA_name = "sun/security/provider/SHA";
6336 }
6337 break;
6338 case 1:
6339 if (UseSHA256Intrinsics) {
6340 // we want to do an instanceof comparison against the SHA2 class
6341 klass_SHA_name = "sun/security/provider/SHA2";
6342 }
6343 break;
6344 case 2:
6345 if (UseSHA512Intrinsics) {
6346 // we want to do an instanceof comparison against the SHA5 class
6347 klass_SHA_name = "sun/security/provider/SHA5";
6348 }
6349 break;
6350 default:
6351 fatal("unknown SHA intrinsic predicate: %d", predicate);
6352 }
6353
6354 ciKlass* klass_SHA = NULL;
6355 if (klass_SHA_name != NULL) {
6356 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6357 }
6358 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6359 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6360 Node* ctrl = control();
6361 set_control(top()); // no intrinsic path
6362 return ctrl;
6363 }
6364 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6365
6366 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6367 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6368 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6369 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6370
6371 return instof_false; // even if it is NULL
6372 }
6373
6374 bool LibraryCallKit::inline_profileBoolean() {
6375 Node* counts = argument(1);
6376 const TypeAryPtr* ary = NULL;
6377 ciArray* aobj = NULL;
6378 if (counts->is_Con()
6379 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6380 && (aobj = ary->const_oop()->as_array()) != NULL
6381 && (aobj->length() == 2)) {
6382 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6383 jint false_cnt = aobj->element_value(0).as_int();
6384 jint true_cnt = aobj->element_value(1).as_int();
6385
6386 if (C->log() != NULL) {
6387 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6388 false_cnt, true_cnt);
6389 }
6390
6391 if (false_cnt + true_cnt == 0) {
6392 // According to profile, never executed.
6393 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6394 Deoptimization::Action_reinterpret);
6395 return true;
6396 }
6397
6398 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6399 // is a number of each value occurrences.
6400 Node* result = argument(0);
6401 if (false_cnt == 0 || true_cnt == 0) {
6402 // According to profile, one value has been never seen.
6403 int expected_val = (false_cnt == 0) ? 1 : 0;
6404
6405 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6406 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6407
6408 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6409 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6410 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6411
6412 { // Slow path: uncommon trap for never seen value and then reexecute
6413 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6414 // the value has been seen at least once.
6415 PreserveJVMState pjvms(this);
6416 PreserveReexecuteState preexecs(this);
6417 jvms()->set_should_reexecute(true);
6418
6419 set_control(slow_path);
6420 set_i_o(i_o());
6421
6422 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6423 Deoptimization::Action_reinterpret);
6424 }
6425 // The guard for never seen value enables sharpening of the result and
6426 // returning a constant. It allows to eliminate branches on the same value
6427 // later on.
6428 set_control(fast_path);
6429 result = intcon(expected_val);
6430 }
6431 // Stop profiling.
6432 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6433 // By replacing method body with profile data (represented as ProfileBooleanNode
6434 // on IR level) we effectively disable profiling.
6435 // It enables full speed execution once optimized code is generated.
6436 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6437 C->record_for_igvn(profile);
6438 set_result(profile);
6439 return true;
6440 } else {
6441 // Continue profiling.
6442 // Profile data isn't available at the moment. So, execute method's bytecode version.
6443 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6444 // is compiled and counters aren't available since corresponding MethodHandle
6445 // isn't a compile-time constant.
6446 return false;
6447 }
6448 }
6449
6450 bool LibraryCallKit::inline_isCompileConstant() {
6451 Node* n = argument(0);
6452 set_result(n->is_Con() ? intcon(1) : intcon(0));
6453 return true;
6454 }