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 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
330 !C->directive()->is_intrinsic_disabled(mh) &&
331 !vmIntrinsics::is_disabled_by_flags(mh);
332
333 }
334
335 if (is_available) {
336 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
337 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
338 return new LibraryIntrinsic(m, is_virtual,
339 vmIntrinsics::predicates_needed(id),
340 vmIntrinsics::does_virtual_dispatch(id),
341 (vmIntrinsics::ID) id);
342 } else {
343 return NULL;
344 }
345 }
346
347 //----------------------register_library_intrinsics-----------------------
348 // Initialize this file's data structures, for each Compile instance.
349 void Compile::register_library_intrinsics() {
350 // Nothing to do here.
351 }
352
353 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
354 LibraryCallKit kit(jvms, this);
355 Compile* C = kit.C;
356 int nodes = C->unique();
357 #ifndef PRODUCT
358 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
359 char buf[1000];
360 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
361 tty->print_cr("Intrinsic %s", str);
362 }
363 #endif
364 ciMethod* callee = kit.callee();
365 const int bci = kit.bci();
366
367 // Try to inline the intrinsic.
368 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
369 kit.try_to_inline(_last_predicate)) {
370 if (C->print_intrinsics() || C->print_inlining()) {
371 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
372 }
373 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
374 if (C->log()) {
375 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
376 vmIntrinsics::name_at(intrinsic_id()),
377 (is_virtual() ? " virtual='1'" : ""),
378 C->unique() - nodes);
379 }
380 // Push the result from the inlined method onto the stack.
381 kit.push_result();
382 C->print_inlining_update(this);
383 return kit.transfer_exceptions_into_jvms();
384 }
385
386 // The intrinsic bailed out
387 if (C->print_intrinsics() || C->print_inlining()) {
388 if (jvms->has_method()) {
389 // Not a root compile.
390 const char* msg;
391 if (callee->intrinsic_candidate()) {
392 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
393 } else {
394 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
395 : "failed to inline (intrinsic), method not annotated";
396 }
397 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
398 } else {
399 // Root compile
400 tty->print("Did not generate intrinsic %s%s at bci:%d in",
401 vmIntrinsics::name_at(intrinsic_id()),
402 (is_virtual() ? " (virtual)" : ""), bci);
403 }
404 }
405 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
406 C->print_inlining_update(this);
407 return NULL;
408 }
409
410 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
411 LibraryCallKit kit(jvms, this);
412 Compile* C = kit.C;
413 int nodes = C->unique();
414 _last_predicate = predicate;
415 #ifndef PRODUCT
416 assert(is_predicated() && predicate < predicates_count(), "sanity");
417 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
418 char buf[1000];
419 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
420 tty->print_cr("Predicate for intrinsic %s", str);
421 }
422 #endif
423 ciMethod* callee = kit.callee();
424 const int bci = kit.bci();
425
426 Node* slow_ctl = kit.try_to_predicate(predicate);
427 if (!kit.failing()) {
428 if (C->print_intrinsics() || C->print_inlining()) {
429 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
430 }
431 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
432 if (C->log()) {
433 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
434 vmIntrinsics::name_at(intrinsic_id()),
435 (is_virtual() ? " virtual='1'" : ""),
436 C->unique() - nodes);
437 }
438 return slow_ctl; // Could be NULL if the check folds.
439 }
440
441 // The intrinsic bailed out
442 if (C->print_intrinsics() || C->print_inlining()) {
443 if (jvms->has_method()) {
444 // Not a root compile.
445 const char* msg = "failed to generate predicate for intrinsic";
446 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
447 } else {
448 // Root compile
449 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
450 vmIntrinsics::name_at(intrinsic_id()),
451 (is_virtual() ? " (virtual)" : ""), bci);
452 }
453 }
454 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
455 return NULL;
456 }
457
458 bool LibraryCallKit::try_to_inline(int predicate) {
459 // Handle symbolic names for otherwise undistinguished boolean switches:
460 const bool is_store = true;
461 const bool is_native_ptr = true;
462 const bool is_static = true;
463 const bool is_volatile = true;
464
465 if (!jvms()->has_method()) {
466 // Root JVMState has a null method.
467 assert(map()->memory()->Opcode() == Op_Parm, "");
468 // Insert the memory aliasing node
469 set_all_memory(reset_memory());
470 }
471 assert(merged_memory(), "");
472
473
474 switch (intrinsic_id()) {
475 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
476 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
477 case vmIntrinsics::_getClass: return inline_native_getClass();
478
479 case vmIntrinsics::_dsin:
480 case vmIntrinsics::_dcos:
481 case vmIntrinsics::_dtan:
482 case vmIntrinsics::_dabs:
483 case vmIntrinsics::_datan2:
484 case vmIntrinsics::_dsqrt:
485 case vmIntrinsics::_dexp:
486 case vmIntrinsics::_dlog:
487 case vmIntrinsics::_dlog10:
488 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
489
490 case vmIntrinsics::_min:
491 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
492
493 case vmIntrinsics::_notify:
494 case vmIntrinsics::_notifyAll:
495 if (InlineNotify) {
496 return inline_notify(intrinsic_id());
497 }
498 return false;
499
500 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
501 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
502 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
503 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
504 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
505 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
506 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
507 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
508 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
509 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
510 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
511 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
512
513 case vmIntrinsics::_arraycopy: return inline_arraycopy();
514
515 case vmIntrinsics::_compareTo: return inline_string_compareTo();
516 case vmIntrinsics::_indexOf: return inline_string_indexOf();
517 case vmIntrinsics::_equals: return inline_string_equals();
518
519 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
520 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
521 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
522 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
523 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
524 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
525 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
526 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
527 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
528 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
529 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
530 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
531 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
532 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
533 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
534 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
535 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
536 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
537
538 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
539 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
540 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
541 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
542 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
543 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
544 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
545 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
546
547 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
548 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
549 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
550 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
551 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
552 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
553 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
554 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
555
556 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
557 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
558 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
559 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
560 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
561 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
562 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
563 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
564 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
565
566 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
567 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
568 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
569 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
570 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
571 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
572 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
573 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
574 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
575
576 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
577 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
578 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
579 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
580
581 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
582 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
583 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
584 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
585
586 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
587 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
588 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
589
590 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
591 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
592 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
593
594 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
595 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
596 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
597 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
598 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
599
600 case vmIntrinsics::_loadFence:
601 case vmIntrinsics::_storeFence:
602 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
603
604 case vmIntrinsics::_currentThread: return inline_native_currentThread();
605 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
606
607 #ifdef TRACE_HAVE_INTRINSICS
608 case vmIntrinsics::_classID: return inline_native_classID();
609 case vmIntrinsics::_threadID: return inline_native_threadID();
610 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
611 #endif
612 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
613 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
614 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
615 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
616 case vmIntrinsics::_newArray: return inline_native_newArray();
617 case vmIntrinsics::_getLength: return inline_native_getLength();
618 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
619 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
620 case vmIntrinsics::_equalsC: return inline_array_equals();
621 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
622
623 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
624
625 case vmIntrinsics::_isInstance:
626 case vmIntrinsics::_getModifiers:
627 case vmIntrinsics::_isInterface:
628 case vmIntrinsics::_isArray:
629 case vmIntrinsics::_isPrimitive:
630 case vmIntrinsics::_getSuperclass:
631 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
632
633 case vmIntrinsics::_floatToRawIntBits:
634 case vmIntrinsics::_floatToIntBits:
635 case vmIntrinsics::_intBitsToFloat:
636 case vmIntrinsics::_doubleToRawLongBits:
637 case vmIntrinsics::_doubleToLongBits:
638 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
639
640 case vmIntrinsics::_numberOfLeadingZeros_i:
641 case vmIntrinsics::_numberOfLeadingZeros_l:
642 case vmIntrinsics::_numberOfTrailingZeros_i:
643 case vmIntrinsics::_numberOfTrailingZeros_l:
644 case vmIntrinsics::_bitCount_i:
645 case vmIntrinsics::_bitCount_l:
646 case vmIntrinsics::_reverseBytes_i:
647 case vmIntrinsics::_reverseBytes_l:
648 case vmIntrinsics::_reverseBytes_s:
649 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
650
651 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
652
653 case vmIntrinsics::_Reference_get: return inline_reference_get();
654
655 case vmIntrinsics::_Class_cast: return inline_Class_cast();
656
657 case vmIntrinsics::_aescrypt_encryptBlock:
658 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
659
660 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
661 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
662 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
663
664 case vmIntrinsics::_sha_implCompress:
665 case vmIntrinsics::_sha2_implCompress:
666 case vmIntrinsics::_sha5_implCompress:
667 return inline_sha_implCompress(intrinsic_id());
668
669 case vmIntrinsics::_digestBase_implCompressMB:
670 return inline_digestBase_implCompressMB(predicate);
671
672 case vmIntrinsics::_multiplyToLen:
673 return inline_multiplyToLen();
674
675 case vmIntrinsics::_squareToLen:
676 return inline_squareToLen();
677
678 case vmIntrinsics::_mulAdd:
679 return inline_mulAdd();
680
681 case vmIntrinsics::_montgomeryMultiply:
682 return inline_montgomeryMultiply();
683 case vmIntrinsics::_montgomerySquare:
684 return inline_montgomerySquare();
685
686 case vmIntrinsics::_ghash_processBlocks:
687 return inline_ghash_processBlocks();
688
689 case vmIntrinsics::_encodeISOArray:
690 return inline_encodeISOArray();
691
692 case vmIntrinsics::_updateCRC32:
693 return inline_updateCRC32();
694 case vmIntrinsics::_updateBytesCRC32:
695 return inline_updateBytesCRC32();
696 case vmIntrinsics::_updateByteBufferCRC32:
697 return inline_updateByteBufferCRC32();
698
699 case vmIntrinsics::_updateBytesCRC32C:
700 return inline_updateBytesCRC32C();
701 case vmIntrinsics::_updateDirectByteBufferCRC32C:
702 return inline_updateDirectByteBufferCRC32C();
703
704 case vmIntrinsics::_updateBytesAdler32:
705 return inline_updateBytesAdler32();
706 case vmIntrinsics::_updateByteBufferAdler32:
707 return inline_updateByteBufferAdler32();
708
709 case vmIntrinsics::_profileBoolean:
710 return inline_profileBoolean();
711 case vmIntrinsics::_isCompileConstant:
712 return inline_isCompileConstant();
713
714 default:
715 // If you get here, it may be that someone has added a new intrinsic
716 // to the list in vmSymbols.hpp without implementing it here.
717 #ifndef PRODUCT
718 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
719 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
720 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
721 }
722 #endif
723 return false;
724 }
725 }
726
727 Node* LibraryCallKit::try_to_predicate(int predicate) {
728 if (!jvms()->has_method()) {
729 // Root JVMState has a null method.
730 assert(map()->memory()->Opcode() == Op_Parm, "");
731 // Insert the memory aliasing node
732 set_all_memory(reset_memory());
733 }
734 assert(merged_memory(), "");
735
736 switch (intrinsic_id()) {
737 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
738 return inline_cipherBlockChaining_AESCrypt_predicate(false);
739 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
740 return inline_cipherBlockChaining_AESCrypt_predicate(true);
741 case vmIntrinsics::_digestBase_implCompressMB:
742 return inline_digestBase_implCompressMB_predicate(predicate);
743
744 default:
745 // If you get here, it may be that someone has added a new intrinsic
746 // to the list in vmSymbols.hpp without implementing it here.
747 #ifndef PRODUCT
748 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
749 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
750 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
751 }
752 #endif
753 Node* slow_ctl = control();
754 set_control(top()); // No fast path instrinsic
755 return slow_ctl;
756 }
757 }
758
759 //------------------------------set_result-------------------------------
760 // Helper function for finishing intrinsics.
761 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
762 record_for_igvn(region);
763 set_control(_gvn.transform(region));
764 set_result( _gvn.transform(value));
765 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
766 }
767
768 //------------------------------generate_guard---------------------------
769 // Helper function for generating guarded fast-slow graph structures.
770 // The given 'test', if true, guards a slow path. If the test fails
771 // then a fast path can be taken. (We generally hope it fails.)
772 // In all cases, GraphKit::control() is updated to the fast path.
773 // The returned value represents the control for the slow path.
774 // The return value is never 'top'; it is either a valid control
775 // or NULL if it is obvious that the slow path can never be taken.
776 // Also, if region and the slow control are not NULL, the slow edge
777 // is appended to the region.
778 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
779 if (stopped()) {
780 // Already short circuited.
781 return NULL;
782 }
783
784 // Build an if node and its projections.
785 // If test is true we take the slow path, which we assume is uncommon.
786 if (_gvn.type(test) == TypeInt::ZERO) {
787 // The slow branch is never taken. No need to build this guard.
788 return NULL;
789 }
790
791 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
792
793 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
794 if (if_slow == top()) {
795 // The slow branch is never taken. No need to build this guard.
796 return NULL;
797 }
798
799 if (region != NULL)
800 region->add_req(if_slow);
801
802 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
803 set_control(if_fast);
804
805 return if_slow;
806 }
807
808 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
809 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
810 }
811 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
812 return generate_guard(test, region, PROB_FAIR);
813 }
814
815 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
816 Node* *pos_index) {
817 if (stopped())
818 return NULL; // already stopped
819 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
820 return NULL; // index is already adequately typed
821 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
822 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
823 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
824 if (is_neg != NULL && pos_index != NULL) {
825 // Emulate effect of Parse::adjust_map_after_if.
826 Node* ccast = new CastIINode(index, TypeInt::POS);
827 ccast->set_req(0, control());
828 (*pos_index) = _gvn.transform(ccast);
829 }
830 return is_neg;
831 }
832
833 // Make sure that 'position' is a valid limit index, in [0..length].
834 // There are two equivalent plans for checking this:
835 // A. (offset + copyLength) unsigned<= arrayLength
836 // B. offset <= (arrayLength - copyLength)
837 // We require that all of the values above, except for the sum and
838 // difference, are already known to be non-negative.
839 // Plan A is robust in the face of overflow, if offset and copyLength
840 // are both hugely positive.
841 //
842 // Plan B is less direct and intuitive, but it does not overflow at
843 // all, since the difference of two non-negatives is always
844 // representable. Whenever Java methods must perform the equivalent
845 // check they generally use Plan B instead of Plan A.
846 // For the moment we use Plan A.
847 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
848 Node* subseq_length,
849 Node* array_length,
850 RegionNode* region) {
851 if (stopped())
852 return NULL; // already stopped
853 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
854 if (zero_offset && subseq_length->eqv_uncast(array_length))
855 return NULL; // common case of whole-array copy
856 Node* last = subseq_length;
857 if (!zero_offset) // last += offset
858 last = _gvn.transform(new AddINode(last, offset));
859 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
860 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
861 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
862 return is_over;
863 }
864
865
866 //--------------------------generate_current_thread--------------------
867 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
868 ciKlass* thread_klass = env()->Thread_klass();
869 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
870 Node* thread = _gvn.transform(new ThreadLocalNode());
871 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
872 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
873 tls_output = thread;
874 return threadObj;
875 }
876
877
878 //------------------------------make_string_method_node------------------------
879 // Helper method for String intrinsic functions. This version is called
880 // with str1 and str2 pointing to String object nodes.
881 //
882 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
883 Node* no_ctrl = NULL;
884
885 // Get start addr of string
886 Node* str1_value = load_String_value(no_ctrl, str1);
887 Node* str1_offset = load_String_offset(no_ctrl, str1);
888 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
889
890 // Get length of string 1
891 Node* str1_len = load_String_length(no_ctrl, str1);
892
893 Node* str2_value = load_String_value(no_ctrl, str2);
894 Node* str2_offset = load_String_offset(no_ctrl, str2);
895 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
896
897 Node* str2_len = NULL;
898 Node* result = NULL;
899
900 switch (opcode) {
901 case Op_StrIndexOf:
902 // Get length of string 2
903 str2_len = load_String_length(no_ctrl, str2);
904
905 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
906 str1_start, str1_len, str2_start, str2_len);
907 break;
908 case Op_StrComp:
909 // Get length of string 2
910 str2_len = load_String_length(no_ctrl, str2);
911
912 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
913 str1_start, str1_len, str2_start, str2_len);
914 break;
915 case Op_StrEquals:
916 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
917 str1_start, str2_start, str1_len);
918 break;
919 default:
920 ShouldNotReachHere();
921 return NULL;
922 }
923
924 // All these intrinsics have checks.
925 C->set_has_split_ifs(true); // Has chance for split-if optimization
926
927 return _gvn.transform(result);
928 }
929
930 // Helper method for String intrinsic functions. This version is called
931 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
932 // to Int nodes containing the lenghts of str1 and str2.
933 //
934 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
935 Node* result = NULL;
936 switch (opcode) {
937 case Op_StrIndexOf:
938 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
939 str1_start, cnt1, str2_start, cnt2);
940 break;
941 case Op_StrComp:
942 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
943 str1_start, cnt1, str2_start, cnt2);
944 break;
945 case Op_StrEquals:
946 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
947 str1_start, str2_start, cnt1);
948 break;
949 default:
950 ShouldNotReachHere();
951 return NULL;
952 }
953
954 // All these intrinsics have checks.
955 C->set_has_split_ifs(true); // Has chance for split-if optimization
956
957 return _gvn.transform(result);
958 }
959
960 //------------------------------inline_string_compareTo------------------------
961 // public int java.lang.String.compareTo(String anotherString);
962 bool LibraryCallKit::inline_string_compareTo() {
963 Node* receiver = null_check(argument(0));
964 Node* arg = null_check(argument(1));
965 if (stopped()) {
966 return true;
967 }
968 set_result(make_string_method_node(Op_StrComp, receiver, arg));
969 return true;
970 }
971
972 //------------------------------inline_string_equals------------------------
973 bool LibraryCallKit::inline_string_equals() {
974 Node* receiver = null_check_receiver();
975 // NOTE: Do not null check argument for String.equals() because spec
976 // allows to specify NULL as argument.
977 Node* argument = this->argument(1);
978 if (stopped()) {
979 return true;
980 }
981
982 // paths (plus control) merge
983 RegionNode* region = new RegionNode(5);
984 Node* phi = new PhiNode(region, TypeInt::BOOL);
985
986 // does source == target string?
987 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
988 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
989
990 Node* if_eq = generate_slow_guard(bol, NULL);
991 if (if_eq != NULL) {
992 // receiver == argument
993 phi->init_req(2, intcon(1));
994 region->init_req(2, if_eq);
995 }
996
997 // get String klass for instanceOf
998 ciInstanceKlass* klass = env()->String_klass();
999
1000 if (!stopped()) {
1001 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1002 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1003 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1004
1005 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1006 //instanceOf == true, fallthrough
1007
1008 if (inst_false != NULL) {
1009 phi->init_req(3, intcon(0));
1010 region->init_req(3, inst_false);
1011 }
1012 }
1013
1014 if (!stopped()) {
1015 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1016
1017 // Properly cast the argument to String
1018 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1019 // This path is taken only when argument's type is String:NotNull.
1020 argument = cast_not_null(argument, false);
1021
1022 Node* no_ctrl = NULL;
1023
1024 // Get start addr of receiver
1025 Node* receiver_val = load_String_value(no_ctrl, receiver);
1026 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1027 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1028
1029 // Get length of receiver
1030 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1031
1032 // Get start addr of argument
1033 Node* argument_val = load_String_value(no_ctrl, argument);
1034 Node* argument_offset = load_String_offset(no_ctrl, argument);
1035 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1036
1037 // Get length of argument
1038 Node* argument_cnt = load_String_length(no_ctrl, argument);
1039
1040 // Check for receiver count != argument count
1041 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1042 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1043 Node* if_ne = generate_slow_guard(bol, NULL);
1044 if (if_ne != NULL) {
1045 phi->init_req(4, intcon(0));
1046 region->init_req(4, if_ne);
1047 }
1048
1049 // Check for count == 0 is done by assembler code for StrEquals.
1050
1051 if (!stopped()) {
1052 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1053 phi->init_req(1, equals);
1054 region->init_req(1, control());
1055 }
1056 }
1057
1058 // post merge
1059 set_control(_gvn.transform(region));
1060 record_for_igvn(region);
1061
1062 set_result(_gvn.transform(phi));
1063 return true;
1064 }
1065
1066 //------------------------------inline_array_equals----------------------------
1067 bool LibraryCallKit::inline_array_equals() {
1068 Node* arg1 = argument(0);
1069 Node* arg2 = argument(1);
1070 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1071 return true;
1072 }
1073
1074 // Java version of String.indexOf(constant string)
1075 // class StringDecl {
1076 // StringDecl(char[] ca) {
1077 // offset = 0;
1078 // count = ca.length;
1079 // value = ca;
1080 // }
1081 // int offset;
1082 // int count;
1083 // char[] value;
1084 // }
1085 //
1086 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1087 // int targetOffset, int cache_i, int md2) {
1088 // int cache = cache_i;
1089 // int sourceOffset = string_object.offset;
1090 // int sourceCount = string_object.count;
1091 // int targetCount = target_object.length;
1092 //
1093 // int targetCountLess1 = targetCount - 1;
1094 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1095 //
1096 // char[] source = string_object.value;
1097 // char[] target = target_object;
1098 // int lastChar = target[targetCountLess1];
1099 //
1100 // outer_loop:
1101 // for (int i = sourceOffset; i < sourceEnd; ) {
1102 // int src = source[i + targetCountLess1];
1103 // if (src == lastChar) {
1104 // // With random strings and a 4-character alphabet,
1105 // // reverse matching at this point sets up 0.8% fewer
1106 // // frames, but (paradoxically) makes 0.3% more probes.
1107 // // Since those probes are nearer the lastChar probe,
1108 // // there is may be a net D$ win with reverse matching.
1109 // // But, reversing loop inhibits unroll of inner loop
1110 // // for unknown reason. So, does running outer loop from
1111 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1112 // for (int j = 0; j < targetCountLess1; j++) {
1113 // if (target[targetOffset + j] != source[i+j]) {
1114 // if ((cache & (1 << source[i+j])) == 0) {
1115 // if (md2 < j+1) {
1116 // i += j+1;
1117 // continue outer_loop;
1118 // }
1119 // }
1120 // i += md2;
1121 // continue outer_loop;
1122 // }
1123 // }
1124 // return i - sourceOffset;
1125 // }
1126 // if ((cache & (1 << src)) == 0) {
1127 // i += targetCountLess1;
1128 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1129 // i++;
1130 // }
1131 // return -1;
1132 // }
1133
1134 //------------------------------string_indexOf------------------------
1135 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1136 jint cache_i, jint md2_i) {
1137
1138 Node* no_ctrl = NULL;
1139 float likely = PROB_LIKELY(0.9);
1140 float unlikely = PROB_UNLIKELY(0.9);
1141
1142 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1143
1144 Node* source = load_String_value(no_ctrl, string_object);
1145 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1146 Node* sourceCount = load_String_length(no_ctrl, string_object);
1147
1148 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1149 jint target_length = target_array->length();
1150 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1151 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1152
1153 // String.value field is known to be @Stable.
1154 if (UseImplicitStableValues) {
1155 target = cast_array_to_stable(target, target_type);
1156 }
1157
1158 IdealKit kit(this, false, true);
1159 #define __ kit.
1160 Node* zero = __ ConI(0);
1161 Node* one = __ ConI(1);
1162 Node* cache = __ ConI(cache_i);
1163 Node* md2 = __ ConI(md2_i);
1164 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1165 Node* targetCountLess1 = __ ConI(target_length - 1);
1166 Node* targetOffset = __ ConI(targetOffset_i);
1167 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1168
1169 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1170 Node* outer_loop = __ make_label(2 /* goto */);
1171 Node* return_ = __ make_label(1);
1172
1173 __ set(rtn,__ ConI(-1));
1174 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1175 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1176 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1177 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1178 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1179 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1180 Node* tpj = __ AddI(targetOffset, __ value(j));
1181 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1182 Node* ipj = __ AddI(__ value(i), __ value(j));
1183 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1184 __ if_then(targ, BoolTest::ne, src2); {
1185 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1186 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1187 __ increment(i, __ AddI(__ value(j), one));
1188 __ goto_(outer_loop);
1189 } __ end_if(); __ dead(j);
1190 }__ end_if(); __ dead(j);
1191 __ increment(i, md2);
1192 __ goto_(outer_loop);
1193 }__ end_if();
1194 __ increment(j, one);
1195 }__ end_loop(); __ dead(j);
1196 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1197 __ goto_(return_);
1198 }__ end_if();
1199 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1200 __ increment(i, targetCountLess1);
1201 }__ end_if();
1202 __ increment(i, one);
1203 __ bind(outer_loop);
1204 }__ end_loop(); __ dead(i);
1205 __ bind(return_);
1206
1207 // Final sync IdealKit and GraphKit.
1208 final_sync(kit);
1209 Node* result = __ value(rtn);
1210 #undef __
1211 C->set_has_loops(true);
1212 return result;
1213 }
1214
1215 //------------------------------inline_string_indexOf------------------------
1216 bool LibraryCallKit::inline_string_indexOf() {
1217 Node* receiver = argument(0);
1218 Node* arg = argument(1);
1219
1220 Node* result;
1221 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1222 UseSSE42Intrinsics) {
1223 // Generate SSE4.2 version of indexOf
1224 // We currently only have match rules that use SSE4.2
1225
1226 receiver = null_check(receiver);
1227 arg = null_check(arg);
1228 if (stopped()) {
1229 return true;
1230 }
1231
1232 // Make the merge point
1233 RegionNode* result_rgn = new RegionNode(4);
1234 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1235 Node* no_ctrl = NULL;
1236
1237 // Get start addr of source string
1238 Node* source = load_String_value(no_ctrl, receiver);
1239 Node* source_offset = load_String_offset(no_ctrl, receiver);
1240 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1241
1242 // Get length of source string
1243 Node* source_cnt = load_String_length(no_ctrl, receiver);
1244
1245 // Get start addr of substring
1246 Node* substr = load_String_value(no_ctrl, arg);
1247 Node* substr_offset = load_String_offset(no_ctrl, arg);
1248 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1249
1250 // Get length of source string
1251 Node* substr_cnt = load_String_length(no_ctrl, arg);
1252
1253 // Check for substr count > string count
1254 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1255 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1256 Node* if_gt = generate_slow_guard(bol, NULL);
1257 if (if_gt != NULL) {
1258 result_phi->init_req(2, intcon(-1));
1259 result_rgn->init_req(2, if_gt);
1260 }
1261
1262 if (!stopped()) {
1263 // Check for substr count == 0
1264 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1265 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1266 Node* if_zero = generate_slow_guard(bol, NULL);
1267 if (if_zero != NULL) {
1268 result_phi->init_req(3, intcon(0));
1269 result_rgn->init_req(3, if_zero);
1270 }
1271 }
1272
1273 if (!stopped()) {
1274 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1275 result_phi->init_req(1, result);
1276 result_rgn->init_req(1, control());
1277 }
1278 set_control(_gvn.transform(result_rgn));
1279 record_for_igvn(result_rgn);
1280 result = _gvn.transform(result_phi);
1281
1282 } else { // Use LibraryCallKit::string_indexOf
1283 // don't intrinsify if argument isn't a constant string.
1284 if (!arg->is_Con()) {
1285 return false;
1286 }
1287 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1288 if (str_type == NULL) {
1289 return false;
1290 }
1291 ciInstanceKlass* klass = env()->String_klass();
1292 ciObject* str_const = str_type->const_oop();
1293 if (str_const == NULL || str_const->klass() != klass) {
1294 return false;
1295 }
1296 ciInstance* str = str_const->as_instance();
1297 assert(str != NULL, "must be instance");
1298
1299 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1300 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1301
1302 int o;
1303 int c;
1304 if (java_lang_String::has_offset_field()) {
1305 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1306 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1307 } else {
1308 o = 0;
1309 c = pat->length();
1310 }
1311
1312 // constant strings have no offset and count == length which
1313 // simplifies the resulting code somewhat so lets optimize for that.
1314 if (o != 0 || c != pat->length()) {
1315 return false;
1316 }
1317
1318 receiver = null_check(receiver, T_OBJECT);
1319 // NOTE: No null check on the argument is needed since it's a constant String oop.
1320 if (stopped()) {
1321 return true;
1322 }
1323
1324 // The null string as a pattern always returns 0 (match at beginning of string)
1325 if (c == 0) {
1326 set_result(intcon(0));
1327 return true;
1328 }
1329
1330 // Generate default indexOf
1331 jchar lastChar = pat->char_at(o + (c - 1));
1332 int cache = 0;
1333 int i;
1334 for (i = 0; i < c - 1; i++) {
1335 assert(i < pat->length(), "out of range");
1336 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1337 }
1338
1339 int md2 = c;
1340 for (i = 0; i < c - 1; i++) {
1341 assert(i < pat->length(), "out of range");
1342 if (pat->char_at(o + i) == lastChar) {
1343 md2 = (c - 1) - i;
1344 }
1345 }
1346
1347 result = string_indexOf(receiver, pat, o, cache, md2);
1348 }
1349 set_result(result);
1350 return true;
1351 }
1352
1353 //--------------------------round_double_node--------------------------------
1354 // Round a double node if necessary.
1355 Node* LibraryCallKit::round_double_node(Node* n) {
1356 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1357 n = _gvn.transform(new RoundDoubleNode(0, n));
1358 return n;
1359 }
1360
1361 //------------------------------inline_math-----------------------------------
1362 // public static double Math.abs(double)
1363 // public static double Math.sqrt(double)
1364 // public static double Math.log(double)
1365 // public static double Math.log10(double)
1366 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1367 Node* arg = round_double_node(argument(0));
1368 Node* n;
1369 switch (id) {
1370 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1371 case vmIntrinsics::_dsqrt: n = new SqrtDNode(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:
1756 return StubRoutines::dlog() != NULL ?
1757 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1758 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1759 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1760 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1761
1762 // These intrinsics are supported on all hardware
1763 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1764 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1765
1766 case vmIntrinsics::_dexp:
1767 return StubRoutines::dexp() != NULL ?
1768 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1769 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1770 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1771 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1772 #undef FN_PTR
1773
1774 // These intrinsics are not yet correctly implemented
1775 case vmIntrinsics::_datan2:
1776 return false;
1777
1778 default:
1779 fatal_unexpected_iid(id);
1780 return false;
1781 }
1782 }
1783
1784 static bool is_simple_name(Node* n) {
1785 return (n->req() == 1 // constant
1786 || (n->is_Type() && n->as_Type()->type()->singleton())
1787 || n->is_Proj() // parameter or return value
1788 || n->is_Phi() // local of some sort
1789 );
1790 }
1791
1792 //----------------------------inline_notify-----------------------------------*
1793 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1794 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1795 address func;
1796 if (id == vmIntrinsics::_notify) {
1797 func = OptoRuntime::monitor_notify_Java();
1798 } else {
1799 func = OptoRuntime::monitor_notifyAll_Java();
1800 }
1801 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1802 make_slow_call_ex(call, env()->Throwable_klass(), false);
1803 return true;
1804 }
1805
1806
1807 //----------------------------inline_min_max-----------------------------------
1808 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1809 set_result(generate_min_max(id, argument(0), argument(1)));
1810 return true;
1811 }
1812
1813 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1814 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1815 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1816 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1817 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1818
1819 {
1820 PreserveJVMState pjvms(this);
1821 PreserveReexecuteState preexecs(this);
1822 jvms()->set_should_reexecute(true);
1823
1824 set_control(slow_path);
1825 set_i_o(i_o());
1826
1827 uncommon_trap(Deoptimization::Reason_intrinsic,
1828 Deoptimization::Action_none);
1829 }
1830
1831 set_control(fast_path);
1832 set_result(math);
1833 }
1834
1835 template <typename OverflowOp>
1836 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1837 typedef typename OverflowOp::MathOp MathOp;
1838
1839 MathOp* mathOp = new MathOp(arg1, arg2);
1840 Node* operation = _gvn.transform( mathOp );
1841 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1842 inline_math_mathExact(operation, ofcheck);
1843 return true;
1844 }
1845
1846 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1847 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1848 }
1849
1850 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1851 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1852 }
1853
1854 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1855 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1856 }
1857
1858 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1859 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1860 }
1861
1862 bool LibraryCallKit::inline_math_negateExactI() {
1863 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1864 }
1865
1866 bool LibraryCallKit::inline_math_negateExactL() {
1867 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1868 }
1869
1870 bool LibraryCallKit::inline_math_multiplyExactI() {
1871 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1872 }
1873
1874 bool LibraryCallKit::inline_math_multiplyExactL() {
1875 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1876 }
1877
1878 Node*
1879 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1880 // These are the candidate return value:
1881 Node* xvalue = x0;
1882 Node* yvalue = y0;
1883
1884 if (xvalue == yvalue) {
1885 return xvalue;
1886 }
1887
1888 bool want_max = (id == vmIntrinsics::_max);
1889
1890 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1891 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1892 if (txvalue == NULL || tyvalue == NULL) return top();
1893 // This is not really necessary, but it is consistent with a
1894 // hypothetical MaxINode::Value method:
1895 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1896
1897 // %%% This folding logic should (ideally) be in a different place.
1898 // Some should be inside IfNode, and there to be a more reliable
1899 // transformation of ?: style patterns into cmoves. We also want
1900 // more powerful optimizations around cmove and min/max.
1901
1902 // Try to find a dominating comparison of these guys.
1903 // It can simplify the index computation for Arrays.copyOf
1904 // and similar uses of System.arraycopy.
1905 // First, compute the normalized version of CmpI(x, y).
1906 int cmp_op = Op_CmpI;
1907 Node* xkey = xvalue;
1908 Node* ykey = yvalue;
1909 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1910 if (ideal_cmpxy->is_Cmp()) {
1911 // E.g., if we have CmpI(length - offset, count),
1912 // it might idealize to CmpI(length, count + offset)
1913 cmp_op = ideal_cmpxy->Opcode();
1914 xkey = ideal_cmpxy->in(1);
1915 ykey = ideal_cmpxy->in(2);
1916 }
1917
1918 // Start by locating any relevant comparisons.
1919 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1920 Node* cmpxy = NULL;
1921 Node* cmpyx = NULL;
1922 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1923 Node* cmp = start_from->fast_out(k);
1924 if (cmp->outcnt() > 0 && // must have prior uses
1925 cmp->in(0) == NULL && // must be context-independent
1926 cmp->Opcode() == cmp_op) { // right kind of compare
1927 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1928 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1929 }
1930 }
1931
1932 const int NCMPS = 2;
1933 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1934 int cmpn;
1935 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1936 if (cmps[cmpn] != NULL) break; // find a result
1937 }
1938 if (cmpn < NCMPS) {
1939 // Look for a dominating test that tells us the min and max.
1940 int depth = 0; // Limit search depth for speed
1941 Node* dom = control();
1942 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1943 if (++depth >= 100) break;
1944 Node* ifproj = dom;
1945 if (!ifproj->is_Proj()) continue;
1946 Node* iff = ifproj->in(0);
1947 if (!iff->is_If()) continue;
1948 Node* bol = iff->in(1);
1949 if (!bol->is_Bool()) continue;
1950 Node* cmp = bol->in(1);
1951 if (cmp == NULL) continue;
1952 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1953 if (cmps[cmpn] == cmp) break;
1954 if (cmpn == NCMPS) continue;
1955 BoolTest::mask btest = bol->as_Bool()->_test._test;
1956 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1957 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1958 // At this point, we know that 'x btest y' is true.
1959 switch (btest) {
1960 case BoolTest::eq:
1961 // They are proven equal, so we can collapse the min/max.
1962 // Either value is the answer. Choose the simpler.
1963 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1964 return yvalue;
1965 return xvalue;
1966 case BoolTest::lt: // x < y
1967 case BoolTest::le: // x <= y
1968 return (want_max ? yvalue : xvalue);
1969 case BoolTest::gt: // x > y
1970 case BoolTest::ge: // x >= y
1971 return (want_max ? xvalue : yvalue);
1972 }
1973 }
1974 }
1975
1976 // We failed to find a dominating test.
1977 // Let's pick a test that might GVN with prior tests.
1978 Node* best_bol = NULL;
1979 BoolTest::mask best_btest = BoolTest::illegal;
1980 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1981 Node* cmp = cmps[cmpn];
1982 if (cmp == NULL) continue;
1983 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1984 Node* bol = cmp->fast_out(j);
1985 if (!bol->is_Bool()) continue;
1986 BoolTest::mask btest = bol->as_Bool()->_test._test;
1987 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1988 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1989 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1990 best_bol = bol->as_Bool();
1991 best_btest = btest;
1992 }
1993 }
1994 }
1995
1996 Node* answer_if_true = NULL;
1997 Node* answer_if_false = NULL;
1998 switch (best_btest) {
1999 default:
2000 if (cmpxy == NULL)
2001 cmpxy = ideal_cmpxy;
2002 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2003 // and fall through:
2004 case BoolTest::lt: // x < y
2005 case BoolTest::le: // x <= y
2006 answer_if_true = (want_max ? yvalue : xvalue);
2007 answer_if_false = (want_max ? xvalue : yvalue);
2008 break;
2009 case BoolTest::gt: // x > y
2010 case BoolTest::ge: // x >= y
2011 answer_if_true = (want_max ? xvalue : yvalue);
2012 answer_if_false = (want_max ? yvalue : xvalue);
2013 break;
2014 }
2015
2016 jint hi, lo;
2017 if (want_max) {
2018 // We can sharpen the minimum.
2019 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2020 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2021 } else {
2022 // We can sharpen the maximum.
2023 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2024 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2025 }
2026
2027 // Use a flow-free graph structure, to avoid creating excess control edges
2028 // which could hinder other optimizations.
2029 // Since Math.min/max is often used with arraycopy, we want
2030 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2031 Node* cmov = CMoveNode::make(NULL, best_bol,
2032 answer_if_false, answer_if_true,
2033 TypeInt::make(lo, hi, widen));
2034
2035 return _gvn.transform(cmov);
2036
2037 /*
2038 // This is not as desirable as it may seem, since Min and Max
2039 // nodes do not have a full set of optimizations.
2040 // And they would interfere, anyway, with 'if' optimizations
2041 // and with CMoveI canonical forms.
2042 switch (id) {
2043 case vmIntrinsics::_min:
2044 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2045 case vmIntrinsics::_max:
2046 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2047 default:
2048 ShouldNotReachHere();
2049 }
2050 */
2051 }
2052
2053 inline int
2054 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2055 const TypePtr* base_type = TypePtr::NULL_PTR;
2056 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2057 if (base_type == NULL) {
2058 // Unknown type.
2059 return Type::AnyPtr;
2060 } else if (base_type == TypePtr::NULL_PTR) {
2061 // Since this is a NULL+long form, we have to switch to a rawptr.
2062 base = _gvn.transform(new CastX2PNode(offset));
2063 offset = MakeConX(0);
2064 return Type::RawPtr;
2065 } else if (base_type->base() == Type::RawPtr) {
2066 return Type::RawPtr;
2067 } else if (base_type->isa_oopptr()) {
2068 // Base is never null => always a heap address.
2069 if (base_type->ptr() == TypePtr::NotNull) {
2070 return Type::OopPtr;
2071 }
2072 // Offset is small => always a heap address.
2073 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2074 if (offset_type != NULL &&
2075 base_type->offset() == 0 && // (should always be?)
2076 offset_type->_lo >= 0 &&
2077 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2078 return Type::OopPtr;
2079 }
2080 // Otherwise, it might either be oop+off or NULL+addr.
2081 return Type::AnyPtr;
2082 } else {
2083 // No information:
2084 return Type::AnyPtr;
2085 }
2086 }
2087
2088 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2089 int kind = classify_unsafe_addr(base, offset);
2090 if (kind == Type::RawPtr) {
2091 return basic_plus_adr(top(), base, offset);
2092 } else {
2093 return basic_plus_adr(base, offset);
2094 }
2095 }
2096
2097 //--------------------------inline_number_methods-----------------------------
2098 // inline int Integer.numberOfLeadingZeros(int)
2099 // inline int Long.numberOfLeadingZeros(long)
2100 //
2101 // inline int Integer.numberOfTrailingZeros(int)
2102 // inline int Long.numberOfTrailingZeros(long)
2103 //
2104 // inline int Integer.bitCount(int)
2105 // inline int Long.bitCount(long)
2106 //
2107 // inline char Character.reverseBytes(char)
2108 // inline short Short.reverseBytes(short)
2109 // inline int Integer.reverseBytes(int)
2110 // inline long Long.reverseBytes(long)
2111 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2112 Node* arg = argument(0);
2113 Node* n;
2114 switch (id) {
2115 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2116 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2117 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2118 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2119 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2120 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2121 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2122 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2123 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2124 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2125 default: fatal_unexpected_iid(id); break;
2126 }
2127 set_result(_gvn.transform(n));
2128 return true;
2129 }
2130
2131 //----------------------------inline_unsafe_access----------------------------
2132
2133 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2134
2135 // Helper that guards and inserts a pre-barrier.
2136 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2137 Node* pre_val, bool need_mem_bar) {
2138 // We could be accessing the referent field of a reference object. If so, when G1
2139 // is enabled, we need to log the value in the referent field in an SATB buffer.
2140 // This routine performs some compile time filters and generates suitable
2141 // runtime filters that guard the pre-barrier code.
2142 // Also add memory barrier for non volatile load from the referent field
2143 // to prevent commoning of loads across safepoint.
2144 if (!UseG1GC && !need_mem_bar)
2145 return;
2146
2147 // Some compile time checks.
2148
2149 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2150 const TypeX* otype = offset->find_intptr_t_type();
2151 if (otype != NULL && otype->is_con() &&
2152 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2153 // Constant offset but not the reference_offset so just return
2154 return;
2155 }
2156
2157 // We only need to generate the runtime guards for instances.
2158 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2159 if (btype != NULL) {
2160 if (btype->isa_aryptr()) {
2161 // Array type so nothing to do
2162 return;
2163 }
2164
2165 const TypeInstPtr* itype = btype->isa_instptr();
2166 if (itype != NULL) {
2167 // Can the klass of base_oop be statically determined to be
2168 // _not_ a sub-class of Reference and _not_ Object?
2169 ciKlass* klass = itype->klass();
2170 if ( klass->is_loaded() &&
2171 !klass->is_subtype_of(env()->Reference_klass()) &&
2172 !env()->Object_klass()->is_subtype_of(klass)) {
2173 return;
2174 }
2175 }
2176 }
2177
2178 // The compile time filters did not reject base_oop/offset so
2179 // we need to generate the following runtime filters
2180 //
2181 // if (offset == java_lang_ref_Reference::_reference_offset) {
2182 // if (instance_of(base, java.lang.ref.Reference)) {
2183 // pre_barrier(_, pre_val, ...);
2184 // }
2185 // }
2186
2187 float likely = PROB_LIKELY( 0.999);
2188 float unlikely = PROB_UNLIKELY(0.999);
2189
2190 IdealKit ideal(this);
2191 #define __ ideal.
2192
2193 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2194
2195 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2196 // Update graphKit memory and control from IdealKit.
2197 sync_kit(ideal);
2198
2199 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2200 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2201
2202 // Update IdealKit memory and control from graphKit.
2203 __ sync_kit(this);
2204
2205 Node* one = __ ConI(1);
2206 // is_instof == 0 if base_oop == NULL
2207 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2208
2209 // Update graphKit from IdeakKit.
2210 sync_kit(ideal);
2211
2212 // Use the pre-barrier to record the value in the referent field
2213 pre_barrier(false /* do_load */,
2214 __ ctrl(),
2215 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2216 pre_val /* pre_val */,
2217 T_OBJECT);
2218 if (need_mem_bar) {
2219 // Add memory barrier to prevent commoning reads from this field
2220 // across safepoint since GC can change its value.
2221 insert_mem_bar(Op_MemBarCPUOrder);
2222 }
2223 // Update IdealKit from graphKit.
2224 __ sync_kit(this);
2225
2226 } __ end_if(); // _ref_type != ref_none
2227 } __ end_if(); // offset == referent_offset
2228
2229 // Final sync IdealKit and GraphKit.
2230 final_sync(ideal);
2231 #undef __
2232 }
2233
2234
2235 // Interpret Unsafe.fieldOffset cookies correctly:
2236 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2237
2238 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2239 // Attempt to infer a sharper value type from the offset and base type.
2240 ciKlass* sharpened_klass = NULL;
2241
2242 // See if it is an instance field, with an object type.
2243 if (alias_type->field() != NULL) {
2244 assert(!is_native_ptr, "native pointer op cannot use a java address");
2245 if (alias_type->field()->type()->is_klass()) {
2246 sharpened_klass = alias_type->field()->type()->as_klass();
2247 }
2248 }
2249
2250 // See if it is a narrow oop array.
2251 if (adr_type->isa_aryptr()) {
2252 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2253 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2254 if (elem_type != NULL) {
2255 sharpened_klass = elem_type->klass();
2256 }
2257 }
2258 }
2259
2260 // The sharpened class might be unloaded if there is no class loader
2261 // contraint in place.
2262 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2263 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2264
2265 #ifndef PRODUCT
2266 if (C->print_intrinsics() || C->print_inlining()) {
2267 tty->print(" from base type: "); adr_type->dump();
2268 tty->print(" sharpened value: "); tjp->dump();
2269 }
2270 #endif
2271 // Sharpen the value type.
2272 return tjp;
2273 }
2274 return NULL;
2275 }
2276
2277 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2278 if (callee()->is_static()) return false; // caller must have the capability!
2279
2280 #ifndef PRODUCT
2281 {
2282 ResourceMark rm;
2283 // Check the signatures.
2284 ciSignature* sig = callee()->signature();
2285 #ifdef ASSERT
2286 if (!is_store) {
2287 // Object getObject(Object base, int/long offset), etc.
2288 BasicType rtype = sig->return_type()->basic_type();
2289 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2290 rtype = T_ADDRESS; // it is really a C void*
2291 assert(rtype == type, "getter must return the expected value");
2292 if (!is_native_ptr) {
2293 assert(sig->count() == 2, "oop getter has 2 arguments");
2294 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2295 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2296 } else {
2297 assert(sig->count() == 1, "native getter has 1 argument");
2298 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2299 }
2300 } else {
2301 // void putObject(Object base, int/long offset, Object x), etc.
2302 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2303 if (!is_native_ptr) {
2304 assert(sig->count() == 3, "oop putter has 3 arguments");
2305 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2306 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2307 } else {
2308 assert(sig->count() == 2, "native putter has 2 arguments");
2309 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2310 }
2311 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2312 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2313 vtype = T_ADDRESS; // it is really a C void*
2314 assert(vtype == type, "putter must accept the expected value");
2315 }
2316 #endif // ASSERT
2317 }
2318 #endif //PRODUCT
2319
2320 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2321
2322 Node* receiver = argument(0); // type: oop
2323
2324 // Build address expression.
2325 Node* adr;
2326 Node* heap_base_oop = top();
2327 Node* offset = top();
2328 Node* val;
2329
2330 if (!is_native_ptr) {
2331 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2332 Node* base = argument(1); // type: oop
2333 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2334 offset = argument(2); // type: long
2335 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2336 // to be plain byte offsets, which are also the same as those accepted
2337 // by oopDesc::field_base.
2338 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2339 "fieldOffset must be byte-scaled");
2340 // 32-bit machines ignore the high half!
2341 offset = ConvL2X(offset);
2342 adr = make_unsafe_address(base, offset);
2343 heap_base_oop = base;
2344 val = is_store ? argument(4) : NULL;
2345 } else {
2346 Node* ptr = argument(1); // type: long
2347 ptr = ConvL2X(ptr); // adjust Java long to machine word
2348 adr = make_unsafe_address(NULL, ptr);
2349 val = is_store ? argument(3) : NULL;
2350 }
2351
2352 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2353
2354 // First guess at the value type.
2355 const Type *value_type = Type::get_const_basic_type(type);
2356
2357 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2358 // there was not enough information to nail it down.
2359 Compile::AliasType* alias_type = C->alias_type(adr_type);
2360 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2361
2362 // We will need memory barriers unless we can determine a unique
2363 // alias category for this reference. (Note: If for some reason
2364 // the barriers get omitted and the unsafe reference begins to "pollute"
2365 // the alias analysis of the rest of the graph, either Compile::can_alias
2366 // or Compile::must_alias will throw a diagnostic assert.)
2367 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2368
2369 // If we are reading the value of the referent field of a Reference
2370 // object (either by using Unsafe directly or through reflection)
2371 // then, if G1 is enabled, we need to record the referent in an
2372 // SATB log buffer using the pre-barrier mechanism.
2373 // Also we need to add memory barrier to prevent commoning reads
2374 // from this field across safepoint since GC can change its value.
2375 bool need_read_barrier = !is_native_ptr && !is_store &&
2376 offset != top() && heap_base_oop != top();
2377
2378 if (!is_store && type == T_OBJECT) {
2379 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2380 if (tjp != NULL) {
2381 value_type = tjp;
2382 }
2383 }
2384
2385 receiver = null_check(receiver);
2386 if (stopped()) {
2387 return true;
2388 }
2389 // Heap pointers get a null-check from the interpreter,
2390 // as a courtesy. However, this is not guaranteed by Unsafe,
2391 // and it is not possible to fully distinguish unintended nulls
2392 // from intended ones in this API.
2393
2394 if (is_volatile) {
2395 // We need to emit leading and trailing CPU membars (see below) in
2396 // addition to memory membars when is_volatile. This is a little
2397 // too strong, but avoids the need to insert per-alias-type
2398 // volatile membars (for stores; compare Parse::do_put_xxx), which
2399 // we cannot do effectively here because we probably only have a
2400 // rough approximation of type.
2401 need_mem_bar = true;
2402 // For Stores, place a memory ordering barrier now.
2403 if (is_store) {
2404 insert_mem_bar(Op_MemBarRelease);
2405 } else {
2406 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2407 insert_mem_bar(Op_MemBarVolatile);
2408 }
2409 }
2410 }
2411
2412 // Memory barrier to prevent normal and 'unsafe' accesses from
2413 // bypassing each other. Happens after null checks, so the
2414 // exception paths do not take memory state from the memory barrier,
2415 // so there's no problems making a strong assert about mixing users
2416 // of safe & unsafe memory.
2417 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2418
2419 if (!is_store) {
2420 Node* p = NULL;
2421 // Try to constant fold a load from a constant field
2422 ciField* field = alias_type->field();
2423 if (heap_base_oop != top() &&
2424 field != NULL && field->is_constant() && field->layout_type() == type) {
2425 // final or stable field
2426 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2427 if (con_type != NULL) {
2428 p = makecon(con_type);
2429 }
2430 }
2431 if (p == NULL) {
2432 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2433 // To be valid, unsafe loads may depend on other conditions than
2434 // the one that guards them: pin the Load node
2435 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2436 // load value
2437 switch (type) {
2438 case T_BOOLEAN:
2439 case T_CHAR:
2440 case T_BYTE:
2441 case T_SHORT:
2442 case T_INT:
2443 case T_LONG:
2444 case T_FLOAT:
2445 case T_DOUBLE:
2446 break;
2447 case T_OBJECT:
2448 if (need_read_barrier) {
2449 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2450 }
2451 break;
2452 case T_ADDRESS:
2453 // Cast to an int type.
2454 p = _gvn.transform(new CastP2XNode(NULL, p));
2455 p = ConvX2UL(p);
2456 break;
2457 default:
2458 fatal("unexpected type %d: %s", type, type2name(type));
2459 break;
2460 }
2461 }
2462 // The load node has the control of the preceding MemBarCPUOrder. All
2463 // following nodes will have the control of the MemBarCPUOrder inserted at
2464 // the end of this method. So, pushing the load onto the stack at a later
2465 // point is fine.
2466 set_result(p);
2467 } else {
2468 // place effect of store into memory
2469 switch (type) {
2470 case T_DOUBLE:
2471 val = dstore_rounding(val);
2472 break;
2473 case T_ADDRESS:
2474 // Repackage the long as a pointer.
2475 val = ConvL2X(val);
2476 val = _gvn.transform(new CastX2PNode(val));
2477 break;
2478 }
2479
2480 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2481 if (type != T_OBJECT ) {
2482 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2483 } else {
2484 // Possibly an oop being stored to Java heap or native memory
2485 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2486 // oop to Java heap.
2487 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2488 } else {
2489 // We can't tell at compile time if we are storing in the Java heap or outside
2490 // of it. So we need to emit code to conditionally do the proper type of
2491 // store.
2492
2493 IdealKit ideal(this);
2494 #define __ ideal.
2495 // QQQ who knows what probability is here??
2496 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2497 // Sync IdealKit and graphKit.
2498 sync_kit(ideal);
2499 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2500 // Update IdealKit memory.
2501 __ sync_kit(this);
2502 } __ else_(); {
2503 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2504 } __ end_if();
2505 // Final sync IdealKit and GraphKit.
2506 final_sync(ideal);
2507 #undef __
2508 }
2509 }
2510 }
2511
2512 if (is_volatile) {
2513 if (!is_store) {
2514 insert_mem_bar(Op_MemBarAcquire);
2515 } else {
2516 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2517 insert_mem_bar(Op_MemBarVolatile);
2518 }
2519 }
2520 }
2521
2522 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2523
2524 return true;
2525 }
2526
2527 //----------------------------inline_unsafe_load_store----------------------------
2528 // This method serves a couple of different customers (depending on LoadStoreKind):
2529 //
2530 // LS_cmpxchg:
2531 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2532 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2533 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2534 //
2535 // LS_xadd:
2536 // public int getAndAddInt( Object o, long offset, int delta)
2537 // public long getAndAddLong(Object o, long offset, long delta)
2538 //
2539 // LS_xchg:
2540 // int getAndSet(Object o, long offset, int newValue)
2541 // long getAndSet(Object o, long offset, long newValue)
2542 // Object getAndSet(Object o, long offset, Object newValue)
2543 //
2544 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2545 // This basic scheme here is the same as inline_unsafe_access, but
2546 // differs in enough details that combining them would make the code
2547 // overly confusing. (This is a true fact! I originally combined
2548 // them, but even I was confused by it!) As much code/comments as
2549 // possible are retained from inline_unsafe_access though to make
2550 // the correspondences clearer. - dl
2551
2552 if (callee()->is_static()) return false; // caller must have the capability!
2553
2554 #ifndef PRODUCT
2555 BasicType rtype;
2556 {
2557 ResourceMark rm;
2558 // Check the signatures.
2559 ciSignature* sig = callee()->signature();
2560 rtype = sig->return_type()->basic_type();
2561 if (kind == LS_xadd || kind == LS_xchg) {
2562 // Check the signatures.
2563 #ifdef ASSERT
2564 assert(rtype == type, "get and set must return the expected type");
2565 assert(sig->count() == 3, "get and set has 3 arguments");
2566 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2567 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2568 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2569 #endif // ASSERT
2570 } else if (kind == LS_cmpxchg) {
2571 // Check the signatures.
2572 #ifdef ASSERT
2573 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2574 assert(sig->count() == 4, "CAS has 4 arguments");
2575 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2576 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2577 #endif // ASSERT
2578 } else {
2579 ShouldNotReachHere();
2580 }
2581 }
2582 #endif //PRODUCT
2583
2584 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2585
2586 // Get arguments:
2587 Node* receiver = NULL;
2588 Node* base = NULL;
2589 Node* offset = NULL;
2590 Node* oldval = NULL;
2591 Node* newval = NULL;
2592 if (kind == LS_cmpxchg) {
2593 const bool two_slot_type = type2size[type] == 2;
2594 receiver = argument(0); // type: oop
2595 base = argument(1); // type: oop
2596 offset = argument(2); // type: long
2597 oldval = argument(4); // type: oop, int, or long
2598 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2599 } else if (kind == LS_xadd || kind == LS_xchg){
2600 receiver = argument(0); // type: oop
2601 base = argument(1); // type: oop
2602 offset = argument(2); // type: long
2603 oldval = NULL;
2604 newval = argument(4); // type: oop, int, or long
2605 }
2606
2607 // Null check receiver.
2608 receiver = null_check(receiver);
2609 if (stopped()) {
2610 return true;
2611 }
2612
2613 // Build field offset expression.
2614 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2615 // to be plain byte offsets, which are also the same as those accepted
2616 // by oopDesc::field_base.
2617 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2618 // 32-bit machines ignore the high half of long offsets
2619 offset = ConvL2X(offset);
2620 Node* adr = make_unsafe_address(base, offset);
2621 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2622
2623 // For CAS, unlike inline_unsafe_access, there seems no point in
2624 // trying to refine types. Just use the coarse types here.
2625 const Type *value_type = Type::get_const_basic_type(type);
2626 Compile::AliasType* alias_type = C->alias_type(adr_type);
2627 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2628
2629 if (kind == LS_xchg && type == T_OBJECT) {
2630 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2631 if (tjp != NULL) {
2632 value_type = tjp;
2633 }
2634 }
2635
2636 int alias_idx = C->get_alias_index(adr_type);
2637
2638 // Memory-model-wise, a LoadStore acts like a little synchronized
2639 // block, so needs barriers on each side. These don't translate
2640 // into actual barriers on most machines, but we still need rest of
2641 // compiler to respect ordering.
2642
2643 insert_mem_bar(Op_MemBarRelease);
2644 insert_mem_bar(Op_MemBarCPUOrder);
2645
2646 // 4984716: MemBars must be inserted before this
2647 // memory node in order to avoid a false
2648 // dependency which will confuse the scheduler.
2649 Node *mem = memory(alias_idx);
2650
2651 // For now, we handle only those cases that actually exist: ints,
2652 // longs, and Object. Adding others should be straightforward.
2653 Node* load_store;
2654 switch(type) {
2655 case T_INT:
2656 if (kind == LS_xadd) {
2657 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2658 } else if (kind == LS_xchg) {
2659 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2660 } else if (kind == LS_cmpxchg) {
2661 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2662 } else {
2663 ShouldNotReachHere();
2664 }
2665 break;
2666 case T_LONG:
2667 if (kind == LS_xadd) {
2668 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2669 } else if (kind == LS_xchg) {
2670 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2671 } else if (kind == LS_cmpxchg) {
2672 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2673 } else {
2674 ShouldNotReachHere();
2675 }
2676 break;
2677 case T_OBJECT:
2678 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2679 // could be delayed during Parse (for example, in adjust_map_after_if()).
2680 // Execute transformation here to avoid barrier generation in such case.
2681 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2682 newval = _gvn.makecon(TypePtr::NULL_PTR);
2683
2684 // Reference stores need a store barrier.
2685 if (kind == LS_xchg) {
2686 // If pre-barrier must execute before the oop store, old value will require do_load here.
2687 if (!can_move_pre_barrier()) {
2688 pre_barrier(true /* do_load*/,
2689 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2690 NULL /* pre_val*/,
2691 T_OBJECT);
2692 } // Else move pre_barrier to use load_store value, see below.
2693 } else if (kind == LS_cmpxchg) {
2694 // Same as for newval above:
2695 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2696 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2697 }
2698 // The only known value which might get overwritten is oldval.
2699 pre_barrier(false /* do_load */,
2700 control(), NULL, NULL, max_juint, NULL, NULL,
2701 oldval /* pre_val */,
2702 T_OBJECT);
2703 } else {
2704 ShouldNotReachHere();
2705 }
2706
2707 #ifdef _LP64
2708 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2709 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2710 if (kind == LS_xchg) {
2711 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2712 newval_enc, adr_type, value_type->make_narrowoop()));
2713 } else {
2714 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2715 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2716 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2717 newval_enc, oldval_enc));
2718 }
2719 } else
2720 #endif
2721 {
2722 if (kind == LS_xchg) {
2723 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2724 } else {
2725 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2726 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2727 }
2728 }
2729 if (kind == LS_cmpxchg) {
2730 // Emit the post barrier only when the actual store happened.
2731 // This makes sense to check only for compareAndSet that can fail to set the value.
2732 // CAS success path is marked more likely since we anticipate this is a performance
2733 // critical path, while CAS failure path can use the penalty for going through unlikely
2734 // path as backoff. Which is still better than doing a store barrier there.
2735 IdealKit ideal(this);
2736 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2737 sync_kit(ideal);
2738 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2739 ideal.sync_kit(this);
2740 } ideal.end_if();
2741 final_sync(ideal);
2742 } else {
2743 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2744 }
2745 break;
2746 default:
2747 fatal("unexpected type %d: %s", type, type2name(type));
2748 break;
2749 }
2750
2751 // SCMemProjNodes represent the memory state of a LoadStore. Their
2752 // main role is to prevent LoadStore nodes from being optimized away
2753 // when their results aren't used.
2754 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2755 set_memory(proj, alias_idx);
2756
2757 if (type == T_OBJECT && kind == LS_xchg) {
2758 #ifdef _LP64
2759 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2760 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2761 }
2762 #endif
2763 if (can_move_pre_barrier()) {
2764 // Don't need to load pre_val. The old value is returned by load_store.
2765 // The pre_barrier can execute after the xchg as long as no safepoint
2766 // gets inserted between them.
2767 pre_barrier(false /* do_load */,
2768 control(), NULL, NULL, max_juint, NULL, NULL,
2769 load_store /* pre_val */,
2770 T_OBJECT);
2771 }
2772 }
2773
2774 // Add the trailing membar surrounding the access
2775 insert_mem_bar(Op_MemBarCPUOrder);
2776 insert_mem_bar(Op_MemBarAcquire);
2777
2778 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2779 set_result(load_store);
2780 return true;
2781 }
2782
2783 //----------------------------inline_unsafe_ordered_store----------------------
2784 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2785 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2786 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2787 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2788 // This is another variant of inline_unsafe_access, differing in
2789 // that it always issues store-store ("release") barrier and ensures
2790 // store-atomicity (which only matters for "long").
2791
2792 if (callee()->is_static()) return false; // caller must have the capability!
2793
2794 #ifndef PRODUCT
2795 {
2796 ResourceMark rm;
2797 // Check the signatures.
2798 ciSignature* sig = callee()->signature();
2799 #ifdef ASSERT
2800 BasicType rtype = sig->return_type()->basic_type();
2801 assert(rtype == T_VOID, "must return void");
2802 assert(sig->count() == 3, "has 3 arguments");
2803 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2804 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2805 #endif // ASSERT
2806 }
2807 #endif //PRODUCT
2808
2809 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2810
2811 // Get arguments:
2812 Node* receiver = argument(0); // type: oop
2813 Node* base = argument(1); // type: oop
2814 Node* offset = argument(2); // type: long
2815 Node* val = argument(4); // type: oop, int, or long
2816
2817 // Null check receiver.
2818 receiver = null_check(receiver);
2819 if (stopped()) {
2820 return true;
2821 }
2822
2823 // Build field offset expression.
2824 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2825 // 32-bit machines ignore the high half of long offsets
2826 offset = ConvL2X(offset);
2827 Node* adr = make_unsafe_address(base, offset);
2828 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2829 const Type *value_type = Type::get_const_basic_type(type);
2830 Compile::AliasType* alias_type = C->alias_type(adr_type);
2831
2832 insert_mem_bar(Op_MemBarRelease);
2833 insert_mem_bar(Op_MemBarCPUOrder);
2834 // Ensure that the store is atomic for longs:
2835 const bool require_atomic_access = true;
2836 Node* store;
2837 if (type == T_OBJECT) // reference stores need a store barrier.
2838 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2839 else {
2840 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2841 }
2842 insert_mem_bar(Op_MemBarCPUOrder);
2843 return true;
2844 }
2845
2846 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2847 // Regardless of form, don't allow previous ld/st to move down,
2848 // then issue acquire, release, or volatile mem_bar.
2849 insert_mem_bar(Op_MemBarCPUOrder);
2850 switch(id) {
2851 case vmIntrinsics::_loadFence:
2852 insert_mem_bar(Op_LoadFence);
2853 return true;
2854 case vmIntrinsics::_storeFence:
2855 insert_mem_bar(Op_StoreFence);
2856 return true;
2857 case vmIntrinsics::_fullFence:
2858 insert_mem_bar(Op_MemBarVolatile);
2859 return true;
2860 default:
2861 fatal_unexpected_iid(id);
2862 return false;
2863 }
2864 }
2865
2866 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2867 if (!kls->is_Con()) {
2868 return true;
2869 }
2870 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2871 if (klsptr == NULL) {
2872 return true;
2873 }
2874 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2875 // don't need a guard for a klass that is already initialized
2876 return !ik->is_initialized();
2877 }
2878
2879 //----------------------------inline_unsafe_allocate---------------------------
2880 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2881 bool LibraryCallKit::inline_unsafe_allocate() {
2882 if (callee()->is_static()) return false; // caller must have the capability!
2883
2884 null_check_receiver(); // null-check, then ignore
2885 Node* cls = null_check(argument(1));
2886 if (stopped()) return true;
2887
2888 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2889 kls = null_check(kls);
2890 if (stopped()) return true; // argument was like int.class
2891
2892 Node* test = NULL;
2893 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2894 // Note: The argument might still be an illegal value like
2895 // Serializable.class or Object[].class. The runtime will handle it.
2896 // But we must make an explicit check for initialization.
2897 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2898 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2899 // can generate code to load it as unsigned byte.
2900 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2901 Node* bits = intcon(InstanceKlass::fully_initialized);
2902 test = _gvn.transform(new SubINode(inst, bits));
2903 // The 'test' is non-zero if we need to take a slow path.
2904 }
2905
2906 Node* obj = new_instance(kls, test);
2907 set_result(obj);
2908 return true;
2909 }
2910
2911 #ifdef TRACE_HAVE_INTRINSICS
2912 /*
2913 * oop -> myklass
2914 * myklass->trace_id |= USED
2915 * return myklass->trace_id & ~0x3
2916 */
2917 bool LibraryCallKit::inline_native_classID() {
2918 null_check_receiver(); // null-check, then ignore
2919 Node* cls = null_check(argument(1), T_OBJECT);
2920 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2921 kls = null_check(kls, T_OBJECT);
2922 ByteSize offset = TRACE_ID_OFFSET;
2923 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2924 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2925 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2926 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2927 Node* clsused = longcon(0x01l); // set the class bit
2928 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2929
2930 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2931 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2932 set_result(andl);
2933 return true;
2934 }
2935
2936 bool LibraryCallKit::inline_native_threadID() {
2937 Node* tls_ptr = NULL;
2938 Node* cur_thr = generate_current_thread(tls_ptr);
2939 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2940 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2941 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2942
2943 Node* threadid = NULL;
2944 size_t thread_id_size = OSThread::thread_id_size();
2945 if (thread_id_size == (size_t) BytesPerLong) {
2946 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2947 } else if (thread_id_size == (size_t) BytesPerInt) {
2948 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2949 } else {
2950 ShouldNotReachHere();
2951 }
2952 set_result(threadid);
2953 return true;
2954 }
2955 #endif
2956
2957 //------------------------inline_native_time_funcs--------------
2958 // inline code for System.currentTimeMillis() and System.nanoTime()
2959 // these have the same type and signature
2960 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2961 const TypeFunc* tf = OptoRuntime::void_long_Type();
2962 const TypePtr* no_memory_effects = NULL;
2963 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2964 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2965 #ifdef ASSERT
2966 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2967 assert(value_top == top(), "second value must be top");
2968 #endif
2969 set_result(value);
2970 return true;
2971 }
2972
2973 //------------------------inline_native_currentThread------------------
2974 bool LibraryCallKit::inline_native_currentThread() {
2975 Node* junk = NULL;
2976 set_result(generate_current_thread(junk));
2977 return true;
2978 }
2979
2980 //------------------------inline_native_isInterrupted------------------
2981 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2982 bool LibraryCallKit::inline_native_isInterrupted() {
2983 // Add a fast path to t.isInterrupted(clear_int):
2984 // (t == Thread.current() &&
2985 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2986 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2987 // So, in the common case that the interrupt bit is false,
2988 // we avoid making a call into the VM. Even if the interrupt bit
2989 // is true, if the clear_int argument is false, we avoid the VM call.
2990 // However, if the receiver is not currentThread, we must call the VM,
2991 // because there must be some locking done around the operation.
2992
2993 // We only go to the fast case code if we pass two guards.
2994 // Paths which do not pass are accumulated in the slow_region.
2995
2996 enum {
2997 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2998 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2999 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3000 PATH_LIMIT
3001 };
3002
3003 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3004 // out of the function.
3005 insert_mem_bar(Op_MemBarCPUOrder);
3006
3007 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3008 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3009
3010 RegionNode* slow_region = new RegionNode(1);
3011 record_for_igvn(slow_region);
3012
3013 // (a) Receiving thread must be the current thread.
3014 Node* rec_thr = argument(0);
3015 Node* tls_ptr = NULL;
3016 Node* cur_thr = generate_current_thread(tls_ptr);
3017 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3018 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3019
3020 generate_slow_guard(bol_thr, slow_region);
3021
3022 // (b) Interrupt bit on TLS must be false.
3023 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3024 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3025 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3026
3027 // Set the control input on the field _interrupted read to prevent it floating up.
3028 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3029 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3030 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3031
3032 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3033
3034 // First fast path: if (!TLS._interrupted) return false;
3035 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3036 result_rgn->init_req(no_int_result_path, false_bit);
3037 result_val->init_req(no_int_result_path, intcon(0));
3038
3039 // drop through to next case
3040 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3041
3042 #ifndef TARGET_OS_FAMILY_windows
3043 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3044 Node* clr_arg = argument(1);
3045 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3046 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3047 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3048
3049 // Second fast path: ... else if (!clear_int) return true;
3050 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3051 result_rgn->init_req(no_clear_result_path, false_arg);
3052 result_val->init_req(no_clear_result_path, intcon(1));
3053
3054 // drop through to next case
3055 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3056 #else
3057 // To return true on Windows you must read the _interrupted field
3058 // and check the the event state i.e. take the slow path.
3059 #endif // TARGET_OS_FAMILY_windows
3060
3061 // (d) Otherwise, go to the slow path.
3062 slow_region->add_req(control());
3063 set_control( _gvn.transform(slow_region));
3064
3065 if (stopped()) {
3066 // There is no slow path.
3067 result_rgn->init_req(slow_result_path, top());
3068 result_val->init_req(slow_result_path, top());
3069 } else {
3070 // non-virtual because it is a private non-static
3071 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3072
3073 Node* slow_val = set_results_for_java_call(slow_call);
3074 // this->control() comes from set_results_for_java_call
3075
3076 Node* fast_io = slow_call->in(TypeFunc::I_O);
3077 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3078
3079 // These two phis are pre-filled with copies of of the fast IO and Memory
3080 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3081 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3082
3083 result_rgn->init_req(slow_result_path, control());
3084 result_io ->init_req(slow_result_path, i_o());
3085 result_mem->init_req(slow_result_path, reset_memory());
3086 result_val->init_req(slow_result_path, slow_val);
3087
3088 set_all_memory(_gvn.transform(result_mem));
3089 set_i_o( _gvn.transform(result_io));
3090 }
3091
3092 C->set_has_split_ifs(true); // Has chance for split-if optimization
3093 set_result(result_rgn, result_val);
3094 return true;
3095 }
3096
3097 //---------------------------load_mirror_from_klass----------------------------
3098 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3099 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3100 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3101 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3102 }
3103
3104 //-----------------------load_klass_from_mirror_common-------------------------
3105 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3106 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3107 // and branch to the given path on the region.
3108 // If never_see_null, take an uncommon trap on null, so we can optimistically
3109 // compile for the non-null case.
3110 // If the region is NULL, force never_see_null = true.
3111 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3112 bool never_see_null,
3113 RegionNode* region,
3114 int null_path,
3115 int offset) {
3116 if (region == NULL) never_see_null = true;
3117 Node* p = basic_plus_adr(mirror, offset);
3118 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3119 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3120 Node* null_ctl = top();
3121 kls = null_check_oop(kls, &null_ctl, never_see_null);
3122 if (region != NULL) {
3123 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3124 region->init_req(null_path, null_ctl);
3125 } else {
3126 assert(null_ctl == top(), "no loose ends");
3127 }
3128 return kls;
3129 }
3130
3131 //--------------------(inline_native_Class_query helpers)---------------------
3132 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3133 // Fall through if (mods & mask) == bits, take the guard otherwise.
3134 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3135 // Branch around if the given klass has the given modifier bit set.
3136 // Like generate_guard, adds a new path onto the region.
3137 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3138 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3139 Node* mask = intcon(modifier_mask);
3140 Node* bits = intcon(modifier_bits);
3141 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3142 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3143 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3144 return generate_fair_guard(bol, region);
3145 }
3146 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3147 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3148 }
3149
3150 //-------------------------inline_native_Class_query-------------------
3151 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3152 const Type* return_type = TypeInt::BOOL;
3153 Node* prim_return_value = top(); // what happens if it's a primitive class?
3154 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3155 bool expect_prim = false; // most of these guys expect to work on refs
3156
3157 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3158
3159 Node* mirror = argument(0);
3160 Node* obj = top();
3161
3162 switch (id) {
3163 case vmIntrinsics::_isInstance:
3164 // nothing is an instance of a primitive type
3165 prim_return_value = intcon(0);
3166 obj = argument(1);
3167 break;
3168 case vmIntrinsics::_getModifiers:
3169 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3170 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3171 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3172 break;
3173 case vmIntrinsics::_isInterface:
3174 prim_return_value = intcon(0);
3175 break;
3176 case vmIntrinsics::_isArray:
3177 prim_return_value = intcon(0);
3178 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3179 break;
3180 case vmIntrinsics::_isPrimitive:
3181 prim_return_value = intcon(1);
3182 expect_prim = true; // obviously
3183 break;
3184 case vmIntrinsics::_getSuperclass:
3185 prim_return_value = null();
3186 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3187 break;
3188 case vmIntrinsics::_getClassAccessFlags:
3189 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3190 return_type = TypeInt::INT; // not bool! 6297094
3191 break;
3192 default:
3193 fatal_unexpected_iid(id);
3194 break;
3195 }
3196
3197 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3198 if (mirror_con == NULL) return false; // cannot happen?
3199
3200 #ifndef PRODUCT
3201 if (C->print_intrinsics() || C->print_inlining()) {
3202 ciType* k = mirror_con->java_mirror_type();
3203 if (k) {
3204 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3205 k->print_name();
3206 tty->cr();
3207 }
3208 }
3209 #endif
3210
3211 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3212 RegionNode* region = new RegionNode(PATH_LIMIT);
3213 record_for_igvn(region);
3214 PhiNode* phi = new PhiNode(region, return_type);
3215
3216 // The mirror will never be null of Reflection.getClassAccessFlags, however
3217 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3218 // if it is. See bug 4774291.
3219
3220 // For Reflection.getClassAccessFlags(), the null check occurs in
3221 // the wrong place; see inline_unsafe_access(), above, for a similar
3222 // situation.
3223 mirror = null_check(mirror);
3224 // If mirror or obj is dead, only null-path is taken.
3225 if (stopped()) return true;
3226
3227 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3228
3229 // Now load the mirror's klass metaobject, and null-check it.
3230 // Side-effects region with the control path if the klass is null.
3231 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3232 // If kls is null, we have a primitive mirror.
3233 phi->init_req(_prim_path, prim_return_value);
3234 if (stopped()) { set_result(region, phi); return true; }
3235 bool safe_for_replace = (region->in(_prim_path) == top());
3236
3237 Node* p; // handy temp
3238 Node* null_ctl;
3239
3240 // Now that we have the non-null klass, we can perform the real query.
3241 // For constant classes, the query will constant-fold in LoadNode::Value.
3242 Node* query_value = top();
3243 switch (id) {
3244 case vmIntrinsics::_isInstance:
3245 // nothing is an instance of a primitive type
3246 query_value = gen_instanceof(obj, kls, safe_for_replace);
3247 break;
3248
3249 case vmIntrinsics::_getModifiers:
3250 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3251 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3252 break;
3253
3254 case vmIntrinsics::_isInterface:
3255 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3256 if (generate_interface_guard(kls, region) != NULL)
3257 // A guard was added. If the guard is taken, it was an interface.
3258 phi->add_req(intcon(1));
3259 // If we fall through, it's a plain class.
3260 query_value = intcon(0);
3261 break;
3262
3263 case vmIntrinsics::_isArray:
3264 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3265 if (generate_array_guard(kls, region) != NULL)
3266 // A guard was added. If the guard is taken, it was an array.
3267 phi->add_req(intcon(1));
3268 // If we fall through, it's a plain class.
3269 query_value = intcon(0);
3270 break;
3271
3272 case vmIntrinsics::_isPrimitive:
3273 query_value = intcon(0); // "normal" path produces false
3274 break;
3275
3276 case vmIntrinsics::_getSuperclass:
3277 // The rules here are somewhat unfortunate, but we can still do better
3278 // with random logic than with a JNI call.
3279 // Interfaces store null or Object as _super, but must report null.
3280 // Arrays store an intermediate super as _super, but must report Object.
3281 // Other types can report the actual _super.
3282 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3283 if (generate_interface_guard(kls, region) != NULL)
3284 // A guard was added. If the guard is taken, it was an interface.
3285 phi->add_req(null());
3286 if (generate_array_guard(kls, region) != NULL)
3287 // A guard was added. If the guard is taken, it was an array.
3288 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3289 // If we fall through, it's a plain class. Get its _super.
3290 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3291 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3292 null_ctl = top();
3293 kls = null_check_oop(kls, &null_ctl);
3294 if (null_ctl != top()) {
3295 // If the guard is taken, Object.superClass is null (both klass and mirror).
3296 region->add_req(null_ctl);
3297 phi ->add_req(null());
3298 }
3299 if (!stopped()) {
3300 query_value = load_mirror_from_klass(kls);
3301 }
3302 break;
3303
3304 case vmIntrinsics::_getClassAccessFlags:
3305 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3306 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3307 break;
3308
3309 default:
3310 fatal_unexpected_iid(id);
3311 break;
3312 }
3313
3314 // Fall-through is the normal case of a query to a real class.
3315 phi->init_req(1, query_value);
3316 region->init_req(1, control());
3317
3318 C->set_has_split_ifs(true); // Has chance for split-if optimization
3319 set_result(region, phi);
3320 return true;
3321 }
3322
3323 //-------------------------inline_Class_cast-------------------
3324 bool LibraryCallKit::inline_Class_cast() {
3325 Node* mirror = argument(0); // Class
3326 Node* obj = argument(1);
3327 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3328 if (mirror_con == NULL) {
3329 return false; // dead path (mirror->is_top()).
3330 }
3331 if (obj == NULL || obj->is_top()) {
3332 return false; // dead path
3333 }
3334 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3335
3336 // First, see if Class.cast() can be folded statically.
3337 // java_mirror_type() returns non-null for compile-time Class constants.
3338 ciType* tm = mirror_con->java_mirror_type();
3339 if (tm != NULL && tm->is_klass() &&
3340 tp != NULL && tp->klass() != NULL) {
3341 if (!tp->klass()->is_loaded()) {
3342 // Don't use intrinsic when class is not loaded.
3343 return false;
3344 } else {
3345 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3346 if (static_res == Compile::SSC_always_true) {
3347 // isInstance() is true - fold the code.
3348 set_result(obj);
3349 return true;
3350 } else if (static_res == Compile::SSC_always_false) {
3351 // Don't use intrinsic, have to throw ClassCastException.
3352 // If the reference is null, the non-intrinsic bytecode will
3353 // be optimized appropriately.
3354 return false;
3355 }
3356 }
3357 }
3358
3359 // Bailout intrinsic and do normal inlining if exception path is frequent.
3360 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3361 return false;
3362 }
3363
3364 // Generate dynamic checks.
3365 // Class.cast() is java implementation of _checkcast bytecode.
3366 // Do checkcast (Parse::do_checkcast()) optimizations here.
3367
3368 mirror = null_check(mirror);
3369 // If mirror is dead, only null-path is taken.
3370 if (stopped()) {
3371 return true;
3372 }
3373
3374 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3375 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3376 RegionNode* region = new RegionNode(PATH_LIMIT);
3377 record_for_igvn(region);
3378
3379 // Now load the mirror's klass metaobject, and null-check it.
3380 // If kls is null, we have a primitive mirror and
3381 // nothing is an instance of a primitive type.
3382 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3383
3384 Node* res = top();
3385 if (!stopped()) {
3386 Node* bad_type_ctrl = top();
3387 // Do checkcast optimizations.
3388 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3389 region->init_req(_bad_type_path, bad_type_ctrl);
3390 }
3391 if (region->in(_prim_path) != top() ||
3392 region->in(_bad_type_path) != top()) {
3393 // Let Interpreter throw ClassCastException.
3394 PreserveJVMState pjvms(this);
3395 set_control(_gvn.transform(region));
3396 uncommon_trap(Deoptimization::Reason_intrinsic,
3397 Deoptimization::Action_maybe_recompile);
3398 }
3399 if (!stopped()) {
3400 set_result(res);
3401 }
3402 return true;
3403 }
3404
3405
3406 //--------------------------inline_native_subtype_check------------------------
3407 // This intrinsic takes the JNI calls out of the heart of
3408 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3409 bool LibraryCallKit::inline_native_subtype_check() {
3410 // Pull both arguments off the stack.
3411 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3412 args[0] = argument(0);
3413 args[1] = argument(1);
3414 Node* klasses[2]; // corresponding Klasses: superk, subk
3415 klasses[0] = klasses[1] = top();
3416
3417 enum {
3418 // A full decision tree on {superc is prim, subc is prim}:
3419 _prim_0_path = 1, // {P,N} => false
3420 // {P,P} & superc!=subc => false
3421 _prim_same_path, // {P,P} & superc==subc => true
3422 _prim_1_path, // {N,P} => false
3423 _ref_subtype_path, // {N,N} & subtype check wins => true
3424 _both_ref_path, // {N,N} & subtype check loses => false
3425 PATH_LIMIT
3426 };
3427
3428 RegionNode* region = new RegionNode(PATH_LIMIT);
3429 Node* phi = new PhiNode(region, TypeInt::BOOL);
3430 record_for_igvn(region);
3431
3432 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3433 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3434 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3435
3436 // First null-check both mirrors and load each mirror's klass metaobject.
3437 int which_arg;
3438 for (which_arg = 0; which_arg <= 1; which_arg++) {
3439 Node* arg = args[which_arg];
3440 arg = null_check(arg);
3441 if (stopped()) break;
3442 args[which_arg] = arg;
3443
3444 Node* p = basic_plus_adr(arg, class_klass_offset);
3445 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3446 klasses[which_arg] = _gvn.transform(kls);
3447 }
3448
3449 // Having loaded both klasses, test each for null.
3450 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3451 for (which_arg = 0; which_arg <= 1; which_arg++) {
3452 Node* kls = klasses[which_arg];
3453 Node* null_ctl = top();
3454 kls = null_check_oop(kls, &null_ctl, never_see_null);
3455 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3456 region->init_req(prim_path, null_ctl);
3457 if (stopped()) break;
3458 klasses[which_arg] = kls;
3459 }
3460
3461 if (!stopped()) {
3462 // now we have two reference types, in klasses[0..1]
3463 Node* subk = klasses[1]; // the argument to isAssignableFrom
3464 Node* superk = klasses[0]; // the receiver
3465 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3466 // now we have a successful reference subtype check
3467 region->set_req(_ref_subtype_path, control());
3468 }
3469
3470 // If both operands are primitive (both klasses null), then
3471 // we must return true when they are identical primitives.
3472 // It is convenient to test this after the first null klass check.
3473 set_control(region->in(_prim_0_path)); // go back to first null check
3474 if (!stopped()) {
3475 // Since superc is primitive, make a guard for the superc==subc case.
3476 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3477 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3478 generate_guard(bol_eq, region, PROB_FAIR);
3479 if (region->req() == PATH_LIMIT+1) {
3480 // A guard was added. If the added guard is taken, superc==subc.
3481 region->swap_edges(PATH_LIMIT, _prim_same_path);
3482 region->del_req(PATH_LIMIT);
3483 }
3484 region->set_req(_prim_0_path, control()); // Not equal after all.
3485 }
3486
3487 // these are the only paths that produce 'true':
3488 phi->set_req(_prim_same_path, intcon(1));
3489 phi->set_req(_ref_subtype_path, intcon(1));
3490
3491 // pull together the cases:
3492 assert(region->req() == PATH_LIMIT, "sane region");
3493 for (uint i = 1; i < region->req(); i++) {
3494 Node* ctl = region->in(i);
3495 if (ctl == NULL || ctl == top()) {
3496 region->set_req(i, top());
3497 phi ->set_req(i, top());
3498 } else if (phi->in(i) == NULL) {
3499 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3500 }
3501 }
3502
3503 set_control(_gvn.transform(region));
3504 set_result(_gvn.transform(phi));
3505 return true;
3506 }
3507
3508 //---------------------generate_array_guard_common------------------------
3509 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3510 bool obj_array, bool not_array) {
3511
3512 if (stopped()) {
3513 return NULL;
3514 }
3515
3516 // If obj_array/non_array==false/false:
3517 // Branch around if the given klass is in fact an array (either obj or prim).
3518 // If obj_array/non_array==false/true:
3519 // Branch around if the given klass is not an array klass of any kind.
3520 // If obj_array/non_array==true/true:
3521 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3522 // If obj_array/non_array==true/false:
3523 // Branch around if the kls is an oop array (Object[] or subtype)
3524 //
3525 // Like generate_guard, adds a new path onto the region.
3526 jint layout_con = 0;
3527 Node* layout_val = get_layout_helper(kls, layout_con);
3528 if (layout_val == NULL) {
3529 bool query = (obj_array
3530 ? Klass::layout_helper_is_objArray(layout_con)
3531 : Klass::layout_helper_is_array(layout_con));
3532 if (query == not_array) {
3533 return NULL; // never a branch
3534 } else { // always a branch
3535 Node* always_branch = control();
3536 if (region != NULL)
3537 region->add_req(always_branch);
3538 set_control(top());
3539 return always_branch;
3540 }
3541 }
3542 // Now test the correct condition.
3543 jint nval = (obj_array
3544 ? ((jint)Klass::_lh_array_tag_type_value
3545 << Klass::_lh_array_tag_shift)
3546 : Klass::_lh_neutral_value);
3547 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3548 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3549 // invert the test if we are looking for a non-array
3550 if (not_array) btest = BoolTest(btest).negate();
3551 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3552 return generate_fair_guard(bol, region);
3553 }
3554
3555
3556 //-----------------------inline_native_newArray--------------------------
3557 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3558 bool LibraryCallKit::inline_native_newArray() {
3559 Node* mirror = argument(0);
3560 Node* count_val = argument(1);
3561
3562 mirror = null_check(mirror);
3563 // If mirror or obj is dead, only null-path is taken.
3564 if (stopped()) return true;
3565
3566 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3567 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3568 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3569 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3570 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3571
3572 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3573 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3574 result_reg, _slow_path);
3575 Node* normal_ctl = control();
3576 Node* no_array_ctl = result_reg->in(_slow_path);
3577
3578 // Generate code for the slow case. We make a call to newArray().
3579 set_control(no_array_ctl);
3580 if (!stopped()) {
3581 // Either the input type is void.class, or else the
3582 // array klass has not yet been cached. Either the
3583 // ensuing call will throw an exception, or else it
3584 // will cache the array klass for next time.
3585 PreserveJVMState pjvms(this);
3586 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3587 Node* slow_result = set_results_for_java_call(slow_call);
3588 // this->control() comes from set_results_for_java_call
3589 result_reg->set_req(_slow_path, control());
3590 result_val->set_req(_slow_path, slow_result);
3591 result_io ->set_req(_slow_path, i_o());
3592 result_mem->set_req(_slow_path, reset_memory());
3593 }
3594
3595 set_control(normal_ctl);
3596 if (!stopped()) {
3597 // Normal case: The array type has been cached in the java.lang.Class.
3598 // The following call works fine even if the array type is polymorphic.
3599 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3600 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3601 result_reg->init_req(_normal_path, control());
3602 result_val->init_req(_normal_path, obj);
3603 result_io ->init_req(_normal_path, i_o());
3604 result_mem->init_req(_normal_path, reset_memory());
3605 }
3606
3607 // Return the combined state.
3608 set_i_o( _gvn.transform(result_io) );
3609 set_all_memory( _gvn.transform(result_mem));
3610
3611 C->set_has_split_ifs(true); // Has chance for split-if optimization
3612 set_result(result_reg, result_val);
3613 return true;
3614 }
3615
3616 //----------------------inline_native_getLength--------------------------
3617 // public static native int java.lang.reflect.Array.getLength(Object array);
3618 bool LibraryCallKit::inline_native_getLength() {
3619 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3620
3621 Node* array = null_check(argument(0));
3622 // If array is dead, only null-path is taken.
3623 if (stopped()) return true;
3624
3625 // Deoptimize if it is a non-array.
3626 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3627
3628 if (non_array != NULL) {
3629 PreserveJVMState pjvms(this);
3630 set_control(non_array);
3631 uncommon_trap(Deoptimization::Reason_intrinsic,
3632 Deoptimization::Action_maybe_recompile);
3633 }
3634
3635 // If control is dead, only non-array-path is taken.
3636 if (stopped()) return true;
3637
3638 // The works fine even if the array type is polymorphic.
3639 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3640 Node* result = load_array_length(array);
3641
3642 C->set_has_split_ifs(true); // Has chance for split-if optimization
3643 set_result(result);
3644 return true;
3645 }
3646
3647 //------------------------inline_array_copyOf----------------------------
3648 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3649 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3650 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3651 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3652
3653 // Get the arguments.
3654 Node* original = argument(0);
3655 Node* start = is_copyOfRange? argument(1): intcon(0);
3656 Node* end = is_copyOfRange? argument(2): argument(1);
3657 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3658
3659 Node* newcopy;
3660
3661 // Set the original stack and the reexecute bit for the interpreter to reexecute
3662 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3663 { PreserveReexecuteState preexecs(this);
3664 jvms()->set_should_reexecute(true);
3665
3666 array_type_mirror = null_check(array_type_mirror);
3667 original = null_check(original);
3668
3669 // Check if a null path was taken unconditionally.
3670 if (stopped()) return true;
3671
3672 Node* orig_length = load_array_length(original);
3673
3674 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3675 klass_node = null_check(klass_node);
3676
3677 RegionNode* bailout = new RegionNode(1);
3678 record_for_igvn(bailout);
3679
3680 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3681 // Bail out if that is so.
3682 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3683 if (not_objArray != NULL) {
3684 // Improve the klass node's type from the new optimistic assumption:
3685 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3686 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3687 Node* cast = new CastPPNode(klass_node, akls);
3688 cast->init_req(0, control());
3689 klass_node = _gvn.transform(cast);
3690 }
3691
3692 // Bail out if either start or end is negative.
3693 generate_negative_guard(start, bailout, &start);
3694 generate_negative_guard(end, bailout, &end);
3695
3696 Node* length = end;
3697 if (_gvn.type(start) != TypeInt::ZERO) {
3698 length = _gvn.transform(new SubINode(end, start));
3699 }
3700
3701 // Bail out if length is negative.
3702 // Without this the new_array would throw
3703 // NegativeArraySizeException but IllegalArgumentException is what
3704 // should be thrown
3705 generate_negative_guard(length, bailout, &length);
3706
3707 if (bailout->req() > 1) {
3708 PreserveJVMState pjvms(this);
3709 set_control(_gvn.transform(bailout));
3710 uncommon_trap(Deoptimization::Reason_intrinsic,
3711 Deoptimization::Action_maybe_recompile);
3712 }
3713
3714 if (!stopped()) {
3715 // How many elements will we copy from the original?
3716 // The answer is MinI(orig_length - start, length).
3717 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3718 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3719
3720 // Generate a direct call to the right arraycopy function(s).
3721 // We know the copy is disjoint but we might not know if the
3722 // oop stores need checking.
3723 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3724 // This will fail a store-check if x contains any non-nulls.
3725
3726 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3727 // loads/stores but it is legal only if we're sure the
3728 // Arrays.copyOf would succeed. So we need all input arguments
3729 // to the copyOf to be validated, including that the copy to the
3730 // new array won't trigger an ArrayStoreException. That subtype
3731 // check can be optimized if we know something on the type of
3732 // the input array from type speculation.
3733 if (_gvn.type(klass_node)->singleton()) {
3734 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3735 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3736
3737 int test = C->static_subtype_check(superk, subk);
3738 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3739 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3740 if (t_original->speculative_type() != NULL) {
3741 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3742 }
3743 }
3744 }
3745
3746 bool validated = false;
3747 // Reason_class_check rather than Reason_intrinsic because we
3748 // want to intrinsify even if this traps.
3749 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3750 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3751 klass_node);
3752
3753 if (not_subtype_ctrl != top()) {
3754 PreserveJVMState pjvms(this);
3755 set_control(not_subtype_ctrl);
3756 uncommon_trap(Deoptimization::Reason_class_check,
3757 Deoptimization::Action_make_not_entrant);
3758 assert(stopped(), "Should be stopped");
3759 }
3760 validated = true;
3761 }
3762
3763 if (!stopped()) {
3764 newcopy = new_array(klass_node, length, 0); // no arguments to push
3765
3766 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3767 load_object_klass(original), klass_node);
3768 if (!is_copyOfRange) {
3769 ac->set_copyof(validated);
3770 } else {
3771 ac->set_copyofrange(validated);
3772 }
3773 Node* n = _gvn.transform(ac);
3774 if (n == ac) {
3775 ac->connect_outputs(this);
3776 } else {
3777 assert(validated, "shouldn't transform if all arguments not validated");
3778 set_all_memory(n);
3779 }
3780 }
3781 }
3782 } // original reexecute is set back here
3783
3784 C->set_has_split_ifs(true); // Has chance for split-if optimization
3785 if (!stopped()) {
3786 set_result(newcopy);
3787 }
3788 return true;
3789 }
3790
3791
3792 //----------------------generate_virtual_guard---------------------------
3793 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3794 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3795 RegionNode* slow_region) {
3796 ciMethod* method = callee();
3797 int vtable_index = method->vtable_index();
3798 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3799 "bad index %d", vtable_index);
3800 // Get the Method* out of the appropriate vtable entry.
3801 int entry_offset = (InstanceKlass::vtable_start_offset() +
3802 vtable_index*vtableEntry::size()) * wordSize +
3803 vtableEntry::method_offset_in_bytes();
3804 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3805 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3806
3807 // Compare the target method with the expected method (e.g., Object.hashCode).
3808 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3809
3810 Node* native_call = makecon(native_call_addr);
3811 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3812 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3813
3814 return generate_slow_guard(test_native, slow_region);
3815 }
3816
3817 //-----------------------generate_method_call----------------------------
3818 // Use generate_method_call to make a slow-call to the real
3819 // method if the fast path fails. An alternative would be to
3820 // use a stub like OptoRuntime::slow_arraycopy_Java.
3821 // This only works for expanding the current library call,
3822 // not another intrinsic. (E.g., don't use this for making an
3823 // arraycopy call inside of the copyOf intrinsic.)
3824 CallJavaNode*
3825 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3826 // When compiling the intrinsic method itself, do not use this technique.
3827 guarantee(callee() != C->method(), "cannot make slow-call to self");
3828
3829 ciMethod* method = callee();
3830 // ensure the JVMS we have will be correct for this call
3831 guarantee(method_id == method->intrinsic_id(), "must match");
3832
3833 const TypeFunc* tf = TypeFunc::make(method);
3834 CallJavaNode* slow_call;
3835 if (is_static) {
3836 assert(!is_virtual, "");
3837 slow_call = new CallStaticJavaNode(C, tf,
3838 SharedRuntime::get_resolve_static_call_stub(),
3839 method, bci());
3840 } else if (is_virtual) {
3841 null_check_receiver();
3842 int vtable_index = Method::invalid_vtable_index;
3843 if (UseInlineCaches) {
3844 // Suppress the vtable call
3845 } else {
3846 // hashCode and clone are not a miranda methods,
3847 // so the vtable index is fixed.
3848 // No need to use the linkResolver to get it.
3849 vtable_index = method->vtable_index();
3850 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3851 "bad index %d", vtable_index);
3852 }
3853 slow_call = new CallDynamicJavaNode(tf,
3854 SharedRuntime::get_resolve_virtual_call_stub(),
3855 method, vtable_index, bci());
3856 } else { // neither virtual nor static: opt_virtual
3857 null_check_receiver();
3858 slow_call = new CallStaticJavaNode(C, tf,
3859 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3860 method, bci());
3861 slow_call->set_optimized_virtual(true);
3862 }
3863 set_arguments_for_java_call(slow_call);
3864 set_edges_for_java_call(slow_call);
3865 return slow_call;
3866 }
3867
3868
3869 /**
3870 * Build special case code for calls to hashCode on an object. This call may
3871 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3872 * slightly different code.
3873 */
3874 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3875 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3876 assert(!(is_virtual && is_static), "either virtual, special, or static");
3877
3878 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3879
3880 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3881 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3882 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3883 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3884 Node* obj = NULL;
3885 if (!is_static) {
3886 // Check for hashing null object
3887 obj = null_check_receiver();
3888 if (stopped()) return true; // unconditionally null
3889 result_reg->init_req(_null_path, top());
3890 result_val->init_req(_null_path, top());
3891 } else {
3892 // Do a null check, and return zero if null.
3893 // System.identityHashCode(null) == 0
3894 obj = argument(0);
3895 Node* null_ctl = top();
3896 obj = null_check_oop(obj, &null_ctl);
3897 result_reg->init_req(_null_path, null_ctl);
3898 result_val->init_req(_null_path, _gvn.intcon(0));
3899 }
3900
3901 // Unconditionally null? Then return right away.
3902 if (stopped()) {
3903 set_control( result_reg->in(_null_path));
3904 if (!stopped())
3905 set_result(result_val->in(_null_path));
3906 return true;
3907 }
3908
3909 // We only go to the fast case code if we pass a number of guards. The
3910 // paths which do not pass are accumulated in the slow_region.
3911 RegionNode* slow_region = new RegionNode(1);
3912 record_for_igvn(slow_region);
3913
3914 // If this is a virtual call, we generate a funny guard. We pull out
3915 // the vtable entry corresponding to hashCode() from the target object.
3916 // If the target method which we are calling happens to be the native
3917 // Object hashCode() method, we pass the guard. We do not need this
3918 // guard for non-virtual calls -- the caller is known to be the native
3919 // Object hashCode().
3920 if (is_virtual) {
3921 // After null check, get the object's klass.
3922 Node* obj_klass = load_object_klass(obj);
3923 generate_virtual_guard(obj_klass, slow_region);
3924 }
3925
3926 // Get the header out of the object, use LoadMarkNode when available
3927 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3928 // The control of the load must be NULL. Otherwise, the load can move before
3929 // the null check after castPP removal.
3930 Node* no_ctrl = NULL;
3931 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3932
3933 // Test the header to see if it is unlocked.
3934 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3935 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3936 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3937 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3938 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3939
3940 generate_slow_guard(test_unlocked, slow_region);
3941
3942 // Get the hash value and check to see that it has been properly assigned.
3943 // We depend on hash_mask being at most 32 bits and avoid the use of
3944 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3945 // vm: see markOop.hpp.
3946 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3947 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3948 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3949 // This hack lets the hash bits live anywhere in the mark object now, as long
3950 // as the shift drops the relevant bits into the low 32 bits. Note that
3951 // Java spec says that HashCode is an int so there's no point in capturing
3952 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3953 hshifted_header = ConvX2I(hshifted_header);
3954 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3955
3956 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3957 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3958 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3959
3960 generate_slow_guard(test_assigned, slow_region);
3961
3962 Node* init_mem = reset_memory();
3963 // fill in the rest of the null path:
3964 result_io ->init_req(_null_path, i_o());
3965 result_mem->init_req(_null_path, init_mem);
3966
3967 result_val->init_req(_fast_path, hash_val);
3968 result_reg->init_req(_fast_path, control());
3969 result_io ->init_req(_fast_path, i_o());
3970 result_mem->init_req(_fast_path, init_mem);
3971
3972 // Generate code for the slow case. We make a call to hashCode().
3973 set_control(_gvn.transform(slow_region));
3974 if (!stopped()) {
3975 // No need for PreserveJVMState, because we're using up the present state.
3976 set_all_memory(init_mem);
3977 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3978 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3979 Node* slow_result = set_results_for_java_call(slow_call);
3980 // this->control() comes from set_results_for_java_call
3981 result_reg->init_req(_slow_path, control());
3982 result_val->init_req(_slow_path, slow_result);
3983 result_io ->set_req(_slow_path, i_o());
3984 result_mem ->set_req(_slow_path, reset_memory());
3985 }
3986
3987 // Return the combined state.
3988 set_i_o( _gvn.transform(result_io) );
3989 set_all_memory( _gvn.transform(result_mem));
3990
3991 set_result(result_reg, result_val);
3992 return true;
3993 }
3994
3995 //---------------------------inline_native_getClass----------------------------
3996 // public final native Class<?> java.lang.Object.getClass();
3997 //
3998 // Build special case code for calls to getClass on an object.
3999 bool LibraryCallKit::inline_native_getClass() {
4000 Node* obj = null_check_receiver();
4001 if (stopped()) return true;
4002 set_result(load_mirror_from_klass(load_object_klass(obj)));
4003 return true;
4004 }
4005
4006 //-----------------inline_native_Reflection_getCallerClass---------------------
4007 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4008 //
4009 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4010 //
4011 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4012 // in that it must skip particular security frames and checks for
4013 // caller sensitive methods.
4014 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4015 #ifndef PRODUCT
4016 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4017 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4018 }
4019 #endif
4020
4021 if (!jvms()->has_method()) {
4022 #ifndef PRODUCT
4023 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4024 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4025 }
4026 #endif
4027 return false;
4028 }
4029
4030 // Walk back up the JVM state to find the caller at the required
4031 // depth.
4032 JVMState* caller_jvms = jvms();
4033
4034 // Cf. JVM_GetCallerClass
4035 // NOTE: Start the loop at depth 1 because the current JVM state does
4036 // not include the Reflection.getCallerClass() frame.
4037 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4038 ciMethod* m = caller_jvms->method();
4039 switch (n) {
4040 case 0:
4041 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4042 break;
4043 case 1:
4044 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4045 if (!m->caller_sensitive()) {
4046 #ifndef PRODUCT
4047 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4048 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4049 }
4050 #endif
4051 return false; // bail-out; let JVM_GetCallerClass do the work
4052 }
4053 break;
4054 default:
4055 if (!m->is_ignored_by_security_stack_walk()) {
4056 // We have reached the desired frame; return the holder class.
4057 // Acquire method holder as java.lang.Class and push as constant.
4058 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4059 ciInstance* caller_mirror = caller_klass->java_mirror();
4060 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4061
4062 #ifndef PRODUCT
4063 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4064 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());
4065 tty->print_cr(" JVM state at this point:");
4066 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4067 ciMethod* m = jvms()->of_depth(i)->method();
4068 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4069 }
4070 }
4071 #endif
4072 return true;
4073 }
4074 break;
4075 }
4076 }
4077
4078 #ifndef PRODUCT
4079 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4080 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4081 tty->print_cr(" JVM state at this point:");
4082 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4083 ciMethod* m = jvms()->of_depth(i)->method();
4084 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4085 }
4086 }
4087 #endif
4088
4089 return false; // bail-out; let JVM_GetCallerClass do the work
4090 }
4091
4092 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4093 Node* arg = argument(0);
4094 Node* result;
4095
4096 switch (id) {
4097 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4098 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4099 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4100 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4101
4102 case vmIntrinsics::_doubleToLongBits: {
4103 // two paths (plus control) merge in a wood
4104 RegionNode *r = new RegionNode(3);
4105 Node *phi = new PhiNode(r, TypeLong::LONG);
4106
4107 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4108 // Build the boolean node
4109 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4110
4111 // Branch either way.
4112 // NaN case is less traveled, which makes all the difference.
4113 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4114 Node *opt_isnan = _gvn.transform(ifisnan);
4115 assert( opt_isnan->is_If(), "Expect an IfNode");
4116 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4117 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4118
4119 set_control(iftrue);
4120
4121 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4122 Node *slow_result = longcon(nan_bits); // return NaN
4123 phi->init_req(1, _gvn.transform( slow_result ));
4124 r->init_req(1, iftrue);
4125
4126 // Else fall through
4127 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4128 set_control(iffalse);
4129
4130 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4131 r->init_req(2, iffalse);
4132
4133 // Post merge
4134 set_control(_gvn.transform(r));
4135 record_for_igvn(r);
4136
4137 C->set_has_split_ifs(true); // Has chance for split-if optimization
4138 result = phi;
4139 assert(result->bottom_type()->isa_long(), "must be");
4140 break;
4141 }
4142
4143 case vmIntrinsics::_floatToIntBits: {
4144 // two paths (plus control) merge in a wood
4145 RegionNode *r = new RegionNode(3);
4146 Node *phi = new PhiNode(r, TypeInt::INT);
4147
4148 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4149 // Build the boolean node
4150 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4151
4152 // Branch either way.
4153 // NaN case is less traveled, which makes all the difference.
4154 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4155 Node *opt_isnan = _gvn.transform(ifisnan);
4156 assert( opt_isnan->is_If(), "Expect an IfNode");
4157 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4158 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4159
4160 set_control(iftrue);
4161
4162 static const jint nan_bits = 0x7fc00000;
4163 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4164 phi->init_req(1, _gvn.transform( slow_result ));
4165 r->init_req(1, iftrue);
4166
4167 // Else fall through
4168 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4169 set_control(iffalse);
4170
4171 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4172 r->init_req(2, iffalse);
4173
4174 // Post merge
4175 set_control(_gvn.transform(r));
4176 record_for_igvn(r);
4177
4178 C->set_has_split_ifs(true); // Has chance for split-if optimization
4179 result = phi;
4180 assert(result->bottom_type()->isa_int(), "must be");
4181 break;
4182 }
4183
4184 default:
4185 fatal_unexpected_iid(id);
4186 break;
4187 }
4188 set_result(_gvn.transform(result));
4189 return true;
4190 }
4191
4192 #ifdef _LP64
4193 #define XTOP ,top() /*additional argument*/
4194 #else //_LP64
4195 #define XTOP /*no additional argument*/
4196 #endif //_LP64
4197
4198 //----------------------inline_unsafe_copyMemory-------------------------
4199 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4200 bool LibraryCallKit::inline_unsafe_copyMemory() {
4201 if (callee()->is_static()) return false; // caller must have the capability!
4202 null_check_receiver(); // null-check receiver
4203 if (stopped()) return true;
4204
4205 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4206
4207 Node* src_ptr = argument(1); // type: oop
4208 Node* src_off = ConvL2X(argument(2)); // type: long
4209 Node* dst_ptr = argument(4); // type: oop
4210 Node* dst_off = ConvL2X(argument(5)); // type: long
4211 Node* size = ConvL2X(argument(7)); // type: long
4212
4213 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4214 "fieldOffset must be byte-scaled");
4215
4216 Node* src = make_unsafe_address(src_ptr, src_off);
4217 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4218
4219 // Conservatively insert a memory barrier on all memory slices.
4220 // Do not let writes of the copy source or destination float below the copy.
4221 insert_mem_bar(Op_MemBarCPUOrder);
4222
4223 // Call it. Note that the length argument is not scaled.
4224 make_runtime_call(RC_LEAF|RC_NO_FP,
4225 OptoRuntime::fast_arraycopy_Type(),
4226 StubRoutines::unsafe_arraycopy(),
4227 "unsafe_arraycopy",
4228 TypeRawPtr::BOTTOM,
4229 src, dst, size XTOP);
4230
4231 // Do not let reads of the copy destination float above the copy.
4232 insert_mem_bar(Op_MemBarCPUOrder);
4233
4234 return true;
4235 }
4236
4237 //------------------------clone_coping-----------------------------------
4238 // Helper function for inline_native_clone.
4239 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4240 assert(obj_size != NULL, "");
4241 Node* raw_obj = alloc_obj->in(1);
4242 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4243
4244 AllocateNode* alloc = NULL;
4245 if (ReduceBulkZeroing) {
4246 // We will be completely responsible for initializing this object -
4247 // mark Initialize node as complete.
4248 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4249 // The object was just allocated - there should be no any stores!
4250 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4251 // Mark as complete_with_arraycopy so that on AllocateNode
4252 // expansion, we know this AllocateNode is initialized by an array
4253 // copy and a StoreStore barrier exists after the array copy.
4254 alloc->initialization()->set_complete_with_arraycopy();
4255 }
4256
4257 // Copy the fastest available way.
4258 // TODO: generate fields copies for small objects instead.
4259 Node* src = obj;
4260 Node* dest = alloc_obj;
4261 Node* size = _gvn.transform(obj_size);
4262
4263 // Exclude the header but include array length to copy by 8 bytes words.
4264 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4265 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4266 instanceOopDesc::base_offset_in_bytes();
4267 // base_off:
4268 // 8 - 32-bit VM
4269 // 12 - 64-bit VM, compressed klass
4270 // 16 - 64-bit VM, normal klass
4271 if (base_off % BytesPerLong != 0) {
4272 assert(UseCompressedClassPointers, "");
4273 if (is_array) {
4274 // Exclude length to copy by 8 bytes words.
4275 base_off += sizeof(int);
4276 } else {
4277 // Include klass to copy by 8 bytes words.
4278 base_off = instanceOopDesc::klass_offset_in_bytes();
4279 }
4280 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4281 }
4282 src = basic_plus_adr(src, base_off);
4283 dest = basic_plus_adr(dest, base_off);
4284
4285 // Compute the length also, if needed:
4286 Node* countx = size;
4287 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4288 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4289
4290 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4291
4292 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4293 ac->set_clonebasic();
4294 Node* n = _gvn.transform(ac);
4295 if (n == ac) {
4296 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4297 } else {
4298 set_all_memory(n);
4299 }
4300
4301 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4302 if (card_mark) {
4303 assert(!is_array, "");
4304 // Put in store barrier for any and all oops we are sticking
4305 // into this object. (We could avoid this if we could prove
4306 // that the object type contains no oop fields at all.)
4307 Node* no_particular_value = NULL;
4308 Node* no_particular_field = NULL;
4309 int raw_adr_idx = Compile::AliasIdxRaw;
4310 post_barrier(control(),
4311 memory(raw_adr_type),
4312 alloc_obj,
4313 no_particular_field,
4314 raw_adr_idx,
4315 no_particular_value,
4316 T_OBJECT,
4317 false);
4318 }
4319
4320 // Do not let reads from the cloned object float above the arraycopy.
4321 if (alloc != NULL) {
4322 // Do not let stores that initialize this object be reordered with
4323 // a subsequent store that would make this object accessible by
4324 // other threads.
4325 // Record what AllocateNode this StoreStore protects so that
4326 // escape analysis can go from the MemBarStoreStoreNode to the
4327 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4328 // based on the escape status of the AllocateNode.
4329 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4330 } else {
4331 insert_mem_bar(Op_MemBarCPUOrder);
4332 }
4333 }
4334
4335 //------------------------inline_native_clone----------------------------
4336 // protected native Object java.lang.Object.clone();
4337 //
4338 // Here are the simple edge cases:
4339 // null receiver => normal trap
4340 // virtual and clone was overridden => slow path to out-of-line clone
4341 // not cloneable or finalizer => slow path to out-of-line Object.clone
4342 //
4343 // The general case has two steps, allocation and copying.
4344 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4345 //
4346 // Copying also has two cases, oop arrays and everything else.
4347 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4348 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4349 //
4350 // These steps fold up nicely if and when the cloned object's klass
4351 // can be sharply typed as an object array, a type array, or an instance.
4352 //
4353 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4354 PhiNode* result_val;
4355
4356 // Set the reexecute bit for the interpreter to reexecute
4357 // the bytecode that invokes Object.clone if deoptimization happens.
4358 { PreserveReexecuteState preexecs(this);
4359 jvms()->set_should_reexecute(true);
4360
4361 Node* obj = null_check_receiver();
4362 if (stopped()) return true;
4363
4364 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4365
4366 // If we are going to clone an instance, we need its exact type to
4367 // know the number and types of fields to convert the clone to
4368 // loads/stores. Maybe a speculative type can help us.
4369 if (!obj_type->klass_is_exact() &&
4370 obj_type->speculative_type() != NULL &&
4371 obj_type->speculative_type()->is_instance_klass()) {
4372 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4373 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4374 !spec_ik->has_injected_fields()) {
4375 ciKlass* k = obj_type->klass();
4376 if (!k->is_instance_klass() ||
4377 k->as_instance_klass()->is_interface() ||
4378 k->as_instance_klass()->has_subklass()) {
4379 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4380 }
4381 }
4382 }
4383
4384 Node* obj_klass = load_object_klass(obj);
4385 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4386 const TypeOopPtr* toop = ((tklass != NULL)
4387 ? tklass->as_instance_type()
4388 : TypeInstPtr::NOTNULL);
4389
4390 // Conservatively insert a memory barrier on all memory slices.
4391 // Do not let writes into the original float below the clone.
4392 insert_mem_bar(Op_MemBarCPUOrder);
4393
4394 // paths into result_reg:
4395 enum {
4396 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4397 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4398 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4399 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4400 PATH_LIMIT
4401 };
4402 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4403 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4404 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4405 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4406 record_for_igvn(result_reg);
4407
4408 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4409 int raw_adr_idx = Compile::AliasIdxRaw;
4410
4411 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4412 if (array_ctl != NULL) {
4413 // It's an array.
4414 PreserveJVMState pjvms(this);
4415 set_control(array_ctl);
4416 Node* obj_length = load_array_length(obj);
4417 Node* obj_size = NULL;
4418 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4419
4420 if (!use_ReduceInitialCardMarks()) {
4421 // If it is an oop array, it requires very special treatment,
4422 // because card marking is required on each card of the array.
4423 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4424 if (is_obja != NULL) {
4425 PreserveJVMState pjvms2(this);
4426 set_control(is_obja);
4427 // Generate a direct call to the right arraycopy function(s).
4428 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4429 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4430 ac->set_cloneoop();
4431 Node* n = _gvn.transform(ac);
4432 assert(n == ac, "cannot disappear");
4433 ac->connect_outputs(this);
4434
4435 result_reg->init_req(_objArray_path, control());
4436 result_val->init_req(_objArray_path, alloc_obj);
4437 result_i_o ->set_req(_objArray_path, i_o());
4438 result_mem ->set_req(_objArray_path, reset_memory());
4439 }
4440 }
4441 // Otherwise, there are no card marks to worry about.
4442 // (We can dispense with card marks if we know the allocation
4443 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4444 // causes the non-eden paths to take compensating steps to
4445 // simulate a fresh allocation, so that no further
4446 // card marks are required in compiled code to initialize
4447 // the object.)
4448
4449 if (!stopped()) {
4450 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4451
4452 // Present the results of the copy.
4453 result_reg->init_req(_array_path, control());
4454 result_val->init_req(_array_path, alloc_obj);
4455 result_i_o ->set_req(_array_path, i_o());
4456 result_mem ->set_req(_array_path, reset_memory());
4457 }
4458 }
4459
4460 // We only go to the instance fast case code if we pass a number of guards.
4461 // The paths which do not pass are accumulated in the slow_region.
4462 RegionNode* slow_region = new RegionNode(1);
4463 record_for_igvn(slow_region);
4464 if (!stopped()) {
4465 // It's an instance (we did array above). Make the slow-path tests.
4466 // If this is a virtual call, we generate a funny guard. We grab
4467 // the vtable entry corresponding to clone() from the target object.
4468 // If the target method which we are calling happens to be the
4469 // Object clone() method, we pass the guard. We do not need this
4470 // guard for non-virtual calls; the caller is known to be the native
4471 // Object clone().
4472 if (is_virtual) {
4473 generate_virtual_guard(obj_klass, slow_region);
4474 }
4475
4476 // The object must be cloneable and must not have a finalizer.
4477 // Both of these conditions may be checked in a single test.
4478 // We could optimize the cloneable test further, but we don't care.
4479 generate_access_flags_guard(obj_klass,
4480 // Test both conditions:
4481 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4482 // Must be cloneable but not finalizer:
4483 JVM_ACC_IS_CLONEABLE,
4484 slow_region);
4485 }
4486
4487 if (!stopped()) {
4488 // It's an instance, and it passed the slow-path tests.
4489 PreserveJVMState pjvms(this);
4490 Node* obj_size = NULL;
4491 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4492 // is reexecuted if deoptimization occurs and there could be problems when merging
4493 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4494 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4495
4496 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4497
4498 // Present the results of the slow call.
4499 result_reg->init_req(_instance_path, control());
4500 result_val->init_req(_instance_path, alloc_obj);
4501 result_i_o ->set_req(_instance_path, i_o());
4502 result_mem ->set_req(_instance_path, reset_memory());
4503 }
4504
4505 // Generate code for the slow case. We make a call to clone().
4506 set_control(_gvn.transform(slow_region));
4507 if (!stopped()) {
4508 PreserveJVMState pjvms(this);
4509 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4510 Node* slow_result = set_results_for_java_call(slow_call);
4511 // this->control() comes from set_results_for_java_call
4512 result_reg->init_req(_slow_path, control());
4513 result_val->init_req(_slow_path, slow_result);
4514 result_i_o ->set_req(_slow_path, i_o());
4515 result_mem ->set_req(_slow_path, reset_memory());
4516 }
4517
4518 // Return the combined state.
4519 set_control( _gvn.transform(result_reg));
4520 set_i_o( _gvn.transform(result_i_o));
4521 set_all_memory( _gvn.transform(result_mem));
4522 } // original reexecute is set back here
4523
4524 set_result(_gvn.transform(result_val));
4525 return true;
4526 }
4527
4528 // If we have a tighly coupled allocation, the arraycopy may take care
4529 // of the array initialization. If one of the guards we insert between
4530 // the allocation and the arraycopy causes a deoptimization, an
4531 // unitialized array will escape the compiled method. To prevent that
4532 // we set the JVM state for uncommon traps between the allocation and
4533 // the arraycopy to the state before the allocation so, in case of
4534 // deoptimization, we'll reexecute the allocation and the
4535 // initialization.
4536 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4537 if (alloc != NULL) {
4538 ciMethod* trap_method = alloc->jvms()->method();
4539 int trap_bci = alloc->jvms()->bci();
4540
4541 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4542 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4543 // Make sure there's no store between the allocation and the
4544 // arraycopy otherwise visible side effects could be rexecuted
4545 // in case of deoptimization and cause incorrect execution.
4546 bool no_interfering_store = true;
4547 Node* mem = alloc->in(TypeFunc::Memory);
4548 if (mem->is_MergeMem()) {
4549 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4550 Node* n = mms.memory();
4551 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4552 assert(n->is_Store(), "what else?");
4553 no_interfering_store = false;
4554 break;
4555 }
4556 }
4557 } else {
4558 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4559 Node* n = mms.memory();
4560 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4561 assert(n->is_Store(), "what else?");
4562 no_interfering_store = false;
4563 break;
4564 }
4565 }
4566 }
4567
4568 if (no_interfering_store) {
4569 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4570 uint size = alloc->req();
4571 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4572 old_jvms->set_map(sfpt);
4573 for (uint i = 0; i < size; i++) {
4574 sfpt->init_req(i, alloc->in(i));
4575 }
4576 // re-push array length for deoptimization
4577 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4578 old_jvms->set_sp(old_jvms->sp()+1);
4579 old_jvms->set_monoff(old_jvms->monoff()+1);
4580 old_jvms->set_scloff(old_jvms->scloff()+1);
4581 old_jvms->set_endoff(old_jvms->endoff()+1);
4582 old_jvms->set_should_reexecute(true);
4583
4584 sfpt->set_i_o(map()->i_o());
4585 sfpt->set_memory(map()->memory());
4586 sfpt->set_control(map()->control());
4587
4588 JVMState* saved_jvms = jvms();
4589 saved_reexecute_sp = _reexecute_sp;
4590
4591 set_jvms(sfpt->jvms());
4592 _reexecute_sp = jvms()->sp();
4593
4594 return saved_jvms;
4595 }
4596 }
4597 }
4598 return NULL;
4599 }
4600
4601 // In case of a deoptimization, we restart execution at the
4602 // allocation, allocating a new array. We would leave an uninitialized
4603 // array in the heap that GCs wouldn't expect. Move the allocation
4604 // after the traps so we don't allocate the array if we
4605 // deoptimize. This is possible because tightly_coupled_allocation()
4606 // guarantees there's no observer of the allocated array at this point
4607 // and the control flow is simple enough.
4608 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4609 if (saved_jvms != NULL && !stopped()) {
4610 assert(alloc != NULL, "only with a tightly coupled allocation");
4611 // restore JVM state to the state at the arraycopy
4612 saved_jvms->map()->set_control(map()->control());
4613 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4614 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4615 // If we've improved the types of some nodes (null check) while
4616 // emitting the guards, propagate them to the current state
4617 map()->replaced_nodes().apply(saved_jvms->map());
4618 set_jvms(saved_jvms);
4619 _reexecute_sp = saved_reexecute_sp;
4620
4621 // Remove the allocation from above the guards
4622 CallProjections callprojs;
4623 alloc->extract_projections(&callprojs, true);
4624 InitializeNode* init = alloc->initialization();
4625 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4626 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4627 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4628 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4629
4630 // move the allocation here (after the guards)
4631 _gvn.hash_delete(alloc);
4632 alloc->set_req(TypeFunc::Control, control());
4633 alloc->set_req(TypeFunc::I_O, i_o());
4634 Node *mem = reset_memory();
4635 set_all_memory(mem);
4636 alloc->set_req(TypeFunc::Memory, mem);
4637 set_control(init->proj_out(TypeFunc::Control));
4638 set_i_o(callprojs.fallthrough_ioproj);
4639
4640 // Update memory as done in GraphKit::set_output_for_allocation()
4641 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4642 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4643 if (ary_type->isa_aryptr() && length_type != NULL) {
4644 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4645 }
4646 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4647 int elemidx = C->get_alias_index(telemref);
4648 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4649 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4650
4651 Node* allocx = _gvn.transform(alloc);
4652 assert(allocx == alloc, "where has the allocation gone?");
4653 assert(dest->is_CheckCastPP(), "not an allocation result?");
4654
4655 _gvn.hash_delete(dest);
4656 dest->set_req(0, control());
4657 Node* destx = _gvn.transform(dest);
4658 assert(destx == dest, "where has the allocation result gone?");
4659 }
4660 }
4661
4662
4663 //------------------------------inline_arraycopy-----------------------
4664 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4665 // Object dest, int destPos,
4666 // int length);
4667 bool LibraryCallKit::inline_arraycopy() {
4668 // Get the arguments.
4669 Node* src = argument(0); // type: oop
4670 Node* src_offset = argument(1); // type: int
4671 Node* dest = argument(2); // type: oop
4672 Node* dest_offset = argument(3); // type: int
4673 Node* length = argument(4); // type: int
4674
4675
4676 // Check for allocation before we add nodes that would confuse
4677 // tightly_coupled_allocation()
4678 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4679
4680 int saved_reexecute_sp = -1;
4681 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4682 // See arraycopy_restore_alloc_state() comment
4683 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4684 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4685 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4686 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4687
4688 // The following tests must be performed
4689 // (1) src and dest are arrays.
4690 // (2) src and dest arrays must have elements of the same BasicType
4691 // (3) src and dest must not be null.
4692 // (4) src_offset must not be negative.
4693 // (5) dest_offset must not be negative.
4694 // (6) length must not be negative.
4695 // (7) src_offset + length must not exceed length of src.
4696 // (8) dest_offset + length must not exceed length of dest.
4697 // (9) each element of an oop array must be assignable
4698
4699 // (3) src and dest must not be null.
4700 // always do this here because we need the JVM state for uncommon traps
4701 Node* null_ctl = top();
4702 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4703 assert(null_ctl->is_top(), "no null control here");
4704 dest = null_check(dest, T_ARRAY);
4705
4706 if (!can_emit_guards) {
4707 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4708 // guards but the arraycopy node could still take advantage of a
4709 // tightly allocated allocation. tightly_coupled_allocation() is
4710 // called again to make sure it takes the null check above into
4711 // account: the null check is mandatory and if it caused an
4712 // uncommon trap to be emitted then the allocation can't be
4713 // considered tightly coupled in this context.
4714 alloc = tightly_coupled_allocation(dest, NULL);
4715 }
4716
4717 bool validated = false;
4718
4719 const Type* src_type = _gvn.type(src);
4720 const Type* dest_type = _gvn.type(dest);
4721 const TypeAryPtr* top_src = src_type->isa_aryptr();
4722 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4723
4724 // Do we have the type of src?
4725 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4726 // Do we have the type of dest?
4727 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4728 // Is the type for src from speculation?
4729 bool src_spec = false;
4730 // Is the type for dest from speculation?
4731 bool dest_spec = false;
4732
4733 if ((!has_src || !has_dest) && can_emit_guards) {
4734 // We don't have sufficient type information, let's see if
4735 // speculative types can help. We need to have types for both src
4736 // and dest so that it pays off.
4737
4738 // Do we already have or could we have type information for src
4739 bool could_have_src = has_src;
4740 // Do we already have or could we have type information for dest
4741 bool could_have_dest = has_dest;
4742
4743 ciKlass* src_k = NULL;
4744 if (!has_src) {
4745 src_k = src_type->speculative_type_not_null();
4746 if (src_k != NULL && src_k->is_array_klass()) {
4747 could_have_src = true;
4748 }
4749 }
4750
4751 ciKlass* dest_k = NULL;
4752 if (!has_dest) {
4753 dest_k = dest_type->speculative_type_not_null();
4754 if (dest_k != NULL && dest_k->is_array_klass()) {
4755 could_have_dest = true;
4756 }
4757 }
4758
4759 if (could_have_src && could_have_dest) {
4760 // This is going to pay off so emit the required guards
4761 if (!has_src) {
4762 src = maybe_cast_profiled_obj(src, src_k, true);
4763 src_type = _gvn.type(src);
4764 top_src = src_type->isa_aryptr();
4765 has_src = (top_src != NULL && top_src->klass() != NULL);
4766 src_spec = true;
4767 }
4768 if (!has_dest) {
4769 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4770 dest_type = _gvn.type(dest);
4771 top_dest = dest_type->isa_aryptr();
4772 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4773 dest_spec = true;
4774 }
4775 }
4776 }
4777
4778 if (has_src && has_dest && can_emit_guards) {
4779 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4780 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4781 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4782 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4783
4784 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4785 // If both arrays are object arrays then having the exact types
4786 // for both will remove the need for a subtype check at runtime
4787 // before the call and may make it possible to pick a faster copy
4788 // routine (without a subtype check on every element)
4789 // Do we have the exact type of src?
4790 bool could_have_src = src_spec;
4791 // Do we have the exact type of dest?
4792 bool could_have_dest = dest_spec;
4793 ciKlass* src_k = top_src->klass();
4794 ciKlass* dest_k = top_dest->klass();
4795 if (!src_spec) {
4796 src_k = src_type->speculative_type_not_null();
4797 if (src_k != NULL && src_k->is_array_klass()) {
4798 could_have_src = true;
4799 }
4800 }
4801 if (!dest_spec) {
4802 dest_k = dest_type->speculative_type_not_null();
4803 if (dest_k != NULL && dest_k->is_array_klass()) {
4804 could_have_dest = true;
4805 }
4806 }
4807 if (could_have_src && could_have_dest) {
4808 // If we can have both exact types, emit the missing guards
4809 if (could_have_src && !src_spec) {
4810 src = maybe_cast_profiled_obj(src, src_k, true);
4811 }
4812 if (could_have_dest && !dest_spec) {
4813 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4814 }
4815 }
4816 }
4817 }
4818
4819 ciMethod* trap_method = method();
4820 int trap_bci = bci();
4821 if (saved_jvms != NULL) {
4822 trap_method = alloc->jvms()->method();
4823 trap_bci = alloc->jvms()->bci();
4824 }
4825
4826 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4827 can_emit_guards &&
4828 !src->is_top() && !dest->is_top()) {
4829 // validate arguments: enables transformation the ArrayCopyNode
4830 validated = true;
4831
4832 RegionNode* slow_region = new RegionNode(1);
4833 record_for_igvn(slow_region);
4834
4835 // (1) src and dest are arrays.
4836 generate_non_array_guard(load_object_klass(src), slow_region);
4837 generate_non_array_guard(load_object_klass(dest), slow_region);
4838
4839 // (2) src and dest arrays must have elements of the same BasicType
4840 // done at macro expansion or at Ideal transformation time
4841
4842 // (4) src_offset must not be negative.
4843 generate_negative_guard(src_offset, slow_region);
4844
4845 // (5) dest_offset must not be negative.
4846 generate_negative_guard(dest_offset, slow_region);
4847
4848 // (7) src_offset + length must not exceed length of src.
4849 generate_limit_guard(src_offset, length,
4850 load_array_length(src),
4851 slow_region);
4852
4853 // (8) dest_offset + length must not exceed length of dest.
4854 generate_limit_guard(dest_offset, length,
4855 load_array_length(dest),
4856 slow_region);
4857
4858 // (9) each element of an oop array must be assignable
4859 Node* src_klass = load_object_klass(src);
4860 Node* dest_klass = load_object_klass(dest);
4861 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4862
4863 if (not_subtype_ctrl != top()) {
4864 PreserveJVMState pjvms(this);
4865 set_control(not_subtype_ctrl);
4866 uncommon_trap(Deoptimization::Reason_intrinsic,
4867 Deoptimization::Action_make_not_entrant);
4868 assert(stopped(), "Should be stopped");
4869 }
4870 {
4871 PreserveJVMState pjvms(this);
4872 set_control(_gvn.transform(slow_region));
4873 uncommon_trap(Deoptimization::Reason_intrinsic,
4874 Deoptimization::Action_make_not_entrant);
4875 assert(stopped(), "Should be stopped");
4876 }
4877 }
4878
4879 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4880
4881 if (stopped()) {
4882 return true;
4883 }
4884
4885 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4886 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4887 // so the compiler has a chance to eliminate them: during macro expansion,
4888 // we have to set their control (CastPP nodes are eliminated).
4889 load_object_klass(src), load_object_klass(dest),
4890 load_array_length(src), load_array_length(dest));
4891
4892 ac->set_arraycopy(validated);
4893
4894 Node* n = _gvn.transform(ac);
4895 if (n == ac) {
4896 ac->connect_outputs(this);
4897 } else {
4898 assert(validated, "shouldn't transform if all arguments not validated");
4899 set_all_memory(n);
4900 }
4901
4902 return true;
4903 }
4904
4905
4906 // Helper function which determines if an arraycopy immediately follows
4907 // an allocation, with no intervening tests or other escapes for the object.
4908 AllocateArrayNode*
4909 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4910 RegionNode* slow_region) {
4911 if (stopped()) return NULL; // no fast path
4912 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4913
4914 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4915 if (alloc == NULL) return NULL;
4916
4917 Node* rawmem = memory(Compile::AliasIdxRaw);
4918 // Is the allocation's memory state untouched?
4919 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4920 // Bail out if there have been raw-memory effects since the allocation.
4921 // (Example: There might have been a call or safepoint.)
4922 return NULL;
4923 }
4924 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4925 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4926 return NULL;
4927 }
4928
4929 // There must be no unexpected observers of this allocation.
4930 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4931 Node* obs = ptr->fast_out(i);
4932 if (obs != this->map()) {
4933 return NULL;
4934 }
4935 }
4936
4937 // This arraycopy must unconditionally follow the allocation of the ptr.
4938 Node* alloc_ctl = ptr->in(0);
4939 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4940
4941 Node* ctl = control();
4942 while (ctl != alloc_ctl) {
4943 // There may be guards which feed into the slow_region.
4944 // Any other control flow means that we might not get a chance
4945 // to finish initializing the allocated object.
4946 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4947 IfNode* iff = ctl->in(0)->as_If();
4948 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4949 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4950 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4951 ctl = iff->in(0); // This test feeds the known slow_region.
4952 continue;
4953 }
4954 // One more try: Various low-level checks bottom out in
4955 // uncommon traps. If the debug-info of the trap omits
4956 // any reference to the allocation, as we've already
4957 // observed, then there can be no objection to the trap.
4958 bool found_trap = false;
4959 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4960 Node* obs = not_ctl->fast_out(j);
4961 if (obs->in(0) == not_ctl && obs->is_Call() &&
4962 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4963 found_trap = true; break;
4964 }
4965 }
4966 if (found_trap) {
4967 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4968 continue;
4969 }
4970 }
4971 return NULL;
4972 }
4973
4974 // If we get this far, we have an allocation which immediately
4975 // precedes the arraycopy, and we can take over zeroing the new object.
4976 // The arraycopy will finish the initialization, and provide
4977 // a new control state to which we will anchor the destination pointer.
4978
4979 return alloc;
4980 }
4981
4982 //-------------inline_encodeISOArray-----------------------------------
4983 // encode char[] to byte[] in ISO_8859_1
4984 bool LibraryCallKit::inline_encodeISOArray() {
4985 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4986 // no receiver since it is static method
4987 Node *src = argument(0);
4988 Node *src_offset = argument(1);
4989 Node *dst = argument(2);
4990 Node *dst_offset = argument(3);
4991 Node *length = argument(4);
4992
4993 const Type* src_type = src->Value(&_gvn);
4994 const Type* dst_type = dst->Value(&_gvn);
4995 const TypeAryPtr* top_src = src_type->isa_aryptr();
4996 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4997 if (top_src == NULL || top_src->klass() == NULL ||
4998 top_dest == NULL || top_dest->klass() == NULL) {
4999 // failed array check
5000 return false;
5001 }
5002
5003 // Figure out the size and type of the elements we will be copying.
5004 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5005 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5006 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5007 return false;
5008 }
5009 Node* src_start = array_element_address(src, src_offset, src_elem);
5010 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5011 // 'src_start' points to src array + scaled offset
5012 // 'dst_start' points to dst array + scaled offset
5013
5014 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5015 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5016 enc = _gvn.transform(enc);
5017 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5018 set_memory(res_mem, mtype);
5019 set_result(enc);
5020 return true;
5021 }
5022
5023 //-------------inline_multiplyToLen-----------------------------------
5024 bool LibraryCallKit::inline_multiplyToLen() {
5025 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5026
5027 address stubAddr = StubRoutines::multiplyToLen();
5028 if (stubAddr == NULL) {
5029 return false; // Intrinsic's stub is not implemented on this platform
5030 }
5031 const char* stubName = "multiplyToLen";
5032
5033 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5034
5035 // no receiver because it is a static method
5036 Node* x = argument(0);
5037 Node* xlen = argument(1);
5038 Node* y = argument(2);
5039 Node* ylen = argument(3);
5040 Node* z = argument(4);
5041
5042 const Type* x_type = x->Value(&_gvn);
5043 const Type* y_type = y->Value(&_gvn);
5044 const TypeAryPtr* top_x = x_type->isa_aryptr();
5045 const TypeAryPtr* top_y = y_type->isa_aryptr();
5046 if (top_x == NULL || top_x->klass() == NULL ||
5047 top_y == NULL || top_y->klass() == NULL) {
5048 // failed array check
5049 return false;
5050 }
5051
5052 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5053 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5054 if (x_elem != T_INT || y_elem != T_INT) {
5055 return false;
5056 }
5057
5058 // Set the original stack and the reexecute bit for the interpreter to reexecute
5059 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5060 // on the return from z array allocation in runtime.
5061 { PreserveReexecuteState preexecs(this);
5062 jvms()->set_should_reexecute(true);
5063
5064 Node* x_start = array_element_address(x, intcon(0), x_elem);
5065 Node* y_start = array_element_address(y, intcon(0), y_elem);
5066 // 'x_start' points to x array + scaled xlen
5067 // 'y_start' points to y array + scaled ylen
5068
5069 // Allocate the result array
5070 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5071 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5072 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5073
5074 IdealKit ideal(this);
5075
5076 #define __ ideal.
5077 Node* one = __ ConI(1);
5078 Node* zero = __ ConI(0);
5079 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5080 __ set(need_alloc, zero);
5081 __ set(z_alloc, z);
5082 __ if_then(z, BoolTest::eq, null()); {
5083 __ increment (need_alloc, one);
5084 } __ else_(); {
5085 // Update graphKit memory and control from IdealKit.
5086 sync_kit(ideal);
5087 Node* zlen_arg = load_array_length(z);
5088 // Update IdealKit memory and control from graphKit.
5089 __ sync_kit(this);
5090 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5091 __ increment (need_alloc, one);
5092 } __ end_if();
5093 } __ end_if();
5094
5095 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5096 // Update graphKit memory and control from IdealKit.
5097 sync_kit(ideal);
5098 Node * narr = new_array(klass_node, zlen, 1);
5099 // Update IdealKit memory and control from graphKit.
5100 __ sync_kit(this);
5101 __ set(z_alloc, narr);
5102 } __ end_if();
5103
5104 sync_kit(ideal);
5105 z = __ value(z_alloc);
5106 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5107 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5108 // Final sync IdealKit and GraphKit.
5109 final_sync(ideal);
5110 #undef __
5111
5112 Node* z_start = array_element_address(z, intcon(0), T_INT);
5113
5114 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5115 OptoRuntime::multiplyToLen_Type(),
5116 stubAddr, stubName, TypePtr::BOTTOM,
5117 x_start, xlen, y_start, ylen, z_start, zlen);
5118 } // original reexecute is set back here
5119
5120 C->set_has_split_ifs(true); // Has chance for split-if optimization
5121 set_result(z);
5122 return true;
5123 }
5124
5125 //-------------inline_squareToLen------------------------------------
5126 bool LibraryCallKit::inline_squareToLen() {
5127 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5128
5129 address stubAddr = StubRoutines::squareToLen();
5130 if (stubAddr == NULL) {
5131 return false; // Intrinsic's stub is not implemented on this platform
5132 }
5133 const char* stubName = "squareToLen";
5134
5135 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5136
5137 Node* x = argument(0);
5138 Node* len = argument(1);
5139 Node* z = argument(2);
5140 Node* zlen = argument(3);
5141
5142 const Type* x_type = x->Value(&_gvn);
5143 const Type* z_type = z->Value(&_gvn);
5144 const TypeAryPtr* top_x = x_type->isa_aryptr();
5145 const TypeAryPtr* top_z = z_type->isa_aryptr();
5146 if (top_x == NULL || top_x->klass() == NULL ||
5147 top_z == NULL || top_z->klass() == NULL) {
5148 // failed array check
5149 return false;
5150 }
5151
5152 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5153 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5154 if (x_elem != T_INT || z_elem != T_INT) {
5155 return false;
5156 }
5157
5158
5159 Node* x_start = array_element_address(x, intcon(0), x_elem);
5160 Node* z_start = array_element_address(z, intcon(0), z_elem);
5161
5162 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5163 OptoRuntime::squareToLen_Type(),
5164 stubAddr, stubName, TypePtr::BOTTOM,
5165 x_start, len, z_start, zlen);
5166
5167 set_result(z);
5168 return true;
5169 }
5170
5171 //-------------inline_mulAdd------------------------------------------
5172 bool LibraryCallKit::inline_mulAdd() {
5173 assert(UseMulAddIntrinsic, "not implementated on this platform");
5174
5175 address stubAddr = StubRoutines::mulAdd();
5176 if (stubAddr == NULL) {
5177 return false; // Intrinsic's stub is not implemented on this platform
5178 }
5179 const char* stubName = "mulAdd";
5180
5181 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5182
5183 Node* out = argument(0);
5184 Node* in = argument(1);
5185 Node* offset = argument(2);
5186 Node* len = argument(3);
5187 Node* k = argument(4);
5188
5189 const Type* out_type = out->Value(&_gvn);
5190 const Type* in_type = in->Value(&_gvn);
5191 const TypeAryPtr* top_out = out_type->isa_aryptr();
5192 const TypeAryPtr* top_in = in_type->isa_aryptr();
5193 if (top_out == NULL || top_out->klass() == NULL ||
5194 top_in == NULL || top_in->klass() == NULL) {
5195 // failed array check
5196 return false;
5197 }
5198
5199 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5200 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5201 if (out_elem != T_INT || in_elem != T_INT) {
5202 return false;
5203 }
5204
5205 Node* outlen = load_array_length(out);
5206 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5207 Node* out_start = array_element_address(out, intcon(0), out_elem);
5208 Node* in_start = array_element_address(in, intcon(0), in_elem);
5209
5210 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5211 OptoRuntime::mulAdd_Type(),
5212 stubAddr, stubName, TypePtr::BOTTOM,
5213 out_start,in_start, new_offset, len, k);
5214 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5215 set_result(result);
5216 return true;
5217 }
5218
5219 //-------------inline_montgomeryMultiply-----------------------------------
5220 bool LibraryCallKit::inline_montgomeryMultiply() {
5221 address stubAddr = StubRoutines::montgomeryMultiply();
5222 if (stubAddr == NULL) {
5223 return false; // Intrinsic's stub is not implemented on this platform
5224 }
5225
5226 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5227 const char* stubName = "montgomery_square";
5228
5229 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5230
5231 Node* a = argument(0);
5232 Node* b = argument(1);
5233 Node* n = argument(2);
5234 Node* len = argument(3);
5235 Node* inv = argument(4);
5236 Node* m = argument(6);
5237
5238 const Type* a_type = a->Value(&_gvn);
5239 const TypeAryPtr* top_a = a_type->isa_aryptr();
5240 const Type* b_type = b->Value(&_gvn);
5241 const TypeAryPtr* top_b = b_type->isa_aryptr();
5242 const Type* n_type = a->Value(&_gvn);
5243 const TypeAryPtr* top_n = n_type->isa_aryptr();
5244 const Type* m_type = a->Value(&_gvn);
5245 const TypeAryPtr* top_m = m_type->isa_aryptr();
5246 if (top_a == NULL || top_a->klass() == NULL ||
5247 top_b == NULL || top_b->klass() == NULL ||
5248 top_n == NULL || top_n->klass() == NULL ||
5249 top_m == NULL || top_m->klass() == NULL) {
5250 // failed array check
5251 return false;
5252 }
5253
5254 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5255 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5256 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5257 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5258 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5259 return false;
5260 }
5261
5262 // Make the call
5263 {
5264 Node* a_start = array_element_address(a, intcon(0), a_elem);
5265 Node* b_start = array_element_address(b, intcon(0), b_elem);
5266 Node* n_start = array_element_address(n, intcon(0), n_elem);
5267 Node* m_start = array_element_address(m, intcon(0), m_elem);
5268
5269 Node* call = make_runtime_call(RC_LEAF,
5270 OptoRuntime::montgomeryMultiply_Type(),
5271 stubAddr, stubName, TypePtr::BOTTOM,
5272 a_start, b_start, n_start, len, inv, top(),
5273 m_start);
5274 set_result(m);
5275 }
5276
5277 return true;
5278 }
5279
5280 bool LibraryCallKit::inline_montgomerySquare() {
5281 address stubAddr = StubRoutines::montgomerySquare();
5282 if (stubAddr == NULL) {
5283 return false; // Intrinsic's stub is not implemented on this platform
5284 }
5285
5286 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5287 const char* stubName = "montgomery_square";
5288
5289 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5290
5291 Node* a = argument(0);
5292 Node* n = argument(1);
5293 Node* len = argument(2);
5294 Node* inv = argument(3);
5295 Node* m = argument(5);
5296
5297 const Type* a_type = a->Value(&_gvn);
5298 const TypeAryPtr* top_a = a_type->isa_aryptr();
5299 const Type* n_type = a->Value(&_gvn);
5300 const TypeAryPtr* top_n = n_type->isa_aryptr();
5301 const Type* m_type = a->Value(&_gvn);
5302 const TypeAryPtr* top_m = m_type->isa_aryptr();
5303 if (top_a == NULL || top_a->klass() == NULL ||
5304 top_n == NULL || top_n->klass() == NULL ||
5305 top_m == NULL || top_m->klass() == NULL) {
5306 // failed array check
5307 return false;
5308 }
5309
5310 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5311 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5312 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5313 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5314 return false;
5315 }
5316
5317 // Make the call
5318 {
5319 Node* a_start = array_element_address(a, intcon(0), a_elem);
5320 Node* n_start = array_element_address(n, intcon(0), n_elem);
5321 Node* m_start = array_element_address(m, intcon(0), m_elem);
5322
5323 Node* call = make_runtime_call(RC_LEAF,
5324 OptoRuntime::montgomerySquare_Type(),
5325 stubAddr, stubName, TypePtr::BOTTOM,
5326 a_start, n_start, len, inv, top(),
5327 m_start);
5328 set_result(m);
5329 }
5330
5331 return true;
5332 }
5333
5334
5335 /**
5336 * Calculate CRC32 for byte.
5337 * int java.util.zip.CRC32.update(int crc, int b)
5338 */
5339 bool LibraryCallKit::inline_updateCRC32() {
5340 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5341 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5342 // no receiver since it is static method
5343 Node* crc = argument(0); // type: int
5344 Node* b = argument(1); // type: int
5345
5346 /*
5347 * int c = ~ crc;
5348 * b = timesXtoThe32[(b ^ c) & 0xFF];
5349 * b = b ^ (c >>> 8);
5350 * crc = ~b;
5351 */
5352
5353 Node* M1 = intcon(-1);
5354 crc = _gvn.transform(new XorINode(crc, M1));
5355 Node* result = _gvn.transform(new XorINode(crc, b));
5356 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5357
5358 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5359 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5360 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5361 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5362
5363 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5364 result = _gvn.transform(new XorINode(crc, result));
5365 result = _gvn.transform(new XorINode(result, M1));
5366 set_result(result);
5367 return true;
5368 }
5369
5370 /**
5371 * Calculate CRC32 for byte[] array.
5372 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5373 */
5374 bool LibraryCallKit::inline_updateBytesCRC32() {
5375 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5376 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5377 // no receiver since it is static method
5378 Node* crc = argument(0); // type: int
5379 Node* src = argument(1); // type: oop
5380 Node* offset = argument(2); // type: int
5381 Node* length = argument(3); // type: int
5382
5383 const Type* src_type = src->Value(&_gvn);
5384 const TypeAryPtr* top_src = src_type->isa_aryptr();
5385 if (top_src == NULL || top_src->klass() == NULL) {
5386 // failed array check
5387 return false;
5388 }
5389
5390 // Figure out the size and type of the elements we will be copying.
5391 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5392 if (src_elem != T_BYTE) {
5393 return false;
5394 }
5395
5396 // 'src_start' points to src array + scaled offset
5397 Node* src_start = array_element_address(src, offset, src_elem);
5398
5399 // We assume that range check is done by caller.
5400 // TODO: generate range check (offset+length < src.length) in debug VM.
5401
5402 // Call the stub.
5403 address stubAddr = StubRoutines::updateBytesCRC32();
5404 const char *stubName = "updateBytesCRC32";
5405
5406 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5407 stubAddr, stubName, TypePtr::BOTTOM,
5408 crc, src_start, length);
5409 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5410 set_result(result);
5411 return true;
5412 }
5413
5414 /**
5415 * Calculate CRC32 for ByteBuffer.
5416 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5417 */
5418 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5419 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5420 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5421 // no receiver since it is static method
5422 Node* crc = argument(0); // type: int
5423 Node* src = argument(1); // type: long
5424 Node* offset = argument(3); // type: int
5425 Node* length = argument(4); // type: int
5426
5427 src = ConvL2X(src); // adjust Java long to machine word
5428 Node* base = _gvn.transform(new CastX2PNode(src));
5429 offset = ConvI2X(offset);
5430
5431 // 'src_start' points to src array + scaled offset
5432 Node* src_start = basic_plus_adr(top(), base, offset);
5433
5434 // Call the stub.
5435 address stubAddr = StubRoutines::updateBytesCRC32();
5436 const char *stubName = "updateBytesCRC32";
5437
5438 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5439 stubAddr, stubName, TypePtr::BOTTOM,
5440 crc, src_start, length);
5441 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5442 set_result(result);
5443 return true;
5444 }
5445
5446 //------------------------------get_table_from_crc32c_class-----------------------
5447 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5448 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5449 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5450
5451 return table;
5452 }
5453
5454 //------------------------------inline_updateBytesCRC32C-----------------------
5455 //
5456 // Calculate CRC32C for byte[] array.
5457 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5458 //
5459 bool LibraryCallKit::inline_updateBytesCRC32C() {
5460 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5461 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5462 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5463 // no receiver since it is a static method
5464 Node* crc = argument(0); // type: int
5465 Node* src = argument(1); // type: oop
5466 Node* offset = argument(2); // type: int
5467 Node* end = argument(3); // type: int
5468
5469 Node* length = _gvn.transform(new SubINode(end, offset));
5470
5471 const Type* src_type = src->Value(&_gvn);
5472 const TypeAryPtr* top_src = src_type->isa_aryptr();
5473 if (top_src == NULL || top_src->klass() == NULL) {
5474 // failed array check
5475 return false;
5476 }
5477
5478 // Figure out the size and type of the elements we will be copying.
5479 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5480 if (src_elem != T_BYTE) {
5481 return false;
5482 }
5483
5484 // 'src_start' points to src array + scaled offset
5485 Node* src_start = array_element_address(src, offset, src_elem);
5486
5487 // static final int[] byteTable in class CRC32C
5488 Node* table = get_table_from_crc32c_class(callee()->holder());
5489 Node* table_start = array_element_address(table, intcon(0), T_INT);
5490
5491 // We assume that range check is done by caller.
5492 // TODO: generate range check (offset+length < src.length) in debug VM.
5493
5494 // Call the stub.
5495 address stubAddr = StubRoutines::updateBytesCRC32C();
5496 const char *stubName = "updateBytesCRC32C";
5497
5498 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5499 stubAddr, stubName, TypePtr::BOTTOM,
5500 crc, src_start, length, table_start);
5501 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5502 set_result(result);
5503 return true;
5504 }
5505
5506 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5507 //
5508 // Calculate CRC32C for DirectByteBuffer.
5509 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5510 //
5511 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5512 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5513 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5514 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5515 // no receiver since it is a static method
5516 Node* crc = argument(0); // type: int
5517 Node* src = argument(1); // type: long
5518 Node* offset = argument(3); // type: int
5519 Node* end = argument(4); // type: int
5520
5521 Node* length = _gvn.transform(new SubINode(end, offset));
5522
5523 src = ConvL2X(src); // adjust Java long to machine word
5524 Node* base = _gvn.transform(new CastX2PNode(src));
5525 offset = ConvI2X(offset);
5526
5527 // 'src_start' points to src array + scaled offset
5528 Node* src_start = basic_plus_adr(top(), base, offset);
5529
5530 // static final int[] byteTable in class CRC32C
5531 Node* table = get_table_from_crc32c_class(callee()->holder());
5532 Node* table_start = array_element_address(table, intcon(0), T_INT);
5533
5534 // Call the stub.
5535 address stubAddr = StubRoutines::updateBytesCRC32C();
5536 const char *stubName = "updateBytesCRC32C";
5537
5538 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5539 stubAddr, stubName, TypePtr::BOTTOM,
5540 crc, src_start, length, table_start);
5541 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5542 set_result(result);
5543 return true;
5544 }
5545
5546 //------------------------------inline_updateBytesAdler32----------------------
5547 //
5548 // Calculate Adler32 checksum for byte[] array.
5549 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5550 //
5551 bool LibraryCallKit::inline_updateBytesAdler32() {
5552 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5553 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5554 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5555 // no receiver since it is static method
5556 Node* crc = argument(0); // type: int
5557 Node* src = argument(1); // type: oop
5558 Node* offset = argument(2); // type: int
5559 Node* length = argument(3); // type: int
5560
5561 const Type* src_type = src->Value(&_gvn);
5562 const TypeAryPtr* top_src = src_type->isa_aryptr();
5563 if (top_src == NULL || top_src->klass() == NULL) {
5564 // failed array check
5565 return false;
5566 }
5567
5568 // Figure out the size and type of the elements we will be copying.
5569 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5570 if (src_elem != T_BYTE) {
5571 return false;
5572 }
5573
5574 // 'src_start' points to src array + scaled offset
5575 Node* src_start = array_element_address(src, offset, src_elem);
5576
5577 // We assume that range check is done by caller.
5578 // TODO: generate range check (offset+length < src.length) in debug VM.
5579
5580 // Call the stub.
5581 address stubAddr = StubRoutines::updateBytesAdler32();
5582 const char *stubName = "updateBytesAdler32";
5583
5584 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5585 stubAddr, stubName, TypePtr::BOTTOM,
5586 crc, src_start, length);
5587 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5588 set_result(result);
5589 return true;
5590 }
5591
5592 //------------------------------inline_updateByteBufferAdler32---------------
5593 //
5594 // Calculate Adler32 checksum for DirectByteBuffer.
5595 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5596 //
5597 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5598 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5599 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5600 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5601 // no receiver since it is static method
5602 Node* crc = argument(0); // type: int
5603 Node* src = argument(1); // type: long
5604 Node* offset = argument(3); // type: int
5605 Node* length = argument(4); // type: int
5606
5607 src = ConvL2X(src); // adjust Java long to machine word
5608 Node* base = _gvn.transform(new CastX2PNode(src));
5609 offset = ConvI2X(offset);
5610
5611 // 'src_start' points to src array + scaled offset
5612 Node* src_start = basic_plus_adr(top(), base, offset);
5613
5614 // Call the stub.
5615 address stubAddr = StubRoutines::updateBytesAdler32();
5616 const char *stubName = "updateBytesAdler32";
5617
5618 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5619 stubAddr, stubName, TypePtr::BOTTOM,
5620 crc, src_start, length);
5621
5622 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5623 set_result(result);
5624 return true;
5625 }
5626
5627 //----------------------------inline_reference_get----------------------------
5628 // public T java.lang.ref.Reference.get();
5629 bool LibraryCallKit::inline_reference_get() {
5630 const int referent_offset = java_lang_ref_Reference::referent_offset;
5631 guarantee(referent_offset > 0, "should have already been set");
5632
5633 // Get the argument:
5634 Node* reference_obj = null_check_receiver();
5635 if (stopped()) return true;
5636
5637 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5638
5639 ciInstanceKlass* klass = env()->Object_klass();
5640 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5641
5642 Node* no_ctrl = NULL;
5643 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5644
5645 // Use the pre-barrier to record the value in the referent field
5646 pre_barrier(false /* do_load */,
5647 control(),
5648 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5649 result /* pre_val */,
5650 T_OBJECT);
5651
5652 // Add memory barrier to prevent commoning reads from this field
5653 // across safepoint since GC can change its value.
5654 insert_mem_bar(Op_MemBarCPUOrder);
5655
5656 set_result(result);
5657 return true;
5658 }
5659
5660
5661 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5662 bool is_exact=true, bool is_static=false,
5663 ciInstanceKlass * fromKls=NULL) {
5664 if (fromKls == NULL) {
5665 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5666 assert(tinst != NULL, "obj is null");
5667 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5668 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5669 fromKls = tinst->klass()->as_instance_klass();
5670 } else {
5671 assert(is_static, "only for static field access");
5672 }
5673 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5674 ciSymbol::make(fieldTypeString),
5675 is_static);
5676
5677 assert (field != NULL, "undefined field");
5678 if (field == NULL) return (Node *) NULL;
5679
5680 if (is_static) {
5681 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5682 fromObj = makecon(tip);
5683 }
5684
5685 // Next code copied from Parse::do_get_xxx():
5686
5687 // Compute address and memory type.
5688 int offset = field->offset_in_bytes();
5689 bool is_vol = field->is_volatile();
5690 ciType* field_klass = field->type();
5691 assert(field_klass->is_loaded(), "should be loaded");
5692 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5693 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5694 BasicType bt = field->layout_type();
5695
5696 // Build the resultant type of the load
5697 const Type *type;
5698 if (bt == T_OBJECT) {
5699 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5700 } else {
5701 type = Type::get_const_basic_type(bt);
5702 }
5703
5704 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5705 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5706 }
5707 // Build the load.
5708 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5709 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5710 // If reference is volatile, prevent following memory ops from
5711 // floating up past the volatile read. Also prevents commoning
5712 // another volatile read.
5713 if (is_vol) {
5714 // Memory barrier includes bogus read of value to force load BEFORE membar
5715 insert_mem_bar(Op_MemBarAcquire, loadedField);
5716 }
5717 return loadedField;
5718 }
5719
5720
5721 //------------------------------inline_aescrypt_Block-----------------------
5722 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5723 address stubAddr;
5724 const char *stubName;
5725 assert(UseAES, "need AES instruction support");
5726
5727 switch(id) {
5728 case vmIntrinsics::_aescrypt_encryptBlock:
5729 stubAddr = StubRoutines::aescrypt_encryptBlock();
5730 stubName = "aescrypt_encryptBlock";
5731 break;
5732 case vmIntrinsics::_aescrypt_decryptBlock:
5733 stubAddr = StubRoutines::aescrypt_decryptBlock();
5734 stubName = "aescrypt_decryptBlock";
5735 break;
5736 }
5737 if (stubAddr == NULL) return false;
5738
5739 Node* aescrypt_object = argument(0);
5740 Node* src = argument(1);
5741 Node* src_offset = argument(2);
5742 Node* dest = argument(3);
5743 Node* dest_offset = argument(4);
5744
5745 // (1) src and dest are arrays.
5746 const Type* src_type = src->Value(&_gvn);
5747 const Type* dest_type = dest->Value(&_gvn);
5748 const TypeAryPtr* top_src = src_type->isa_aryptr();
5749 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5750 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5751
5752 // for the quick and dirty code we will skip all the checks.
5753 // we are just trying to get the call to be generated.
5754 Node* src_start = src;
5755 Node* dest_start = dest;
5756 if (src_offset != NULL || dest_offset != NULL) {
5757 assert(src_offset != NULL && dest_offset != NULL, "");
5758 src_start = array_element_address(src, src_offset, T_BYTE);
5759 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5760 }
5761
5762 // now need to get the start of its expanded key array
5763 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5764 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5765 if (k_start == NULL) return false;
5766
5767 if (Matcher::pass_original_key_for_aes()) {
5768 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5769 // compatibility issues between Java key expansion and SPARC crypto instructions
5770 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5771 if (original_k_start == NULL) return false;
5772
5773 // Call the stub.
5774 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5775 stubAddr, stubName, TypePtr::BOTTOM,
5776 src_start, dest_start, k_start, original_k_start);
5777 } else {
5778 // Call the stub.
5779 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5780 stubAddr, stubName, TypePtr::BOTTOM,
5781 src_start, dest_start, k_start);
5782 }
5783
5784 return true;
5785 }
5786
5787 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5788 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5789 address stubAddr;
5790 const char *stubName;
5791
5792 assert(UseAES, "need AES instruction support");
5793
5794 switch(id) {
5795 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5796 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5797 stubName = "cipherBlockChaining_encryptAESCrypt";
5798 break;
5799 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5800 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5801 stubName = "cipherBlockChaining_decryptAESCrypt";
5802 break;
5803 }
5804 if (stubAddr == NULL) return false;
5805
5806 Node* cipherBlockChaining_object = argument(0);
5807 Node* src = argument(1);
5808 Node* src_offset = argument(2);
5809 Node* len = argument(3);
5810 Node* dest = argument(4);
5811 Node* dest_offset = argument(5);
5812
5813 // (1) src and dest are arrays.
5814 const Type* src_type = src->Value(&_gvn);
5815 const Type* dest_type = dest->Value(&_gvn);
5816 const TypeAryPtr* top_src = src_type->isa_aryptr();
5817 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5818 assert (top_src != NULL && top_src->klass() != NULL
5819 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5820
5821 // checks are the responsibility of the caller
5822 Node* src_start = src;
5823 Node* dest_start = dest;
5824 if (src_offset != NULL || dest_offset != NULL) {
5825 assert(src_offset != NULL && dest_offset != NULL, "");
5826 src_start = array_element_address(src, src_offset, T_BYTE);
5827 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5828 }
5829
5830 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5831 // (because of the predicated logic executed earlier).
5832 // so we cast it here safely.
5833 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5834
5835 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5836 if (embeddedCipherObj == NULL) return false;
5837
5838 // cast it to what we know it will be at runtime
5839 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5840 assert(tinst != NULL, "CBC obj is null");
5841 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5842 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5843 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5844
5845 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5846 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5847 const TypeOopPtr* xtype = aklass->as_instance_type();
5848 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5849 aescrypt_object = _gvn.transform(aescrypt_object);
5850
5851 // we need to get the start of the aescrypt_object's expanded key array
5852 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5853 if (k_start == NULL) return false;
5854
5855 // similarly, get the start address of the r vector
5856 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5857 if (objRvec == NULL) return false;
5858 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5859
5860 Node* cbcCrypt;
5861 if (Matcher::pass_original_key_for_aes()) {
5862 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5863 // compatibility issues between Java key expansion and SPARC crypto instructions
5864 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5865 if (original_k_start == NULL) return false;
5866
5867 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5868 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5869 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5870 stubAddr, stubName, TypePtr::BOTTOM,
5871 src_start, dest_start, k_start, r_start, len, original_k_start);
5872 } else {
5873 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5874 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5875 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5876 stubAddr, stubName, TypePtr::BOTTOM,
5877 src_start, dest_start, k_start, r_start, len);
5878 }
5879
5880 // return cipher length (int)
5881 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5882 set_result(retvalue);
5883 return true;
5884 }
5885
5886 //------------------------------get_key_start_from_aescrypt_object-----------------------
5887 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5888 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5889 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5890 if (objAESCryptKey == NULL) return (Node *) NULL;
5891
5892 // now have the array, need to get the start address of the K array
5893 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5894 return k_start;
5895 }
5896
5897 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5898 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5899 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5900 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5901 if (objAESCryptKey == NULL) return (Node *) NULL;
5902
5903 // now have the array, need to get the start address of the lastKey array
5904 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5905 return original_k_start;
5906 }
5907
5908 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5909 // Return node representing slow path of predicate check.
5910 // the pseudo code we want to emulate with this predicate is:
5911 // for encryption:
5912 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5913 // for decryption:
5914 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5915 // note cipher==plain is more conservative than the original java code but that's OK
5916 //
5917 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5918 // The receiver was checked for NULL already.
5919 Node* objCBC = argument(0);
5920
5921 // Load embeddedCipher field of CipherBlockChaining object.
5922 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5923
5924 // get AESCrypt klass for instanceOf check
5925 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5926 // will have same classloader as CipherBlockChaining object
5927 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5928 assert(tinst != NULL, "CBCobj is null");
5929 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5930
5931 // we want to do an instanceof comparison against the AESCrypt class
5932 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5933 if (!klass_AESCrypt->is_loaded()) {
5934 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5935 Node* ctrl = control();
5936 set_control(top()); // no regular fast path
5937 return ctrl;
5938 }
5939 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5940
5941 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5942 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5943 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5944
5945 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5946
5947 // for encryption, we are done
5948 if (!decrypting)
5949 return instof_false; // even if it is NULL
5950
5951 // for decryption, we need to add a further check to avoid
5952 // taking the intrinsic path when cipher and plain are the same
5953 // see the original java code for why.
5954 RegionNode* region = new RegionNode(3);
5955 region->init_req(1, instof_false);
5956 Node* src = argument(1);
5957 Node* dest = argument(4);
5958 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5959 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5960 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5961 region->init_req(2, src_dest_conjoint);
5962
5963 record_for_igvn(region);
5964 return _gvn.transform(region);
5965 }
5966
5967 //------------------------------inline_ghash_processBlocks
5968 bool LibraryCallKit::inline_ghash_processBlocks() {
5969 address stubAddr;
5970 const char *stubName;
5971 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5972
5973 stubAddr = StubRoutines::ghash_processBlocks();
5974 stubName = "ghash_processBlocks";
5975
5976 Node* data = argument(0);
5977 Node* offset = argument(1);
5978 Node* len = argument(2);
5979 Node* state = argument(3);
5980 Node* subkeyH = argument(4);
5981
5982 Node* state_start = array_element_address(state, intcon(0), T_LONG);
5983 assert(state_start, "state is NULL");
5984 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
5985 assert(subkeyH_start, "subkeyH is NULL");
5986 Node* data_start = array_element_address(data, offset, T_BYTE);
5987 assert(data_start, "data is NULL");
5988
5989 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
5990 OptoRuntime::ghash_processBlocks_Type(),
5991 stubAddr, stubName, TypePtr::BOTTOM,
5992 state_start, subkeyH_start, data_start, len);
5993 return true;
5994 }
5995
5996 //------------------------------inline_sha_implCompress-----------------------
5997 //
5998 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5999 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6000 //
6001 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6002 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6003 //
6004 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6005 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6006 //
6007 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6008 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6009
6010 Node* sha_obj = argument(0);
6011 Node* src = argument(1); // type oop
6012 Node* ofs = argument(2); // type int
6013
6014 const Type* src_type = src->Value(&_gvn);
6015 const TypeAryPtr* top_src = src_type->isa_aryptr();
6016 if (top_src == NULL || top_src->klass() == NULL) {
6017 // failed array check
6018 return false;
6019 }
6020 // Figure out the size and type of the elements we will be copying.
6021 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6022 if (src_elem != T_BYTE) {
6023 return false;
6024 }
6025 // 'src_start' points to src array + offset
6026 Node* src_start = array_element_address(src, ofs, src_elem);
6027 Node* state = NULL;
6028 address stubAddr;
6029 const char *stubName;
6030
6031 switch(id) {
6032 case vmIntrinsics::_sha_implCompress:
6033 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6034 state = get_state_from_sha_object(sha_obj);
6035 stubAddr = StubRoutines::sha1_implCompress();
6036 stubName = "sha1_implCompress";
6037 break;
6038 case vmIntrinsics::_sha2_implCompress:
6039 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6040 state = get_state_from_sha_object(sha_obj);
6041 stubAddr = StubRoutines::sha256_implCompress();
6042 stubName = "sha256_implCompress";
6043 break;
6044 case vmIntrinsics::_sha5_implCompress:
6045 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6046 state = get_state_from_sha5_object(sha_obj);
6047 stubAddr = StubRoutines::sha512_implCompress();
6048 stubName = "sha512_implCompress";
6049 break;
6050 default:
6051 fatal_unexpected_iid(id);
6052 return false;
6053 }
6054 if (state == NULL) return false;
6055
6056 // Call the stub.
6057 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6058 stubAddr, stubName, TypePtr::BOTTOM,
6059 src_start, state);
6060
6061 return true;
6062 }
6063
6064 //------------------------------inline_digestBase_implCompressMB-----------------------
6065 //
6066 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6067 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6068 //
6069 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6070 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6071 "need SHA1/SHA256/SHA512 instruction support");
6072 assert((uint)predicate < 3, "sanity");
6073 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6074
6075 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6076 Node* src = argument(1); // byte[] array
6077 Node* ofs = argument(2); // type int
6078 Node* limit = argument(3); // type int
6079
6080 const Type* src_type = src->Value(&_gvn);
6081 const TypeAryPtr* top_src = src_type->isa_aryptr();
6082 if (top_src == NULL || top_src->klass() == NULL) {
6083 // failed array check
6084 return false;
6085 }
6086 // Figure out the size and type of the elements we will be copying.
6087 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6088 if (src_elem != T_BYTE) {
6089 return false;
6090 }
6091 // 'src_start' points to src array + offset
6092 Node* src_start = array_element_address(src, ofs, src_elem);
6093
6094 const char* klass_SHA_name = NULL;
6095 const char* stub_name = NULL;
6096 address stub_addr = NULL;
6097 bool long_state = false;
6098
6099 switch (predicate) {
6100 case 0:
6101 if (UseSHA1Intrinsics) {
6102 klass_SHA_name = "sun/security/provider/SHA";
6103 stub_name = "sha1_implCompressMB";
6104 stub_addr = StubRoutines::sha1_implCompressMB();
6105 }
6106 break;
6107 case 1:
6108 if (UseSHA256Intrinsics) {
6109 klass_SHA_name = "sun/security/provider/SHA2";
6110 stub_name = "sha256_implCompressMB";
6111 stub_addr = StubRoutines::sha256_implCompressMB();
6112 }
6113 break;
6114 case 2:
6115 if (UseSHA512Intrinsics) {
6116 klass_SHA_name = "sun/security/provider/SHA5";
6117 stub_name = "sha512_implCompressMB";
6118 stub_addr = StubRoutines::sha512_implCompressMB();
6119 long_state = true;
6120 }
6121 break;
6122 default:
6123 fatal("unknown SHA intrinsic predicate: %d", predicate);
6124 }
6125 if (klass_SHA_name != NULL) {
6126 // get DigestBase klass to lookup for SHA klass
6127 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6128 assert(tinst != NULL, "digestBase_obj is not instance???");
6129 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6130
6131 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6132 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6133 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6134 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6135 }
6136 return false;
6137 }
6138 //------------------------------inline_sha_implCompressMB-----------------------
6139 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6140 bool long_state, address stubAddr, const char *stubName,
6141 Node* src_start, Node* ofs, Node* limit) {
6142 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6143 const TypeOopPtr* xtype = aklass->as_instance_type();
6144 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6145 sha_obj = _gvn.transform(sha_obj);
6146
6147 Node* state;
6148 if (long_state) {
6149 state = get_state_from_sha5_object(sha_obj);
6150 } else {
6151 state = get_state_from_sha_object(sha_obj);
6152 }
6153 if (state == NULL) return false;
6154
6155 // Call the stub.
6156 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6157 OptoRuntime::digestBase_implCompressMB_Type(),
6158 stubAddr, stubName, TypePtr::BOTTOM,
6159 src_start, state, ofs, limit);
6160 // return ofs (int)
6161 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6162 set_result(result);
6163
6164 return true;
6165 }
6166
6167 //------------------------------get_state_from_sha_object-----------------------
6168 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6169 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6170 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6171 if (sha_state == NULL) return (Node *) NULL;
6172
6173 // now have the array, need to get the start address of the state array
6174 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6175 return state;
6176 }
6177
6178 //------------------------------get_state_from_sha5_object-----------------------
6179 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6180 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6181 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6182 if (sha_state == NULL) return (Node *) NULL;
6183
6184 // now have the array, need to get the start address of the state array
6185 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6186 return state;
6187 }
6188
6189 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6190 // Return node representing slow path of predicate check.
6191 // the pseudo code we want to emulate with this predicate is:
6192 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6193 //
6194 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6195 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6196 "need SHA1/SHA256/SHA512 instruction support");
6197 assert((uint)predicate < 3, "sanity");
6198
6199 // The receiver was checked for NULL already.
6200 Node* digestBaseObj = argument(0);
6201
6202 // get DigestBase klass for instanceOf check
6203 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6204 assert(tinst != NULL, "digestBaseObj is null");
6205 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6206
6207 const char* klass_SHA_name = NULL;
6208 switch (predicate) {
6209 case 0:
6210 if (UseSHA1Intrinsics) {
6211 // we want to do an instanceof comparison against the SHA class
6212 klass_SHA_name = "sun/security/provider/SHA";
6213 }
6214 break;
6215 case 1:
6216 if (UseSHA256Intrinsics) {
6217 // we want to do an instanceof comparison against the SHA2 class
6218 klass_SHA_name = "sun/security/provider/SHA2";
6219 }
6220 break;
6221 case 2:
6222 if (UseSHA512Intrinsics) {
6223 // we want to do an instanceof comparison against the SHA5 class
6224 klass_SHA_name = "sun/security/provider/SHA5";
6225 }
6226 break;
6227 default:
6228 fatal("unknown SHA intrinsic predicate: %d", predicate);
6229 }
6230
6231 ciKlass* klass_SHA = NULL;
6232 if (klass_SHA_name != NULL) {
6233 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6234 }
6235 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6236 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6237 Node* ctrl = control();
6238 set_control(top()); // no intrinsic path
6239 return ctrl;
6240 }
6241 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6242
6243 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6244 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6245 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6246 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6247
6248 return instof_false; // even if it is NULL
6249 }
6250
6251 bool LibraryCallKit::inline_profileBoolean() {
6252 Node* counts = argument(1);
6253 const TypeAryPtr* ary = NULL;
6254 ciArray* aobj = NULL;
6255 if (counts->is_Con()
6256 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6257 && (aobj = ary->const_oop()->as_array()) != NULL
6258 && (aobj->length() == 2)) {
6259 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6260 jint false_cnt = aobj->element_value(0).as_int();
6261 jint true_cnt = aobj->element_value(1).as_int();
6262
6263 if (C->log() != NULL) {
6264 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6265 false_cnt, true_cnt);
6266 }
6267
6268 if (false_cnt + true_cnt == 0) {
6269 // According to profile, never executed.
6270 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6271 Deoptimization::Action_reinterpret);
6272 return true;
6273 }
6274
6275 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6276 // is a number of each value occurrences.
6277 Node* result = argument(0);
6278 if (false_cnt == 0 || true_cnt == 0) {
6279 // According to profile, one value has been never seen.
6280 int expected_val = (false_cnt == 0) ? 1 : 0;
6281
6282 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6283 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6284
6285 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6286 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6287 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6288
6289 { // Slow path: uncommon trap for never seen value and then reexecute
6290 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6291 // the value has been seen at least once.
6292 PreserveJVMState pjvms(this);
6293 PreserveReexecuteState preexecs(this);
6294 jvms()->set_should_reexecute(true);
6295
6296 set_control(slow_path);
6297 set_i_o(i_o());
6298
6299 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6300 Deoptimization::Action_reinterpret);
6301 }
6302 // The guard for never seen value enables sharpening of the result and
6303 // returning a constant. It allows to eliminate branches on the same value
6304 // later on.
6305 set_control(fast_path);
6306 result = intcon(expected_val);
6307 }
6308 // Stop profiling.
6309 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6310 // By replacing method body with profile data (represented as ProfileBooleanNode
6311 // on IR level) we effectively disable profiling.
6312 // It enables full speed execution once optimized code is generated.
6313 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6314 C->record_for_igvn(profile);
6315 set_result(profile);
6316 return true;
6317 } else {
6318 // Continue profiling.
6319 // Profile data isn't available at the moment. So, execute method's bytecode version.
6320 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6321 // is compiled and counters aren't available since corresponding MethodHandle
6322 // isn't a compile-time constant.
6323 return false;
6324 }
6325 }
6326
6327 bool LibraryCallKit::inline_isCompileConstant() {
6328 Node* n = argument(0);
6329 set_result(n->is_Con() ? intcon(1) : intcon(0));
6330 return true;
6331 }