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);
204 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
205 bool inline_string_compareTo();
206 bool inline_string_indexOf();
207 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
208 bool inline_string_equals();
209 Node* round_double_node(Node* n);
210 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
211 bool inline_math_native(vmIntrinsics::ID id);
212 bool inline_trig(vmIntrinsics::ID id);
213 bool inline_math(vmIntrinsics::ID id);
214 template <typename OverflowOp>
215 bool inline_math_overflow(Node* arg1, Node* arg2);
216 void inline_math_mathExact(Node* math, Node* test);
217 bool inline_math_addExactI(bool is_increment);
218 bool inline_math_addExactL(bool is_increment);
219 bool inline_math_multiplyExactI();
220 bool inline_math_multiplyExactL();
221 bool inline_math_negateExactI();
222 bool inline_math_negateExactL();
223 bool inline_math_subtractExactI(bool is_decrement);
224 bool inline_math_subtractExactL(bool is_decrement);
225 bool inline_pow();
226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
227 bool inline_min_max(vmIntrinsics::ID id);
228 bool inline_notify(vmIntrinsics::ID id);
229 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
230 // This returns Type::AnyPtr, RawPtr, or OopPtr.
231 int classify_unsafe_addr(Node* &base, Node* &offset);
232 Node* make_unsafe_address(Node* base, Node* offset);
233 // Helper for inline_unsafe_access.
234 // Generates the guards that check whether the result of
235 // Unsafe.getObject should be recorded in an SATB log buffer.
236 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
237 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
238 static bool klass_needs_init_guard(Node* kls);
239 bool inline_unsafe_allocate();
240 bool inline_unsafe_copyMemory();
241 bool inline_native_currentThread();
242 #ifdef TRACE_HAVE_INTRINSICS
243 bool inline_native_classID();
244 bool inline_native_threadID();
245 #endif
246 bool inline_native_time_funcs(address method, const char* funcName);
247 bool inline_native_isInterrupted();
248 bool inline_native_Class_query(vmIntrinsics::ID id);
249 bool inline_native_subtype_check();
250
251 bool inline_native_newArray();
252 bool inline_native_getLength();
253 bool inline_array_copyOf(bool is_copyOfRange);
254 bool inline_array_equals();
255 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
256 bool inline_native_clone(bool is_virtual);
257 bool inline_native_Reflection_getCallerClass();
258 // Helper function for inlining native object hash method
259 bool inline_native_hashcode(bool is_virtual, bool is_static);
260 bool inline_native_getClass();
261
262 // Helper functions for inlining arraycopy
263 bool inline_arraycopy();
264 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
265 RegionNode* slow_region);
266 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
267 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
268
269 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
270 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
271 bool inline_unsafe_ordered_store(BasicType type);
272 bool inline_unsafe_fence(vmIntrinsics::ID id);
273 bool inline_fp_conversions(vmIntrinsics::ID id);
274 bool inline_number_methods(vmIntrinsics::ID id);
275 bool inline_reference_get();
276 bool inline_Class_cast();
277 bool inline_aescrypt_Block(vmIntrinsics::ID id);
278 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
279 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
280 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
281 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
282 bool inline_ghash_processBlocks();
283 bool inline_sha_implCompress(vmIntrinsics::ID id);
284 bool inline_digestBase_implCompressMB(int predicate);
285 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
286 bool long_state, address stubAddr, const char *stubName,
287 Node* src_start, Node* ofs, Node* limit);
288 Node* get_state_from_sha_object(Node *sha_object);
289 Node* get_state_from_sha5_object(Node *sha_object);
290 Node* inline_digestBase_implCompressMB_predicate(int predicate);
291 bool inline_encodeISOArray();
292 bool inline_updateCRC32();
293 bool inline_updateBytesCRC32();
294 bool inline_updateByteBufferCRC32();
295 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
296 bool inline_updateBytesCRC32C();
297 bool inline_updateDirectByteBufferCRC32C();
298 bool inline_updateBytesAdler32();
299 bool inline_updateByteBufferAdler32();
300 bool inline_multiplyToLen();
301 bool inline_squareToLen();
302 bool inline_mulAdd();
303 bool inline_montgomeryMultiply();
304 bool inline_montgomerySquare();
305
306 bool inline_profileBoolean();
307 bool inline_isCompileConstant();
308 };
309
310 //---------------------------make_vm_intrinsic----------------------------
311 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
312 vmIntrinsics::ID id = m->intrinsic_id();
313 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
314
315 if (!m->is_loaded()) {
316 // Do not attempt to inline unloaded methods.
317 return NULL;
318 }
319
320 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
321 bool is_available = false;
322
323 {
324 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
325 // the compiler must transition to '_thread_in_vm' state because both
326 // methods access VM-internal data.
327 VM_ENTRY_MARK;
328 methodHandle mh(THREAD, m->get_Method());
329 methodHandle ct(THREAD, method()->get_Method());
330 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
331 !vmIntrinsics::is_disabled_by_flags(mh, ct);
332 }
333
334 if (is_available) {
335 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
336 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
337 return new LibraryIntrinsic(m, is_virtual,
338 vmIntrinsics::predicates_needed(id),
339 vmIntrinsics::does_virtual_dispatch(id),
340 (vmIntrinsics::ID) id);
341 } else {
342 return NULL;
343 }
344 }
345
346 //----------------------register_library_intrinsics-----------------------
347 // Initialize this file's data structures, for each Compile instance.
348 void Compile::register_library_intrinsics() {
349 // Nothing to do here.
350 }
351
352 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
353 LibraryCallKit kit(jvms, this);
354 Compile* C = kit.C;
355 int nodes = C->unique();
356 #ifndef PRODUCT
357 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
358 char buf[1000];
359 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
360 tty->print_cr("Intrinsic %s", str);
361 }
362 #endif
363 ciMethod* callee = kit.callee();
364 const int bci = kit.bci();
365
366 // Try to inline the intrinsic.
367 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
368 kit.try_to_inline(_last_predicate)) {
369 if (C->print_intrinsics() || C->print_inlining()) {
370 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
371 }
372 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
373 if (C->log()) {
374 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
375 vmIntrinsics::name_at(intrinsic_id()),
376 (is_virtual() ? " virtual='1'" : ""),
377 C->unique() - nodes);
378 }
379 // Push the result from the inlined method onto the stack.
380 kit.push_result();
381 C->print_inlining_update(this);
382 return kit.transfer_exceptions_into_jvms();
383 }
384
385 // The intrinsic bailed out
386 if (C->print_intrinsics() || C->print_inlining()) {
387 if (jvms->has_method()) {
388 // Not a root compile.
389 const char* msg;
390 if (callee->intrinsic_candidate()) {
391 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
392 } else {
393 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
394 : "failed to inline (intrinsic), method not annotated";
395 }
396 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
397 } else {
398 // Root compile
399 tty->print("Did not generate intrinsic %s%s at bci:%d in",
400 vmIntrinsics::name_at(intrinsic_id()),
401 (is_virtual() ? " (virtual)" : ""), bci);
402 }
403 }
404 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
405 C->print_inlining_update(this);
406 return NULL;
407 }
408
409 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
410 LibraryCallKit kit(jvms, this);
411 Compile* C = kit.C;
412 int nodes = C->unique();
413 _last_predicate = predicate;
414 #ifndef PRODUCT
415 assert(is_predicated() && predicate < predicates_count(), "sanity");
416 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
417 char buf[1000];
418 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
419 tty->print_cr("Predicate for intrinsic %s", str);
420 }
421 #endif
422 ciMethod* callee = kit.callee();
423 const int bci = kit.bci();
424
425 Node* slow_ctl = kit.try_to_predicate(predicate);
426 if (!kit.failing()) {
427 if (C->print_intrinsics() || C->print_inlining()) {
428 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
429 }
430 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
431 if (C->log()) {
432 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
433 vmIntrinsics::name_at(intrinsic_id()),
434 (is_virtual() ? " virtual='1'" : ""),
435 C->unique() - nodes);
436 }
437 return slow_ctl; // Could be NULL if the check folds.
438 }
439
440 // The intrinsic bailed out
441 if (C->print_intrinsics() || C->print_inlining()) {
442 if (jvms->has_method()) {
443 // Not a root compile.
444 const char* msg = "failed to generate predicate for intrinsic";
445 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
446 } else {
447 // Root compile
448 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
449 vmIntrinsics::name_at(intrinsic_id()),
450 (is_virtual() ? " (virtual)" : ""), bci);
451 }
452 }
453 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
454 return NULL;
455 }
456
457 bool LibraryCallKit::try_to_inline(int predicate) {
458 // Handle symbolic names for otherwise undistinguished boolean switches:
459 const bool is_store = true;
460 const bool is_native_ptr = true;
461 const bool is_static = true;
462 const bool is_volatile = true;
463
464 if (!jvms()->has_method()) {
465 // Root JVMState has a null method.
466 assert(map()->memory()->Opcode() == Op_Parm, "");
467 // Insert the memory aliasing node
468 set_all_memory(reset_memory());
469 }
470 assert(merged_memory(), "");
471
472
473 switch (intrinsic_id()) {
474 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
475 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
476 case vmIntrinsics::_getClass: return inline_native_getClass();
477
478 case vmIntrinsics::_dsin:
479 case vmIntrinsics::_dcos:
480 case vmIntrinsics::_dtan:
481 case vmIntrinsics::_dabs:
482 case vmIntrinsics::_datan2:
483 case vmIntrinsics::_dsqrt:
484 case vmIntrinsics::_dexp:
485 case vmIntrinsics::_dlog:
486 case vmIntrinsics::_dlog10:
487 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
488
489 case vmIntrinsics::_min:
490 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
491
492 case vmIntrinsics::_notify:
493 case vmIntrinsics::_notifyAll:
494 if (InlineNotify) {
495 return inline_notify(intrinsic_id());
496 }
497 return false;
498
499 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
500 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
501 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
502 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
503 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
504 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
505 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
506 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
507 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
508 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
509 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
510 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
511
512 case vmIntrinsics::_arraycopy: return inline_arraycopy();
513
514 case vmIntrinsics::_compareTo: return inline_string_compareTo();
515 case vmIntrinsics::_indexOf: return inline_string_indexOf();
516 case vmIntrinsics::_equals: return inline_string_equals();
517
518 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
519 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
520 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
521 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
522 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
523 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
524 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
525 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
526 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
527 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
528 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
529 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
530 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
531 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
532 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
533 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
534 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
535 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
536
537 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
538 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
539 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
540 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
541 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
542 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
543 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
544 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
545
546 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
547 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
548 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
549 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
550 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
551 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
552 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
553 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
554
555 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
556 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
557 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
558 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
559 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
560 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
561 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
562 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
563 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
564
565 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
566 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
567 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
568 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
569 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
570 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
571 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
572 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
573 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
574
575 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
576 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
577 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
578 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
579
580 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
581 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
582 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
583 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
584
585 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
586 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
587 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
588
589 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
590 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
591 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
592
593 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
594 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
595 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
596 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
597 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
598
599 case vmIntrinsics::_loadFence:
600 case vmIntrinsics::_storeFence:
601 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
602
603 case vmIntrinsics::_currentThread: return inline_native_currentThread();
604 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
605
606 #ifdef TRACE_HAVE_INTRINSICS
607 case vmIntrinsics::_classID: return inline_native_classID();
608 case vmIntrinsics::_threadID: return inline_native_threadID();
609 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
610 #endif
611 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
612 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
613 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
614 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
615 case vmIntrinsics::_newArray: return inline_native_newArray();
616 case vmIntrinsics::_getLength: return inline_native_getLength();
617 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
618 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
619 case vmIntrinsics::_equalsC: return inline_array_equals();
620 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
621
622 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
623
624 case vmIntrinsics::_isInstance:
625 case vmIntrinsics::_getModifiers:
626 case vmIntrinsics::_isInterface:
627 case vmIntrinsics::_isArray:
628 case vmIntrinsics::_isPrimitive:
629 case vmIntrinsics::_getSuperclass:
630 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
631
632 case vmIntrinsics::_floatToRawIntBits:
633 case vmIntrinsics::_floatToIntBits:
634 case vmIntrinsics::_intBitsToFloat:
635 case vmIntrinsics::_doubleToRawLongBits:
636 case vmIntrinsics::_doubleToLongBits:
637 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
638
639 case vmIntrinsics::_numberOfLeadingZeros_i:
640 case vmIntrinsics::_numberOfLeadingZeros_l:
641 case vmIntrinsics::_numberOfTrailingZeros_i:
642 case vmIntrinsics::_numberOfTrailingZeros_l:
643 case vmIntrinsics::_bitCount_i:
644 case vmIntrinsics::_bitCount_l:
645 case vmIntrinsics::_reverseBytes_i:
646 case vmIntrinsics::_reverseBytes_l:
647 case vmIntrinsics::_reverseBytes_s:
648 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
649
650 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
651
652 case vmIntrinsics::_Reference_get: return inline_reference_get();
653
654 case vmIntrinsics::_Class_cast: return inline_Class_cast();
655
656 case vmIntrinsics::_aescrypt_encryptBlock:
657 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
658
659 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
660 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
661 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
662
663 case vmIntrinsics::_sha_implCompress:
664 case vmIntrinsics::_sha2_implCompress:
665 case vmIntrinsics::_sha5_implCompress:
666 return inline_sha_implCompress(intrinsic_id());
667
668 case vmIntrinsics::_digestBase_implCompressMB:
669 return inline_digestBase_implCompressMB(predicate);
670
671 case vmIntrinsics::_multiplyToLen:
672 return inline_multiplyToLen();
673
674 case vmIntrinsics::_squareToLen:
675 return inline_squareToLen();
676
677 case vmIntrinsics::_mulAdd:
678 return inline_mulAdd();
679
680 case vmIntrinsics::_montgomeryMultiply:
681 return inline_montgomeryMultiply();
682 case vmIntrinsics::_montgomerySquare:
683 return inline_montgomerySquare();
684
685 case vmIntrinsics::_ghash_processBlocks:
686 return inline_ghash_processBlocks();
687
688 case vmIntrinsics::_encodeISOArray:
689 return inline_encodeISOArray();
690
691 case vmIntrinsics::_updateCRC32:
692 return inline_updateCRC32();
693 case vmIntrinsics::_updateBytesCRC32:
694 return inline_updateBytesCRC32();
695 case vmIntrinsics::_updateByteBufferCRC32:
696 return inline_updateByteBufferCRC32();
697
698 case vmIntrinsics::_updateBytesCRC32C:
699 return inline_updateBytesCRC32C();
700 case vmIntrinsics::_updateDirectByteBufferCRC32C:
701 return inline_updateDirectByteBufferCRC32C();
702
703 case vmIntrinsics::_updateBytesAdler32:
704 return inline_updateBytesAdler32();
705 case vmIntrinsics::_updateByteBufferAdler32:
706 return inline_updateByteBufferAdler32();
707
708 case vmIntrinsics::_profileBoolean:
709 return inline_profileBoolean();
710 case vmIntrinsics::_isCompileConstant:
711 return inline_isCompileConstant();
712
713 default:
714 // If you get here, it may be that someone has added a new intrinsic
715 // to the list in vmSymbols.hpp without implementing it here.
716 #ifndef PRODUCT
717 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
718 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
719 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
720 }
721 #endif
722 return false;
723 }
724 }
725
726 Node* LibraryCallKit::try_to_predicate(int predicate) {
727 if (!jvms()->has_method()) {
728 // Root JVMState has a null method.
729 assert(map()->memory()->Opcode() == Op_Parm, "");
730 // Insert the memory aliasing node
731 set_all_memory(reset_memory());
732 }
733 assert(merged_memory(), "");
734
735 switch (intrinsic_id()) {
736 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
737 return inline_cipherBlockChaining_AESCrypt_predicate(false);
738 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
739 return inline_cipherBlockChaining_AESCrypt_predicate(true);
740 case vmIntrinsics::_digestBase_implCompressMB:
741 return inline_digestBase_implCompressMB_predicate(predicate);
742
743 default:
744 // If you get here, it may be that someone has added a new intrinsic
745 // to the list in vmSymbols.hpp without implementing it here.
746 #ifndef PRODUCT
747 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
748 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
749 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
750 }
751 #endif
752 Node* slow_ctl = control();
753 set_control(top()); // No fast path instrinsic
754 return slow_ctl;
755 }
756 }
757
758 //------------------------------set_result-------------------------------
759 // Helper function for finishing intrinsics.
760 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
761 record_for_igvn(region);
762 set_control(_gvn.transform(region));
763 set_result( _gvn.transform(value));
764 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
765 }
766
767 //------------------------------generate_guard---------------------------
768 // Helper function for generating guarded fast-slow graph structures.
769 // The given 'test', if true, guards a slow path. If the test fails
770 // then a fast path can be taken. (We generally hope it fails.)
771 // In all cases, GraphKit::control() is updated to the fast path.
772 // The returned value represents the control for the slow path.
773 // The return value is never 'top'; it is either a valid control
774 // or NULL if it is obvious that the slow path can never be taken.
775 // Also, if region and the slow control are not NULL, the slow edge
776 // is appended to the region.
777 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
778 if (stopped()) {
779 // Already short circuited.
780 return NULL;
781 }
782
783 // Build an if node and its projections.
784 // If test is true we take the slow path, which we assume is uncommon.
785 if (_gvn.type(test) == TypeInt::ZERO) {
786 // The slow branch is never taken. No need to build this guard.
787 return NULL;
788 }
789
790 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
791
792 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
793 if (if_slow == top()) {
794 // The slow branch is never taken. No need to build this guard.
795 return NULL;
796 }
797
798 if (region != NULL)
799 region->add_req(if_slow);
800
801 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
802 set_control(if_fast);
803
804 return if_slow;
805 }
806
807 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
808 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
809 }
810 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
811 return generate_guard(test, region, PROB_FAIR);
812 }
813
814 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
815 Node* *pos_index) {
816 if (stopped())
817 return NULL; // already stopped
818 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
819 return NULL; // index is already adequately typed
820 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
821 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
822 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
823 if (is_neg != NULL && pos_index != NULL) {
824 // Emulate effect of Parse::adjust_map_after_if.
825 Node* ccast = new CastIINode(index, TypeInt::POS);
826 ccast->set_req(0, control());
827 (*pos_index) = _gvn.transform(ccast);
828 }
829 return is_neg;
830 }
831
832 // Make sure that 'position' is a valid limit index, in [0..length].
833 // There are two equivalent plans for checking this:
834 // A. (offset + copyLength) unsigned<= arrayLength
835 // B. offset <= (arrayLength - copyLength)
836 // We require that all of the values above, except for the sum and
837 // difference, are already known to be non-negative.
838 // Plan A is robust in the face of overflow, if offset and copyLength
839 // are both hugely positive.
840 //
841 // Plan B is less direct and intuitive, but it does not overflow at
842 // all, since the difference of two non-negatives is always
843 // representable. Whenever Java methods must perform the equivalent
844 // check they generally use Plan B instead of Plan A.
845 // For the moment we use Plan A.
846 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
847 Node* subseq_length,
848 Node* array_length,
849 RegionNode* region) {
850 if (stopped())
851 return NULL; // already stopped
852 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
853 if (zero_offset && subseq_length->eqv_uncast(array_length))
854 return NULL; // common case of whole-array copy
855 Node* last = subseq_length;
856 if (!zero_offset) // last += offset
857 last = _gvn.transform(new AddINode(last, offset));
858 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
859 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
860 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
861 return is_over;
862 }
863
864
865 //--------------------------generate_current_thread--------------------
866 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
867 ciKlass* thread_klass = env()->Thread_klass();
868 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
869 Node* thread = _gvn.transform(new ThreadLocalNode());
870 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
871 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
872 tls_output = thread;
873 return threadObj;
874 }
875
876
877 //------------------------------make_string_method_node------------------------
878 // Helper method for String intrinsic functions. This version is called
879 // with str1 and str2 pointing to String object nodes.
880 //
881 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
882 Node* no_ctrl = NULL;
883
884 // Get start addr of string
885 Node* str1_value = load_String_value(no_ctrl, str1);
886 Node* str1_offset = load_String_offset(no_ctrl, str1);
887 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
888
889 // Get length of string 1
890 Node* str1_len = load_String_length(no_ctrl, str1);
891
892 Node* str2_value = load_String_value(no_ctrl, str2);
893 Node* str2_offset = load_String_offset(no_ctrl, str2);
894 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
895
896 Node* str2_len = NULL;
897 Node* result = NULL;
898
899 switch (opcode) {
900 case Op_StrIndexOf:
901 // Get length of string 2
902 str2_len = load_String_length(no_ctrl, str2);
903
904 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
905 str1_start, str1_len, str2_start, str2_len);
906 break;
907 case Op_StrComp:
908 // Get length of string 2
909 str2_len = load_String_length(no_ctrl, str2);
910
911 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
912 str1_start, str1_len, str2_start, str2_len);
913 break;
914 case Op_StrEquals:
915 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
916 str1_start, str2_start, str1_len);
917 break;
918 default:
919 ShouldNotReachHere();
920 return NULL;
921 }
922
923 // All these intrinsics have checks.
924 C->set_has_split_ifs(true); // Has chance for split-if optimization
925
926 return _gvn.transform(result);
927 }
928
929 // Helper method for String intrinsic functions. This version is called
930 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
931 // to Int nodes containing the lenghts of str1 and str2.
932 //
933 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
934 Node* result = NULL;
935 switch (opcode) {
936 case Op_StrIndexOf:
937 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
938 str1_start, cnt1, str2_start, cnt2);
939 break;
940 case Op_StrComp:
941 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
942 str1_start, cnt1, str2_start, cnt2);
943 break;
944 case Op_StrEquals:
945 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
946 str1_start, str2_start, cnt1);
947 break;
948 default:
949 ShouldNotReachHere();
950 return NULL;
951 }
952
953 // All these intrinsics have checks.
954 C->set_has_split_ifs(true); // Has chance for split-if optimization
955
956 return _gvn.transform(result);
957 }
958
959 //------------------------------inline_string_compareTo------------------------
960 // public int java.lang.String.compareTo(String anotherString);
961 bool LibraryCallKit::inline_string_compareTo() {
962 Node* receiver = null_check(argument(0));
963 Node* arg = null_check(argument(1));
964 if (stopped()) {
965 return true;
966 }
967 set_result(make_string_method_node(Op_StrComp, receiver, arg));
968 return true;
969 }
970
971 //------------------------------inline_string_equals------------------------
972 bool LibraryCallKit::inline_string_equals() {
973 Node* receiver = null_check_receiver();
974 // NOTE: Do not null check argument for String.equals() because spec
975 // allows to specify NULL as argument.
976 Node* argument = this->argument(1);
977 if (stopped()) {
978 return true;
979 }
980
981 // paths (plus control) merge
982 RegionNode* region = new RegionNode(5);
983 Node* phi = new PhiNode(region, TypeInt::BOOL);
984
985 // does source == target string?
986 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
987 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
988
989 Node* if_eq = generate_slow_guard(bol, NULL);
990 if (if_eq != NULL) {
991 // receiver == argument
992 phi->init_req(2, intcon(1));
993 region->init_req(2, if_eq);
994 }
995
996 // get String klass for instanceOf
997 ciInstanceKlass* klass = env()->String_klass();
998
999 if (!stopped()) {
1000 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1001 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1002 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1003
1004 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1005 //instanceOf == true, fallthrough
1006
1007 if (inst_false != NULL) {
1008 phi->init_req(3, intcon(0));
1009 region->init_req(3, inst_false);
1010 }
1011 }
1012
1013 if (!stopped()) {
1014 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1015
1016 // Properly cast the argument to String
1017 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1018 // This path is taken only when argument's type is String:NotNull.
1019 argument = cast_not_null(argument, false);
1020
1021 Node* no_ctrl = NULL;
1022
1023 // Get start addr of receiver
1024 Node* receiver_val = load_String_value(no_ctrl, receiver);
1025 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1026 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1027
1028 // Get length of receiver
1029 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1030
1031 // Get start addr of argument
1032 Node* argument_val = load_String_value(no_ctrl, argument);
1033 Node* argument_offset = load_String_offset(no_ctrl, argument);
1034 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1035
1036 // Get length of argument
1037 Node* argument_cnt = load_String_length(no_ctrl, argument);
1038
1039 // Check for receiver count != argument count
1040 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1041 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1042 Node* if_ne = generate_slow_guard(bol, NULL);
1043 if (if_ne != NULL) {
1044 phi->init_req(4, intcon(0));
1045 region->init_req(4, if_ne);
1046 }
1047
1048 // Check for count == 0 is done by assembler code for StrEquals.
1049
1050 if (!stopped()) {
1051 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1052 phi->init_req(1, equals);
1053 region->init_req(1, control());
1054 }
1055 }
1056
1057 // post merge
1058 set_control(_gvn.transform(region));
1059 record_for_igvn(region);
1060
1061 set_result(_gvn.transform(phi));
1062 return true;
1063 }
1064
1065 //------------------------------inline_array_equals----------------------------
1066 bool LibraryCallKit::inline_array_equals() {
1067 Node* arg1 = argument(0);
1068 Node* arg2 = argument(1);
1069 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1070 return true;
1071 }
1072
1073 // Java version of String.indexOf(constant string)
1074 // class StringDecl {
1075 // StringDecl(char[] ca) {
1076 // offset = 0;
1077 // count = ca.length;
1078 // value = ca;
1079 // }
1080 // int offset;
1081 // int count;
1082 // char[] value;
1083 // }
1084 //
1085 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1086 // int targetOffset, int cache_i, int md2) {
1087 // int cache = cache_i;
1088 // int sourceOffset = string_object.offset;
1089 // int sourceCount = string_object.count;
1090 // int targetCount = target_object.length;
1091 //
1092 // int targetCountLess1 = targetCount - 1;
1093 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1094 //
1095 // char[] source = string_object.value;
1096 // char[] target = target_object;
1097 // int lastChar = target[targetCountLess1];
1098 //
1099 // outer_loop:
1100 // for (int i = sourceOffset; i < sourceEnd; ) {
1101 // int src = source[i + targetCountLess1];
1102 // if (src == lastChar) {
1103 // // With random strings and a 4-character alphabet,
1104 // // reverse matching at this point sets up 0.8% fewer
1105 // // frames, but (paradoxically) makes 0.3% more probes.
1106 // // Since those probes are nearer the lastChar probe,
1107 // // there is may be a net D$ win with reverse matching.
1108 // // But, reversing loop inhibits unroll of inner loop
1109 // // for unknown reason. So, does running outer loop from
1110 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1111 // for (int j = 0; j < targetCountLess1; j++) {
1112 // if (target[targetOffset + j] != source[i+j]) {
1113 // if ((cache & (1 << source[i+j])) == 0) {
1114 // if (md2 < j+1) {
1115 // i += j+1;
1116 // continue outer_loop;
1117 // }
1118 // }
1119 // i += md2;
1120 // continue outer_loop;
1121 // }
1122 // }
1123 // return i - sourceOffset;
1124 // }
1125 // if ((cache & (1 << src)) == 0) {
1126 // i += targetCountLess1;
1127 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1128 // i++;
1129 // }
1130 // return -1;
1131 // }
1132
1133 //------------------------------string_indexOf------------------------
1134 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1135 jint cache_i, jint md2_i) {
1136
1137 Node* no_ctrl = NULL;
1138 float likely = PROB_LIKELY(0.9);
1139 float unlikely = PROB_UNLIKELY(0.9);
1140
1141 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1142
1143 Node* source = load_String_value(no_ctrl, string_object);
1144 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1145 Node* sourceCount = load_String_length(no_ctrl, string_object);
1146
1147 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1148 jint target_length = target_array->length();
1149 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1150 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1151
1152 // String.value field is known to be @Stable.
1153 if (UseImplicitStableValues) {
1154 target = cast_array_to_stable(target, target_type);
1155 }
1156
1157 IdealKit kit(this, false, true);
1158 #define __ kit.
1159 Node* zero = __ ConI(0);
1160 Node* one = __ ConI(1);
1161 Node* cache = __ ConI(cache_i);
1162 Node* md2 = __ ConI(md2_i);
1163 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1164 Node* targetCountLess1 = __ ConI(target_length - 1);
1165 Node* targetOffset = __ ConI(targetOffset_i);
1166 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1167
1168 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1169 Node* outer_loop = __ make_label(2 /* goto */);
1170 Node* return_ = __ make_label(1);
1171
1172 __ set(rtn,__ ConI(-1));
1173 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1174 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1175 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1176 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1177 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1178 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1179 Node* tpj = __ AddI(targetOffset, __ value(j));
1180 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1181 Node* ipj = __ AddI(__ value(i), __ value(j));
1182 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1183 __ if_then(targ, BoolTest::ne, src2); {
1184 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1185 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1186 __ increment(i, __ AddI(__ value(j), one));
1187 __ goto_(outer_loop);
1188 } __ end_if(); __ dead(j);
1189 }__ end_if(); __ dead(j);
1190 __ increment(i, md2);
1191 __ goto_(outer_loop);
1192 }__ end_if();
1193 __ increment(j, one);
1194 }__ end_loop(); __ dead(j);
1195 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1196 __ goto_(return_);
1197 }__ end_if();
1198 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1199 __ increment(i, targetCountLess1);
1200 }__ end_if();
1201 __ increment(i, one);
1202 __ bind(outer_loop);
1203 }__ end_loop(); __ dead(i);
1204 __ bind(return_);
1205
1206 // Final sync IdealKit and GraphKit.
1207 final_sync(kit);
1208 Node* result = __ value(rtn);
1209 #undef __
1210 C->set_has_loops(true);
1211 return result;
1212 }
1213
1214 //------------------------------inline_string_indexOf------------------------
1215 bool LibraryCallKit::inline_string_indexOf() {
1216 Node* receiver = argument(0);
1217 Node* arg = argument(1);
1218
1219 Node* result;
1220 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1221 UseSSE42Intrinsics) {
1222 // Generate SSE4.2 version of indexOf
1223 // We currently only have match rules that use SSE4.2
1224
1225 receiver = null_check(receiver);
1226 arg = null_check(arg);
1227 if (stopped()) {
1228 return true;
1229 }
1230
1231 // Make the merge point
1232 RegionNode* result_rgn = new RegionNode(4);
1233 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1234 Node* no_ctrl = NULL;
1235
1236 // Get start addr of source string
1237 Node* source = load_String_value(no_ctrl, receiver);
1238 Node* source_offset = load_String_offset(no_ctrl, receiver);
1239 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1240
1241 // Get length of source string
1242 Node* source_cnt = load_String_length(no_ctrl, receiver);
1243
1244 // Get start addr of substring
1245 Node* substr = load_String_value(no_ctrl, arg);
1246 Node* substr_offset = load_String_offset(no_ctrl, arg);
1247 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1248
1249 // Get length of source string
1250 Node* substr_cnt = load_String_length(no_ctrl, arg);
1251
1252 // Check for substr count > string count
1253 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1254 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1255 Node* if_gt = generate_slow_guard(bol, NULL);
1256 if (if_gt != NULL) {
1257 result_phi->init_req(2, intcon(-1));
1258 result_rgn->init_req(2, if_gt);
1259 }
1260
1261 if (!stopped()) {
1262 // Check for substr count == 0
1263 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1264 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1265 Node* if_zero = generate_slow_guard(bol, NULL);
1266 if (if_zero != NULL) {
1267 result_phi->init_req(3, intcon(0));
1268 result_rgn->init_req(3, if_zero);
1269 }
1270 }
1271
1272 if (!stopped()) {
1273 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1274 result_phi->init_req(1, result);
1275 result_rgn->init_req(1, control());
1276 }
1277 set_control(_gvn.transform(result_rgn));
1278 record_for_igvn(result_rgn);
1279 result = _gvn.transform(result_phi);
1280
1281 } else { // Use LibraryCallKit::string_indexOf
1282 // don't intrinsify if argument isn't a constant string.
1283 if (!arg->is_Con()) {
1284 return false;
1285 }
1286 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1287 if (str_type == NULL) {
1288 return false;
1289 }
1290 ciInstanceKlass* klass = env()->String_klass();
1291 ciObject* str_const = str_type->const_oop();
1292 if (str_const == NULL || str_const->klass() != klass) {
1293 return false;
1294 }
1295 ciInstance* str = str_const->as_instance();
1296 assert(str != NULL, "must be instance");
1297
1298 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1299 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1300
1301 int o;
1302 int c;
1303 if (java_lang_String::has_offset_field()) {
1304 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1305 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1306 } else {
1307 o = 0;
1308 c = pat->length();
1309 }
1310
1311 // constant strings have no offset and count == length which
1312 // simplifies the resulting code somewhat so lets optimize for that.
1313 if (o != 0 || c != pat->length()) {
1314 return false;
1315 }
1316
1317 receiver = null_check(receiver, T_OBJECT);
1318 // NOTE: No null check on the argument is needed since it's a constant String oop.
1319 if (stopped()) {
1320 return true;
1321 }
1322
1323 // The null string as a pattern always returns 0 (match at beginning of string)
1324 if (c == 0) {
1325 set_result(intcon(0));
1326 return true;
1327 }
1328
1329 // Generate default indexOf
1330 jchar lastChar = pat->char_at(o + (c - 1));
1331 int cache = 0;
1332 int i;
1333 for (i = 0; i < c - 1; i++) {
1334 assert(i < pat->length(), "out of range");
1335 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1336 }
1337
1338 int md2 = c;
1339 for (i = 0; i < c - 1; i++) {
1340 assert(i < pat->length(), "out of range");
1341 if (pat->char_at(o + i) == lastChar) {
1342 md2 = (c - 1) - i;
1343 }
1344 }
1345
1346 result = string_indexOf(receiver, pat, o, cache, md2);
1347 }
1348 set_result(result);
1349 return true;
1350 }
1351
1352 //--------------------------round_double_node--------------------------------
1353 // Round a double node if necessary.
1354 Node* LibraryCallKit::round_double_node(Node* n) {
1355 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1356 n = _gvn.transform(new RoundDoubleNode(0, n));
1357 return n;
1358 }
1359
1360 //------------------------------inline_math-----------------------------------
1361 // public static double Math.abs(double)
1362 // public static double Math.sqrt(double)
1363 // public static double Math.log(double)
1364 // public static double Math.log10(double)
1365 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1366 Node* arg = round_double_node(argument(0));
1367 Node* n;
1368 switch (id) {
1369 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1370 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1371 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1372 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1373 default: fatal_unexpected_iid(id); break;
1374 }
1375 set_result(_gvn.transform(n));
1376 return true;
1377 }
1378
1379 //------------------------------inline_trig----------------------------------
1380 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1381 // argument reduction which will turn into a fast/slow diamond.
1382 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1383 Node* arg = round_double_node(argument(0));
1384 Node* n = NULL;
1385
1386 switch (id) {
1387 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1388 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1389 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1390 default: fatal_unexpected_iid(id); break;
1391 }
1392 n = _gvn.transform(n);
1393
1394 // Rounding required? Check for argument reduction!
1395 if (Matcher::strict_fp_requires_explicit_rounding) {
1396 static const double pi_4 = 0.7853981633974483;
1397 static const double neg_pi_4 = -0.7853981633974483;
1398 // pi/2 in 80-bit extended precision
1399 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1400 // -pi/2 in 80-bit extended precision
1401 // 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};
1402 // Cutoff value for using this argument reduction technique
1403 //static const double pi_2_minus_epsilon = 1.564660403643354;
1404 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1405
1406 // Pseudocode for sin:
1407 // if (x <= Math.PI / 4.0) {
1408 // if (x >= -Math.PI / 4.0) return fsin(x);
1409 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1410 // } else {
1411 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1412 // }
1413 // return StrictMath.sin(x);
1414
1415 // Pseudocode for cos:
1416 // if (x <= Math.PI / 4.0) {
1417 // if (x >= -Math.PI / 4.0) return fcos(x);
1418 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1419 // } else {
1420 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1421 // }
1422 // return StrictMath.cos(x);
1423
1424 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1425 // requires a special machine instruction to load it. Instead we'll try
1426 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1427 // probably do the math inside the SIN encoding.
1428
1429 // Make the merge point
1430 RegionNode* r = new RegionNode(3);
1431 Node* phi = new PhiNode(r, Type::DOUBLE);
1432
1433 // Flatten arg so we need only 1 test
1434 Node *abs = _gvn.transform(new AbsDNode(arg));
1435 // Node for PI/4 constant
1436 Node *pi4 = makecon(TypeD::make(pi_4));
1437 // Check PI/4 : abs(arg)
1438 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1439 // Check: If PI/4 < abs(arg) then go slow
1440 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1441 // Branch either way
1442 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1443 set_control(opt_iff(r,iff));
1444
1445 // Set fast path result
1446 phi->init_req(2, n);
1447
1448 // Slow path - non-blocking leaf call
1449 Node* call = NULL;
1450 switch (id) {
1451 case vmIntrinsics::_dsin:
1452 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1453 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1454 "Sin", NULL, arg, top());
1455 break;
1456 case vmIntrinsics::_dcos:
1457 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1458 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1459 "Cos", NULL, arg, top());
1460 break;
1461 case vmIntrinsics::_dtan:
1462 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1463 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1464 "Tan", NULL, arg, top());
1465 break;
1466 }
1467 assert(control()->in(0) == call, "");
1468 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1469 r->init_req(1, control());
1470 phi->init_req(1, slow_result);
1471
1472 // Post-merge
1473 set_control(_gvn.transform(r));
1474 record_for_igvn(r);
1475 n = _gvn.transform(phi);
1476
1477 C->set_has_split_ifs(true); // Has chance for split-if optimization
1478 }
1479 set_result(n);
1480 return true;
1481 }
1482
1483 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1484 //-------------------
1485 //result=(result.isNaN())? funcAddr():result;
1486 // Check: If isNaN() by checking result!=result? then either trap
1487 // or go to runtime
1488 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1489 // Build the boolean node
1490 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1491
1492 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1493 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1494 // The pow or exp intrinsic returned a NaN, which requires a call
1495 // to the runtime. Recompile with the runtime call.
1496 uncommon_trap(Deoptimization::Reason_intrinsic,
1497 Deoptimization::Action_make_not_entrant);
1498 }
1499 return result;
1500 } else {
1501 // If this inlining ever returned NaN in the past, we compile a call
1502 // to the runtime to properly handle corner cases
1503
1504 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1505 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1506 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1507
1508 if (!if_slow->is_top()) {
1509 RegionNode* result_region = new RegionNode(3);
1510 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1511
1512 result_region->init_req(1, if_fast);
1513 result_val->init_req(1, result);
1514
1515 set_control(if_slow);
1516
1517 const TypePtr* no_memory_effects = NULL;
1518 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1519 no_memory_effects,
1520 x, top(), y, y ? top() : NULL);
1521 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1522 #ifdef ASSERT
1523 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1524 assert(value_top == top(), "second value must be top");
1525 #endif
1526
1527 result_region->init_req(2, control());
1528 result_val->init_req(2, value);
1529 set_control(_gvn.transform(result_region));
1530 return _gvn.transform(result_val);
1531 } else {
1532 return result;
1533 }
1534 }
1535 }
1536
1537 //------------------------------inline_pow-------------------------------------
1538 // Inline power instructions, if possible.
1539 bool LibraryCallKit::inline_pow() {
1540 // Pseudocode for pow
1541 // if (y == 2) {
1542 // return x * x;
1543 // } else {
1544 // if (x <= 0.0) {
1545 // long longy = (long)y;
1546 // if ((double)longy == y) { // if y is long
1547 // if (y + 1 == y) longy = 0; // huge number: even
1548 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1549 // } else {
1550 // result = NaN;
1551 // }
1552 // } else {
1553 // result = DPow(x,y);
1554 // }
1555 // if (result != result)? {
1556 // result = uncommon_trap() or runtime_call();
1557 // }
1558 // return result;
1559 // }
1560
1561 Node* x = round_double_node(argument(0));
1562 Node* y = round_double_node(argument(2));
1563
1564 Node* result = NULL;
1565
1566 Node* const_two_node = makecon(TypeD::make(2.0));
1567 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1568 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1569 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1570 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1571 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1572
1573 RegionNode* region_node = new RegionNode(3);
1574 region_node->init_req(1, if_true);
1575
1576 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1577 // special case for x^y where y == 2, we can convert it to x * x
1578 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1579
1580 // set control to if_false since we will now process the false branch
1581 set_control(if_false);
1582
1583 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1584 // Short form: skip the fancy tests and just check for NaN result.
1585 result = _gvn.transform(new PowDNode(C, control(), x, y));
1586 } else {
1587 // If this inlining ever returned NaN in the past, include all
1588 // checks + call to the runtime.
1589
1590 // Set the merge point for If node with condition of (x <= 0.0)
1591 // There are four possible paths to region node and phi node
1592 RegionNode *r = new RegionNode(4);
1593 Node *phi = new PhiNode(r, Type::DOUBLE);
1594
1595 // Build the first if node: if (x <= 0.0)
1596 // Node for 0 constant
1597 Node *zeronode = makecon(TypeD::ZERO);
1598 // Check x:0
1599 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1600 // Check: If (x<=0) then go complex path
1601 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1602 // Branch either way
1603 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1604 // Fast path taken; set region slot 3
1605 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1606 r->init_req(3,fast_taken); // Capture fast-control
1607
1608 // Fast path not-taken, i.e. slow path
1609 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1610
1611 // Set fast path result
1612 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1613 phi->init_req(3, fast_result);
1614
1615 // Complex path
1616 // Build the second if node (if y is long)
1617 // Node for (long)y
1618 Node *longy = _gvn.transform(new ConvD2LNode(y));
1619 // Node for (double)((long) y)
1620 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1621 // Check (double)((long) y) : y
1622 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1623 // Check if (y isn't long) then go to slow path
1624
1625 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1626 // Branch either way
1627 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1628 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1629
1630 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1631
1632 // Calculate DPow(abs(x), y)*(1 & (long)y)
1633 // Node for constant 1
1634 Node *conone = longcon(1);
1635 // 1& (long)y
1636 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1637
1638 // A huge number is always even. Detect a huge number by checking
1639 // if y + 1 == y and set integer to be tested for parity to 0.
1640 // Required for corner case:
1641 // (long)9.223372036854776E18 = max_jlong
1642 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1643 // max_jlong is odd but 9.223372036854776E18 is even
1644 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1645 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1646 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1647 Node* correctedsign = NULL;
1648 if (ConditionalMoveLimit != 0) {
1649 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1650 } else {
1651 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1652 RegionNode *r = new RegionNode(3);
1653 Node *phi = new PhiNode(r, TypeLong::LONG);
1654 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1655 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1656 phi->init_req(1, signnode);
1657 phi->init_req(2, longcon(0));
1658 correctedsign = _gvn.transform(phi);
1659 ylong_path = _gvn.transform(r);
1660 record_for_igvn(r);
1661 }
1662
1663 // zero node
1664 Node *conzero = longcon(0);
1665 // Check (1&(long)y)==0?
1666 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1667 // Check if (1&(long)y)!=0?, if so the result is negative
1668 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1669 // abs(x)
1670 Node *absx=_gvn.transform(new AbsDNode(x));
1671 // abs(x)^y
1672 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1673 // -abs(x)^y
1674 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1675 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1676 Node *signresult = NULL;
1677 if (ConditionalMoveLimit != 0) {
1678 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1679 } else {
1680 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1681 RegionNode *r = new RegionNode(3);
1682 Node *phi = new PhiNode(r, Type::DOUBLE);
1683 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1684 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1685 phi->init_req(1, absxpowy);
1686 phi->init_req(2, negabsxpowy);
1687 signresult = _gvn.transform(phi);
1688 ylong_path = _gvn.transform(r);
1689 record_for_igvn(r);
1690 }
1691 // Set complex path fast result
1692 r->init_req(2, ylong_path);
1693 phi->init_req(2, signresult);
1694
1695 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1696 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1697 r->init_req(1,slow_path);
1698 phi->init_req(1,slow_result);
1699
1700 // Post merge
1701 set_control(_gvn.transform(r));
1702 record_for_igvn(r);
1703 result = _gvn.transform(phi);
1704 }
1705
1706 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1707
1708 // control from finish_pow_exp is now input to the region node
1709 region_node->set_req(2, control());
1710 // the result from finish_pow_exp is now input to the phi node
1711 phi_node->init_req(2, result);
1712 set_control(_gvn.transform(region_node));
1713 record_for_igvn(region_node);
1714 set_result(_gvn.transform(phi_node));
1715
1716 C->set_has_split_ifs(true); // Has chance for split-if optimization
1717 return true;
1718 }
1719
1720 //------------------------------runtime_math-----------------------------
1721 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1722 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1723 "must be (DD)D or (D)D type");
1724
1725 // Inputs
1726 Node* a = round_double_node(argument(0));
1727 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1728
1729 const TypePtr* no_memory_effects = NULL;
1730 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1731 no_memory_effects,
1732 a, top(), b, b ? top() : NULL);
1733 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1734 #ifdef ASSERT
1735 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1736 assert(value_top == top(), "second value must be top");
1737 #endif
1738
1739 set_result(value);
1740 return true;
1741 }
1742
1743 //------------------------------inline_math_native-----------------------------
1744 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1745 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1746 switch (id) {
1747 // These intrinsics are not properly supported on all hardware
1748 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1749 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1750 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1751 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1752 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1753 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1754
1755 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1756 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1757 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1758 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1759
1760 // These intrinsics are supported on all hardware
1761 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1762 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1763
1764 case vmIntrinsics::_dexp:
1765 return StubRoutines::dexp() != NULL ?
1766 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1767 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1768 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1769 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1770 #undef FN_PTR
1771
1772 // These intrinsics are not yet correctly implemented
1773 case vmIntrinsics::_datan2:
1774 return false;
1775
1776 default:
1777 fatal_unexpected_iid(id);
1778 return false;
1779 }
1780 }
1781
1782 static bool is_simple_name(Node* n) {
1783 return (n->req() == 1 // constant
1784 || (n->is_Type() && n->as_Type()->type()->singleton())
1785 || n->is_Proj() // parameter or return value
1786 || n->is_Phi() // local of some sort
1787 );
1788 }
1789
1790 //----------------------------inline_notify-----------------------------------*
1791 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1792 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1793 address func;
1794 if (id == vmIntrinsics::_notify) {
1795 func = OptoRuntime::monitor_notify_Java();
1796 } else {
1797 func = OptoRuntime::monitor_notifyAll_Java();
1798 }
1799 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1800 make_slow_call_ex(call, env()->Throwable_klass(), false);
1801 return true;
1802 }
1803
1804
1805 //----------------------------inline_min_max-----------------------------------
1806 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1807 set_result(generate_min_max(id, argument(0), argument(1)));
1808 return true;
1809 }
1810
1811 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1812 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1813 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1814 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1815 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1816
1817 {
1818 PreserveJVMState pjvms(this);
1819 PreserveReexecuteState preexecs(this);
1820 jvms()->set_should_reexecute(true);
1821
1822 set_control(slow_path);
1823 set_i_o(i_o());
1824
1825 uncommon_trap(Deoptimization::Reason_intrinsic,
1826 Deoptimization::Action_none);
1827 }
1828
1829 set_control(fast_path);
1830 set_result(math);
1831 }
1832
1833 template <typename OverflowOp>
1834 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1835 typedef typename OverflowOp::MathOp MathOp;
1836
1837 MathOp* mathOp = new MathOp(arg1, arg2);
1838 Node* operation = _gvn.transform( mathOp );
1839 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1840 inline_math_mathExact(operation, ofcheck);
1841 return true;
1842 }
1843
1844 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1845 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1846 }
1847
1848 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1849 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1850 }
1851
1852 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1853 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1854 }
1855
1856 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1857 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1858 }
1859
1860 bool LibraryCallKit::inline_math_negateExactI() {
1861 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1862 }
1863
1864 bool LibraryCallKit::inline_math_negateExactL() {
1865 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1866 }
1867
1868 bool LibraryCallKit::inline_math_multiplyExactI() {
1869 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1870 }
1871
1872 bool LibraryCallKit::inline_math_multiplyExactL() {
1873 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1874 }
1875
1876 Node*
1877 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1878 // These are the candidate return value:
1879 Node* xvalue = x0;
1880 Node* yvalue = y0;
1881
1882 if (xvalue == yvalue) {
1883 return xvalue;
1884 }
1885
1886 bool want_max = (id == vmIntrinsics::_max);
1887
1888 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1889 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1890 if (txvalue == NULL || tyvalue == NULL) return top();
1891 // This is not really necessary, but it is consistent with a
1892 // hypothetical MaxINode::Value method:
1893 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1894
1895 // %%% This folding logic should (ideally) be in a different place.
1896 // Some should be inside IfNode, and there to be a more reliable
1897 // transformation of ?: style patterns into cmoves. We also want
1898 // more powerful optimizations around cmove and min/max.
1899
1900 // Try to find a dominating comparison of these guys.
1901 // It can simplify the index computation for Arrays.copyOf
1902 // and similar uses of System.arraycopy.
1903 // First, compute the normalized version of CmpI(x, y).
1904 int cmp_op = Op_CmpI;
1905 Node* xkey = xvalue;
1906 Node* ykey = yvalue;
1907 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1908 if (ideal_cmpxy->is_Cmp()) {
1909 // E.g., if we have CmpI(length - offset, count),
1910 // it might idealize to CmpI(length, count + offset)
1911 cmp_op = ideal_cmpxy->Opcode();
1912 xkey = ideal_cmpxy->in(1);
1913 ykey = ideal_cmpxy->in(2);
1914 }
1915
1916 // Start by locating any relevant comparisons.
1917 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1918 Node* cmpxy = NULL;
1919 Node* cmpyx = NULL;
1920 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1921 Node* cmp = start_from->fast_out(k);
1922 if (cmp->outcnt() > 0 && // must have prior uses
1923 cmp->in(0) == NULL && // must be context-independent
1924 cmp->Opcode() == cmp_op) { // right kind of compare
1925 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1926 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1927 }
1928 }
1929
1930 const int NCMPS = 2;
1931 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1932 int cmpn;
1933 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1934 if (cmps[cmpn] != NULL) break; // find a result
1935 }
1936 if (cmpn < NCMPS) {
1937 // Look for a dominating test that tells us the min and max.
1938 int depth = 0; // Limit search depth for speed
1939 Node* dom = control();
1940 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1941 if (++depth >= 100) break;
1942 Node* ifproj = dom;
1943 if (!ifproj->is_Proj()) continue;
1944 Node* iff = ifproj->in(0);
1945 if (!iff->is_If()) continue;
1946 Node* bol = iff->in(1);
1947 if (!bol->is_Bool()) continue;
1948 Node* cmp = bol->in(1);
1949 if (cmp == NULL) continue;
1950 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1951 if (cmps[cmpn] == cmp) break;
1952 if (cmpn == NCMPS) continue;
1953 BoolTest::mask btest = bol->as_Bool()->_test._test;
1954 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1955 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1956 // At this point, we know that 'x btest y' is true.
1957 switch (btest) {
1958 case BoolTest::eq:
1959 // They are proven equal, so we can collapse the min/max.
1960 // Either value is the answer. Choose the simpler.
1961 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1962 return yvalue;
1963 return xvalue;
1964 case BoolTest::lt: // x < y
1965 case BoolTest::le: // x <= y
1966 return (want_max ? yvalue : xvalue);
1967 case BoolTest::gt: // x > y
1968 case BoolTest::ge: // x >= y
1969 return (want_max ? xvalue : yvalue);
1970 }
1971 }
1972 }
1973
1974 // We failed to find a dominating test.
1975 // Let's pick a test that might GVN with prior tests.
1976 Node* best_bol = NULL;
1977 BoolTest::mask best_btest = BoolTest::illegal;
1978 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1979 Node* cmp = cmps[cmpn];
1980 if (cmp == NULL) continue;
1981 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1982 Node* bol = cmp->fast_out(j);
1983 if (!bol->is_Bool()) continue;
1984 BoolTest::mask btest = bol->as_Bool()->_test._test;
1985 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1986 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1987 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1988 best_bol = bol->as_Bool();
1989 best_btest = btest;
1990 }
1991 }
1992 }
1993
1994 Node* answer_if_true = NULL;
1995 Node* answer_if_false = NULL;
1996 switch (best_btest) {
1997 default:
1998 if (cmpxy == NULL)
1999 cmpxy = ideal_cmpxy;
2000 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2001 // and fall through:
2002 case BoolTest::lt: // x < y
2003 case BoolTest::le: // x <= y
2004 answer_if_true = (want_max ? yvalue : xvalue);
2005 answer_if_false = (want_max ? xvalue : yvalue);
2006 break;
2007 case BoolTest::gt: // x > y
2008 case BoolTest::ge: // x >= y
2009 answer_if_true = (want_max ? xvalue : yvalue);
2010 answer_if_false = (want_max ? yvalue : xvalue);
2011 break;
2012 }
2013
2014 jint hi, lo;
2015 if (want_max) {
2016 // We can sharpen the minimum.
2017 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2018 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2019 } else {
2020 // We can sharpen the maximum.
2021 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2022 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2023 }
2024
2025 // Use a flow-free graph structure, to avoid creating excess control edges
2026 // which could hinder other optimizations.
2027 // Since Math.min/max is often used with arraycopy, we want
2028 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2029 Node* cmov = CMoveNode::make(NULL, best_bol,
2030 answer_if_false, answer_if_true,
2031 TypeInt::make(lo, hi, widen));
2032
2033 return _gvn.transform(cmov);
2034
2035 /*
2036 // This is not as desirable as it may seem, since Min and Max
2037 // nodes do not have a full set of optimizations.
2038 // And they would interfere, anyway, with 'if' optimizations
2039 // and with CMoveI canonical forms.
2040 switch (id) {
2041 case vmIntrinsics::_min:
2042 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2043 case vmIntrinsics::_max:
2044 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2045 default:
2046 ShouldNotReachHere();
2047 }
2048 */
2049 }
2050
2051 inline int
2052 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2053 const TypePtr* base_type = TypePtr::NULL_PTR;
2054 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2055 if (base_type == NULL) {
2056 // Unknown type.
2057 return Type::AnyPtr;
2058 } else if (base_type == TypePtr::NULL_PTR) {
2059 // Since this is a NULL+long form, we have to switch to a rawptr.
2060 base = _gvn.transform(new CastX2PNode(offset));
2061 offset = MakeConX(0);
2062 return Type::RawPtr;
2063 } else if (base_type->base() == Type::RawPtr) {
2064 return Type::RawPtr;
2065 } else if (base_type->isa_oopptr()) {
2066 // Base is never null => always a heap address.
2067 if (base_type->ptr() == TypePtr::NotNull) {
2068 return Type::OopPtr;
2069 }
2070 // Offset is small => always a heap address.
2071 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2072 if (offset_type != NULL &&
2073 base_type->offset() == 0 && // (should always be?)
2074 offset_type->_lo >= 0 &&
2075 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2076 return Type::OopPtr;
2077 }
2078 // Otherwise, it might either be oop+off or NULL+addr.
2079 return Type::AnyPtr;
2080 } else {
2081 // No information:
2082 return Type::AnyPtr;
2083 }
2084 }
2085
2086 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2087 int kind = classify_unsafe_addr(base, offset);
2088 if (kind == Type::RawPtr) {
2089 return basic_plus_adr(top(), base, offset);
2090 } else {
2091 return basic_plus_adr(base, offset);
2092 }
2093 }
2094
2095 //--------------------------inline_number_methods-----------------------------
2096 // inline int Integer.numberOfLeadingZeros(int)
2097 // inline int Long.numberOfLeadingZeros(long)
2098 //
2099 // inline int Integer.numberOfTrailingZeros(int)
2100 // inline int Long.numberOfTrailingZeros(long)
2101 //
2102 // inline int Integer.bitCount(int)
2103 // inline int Long.bitCount(long)
2104 //
2105 // inline char Character.reverseBytes(char)
2106 // inline short Short.reverseBytes(short)
2107 // inline int Integer.reverseBytes(int)
2108 // inline long Long.reverseBytes(long)
2109 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2110 Node* arg = argument(0);
2111 Node* n;
2112 switch (id) {
2113 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2114 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2115 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2116 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2117 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2118 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2119 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2120 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2121 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2122 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2123 default: fatal_unexpected_iid(id); break;
2124 }
2125 set_result(_gvn.transform(n));
2126 return true;
2127 }
2128
2129 //----------------------------inline_unsafe_access----------------------------
2130
2131 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2132
2133 // Helper that guards and inserts a pre-barrier.
2134 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2135 Node* pre_val, bool need_mem_bar) {
2136 // We could be accessing the referent field of a reference object. If so, when G1
2137 // is enabled, we need to log the value in the referent field in an SATB buffer.
2138 // This routine performs some compile time filters and generates suitable
2139 // runtime filters that guard the pre-barrier code.
2140 // Also add memory barrier for non volatile load from the referent field
2141 // to prevent commoning of loads across safepoint.
2142 if (!UseG1GC && !need_mem_bar)
2143 return;
2144
2145 // Some compile time checks.
2146
2147 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2148 const TypeX* otype = offset->find_intptr_t_type();
2149 if (otype != NULL && otype->is_con() &&
2150 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2151 // Constant offset but not the reference_offset so just return
2152 return;
2153 }
2154
2155 // We only need to generate the runtime guards for instances.
2156 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2157 if (btype != NULL) {
2158 if (btype->isa_aryptr()) {
2159 // Array type so nothing to do
2160 return;
2161 }
2162
2163 const TypeInstPtr* itype = btype->isa_instptr();
2164 if (itype != NULL) {
2165 // Can the klass of base_oop be statically determined to be
2166 // _not_ a sub-class of Reference and _not_ Object?
2167 ciKlass* klass = itype->klass();
2168 if ( klass->is_loaded() &&
2169 !klass->is_subtype_of(env()->Reference_klass()) &&
2170 !env()->Object_klass()->is_subtype_of(klass)) {
2171 return;
2172 }
2173 }
2174 }
2175
2176 // The compile time filters did not reject base_oop/offset so
2177 // we need to generate the following runtime filters
2178 //
2179 // if (offset == java_lang_ref_Reference::_reference_offset) {
2180 // if (instance_of(base, java.lang.ref.Reference)) {
2181 // pre_barrier(_, pre_val, ...);
2182 // }
2183 // }
2184
2185 float likely = PROB_LIKELY( 0.999);
2186 float unlikely = PROB_UNLIKELY(0.999);
2187
2188 IdealKit ideal(this);
2189 #define __ ideal.
2190
2191 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2192
2193 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2194 // Update graphKit memory and control from IdealKit.
2195 sync_kit(ideal);
2196
2197 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2198 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2199
2200 // Update IdealKit memory and control from graphKit.
2201 __ sync_kit(this);
2202
2203 Node* one = __ ConI(1);
2204 // is_instof == 0 if base_oop == NULL
2205 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2206
2207 // Update graphKit from IdeakKit.
2208 sync_kit(ideal);
2209
2210 // Use the pre-barrier to record the value in the referent field
2211 pre_barrier(false /* do_load */,
2212 __ ctrl(),
2213 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2214 pre_val /* pre_val */,
2215 T_OBJECT);
2216 if (need_mem_bar) {
2217 // Add memory barrier to prevent commoning reads from this field
2218 // across safepoint since GC can change its value.
2219 insert_mem_bar(Op_MemBarCPUOrder);
2220 }
2221 // Update IdealKit from graphKit.
2222 __ sync_kit(this);
2223
2224 } __ end_if(); // _ref_type != ref_none
2225 } __ end_if(); // offset == referent_offset
2226
2227 // Final sync IdealKit and GraphKit.
2228 final_sync(ideal);
2229 #undef __
2230 }
2231
2232
2233 // Interpret Unsafe.fieldOffset cookies correctly:
2234 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2235
2236 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2237 // Attempt to infer a sharper value type from the offset and base type.
2238 ciKlass* sharpened_klass = NULL;
2239
2240 // See if it is an instance field, with an object type.
2241 if (alias_type->field() != NULL) {
2242 assert(!is_native_ptr, "native pointer op cannot use a java address");
2243 if (alias_type->field()->type()->is_klass()) {
2244 sharpened_klass = alias_type->field()->type()->as_klass();
2245 }
2246 }
2247
2248 // See if it is a narrow oop array.
2249 if (adr_type->isa_aryptr()) {
2250 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2251 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2252 if (elem_type != NULL) {
2253 sharpened_klass = elem_type->klass();
2254 }
2255 }
2256 }
2257
2258 // The sharpened class might be unloaded if there is no class loader
2259 // contraint in place.
2260 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2261 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2262
2263 #ifndef PRODUCT
2264 if (C->print_intrinsics() || C->print_inlining()) {
2265 tty->print(" from base type: "); adr_type->dump();
2266 tty->print(" sharpened value: "); tjp->dump();
2267 }
2268 #endif
2269 // Sharpen the value type.
2270 return tjp;
2271 }
2272 return NULL;
2273 }
2274
2275 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2276 if (callee()->is_static()) return false; // caller must have the capability!
2277
2278 #ifndef PRODUCT
2279 {
2280 ResourceMark rm;
2281 // Check the signatures.
2282 ciSignature* sig = callee()->signature();
2283 #ifdef ASSERT
2284 if (!is_store) {
2285 // Object getObject(Object base, int/long offset), etc.
2286 BasicType rtype = sig->return_type()->basic_type();
2287 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2288 rtype = T_ADDRESS; // it is really a C void*
2289 assert(rtype == type, "getter must return the expected value");
2290 if (!is_native_ptr) {
2291 assert(sig->count() == 2, "oop getter has 2 arguments");
2292 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2293 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2294 } else {
2295 assert(sig->count() == 1, "native getter has 1 argument");
2296 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2297 }
2298 } else {
2299 // void putObject(Object base, int/long offset, Object x), etc.
2300 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2301 if (!is_native_ptr) {
2302 assert(sig->count() == 3, "oop putter has 3 arguments");
2303 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2304 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2305 } else {
2306 assert(sig->count() == 2, "native putter has 2 arguments");
2307 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2308 }
2309 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2310 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2311 vtype = T_ADDRESS; // it is really a C void*
2312 assert(vtype == type, "putter must accept the expected value");
2313 }
2314 #endif // ASSERT
2315 }
2316 #endif //PRODUCT
2317
2318 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2319
2320 Node* receiver = argument(0); // type: oop
2321
2322 // Build address expression.
2323 Node* adr;
2324 Node* heap_base_oop = top();
2325 Node* offset = top();
2326 Node* val;
2327
2328 if (!is_native_ptr) {
2329 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2330 Node* base = argument(1); // type: oop
2331 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2332 offset = argument(2); // type: long
2333 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2334 // to be plain byte offsets, which are also the same as those accepted
2335 // by oopDesc::field_base.
2336 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2337 "fieldOffset must be byte-scaled");
2338 // 32-bit machines ignore the high half!
2339 offset = ConvL2X(offset);
2340 adr = make_unsafe_address(base, offset);
2341 heap_base_oop = base;
2342 val = is_store ? argument(4) : NULL;
2343 } else {
2344 Node* ptr = argument(1); // type: long
2345 ptr = ConvL2X(ptr); // adjust Java long to machine word
2346 adr = make_unsafe_address(NULL, ptr);
2347 val = is_store ? argument(3) : NULL;
2348 }
2349
2350 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2351
2352 // First guess at the value type.
2353 const Type *value_type = Type::get_const_basic_type(type);
2354
2355 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2356 // there was not enough information to nail it down.
2357 Compile::AliasType* alias_type = C->alias_type(adr_type);
2358 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2359
2360 // We will need memory barriers unless we can determine a unique
2361 // alias category for this reference. (Note: If for some reason
2362 // the barriers get omitted and the unsafe reference begins to "pollute"
2363 // the alias analysis of the rest of the graph, either Compile::can_alias
2364 // or Compile::must_alias will throw a diagnostic assert.)
2365 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2366
2367 // If we are reading the value of the referent field of a Reference
2368 // object (either by using Unsafe directly or through reflection)
2369 // then, if G1 is enabled, we need to record the referent in an
2370 // SATB log buffer using the pre-barrier mechanism.
2371 // Also we need to add memory barrier to prevent commoning reads
2372 // from this field across safepoint since GC can change its value.
2373 bool need_read_barrier = !is_native_ptr && !is_store &&
2374 offset != top() && heap_base_oop != top();
2375
2376 if (!is_store && type == T_OBJECT) {
2377 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2378 if (tjp != NULL) {
2379 value_type = tjp;
2380 }
2381 }
2382
2383 receiver = null_check(receiver);
2384 if (stopped()) {
2385 return true;
2386 }
2387 // Heap pointers get a null-check from the interpreter,
2388 // as a courtesy. However, this is not guaranteed by Unsafe,
2389 // and it is not possible to fully distinguish unintended nulls
2390 // from intended ones in this API.
2391
2392 if (is_volatile) {
2393 // We need to emit leading and trailing CPU membars (see below) in
2394 // addition to memory membars when is_volatile. This is a little
2395 // too strong, but avoids the need to insert per-alias-type
2396 // volatile membars (for stores; compare Parse::do_put_xxx), which
2397 // we cannot do effectively here because we probably only have a
2398 // rough approximation of type.
2399 need_mem_bar = true;
2400 // For Stores, place a memory ordering barrier now.
2401 if (is_store) {
2402 insert_mem_bar(Op_MemBarRelease);
2403 } else {
2404 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2405 insert_mem_bar(Op_MemBarVolatile);
2406 }
2407 }
2408 }
2409
2410 // Memory barrier to prevent normal and 'unsafe' accesses from
2411 // bypassing each other. Happens after null checks, so the
2412 // exception paths do not take memory state from the memory barrier,
2413 // so there's no problems making a strong assert about mixing users
2414 // of safe & unsafe memory.
2415 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2416
2417 if (!is_store) {
2418 Node* p = NULL;
2419 // Try to constant fold a load from a constant field
2420 ciField* field = alias_type->field();
2421 if (heap_base_oop != top() &&
2422 field != NULL && field->is_constant() && field->layout_type() == type) {
2423 // final or stable field
2424 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2425 if (con_type != NULL) {
2426 p = makecon(con_type);
2427 }
2428 }
2429 if (p == NULL) {
2430 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2431 // To be valid, unsafe loads may depend on other conditions than
2432 // the one that guards them: pin the Load node
2433 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2434 // load value
2435 switch (type) {
2436 case T_BOOLEAN:
2437 case T_CHAR:
2438 case T_BYTE:
2439 case T_SHORT:
2440 case T_INT:
2441 case T_LONG:
2442 case T_FLOAT:
2443 case T_DOUBLE:
2444 break;
2445 case T_OBJECT:
2446 if (need_read_barrier) {
2447 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2448 }
2449 break;
2450 case T_ADDRESS:
2451 // Cast to an int type.
2452 p = _gvn.transform(new CastP2XNode(NULL, p));
2453 p = ConvX2UL(p);
2454 break;
2455 default:
2456 fatal("unexpected type %d: %s", type, type2name(type));
2457 break;
2458 }
2459 }
2460 // The load node has the control of the preceding MemBarCPUOrder. All
2461 // following nodes will have the control of the MemBarCPUOrder inserted at
2462 // the end of this method. So, pushing the load onto the stack at a later
2463 // point is fine.
2464 set_result(p);
2465 } else {
2466 // place effect of store into memory
2467 switch (type) {
2468 case T_DOUBLE:
2469 val = dstore_rounding(val);
2470 break;
2471 case T_ADDRESS:
2472 // Repackage the long as a pointer.
2473 val = ConvL2X(val);
2474 val = _gvn.transform(new CastX2PNode(val));
2475 break;
2476 }
2477
2478 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2479 if (type != T_OBJECT ) {
2480 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2481 } else {
2482 // Possibly an oop being stored to Java heap or native memory
2483 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2484 // oop to Java heap.
2485 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2486 } else {
2487 // We can't tell at compile time if we are storing in the Java heap or outside
2488 // of it. So we need to emit code to conditionally do the proper type of
2489 // store.
2490
2491 IdealKit ideal(this);
2492 #define __ ideal.
2493 // QQQ who knows what probability is here??
2494 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2495 // Sync IdealKit and graphKit.
2496 sync_kit(ideal);
2497 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2498 // Update IdealKit memory.
2499 __ sync_kit(this);
2500 } __ else_(); {
2501 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2502 } __ end_if();
2503 // Final sync IdealKit and GraphKit.
2504 final_sync(ideal);
2505 #undef __
2506 }
2507 }
2508 }
2509
2510 if (is_volatile) {
2511 if (!is_store) {
2512 insert_mem_bar(Op_MemBarAcquire);
2513 } else {
2514 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2515 insert_mem_bar(Op_MemBarVolatile);
2516 }
2517 }
2518 }
2519
2520 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2521
2522 return true;
2523 }
2524
2525 //----------------------------inline_unsafe_load_store----------------------------
2526 // This method serves a couple of different customers (depending on LoadStoreKind):
2527 //
2528 // LS_cmpxchg:
2529 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2530 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2531 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2532 //
2533 // LS_xadd:
2534 // public int getAndAddInt( Object o, long offset, int delta)
2535 // public long getAndAddLong(Object o, long offset, long delta)
2536 //
2537 // LS_xchg:
2538 // int getAndSet(Object o, long offset, int newValue)
2539 // long getAndSet(Object o, long offset, long newValue)
2540 // Object getAndSet(Object o, long offset, Object newValue)
2541 //
2542 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2543 // This basic scheme here is the same as inline_unsafe_access, but
2544 // differs in enough details that combining them would make the code
2545 // overly confusing. (This is a true fact! I originally combined
2546 // them, but even I was confused by it!) As much code/comments as
2547 // possible are retained from inline_unsafe_access though to make
2548 // the correspondences clearer. - dl
2549
2550 if (callee()->is_static()) return false; // caller must have the capability!
2551
2552 #ifndef PRODUCT
2553 BasicType rtype;
2554 {
2555 ResourceMark rm;
2556 // Check the signatures.
2557 ciSignature* sig = callee()->signature();
2558 rtype = sig->return_type()->basic_type();
2559 if (kind == LS_xadd || kind == LS_xchg) {
2560 // Check the signatures.
2561 #ifdef ASSERT
2562 assert(rtype == type, "get and set must return the expected type");
2563 assert(sig->count() == 3, "get and set has 3 arguments");
2564 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2565 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2566 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2567 #endif // ASSERT
2568 } else if (kind == LS_cmpxchg) {
2569 // Check the signatures.
2570 #ifdef ASSERT
2571 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2572 assert(sig->count() == 4, "CAS has 4 arguments");
2573 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2574 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2575 #endif // ASSERT
2576 } else {
2577 ShouldNotReachHere();
2578 }
2579 }
2580 #endif //PRODUCT
2581
2582 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2583
2584 // Get arguments:
2585 Node* receiver = NULL;
2586 Node* base = NULL;
2587 Node* offset = NULL;
2588 Node* oldval = NULL;
2589 Node* newval = NULL;
2590 if (kind == LS_cmpxchg) {
2591 const bool two_slot_type = type2size[type] == 2;
2592 receiver = argument(0); // type: oop
2593 base = argument(1); // type: oop
2594 offset = argument(2); // type: long
2595 oldval = argument(4); // type: oop, int, or long
2596 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2597 } else if (kind == LS_xadd || kind == LS_xchg){
2598 receiver = argument(0); // type: oop
2599 base = argument(1); // type: oop
2600 offset = argument(2); // type: long
2601 oldval = NULL;
2602 newval = argument(4); // type: oop, int, or long
2603 }
2604
2605 // Null check receiver.
2606 receiver = null_check(receiver);
2607 if (stopped()) {
2608 return true;
2609 }
2610
2611 // Build field offset expression.
2612 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2613 // to be plain byte offsets, which are also the same as those accepted
2614 // by oopDesc::field_base.
2615 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2616 // 32-bit machines ignore the high half of long offsets
2617 offset = ConvL2X(offset);
2618 Node* adr = make_unsafe_address(base, offset);
2619 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2620
2621 // For CAS, unlike inline_unsafe_access, there seems no point in
2622 // trying to refine types. Just use the coarse types here.
2623 const Type *value_type = Type::get_const_basic_type(type);
2624 Compile::AliasType* alias_type = C->alias_type(adr_type);
2625 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2626
2627 if (kind == LS_xchg && type == T_OBJECT) {
2628 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2629 if (tjp != NULL) {
2630 value_type = tjp;
2631 }
2632 }
2633
2634 int alias_idx = C->get_alias_index(adr_type);
2635
2636 // Memory-model-wise, a LoadStore acts like a little synchronized
2637 // block, so needs barriers on each side. These don't translate
2638 // into actual barriers on most machines, but we still need rest of
2639 // compiler to respect ordering.
2640
2641 insert_mem_bar(Op_MemBarRelease);
2642 insert_mem_bar(Op_MemBarCPUOrder);
2643
2644 // 4984716: MemBars must be inserted before this
2645 // memory node in order to avoid a false
2646 // dependency which will confuse the scheduler.
2647 Node *mem = memory(alias_idx);
2648
2649 // For now, we handle only those cases that actually exist: ints,
2650 // longs, and Object. Adding others should be straightforward.
2651 Node* load_store;
2652 switch(type) {
2653 case T_INT:
2654 if (kind == LS_xadd) {
2655 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2656 } else if (kind == LS_xchg) {
2657 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2658 } else if (kind == LS_cmpxchg) {
2659 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2660 } else {
2661 ShouldNotReachHere();
2662 }
2663 break;
2664 case T_LONG:
2665 if (kind == LS_xadd) {
2666 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2667 } else if (kind == LS_xchg) {
2668 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2669 } else if (kind == LS_cmpxchg) {
2670 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2671 } else {
2672 ShouldNotReachHere();
2673 }
2674 break;
2675 case T_OBJECT:
2676 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2677 // could be delayed during Parse (for example, in adjust_map_after_if()).
2678 // Execute transformation here to avoid barrier generation in such case.
2679 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2680 newval = _gvn.makecon(TypePtr::NULL_PTR);
2681
2682 // Reference stores need a store barrier.
2683 if (kind == LS_xchg) {
2684 // If pre-barrier must execute before the oop store, old value will require do_load here.
2685 if (!can_move_pre_barrier()) {
2686 pre_barrier(true /* do_load*/,
2687 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2688 NULL /* pre_val*/,
2689 T_OBJECT);
2690 } // Else move pre_barrier to use load_store value, see below.
2691 } else if (kind == LS_cmpxchg) {
2692 // Same as for newval above:
2693 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2694 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2695 }
2696 // The only known value which might get overwritten is oldval.
2697 pre_barrier(false /* do_load */,
2698 control(), NULL, NULL, max_juint, NULL, NULL,
2699 oldval /* pre_val */,
2700 T_OBJECT);
2701 } else {
2702 ShouldNotReachHere();
2703 }
2704
2705 #ifdef _LP64
2706 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2707 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2708 if (kind == LS_xchg) {
2709 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2710 newval_enc, adr_type, value_type->make_narrowoop()));
2711 } else {
2712 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2713 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2714 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2715 newval_enc, oldval_enc));
2716 }
2717 } else
2718 #endif
2719 {
2720 if (kind == LS_xchg) {
2721 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2722 } else {
2723 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2724 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2725 }
2726 }
2727 if (kind == LS_cmpxchg) {
2728 // Emit the post barrier only when the actual store happened.
2729 // This makes sense to check only for compareAndSet that can fail to set the value.
2730 // CAS success path is marked more likely since we anticipate this is a performance
2731 // critical path, while CAS failure path can use the penalty for going through unlikely
2732 // path as backoff. Which is still better than doing a store barrier there.
2733 IdealKit ideal(this);
2734 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2735 sync_kit(ideal);
2736 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2737 ideal.sync_kit(this);
2738 } ideal.end_if();
2739 final_sync(ideal);
2740 } else {
2741 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2742 }
2743 break;
2744 default:
2745 fatal("unexpected type %d: %s", type, type2name(type));
2746 break;
2747 }
2748
2749 // SCMemProjNodes represent the memory state of a LoadStore. Their
2750 // main role is to prevent LoadStore nodes from being optimized away
2751 // when their results aren't used.
2752 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2753 set_memory(proj, alias_idx);
2754
2755 if (type == T_OBJECT && kind == LS_xchg) {
2756 #ifdef _LP64
2757 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2758 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2759 }
2760 #endif
2761 if (can_move_pre_barrier()) {
2762 // Don't need to load pre_val. The old value is returned by load_store.
2763 // The pre_barrier can execute after the xchg as long as no safepoint
2764 // gets inserted between them.
2765 pre_barrier(false /* do_load */,
2766 control(), NULL, NULL, max_juint, NULL, NULL,
2767 load_store /* pre_val */,
2768 T_OBJECT);
2769 }
2770 }
2771
2772 // Add the trailing membar surrounding the access
2773 insert_mem_bar(Op_MemBarCPUOrder);
2774 insert_mem_bar(Op_MemBarAcquire);
2775
2776 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2777 set_result(load_store);
2778 return true;
2779 }
2780
2781 //----------------------------inline_unsafe_ordered_store----------------------
2782 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2783 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2784 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2785 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2786 // This is another variant of inline_unsafe_access, differing in
2787 // that it always issues store-store ("release") barrier and ensures
2788 // store-atomicity (which only matters for "long").
2789
2790 if (callee()->is_static()) return false; // caller must have the capability!
2791
2792 #ifndef PRODUCT
2793 {
2794 ResourceMark rm;
2795 // Check the signatures.
2796 ciSignature* sig = callee()->signature();
2797 #ifdef ASSERT
2798 BasicType rtype = sig->return_type()->basic_type();
2799 assert(rtype == T_VOID, "must return void");
2800 assert(sig->count() == 3, "has 3 arguments");
2801 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2802 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2803 #endif // ASSERT
2804 }
2805 #endif //PRODUCT
2806
2807 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2808
2809 // Get arguments:
2810 Node* receiver = argument(0); // type: oop
2811 Node* base = argument(1); // type: oop
2812 Node* offset = argument(2); // type: long
2813 Node* val = argument(4); // type: oop, int, or long
2814
2815 // Null check receiver.
2816 receiver = null_check(receiver);
2817 if (stopped()) {
2818 return true;
2819 }
2820
2821 // Build field offset expression.
2822 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2823 // 32-bit machines ignore the high half of long offsets
2824 offset = ConvL2X(offset);
2825 Node* adr = make_unsafe_address(base, offset);
2826 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2827 const Type *value_type = Type::get_const_basic_type(type);
2828 Compile::AliasType* alias_type = C->alias_type(adr_type);
2829
2830 insert_mem_bar(Op_MemBarRelease);
2831 insert_mem_bar(Op_MemBarCPUOrder);
2832 // Ensure that the store is atomic for longs:
2833 const bool require_atomic_access = true;
2834 Node* store;
2835 if (type == T_OBJECT) // reference stores need a store barrier.
2836 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2837 else {
2838 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2839 }
2840 insert_mem_bar(Op_MemBarCPUOrder);
2841 return true;
2842 }
2843
2844 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2845 // Regardless of form, don't allow previous ld/st to move down,
2846 // then issue acquire, release, or volatile mem_bar.
2847 insert_mem_bar(Op_MemBarCPUOrder);
2848 switch(id) {
2849 case vmIntrinsics::_loadFence:
2850 insert_mem_bar(Op_LoadFence);
2851 return true;
2852 case vmIntrinsics::_storeFence:
2853 insert_mem_bar(Op_StoreFence);
2854 return true;
2855 case vmIntrinsics::_fullFence:
2856 insert_mem_bar(Op_MemBarVolatile);
2857 return true;
2858 default:
2859 fatal_unexpected_iid(id);
2860 return false;
2861 }
2862 }
2863
2864 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2865 if (!kls->is_Con()) {
2866 return true;
2867 }
2868 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2869 if (klsptr == NULL) {
2870 return true;
2871 }
2872 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2873 // don't need a guard for a klass that is already initialized
2874 return !ik->is_initialized();
2875 }
2876
2877 //----------------------------inline_unsafe_allocate---------------------------
2878 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2879 bool LibraryCallKit::inline_unsafe_allocate() {
2880 if (callee()->is_static()) return false; // caller must have the capability!
2881
2882 null_check_receiver(); // null-check, then ignore
2883 Node* cls = null_check(argument(1));
2884 if (stopped()) return true;
2885
2886 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2887 kls = null_check(kls);
2888 if (stopped()) return true; // argument was like int.class
2889
2890 Node* test = NULL;
2891 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2892 // Note: The argument might still be an illegal value like
2893 // Serializable.class or Object[].class. The runtime will handle it.
2894 // But we must make an explicit check for initialization.
2895 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2896 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2897 // can generate code to load it as unsigned byte.
2898 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2899 Node* bits = intcon(InstanceKlass::fully_initialized);
2900 test = _gvn.transform(new SubINode(inst, bits));
2901 // The 'test' is non-zero if we need to take a slow path.
2902 }
2903
2904 Node* obj = new_instance(kls, test);
2905 set_result(obj);
2906 return true;
2907 }
2908
2909 #ifdef TRACE_HAVE_INTRINSICS
2910 /*
2911 * oop -> myklass
2912 * myklass->trace_id |= USED
2913 * return myklass->trace_id & ~0x3
2914 */
2915 bool LibraryCallKit::inline_native_classID() {
2916 null_check_receiver(); // null-check, then ignore
2917 Node* cls = null_check(argument(1), T_OBJECT);
2918 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2919 kls = null_check(kls, T_OBJECT);
2920 ByteSize offset = TRACE_ID_OFFSET;
2921 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2922 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2923 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2924 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2925 Node* clsused = longcon(0x01l); // set the class bit
2926 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2927
2928 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2929 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2930 set_result(andl);
2931 return true;
2932 }
2933
2934 bool LibraryCallKit::inline_native_threadID() {
2935 Node* tls_ptr = NULL;
2936 Node* cur_thr = generate_current_thread(tls_ptr);
2937 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2938 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2939 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2940
2941 Node* threadid = NULL;
2942 size_t thread_id_size = OSThread::thread_id_size();
2943 if (thread_id_size == (size_t) BytesPerLong) {
2944 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2945 } else if (thread_id_size == (size_t) BytesPerInt) {
2946 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2947 } else {
2948 ShouldNotReachHere();
2949 }
2950 set_result(threadid);
2951 return true;
2952 }
2953 #endif
2954
2955 //------------------------inline_native_time_funcs--------------
2956 // inline code for System.currentTimeMillis() and System.nanoTime()
2957 // these have the same type and signature
2958 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2959 const TypeFunc* tf = OptoRuntime::void_long_Type();
2960 const TypePtr* no_memory_effects = NULL;
2961 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2962 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2963 #ifdef ASSERT
2964 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2965 assert(value_top == top(), "second value must be top");
2966 #endif
2967 set_result(value);
2968 return true;
2969 }
2970
2971 //------------------------inline_native_currentThread------------------
2972 bool LibraryCallKit::inline_native_currentThread() {
2973 Node* junk = NULL;
2974 set_result(generate_current_thread(junk));
2975 return true;
2976 }
2977
2978 //------------------------inline_native_isInterrupted------------------
2979 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2980 bool LibraryCallKit::inline_native_isInterrupted() {
2981 // Add a fast path to t.isInterrupted(clear_int):
2982 // (t == Thread.current() &&
2983 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2984 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2985 // So, in the common case that the interrupt bit is false,
2986 // we avoid making a call into the VM. Even if the interrupt bit
2987 // is true, if the clear_int argument is false, we avoid the VM call.
2988 // However, if the receiver is not currentThread, we must call the VM,
2989 // because there must be some locking done around the operation.
2990
2991 // We only go to the fast case code if we pass two guards.
2992 // Paths which do not pass are accumulated in the slow_region.
2993
2994 enum {
2995 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2996 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2997 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2998 PATH_LIMIT
2999 };
3000
3001 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3002 // out of the function.
3003 insert_mem_bar(Op_MemBarCPUOrder);
3004
3005 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3006 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3007
3008 RegionNode* slow_region = new RegionNode(1);
3009 record_for_igvn(slow_region);
3010
3011 // (a) Receiving thread must be the current thread.
3012 Node* rec_thr = argument(0);
3013 Node* tls_ptr = NULL;
3014 Node* cur_thr = generate_current_thread(tls_ptr);
3015 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3016 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3017
3018 generate_slow_guard(bol_thr, slow_region);
3019
3020 // (b) Interrupt bit on TLS must be false.
3021 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3022 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3023 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3024
3025 // Set the control input on the field _interrupted read to prevent it floating up.
3026 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3027 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3028 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3029
3030 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3031
3032 // First fast path: if (!TLS._interrupted) return false;
3033 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3034 result_rgn->init_req(no_int_result_path, false_bit);
3035 result_val->init_req(no_int_result_path, intcon(0));
3036
3037 // drop through to next case
3038 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3039
3040 #ifndef TARGET_OS_FAMILY_windows
3041 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3042 Node* clr_arg = argument(1);
3043 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3044 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3045 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3046
3047 // Second fast path: ... else if (!clear_int) return true;
3048 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3049 result_rgn->init_req(no_clear_result_path, false_arg);
3050 result_val->init_req(no_clear_result_path, intcon(1));
3051
3052 // drop through to next case
3053 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3054 #else
3055 // To return true on Windows you must read the _interrupted field
3056 // and check the the event state i.e. take the slow path.
3057 #endif // TARGET_OS_FAMILY_windows
3058
3059 // (d) Otherwise, go to the slow path.
3060 slow_region->add_req(control());
3061 set_control( _gvn.transform(slow_region));
3062
3063 if (stopped()) {
3064 // There is no slow path.
3065 result_rgn->init_req(slow_result_path, top());
3066 result_val->init_req(slow_result_path, top());
3067 } else {
3068 // non-virtual because it is a private non-static
3069 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3070
3071 Node* slow_val = set_results_for_java_call(slow_call);
3072 // this->control() comes from set_results_for_java_call
3073
3074 Node* fast_io = slow_call->in(TypeFunc::I_O);
3075 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3076
3077 // These two phis are pre-filled with copies of of the fast IO and Memory
3078 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3079 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3080
3081 result_rgn->init_req(slow_result_path, control());
3082 result_io ->init_req(slow_result_path, i_o());
3083 result_mem->init_req(slow_result_path, reset_memory());
3084 result_val->init_req(slow_result_path, slow_val);
3085
3086 set_all_memory(_gvn.transform(result_mem));
3087 set_i_o( _gvn.transform(result_io));
3088 }
3089
3090 C->set_has_split_ifs(true); // Has chance for split-if optimization
3091 set_result(result_rgn, result_val);
3092 return true;
3093 }
3094
3095 //---------------------------load_mirror_from_klass----------------------------
3096 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3097 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3098 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3099 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3100 }
3101
3102 //-----------------------load_klass_from_mirror_common-------------------------
3103 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3104 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3105 // and branch to the given path on the region.
3106 // If never_see_null, take an uncommon trap on null, so we can optimistically
3107 // compile for the non-null case.
3108 // If the region is NULL, force never_see_null = true.
3109 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3110 bool never_see_null,
3111 RegionNode* region,
3112 int null_path,
3113 int offset) {
3114 if (region == NULL) never_see_null = true;
3115 Node* p = basic_plus_adr(mirror, offset);
3116 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3117 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3118 Node* null_ctl = top();
3119 kls = null_check_oop(kls, &null_ctl, never_see_null);
3120 if (region != NULL) {
3121 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3122 region->init_req(null_path, null_ctl);
3123 } else {
3124 assert(null_ctl == top(), "no loose ends");
3125 }
3126 return kls;
3127 }
3128
3129 //--------------------(inline_native_Class_query helpers)---------------------
3130 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3131 // Fall through if (mods & mask) == bits, take the guard otherwise.
3132 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3133 // Branch around if the given klass has the given modifier bit set.
3134 // Like generate_guard, adds a new path onto the region.
3135 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3136 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3137 Node* mask = intcon(modifier_mask);
3138 Node* bits = intcon(modifier_bits);
3139 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3140 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3141 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3142 return generate_fair_guard(bol, region);
3143 }
3144 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3145 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3146 }
3147
3148 //-------------------------inline_native_Class_query-------------------
3149 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3150 const Type* return_type = TypeInt::BOOL;
3151 Node* prim_return_value = top(); // what happens if it's a primitive class?
3152 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3153 bool expect_prim = false; // most of these guys expect to work on refs
3154
3155 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3156
3157 Node* mirror = argument(0);
3158 Node* obj = top();
3159
3160 switch (id) {
3161 case vmIntrinsics::_isInstance:
3162 // nothing is an instance of a primitive type
3163 prim_return_value = intcon(0);
3164 obj = argument(1);
3165 break;
3166 case vmIntrinsics::_getModifiers:
3167 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3168 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3169 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3170 break;
3171 case vmIntrinsics::_isInterface:
3172 prim_return_value = intcon(0);
3173 break;
3174 case vmIntrinsics::_isArray:
3175 prim_return_value = intcon(0);
3176 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3177 break;
3178 case vmIntrinsics::_isPrimitive:
3179 prim_return_value = intcon(1);
3180 expect_prim = true; // obviously
3181 break;
3182 case vmIntrinsics::_getSuperclass:
3183 prim_return_value = null();
3184 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3185 break;
3186 case vmIntrinsics::_getClassAccessFlags:
3187 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3188 return_type = TypeInt::INT; // not bool! 6297094
3189 break;
3190 default:
3191 fatal_unexpected_iid(id);
3192 break;
3193 }
3194
3195 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3196 if (mirror_con == NULL) return false; // cannot happen?
3197
3198 #ifndef PRODUCT
3199 if (C->print_intrinsics() || C->print_inlining()) {
3200 ciType* k = mirror_con->java_mirror_type();
3201 if (k) {
3202 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3203 k->print_name();
3204 tty->cr();
3205 }
3206 }
3207 #endif
3208
3209 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3210 RegionNode* region = new RegionNode(PATH_LIMIT);
3211 record_for_igvn(region);
3212 PhiNode* phi = new PhiNode(region, return_type);
3213
3214 // The mirror will never be null of Reflection.getClassAccessFlags, however
3215 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3216 // if it is. See bug 4774291.
3217
3218 // For Reflection.getClassAccessFlags(), the null check occurs in
3219 // the wrong place; see inline_unsafe_access(), above, for a similar
3220 // situation.
3221 mirror = null_check(mirror);
3222 // If mirror or obj is dead, only null-path is taken.
3223 if (stopped()) return true;
3224
3225 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3226
3227 // Now load the mirror's klass metaobject, and null-check it.
3228 // Side-effects region with the control path if the klass is null.
3229 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3230 // If kls is null, we have a primitive mirror.
3231 phi->init_req(_prim_path, prim_return_value);
3232 if (stopped()) { set_result(region, phi); return true; }
3233 bool safe_for_replace = (region->in(_prim_path) == top());
3234
3235 Node* p; // handy temp
3236 Node* null_ctl;
3237
3238 // Now that we have the non-null klass, we can perform the real query.
3239 // For constant classes, the query will constant-fold in LoadNode::Value.
3240 Node* query_value = top();
3241 switch (id) {
3242 case vmIntrinsics::_isInstance:
3243 // nothing is an instance of a primitive type
3244 query_value = gen_instanceof(obj, kls, safe_for_replace);
3245 break;
3246
3247 case vmIntrinsics::_getModifiers:
3248 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3249 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3250 break;
3251
3252 case vmIntrinsics::_isInterface:
3253 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3254 if (generate_interface_guard(kls, region) != NULL)
3255 // A guard was added. If the guard is taken, it was an interface.
3256 phi->add_req(intcon(1));
3257 // If we fall through, it's a plain class.
3258 query_value = intcon(0);
3259 break;
3260
3261 case vmIntrinsics::_isArray:
3262 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3263 if (generate_array_guard(kls, region) != NULL)
3264 // A guard was added. If the guard is taken, it was an array.
3265 phi->add_req(intcon(1));
3266 // If we fall through, it's a plain class.
3267 query_value = intcon(0);
3268 break;
3269
3270 case vmIntrinsics::_isPrimitive:
3271 query_value = intcon(0); // "normal" path produces false
3272 break;
3273
3274 case vmIntrinsics::_getSuperclass:
3275 // The rules here are somewhat unfortunate, but we can still do better
3276 // with random logic than with a JNI call.
3277 // Interfaces store null or Object as _super, but must report null.
3278 // Arrays store an intermediate super as _super, but must report Object.
3279 // Other types can report the actual _super.
3280 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3281 if (generate_interface_guard(kls, region) != NULL)
3282 // A guard was added. If the guard is taken, it was an interface.
3283 phi->add_req(null());
3284 if (generate_array_guard(kls, region) != NULL)
3285 // A guard was added. If the guard is taken, it was an array.
3286 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3287 // If we fall through, it's a plain class. Get its _super.
3288 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3289 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3290 null_ctl = top();
3291 kls = null_check_oop(kls, &null_ctl);
3292 if (null_ctl != top()) {
3293 // If the guard is taken, Object.superClass is null (both klass and mirror).
3294 region->add_req(null_ctl);
3295 phi ->add_req(null());
3296 }
3297 if (!stopped()) {
3298 query_value = load_mirror_from_klass(kls);
3299 }
3300 break;
3301
3302 case vmIntrinsics::_getClassAccessFlags:
3303 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3304 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3305 break;
3306
3307 default:
3308 fatal_unexpected_iid(id);
3309 break;
3310 }
3311
3312 // Fall-through is the normal case of a query to a real class.
3313 phi->init_req(1, query_value);
3314 region->init_req(1, control());
3315
3316 C->set_has_split_ifs(true); // Has chance for split-if optimization
3317 set_result(region, phi);
3318 return true;
3319 }
3320
3321 //-------------------------inline_Class_cast-------------------
3322 bool LibraryCallKit::inline_Class_cast() {
3323 Node* mirror = argument(0); // Class
3324 Node* obj = argument(1);
3325 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3326 if (mirror_con == NULL) {
3327 return false; // dead path (mirror->is_top()).
3328 }
3329 if (obj == NULL || obj->is_top()) {
3330 return false; // dead path
3331 }
3332 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3333
3334 // First, see if Class.cast() can be folded statically.
3335 // java_mirror_type() returns non-null for compile-time Class constants.
3336 ciType* tm = mirror_con->java_mirror_type();
3337 if (tm != NULL && tm->is_klass() &&
3338 tp != NULL && tp->klass() != NULL) {
3339 if (!tp->klass()->is_loaded()) {
3340 // Don't use intrinsic when class is not loaded.
3341 return false;
3342 } else {
3343 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3344 if (static_res == Compile::SSC_always_true) {
3345 // isInstance() is true - fold the code.
3346 set_result(obj);
3347 return true;
3348 } else if (static_res == Compile::SSC_always_false) {
3349 // Don't use intrinsic, have to throw ClassCastException.
3350 // If the reference is null, the non-intrinsic bytecode will
3351 // be optimized appropriately.
3352 return false;
3353 }
3354 }
3355 }
3356
3357 // Bailout intrinsic and do normal inlining if exception path is frequent.
3358 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3359 return false;
3360 }
3361
3362 // Generate dynamic checks.
3363 // Class.cast() is java implementation of _checkcast bytecode.
3364 // Do checkcast (Parse::do_checkcast()) optimizations here.
3365
3366 mirror = null_check(mirror);
3367 // If mirror is dead, only null-path is taken.
3368 if (stopped()) {
3369 return true;
3370 }
3371
3372 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3373 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3374 RegionNode* region = new RegionNode(PATH_LIMIT);
3375 record_for_igvn(region);
3376
3377 // Now load the mirror's klass metaobject, and null-check it.
3378 // If kls is null, we have a primitive mirror and
3379 // nothing is an instance of a primitive type.
3380 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3381
3382 Node* res = top();
3383 if (!stopped()) {
3384 Node* bad_type_ctrl = top();
3385 // Do checkcast optimizations.
3386 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3387 region->init_req(_bad_type_path, bad_type_ctrl);
3388 }
3389 if (region->in(_prim_path) != top() ||
3390 region->in(_bad_type_path) != top()) {
3391 // Let Interpreter throw ClassCastException.
3392 PreserveJVMState pjvms(this);
3393 set_control(_gvn.transform(region));
3394 uncommon_trap(Deoptimization::Reason_intrinsic,
3395 Deoptimization::Action_maybe_recompile);
3396 }
3397 if (!stopped()) {
3398 set_result(res);
3399 }
3400 return true;
3401 }
3402
3403
3404 //--------------------------inline_native_subtype_check------------------------
3405 // This intrinsic takes the JNI calls out of the heart of
3406 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3407 bool LibraryCallKit::inline_native_subtype_check() {
3408 // Pull both arguments off the stack.
3409 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3410 args[0] = argument(0);
3411 args[1] = argument(1);
3412 Node* klasses[2]; // corresponding Klasses: superk, subk
3413 klasses[0] = klasses[1] = top();
3414
3415 enum {
3416 // A full decision tree on {superc is prim, subc is prim}:
3417 _prim_0_path = 1, // {P,N} => false
3418 // {P,P} & superc!=subc => false
3419 _prim_same_path, // {P,P} & superc==subc => true
3420 _prim_1_path, // {N,P} => false
3421 _ref_subtype_path, // {N,N} & subtype check wins => true
3422 _both_ref_path, // {N,N} & subtype check loses => false
3423 PATH_LIMIT
3424 };
3425
3426 RegionNode* region = new RegionNode(PATH_LIMIT);
3427 Node* phi = new PhiNode(region, TypeInt::BOOL);
3428 record_for_igvn(region);
3429
3430 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3431 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3432 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3433
3434 // First null-check both mirrors and load each mirror's klass metaobject.
3435 int which_arg;
3436 for (which_arg = 0; which_arg <= 1; which_arg++) {
3437 Node* arg = args[which_arg];
3438 arg = null_check(arg);
3439 if (stopped()) break;
3440 args[which_arg] = arg;
3441
3442 Node* p = basic_plus_adr(arg, class_klass_offset);
3443 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3444 klasses[which_arg] = _gvn.transform(kls);
3445 }
3446
3447 // Having loaded both klasses, test each for null.
3448 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3449 for (which_arg = 0; which_arg <= 1; which_arg++) {
3450 Node* kls = klasses[which_arg];
3451 Node* null_ctl = top();
3452 kls = null_check_oop(kls, &null_ctl, never_see_null);
3453 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3454 region->init_req(prim_path, null_ctl);
3455 if (stopped()) break;
3456 klasses[which_arg] = kls;
3457 }
3458
3459 if (!stopped()) {
3460 // now we have two reference types, in klasses[0..1]
3461 Node* subk = klasses[1]; // the argument to isAssignableFrom
3462 Node* superk = klasses[0]; // the receiver
3463 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3464 // now we have a successful reference subtype check
3465 region->set_req(_ref_subtype_path, control());
3466 }
3467
3468 // If both operands are primitive (both klasses null), then
3469 // we must return true when they are identical primitives.
3470 // It is convenient to test this after the first null klass check.
3471 set_control(region->in(_prim_0_path)); // go back to first null check
3472 if (!stopped()) {
3473 // Since superc is primitive, make a guard for the superc==subc case.
3474 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3475 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3476 generate_guard(bol_eq, region, PROB_FAIR);
3477 if (region->req() == PATH_LIMIT+1) {
3478 // A guard was added. If the added guard is taken, superc==subc.
3479 region->swap_edges(PATH_LIMIT, _prim_same_path);
3480 region->del_req(PATH_LIMIT);
3481 }
3482 region->set_req(_prim_0_path, control()); // Not equal after all.
3483 }
3484
3485 // these are the only paths that produce 'true':
3486 phi->set_req(_prim_same_path, intcon(1));
3487 phi->set_req(_ref_subtype_path, intcon(1));
3488
3489 // pull together the cases:
3490 assert(region->req() == PATH_LIMIT, "sane region");
3491 for (uint i = 1; i < region->req(); i++) {
3492 Node* ctl = region->in(i);
3493 if (ctl == NULL || ctl == top()) {
3494 region->set_req(i, top());
3495 phi ->set_req(i, top());
3496 } else if (phi->in(i) == NULL) {
3497 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3498 }
3499 }
3500
3501 set_control(_gvn.transform(region));
3502 set_result(_gvn.transform(phi));
3503 return true;
3504 }
3505
3506 //---------------------generate_array_guard_common------------------------
3507 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3508 bool obj_array, bool not_array) {
3509
3510 if (stopped()) {
3511 return NULL;
3512 }
3513
3514 // If obj_array/non_array==false/false:
3515 // Branch around if the given klass is in fact an array (either obj or prim).
3516 // If obj_array/non_array==false/true:
3517 // Branch around if the given klass is not an array klass of any kind.
3518 // If obj_array/non_array==true/true:
3519 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3520 // If obj_array/non_array==true/false:
3521 // Branch around if the kls is an oop array (Object[] or subtype)
3522 //
3523 // Like generate_guard, adds a new path onto the region.
3524 jint layout_con = 0;
3525 Node* layout_val = get_layout_helper(kls, layout_con);
3526 if (layout_val == NULL) {
3527 bool query = (obj_array
3528 ? Klass::layout_helper_is_objArray(layout_con)
3529 : Klass::layout_helper_is_array(layout_con));
3530 if (query == not_array) {
3531 return NULL; // never a branch
3532 } else { // always a branch
3533 Node* always_branch = control();
3534 if (region != NULL)
3535 region->add_req(always_branch);
3536 set_control(top());
3537 return always_branch;
3538 }
3539 }
3540 // Now test the correct condition.
3541 jint nval = (obj_array
3542 ? ((jint)Klass::_lh_array_tag_type_value
3543 << Klass::_lh_array_tag_shift)
3544 : Klass::_lh_neutral_value);
3545 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3546 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3547 // invert the test if we are looking for a non-array
3548 if (not_array) btest = BoolTest(btest).negate();
3549 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3550 return generate_fair_guard(bol, region);
3551 }
3552
3553
3554 //-----------------------inline_native_newArray--------------------------
3555 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3556 bool LibraryCallKit::inline_native_newArray() {
3557 Node* mirror = argument(0);
3558 Node* count_val = argument(1);
3559
3560 mirror = null_check(mirror);
3561 // If mirror or obj is dead, only null-path is taken.
3562 if (stopped()) return true;
3563
3564 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3565 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3566 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3567 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3568 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3569
3570 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3571 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3572 result_reg, _slow_path);
3573 Node* normal_ctl = control();
3574 Node* no_array_ctl = result_reg->in(_slow_path);
3575
3576 // Generate code for the slow case. We make a call to newArray().
3577 set_control(no_array_ctl);
3578 if (!stopped()) {
3579 // Either the input type is void.class, or else the
3580 // array klass has not yet been cached. Either the
3581 // ensuing call will throw an exception, or else it
3582 // will cache the array klass for next time.
3583 PreserveJVMState pjvms(this);
3584 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3585 Node* slow_result = set_results_for_java_call(slow_call);
3586 // this->control() comes from set_results_for_java_call
3587 result_reg->set_req(_slow_path, control());
3588 result_val->set_req(_slow_path, slow_result);
3589 result_io ->set_req(_slow_path, i_o());
3590 result_mem->set_req(_slow_path, reset_memory());
3591 }
3592
3593 set_control(normal_ctl);
3594 if (!stopped()) {
3595 // Normal case: The array type has been cached in the java.lang.Class.
3596 // The following call works fine even if the array type is polymorphic.
3597 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3598 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3599 result_reg->init_req(_normal_path, control());
3600 result_val->init_req(_normal_path, obj);
3601 result_io ->init_req(_normal_path, i_o());
3602 result_mem->init_req(_normal_path, reset_memory());
3603 }
3604
3605 // Return the combined state.
3606 set_i_o( _gvn.transform(result_io) );
3607 set_all_memory( _gvn.transform(result_mem));
3608
3609 C->set_has_split_ifs(true); // Has chance for split-if optimization
3610 set_result(result_reg, result_val);
3611 return true;
3612 }
3613
3614 //----------------------inline_native_getLength--------------------------
3615 // public static native int java.lang.reflect.Array.getLength(Object array);
3616 bool LibraryCallKit::inline_native_getLength() {
3617 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3618
3619 Node* array = null_check(argument(0));
3620 // If array is dead, only null-path is taken.
3621 if (stopped()) return true;
3622
3623 // Deoptimize if it is a non-array.
3624 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3625
3626 if (non_array != NULL) {
3627 PreserveJVMState pjvms(this);
3628 set_control(non_array);
3629 uncommon_trap(Deoptimization::Reason_intrinsic,
3630 Deoptimization::Action_maybe_recompile);
3631 }
3632
3633 // If control is dead, only non-array-path is taken.
3634 if (stopped()) return true;
3635
3636 // The works fine even if the array type is polymorphic.
3637 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3638 Node* result = load_array_length(array);
3639
3640 C->set_has_split_ifs(true); // Has chance for split-if optimization
3641 set_result(result);
3642 return true;
3643 }
3644
3645 //------------------------inline_array_copyOf----------------------------
3646 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3647 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3648 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3649 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3650
3651 // Get the arguments.
3652 Node* original = argument(0);
3653 Node* start = is_copyOfRange? argument(1): intcon(0);
3654 Node* end = is_copyOfRange? argument(2): argument(1);
3655 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3656
3657 Node* newcopy;
3658
3659 // Set the original stack and the reexecute bit for the interpreter to reexecute
3660 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3661 { PreserveReexecuteState preexecs(this);
3662 jvms()->set_should_reexecute(true);
3663
3664 array_type_mirror = null_check(array_type_mirror);
3665 original = null_check(original);
3666
3667 // Check if a null path was taken unconditionally.
3668 if (stopped()) return true;
3669
3670 Node* orig_length = load_array_length(original);
3671
3672 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3673 klass_node = null_check(klass_node);
3674
3675 RegionNode* bailout = new RegionNode(1);
3676 record_for_igvn(bailout);
3677
3678 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3679 // Bail out if that is so.
3680 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3681 if (not_objArray != NULL) {
3682 // Improve the klass node's type from the new optimistic assumption:
3683 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3684 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3685 Node* cast = new CastPPNode(klass_node, akls);
3686 cast->init_req(0, control());
3687 klass_node = _gvn.transform(cast);
3688 }
3689
3690 // Bail out if either start or end is negative.
3691 generate_negative_guard(start, bailout, &start);
3692 generate_negative_guard(end, bailout, &end);
3693
3694 Node* length = end;
3695 if (_gvn.type(start) != TypeInt::ZERO) {
3696 length = _gvn.transform(new SubINode(end, start));
3697 }
3698
3699 // Bail out if length is negative.
3700 // Without this the new_array would throw
3701 // NegativeArraySizeException but IllegalArgumentException is what
3702 // should be thrown
3703 generate_negative_guard(length, bailout, &length);
3704
3705 if (bailout->req() > 1) {
3706 PreserveJVMState pjvms(this);
3707 set_control(_gvn.transform(bailout));
3708 uncommon_trap(Deoptimization::Reason_intrinsic,
3709 Deoptimization::Action_maybe_recompile);
3710 }
3711
3712 if (!stopped()) {
3713 // How many elements will we copy from the original?
3714 // The answer is MinI(orig_length - start, length).
3715 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3716 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3717
3718 // Generate a direct call to the right arraycopy function(s).
3719 // We know the copy is disjoint but we might not know if the
3720 // oop stores need checking.
3721 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3722 // This will fail a store-check if x contains any non-nulls.
3723
3724 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3725 // loads/stores but it is legal only if we're sure the
3726 // Arrays.copyOf would succeed. So we need all input arguments
3727 // to the copyOf to be validated, including that the copy to the
3728 // new array won't trigger an ArrayStoreException. That subtype
3729 // check can be optimized if we know something on the type of
3730 // the input array from type speculation.
3731 if (_gvn.type(klass_node)->singleton()) {
3732 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3733 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3734
3735 int test = C->static_subtype_check(superk, subk);
3736 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3737 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3738 if (t_original->speculative_type() != NULL) {
3739 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3740 }
3741 }
3742 }
3743
3744 bool validated = false;
3745 // Reason_class_check rather than Reason_intrinsic because we
3746 // want to intrinsify even if this traps.
3747 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3748 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3749 klass_node);
3750
3751 if (not_subtype_ctrl != top()) {
3752 PreserveJVMState pjvms(this);
3753 set_control(not_subtype_ctrl);
3754 uncommon_trap(Deoptimization::Reason_class_check,
3755 Deoptimization::Action_make_not_entrant);
3756 assert(stopped(), "Should be stopped");
3757 }
3758 validated = true;
3759 }
3760
3761 if (!stopped()) {
3762 newcopy = new_array(klass_node, length, 0); // no arguments to push
3763
3764 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3765 load_object_klass(original), klass_node);
3766 if (!is_copyOfRange) {
3767 ac->set_copyof(validated);
3768 } else {
3769 ac->set_copyofrange(validated);
3770 }
3771 Node* n = _gvn.transform(ac);
3772 if (n == ac) {
3773 ac->connect_outputs(this);
3774 } else {
3775 assert(validated, "shouldn't transform if all arguments not validated");
3776 set_all_memory(n);
3777 }
3778 }
3779 }
3780 } // original reexecute is set back here
3781
3782 C->set_has_split_ifs(true); // Has chance for split-if optimization
3783 if (!stopped()) {
3784 set_result(newcopy);
3785 }
3786 return true;
3787 }
3788
3789
3790 //----------------------generate_virtual_guard---------------------------
3791 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3792 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3793 RegionNode* slow_region) {
3794 ciMethod* method = callee();
3795 int vtable_index = method->vtable_index();
3796 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3797 "bad index %d", vtable_index);
3798 // Get the Method* out of the appropriate vtable entry.
3799 int entry_offset = (InstanceKlass::vtable_start_offset() +
3800 vtable_index*vtableEntry::size()) * wordSize +
3801 vtableEntry::method_offset_in_bytes();
3802 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3803 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3804
3805 // Compare the target method with the expected method (e.g., Object.hashCode).
3806 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3807
3808 Node* native_call = makecon(native_call_addr);
3809 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3810 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3811
3812 return generate_slow_guard(test_native, slow_region);
3813 }
3814
3815 //-----------------------generate_method_call----------------------------
3816 // Use generate_method_call to make a slow-call to the real
3817 // method if the fast path fails. An alternative would be to
3818 // use a stub like OptoRuntime::slow_arraycopy_Java.
3819 // This only works for expanding the current library call,
3820 // not another intrinsic. (E.g., don't use this for making an
3821 // arraycopy call inside of the copyOf intrinsic.)
3822 CallJavaNode*
3823 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3824 // When compiling the intrinsic method itself, do not use this technique.
3825 guarantee(callee() != C->method(), "cannot make slow-call to self");
3826
3827 ciMethod* method = callee();
3828 // ensure the JVMS we have will be correct for this call
3829 guarantee(method_id == method->intrinsic_id(), "must match");
3830
3831 const TypeFunc* tf = TypeFunc::make(method);
3832 CallJavaNode* slow_call;
3833 if (is_static) {
3834 assert(!is_virtual, "");
3835 slow_call = new CallStaticJavaNode(C, tf,
3836 SharedRuntime::get_resolve_static_call_stub(),
3837 method, bci());
3838 } else if (is_virtual) {
3839 null_check_receiver();
3840 int vtable_index = Method::invalid_vtable_index;
3841 if (UseInlineCaches) {
3842 // Suppress the vtable call
3843 } else {
3844 // hashCode and clone are not a miranda methods,
3845 // so the vtable index is fixed.
3846 // No need to use the linkResolver to get it.
3847 vtable_index = method->vtable_index();
3848 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3849 "bad index %d", vtable_index);
3850 }
3851 slow_call = new CallDynamicJavaNode(tf,
3852 SharedRuntime::get_resolve_virtual_call_stub(),
3853 method, vtable_index, bci());
3854 } else { // neither virtual nor static: opt_virtual
3855 null_check_receiver();
3856 slow_call = new CallStaticJavaNode(C, tf,
3857 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3858 method, bci());
3859 slow_call->set_optimized_virtual(true);
3860 }
3861 set_arguments_for_java_call(slow_call);
3862 set_edges_for_java_call(slow_call);
3863 return slow_call;
3864 }
3865
3866
3867 /**
3868 * Build special case code for calls to hashCode on an object. This call may
3869 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3870 * slightly different code.
3871 */
3872 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3873 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3874 assert(!(is_virtual && is_static), "either virtual, special, or static");
3875
3876 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3877
3878 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3879 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3880 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3881 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3882 Node* obj = NULL;
3883 if (!is_static) {
3884 // Check for hashing null object
3885 obj = null_check_receiver();
3886 if (stopped()) return true; // unconditionally null
3887 result_reg->init_req(_null_path, top());
3888 result_val->init_req(_null_path, top());
3889 } else {
3890 // Do a null check, and return zero if null.
3891 // System.identityHashCode(null) == 0
3892 obj = argument(0);
3893 Node* null_ctl = top();
3894 obj = null_check_oop(obj, &null_ctl);
3895 result_reg->init_req(_null_path, null_ctl);
3896 result_val->init_req(_null_path, _gvn.intcon(0));
3897 }
3898
3899 // Unconditionally null? Then return right away.
3900 if (stopped()) {
3901 set_control( result_reg->in(_null_path));
3902 if (!stopped())
3903 set_result(result_val->in(_null_path));
3904 return true;
3905 }
3906
3907 // We only go to the fast case code if we pass a number of guards. The
3908 // paths which do not pass are accumulated in the slow_region.
3909 RegionNode* slow_region = new RegionNode(1);
3910 record_for_igvn(slow_region);
3911
3912 // If this is a virtual call, we generate a funny guard. We pull out
3913 // the vtable entry corresponding to hashCode() from the target object.
3914 // If the target method which we are calling happens to be the native
3915 // Object hashCode() method, we pass the guard. We do not need this
3916 // guard for non-virtual calls -- the caller is known to be the native
3917 // Object hashCode().
3918 if (is_virtual) {
3919 // After null check, get the object's klass.
3920 Node* obj_klass = load_object_klass(obj);
3921 generate_virtual_guard(obj_klass, slow_region);
3922 }
3923
3924 // Get the header out of the object, use LoadMarkNode when available
3925 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3926 // The control of the load must be NULL. Otherwise, the load can move before
3927 // the null check after castPP removal.
3928 Node* no_ctrl = NULL;
3929 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3930
3931 // Test the header to see if it is unlocked.
3932 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3933 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3934 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3935 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3936 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3937
3938 generate_slow_guard(test_unlocked, slow_region);
3939
3940 // Get the hash value and check to see that it has been properly assigned.
3941 // We depend on hash_mask being at most 32 bits and avoid the use of
3942 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3943 // vm: see markOop.hpp.
3944 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3945 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3946 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3947 // This hack lets the hash bits live anywhere in the mark object now, as long
3948 // as the shift drops the relevant bits into the low 32 bits. Note that
3949 // Java spec says that HashCode is an int so there's no point in capturing
3950 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3951 hshifted_header = ConvX2I(hshifted_header);
3952 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3953
3954 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3955 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3956 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3957
3958 generate_slow_guard(test_assigned, slow_region);
3959
3960 Node* init_mem = reset_memory();
3961 // fill in the rest of the null path:
3962 result_io ->init_req(_null_path, i_o());
3963 result_mem->init_req(_null_path, init_mem);
3964
3965 result_val->init_req(_fast_path, hash_val);
3966 result_reg->init_req(_fast_path, control());
3967 result_io ->init_req(_fast_path, i_o());
3968 result_mem->init_req(_fast_path, init_mem);
3969
3970 // Generate code for the slow case. We make a call to hashCode().
3971 set_control(_gvn.transform(slow_region));
3972 if (!stopped()) {
3973 // No need for PreserveJVMState, because we're using up the present state.
3974 set_all_memory(init_mem);
3975 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3976 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3977 Node* slow_result = set_results_for_java_call(slow_call);
3978 // this->control() comes from set_results_for_java_call
3979 result_reg->init_req(_slow_path, control());
3980 result_val->init_req(_slow_path, slow_result);
3981 result_io ->set_req(_slow_path, i_o());
3982 result_mem ->set_req(_slow_path, reset_memory());
3983 }
3984
3985 // Return the combined state.
3986 set_i_o( _gvn.transform(result_io) );
3987 set_all_memory( _gvn.transform(result_mem));
3988
3989 set_result(result_reg, result_val);
3990 return true;
3991 }
3992
3993 //---------------------------inline_native_getClass----------------------------
3994 // public final native Class<?> java.lang.Object.getClass();
3995 //
3996 // Build special case code for calls to getClass on an object.
3997 bool LibraryCallKit::inline_native_getClass() {
3998 Node* obj = null_check_receiver();
3999 if (stopped()) return true;
4000 set_result(load_mirror_from_klass(load_object_klass(obj)));
4001 return true;
4002 }
4003
4004 //-----------------inline_native_Reflection_getCallerClass---------------------
4005 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4006 //
4007 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4008 //
4009 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4010 // in that it must skip particular security frames and checks for
4011 // caller sensitive methods.
4012 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4013 #ifndef PRODUCT
4014 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4015 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4016 }
4017 #endif
4018
4019 if (!jvms()->has_method()) {
4020 #ifndef PRODUCT
4021 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4022 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4023 }
4024 #endif
4025 return false;
4026 }
4027
4028 // Walk back up the JVM state to find the caller at the required
4029 // depth.
4030 JVMState* caller_jvms = jvms();
4031
4032 // Cf. JVM_GetCallerClass
4033 // NOTE: Start the loop at depth 1 because the current JVM state does
4034 // not include the Reflection.getCallerClass() frame.
4035 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4036 ciMethod* m = caller_jvms->method();
4037 switch (n) {
4038 case 0:
4039 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4040 break;
4041 case 1:
4042 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4043 if (!m->caller_sensitive()) {
4044 #ifndef PRODUCT
4045 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4046 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4047 }
4048 #endif
4049 return false; // bail-out; let JVM_GetCallerClass do the work
4050 }
4051 break;
4052 default:
4053 if (!m->is_ignored_by_security_stack_walk()) {
4054 // We have reached the desired frame; return the holder class.
4055 // Acquire method holder as java.lang.Class and push as constant.
4056 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4057 ciInstance* caller_mirror = caller_klass->java_mirror();
4058 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4059
4060 #ifndef PRODUCT
4061 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4062 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());
4063 tty->print_cr(" JVM state at this point:");
4064 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4065 ciMethod* m = jvms()->of_depth(i)->method();
4066 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4067 }
4068 }
4069 #endif
4070 return true;
4071 }
4072 break;
4073 }
4074 }
4075
4076 #ifndef PRODUCT
4077 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4078 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4079 tty->print_cr(" JVM state at this point:");
4080 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4081 ciMethod* m = jvms()->of_depth(i)->method();
4082 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4083 }
4084 }
4085 #endif
4086
4087 return false; // bail-out; let JVM_GetCallerClass do the work
4088 }
4089
4090 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4091 Node* arg = argument(0);
4092 Node* result;
4093
4094 switch (id) {
4095 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4096 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4097 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4098 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4099
4100 case vmIntrinsics::_doubleToLongBits: {
4101 // two paths (plus control) merge in a wood
4102 RegionNode *r = new RegionNode(3);
4103 Node *phi = new PhiNode(r, TypeLong::LONG);
4104
4105 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4106 // Build the boolean node
4107 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4108
4109 // Branch either way.
4110 // NaN case is less traveled, which makes all the difference.
4111 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4112 Node *opt_isnan = _gvn.transform(ifisnan);
4113 assert( opt_isnan->is_If(), "Expect an IfNode");
4114 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4115 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4116
4117 set_control(iftrue);
4118
4119 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4120 Node *slow_result = longcon(nan_bits); // return NaN
4121 phi->init_req(1, _gvn.transform( slow_result ));
4122 r->init_req(1, iftrue);
4123
4124 // Else fall through
4125 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4126 set_control(iffalse);
4127
4128 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4129 r->init_req(2, iffalse);
4130
4131 // Post merge
4132 set_control(_gvn.transform(r));
4133 record_for_igvn(r);
4134
4135 C->set_has_split_ifs(true); // Has chance for split-if optimization
4136 result = phi;
4137 assert(result->bottom_type()->isa_long(), "must be");
4138 break;
4139 }
4140
4141 case vmIntrinsics::_floatToIntBits: {
4142 // two paths (plus control) merge in a wood
4143 RegionNode *r = new RegionNode(3);
4144 Node *phi = new PhiNode(r, TypeInt::INT);
4145
4146 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4147 // Build the boolean node
4148 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4149
4150 // Branch either way.
4151 // NaN case is less traveled, which makes all the difference.
4152 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4153 Node *opt_isnan = _gvn.transform(ifisnan);
4154 assert( opt_isnan->is_If(), "Expect an IfNode");
4155 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4156 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4157
4158 set_control(iftrue);
4159
4160 static const jint nan_bits = 0x7fc00000;
4161 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4162 phi->init_req(1, _gvn.transform( slow_result ));
4163 r->init_req(1, iftrue);
4164
4165 // Else fall through
4166 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4167 set_control(iffalse);
4168
4169 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4170 r->init_req(2, iffalse);
4171
4172 // Post merge
4173 set_control(_gvn.transform(r));
4174 record_for_igvn(r);
4175
4176 C->set_has_split_ifs(true); // Has chance for split-if optimization
4177 result = phi;
4178 assert(result->bottom_type()->isa_int(), "must be");
4179 break;
4180 }
4181
4182 default:
4183 fatal_unexpected_iid(id);
4184 break;
4185 }
4186 set_result(_gvn.transform(result));
4187 return true;
4188 }
4189
4190 #ifdef _LP64
4191 #define XTOP ,top() /*additional argument*/
4192 #else //_LP64
4193 #define XTOP /*no additional argument*/
4194 #endif //_LP64
4195
4196 //----------------------inline_unsafe_copyMemory-------------------------
4197 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4198 bool LibraryCallKit::inline_unsafe_copyMemory() {
4199 if (callee()->is_static()) return false; // caller must have the capability!
4200 null_check_receiver(); // null-check receiver
4201 if (stopped()) return true;
4202
4203 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4204
4205 Node* src_ptr = argument(1); // type: oop
4206 Node* src_off = ConvL2X(argument(2)); // type: long
4207 Node* dst_ptr = argument(4); // type: oop
4208 Node* dst_off = ConvL2X(argument(5)); // type: long
4209 Node* size = ConvL2X(argument(7)); // type: long
4210
4211 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4212 "fieldOffset must be byte-scaled");
4213
4214 Node* src = make_unsafe_address(src_ptr, src_off);
4215 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4216
4217 // Conservatively insert a memory barrier on all memory slices.
4218 // Do not let writes of the copy source or destination float below the copy.
4219 insert_mem_bar(Op_MemBarCPUOrder);
4220
4221 // Call it. Note that the length argument is not scaled.
4222 make_runtime_call(RC_LEAF|RC_NO_FP,
4223 OptoRuntime::fast_arraycopy_Type(),
4224 StubRoutines::unsafe_arraycopy(),
4225 "unsafe_arraycopy",
4226 TypeRawPtr::BOTTOM,
4227 src, dst, size XTOP);
4228
4229 // Do not let reads of the copy destination float above the copy.
4230 insert_mem_bar(Op_MemBarCPUOrder);
4231
4232 return true;
4233 }
4234
4235 //------------------------clone_coping-----------------------------------
4236 // Helper function for inline_native_clone.
4237 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4238 assert(obj_size != NULL, "");
4239 Node* raw_obj = alloc_obj->in(1);
4240 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4241
4242 AllocateNode* alloc = NULL;
4243 if (ReduceBulkZeroing) {
4244 // We will be completely responsible for initializing this object -
4245 // mark Initialize node as complete.
4246 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4247 // The object was just allocated - there should be no any stores!
4248 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4249 // Mark as complete_with_arraycopy so that on AllocateNode
4250 // expansion, we know this AllocateNode is initialized by an array
4251 // copy and a StoreStore barrier exists after the array copy.
4252 alloc->initialization()->set_complete_with_arraycopy();
4253 }
4254
4255 // Copy the fastest available way.
4256 // TODO: generate fields copies for small objects instead.
4257 Node* src = obj;
4258 Node* dest = alloc_obj;
4259 Node* size = _gvn.transform(obj_size);
4260
4261 // Exclude the header but include array length to copy by 8 bytes words.
4262 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4263 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4264 instanceOopDesc::base_offset_in_bytes();
4265 // base_off:
4266 // 8 - 32-bit VM
4267 // 12 - 64-bit VM, compressed klass
4268 // 16 - 64-bit VM, normal klass
4269 if (base_off % BytesPerLong != 0) {
4270 assert(UseCompressedClassPointers, "");
4271 if (is_array) {
4272 // Exclude length to copy by 8 bytes words.
4273 base_off += sizeof(int);
4274 } else {
4275 // Include klass to copy by 8 bytes words.
4276 base_off = instanceOopDesc::klass_offset_in_bytes();
4277 }
4278 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4279 }
4280 src = basic_plus_adr(src, base_off);
4281 dest = basic_plus_adr(dest, base_off);
4282
4283 // Compute the length also, if needed:
4284 Node* countx = size;
4285 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4286 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4287
4288 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4289
4290 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4291 ac->set_clonebasic();
4292 Node* n = _gvn.transform(ac);
4293 if (n == ac) {
4294 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4295 } else {
4296 set_all_memory(n);
4297 }
4298
4299 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4300 if (card_mark) {
4301 assert(!is_array, "");
4302 // Put in store barrier for any and all oops we are sticking
4303 // into this object. (We could avoid this if we could prove
4304 // that the object type contains no oop fields at all.)
4305 Node* no_particular_value = NULL;
4306 Node* no_particular_field = NULL;
4307 int raw_adr_idx = Compile::AliasIdxRaw;
4308 post_barrier(control(),
4309 memory(raw_adr_type),
4310 alloc_obj,
4311 no_particular_field,
4312 raw_adr_idx,
4313 no_particular_value,
4314 T_OBJECT,
4315 false);
4316 }
4317
4318 // Do not let reads from the cloned object float above the arraycopy.
4319 if (alloc != NULL) {
4320 // Do not let stores that initialize this object be reordered with
4321 // a subsequent store that would make this object accessible by
4322 // other threads.
4323 // Record what AllocateNode this StoreStore protects so that
4324 // escape analysis can go from the MemBarStoreStoreNode to the
4325 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4326 // based on the escape status of the AllocateNode.
4327 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4328 } else {
4329 insert_mem_bar(Op_MemBarCPUOrder);
4330 }
4331 }
4332
4333 //------------------------inline_native_clone----------------------------
4334 // protected native Object java.lang.Object.clone();
4335 //
4336 // Here are the simple edge cases:
4337 // null receiver => normal trap
4338 // virtual and clone was overridden => slow path to out-of-line clone
4339 // not cloneable or finalizer => slow path to out-of-line Object.clone
4340 //
4341 // The general case has two steps, allocation and copying.
4342 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4343 //
4344 // Copying also has two cases, oop arrays and everything else.
4345 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4346 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4347 //
4348 // These steps fold up nicely if and when the cloned object's klass
4349 // can be sharply typed as an object array, a type array, or an instance.
4350 //
4351 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4352 PhiNode* result_val;
4353
4354 // Set the reexecute bit for the interpreter to reexecute
4355 // the bytecode that invokes Object.clone if deoptimization happens.
4356 { PreserveReexecuteState preexecs(this);
4357 jvms()->set_should_reexecute(true);
4358
4359 Node* obj = null_check_receiver();
4360 if (stopped()) return true;
4361
4362 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4363
4364 // If we are going to clone an instance, we need its exact type to
4365 // know the number and types of fields to convert the clone to
4366 // loads/stores. Maybe a speculative type can help us.
4367 if (!obj_type->klass_is_exact() &&
4368 obj_type->speculative_type() != NULL &&
4369 obj_type->speculative_type()->is_instance_klass()) {
4370 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4371 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4372 !spec_ik->has_injected_fields()) {
4373 ciKlass* k = obj_type->klass();
4374 if (!k->is_instance_klass() ||
4375 k->as_instance_klass()->is_interface() ||
4376 k->as_instance_klass()->has_subklass()) {
4377 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4378 }
4379 }
4380 }
4381
4382 Node* obj_klass = load_object_klass(obj);
4383 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4384 const TypeOopPtr* toop = ((tklass != NULL)
4385 ? tklass->as_instance_type()
4386 : TypeInstPtr::NOTNULL);
4387
4388 // Conservatively insert a memory barrier on all memory slices.
4389 // Do not let writes into the original float below the clone.
4390 insert_mem_bar(Op_MemBarCPUOrder);
4391
4392 // paths into result_reg:
4393 enum {
4394 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4395 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4396 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4397 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4398 PATH_LIMIT
4399 };
4400 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4401 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4402 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4403 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4404 record_for_igvn(result_reg);
4405
4406 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4407 int raw_adr_idx = Compile::AliasIdxRaw;
4408
4409 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4410 if (array_ctl != NULL) {
4411 // It's an array.
4412 PreserveJVMState pjvms(this);
4413 set_control(array_ctl);
4414 Node* obj_length = load_array_length(obj);
4415 Node* obj_size = NULL;
4416 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4417
4418 if (!use_ReduceInitialCardMarks()) {
4419 // If it is an oop array, it requires very special treatment,
4420 // because card marking is required on each card of the array.
4421 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4422 if (is_obja != NULL) {
4423 PreserveJVMState pjvms2(this);
4424 set_control(is_obja);
4425 // Generate a direct call to the right arraycopy function(s).
4426 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4427 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4428 ac->set_cloneoop();
4429 Node* n = _gvn.transform(ac);
4430 assert(n == ac, "cannot disappear");
4431 ac->connect_outputs(this);
4432
4433 result_reg->init_req(_objArray_path, control());
4434 result_val->init_req(_objArray_path, alloc_obj);
4435 result_i_o ->set_req(_objArray_path, i_o());
4436 result_mem ->set_req(_objArray_path, reset_memory());
4437 }
4438 }
4439 // Otherwise, there are no card marks to worry about.
4440 // (We can dispense with card marks if we know the allocation
4441 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4442 // causes the non-eden paths to take compensating steps to
4443 // simulate a fresh allocation, so that no further
4444 // card marks are required in compiled code to initialize
4445 // the object.)
4446
4447 if (!stopped()) {
4448 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4449
4450 // Present the results of the copy.
4451 result_reg->init_req(_array_path, control());
4452 result_val->init_req(_array_path, alloc_obj);
4453 result_i_o ->set_req(_array_path, i_o());
4454 result_mem ->set_req(_array_path, reset_memory());
4455 }
4456 }
4457
4458 // We only go to the instance fast case code if we pass a number of guards.
4459 // The paths which do not pass are accumulated in the slow_region.
4460 RegionNode* slow_region = new RegionNode(1);
4461 record_for_igvn(slow_region);
4462 if (!stopped()) {
4463 // It's an instance (we did array above). Make the slow-path tests.
4464 // If this is a virtual call, we generate a funny guard. We grab
4465 // the vtable entry corresponding to clone() from the target object.
4466 // If the target method which we are calling happens to be the
4467 // Object clone() method, we pass the guard. We do not need this
4468 // guard for non-virtual calls; the caller is known to be the native
4469 // Object clone().
4470 if (is_virtual) {
4471 generate_virtual_guard(obj_klass, slow_region);
4472 }
4473
4474 // The object must be cloneable and must not have a finalizer.
4475 // Both of these conditions may be checked in a single test.
4476 // We could optimize the cloneable test further, but we don't care.
4477 generate_access_flags_guard(obj_klass,
4478 // Test both conditions:
4479 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4480 // Must be cloneable but not finalizer:
4481 JVM_ACC_IS_CLONEABLE,
4482 slow_region);
4483 }
4484
4485 if (!stopped()) {
4486 // It's an instance, and it passed the slow-path tests.
4487 PreserveJVMState pjvms(this);
4488 Node* obj_size = NULL;
4489 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4490 // is reexecuted if deoptimization occurs and there could be problems when merging
4491 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4492 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4493
4494 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4495
4496 // Present the results of the slow call.
4497 result_reg->init_req(_instance_path, control());
4498 result_val->init_req(_instance_path, alloc_obj);
4499 result_i_o ->set_req(_instance_path, i_o());
4500 result_mem ->set_req(_instance_path, reset_memory());
4501 }
4502
4503 // Generate code for the slow case. We make a call to clone().
4504 set_control(_gvn.transform(slow_region));
4505 if (!stopped()) {
4506 PreserveJVMState pjvms(this);
4507 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4508 Node* slow_result = set_results_for_java_call(slow_call);
4509 // this->control() comes from set_results_for_java_call
4510 result_reg->init_req(_slow_path, control());
4511 result_val->init_req(_slow_path, slow_result);
4512 result_i_o ->set_req(_slow_path, i_o());
4513 result_mem ->set_req(_slow_path, reset_memory());
4514 }
4515
4516 // Return the combined state.
4517 set_control( _gvn.transform(result_reg));
4518 set_i_o( _gvn.transform(result_i_o));
4519 set_all_memory( _gvn.transform(result_mem));
4520 } // original reexecute is set back here
4521
4522 set_result(_gvn.transform(result_val));
4523 return true;
4524 }
4525
4526 // If we have a tighly coupled allocation, the arraycopy may take care
4527 // of the array initialization. If one of the guards we insert between
4528 // the allocation and the arraycopy causes a deoptimization, an
4529 // unitialized array will escape the compiled method. To prevent that
4530 // we set the JVM state for uncommon traps between the allocation and
4531 // the arraycopy to the state before the allocation so, in case of
4532 // deoptimization, we'll reexecute the allocation and the
4533 // initialization.
4534 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4535 if (alloc != NULL) {
4536 ciMethod* trap_method = alloc->jvms()->method();
4537 int trap_bci = alloc->jvms()->bci();
4538
4539 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4540 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4541 // Make sure there's no store between the allocation and the
4542 // arraycopy otherwise visible side effects could be rexecuted
4543 // in case of deoptimization and cause incorrect execution.
4544 bool no_interfering_store = true;
4545 Node* mem = alloc->in(TypeFunc::Memory);
4546 if (mem->is_MergeMem()) {
4547 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4548 Node* n = mms.memory();
4549 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4550 assert(n->is_Store(), "what else?");
4551 no_interfering_store = false;
4552 break;
4553 }
4554 }
4555 } else {
4556 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4557 Node* n = mms.memory();
4558 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4559 assert(n->is_Store(), "what else?");
4560 no_interfering_store = false;
4561 break;
4562 }
4563 }
4564 }
4565
4566 if (no_interfering_store) {
4567 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4568 uint size = alloc->req();
4569 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4570 old_jvms->set_map(sfpt);
4571 for (uint i = 0; i < size; i++) {
4572 sfpt->init_req(i, alloc->in(i));
4573 }
4574 // re-push array length for deoptimization
4575 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4576 old_jvms->set_sp(old_jvms->sp()+1);
4577 old_jvms->set_monoff(old_jvms->monoff()+1);
4578 old_jvms->set_scloff(old_jvms->scloff()+1);
4579 old_jvms->set_endoff(old_jvms->endoff()+1);
4580 old_jvms->set_should_reexecute(true);
4581
4582 sfpt->set_i_o(map()->i_o());
4583 sfpt->set_memory(map()->memory());
4584 sfpt->set_control(map()->control());
4585
4586 JVMState* saved_jvms = jvms();
4587 saved_reexecute_sp = _reexecute_sp;
4588
4589 set_jvms(sfpt->jvms());
4590 _reexecute_sp = jvms()->sp();
4591
4592 return saved_jvms;
4593 }
4594 }
4595 }
4596 return NULL;
4597 }
4598
4599 // In case of a deoptimization, we restart execution at the
4600 // allocation, allocating a new array. We would leave an uninitialized
4601 // array in the heap that GCs wouldn't expect. Move the allocation
4602 // after the traps so we don't allocate the array if we
4603 // deoptimize. This is possible because tightly_coupled_allocation()
4604 // guarantees there's no observer of the allocated array at this point
4605 // and the control flow is simple enough.
4606 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4607 if (saved_jvms != NULL && !stopped()) {
4608 assert(alloc != NULL, "only with a tightly coupled allocation");
4609 // restore JVM state to the state at the arraycopy
4610 saved_jvms->map()->set_control(map()->control());
4611 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4612 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4613 // If we've improved the types of some nodes (null check) while
4614 // emitting the guards, propagate them to the current state
4615 map()->replaced_nodes().apply(saved_jvms->map());
4616 set_jvms(saved_jvms);
4617 _reexecute_sp = saved_reexecute_sp;
4618
4619 // Remove the allocation from above the guards
4620 CallProjections callprojs;
4621 alloc->extract_projections(&callprojs, true);
4622 InitializeNode* init = alloc->initialization();
4623 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4624 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4625 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4626 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4627
4628 // move the allocation here (after the guards)
4629 _gvn.hash_delete(alloc);
4630 alloc->set_req(TypeFunc::Control, control());
4631 alloc->set_req(TypeFunc::I_O, i_o());
4632 Node *mem = reset_memory();
4633 set_all_memory(mem);
4634 alloc->set_req(TypeFunc::Memory, mem);
4635 set_control(init->proj_out(TypeFunc::Control));
4636 set_i_o(callprojs.fallthrough_ioproj);
4637
4638 // Update memory as done in GraphKit::set_output_for_allocation()
4639 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4640 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4641 if (ary_type->isa_aryptr() && length_type != NULL) {
4642 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4643 }
4644 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4645 int elemidx = C->get_alias_index(telemref);
4646 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4647 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4648
4649 Node* allocx = _gvn.transform(alloc);
4650 assert(allocx == alloc, "where has the allocation gone?");
4651 assert(dest->is_CheckCastPP(), "not an allocation result?");
4652
4653 _gvn.hash_delete(dest);
4654 dest->set_req(0, control());
4655 Node* destx = _gvn.transform(dest);
4656 assert(destx == dest, "where has the allocation result gone?");
4657 }
4658 }
4659
4660
4661 //------------------------------inline_arraycopy-----------------------
4662 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4663 // Object dest, int destPos,
4664 // int length);
4665 bool LibraryCallKit::inline_arraycopy() {
4666 // Get the arguments.
4667 Node* src = argument(0); // type: oop
4668 Node* src_offset = argument(1); // type: int
4669 Node* dest = argument(2); // type: oop
4670 Node* dest_offset = argument(3); // type: int
4671 Node* length = argument(4); // type: int
4672
4673
4674 // Check for allocation before we add nodes that would confuse
4675 // tightly_coupled_allocation()
4676 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4677
4678 int saved_reexecute_sp = -1;
4679 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4680 // See arraycopy_restore_alloc_state() comment
4681 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4682 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4683 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4684 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4685
4686 // The following tests must be performed
4687 // (1) src and dest are arrays.
4688 // (2) src and dest arrays must have elements of the same BasicType
4689 // (3) src and dest must not be null.
4690 // (4) src_offset must not be negative.
4691 // (5) dest_offset must not be negative.
4692 // (6) length must not be negative.
4693 // (7) src_offset + length must not exceed length of src.
4694 // (8) dest_offset + length must not exceed length of dest.
4695 // (9) each element of an oop array must be assignable
4696
4697 // (3) src and dest must not be null.
4698 // always do this here because we need the JVM state for uncommon traps
4699 Node* null_ctl = top();
4700 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4701 assert(null_ctl->is_top(), "no null control here");
4702 dest = null_check(dest, T_ARRAY);
4703
4704 if (!can_emit_guards) {
4705 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4706 // guards but the arraycopy node could still take advantage of a
4707 // tightly allocated allocation. tightly_coupled_allocation() is
4708 // called again to make sure it takes the null check above into
4709 // account: the null check is mandatory and if it caused an
4710 // uncommon trap to be emitted then the allocation can't be
4711 // considered tightly coupled in this context.
4712 alloc = tightly_coupled_allocation(dest, NULL);
4713 }
4714
4715 bool validated = false;
4716
4717 const Type* src_type = _gvn.type(src);
4718 const Type* dest_type = _gvn.type(dest);
4719 const TypeAryPtr* top_src = src_type->isa_aryptr();
4720 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4721
4722 // Do we have the type of src?
4723 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4724 // Do we have the type of dest?
4725 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4726 // Is the type for src from speculation?
4727 bool src_spec = false;
4728 // Is the type for dest from speculation?
4729 bool dest_spec = false;
4730
4731 if ((!has_src || !has_dest) && can_emit_guards) {
4732 // We don't have sufficient type information, let's see if
4733 // speculative types can help. We need to have types for both src
4734 // and dest so that it pays off.
4735
4736 // Do we already have or could we have type information for src
4737 bool could_have_src = has_src;
4738 // Do we already have or could we have type information for dest
4739 bool could_have_dest = has_dest;
4740
4741 ciKlass* src_k = NULL;
4742 if (!has_src) {
4743 src_k = src_type->speculative_type_not_null();
4744 if (src_k != NULL && src_k->is_array_klass()) {
4745 could_have_src = true;
4746 }
4747 }
4748
4749 ciKlass* dest_k = NULL;
4750 if (!has_dest) {
4751 dest_k = dest_type->speculative_type_not_null();
4752 if (dest_k != NULL && dest_k->is_array_klass()) {
4753 could_have_dest = true;
4754 }
4755 }
4756
4757 if (could_have_src && could_have_dest) {
4758 // This is going to pay off so emit the required guards
4759 if (!has_src) {
4760 src = maybe_cast_profiled_obj(src, src_k, true);
4761 src_type = _gvn.type(src);
4762 top_src = src_type->isa_aryptr();
4763 has_src = (top_src != NULL && top_src->klass() != NULL);
4764 src_spec = true;
4765 }
4766 if (!has_dest) {
4767 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4768 dest_type = _gvn.type(dest);
4769 top_dest = dest_type->isa_aryptr();
4770 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4771 dest_spec = true;
4772 }
4773 }
4774 }
4775
4776 if (has_src && has_dest && can_emit_guards) {
4777 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4778 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4779 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4780 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4781
4782 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4783 // If both arrays are object arrays then having the exact types
4784 // for both will remove the need for a subtype check at runtime
4785 // before the call and may make it possible to pick a faster copy
4786 // routine (without a subtype check on every element)
4787 // Do we have the exact type of src?
4788 bool could_have_src = src_spec;
4789 // Do we have the exact type of dest?
4790 bool could_have_dest = dest_spec;
4791 ciKlass* src_k = top_src->klass();
4792 ciKlass* dest_k = top_dest->klass();
4793 if (!src_spec) {
4794 src_k = src_type->speculative_type_not_null();
4795 if (src_k != NULL && src_k->is_array_klass()) {
4796 could_have_src = true;
4797 }
4798 }
4799 if (!dest_spec) {
4800 dest_k = dest_type->speculative_type_not_null();
4801 if (dest_k != NULL && dest_k->is_array_klass()) {
4802 could_have_dest = true;
4803 }
4804 }
4805 if (could_have_src && could_have_dest) {
4806 // If we can have both exact types, emit the missing guards
4807 if (could_have_src && !src_spec) {
4808 src = maybe_cast_profiled_obj(src, src_k, true);
4809 }
4810 if (could_have_dest && !dest_spec) {
4811 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4812 }
4813 }
4814 }
4815 }
4816
4817 ciMethod* trap_method = method();
4818 int trap_bci = bci();
4819 if (saved_jvms != NULL) {
4820 trap_method = alloc->jvms()->method();
4821 trap_bci = alloc->jvms()->bci();
4822 }
4823
4824 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4825 can_emit_guards &&
4826 !src->is_top() && !dest->is_top()) {
4827 // validate arguments: enables transformation the ArrayCopyNode
4828 validated = true;
4829
4830 RegionNode* slow_region = new RegionNode(1);
4831 record_for_igvn(slow_region);
4832
4833 // (1) src and dest are arrays.
4834 generate_non_array_guard(load_object_klass(src), slow_region);
4835 generate_non_array_guard(load_object_klass(dest), slow_region);
4836
4837 // (2) src and dest arrays must have elements of the same BasicType
4838 // done at macro expansion or at Ideal transformation time
4839
4840 // (4) src_offset must not be negative.
4841 generate_negative_guard(src_offset, slow_region);
4842
4843 // (5) dest_offset must not be negative.
4844 generate_negative_guard(dest_offset, slow_region);
4845
4846 // (7) src_offset + length must not exceed length of src.
4847 generate_limit_guard(src_offset, length,
4848 load_array_length(src),
4849 slow_region);
4850
4851 // (8) dest_offset + length must not exceed length of dest.
4852 generate_limit_guard(dest_offset, length,
4853 load_array_length(dest),
4854 slow_region);
4855
4856 // (9) each element of an oop array must be assignable
4857 Node* src_klass = load_object_klass(src);
4858 Node* dest_klass = load_object_klass(dest);
4859 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4860
4861 if (not_subtype_ctrl != top()) {
4862 PreserveJVMState pjvms(this);
4863 set_control(not_subtype_ctrl);
4864 uncommon_trap(Deoptimization::Reason_intrinsic,
4865 Deoptimization::Action_make_not_entrant);
4866 assert(stopped(), "Should be stopped");
4867 }
4868 {
4869 PreserveJVMState pjvms(this);
4870 set_control(_gvn.transform(slow_region));
4871 uncommon_trap(Deoptimization::Reason_intrinsic,
4872 Deoptimization::Action_make_not_entrant);
4873 assert(stopped(), "Should be stopped");
4874 }
4875 }
4876
4877 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4878
4879 if (stopped()) {
4880 return true;
4881 }
4882
4883 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4884 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4885 // so the compiler has a chance to eliminate them: during macro expansion,
4886 // we have to set their control (CastPP nodes are eliminated).
4887 load_object_klass(src), load_object_klass(dest),
4888 load_array_length(src), load_array_length(dest));
4889
4890 ac->set_arraycopy(validated);
4891
4892 Node* n = _gvn.transform(ac);
4893 if (n == ac) {
4894 ac->connect_outputs(this);
4895 } else {
4896 assert(validated, "shouldn't transform if all arguments not validated");
4897 set_all_memory(n);
4898 }
4899
4900 return true;
4901 }
4902
4903
4904 // Helper function which determines if an arraycopy immediately follows
4905 // an allocation, with no intervening tests or other escapes for the object.
4906 AllocateArrayNode*
4907 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4908 RegionNode* slow_region) {
4909 if (stopped()) return NULL; // no fast path
4910 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4911
4912 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4913 if (alloc == NULL) return NULL;
4914
4915 Node* rawmem = memory(Compile::AliasIdxRaw);
4916 // Is the allocation's memory state untouched?
4917 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4918 // Bail out if there have been raw-memory effects since the allocation.
4919 // (Example: There might have been a call or safepoint.)
4920 return NULL;
4921 }
4922 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4923 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4924 return NULL;
4925 }
4926
4927 // There must be no unexpected observers of this allocation.
4928 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4929 Node* obs = ptr->fast_out(i);
4930 if (obs != this->map()) {
4931 return NULL;
4932 }
4933 }
4934
4935 // This arraycopy must unconditionally follow the allocation of the ptr.
4936 Node* alloc_ctl = ptr->in(0);
4937 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4938
4939 Node* ctl = control();
4940 while (ctl != alloc_ctl) {
4941 // There may be guards which feed into the slow_region.
4942 // Any other control flow means that we might not get a chance
4943 // to finish initializing the allocated object.
4944 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4945 IfNode* iff = ctl->in(0)->as_If();
4946 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4947 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4948 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4949 ctl = iff->in(0); // This test feeds the known slow_region.
4950 continue;
4951 }
4952 // One more try: Various low-level checks bottom out in
4953 // uncommon traps. If the debug-info of the trap omits
4954 // any reference to the allocation, as we've already
4955 // observed, then there can be no objection to the trap.
4956 bool found_trap = false;
4957 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4958 Node* obs = not_ctl->fast_out(j);
4959 if (obs->in(0) == not_ctl && obs->is_Call() &&
4960 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4961 found_trap = true; break;
4962 }
4963 }
4964 if (found_trap) {
4965 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4966 continue;
4967 }
4968 }
4969 return NULL;
4970 }
4971
4972 // If we get this far, we have an allocation which immediately
4973 // precedes the arraycopy, and we can take over zeroing the new object.
4974 // The arraycopy will finish the initialization, and provide
4975 // a new control state to which we will anchor the destination pointer.
4976
4977 return alloc;
4978 }
4979
4980 //-------------inline_encodeISOArray-----------------------------------
4981 // encode char[] to byte[] in ISO_8859_1
4982 bool LibraryCallKit::inline_encodeISOArray() {
4983 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4984 // no receiver since it is static method
4985 Node *src = argument(0);
4986 Node *src_offset = argument(1);
4987 Node *dst = argument(2);
4988 Node *dst_offset = argument(3);
4989 Node *length = argument(4);
4990
4991 const Type* src_type = src->Value(&_gvn);
4992 const Type* dst_type = dst->Value(&_gvn);
4993 const TypeAryPtr* top_src = src_type->isa_aryptr();
4994 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4995 if (top_src == NULL || top_src->klass() == NULL ||
4996 top_dest == NULL || top_dest->klass() == NULL) {
4997 // failed array check
4998 return false;
4999 }
5000
5001 // Figure out the size and type of the elements we will be copying.
5002 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5003 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5004 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5005 return false;
5006 }
5007 Node* src_start = array_element_address(src, src_offset, src_elem);
5008 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5009 // 'src_start' points to src array + scaled offset
5010 // 'dst_start' points to dst array + scaled offset
5011
5012 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5013 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5014 enc = _gvn.transform(enc);
5015 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5016 set_memory(res_mem, mtype);
5017 set_result(enc);
5018 return true;
5019 }
5020
5021 //-------------inline_multiplyToLen-----------------------------------
5022 bool LibraryCallKit::inline_multiplyToLen() {
5023 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5024
5025 address stubAddr = StubRoutines::multiplyToLen();
5026 if (stubAddr == NULL) {
5027 return false; // Intrinsic's stub is not implemented on this platform
5028 }
5029 const char* stubName = "multiplyToLen";
5030
5031 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5032
5033 // no receiver because it is a static method
5034 Node* x = argument(0);
5035 Node* xlen = argument(1);
5036 Node* y = argument(2);
5037 Node* ylen = argument(3);
5038 Node* z = argument(4);
5039
5040 const Type* x_type = x->Value(&_gvn);
5041 const Type* y_type = y->Value(&_gvn);
5042 const TypeAryPtr* top_x = x_type->isa_aryptr();
5043 const TypeAryPtr* top_y = y_type->isa_aryptr();
5044 if (top_x == NULL || top_x->klass() == NULL ||
5045 top_y == NULL || top_y->klass() == NULL) {
5046 // failed array check
5047 return false;
5048 }
5049
5050 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5051 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5052 if (x_elem != T_INT || y_elem != T_INT) {
5053 return false;
5054 }
5055
5056 // Set the original stack and the reexecute bit for the interpreter to reexecute
5057 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5058 // on the return from z array allocation in runtime.
5059 { PreserveReexecuteState preexecs(this);
5060 jvms()->set_should_reexecute(true);
5061
5062 Node* x_start = array_element_address(x, intcon(0), x_elem);
5063 Node* y_start = array_element_address(y, intcon(0), y_elem);
5064 // 'x_start' points to x array + scaled xlen
5065 // 'y_start' points to y array + scaled ylen
5066
5067 // Allocate the result array
5068 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5069 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5070 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5071
5072 IdealKit ideal(this);
5073
5074 #define __ ideal.
5075 Node* one = __ ConI(1);
5076 Node* zero = __ ConI(0);
5077 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5078 __ set(need_alloc, zero);
5079 __ set(z_alloc, z);
5080 __ if_then(z, BoolTest::eq, null()); {
5081 __ increment (need_alloc, one);
5082 } __ else_(); {
5083 // Update graphKit memory and control from IdealKit.
5084 sync_kit(ideal);
5085 Node* zlen_arg = load_array_length(z);
5086 // Update IdealKit memory and control from graphKit.
5087 __ sync_kit(this);
5088 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5089 __ increment (need_alloc, one);
5090 } __ end_if();
5091 } __ end_if();
5092
5093 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5094 // Update graphKit memory and control from IdealKit.
5095 sync_kit(ideal);
5096 Node * narr = new_array(klass_node, zlen, 1);
5097 // Update IdealKit memory and control from graphKit.
5098 __ sync_kit(this);
5099 __ set(z_alloc, narr);
5100 } __ end_if();
5101
5102 sync_kit(ideal);
5103 z = __ value(z_alloc);
5104 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5105 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5106 // Final sync IdealKit and GraphKit.
5107 final_sync(ideal);
5108 #undef __
5109
5110 Node* z_start = array_element_address(z, intcon(0), T_INT);
5111
5112 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5113 OptoRuntime::multiplyToLen_Type(),
5114 stubAddr, stubName, TypePtr::BOTTOM,
5115 x_start, xlen, y_start, ylen, z_start, zlen);
5116 } // original reexecute is set back here
5117
5118 C->set_has_split_ifs(true); // Has chance for split-if optimization
5119 set_result(z);
5120 return true;
5121 }
5122
5123 //-------------inline_squareToLen------------------------------------
5124 bool LibraryCallKit::inline_squareToLen() {
5125 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5126
5127 address stubAddr = StubRoutines::squareToLen();
5128 if (stubAddr == NULL) {
5129 return false; // Intrinsic's stub is not implemented on this platform
5130 }
5131 const char* stubName = "squareToLen";
5132
5133 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5134
5135 Node* x = argument(0);
5136 Node* len = argument(1);
5137 Node* z = argument(2);
5138 Node* zlen = argument(3);
5139
5140 const Type* x_type = x->Value(&_gvn);
5141 const Type* z_type = z->Value(&_gvn);
5142 const TypeAryPtr* top_x = x_type->isa_aryptr();
5143 const TypeAryPtr* top_z = z_type->isa_aryptr();
5144 if (top_x == NULL || top_x->klass() == NULL ||
5145 top_z == NULL || top_z->klass() == NULL) {
5146 // failed array check
5147 return false;
5148 }
5149
5150 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5151 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5152 if (x_elem != T_INT || z_elem != T_INT) {
5153 return false;
5154 }
5155
5156
5157 Node* x_start = array_element_address(x, intcon(0), x_elem);
5158 Node* z_start = array_element_address(z, intcon(0), z_elem);
5159
5160 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5161 OptoRuntime::squareToLen_Type(),
5162 stubAddr, stubName, TypePtr::BOTTOM,
5163 x_start, len, z_start, zlen);
5164
5165 set_result(z);
5166 return true;
5167 }
5168
5169 //-------------inline_mulAdd------------------------------------------
5170 bool LibraryCallKit::inline_mulAdd() {
5171 assert(UseMulAddIntrinsic, "not implementated on this platform");
5172
5173 address stubAddr = StubRoutines::mulAdd();
5174 if (stubAddr == NULL) {
5175 return false; // Intrinsic's stub is not implemented on this platform
5176 }
5177 const char* stubName = "mulAdd";
5178
5179 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5180
5181 Node* out = argument(0);
5182 Node* in = argument(1);
5183 Node* offset = argument(2);
5184 Node* len = argument(3);
5185 Node* k = argument(4);
5186
5187 const Type* out_type = out->Value(&_gvn);
5188 const Type* in_type = in->Value(&_gvn);
5189 const TypeAryPtr* top_out = out_type->isa_aryptr();
5190 const TypeAryPtr* top_in = in_type->isa_aryptr();
5191 if (top_out == NULL || top_out->klass() == NULL ||
5192 top_in == NULL || top_in->klass() == NULL) {
5193 // failed array check
5194 return false;
5195 }
5196
5197 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5198 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5199 if (out_elem != T_INT || in_elem != T_INT) {
5200 return false;
5201 }
5202
5203 Node* outlen = load_array_length(out);
5204 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5205 Node* out_start = array_element_address(out, intcon(0), out_elem);
5206 Node* in_start = array_element_address(in, intcon(0), in_elem);
5207
5208 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5209 OptoRuntime::mulAdd_Type(),
5210 stubAddr, stubName, TypePtr::BOTTOM,
5211 out_start,in_start, new_offset, len, k);
5212 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5213 set_result(result);
5214 return true;
5215 }
5216
5217 //-------------inline_montgomeryMultiply-----------------------------------
5218 bool LibraryCallKit::inline_montgomeryMultiply() {
5219 address stubAddr = StubRoutines::montgomeryMultiply();
5220 if (stubAddr == NULL) {
5221 return false; // Intrinsic's stub is not implemented on this platform
5222 }
5223
5224 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5225 const char* stubName = "montgomery_square";
5226
5227 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5228
5229 Node* a = argument(0);
5230 Node* b = argument(1);
5231 Node* n = argument(2);
5232 Node* len = argument(3);
5233 Node* inv = argument(4);
5234 Node* m = argument(6);
5235
5236 const Type* a_type = a->Value(&_gvn);
5237 const TypeAryPtr* top_a = a_type->isa_aryptr();
5238 const Type* b_type = b->Value(&_gvn);
5239 const TypeAryPtr* top_b = b_type->isa_aryptr();
5240 const Type* n_type = a->Value(&_gvn);
5241 const TypeAryPtr* top_n = n_type->isa_aryptr();
5242 const Type* m_type = a->Value(&_gvn);
5243 const TypeAryPtr* top_m = m_type->isa_aryptr();
5244 if (top_a == NULL || top_a->klass() == NULL ||
5245 top_b == NULL || top_b->klass() == NULL ||
5246 top_n == NULL || top_n->klass() == NULL ||
5247 top_m == NULL || top_m->klass() == NULL) {
5248 // failed array check
5249 return false;
5250 }
5251
5252 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5253 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5254 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5255 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5256 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5257 return false;
5258 }
5259
5260 // Make the call
5261 {
5262 Node* a_start = array_element_address(a, intcon(0), a_elem);
5263 Node* b_start = array_element_address(b, intcon(0), b_elem);
5264 Node* n_start = array_element_address(n, intcon(0), n_elem);
5265 Node* m_start = array_element_address(m, intcon(0), m_elem);
5266
5267 Node* call = make_runtime_call(RC_LEAF,
5268 OptoRuntime::montgomeryMultiply_Type(),
5269 stubAddr, stubName, TypePtr::BOTTOM,
5270 a_start, b_start, n_start, len, inv, top(),
5271 m_start);
5272 set_result(m);
5273 }
5274
5275 return true;
5276 }
5277
5278 bool LibraryCallKit::inline_montgomerySquare() {
5279 address stubAddr = StubRoutines::montgomerySquare();
5280 if (stubAddr == NULL) {
5281 return false; // Intrinsic's stub is not implemented on this platform
5282 }
5283
5284 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5285 const char* stubName = "montgomery_square";
5286
5287 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5288
5289 Node* a = argument(0);
5290 Node* n = argument(1);
5291 Node* len = argument(2);
5292 Node* inv = argument(3);
5293 Node* m = argument(5);
5294
5295 const Type* a_type = a->Value(&_gvn);
5296 const TypeAryPtr* top_a = a_type->isa_aryptr();
5297 const Type* n_type = a->Value(&_gvn);
5298 const TypeAryPtr* top_n = n_type->isa_aryptr();
5299 const Type* m_type = a->Value(&_gvn);
5300 const TypeAryPtr* top_m = m_type->isa_aryptr();
5301 if (top_a == NULL || top_a->klass() == NULL ||
5302 top_n == NULL || top_n->klass() == NULL ||
5303 top_m == NULL || top_m->klass() == NULL) {
5304 // failed array check
5305 return false;
5306 }
5307
5308 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5309 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5310 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5311 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5312 return false;
5313 }
5314
5315 // Make the call
5316 {
5317 Node* a_start = array_element_address(a, intcon(0), a_elem);
5318 Node* n_start = array_element_address(n, intcon(0), n_elem);
5319 Node* m_start = array_element_address(m, intcon(0), m_elem);
5320
5321 Node* call = make_runtime_call(RC_LEAF,
5322 OptoRuntime::montgomerySquare_Type(),
5323 stubAddr, stubName, TypePtr::BOTTOM,
5324 a_start, n_start, len, inv, top(),
5325 m_start);
5326 set_result(m);
5327 }
5328
5329 return true;
5330 }
5331
5332
5333 /**
5334 * Calculate CRC32 for byte.
5335 * int java.util.zip.CRC32.update(int crc, int b)
5336 */
5337 bool LibraryCallKit::inline_updateCRC32() {
5338 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5339 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5340 // no receiver since it is static method
5341 Node* crc = argument(0); // type: int
5342 Node* b = argument(1); // type: int
5343
5344 /*
5345 * int c = ~ crc;
5346 * b = timesXtoThe32[(b ^ c) & 0xFF];
5347 * b = b ^ (c >>> 8);
5348 * crc = ~b;
5349 */
5350
5351 Node* M1 = intcon(-1);
5352 crc = _gvn.transform(new XorINode(crc, M1));
5353 Node* result = _gvn.transform(new XorINode(crc, b));
5354 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5355
5356 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5357 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5358 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5359 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5360
5361 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5362 result = _gvn.transform(new XorINode(crc, result));
5363 result = _gvn.transform(new XorINode(result, M1));
5364 set_result(result);
5365 return true;
5366 }
5367
5368 /**
5369 * Calculate CRC32 for byte[] array.
5370 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5371 */
5372 bool LibraryCallKit::inline_updateBytesCRC32() {
5373 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5374 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5375 // no receiver since it is static method
5376 Node* crc = argument(0); // type: int
5377 Node* src = argument(1); // type: oop
5378 Node* offset = argument(2); // type: int
5379 Node* length = argument(3); // type: int
5380
5381 const Type* src_type = src->Value(&_gvn);
5382 const TypeAryPtr* top_src = src_type->isa_aryptr();
5383 if (top_src == NULL || top_src->klass() == NULL) {
5384 // failed array check
5385 return false;
5386 }
5387
5388 // Figure out the size and type of the elements we will be copying.
5389 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5390 if (src_elem != T_BYTE) {
5391 return false;
5392 }
5393
5394 // 'src_start' points to src array + scaled offset
5395 Node* src_start = array_element_address(src, offset, src_elem);
5396
5397 // We assume that range check is done by caller.
5398 // TODO: generate range check (offset+length < src.length) in debug VM.
5399
5400 // Call the stub.
5401 address stubAddr = StubRoutines::updateBytesCRC32();
5402 const char *stubName = "updateBytesCRC32";
5403
5404 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5405 stubAddr, stubName, TypePtr::BOTTOM,
5406 crc, src_start, length);
5407 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5408 set_result(result);
5409 return true;
5410 }
5411
5412 /**
5413 * Calculate CRC32 for ByteBuffer.
5414 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5415 */
5416 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5417 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5418 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5419 // no receiver since it is static method
5420 Node* crc = argument(0); // type: int
5421 Node* src = argument(1); // type: long
5422 Node* offset = argument(3); // type: int
5423 Node* length = argument(4); // type: int
5424
5425 src = ConvL2X(src); // adjust Java long to machine word
5426 Node* base = _gvn.transform(new CastX2PNode(src));
5427 offset = ConvI2X(offset);
5428
5429 // 'src_start' points to src array + scaled offset
5430 Node* src_start = basic_plus_adr(top(), base, offset);
5431
5432 // Call the stub.
5433 address stubAddr = StubRoutines::updateBytesCRC32();
5434 const char *stubName = "updateBytesCRC32";
5435
5436 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5437 stubAddr, stubName, TypePtr::BOTTOM,
5438 crc, src_start, length);
5439 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5440 set_result(result);
5441 return true;
5442 }
5443
5444 //------------------------------get_table_from_crc32c_class-----------------------
5445 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5446 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5447 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5448
5449 return table;
5450 }
5451
5452 //------------------------------inline_updateBytesCRC32C-----------------------
5453 //
5454 // Calculate CRC32C for byte[] array.
5455 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5456 //
5457 bool LibraryCallKit::inline_updateBytesCRC32C() {
5458 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5459 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5460 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5461 // no receiver since it is a static method
5462 Node* crc = argument(0); // type: int
5463 Node* src = argument(1); // type: oop
5464 Node* offset = argument(2); // type: int
5465 Node* end = argument(3); // type: int
5466
5467 Node* length = _gvn.transform(new SubINode(end, offset));
5468
5469 const Type* src_type = src->Value(&_gvn);
5470 const TypeAryPtr* top_src = src_type->isa_aryptr();
5471 if (top_src == NULL || top_src->klass() == NULL) {
5472 // failed array check
5473 return false;
5474 }
5475
5476 // Figure out the size and type of the elements we will be copying.
5477 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5478 if (src_elem != T_BYTE) {
5479 return false;
5480 }
5481
5482 // 'src_start' points to src array + scaled offset
5483 Node* src_start = array_element_address(src, offset, src_elem);
5484
5485 // static final int[] byteTable in class CRC32C
5486 Node* table = get_table_from_crc32c_class(callee()->holder());
5487 Node* table_start = array_element_address(table, intcon(0), T_INT);
5488
5489 // We assume that range check is done by caller.
5490 // TODO: generate range check (offset+length < src.length) in debug VM.
5491
5492 // Call the stub.
5493 address stubAddr = StubRoutines::updateBytesCRC32C();
5494 const char *stubName = "updateBytesCRC32C";
5495
5496 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5497 stubAddr, stubName, TypePtr::BOTTOM,
5498 crc, src_start, length, table_start);
5499 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5500 set_result(result);
5501 return true;
5502 }
5503
5504 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5505 //
5506 // Calculate CRC32C for DirectByteBuffer.
5507 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5508 //
5509 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5510 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5511 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5512 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5513 // no receiver since it is a static method
5514 Node* crc = argument(0); // type: int
5515 Node* src = argument(1); // type: long
5516 Node* offset = argument(3); // type: int
5517 Node* end = argument(4); // type: int
5518
5519 Node* length = _gvn.transform(new SubINode(end, offset));
5520
5521 src = ConvL2X(src); // adjust Java long to machine word
5522 Node* base = _gvn.transform(new CastX2PNode(src));
5523 offset = ConvI2X(offset);
5524
5525 // 'src_start' points to src array + scaled offset
5526 Node* src_start = basic_plus_adr(top(), base, offset);
5527
5528 // static final int[] byteTable in class CRC32C
5529 Node* table = get_table_from_crc32c_class(callee()->holder());
5530 Node* table_start = array_element_address(table, intcon(0), T_INT);
5531
5532 // Call the stub.
5533 address stubAddr = StubRoutines::updateBytesCRC32C();
5534 const char *stubName = "updateBytesCRC32C";
5535
5536 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5537 stubAddr, stubName, TypePtr::BOTTOM,
5538 crc, src_start, length, table_start);
5539 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5540 set_result(result);
5541 return true;
5542 }
5543
5544 //------------------------------inline_updateBytesAdler32----------------------
5545 //
5546 // Calculate Adler32 checksum for byte[] array.
5547 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5548 //
5549 bool LibraryCallKit::inline_updateBytesAdler32() {
5550 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5551 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5552 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5553 // no receiver since it is static method
5554 Node* crc = argument(0); // type: int
5555 Node* src = argument(1); // type: oop
5556 Node* offset = argument(2); // type: int
5557 Node* length = argument(3); // type: int
5558
5559 const Type* src_type = src->Value(&_gvn);
5560 const TypeAryPtr* top_src = src_type->isa_aryptr();
5561 if (top_src == NULL || top_src->klass() == NULL) {
5562 // failed array check
5563 return false;
5564 }
5565
5566 // Figure out the size and type of the elements we will be copying.
5567 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5568 if (src_elem != T_BYTE) {
5569 return false;
5570 }
5571
5572 // 'src_start' points to src array + scaled offset
5573 Node* src_start = array_element_address(src, offset, src_elem);
5574
5575 // We assume that range check is done by caller.
5576 // TODO: generate range check (offset+length < src.length) in debug VM.
5577
5578 // Call the stub.
5579 address stubAddr = StubRoutines::updateBytesAdler32();
5580 const char *stubName = "updateBytesAdler32";
5581
5582 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5583 stubAddr, stubName, TypePtr::BOTTOM,
5584 crc, src_start, length);
5585 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5586 set_result(result);
5587 return true;
5588 }
5589
5590 //------------------------------inline_updateByteBufferAdler32---------------
5591 //
5592 // Calculate Adler32 checksum for DirectByteBuffer.
5593 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5594 //
5595 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5596 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5597 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5598 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5599 // no receiver since it is static method
5600 Node* crc = argument(0); // type: int
5601 Node* src = argument(1); // type: long
5602 Node* offset = argument(3); // type: int
5603 Node* length = argument(4); // type: int
5604
5605 src = ConvL2X(src); // adjust Java long to machine word
5606 Node* base = _gvn.transform(new CastX2PNode(src));
5607 offset = ConvI2X(offset);
5608
5609 // 'src_start' points to src array + scaled offset
5610 Node* src_start = basic_plus_adr(top(), base, offset);
5611
5612 // Call the stub.
5613 address stubAddr = StubRoutines::updateBytesAdler32();
5614 const char *stubName = "updateBytesAdler32";
5615
5616 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5617 stubAddr, stubName, TypePtr::BOTTOM,
5618 crc, src_start, length);
5619
5620 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5621 set_result(result);
5622 return true;
5623 }
5624
5625 //----------------------------inline_reference_get----------------------------
5626 // public T java.lang.ref.Reference.get();
5627 bool LibraryCallKit::inline_reference_get() {
5628 const int referent_offset = java_lang_ref_Reference::referent_offset;
5629 guarantee(referent_offset > 0, "should have already been set");
5630
5631 // Get the argument:
5632 Node* reference_obj = null_check_receiver();
5633 if (stopped()) return true;
5634
5635 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5636
5637 ciInstanceKlass* klass = env()->Object_klass();
5638 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5639
5640 Node* no_ctrl = NULL;
5641 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5642
5643 // Use the pre-barrier to record the value in the referent field
5644 pre_barrier(false /* do_load */,
5645 control(),
5646 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5647 result /* pre_val */,
5648 T_OBJECT);
5649
5650 // Add memory barrier to prevent commoning reads from this field
5651 // across safepoint since GC can change its value.
5652 insert_mem_bar(Op_MemBarCPUOrder);
5653
5654 set_result(result);
5655 return true;
5656 }
5657
5658
5659 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5660 bool is_exact=true, bool is_static=false,
5661 ciInstanceKlass * fromKls=NULL) {
5662 if (fromKls == NULL) {
5663 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5664 assert(tinst != NULL, "obj is null");
5665 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5666 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5667 fromKls = tinst->klass()->as_instance_klass();
5668 } else {
5669 assert(is_static, "only for static field access");
5670 }
5671 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5672 ciSymbol::make(fieldTypeString),
5673 is_static);
5674
5675 assert (field != NULL, "undefined field");
5676 if (field == NULL) return (Node *) NULL;
5677
5678 if (is_static) {
5679 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5680 fromObj = makecon(tip);
5681 }
5682
5683 // Next code copied from Parse::do_get_xxx():
5684
5685 // Compute address and memory type.
5686 int offset = field->offset_in_bytes();
5687 bool is_vol = field->is_volatile();
5688 ciType* field_klass = field->type();
5689 assert(field_klass->is_loaded(), "should be loaded");
5690 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5691 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5692 BasicType bt = field->layout_type();
5693
5694 // Build the resultant type of the load
5695 const Type *type;
5696 if (bt == T_OBJECT) {
5697 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5698 } else {
5699 type = Type::get_const_basic_type(bt);
5700 }
5701
5702 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5703 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5704 }
5705 // Build the load.
5706 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5707 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5708 // If reference is volatile, prevent following memory ops from
5709 // floating up past the volatile read. Also prevents commoning
5710 // another volatile read.
5711 if (is_vol) {
5712 // Memory barrier includes bogus read of value to force load BEFORE membar
5713 insert_mem_bar(Op_MemBarAcquire, loadedField);
5714 }
5715 return loadedField;
5716 }
5717
5718
5719 //------------------------------inline_aescrypt_Block-----------------------
5720 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5721 address stubAddr;
5722 const char *stubName;
5723 assert(UseAES, "need AES instruction support");
5724
5725 switch(id) {
5726 case vmIntrinsics::_aescrypt_encryptBlock:
5727 stubAddr = StubRoutines::aescrypt_encryptBlock();
5728 stubName = "aescrypt_encryptBlock";
5729 break;
5730 case vmIntrinsics::_aescrypt_decryptBlock:
5731 stubAddr = StubRoutines::aescrypt_decryptBlock();
5732 stubName = "aescrypt_decryptBlock";
5733 break;
5734 }
5735 if (stubAddr == NULL) return false;
5736
5737 Node* aescrypt_object = argument(0);
5738 Node* src = argument(1);
5739 Node* src_offset = argument(2);
5740 Node* dest = argument(3);
5741 Node* dest_offset = argument(4);
5742
5743 // (1) src and dest are arrays.
5744 const Type* src_type = src->Value(&_gvn);
5745 const Type* dest_type = dest->Value(&_gvn);
5746 const TypeAryPtr* top_src = src_type->isa_aryptr();
5747 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5748 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5749
5750 // for the quick and dirty code we will skip all the checks.
5751 // we are just trying to get the call to be generated.
5752 Node* src_start = src;
5753 Node* dest_start = dest;
5754 if (src_offset != NULL || dest_offset != NULL) {
5755 assert(src_offset != NULL && dest_offset != NULL, "");
5756 src_start = array_element_address(src, src_offset, T_BYTE);
5757 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5758 }
5759
5760 // now need to get the start of its expanded key array
5761 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5762 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5763 if (k_start == NULL) return false;
5764
5765 if (Matcher::pass_original_key_for_aes()) {
5766 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5767 // compatibility issues between Java key expansion and SPARC crypto instructions
5768 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5769 if (original_k_start == NULL) return false;
5770
5771 // Call the stub.
5772 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5773 stubAddr, stubName, TypePtr::BOTTOM,
5774 src_start, dest_start, k_start, original_k_start);
5775 } else {
5776 // Call the stub.
5777 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5778 stubAddr, stubName, TypePtr::BOTTOM,
5779 src_start, dest_start, k_start);
5780 }
5781
5782 return true;
5783 }
5784
5785 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5786 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5787 address stubAddr;
5788 const char *stubName;
5789
5790 assert(UseAES, "need AES instruction support");
5791
5792 switch(id) {
5793 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5794 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5795 stubName = "cipherBlockChaining_encryptAESCrypt";
5796 break;
5797 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5798 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5799 stubName = "cipherBlockChaining_decryptAESCrypt";
5800 break;
5801 }
5802 if (stubAddr == NULL) return false;
5803
5804 Node* cipherBlockChaining_object = argument(0);
5805 Node* src = argument(1);
5806 Node* src_offset = argument(2);
5807 Node* len = argument(3);
5808 Node* dest = argument(4);
5809 Node* dest_offset = argument(5);
5810
5811 // (1) src and dest are arrays.
5812 const Type* src_type = src->Value(&_gvn);
5813 const Type* dest_type = dest->Value(&_gvn);
5814 const TypeAryPtr* top_src = src_type->isa_aryptr();
5815 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5816 assert (top_src != NULL && top_src->klass() != NULL
5817 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5818
5819 // checks are the responsibility of the caller
5820 Node* src_start = src;
5821 Node* dest_start = dest;
5822 if (src_offset != NULL || dest_offset != NULL) {
5823 assert(src_offset != NULL && dest_offset != NULL, "");
5824 src_start = array_element_address(src, src_offset, T_BYTE);
5825 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5826 }
5827
5828 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5829 // (because of the predicated logic executed earlier).
5830 // so we cast it here safely.
5831 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5832
5833 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5834 if (embeddedCipherObj == NULL) return false;
5835
5836 // cast it to what we know it will be at runtime
5837 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5838 assert(tinst != NULL, "CBC obj is null");
5839 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5840 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5841 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5842
5843 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5844 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5845 const TypeOopPtr* xtype = aklass->as_instance_type();
5846 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5847 aescrypt_object = _gvn.transform(aescrypt_object);
5848
5849 // we need to get the start of the aescrypt_object's expanded key array
5850 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5851 if (k_start == NULL) return false;
5852
5853 // similarly, get the start address of the r vector
5854 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5855 if (objRvec == NULL) return false;
5856 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5857
5858 Node* cbcCrypt;
5859 if (Matcher::pass_original_key_for_aes()) {
5860 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5861 // compatibility issues between Java key expansion and SPARC crypto instructions
5862 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5863 if (original_k_start == NULL) return false;
5864
5865 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5866 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5867 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5868 stubAddr, stubName, TypePtr::BOTTOM,
5869 src_start, dest_start, k_start, r_start, len, original_k_start);
5870 } else {
5871 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5872 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5873 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5874 stubAddr, stubName, TypePtr::BOTTOM,
5875 src_start, dest_start, k_start, r_start, len);
5876 }
5877
5878 // return cipher length (int)
5879 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5880 set_result(retvalue);
5881 return true;
5882 }
5883
5884 //------------------------------get_key_start_from_aescrypt_object-----------------------
5885 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5886 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5887 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5888 if (objAESCryptKey == NULL) return (Node *) NULL;
5889
5890 // now have the array, need to get the start address of the K array
5891 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5892 return k_start;
5893 }
5894
5895 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5896 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5897 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5898 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5899 if (objAESCryptKey == NULL) return (Node *) NULL;
5900
5901 // now have the array, need to get the start address of the lastKey array
5902 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5903 return original_k_start;
5904 }
5905
5906 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5907 // Return node representing slow path of predicate check.
5908 // the pseudo code we want to emulate with this predicate is:
5909 // for encryption:
5910 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5911 // for decryption:
5912 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5913 // note cipher==plain is more conservative than the original java code but that's OK
5914 //
5915 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5916 // The receiver was checked for NULL already.
5917 Node* objCBC = argument(0);
5918
5919 // Load embeddedCipher field of CipherBlockChaining object.
5920 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5921
5922 // get AESCrypt klass for instanceOf check
5923 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5924 // will have same classloader as CipherBlockChaining object
5925 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5926 assert(tinst != NULL, "CBCobj is null");
5927 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5928
5929 // we want to do an instanceof comparison against the AESCrypt class
5930 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5931 if (!klass_AESCrypt->is_loaded()) {
5932 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5933 Node* ctrl = control();
5934 set_control(top()); // no regular fast path
5935 return ctrl;
5936 }
5937 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5938
5939 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5940 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5941 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5942
5943 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5944
5945 // for encryption, we are done
5946 if (!decrypting)
5947 return instof_false; // even if it is NULL
5948
5949 // for decryption, we need to add a further check to avoid
5950 // taking the intrinsic path when cipher and plain are the same
5951 // see the original java code for why.
5952 RegionNode* region = new RegionNode(3);
5953 region->init_req(1, instof_false);
5954 Node* src = argument(1);
5955 Node* dest = argument(4);
5956 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5957 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5958 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5959 region->init_req(2, src_dest_conjoint);
5960
5961 record_for_igvn(region);
5962 return _gvn.transform(region);
5963 }
5964
5965 //------------------------------inline_ghash_processBlocks
5966 bool LibraryCallKit::inline_ghash_processBlocks() {
5967 address stubAddr;
5968 const char *stubName;
5969 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5970
5971 stubAddr = StubRoutines::ghash_processBlocks();
5972 stubName = "ghash_processBlocks";
5973
5974 Node* data = argument(0);
5975 Node* offset = argument(1);
5976 Node* len = argument(2);
5977 Node* state = argument(3);
5978 Node* subkeyH = argument(4);
5979
5980 Node* state_start = array_element_address(state, intcon(0), T_LONG);
5981 assert(state_start, "state is NULL");
5982 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
5983 assert(subkeyH_start, "subkeyH is NULL");
5984 Node* data_start = array_element_address(data, offset, T_BYTE);
5985 assert(data_start, "data is NULL");
5986
5987 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
5988 OptoRuntime::ghash_processBlocks_Type(),
5989 stubAddr, stubName, TypePtr::BOTTOM,
5990 state_start, subkeyH_start, data_start, len);
5991 return true;
5992 }
5993
5994 //------------------------------inline_sha_implCompress-----------------------
5995 //
5996 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5997 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5998 //
5999 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6000 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6001 //
6002 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6003 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6004 //
6005 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6006 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6007
6008 Node* sha_obj = argument(0);
6009 Node* src = argument(1); // type oop
6010 Node* ofs = argument(2); // type int
6011
6012 const Type* src_type = src->Value(&_gvn);
6013 const TypeAryPtr* top_src = src_type->isa_aryptr();
6014 if (top_src == NULL || top_src->klass() == NULL) {
6015 // failed array check
6016 return false;
6017 }
6018 // Figure out the size and type of the elements we will be copying.
6019 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6020 if (src_elem != T_BYTE) {
6021 return false;
6022 }
6023 // 'src_start' points to src array + offset
6024 Node* src_start = array_element_address(src, ofs, src_elem);
6025 Node* state = NULL;
6026 address stubAddr;
6027 const char *stubName;
6028
6029 switch(id) {
6030 case vmIntrinsics::_sha_implCompress:
6031 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6032 state = get_state_from_sha_object(sha_obj);
6033 stubAddr = StubRoutines::sha1_implCompress();
6034 stubName = "sha1_implCompress";
6035 break;
6036 case vmIntrinsics::_sha2_implCompress:
6037 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6038 state = get_state_from_sha_object(sha_obj);
6039 stubAddr = StubRoutines::sha256_implCompress();
6040 stubName = "sha256_implCompress";
6041 break;
6042 case vmIntrinsics::_sha5_implCompress:
6043 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6044 state = get_state_from_sha5_object(sha_obj);
6045 stubAddr = StubRoutines::sha512_implCompress();
6046 stubName = "sha512_implCompress";
6047 break;
6048 default:
6049 fatal_unexpected_iid(id);
6050 return false;
6051 }
6052 if (state == NULL) return false;
6053
6054 // Call the stub.
6055 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6056 stubAddr, stubName, TypePtr::BOTTOM,
6057 src_start, state);
6058
6059 return true;
6060 }
6061
6062 //------------------------------inline_digestBase_implCompressMB-----------------------
6063 //
6064 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6065 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6066 //
6067 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6068 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6069 "need SHA1/SHA256/SHA512 instruction support");
6070 assert((uint)predicate < 3, "sanity");
6071 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6072
6073 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6074 Node* src = argument(1); // byte[] array
6075 Node* ofs = argument(2); // type int
6076 Node* limit = argument(3); // type int
6077
6078 const Type* src_type = src->Value(&_gvn);
6079 const TypeAryPtr* top_src = src_type->isa_aryptr();
6080 if (top_src == NULL || top_src->klass() == NULL) {
6081 // failed array check
6082 return false;
6083 }
6084 // Figure out the size and type of the elements we will be copying.
6085 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6086 if (src_elem != T_BYTE) {
6087 return false;
6088 }
6089 // 'src_start' points to src array + offset
6090 Node* src_start = array_element_address(src, ofs, src_elem);
6091
6092 const char* klass_SHA_name = NULL;
6093 const char* stub_name = NULL;
6094 address stub_addr = NULL;
6095 bool long_state = false;
6096
6097 switch (predicate) {
6098 case 0:
6099 if (UseSHA1Intrinsics) {
6100 klass_SHA_name = "sun/security/provider/SHA";
6101 stub_name = "sha1_implCompressMB";
6102 stub_addr = StubRoutines::sha1_implCompressMB();
6103 }
6104 break;
6105 case 1:
6106 if (UseSHA256Intrinsics) {
6107 klass_SHA_name = "sun/security/provider/SHA2";
6108 stub_name = "sha256_implCompressMB";
6109 stub_addr = StubRoutines::sha256_implCompressMB();
6110 }
6111 break;
6112 case 2:
6113 if (UseSHA512Intrinsics) {
6114 klass_SHA_name = "sun/security/provider/SHA5";
6115 stub_name = "sha512_implCompressMB";
6116 stub_addr = StubRoutines::sha512_implCompressMB();
6117 long_state = true;
6118 }
6119 break;
6120 default:
6121 fatal("unknown SHA intrinsic predicate: %d", predicate);
6122 }
6123 if (klass_SHA_name != NULL) {
6124 // get DigestBase klass to lookup for SHA klass
6125 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6126 assert(tinst != NULL, "digestBase_obj is not instance???");
6127 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6128
6129 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6130 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6131 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6132 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6133 }
6134 return false;
6135 }
6136 //------------------------------inline_sha_implCompressMB-----------------------
6137 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6138 bool long_state, address stubAddr, const char *stubName,
6139 Node* src_start, Node* ofs, Node* limit) {
6140 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6141 const TypeOopPtr* xtype = aklass->as_instance_type();
6142 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6143 sha_obj = _gvn.transform(sha_obj);
6144
6145 Node* state;
6146 if (long_state) {
6147 state = get_state_from_sha5_object(sha_obj);
6148 } else {
6149 state = get_state_from_sha_object(sha_obj);
6150 }
6151 if (state == NULL) return false;
6152
6153 // Call the stub.
6154 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6155 OptoRuntime::digestBase_implCompressMB_Type(),
6156 stubAddr, stubName, TypePtr::BOTTOM,
6157 src_start, state, ofs, limit);
6158 // return ofs (int)
6159 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6160 set_result(result);
6161
6162 return true;
6163 }
6164
6165 //------------------------------get_state_from_sha_object-----------------------
6166 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6167 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6168 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6169 if (sha_state == NULL) return (Node *) NULL;
6170
6171 // now have the array, need to get the start address of the state array
6172 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6173 return state;
6174 }
6175
6176 //------------------------------get_state_from_sha5_object-----------------------
6177 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6178 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6179 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6180 if (sha_state == NULL) return (Node *) NULL;
6181
6182 // now have the array, need to get the start address of the state array
6183 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6184 return state;
6185 }
6186
6187 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6188 // Return node representing slow path of predicate check.
6189 // the pseudo code we want to emulate with this predicate is:
6190 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6191 //
6192 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6193 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6194 "need SHA1/SHA256/SHA512 instruction support");
6195 assert((uint)predicate < 3, "sanity");
6196
6197 // The receiver was checked for NULL already.
6198 Node* digestBaseObj = argument(0);
6199
6200 // get DigestBase klass for instanceOf check
6201 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6202 assert(tinst != NULL, "digestBaseObj is null");
6203 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6204
6205 const char* klass_SHA_name = NULL;
6206 switch (predicate) {
6207 case 0:
6208 if (UseSHA1Intrinsics) {
6209 // we want to do an instanceof comparison against the SHA class
6210 klass_SHA_name = "sun/security/provider/SHA";
6211 }
6212 break;
6213 case 1:
6214 if (UseSHA256Intrinsics) {
6215 // we want to do an instanceof comparison against the SHA2 class
6216 klass_SHA_name = "sun/security/provider/SHA2";
6217 }
6218 break;
6219 case 2:
6220 if (UseSHA512Intrinsics) {
6221 // we want to do an instanceof comparison against the SHA5 class
6222 klass_SHA_name = "sun/security/provider/SHA5";
6223 }
6224 break;
6225 default:
6226 fatal("unknown SHA intrinsic predicate: %d", predicate);
6227 }
6228
6229 ciKlass* klass_SHA = NULL;
6230 if (klass_SHA_name != NULL) {
6231 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6232 }
6233 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6234 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6235 Node* ctrl = control();
6236 set_control(top()); // no intrinsic path
6237 return ctrl;
6238 }
6239 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6240
6241 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6242 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6243 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6244 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6245
6246 return instof_false; // even if it is NULL
6247 }
6248
6249 bool LibraryCallKit::inline_profileBoolean() {
6250 Node* counts = argument(1);
6251 const TypeAryPtr* ary = NULL;
6252 ciArray* aobj = NULL;
6253 if (counts->is_Con()
6254 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6255 && (aobj = ary->const_oop()->as_array()) != NULL
6256 && (aobj->length() == 2)) {
6257 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6258 jint false_cnt = aobj->element_value(0).as_int();
6259 jint true_cnt = aobj->element_value(1).as_int();
6260
6261 if (C->log() != NULL) {
6262 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6263 false_cnt, true_cnt);
6264 }
6265
6266 if (false_cnt + true_cnt == 0) {
6267 // According to profile, never executed.
6268 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6269 Deoptimization::Action_reinterpret);
6270 return true;
6271 }
6272
6273 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6274 // is a number of each value occurrences.
6275 Node* result = argument(0);
6276 if (false_cnt == 0 || true_cnt == 0) {
6277 // According to profile, one value has been never seen.
6278 int expected_val = (false_cnt == 0) ? 1 : 0;
6279
6280 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6281 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6282
6283 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6284 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6285 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6286
6287 { // Slow path: uncommon trap for never seen value and then reexecute
6288 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6289 // the value has been seen at least once.
6290 PreserveJVMState pjvms(this);
6291 PreserveReexecuteState preexecs(this);
6292 jvms()->set_should_reexecute(true);
6293
6294 set_control(slow_path);
6295 set_i_o(i_o());
6296
6297 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6298 Deoptimization::Action_reinterpret);
6299 }
6300 // The guard for never seen value enables sharpening of the result and
6301 // returning a constant. It allows to eliminate branches on the same value
6302 // later on.
6303 set_control(fast_path);
6304 result = intcon(expected_val);
6305 }
6306 // Stop profiling.
6307 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6308 // By replacing method body with profile data (represented as ProfileBooleanNode
6309 // on IR level) we effectively disable profiling.
6310 // It enables full speed execution once optimized code is generated.
6311 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6312 C->record_for_igvn(profile);
6313 set_result(profile);
6314 return true;
6315 } else {
6316 // Continue profiling.
6317 // Profile data isn't available at the moment. So, execute method's bytecode version.
6318 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6319 // is compiled and counters aren't available since corresponding MethodHandle
6320 // isn't a compile-time constant.
6321 return false;
6322 }
6323 }
6324
6325 bool LibraryCallKit::inline_isCompileConstant() {
6326 Node* n = argument(0);
6327 set_result(n->is_Con() ? intcon(1) : intcon(0));
6328 return true;
6329 }