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 bool inline_deoptimize();
310 };
311
312 //---------------------------make_vm_intrinsic----------------------------
313 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
314 vmIntrinsics::ID id = m->intrinsic_id();
315 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
316
317 if (!m->is_loaded()) {
318 // Do not attempt to inline unloaded methods.
319 return NULL;
320 }
321
322 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
323 bool is_available = false;
324
325 {
326 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
327 // the compiler must transition to '_thread_in_vm' state because both
328 // methods access VM-internal data.
329 VM_ENTRY_MARK;
330 methodHandle mh(THREAD, m->get_Method());
331 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
332 !C->directive()->is_intrinsic_disabled(mh) &&
333 !vmIntrinsics::is_disabled_by_flags(mh);
334
335 }
336
337 if (is_available) {
338 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
339 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
340 return new LibraryIntrinsic(m, is_virtual,
341 vmIntrinsics::predicates_needed(id),
342 vmIntrinsics::does_virtual_dispatch(id),
343 (vmIntrinsics::ID) id);
344 } else {
345 return NULL;
346 }
347 }
348
349 //----------------------register_library_intrinsics-----------------------
350 // Initialize this file's data structures, for each Compile instance.
351 void Compile::register_library_intrinsics() {
352 // Nothing to do here.
353 }
354
355 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
356 LibraryCallKit kit(jvms, this);
357 Compile* C = kit.C;
358 int nodes = C->unique();
359 #ifndef PRODUCT
360 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
361 char buf[1000];
362 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
363 tty->print_cr("Intrinsic %s", str);
364 }
365 #endif
366 ciMethod* callee = kit.callee();
367 const int bci = kit.bci();
368
369 // Try to inline the intrinsic.
370 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
371 kit.try_to_inline(_last_predicate)) {
372 if (C->print_intrinsics() || C->print_inlining()) {
373 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
374 }
375 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
376 if (C->log()) {
377 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
378 vmIntrinsics::name_at(intrinsic_id()),
379 (is_virtual() ? " virtual='1'" : ""),
380 C->unique() - nodes);
381 }
382 // Push the result from the inlined method onto the stack.
383 kit.push_result();
384 C->print_inlining_update(this);
385 return kit.transfer_exceptions_into_jvms();
386 }
387
388 // The intrinsic bailed out
389 if (C->print_intrinsics() || C->print_inlining()) {
390 if (jvms->has_method()) {
391 // Not a root compile.
392 const char* msg;
393 if (callee->intrinsic_candidate()) {
394 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
395 } else {
396 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
397 : "failed to inline (intrinsic), method not annotated";
398 }
399 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
400 } else {
401 // Root compile
402 tty->print("Did not generate intrinsic %s%s at bci:%d in",
403 vmIntrinsics::name_at(intrinsic_id()),
404 (is_virtual() ? " (virtual)" : ""), bci);
405 }
406 }
407 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
408 C->print_inlining_update(this);
409 return NULL;
410 }
411
412 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
413 LibraryCallKit kit(jvms, this);
414 Compile* C = kit.C;
415 int nodes = C->unique();
416 _last_predicate = predicate;
417 #ifndef PRODUCT
418 assert(is_predicated() && predicate < predicates_count(), "sanity");
419 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
420 char buf[1000];
421 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
422 tty->print_cr("Predicate for intrinsic %s", str);
423 }
424 #endif
425 ciMethod* callee = kit.callee();
426 const int bci = kit.bci();
427
428 Node* slow_ctl = kit.try_to_predicate(predicate);
429 if (!kit.failing()) {
430 if (C->print_intrinsics() || C->print_inlining()) {
431 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
432 }
433 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
434 if (C->log()) {
435 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
436 vmIntrinsics::name_at(intrinsic_id()),
437 (is_virtual() ? " virtual='1'" : ""),
438 C->unique() - nodes);
439 }
440 return slow_ctl; // Could be NULL if the check folds.
441 }
442
443 // The intrinsic bailed out
444 if (C->print_intrinsics() || C->print_inlining()) {
445 if (jvms->has_method()) {
446 // Not a root compile.
447 const char* msg = "failed to generate predicate for intrinsic";
448 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
449 } else {
450 // Root compile
451 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
452 vmIntrinsics::name_at(intrinsic_id()),
453 (is_virtual() ? " (virtual)" : ""), bci);
454 }
455 }
456 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
457 return NULL;
458 }
459
460 bool LibraryCallKit::try_to_inline(int predicate) {
461 // Handle symbolic names for otherwise undistinguished boolean switches:
462 const bool is_store = true;
463 const bool is_native_ptr = true;
464 const bool is_static = true;
465 const bool is_volatile = true;
466
467 if (!jvms()->has_method()) {
468 // Root JVMState has a null method.
469 assert(map()->memory()->Opcode() == Op_Parm, "");
470 // Insert the memory aliasing node
471 set_all_memory(reset_memory());
472 }
473 assert(merged_memory(), "");
474
475
476 switch (intrinsic_id()) {
477 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
478 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
479 case vmIntrinsics::_getClass: return inline_native_getClass();
480
481 case vmIntrinsics::_dsin:
482 case vmIntrinsics::_dcos:
483 case vmIntrinsics::_dtan:
484 case vmIntrinsics::_dabs:
485 case vmIntrinsics::_datan2:
486 case vmIntrinsics::_dsqrt:
487 case vmIntrinsics::_dexp:
488 case vmIntrinsics::_dlog:
489 case vmIntrinsics::_dlog10:
490 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
491
492 case vmIntrinsics::_min:
493 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
494
495 case vmIntrinsics::_notify:
496 case vmIntrinsics::_notifyAll:
497 if (InlineNotify) {
498 return inline_notify(intrinsic_id());
499 }
500 return false;
501
502 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
503 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
504 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
505 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
506 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
507 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
508 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
509 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
510 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
511 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
512 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
513 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
514
515 case vmIntrinsics::_arraycopy: return inline_arraycopy();
516
517 case vmIntrinsics::_compareTo: return inline_string_compareTo();
518 case vmIntrinsics::_indexOf: return inline_string_indexOf();
519 case vmIntrinsics::_equals: return inline_string_equals();
520
521 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
522 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
523 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
524 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
525 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
526 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
527 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
528 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
529 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
530 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
531 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
532 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
533 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
534 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
535 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
536 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
537 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
538 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
539
540 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
541 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
542 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
543 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
544 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
545 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
546 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
547 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
548
549 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
550 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
551 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
552 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
553 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
554 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
555 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
556 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
557
558 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
559 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
560 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
561 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
562 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
563 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
564 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
565 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
566 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
567
568 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
569 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
570 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
571 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
572 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
573 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
574 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
575 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
576 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
577
578 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
579 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
580 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
581 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
582
583 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
584 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
585 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
586 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
587
588 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
589 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
590 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
591
592 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
593 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
594 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
595
596 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
597 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
598 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
599 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
600 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
601
602 case vmIntrinsics::_loadFence:
603 case vmIntrinsics::_storeFence:
604 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
605
606 case vmIntrinsics::_currentThread: return inline_native_currentThread();
607 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
608
609 #ifdef TRACE_HAVE_INTRINSICS
610 case vmIntrinsics::_classID: return inline_native_classID();
611 case vmIntrinsics::_threadID: return inline_native_threadID();
612 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
613 #endif
614 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
615 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
616 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
617 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
618 case vmIntrinsics::_newArray: return inline_native_newArray();
619 case vmIntrinsics::_getLength: return inline_native_getLength();
620 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
621 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
622 case vmIntrinsics::_equalsC: return inline_array_equals();
623 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
624
625 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
626
627 case vmIntrinsics::_isInstance:
628 case vmIntrinsics::_getModifiers:
629 case vmIntrinsics::_isInterface:
630 case vmIntrinsics::_isArray:
631 case vmIntrinsics::_isPrimitive:
632 case vmIntrinsics::_getSuperclass:
633 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
634
635 case vmIntrinsics::_floatToRawIntBits:
636 case vmIntrinsics::_floatToIntBits:
637 case vmIntrinsics::_intBitsToFloat:
638 case vmIntrinsics::_doubleToRawLongBits:
639 case vmIntrinsics::_doubleToLongBits:
640 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
641
642 case vmIntrinsics::_numberOfLeadingZeros_i:
643 case vmIntrinsics::_numberOfLeadingZeros_l:
644 case vmIntrinsics::_numberOfTrailingZeros_i:
645 case vmIntrinsics::_numberOfTrailingZeros_l:
646 case vmIntrinsics::_bitCount_i:
647 case vmIntrinsics::_bitCount_l:
648 case vmIntrinsics::_reverseBytes_i:
649 case vmIntrinsics::_reverseBytes_l:
650 case vmIntrinsics::_reverseBytes_s:
651 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
652
653 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
654
655 case vmIntrinsics::_Reference_get: return inline_reference_get();
656
657 case vmIntrinsics::_Class_cast: return inline_Class_cast();
658
659 case vmIntrinsics::_aescrypt_encryptBlock:
660 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
661
662 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
663 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
664 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
665
666 case vmIntrinsics::_sha_implCompress:
667 case vmIntrinsics::_sha2_implCompress:
668 case vmIntrinsics::_sha5_implCompress:
669 return inline_sha_implCompress(intrinsic_id());
670
671 case vmIntrinsics::_digestBase_implCompressMB:
672 return inline_digestBase_implCompressMB(predicate);
673
674 case vmIntrinsics::_multiplyToLen:
675 return inline_multiplyToLen();
676
677 case vmIntrinsics::_squareToLen:
678 return inline_squareToLen();
679
680 case vmIntrinsics::_mulAdd:
681 return inline_mulAdd();
682
683 case vmIntrinsics::_montgomeryMultiply:
684 return inline_montgomeryMultiply();
685 case vmIntrinsics::_montgomerySquare:
686 return inline_montgomerySquare();
687
688 case vmIntrinsics::_ghash_processBlocks:
689 return inline_ghash_processBlocks();
690
691 case vmIntrinsics::_encodeISOArray:
692 return inline_encodeISOArray();
693
694 case vmIntrinsics::_updateCRC32:
695 return inline_updateCRC32();
696 case vmIntrinsics::_updateBytesCRC32:
697 return inline_updateBytesCRC32();
698 case vmIntrinsics::_updateByteBufferCRC32:
699 return inline_updateByteBufferCRC32();
700
701 case vmIntrinsics::_updateBytesCRC32C:
702 return inline_updateBytesCRC32C();
703 case vmIntrinsics::_updateDirectByteBufferCRC32C:
704 return inline_updateDirectByteBufferCRC32C();
705
706 case vmIntrinsics::_updateBytesAdler32:
707 return inline_updateBytesAdler32();
708 case vmIntrinsics::_updateByteBufferAdler32:
709 return inline_updateByteBufferAdler32();
710
711 case vmIntrinsics::_profileBoolean:
712 return inline_profileBoolean();
713 case vmIntrinsics::_isCompileConstant:
714 return inline_isCompileConstant();
715
716 case vmIntrinsics::_deoptimize:
717 return inline_deoptimize();
718
719 default:
720 // If you get here, it may be that someone has added a new intrinsic
721 // to the list in vmSymbols.hpp without implementing it here.
722 #ifndef PRODUCT
723 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
724 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
725 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
726 }
727 #endif
728 return false;
729 }
730 }
731
732 Node* LibraryCallKit::try_to_predicate(int predicate) {
733 if (!jvms()->has_method()) {
734 // Root JVMState has a null method.
735 assert(map()->memory()->Opcode() == Op_Parm, "");
736 // Insert the memory aliasing node
737 set_all_memory(reset_memory());
738 }
739 assert(merged_memory(), "");
740
741 switch (intrinsic_id()) {
742 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
743 return inline_cipherBlockChaining_AESCrypt_predicate(false);
744 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
745 return inline_cipherBlockChaining_AESCrypt_predicate(true);
746 case vmIntrinsics::_digestBase_implCompressMB:
747 return inline_digestBase_implCompressMB_predicate(predicate);
748
749 default:
750 // If you get here, it may be that someone has added a new intrinsic
751 // to the list in vmSymbols.hpp without implementing it here.
752 #ifndef PRODUCT
753 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
754 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
755 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
756 }
757 #endif
758 Node* slow_ctl = control();
759 set_control(top()); // No fast path instrinsic
760 return slow_ctl;
761 }
762 }
763
764 //------------------------------set_result-------------------------------
765 // Helper function for finishing intrinsics.
766 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
767 record_for_igvn(region);
768 set_control(_gvn.transform(region));
769 set_result( _gvn.transform(value));
770 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
771 }
772
773 //------------------------------generate_guard---------------------------
774 // Helper function for generating guarded fast-slow graph structures.
775 // The given 'test', if true, guards a slow path. If the test fails
776 // then a fast path can be taken. (We generally hope it fails.)
777 // In all cases, GraphKit::control() is updated to the fast path.
778 // The returned value represents the control for the slow path.
779 // The return value is never 'top'; it is either a valid control
780 // or NULL if it is obvious that the slow path can never be taken.
781 // Also, if region and the slow control are not NULL, the slow edge
782 // is appended to the region.
783 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
784 if (stopped()) {
785 // Already short circuited.
786 return NULL;
787 }
788
789 // Build an if node and its projections.
790 // If test is true we take the slow path, which we assume is uncommon.
791 if (_gvn.type(test) == TypeInt::ZERO) {
792 // The slow branch is never taken. No need to build this guard.
793 return NULL;
794 }
795
796 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
797
798 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
799 if (if_slow == top()) {
800 // The slow branch is never taken. No need to build this guard.
801 return NULL;
802 }
803
804 if (region != NULL)
805 region->add_req(if_slow);
806
807 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
808 set_control(if_fast);
809
810 return if_slow;
811 }
812
813 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
814 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
815 }
816 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
817 return generate_guard(test, region, PROB_FAIR);
818 }
819
820 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
821 Node* *pos_index) {
822 if (stopped())
823 return NULL; // already stopped
824 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
825 return NULL; // index is already adequately typed
826 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
827 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
828 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
829 if (is_neg != NULL && pos_index != NULL) {
830 // Emulate effect of Parse::adjust_map_after_if.
831 Node* ccast = new CastIINode(index, TypeInt::POS);
832 ccast->set_req(0, control());
833 (*pos_index) = _gvn.transform(ccast);
834 }
835 return is_neg;
836 }
837
838 // Make sure that 'position' is a valid limit index, in [0..length].
839 // There are two equivalent plans for checking this:
840 // A. (offset + copyLength) unsigned<= arrayLength
841 // B. offset <= (arrayLength - copyLength)
842 // We require that all of the values above, except for the sum and
843 // difference, are already known to be non-negative.
844 // Plan A is robust in the face of overflow, if offset and copyLength
845 // are both hugely positive.
846 //
847 // Plan B is less direct and intuitive, but it does not overflow at
848 // all, since the difference of two non-negatives is always
849 // representable. Whenever Java methods must perform the equivalent
850 // check they generally use Plan B instead of Plan A.
851 // For the moment we use Plan A.
852 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
853 Node* subseq_length,
854 Node* array_length,
855 RegionNode* region) {
856 if (stopped())
857 return NULL; // already stopped
858 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
859 if (zero_offset && subseq_length->eqv_uncast(array_length))
860 return NULL; // common case of whole-array copy
861 Node* last = subseq_length;
862 if (!zero_offset) // last += offset
863 last = _gvn.transform(new AddINode(last, offset));
864 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
865 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
866 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
867 return is_over;
868 }
869
870
871 //--------------------------generate_current_thread--------------------
872 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
873 ciKlass* thread_klass = env()->Thread_klass();
874 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
875 Node* thread = _gvn.transform(new ThreadLocalNode());
876 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
877 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
878 tls_output = thread;
879 return threadObj;
880 }
881
882
883 //------------------------------make_string_method_node------------------------
884 // Helper method for String intrinsic functions. This version is called
885 // with str1 and str2 pointing to String object nodes.
886 //
887 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
888 Node* no_ctrl = NULL;
889
890 // Get start addr of string
891 Node* str1_value = load_String_value(no_ctrl, str1);
892 Node* str1_offset = load_String_offset(no_ctrl, str1);
893 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
894
895 // Get length of string 1
896 Node* str1_len = load_String_length(no_ctrl, str1);
897
898 Node* str2_value = load_String_value(no_ctrl, str2);
899 Node* str2_offset = load_String_offset(no_ctrl, str2);
900 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
901
902 Node* str2_len = NULL;
903 Node* result = NULL;
904
905 switch (opcode) {
906 case Op_StrIndexOf:
907 // Get length of string 2
908 str2_len = load_String_length(no_ctrl, str2);
909
910 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
911 str1_start, str1_len, str2_start, str2_len);
912 break;
913 case Op_StrComp:
914 // Get length of string 2
915 str2_len = load_String_length(no_ctrl, str2);
916
917 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
918 str1_start, str1_len, str2_start, str2_len);
919 break;
920 case Op_StrEquals:
921 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
922 str1_start, str2_start, str1_len);
923 break;
924 default:
925 ShouldNotReachHere();
926 return NULL;
927 }
928
929 // All these intrinsics have checks.
930 C->set_has_split_ifs(true); // Has chance for split-if optimization
931
932 return _gvn.transform(result);
933 }
934
935 // Helper method for String intrinsic functions. This version is called
936 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
937 // to Int nodes containing the lenghts of str1 and str2.
938 //
939 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
940 Node* result = NULL;
941 switch (opcode) {
942 case Op_StrIndexOf:
943 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
944 str1_start, cnt1, str2_start, cnt2);
945 break;
946 case Op_StrComp:
947 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
948 str1_start, cnt1, str2_start, cnt2);
949 break;
950 case Op_StrEquals:
951 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
952 str1_start, str2_start, cnt1);
953 break;
954 default:
955 ShouldNotReachHere();
956 return NULL;
957 }
958
959 // All these intrinsics have checks.
960 C->set_has_split_ifs(true); // Has chance for split-if optimization
961
962 return _gvn.transform(result);
963 }
964
965 //------------------------------inline_string_compareTo------------------------
966 // public int java.lang.String.compareTo(String anotherString);
967 bool LibraryCallKit::inline_string_compareTo() {
968 Node* receiver = null_check(argument(0));
969 Node* arg = null_check(argument(1));
970 if (stopped()) {
971 return true;
972 }
973 set_result(make_string_method_node(Op_StrComp, receiver, arg));
974 return true;
975 }
976
977 //------------------------------inline_string_equals------------------------
978 bool LibraryCallKit::inline_string_equals() {
979 Node* receiver = null_check_receiver();
980 // NOTE: Do not null check argument for String.equals() because spec
981 // allows to specify NULL as argument.
982 Node* argument = this->argument(1);
983 if (stopped()) {
984 return true;
985 }
986
987 // paths (plus control) merge
988 RegionNode* region = new RegionNode(5);
989 Node* phi = new PhiNode(region, TypeInt::BOOL);
990
991 // does source == target string?
992 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
993 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
994
995 Node* if_eq = generate_slow_guard(bol, NULL);
996 if (if_eq != NULL) {
997 // receiver == argument
998 phi->init_req(2, intcon(1));
999 region->init_req(2, if_eq);
1000 }
1001
1002 // get String klass for instanceOf
1003 ciInstanceKlass* klass = env()->String_klass();
1004
1005 if (!stopped()) {
1006 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1007 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1008 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1009
1010 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1011 //instanceOf == true, fallthrough
1012
1013 if (inst_false != NULL) {
1014 phi->init_req(3, intcon(0));
1015 region->init_req(3, inst_false);
1016 }
1017 }
1018
1019 if (!stopped()) {
1020 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1021
1022 // Properly cast the argument to String
1023 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1024 // This path is taken only when argument's type is String:NotNull.
1025 argument = cast_not_null(argument, false);
1026
1027 Node* no_ctrl = NULL;
1028
1029 // Get start addr of receiver
1030 Node* receiver_val = load_String_value(no_ctrl, receiver);
1031 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1032 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1033
1034 // Get length of receiver
1035 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1036
1037 // Get start addr of argument
1038 Node* argument_val = load_String_value(no_ctrl, argument);
1039 Node* argument_offset = load_String_offset(no_ctrl, argument);
1040 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1041
1042 // Get length of argument
1043 Node* argument_cnt = load_String_length(no_ctrl, argument);
1044
1045 // Check for receiver count != argument count
1046 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1047 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1048 Node* if_ne = generate_slow_guard(bol, NULL);
1049 if (if_ne != NULL) {
1050 phi->init_req(4, intcon(0));
1051 region->init_req(4, if_ne);
1052 }
1053
1054 // Check for count == 0 is done by assembler code for StrEquals.
1055
1056 if (!stopped()) {
1057 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1058 phi->init_req(1, equals);
1059 region->init_req(1, control());
1060 }
1061 }
1062
1063 // post merge
1064 set_control(_gvn.transform(region));
1065 record_for_igvn(region);
1066
1067 set_result(_gvn.transform(phi));
1068 return true;
1069 }
1070
1071 //------------------------------inline_array_equals----------------------------
1072 bool LibraryCallKit::inline_array_equals() {
1073 Node* arg1 = argument(0);
1074 Node* arg2 = argument(1);
1075 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1076 return true;
1077 }
1078
1079 // Java version of String.indexOf(constant string)
1080 // class StringDecl {
1081 // StringDecl(char[] ca) {
1082 // offset = 0;
1083 // count = ca.length;
1084 // value = ca;
1085 // }
1086 // int offset;
1087 // int count;
1088 // char[] value;
1089 // }
1090 //
1091 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1092 // int targetOffset, int cache_i, int md2) {
1093 // int cache = cache_i;
1094 // int sourceOffset = string_object.offset;
1095 // int sourceCount = string_object.count;
1096 // int targetCount = target_object.length;
1097 //
1098 // int targetCountLess1 = targetCount - 1;
1099 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1100 //
1101 // char[] source = string_object.value;
1102 // char[] target = target_object;
1103 // int lastChar = target[targetCountLess1];
1104 //
1105 // outer_loop:
1106 // for (int i = sourceOffset; i < sourceEnd; ) {
1107 // int src = source[i + targetCountLess1];
1108 // if (src == lastChar) {
1109 // // With random strings and a 4-character alphabet,
1110 // // reverse matching at this point sets up 0.8% fewer
1111 // // frames, but (paradoxically) makes 0.3% more probes.
1112 // // Since those probes are nearer the lastChar probe,
1113 // // there is may be a net D$ win with reverse matching.
1114 // // But, reversing loop inhibits unroll of inner loop
1115 // // for unknown reason. So, does running outer loop from
1116 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1117 // for (int j = 0; j < targetCountLess1; j++) {
1118 // if (target[targetOffset + j] != source[i+j]) {
1119 // if ((cache & (1 << source[i+j])) == 0) {
1120 // if (md2 < j+1) {
1121 // i += j+1;
1122 // continue outer_loop;
1123 // }
1124 // }
1125 // i += md2;
1126 // continue outer_loop;
1127 // }
1128 // }
1129 // return i - sourceOffset;
1130 // }
1131 // if ((cache & (1 << src)) == 0) {
1132 // i += targetCountLess1;
1133 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1134 // i++;
1135 // }
1136 // return -1;
1137 // }
1138
1139 //------------------------------string_indexOf------------------------
1140 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1141 jint cache_i, jint md2_i) {
1142
1143 Node* no_ctrl = NULL;
1144 float likely = PROB_LIKELY(0.9);
1145 float unlikely = PROB_UNLIKELY(0.9);
1146
1147 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1148
1149 Node* source = load_String_value(no_ctrl, string_object);
1150 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1151 Node* sourceCount = load_String_length(no_ctrl, string_object);
1152
1153 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1154 jint target_length = target_array->length();
1155 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1156 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1157
1158 // String.value field is known to be @Stable.
1159 if (UseImplicitStableValues) {
1160 target = cast_array_to_stable(target, target_type);
1161 }
1162
1163 IdealKit kit(this, false, true);
1164 #define __ kit.
1165 Node* zero = __ ConI(0);
1166 Node* one = __ ConI(1);
1167 Node* cache = __ ConI(cache_i);
1168 Node* md2 = __ ConI(md2_i);
1169 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1170 Node* targetCountLess1 = __ ConI(target_length - 1);
1171 Node* targetOffset = __ ConI(targetOffset_i);
1172 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1173
1174 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1175 Node* outer_loop = __ make_label(2 /* goto */);
1176 Node* return_ = __ make_label(1);
1177
1178 __ set(rtn,__ ConI(-1));
1179 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1180 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1181 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1182 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1183 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1184 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1185 Node* tpj = __ AddI(targetOffset, __ value(j));
1186 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1187 Node* ipj = __ AddI(__ value(i), __ value(j));
1188 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1189 __ if_then(targ, BoolTest::ne, src2); {
1190 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1191 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1192 __ increment(i, __ AddI(__ value(j), one));
1193 __ goto_(outer_loop);
1194 } __ end_if(); __ dead(j);
1195 }__ end_if(); __ dead(j);
1196 __ increment(i, md2);
1197 __ goto_(outer_loop);
1198 }__ end_if();
1199 __ increment(j, one);
1200 }__ end_loop(); __ dead(j);
1201 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1202 __ goto_(return_);
1203 }__ end_if();
1204 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1205 __ increment(i, targetCountLess1);
1206 }__ end_if();
1207 __ increment(i, one);
1208 __ bind(outer_loop);
1209 }__ end_loop(); __ dead(i);
1210 __ bind(return_);
1211
1212 // Final sync IdealKit and GraphKit.
1213 final_sync(kit);
1214 Node* result = __ value(rtn);
1215 #undef __
1216 C->set_has_loops(true);
1217 return result;
1218 }
1219
1220 //------------------------------inline_string_indexOf------------------------
1221 bool LibraryCallKit::inline_string_indexOf() {
1222 Node* receiver = argument(0);
1223 Node* arg = argument(1);
1224
1225 Node* result;
1226 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1227 UseSSE42Intrinsics) {
1228 // Generate SSE4.2 version of indexOf
1229 // We currently only have match rules that use SSE4.2
1230
1231 receiver = null_check(receiver);
1232 arg = null_check(arg);
1233 if (stopped()) {
1234 return true;
1235 }
1236
1237 // Make the merge point
1238 RegionNode* result_rgn = new RegionNode(4);
1239 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1240 Node* no_ctrl = NULL;
1241
1242 // Get start addr of source string
1243 Node* source = load_String_value(no_ctrl, receiver);
1244 Node* source_offset = load_String_offset(no_ctrl, receiver);
1245 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1246
1247 // Get length of source string
1248 Node* source_cnt = load_String_length(no_ctrl, receiver);
1249
1250 // Get start addr of substring
1251 Node* substr = load_String_value(no_ctrl, arg);
1252 Node* substr_offset = load_String_offset(no_ctrl, arg);
1253 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1254
1255 // Get length of source string
1256 Node* substr_cnt = load_String_length(no_ctrl, arg);
1257
1258 // Check for substr count > string count
1259 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1260 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1261 Node* if_gt = generate_slow_guard(bol, NULL);
1262 if (if_gt != NULL) {
1263 result_phi->init_req(2, intcon(-1));
1264 result_rgn->init_req(2, if_gt);
1265 }
1266
1267 if (!stopped()) {
1268 // Check for substr count == 0
1269 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1270 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1271 Node* if_zero = generate_slow_guard(bol, NULL);
1272 if (if_zero != NULL) {
1273 result_phi->init_req(3, intcon(0));
1274 result_rgn->init_req(3, if_zero);
1275 }
1276 }
1277
1278 if (!stopped()) {
1279 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1280 result_phi->init_req(1, result);
1281 result_rgn->init_req(1, control());
1282 }
1283 set_control(_gvn.transform(result_rgn));
1284 record_for_igvn(result_rgn);
1285 result = _gvn.transform(result_phi);
1286
1287 } else { // Use LibraryCallKit::string_indexOf
1288 // don't intrinsify if argument isn't a constant string.
1289 if (!arg->is_Con()) {
1290 return false;
1291 }
1292 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1293 if (str_type == NULL) {
1294 return false;
1295 }
1296 ciInstanceKlass* klass = env()->String_klass();
1297 ciObject* str_const = str_type->const_oop();
1298 if (str_const == NULL || str_const->klass() != klass) {
1299 return false;
1300 }
1301 ciInstance* str = str_const->as_instance();
1302 assert(str != NULL, "must be instance");
1303
1304 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1305 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1306
1307 int o;
1308 int c;
1309 if (java_lang_String::has_offset_field()) {
1310 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1311 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1312 } else {
1313 o = 0;
1314 c = pat->length();
1315 }
1316
1317 // constant strings have no offset and count == length which
1318 // simplifies the resulting code somewhat so lets optimize for that.
1319 if (o != 0 || c != pat->length()) {
1320 return false;
1321 }
1322
1323 receiver = null_check(receiver, T_OBJECT);
1324 // NOTE: No null check on the argument is needed since it's a constant String oop.
1325 if (stopped()) {
1326 return true;
1327 }
1328
1329 // The null string as a pattern always returns 0 (match at beginning of string)
1330 if (c == 0) {
1331 set_result(intcon(0));
1332 return true;
1333 }
1334
1335 // Generate default indexOf
1336 jchar lastChar = pat->char_at(o + (c - 1));
1337 int cache = 0;
1338 int i;
1339 for (i = 0; i < c - 1; i++) {
1340 assert(i < pat->length(), "out of range");
1341 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1342 }
1343
1344 int md2 = c;
1345 for (i = 0; i < c - 1; i++) {
1346 assert(i < pat->length(), "out of range");
1347 if (pat->char_at(o + i) == lastChar) {
1348 md2 = (c - 1) - i;
1349 }
1350 }
1351
1352 result = string_indexOf(receiver, pat, o, cache, md2);
1353 }
1354 set_result(result);
1355 return true;
1356 }
1357
1358 //--------------------------round_double_node--------------------------------
1359 // Round a double node if necessary.
1360 Node* LibraryCallKit::round_double_node(Node* n) {
1361 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1362 n = _gvn.transform(new RoundDoubleNode(0, n));
1363 return n;
1364 }
1365
1366 //------------------------------inline_math-----------------------------------
1367 // public static double Math.abs(double)
1368 // public static double Math.sqrt(double)
1369 // public static double Math.log(double)
1370 // public static double Math.log10(double)
1371 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1372 Node* arg = round_double_node(argument(0));
1373 Node* n;
1374 switch (id) {
1375 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1376 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1377 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1378 default: fatal_unexpected_iid(id); break;
1379 }
1380 set_result(_gvn.transform(n));
1381 return true;
1382 }
1383
1384 //------------------------------inline_trig----------------------------------
1385 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1386 // argument reduction which will turn into a fast/slow diamond.
1387 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1388 Node* arg = round_double_node(argument(0));
1389 Node* n = NULL;
1390
1391 switch (id) {
1392 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1393 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1394 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1395 default: fatal_unexpected_iid(id); break;
1396 }
1397 n = _gvn.transform(n);
1398
1399 // Rounding required? Check for argument reduction!
1400 if (Matcher::strict_fp_requires_explicit_rounding) {
1401 static const double pi_4 = 0.7853981633974483;
1402 static const double neg_pi_4 = -0.7853981633974483;
1403 // pi/2 in 80-bit extended precision
1404 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1405 // -pi/2 in 80-bit extended precision
1406 // 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};
1407 // Cutoff value for using this argument reduction technique
1408 //static const double pi_2_minus_epsilon = 1.564660403643354;
1409 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1410
1411 // Pseudocode for sin:
1412 // if (x <= Math.PI / 4.0) {
1413 // if (x >= -Math.PI / 4.0) return fsin(x);
1414 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1415 // } else {
1416 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1417 // }
1418 // return StrictMath.sin(x);
1419
1420 // Pseudocode for cos:
1421 // if (x <= Math.PI / 4.0) {
1422 // if (x >= -Math.PI / 4.0) return fcos(x);
1423 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1424 // } else {
1425 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1426 // }
1427 // return StrictMath.cos(x);
1428
1429 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1430 // requires a special machine instruction to load it. Instead we'll try
1431 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1432 // probably do the math inside the SIN encoding.
1433
1434 // Make the merge point
1435 RegionNode* r = new RegionNode(3);
1436 Node* phi = new PhiNode(r, Type::DOUBLE);
1437
1438 // Flatten arg so we need only 1 test
1439 Node *abs = _gvn.transform(new AbsDNode(arg));
1440 // Node for PI/4 constant
1441 Node *pi4 = makecon(TypeD::make(pi_4));
1442 // Check PI/4 : abs(arg)
1443 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1444 // Check: If PI/4 < abs(arg) then go slow
1445 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1446 // Branch either way
1447 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1448 set_control(opt_iff(r,iff));
1449
1450 // Set fast path result
1451 phi->init_req(2, n);
1452
1453 // Slow path - non-blocking leaf call
1454 Node* call = NULL;
1455 switch (id) {
1456 case vmIntrinsics::_dsin:
1457 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1458 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1459 "Sin", NULL, arg, top());
1460 break;
1461 case vmIntrinsics::_dcos:
1462 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1463 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1464 "Cos", NULL, arg, top());
1465 break;
1466 case vmIntrinsics::_dtan:
1467 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1468 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1469 "Tan", NULL, arg, top());
1470 break;
1471 }
1472 assert(control()->in(0) == call, "");
1473 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1474 r->init_req(1, control());
1475 phi->init_req(1, slow_result);
1476
1477 // Post-merge
1478 set_control(_gvn.transform(r));
1479 record_for_igvn(r);
1480 n = _gvn.transform(phi);
1481
1482 C->set_has_split_ifs(true); // Has chance for split-if optimization
1483 }
1484 set_result(n);
1485 return true;
1486 }
1487
1488 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1489 //-------------------
1490 //result=(result.isNaN())? funcAddr():result;
1491 // Check: If isNaN() by checking result!=result? then either trap
1492 // or go to runtime
1493 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1494 // Build the boolean node
1495 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1496
1497 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1498 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1499 // The pow or exp intrinsic returned a NaN, which requires a call
1500 // to the runtime. Recompile with the runtime call.
1501 uncommon_trap(Deoptimization::Reason_intrinsic,
1502 Deoptimization::Action_make_not_entrant);
1503 }
1504 return result;
1505 } else {
1506 // If this inlining ever returned NaN in the past, we compile a call
1507 // to the runtime to properly handle corner cases
1508
1509 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1510 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1511 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1512
1513 if (!if_slow->is_top()) {
1514 RegionNode* result_region = new RegionNode(3);
1515 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1516
1517 result_region->init_req(1, if_fast);
1518 result_val->init_req(1, result);
1519
1520 set_control(if_slow);
1521
1522 const TypePtr* no_memory_effects = NULL;
1523 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1524 no_memory_effects,
1525 x, top(), y, y ? top() : NULL);
1526 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1527 #ifdef ASSERT
1528 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1529 assert(value_top == top(), "second value must be top");
1530 #endif
1531
1532 result_region->init_req(2, control());
1533 result_val->init_req(2, value);
1534 set_control(_gvn.transform(result_region));
1535 return _gvn.transform(result_val);
1536 } else {
1537 return result;
1538 }
1539 }
1540 }
1541
1542 //------------------------------inline_pow-------------------------------------
1543 // Inline power instructions, if possible.
1544 bool LibraryCallKit::inline_pow() {
1545 // Pseudocode for pow
1546 // if (y == 2) {
1547 // return x * x;
1548 // } else {
1549 // if (x <= 0.0) {
1550 // long longy = (long)y;
1551 // if ((double)longy == y) { // if y is long
1552 // if (y + 1 == y) longy = 0; // huge number: even
1553 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1554 // } else {
1555 // result = NaN;
1556 // }
1557 // } else {
1558 // result = DPow(x,y);
1559 // }
1560 // if (result != result)? {
1561 // result = uncommon_trap() or runtime_call();
1562 // }
1563 // return result;
1564 // }
1565
1566 Node* x = round_double_node(argument(0));
1567 Node* y = round_double_node(argument(2));
1568
1569 Node* result = NULL;
1570
1571 Node* const_two_node = makecon(TypeD::make(2.0));
1572 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1573 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1574 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1575 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1576 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1577
1578 RegionNode* region_node = new RegionNode(3);
1579 region_node->init_req(1, if_true);
1580
1581 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1582 // special case for x^y where y == 2, we can convert it to x * x
1583 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1584
1585 // set control to if_false since we will now process the false branch
1586 set_control(if_false);
1587
1588 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1589 // Short form: skip the fancy tests and just check for NaN result.
1590 result = _gvn.transform(new PowDNode(C, control(), x, y));
1591 } else {
1592 // If this inlining ever returned NaN in the past, include all
1593 // checks + call to the runtime.
1594
1595 // Set the merge point for If node with condition of (x <= 0.0)
1596 // There are four possible paths to region node and phi node
1597 RegionNode *r = new RegionNode(4);
1598 Node *phi = new PhiNode(r, Type::DOUBLE);
1599
1600 // Build the first if node: if (x <= 0.0)
1601 // Node for 0 constant
1602 Node *zeronode = makecon(TypeD::ZERO);
1603 // Check x:0
1604 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1605 // Check: If (x<=0) then go complex path
1606 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1607 // Branch either way
1608 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1609 // Fast path taken; set region slot 3
1610 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1611 r->init_req(3,fast_taken); // Capture fast-control
1612
1613 // Fast path not-taken, i.e. slow path
1614 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1615
1616 // Set fast path result
1617 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1618 phi->init_req(3, fast_result);
1619
1620 // Complex path
1621 // Build the second if node (if y is long)
1622 // Node for (long)y
1623 Node *longy = _gvn.transform(new ConvD2LNode(y));
1624 // Node for (double)((long) y)
1625 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1626 // Check (double)((long) y) : y
1627 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1628 // Check if (y isn't long) then go to slow path
1629
1630 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1631 // Branch either way
1632 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1633 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1634
1635 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1636
1637 // Calculate DPow(abs(x), y)*(1 & (long)y)
1638 // Node for constant 1
1639 Node *conone = longcon(1);
1640 // 1& (long)y
1641 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1642
1643 // A huge number is always even. Detect a huge number by checking
1644 // if y + 1 == y and set integer to be tested for parity to 0.
1645 // Required for corner case:
1646 // (long)9.223372036854776E18 = max_jlong
1647 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1648 // max_jlong is odd but 9.223372036854776E18 is even
1649 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1650 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1651 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1652 Node* correctedsign = NULL;
1653 if (ConditionalMoveLimit != 0) {
1654 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1655 } else {
1656 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1657 RegionNode *r = new RegionNode(3);
1658 Node *phi = new PhiNode(r, TypeLong::LONG);
1659 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1660 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1661 phi->init_req(1, signnode);
1662 phi->init_req(2, longcon(0));
1663 correctedsign = _gvn.transform(phi);
1664 ylong_path = _gvn.transform(r);
1665 record_for_igvn(r);
1666 }
1667
1668 // zero node
1669 Node *conzero = longcon(0);
1670 // Check (1&(long)y)==0?
1671 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1672 // Check if (1&(long)y)!=0?, if so the result is negative
1673 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1674 // abs(x)
1675 Node *absx=_gvn.transform(new AbsDNode(x));
1676 // abs(x)^y
1677 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1678 // -abs(x)^y
1679 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1680 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1681 Node *signresult = NULL;
1682 if (ConditionalMoveLimit != 0) {
1683 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1684 } else {
1685 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1686 RegionNode *r = new RegionNode(3);
1687 Node *phi = new PhiNode(r, Type::DOUBLE);
1688 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1689 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1690 phi->init_req(1, absxpowy);
1691 phi->init_req(2, negabsxpowy);
1692 signresult = _gvn.transform(phi);
1693 ylong_path = _gvn.transform(r);
1694 record_for_igvn(r);
1695 }
1696 // Set complex path fast result
1697 r->init_req(2, ylong_path);
1698 phi->init_req(2, signresult);
1699
1700 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1701 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1702 r->init_req(1,slow_path);
1703 phi->init_req(1,slow_result);
1704
1705 // Post merge
1706 set_control(_gvn.transform(r));
1707 record_for_igvn(r);
1708 result = _gvn.transform(phi);
1709 }
1710
1711 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1712
1713 // control from finish_pow_exp is now input to the region node
1714 region_node->set_req(2, control());
1715 // the result from finish_pow_exp is now input to the phi node
1716 phi_node->init_req(2, result);
1717 set_control(_gvn.transform(region_node));
1718 record_for_igvn(region_node);
1719 set_result(_gvn.transform(phi_node));
1720
1721 C->set_has_split_ifs(true); // Has chance for split-if optimization
1722 return true;
1723 }
1724
1725 //------------------------------runtime_math-----------------------------
1726 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1727 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1728 "must be (DD)D or (D)D type");
1729
1730 // Inputs
1731 Node* a = round_double_node(argument(0));
1732 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1733
1734 const TypePtr* no_memory_effects = NULL;
1735 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1736 no_memory_effects,
1737 a, top(), b, b ? top() : NULL);
1738 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1739 #ifdef ASSERT
1740 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1741 assert(value_top == top(), "second value must be top");
1742 #endif
1743
1744 set_result(value);
1745 return true;
1746 }
1747
1748 //------------------------------inline_math_native-----------------------------
1749 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1750 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1751 switch (id) {
1752 // These intrinsics are not properly supported on all hardware
1753 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1754 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1755 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1756 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1757 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1758 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1759
1760 case vmIntrinsics::_dlog:
1761 return StubRoutines::dlog() != NULL ?
1762 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1763 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1764 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1765 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1766
1767 // These intrinsics are supported on all hardware
1768 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1769 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1770
1771 case vmIntrinsics::_dexp:
1772 return StubRoutines::dexp() != NULL ?
1773 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1774 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1775 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1776 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1777 #undef FN_PTR
1778
1779 // These intrinsics are not yet correctly implemented
1780 case vmIntrinsics::_datan2:
1781 return false;
1782
1783 default:
1784 fatal_unexpected_iid(id);
1785 return false;
1786 }
1787 }
1788
1789 static bool is_simple_name(Node* n) {
1790 return (n->req() == 1 // constant
1791 || (n->is_Type() && n->as_Type()->type()->singleton())
1792 || n->is_Proj() // parameter or return value
1793 || n->is_Phi() // local of some sort
1794 );
1795 }
1796
1797 //----------------------------inline_notify-----------------------------------*
1798 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1799 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1800 address func;
1801 if (id == vmIntrinsics::_notify) {
1802 func = OptoRuntime::monitor_notify_Java();
1803 } else {
1804 func = OptoRuntime::monitor_notifyAll_Java();
1805 }
1806 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1807 make_slow_call_ex(call, env()->Throwable_klass(), false);
1808 return true;
1809 }
1810
1811
1812 //----------------------------inline_min_max-----------------------------------
1813 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1814 set_result(generate_min_max(id, argument(0), argument(1)));
1815 return true;
1816 }
1817
1818 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1819 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1820 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1821 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1822 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1823
1824 {
1825 PreserveJVMState pjvms(this);
1826 PreserveReexecuteState preexecs(this);
1827 jvms()->set_should_reexecute(true);
1828
1829 set_control(slow_path);
1830 set_i_o(i_o());
1831
1832 uncommon_trap(Deoptimization::Reason_intrinsic,
1833 Deoptimization::Action_none);
1834 }
1835
1836 set_control(fast_path);
1837 set_result(math);
1838 }
1839
1840 template <typename OverflowOp>
1841 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1842 typedef typename OverflowOp::MathOp MathOp;
1843
1844 MathOp* mathOp = new MathOp(arg1, arg2);
1845 Node* operation = _gvn.transform( mathOp );
1846 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1847 inline_math_mathExact(operation, ofcheck);
1848 return true;
1849 }
1850
1851 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1852 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1853 }
1854
1855 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1856 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1857 }
1858
1859 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1860 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1861 }
1862
1863 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1864 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1865 }
1866
1867 bool LibraryCallKit::inline_math_negateExactI() {
1868 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1869 }
1870
1871 bool LibraryCallKit::inline_math_negateExactL() {
1872 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1873 }
1874
1875 bool LibraryCallKit::inline_math_multiplyExactI() {
1876 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1877 }
1878
1879 bool LibraryCallKit::inline_math_multiplyExactL() {
1880 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1881 }
1882
1883 Node*
1884 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1885 // These are the candidate return value:
1886 Node* xvalue = x0;
1887 Node* yvalue = y0;
1888
1889 if (xvalue == yvalue) {
1890 return xvalue;
1891 }
1892
1893 bool want_max = (id == vmIntrinsics::_max);
1894
1895 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1896 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1897 if (txvalue == NULL || tyvalue == NULL) return top();
1898 // This is not really necessary, but it is consistent with a
1899 // hypothetical MaxINode::Value method:
1900 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1901
1902 // %%% This folding logic should (ideally) be in a different place.
1903 // Some should be inside IfNode, and there to be a more reliable
1904 // transformation of ?: style patterns into cmoves. We also want
1905 // more powerful optimizations around cmove and min/max.
1906
1907 // Try to find a dominating comparison of these guys.
1908 // It can simplify the index computation for Arrays.copyOf
1909 // and similar uses of System.arraycopy.
1910 // First, compute the normalized version of CmpI(x, y).
1911 int cmp_op = Op_CmpI;
1912 Node* xkey = xvalue;
1913 Node* ykey = yvalue;
1914 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1915 if (ideal_cmpxy->is_Cmp()) {
1916 // E.g., if we have CmpI(length - offset, count),
1917 // it might idealize to CmpI(length, count + offset)
1918 cmp_op = ideal_cmpxy->Opcode();
1919 xkey = ideal_cmpxy->in(1);
1920 ykey = ideal_cmpxy->in(2);
1921 }
1922
1923 // Start by locating any relevant comparisons.
1924 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1925 Node* cmpxy = NULL;
1926 Node* cmpyx = NULL;
1927 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1928 Node* cmp = start_from->fast_out(k);
1929 if (cmp->outcnt() > 0 && // must have prior uses
1930 cmp->in(0) == NULL && // must be context-independent
1931 cmp->Opcode() == cmp_op) { // right kind of compare
1932 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1933 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1934 }
1935 }
1936
1937 const int NCMPS = 2;
1938 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1939 int cmpn;
1940 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1941 if (cmps[cmpn] != NULL) break; // find a result
1942 }
1943 if (cmpn < NCMPS) {
1944 // Look for a dominating test that tells us the min and max.
1945 int depth = 0; // Limit search depth for speed
1946 Node* dom = control();
1947 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1948 if (++depth >= 100) break;
1949 Node* ifproj = dom;
1950 if (!ifproj->is_Proj()) continue;
1951 Node* iff = ifproj->in(0);
1952 if (!iff->is_If()) continue;
1953 Node* bol = iff->in(1);
1954 if (!bol->is_Bool()) continue;
1955 Node* cmp = bol->in(1);
1956 if (cmp == NULL) continue;
1957 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1958 if (cmps[cmpn] == cmp) break;
1959 if (cmpn == NCMPS) continue;
1960 BoolTest::mask btest = bol->as_Bool()->_test._test;
1961 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1962 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1963 // At this point, we know that 'x btest y' is true.
1964 switch (btest) {
1965 case BoolTest::eq:
1966 // They are proven equal, so we can collapse the min/max.
1967 // Either value is the answer. Choose the simpler.
1968 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1969 return yvalue;
1970 return xvalue;
1971 case BoolTest::lt: // x < y
1972 case BoolTest::le: // x <= y
1973 return (want_max ? yvalue : xvalue);
1974 case BoolTest::gt: // x > y
1975 case BoolTest::ge: // x >= y
1976 return (want_max ? xvalue : yvalue);
1977 }
1978 }
1979 }
1980
1981 // We failed to find a dominating test.
1982 // Let's pick a test that might GVN with prior tests.
1983 Node* best_bol = NULL;
1984 BoolTest::mask best_btest = BoolTest::illegal;
1985 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1986 Node* cmp = cmps[cmpn];
1987 if (cmp == NULL) continue;
1988 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1989 Node* bol = cmp->fast_out(j);
1990 if (!bol->is_Bool()) continue;
1991 BoolTest::mask btest = bol->as_Bool()->_test._test;
1992 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1993 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1994 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1995 best_bol = bol->as_Bool();
1996 best_btest = btest;
1997 }
1998 }
1999 }
2000
2001 Node* answer_if_true = NULL;
2002 Node* answer_if_false = NULL;
2003 switch (best_btest) {
2004 default:
2005 if (cmpxy == NULL)
2006 cmpxy = ideal_cmpxy;
2007 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2008 // and fall through:
2009 case BoolTest::lt: // x < y
2010 case BoolTest::le: // x <= y
2011 answer_if_true = (want_max ? yvalue : xvalue);
2012 answer_if_false = (want_max ? xvalue : yvalue);
2013 break;
2014 case BoolTest::gt: // x > y
2015 case BoolTest::ge: // x >= y
2016 answer_if_true = (want_max ? xvalue : yvalue);
2017 answer_if_false = (want_max ? yvalue : xvalue);
2018 break;
2019 }
2020
2021 jint hi, lo;
2022 if (want_max) {
2023 // We can sharpen the minimum.
2024 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2025 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2026 } else {
2027 // We can sharpen the maximum.
2028 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2029 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2030 }
2031
2032 // Use a flow-free graph structure, to avoid creating excess control edges
2033 // which could hinder other optimizations.
2034 // Since Math.min/max is often used with arraycopy, we want
2035 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2036 Node* cmov = CMoveNode::make(NULL, best_bol,
2037 answer_if_false, answer_if_true,
2038 TypeInt::make(lo, hi, widen));
2039
2040 return _gvn.transform(cmov);
2041
2042 /*
2043 // This is not as desirable as it may seem, since Min and Max
2044 // nodes do not have a full set of optimizations.
2045 // And they would interfere, anyway, with 'if' optimizations
2046 // and with CMoveI canonical forms.
2047 switch (id) {
2048 case vmIntrinsics::_min:
2049 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2050 case vmIntrinsics::_max:
2051 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2052 default:
2053 ShouldNotReachHere();
2054 }
2055 */
2056 }
2057
2058 inline int
2059 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2060 const TypePtr* base_type = TypePtr::NULL_PTR;
2061 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2062 if (base_type == NULL) {
2063 // Unknown type.
2064 return Type::AnyPtr;
2065 } else if (base_type == TypePtr::NULL_PTR) {
2066 // Since this is a NULL+long form, we have to switch to a rawptr.
2067 base = _gvn.transform(new CastX2PNode(offset));
2068 offset = MakeConX(0);
2069 return Type::RawPtr;
2070 } else if (base_type->base() == Type::RawPtr) {
2071 return Type::RawPtr;
2072 } else if (base_type->isa_oopptr()) {
2073 // Base is never null => always a heap address.
2074 if (base_type->ptr() == TypePtr::NotNull) {
2075 return Type::OopPtr;
2076 }
2077 // Offset is small => always a heap address.
2078 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2079 if (offset_type != NULL &&
2080 base_type->offset() == 0 && // (should always be?)
2081 offset_type->_lo >= 0 &&
2082 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2083 return Type::OopPtr;
2084 }
2085 // Otherwise, it might either be oop+off or NULL+addr.
2086 return Type::AnyPtr;
2087 } else {
2088 // No information:
2089 return Type::AnyPtr;
2090 }
2091 }
2092
2093 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2094 int kind = classify_unsafe_addr(base, offset);
2095 if (kind == Type::RawPtr) {
2096 return basic_plus_adr(top(), base, offset);
2097 } else {
2098 return basic_plus_adr(base, offset);
2099 }
2100 }
2101
2102 //--------------------------inline_number_methods-----------------------------
2103 // inline int Integer.numberOfLeadingZeros(int)
2104 // inline int Long.numberOfLeadingZeros(long)
2105 //
2106 // inline int Integer.numberOfTrailingZeros(int)
2107 // inline int Long.numberOfTrailingZeros(long)
2108 //
2109 // inline int Integer.bitCount(int)
2110 // inline int Long.bitCount(long)
2111 //
2112 // inline char Character.reverseBytes(char)
2113 // inline short Short.reverseBytes(short)
2114 // inline int Integer.reverseBytes(int)
2115 // inline long Long.reverseBytes(long)
2116 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2117 Node* arg = argument(0);
2118 Node* n;
2119 switch (id) {
2120 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2121 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2122 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2123 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2124 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2125 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2126 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2127 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2128 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2129 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2130 default: fatal_unexpected_iid(id); break;
2131 }
2132 set_result(_gvn.transform(n));
2133 return true;
2134 }
2135
2136 //----------------------------inline_unsafe_access----------------------------
2137
2138 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2139
2140 // Helper that guards and inserts a pre-barrier.
2141 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2142 Node* pre_val, bool need_mem_bar) {
2143 // We could be accessing the referent field of a reference object. If so, when G1
2144 // is enabled, we need to log the value in the referent field in an SATB buffer.
2145 // This routine performs some compile time filters and generates suitable
2146 // runtime filters that guard the pre-barrier code.
2147 // Also add memory barrier for non volatile load from the referent field
2148 // to prevent commoning of loads across safepoint.
2149 if (!UseG1GC && !need_mem_bar)
2150 return;
2151
2152 // Some compile time checks.
2153
2154 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2155 const TypeX* otype = offset->find_intptr_t_type();
2156 if (otype != NULL && otype->is_con() &&
2157 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2158 // Constant offset but not the reference_offset so just return
2159 return;
2160 }
2161
2162 // We only need to generate the runtime guards for instances.
2163 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2164 if (btype != NULL) {
2165 if (btype->isa_aryptr()) {
2166 // Array type so nothing to do
2167 return;
2168 }
2169
2170 const TypeInstPtr* itype = btype->isa_instptr();
2171 if (itype != NULL) {
2172 // Can the klass of base_oop be statically determined to be
2173 // _not_ a sub-class of Reference and _not_ Object?
2174 ciKlass* klass = itype->klass();
2175 if ( klass->is_loaded() &&
2176 !klass->is_subtype_of(env()->Reference_klass()) &&
2177 !env()->Object_klass()->is_subtype_of(klass)) {
2178 return;
2179 }
2180 }
2181 }
2182
2183 // The compile time filters did not reject base_oop/offset so
2184 // we need to generate the following runtime filters
2185 //
2186 // if (offset == java_lang_ref_Reference::_reference_offset) {
2187 // if (instance_of(base, java.lang.ref.Reference)) {
2188 // pre_barrier(_, pre_val, ...);
2189 // }
2190 // }
2191
2192 float likely = PROB_LIKELY( 0.999);
2193 float unlikely = PROB_UNLIKELY(0.999);
2194
2195 IdealKit ideal(this);
2196 #define __ ideal.
2197
2198 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2199
2200 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2201 // Update graphKit memory and control from IdealKit.
2202 sync_kit(ideal);
2203
2204 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2205 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2206
2207 // Update IdealKit memory and control from graphKit.
2208 __ sync_kit(this);
2209
2210 Node* one = __ ConI(1);
2211 // is_instof == 0 if base_oop == NULL
2212 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2213
2214 // Update graphKit from IdeakKit.
2215 sync_kit(ideal);
2216
2217 // Use the pre-barrier to record the value in the referent field
2218 pre_barrier(false /* do_load */,
2219 __ ctrl(),
2220 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2221 pre_val /* pre_val */,
2222 T_OBJECT);
2223 if (need_mem_bar) {
2224 // Add memory barrier to prevent commoning reads from this field
2225 // across safepoint since GC can change its value.
2226 insert_mem_bar(Op_MemBarCPUOrder);
2227 }
2228 // Update IdealKit from graphKit.
2229 __ sync_kit(this);
2230
2231 } __ end_if(); // _ref_type != ref_none
2232 } __ end_if(); // offset == referent_offset
2233
2234 // Final sync IdealKit and GraphKit.
2235 final_sync(ideal);
2236 #undef __
2237 }
2238
2239
2240 // Interpret Unsafe.fieldOffset cookies correctly:
2241 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2242
2243 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2244 // Attempt to infer a sharper value type from the offset and base type.
2245 ciKlass* sharpened_klass = NULL;
2246
2247 // See if it is an instance field, with an object type.
2248 if (alias_type->field() != NULL) {
2249 assert(!is_native_ptr, "native pointer op cannot use a java address");
2250 if (alias_type->field()->type()->is_klass()) {
2251 sharpened_klass = alias_type->field()->type()->as_klass();
2252 }
2253 }
2254
2255 // See if it is a narrow oop array.
2256 if (adr_type->isa_aryptr()) {
2257 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2258 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2259 if (elem_type != NULL) {
2260 sharpened_klass = elem_type->klass();
2261 }
2262 }
2263 }
2264
2265 // The sharpened class might be unloaded if there is no class loader
2266 // contraint in place.
2267 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2268 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2269
2270 #ifndef PRODUCT
2271 if (C->print_intrinsics() || C->print_inlining()) {
2272 tty->print(" from base type: "); adr_type->dump();
2273 tty->print(" sharpened value: "); tjp->dump();
2274 }
2275 #endif
2276 // Sharpen the value type.
2277 return tjp;
2278 }
2279 return NULL;
2280 }
2281
2282 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2283 if (callee()->is_static()) return false; // caller must have the capability!
2284
2285 #ifndef PRODUCT
2286 {
2287 ResourceMark rm;
2288 // Check the signatures.
2289 ciSignature* sig = callee()->signature();
2290 #ifdef ASSERT
2291 if (!is_store) {
2292 // Object getObject(Object base, int/long offset), etc.
2293 BasicType rtype = sig->return_type()->basic_type();
2294 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2295 rtype = T_ADDRESS; // it is really a C void*
2296 assert(rtype == type, "getter must return the expected value");
2297 if (!is_native_ptr) {
2298 assert(sig->count() == 2, "oop getter has 2 arguments");
2299 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2300 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2301 } else {
2302 assert(sig->count() == 1, "native getter has 1 argument");
2303 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2304 }
2305 } else {
2306 // void putObject(Object base, int/long offset, Object x), etc.
2307 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2308 if (!is_native_ptr) {
2309 assert(sig->count() == 3, "oop putter has 3 arguments");
2310 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2311 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2312 } else {
2313 assert(sig->count() == 2, "native putter has 2 arguments");
2314 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2315 }
2316 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2317 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2318 vtype = T_ADDRESS; // it is really a C void*
2319 assert(vtype == type, "putter must accept the expected value");
2320 }
2321 #endif // ASSERT
2322 }
2323 #endif //PRODUCT
2324
2325 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2326
2327 Node* receiver = argument(0); // type: oop
2328
2329 // Build address expression.
2330 Node* adr;
2331 Node* heap_base_oop = top();
2332 Node* offset = top();
2333 Node* val;
2334
2335 if (!is_native_ptr) {
2336 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2337 Node* base = argument(1); // type: oop
2338 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2339 offset = argument(2); // type: long
2340 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2341 // to be plain byte offsets, which are also the same as those accepted
2342 // by oopDesc::field_base.
2343 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2344 "fieldOffset must be byte-scaled");
2345 // 32-bit machines ignore the high half!
2346 offset = ConvL2X(offset);
2347 adr = make_unsafe_address(base, offset);
2348 heap_base_oop = base;
2349 val = is_store ? argument(4) : NULL;
2350 } else {
2351 Node* ptr = argument(1); // type: long
2352 ptr = ConvL2X(ptr); // adjust Java long to machine word
2353 adr = make_unsafe_address(NULL, ptr);
2354 val = is_store ? argument(3) : NULL;
2355 }
2356
2357 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2358
2359 // First guess at the value type.
2360 const Type *value_type = Type::get_const_basic_type(type);
2361
2362 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2363 // there was not enough information to nail it down.
2364 Compile::AliasType* alias_type = C->alias_type(adr_type);
2365 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2366
2367 // We will need memory barriers unless we can determine a unique
2368 // alias category for this reference. (Note: If for some reason
2369 // the barriers get omitted and the unsafe reference begins to "pollute"
2370 // the alias analysis of the rest of the graph, either Compile::can_alias
2371 // or Compile::must_alias will throw a diagnostic assert.)
2372 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2373
2374 // If we are reading the value of the referent field of a Reference
2375 // object (either by using Unsafe directly or through reflection)
2376 // then, if G1 is enabled, we need to record the referent in an
2377 // SATB log buffer using the pre-barrier mechanism.
2378 // Also we need to add memory barrier to prevent commoning reads
2379 // from this field across safepoint since GC can change its value.
2380 bool need_read_barrier = !is_native_ptr && !is_store &&
2381 offset != top() && heap_base_oop != top();
2382
2383 if (!is_store && type == T_OBJECT) {
2384 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2385 if (tjp != NULL) {
2386 value_type = tjp;
2387 }
2388 }
2389
2390 receiver = null_check(receiver);
2391 if (stopped()) {
2392 return true;
2393 }
2394 // Heap pointers get a null-check from the interpreter,
2395 // as a courtesy. However, this is not guaranteed by Unsafe,
2396 // and it is not possible to fully distinguish unintended nulls
2397 // from intended ones in this API.
2398
2399 if (is_volatile) {
2400 // We need to emit leading and trailing CPU membars (see below) in
2401 // addition to memory membars when is_volatile. This is a little
2402 // too strong, but avoids the need to insert per-alias-type
2403 // volatile membars (for stores; compare Parse::do_put_xxx), which
2404 // we cannot do effectively here because we probably only have a
2405 // rough approximation of type.
2406 need_mem_bar = true;
2407 // For Stores, place a memory ordering barrier now.
2408 if (is_store) {
2409 insert_mem_bar(Op_MemBarRelease);
2410 } else {
2411 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2412 insert_mem_bar(Op_MemBarVolatile);
2413 }
2414 }
2415 }
2416
2417 // Memory barrier to prevent normal and 'unsafe' accesses from
2418 // bypassing each other. Happens after null checks, so the
2419 // exception paths do not take memory state from the memory barrier,
2420 // so there's no problems making a strong assert about mixing users
2421 // of safe & unsafe memory.
2422 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2423
2424 if (!is_store) {
2425 Node* p = NULL;
2426 // Try to constant fold a load from a constant field
2427 ciField* field = alias_type->field();
2428 if (heap_base_oop != top() &&
2429 field != NULL && field->is_constant() && field->layout_type() == type) {
2430 // final or stable field
2431 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2432 if (con_type != NULL) {
2433 p = makecon(con_type);
2434 }
2435 }
2436 if (p == NULL) {
2437 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2438 // To be valid, unsafe loads may depend on other conditions than
2439 // the one that guards them: pin the Load node
2440 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2441 // load value
2442 switch (type) {
2443 case T_BOOLEAN:
2444 case T_CHAR:
2445 case T_BYTE:
2446 case T_SHORT:
2447 case T_INT:
2448 case T_LONG:
2449 case T_FLOAT:
2450 case T_DOUBLE:
2451 break;
2452 case T_OBJECT:
2453 if (need_read_barrier) {
2454 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2455 }
2456 break;
2457 case T_ADDRESS:
2458 // Cast to an int type.
2459 p = _gvn.transform(new CastP2XNode(NULL, p));
2460 p = ConvX2UL(p);
2461 break;
2462 default:
2463 fatal("unexpected type %d: %s", type, type2name(type));
2464 break;
2465 }
2466 }
2467 // The load node has the control of the preceding MemBarCPUOrder. All
2468 // following nodes will have the control of the MemBarCPUOrder inserted at
2469 // the end of this method. So, pushing the load onto the stack at a later
2470 // point is fine.
2471 set_result(p);
2472 } else {
2473 // place effect of store into memory
2474 switch (type) {
2475 case T_DOUBLE:
2476 val = dstore_rounding(val);
2477 break;
2478 case T_ADDRESS:
2479 // Repackage the long as a pointer.
2480 val = ConvL2X(val);
2481 val = _gvn.transform(new CastX2PNode(val));
2482 break;
2483 }
2484
2485 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2486 if (type != T_OBJECT ) {
2487 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2488 } else {
2489 // Possibly an oop being stored to Java heap or native memory
2490 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2491 // oop to Java heap.
2492 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2493 } else {
2494 // We can't tell at compile time if we are storing in the Java heap or outside
2495 // of it. So we need to emit code to conditionally do the proper type of
2496 // store.
2497
2498 IdealKit ideal(this);
2499 #define __ ideal.
2500 // QQQ who knows what probability is here??
2501 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2502 // Sync IdealKit and graphKit.
2503 sync_kit(ideal);
2504 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2505 // Update IdealKit memory.
2506 __ sync_kit(this);
2507 } __ else_(); {
2508 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2509 } __ end_if();
2510 // Final sync IdealKit and GraphKit.
2511 final_sync(ideal);
2512 #undef __
2513 }
2514 }
2515 }
2516
2517 if (is_volatile) {
2518 if (!is_store) {
2519 insert_mem_bar(Op_MemBarAcquire);
2520 } else {
2521 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2522 insert_mem_bar(Op_MemBarVolatile);
2523 }
2524 }
2525 }
2526
2527 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2528
2529 return true;
2530 }
2531
2532 //----------------------------inline_unsafe_load_store----------------------------
2533 // This method serves a couple of different customers (depending on LoadStoreKind):
2534 //
2535 // LS_cmpxchg:
2536 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2537 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2538 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2539 //
2540 // LS_xadd:
2541 // public int getAndAddInt( Object o, long offset, int delta)
2542 // public long getAndAddLong(Object o, long offset, long delta)
2543 //
2544 // LS_xchg:
2545 // int getAndSet(Object o, long offset, int newValue)
2546 // long getAndSet(Object o, long offset, long newValue)
2547 // Object getAndSet(Object o, long offset, Object newValue)
2548 //
2549 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2550 // This basic scheme here is the same as inline_unsafe_access, but
2551 // differs in enough details that combining them would make the code
2552 // overly confusing. (This is a true fact! I originally combined
2553 // them, but even I was confused by it!) As much code/comments as
2554 // possible are retained from inline_unsafe_access though to make
2555 // the correspondences clearer. - dl
2556
2557 if (callee()->is_static()) return false; // caller must have the capability!
2558
2559 #ifndef PRODUCT
2560 BasicType rtype;
2561 {
2562 ResourceMark rm;
2563 // Check the signatures.
2564 ciSignature* sig = callee()->signature();
2565 rtype = sig->return_type()->basic_type();
2566 if (kind == LS_xadd || kind == LS_xchg) {
2567 // Check the signatures.
2568 #ifdef ASSERT
2569 assert(rtype == type, "get and set must return the expected type");
2570 assert(sig->count() == 3, "get and set has 3 arguments");
2571 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2572 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2573 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2574 #endif // ASSERT
2575 } else if (kind == LS_cmpxchg) {
2576 // Check the signatures.
2577 #ifdef ASSERT
2578 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2579 assert(sig->count() == 4, "CAS has 4 arguments");
2580 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2581 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2582 #endif // ASSERT
2583 } else {
2584 ShouldNotReachHere();
2585 }
2586 }
2587 #endif //PRODUCT
2588
2589 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2590
2591 // Get arguments:
2592 Node* receiver = NULL;
2593 Node* base = NULL;
2594 Node* offset = NULL;
2595 Node* oldval = NULL;
2596 Node* newval = NULL;
2597 if (kind == LS_cmpxchg) {
2598 const bool two_slot_type = type2size[type] == 2;
2599 receiver = argument(0); // type: oop
2600 base = argument(1); // type: oop
2601 offset = argument(2); // type: long
2602 oldval = argument(4); // type: oop, int, or long
2603 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2604 } else if (kind == LS_xadd || kind == LS_xchg){
2605 receiver = argument(0); // type: oop
2606 base = argument(1); // type: oop
2607 offset = argument(2); // type: long
2608 oldval = NULL;
2609 newval = argument(4); // type: oop, int, or long
2610 }
2611
2612 // Null check receiver.
2613 receiver = null_check(receiver);
2614 if (stopped()) {
2615 return true;
2616 }
2617
2618 // Build field offset expression.
2619 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2620 // to be plain byte offsets, which are also the same as those accepted
2621 // by oopDesc::field_base.
2622 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2623 // 32-bit machines ignore the high half of long offsets
2624 offset = ConvL2X(offset);
2625 Node* adr = make_unsafe_address(base, offset);
2626 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2627
2628 // For CAS, unlike inline_unsafe_access, there seems no point in
2629 // trying to refine types. Just use the coarse types here.
2630 const Type *value_type = Type::get_const_basic_type(type);
2631 Compile::AliasType* alias_type = C->alias_type(adr_type);
2632 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2633
2634 if (kind == LS_xchg && type == T_OBJECT) {
2635 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2636 if (tjp != NULL) {
2637 value_type = tjp;
2638 }
2639 }
2640
2641 int alias_idx = C->get_alias_index(adr_type);
2642
2643 // Memory-model-wise, a LoadStore acts like a little synchronized
2644 // block, so needs barriers on each side. These don't translate
2645 // into actual barriers on most machines, but we still need rest of
2646 // compiler to respect ordering.
2647
2648 insert_mem_bar(Op_MemBarRelease);
2649 insert_mem_bar(Op_MemBarCPUOrder);
2650
2651 // 4984716: MemBars must be inserted before this
2652 // memory node in order to avoid a false
2653 // dependency which will confuse the scheduler.
2654 Node *mem = memory(alias_idx);
2655
2656 // For now, we handle only those cases that actually exist: ints,
2657 // longs, and Object. Adding others should be straightforward.
2658 Node* load_store;
2659 switch(type) {
2660 case T_INT:
2661 if (kind == LS_xadd) {
2662 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2663 } else if (kind == LS_xchg) {
2664 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2665 } else if (kind == LS_cmpxchg) {
2666 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2667 } else {
2668 ShouldNotReachHere();
2669 }
2670 break;
2671 case T_LONG:
2672 if (kind == LS_xadd) {
2673 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2674 } else if (kind == LS_xchg) {
2675 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2676 } else if (kind == LS_cmpxchg) {
2677 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2678 } else {
2679 ShouldNotReachHere();
2680 }
2681 break;
2682 case T_OBJECT:
2683 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2684 // could be delayed during Parse (for example, in adjust_map_after_if()).
2685 // Execute transformation here to avoid barrier generation in such case.
2686 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2687 newval = _gvn.makecon(TypePtr::NULL_PTR);
2688
2689 // Reference stores need a store barrier.
2690 if (kind == LS_xchg) {
2691 // If pre-barrier must execute before the oop store, old value will require do_load here.
2692 if (!can_move_pre_barrier()) {
2693 pre_barrier(true /* do_load*/,
2694 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2695 NULL /* pre_val*/,
2696 T_OBJECT);
2697 } // Else move pre_barrier to use load_store value, see below.
2698 } else if (kind == LS_cmpxchg) {
2699 // Same as for newval above:
2700 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2701 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2702 }
2703 // The only known value which might get overwritten is oldval.
2704 pre_barrier(false /* do_load */,
2705 control(), NULL, NULL, max_juint, NULL, NULL,
2706 oldval /* pre_val */,
2707 T_OBJECT);
2708 } else {
2709 ShouldNotReachHere();
2710 }
2711
2712 #ifdef _LP64
2713 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2714 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2715 if (kind == LS_xchg) {
2716 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2717 newval_enc, adr_type, value_type->make_narrowoop()));
2718 } else {
2719 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2720 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2721 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2722 newval_enc, oldval_enc));
2723 }
2724 } else
2725 #endif
2726 {
2727 if (kind == LS_xchg) {
2728 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2729 } else {
2730 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2731 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2732 }
2733 }
2734 if (kind == LS_cmpxchg) {
2735 // Emit the post barrier only when the actual store happened.
2736 // This makes sense to check only for compareAndSet that can fail to set the value.
2737 // CAS success path is marked more likely since we anticipate this is a performance
2738 // critical path, while CAS failure path can use the penalty for going through unlikely
2739 // path as backoff. Which is still better than doing a store barrier there.
2740 IdealKit ideal(this);
2741 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2742 sync_kit(ideal);
2743 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2744 ideal.sync_kit(this);
2745 } ideal.end_if();
2746 final_sync(ideal);
2747 } else {
2748 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2749 }
2750 break;
2751 default:
2752 fatal("unexpected type %d: %s", type, type2name(type));
2753 break;
2754 }
2755
2756 // SCMemProjNodes represent the memory state of a LoadStore. Their
2757 // main role is to prevent LoadStore nodes from being optimized away
2758 // when their results aren't used.
2759 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2760 set_memory(proj, alias_idx);
2761
2762 if (type == T_OBJECT && kind == LS_xchg) {
2763 #ifdef _LP64
2764 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2765 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2766 }
2767 #endif
2768 if (can_move_pre_barrier()) {
2769 // Don't need to load pre_val. The old value is returned by load_store.
2770 // The pre_barrier can execute after the xchg as long as no safepoint
2771 // gets inserted between them.
2772 pre_barrier(false /* do_load */,
2773 control(), NULL, NULL, max_juint, NULL, NULL,
2774 load_store /* pre_val */,
2775 T_OBJECT);
2776 }
2777 }
2778
2779 // Add the trailing membar surrounding the access
2780 insert_mem_bar(Op_MemBarCPUOrder);
2781 insert_mem_bar(Op_MemBarAcquire);
2782
2783 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2784 set_result(load_store);
2785 return true;
2786 }
2787
2788 //----------------------------inline_unsafe_ordered_store----------------------
2789 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2790 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2791 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2792 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2793 // This is another variant of inline_unsafe_access, differing in
2794 // that it always issues store-store ("release") barrier and ensures
2795 // store-atomicity (which only matters for "long").
2796
2797 if (callee()->is_static()) return false; // caller must have the capability!
2798
2799 #ifndef PRODUCT
2800 {
2801 ResourceMark rm;
2802 // Check the signatures.
2803 ciSignature* sig = callee()->signature();
2804 #ifdef ASSERT
2805 BasicType rtype = sig->return_type()->basic_type();
2806 assert(rtype == T_VOID, "must return void");
2807 assert(sig->count() == 3, "has 3 arguments");
2808 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2809 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2810 #endif // ASSERT
2811 }
2812 #endif //PRODUCT
2813
2814 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2815
2816 // Get arguments:
2817 Node* receiver = argument(0); // type: oop
2818 Node* base = argument(1); // type: oop
2819 Node* offset = argument(2); // type: long
2820 Node* val = argument(4); // type: oop, int, or long
2821
2822 // Null check receiver.
2823 receiver = null_check(receiver);
2824 if (stopped()) {
2825 return true;
2826 }
2827
2828 // Build field offset expression.
2829 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2830 // 32-bit machines ignore the high half of long offsets
2831 offset = ConvL2X(offset);
2832 Node* adr = make_unsafe_address(base, offset);
2833 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2834 const Type *value_type = Type::get_const_basic_type(type);
2835 Compile::AliasType* alias_type = C->alias_type(adr_type);
2836
2837 insert_mem_bar(Op_MemBarRelease);
2838 insert_mem_bar(Op_MemBarCPUOrder);
2839 // Ensure that the store is atomic for longs:
2840 const bool require_atomic_access = true;
2841 Node* store;
2842 if (type == T_OBJECT) // reference stores need a store barrier.
2843 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2844 else {
2845 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2846 }
2847 insert_mem_bar(Op_MemBarCPUOrder);
2848 return true;
2849 }
2850
2851 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2852 // Regardless of form, don't allow previous ld/st to move down,
2853 // then issue acquire, release, or volatile mem_bar.
2854 insert_mem_bar(Op_MemBarCPUOrder);
2855 switch(id) {
2856 case vmIntrinsics::_loadFence:
2857 insert_mem_bar(Op_LoadFence);
2858 return true;
2859 case vmIntrinsics::_storeFence:
2860 insert_mem_bar(Op_StoreFence);
2861 return true;
2862 case vmIntrinsics::_fullFence:
2863 insert_mem_bar(Op_MemBarVolatile);
2864 return true;
2865 default:
2866 fatal_unexpected_iid(id);
2867 return false;
2868 }
2869 }
2870
2871 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2872 if (!kls->is_Con()) {
2873 return true;
2874 }
2875 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2876 if (klsptr == NULL) {
2877 return true;
2878 }
2879 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2880 // don't need a guard for a klass that is already initialized
2881 return !ik->is_initialized();
2882 }
2883
2884 //----------------------------inline_unsafe_allocate---------------------------
2885 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2886 bool LibraryCallKit::inline_unsafe_allocate() {
2887 if (callee()->is_static()) return false; // caller must have the capability!
2888
2889 null_check_receiver(); // null-check, then ignore
2890 Node* cls = null_check(argument(1));
2891 if (stopped()) return true;
2892
2893 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2894 kls = null_check(kls);
2895 if (stopped()) return true; // argument was like int.class
2896
2897 Node* test = NULL;
2898 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2899 // Note: The argument might still be an illegal value like
2900 // Serializable.class or Object[].class. The runtime will handle it.
2901 // But we must make an explicit check for initialization.
2902 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2903 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2904 // can generate code to load it as unsigned byte.
2905 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2906 Node* bits = intcon(InstanceKlass::fully_initialized);
2907 test = _gvn.transform(new SubINode(inst, bits));
2908 // The 'test' is non-zero if we need to take a slow path.
2909 }
2910
2911 Node* obj = new_instance(kls, test);
2912 set_result(obj);
2913 return true;
2914 }
2915
2916 #ifdef TRACE_HAVE_INTRINSICS
2917 /*
2918 * oop -> myklass
2919 * myklass->trace_id |= USED
2920 * return myklass->trace_id & ~0x3
2921 */
2922 bool LibraryCallKit::inline_native_classID() {
2923 null_check_receiver(); // null-check, then ignore
2924 Node* cls = null_check(argument(1), T_OBJECT);
2925 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2926 kls = null_check(kls, T_OBJECT);
2927 ByteSize offset = TRACE_ID_OFFSET;
2928 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2929 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2930 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2931 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2932 Node* clsused = longcon(0x01l); // set the class bit
2933 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2934
2935 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2936 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2937 set_result(andl);
2938 return true;
2939 }
2940
2941 bool LibraryCallKit::inline_native_threadID() {
2942 Node* tls_ptr = NULL;
2943 Node* cur_thr = generate_current_thread(tls_ptr);
2944 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2945 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2946 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2947
2948 Node* threadid = NULL;
2949 size_t thread_id_size = OSThread::thread_id_size();
2950 if (thread_id_size == (size_t) BytesPerLong) {
2951 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2952 } else if (thread_id_size == (size_t) BytesPerInt) {
2953 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2954 } else {
2955 ShouldNotReachHere();
2956 }
2957 set_result(threadid);
2958 return true;
2959 }
2960 #endif
2961
2962 //------------------------inline_native_time_funcs--------------
2963 // inline code for System.currentTimeMillis() and System.nanoTime()
2964 // these have the same type and signature
2965 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2966 const TypeFunc* tf = OptoRuntime::void_long_Type();
2967 const TypePtr* no_memory_effects = NULL;
2968 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2969 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2970 #ifdef ASSERT
2971 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2972 assert(value_top == top(), "second value must be top");
2973 #endif
2974 set_result(value);
2975 return true;
2976 }
2977
2978 //------------------------inline_native_currentThread------------------
2979 bool LibraryCallKit::inline_native_currentThread() {
2980 Node* junk = NULL;
2981 set_result(generate_current_thread(junk));
2982 return true;
2983 }
2984
2985 //------------------------inline_native_isInterrupted------------------
2986 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2987 bool LibraryCallKit::inline_native_isInterrupted() {
2988 // Add a fast path to t.isInterrupted(clear_int):
2989 // (t == Thread.current() &&
2990 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2991 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2992 // So, in the common case that the interrupt bit is false,
2993 // we avoid making a call into the VM. Even if the interrupt bit
2994 // is true, if the clear_int argument is false, we avoid the VM call.
2995 // However, if the receiver is not currentThread, we must call the VM,
2996 // because there must be some locking done around the operation.
2997
2998 // We only go to the fast case code if we pass two guards.
2999 // Paths which do not pass are accumulated in the slow_region.
3000
3001 enum {
3002 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3003 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3004 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3005 PATH_LIMIT
3006 };
3007
3008 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3009 // out of the function.
3010 insert_mem_bar(Op_MemBarCPUOrder);
3011
3012 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3013 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3014
3015 RegionNode* slow_region = new RegionNode(1);
3016 record_for_igvn(slow_region);
3017
3018 // (a) Receiving thread must be the current thread.
3019 Node* rec_thr = argument(0);
3020 Node* tls_ptr = NULL;
3021 Node* cur_thr = generate_current_thread(tls_ptr);
3022 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3023 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3024
3025 generate_slow_guard(bol_thr, slow_region);
3026
3027 // (b) Interrupt bit on TLS must be false.
3028 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3029 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3030 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3031
3032 // Set the control input on the field _interrupted read to prevent it floating up.
3033 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3034 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3035 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3036
3037 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3038
3039 // First fast path: if (!TLS._interrupted) return false;
3040 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3041 result_rgn->init_req(no_int_result_path, false_bit);
3042 result_val->init_req(no_int_result_path, intcon(0));
3043
3044 // drop through to next case
3045 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3046
3047 #ifndef TARGET_OS_FAMILY_windows
3048 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3049 Node* clr_arg = argument(1);
3050 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3051 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3052 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3053
3054 // Second fast path: ... else if (!clear_int) return true;
3055 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3056 result_rgn->init_req(no_clear_result_path, false_arg);
3057 result_val->init_req(no_clear_result_path, intcon(1));
3058
3059 // drop through to next case
3060 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3061 #else
3062 // To return true on Windows you must read the _interrupted field
3063 // and check the the event state i.e. take the slow path.
3064 #endif // TARGET_OS_FAMILY_windows
3065
3066 // (d) Otherwise, go to the slow path.
3067 slow_region->add_req(control());
3068 set_control( _gvn.transform(slow_region));
3069
3070 if (stopped()) {
3071 // There is no slow path.
3072 result_rgn->init_req(slow_result_path, top());
3073 result_val->init_req(slow_result_path, top());
3074 } else {
3075 // non-virtual because it is a private non-static
3076 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3077
3078 Node* slow_val = set_results_for_java_call(slow_call);
3079 // this->control() comes from set_results_for_java_call
3080
3081 Node* fast_io = slow_call->in(TypeFunc::I_O);
3082 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3083
3084 // These two phis are pre-filled with copies of of the fast IO and Memory
3085 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3086 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3087
3088 result_rgn->init_req(slow_result_path, control());
3089 result_io ->init_req(slow_result_path, i_o());
3090 result_mem->init_req(slow_result_path, reset_memory());
3091 result_val->init_req(slow_result_path, slow_val);
3092
3093 set_all_memory(_gvn.transform(result_mem));
3094 set_i_o( _gvn.transform(result_io));
3095 }
3096
3097 C->set_has_split_ifs(true); // Has chance for split-if optimization
3098 set_result(result_rgn, result_val);
3099 return true;
3100 }
3101
3102 //---------------------------load_mirror_from_klass----------------------------
3103 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3104 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3105 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3106 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3107 }
3108
3109 //-----------------------load_klass_from_mirror_common-------------------------
3110 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3111 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3112 // and branch to the given path on the region.
3113 // If never_see_null, take an uncommon trap on null, so we can optimistically
3114 // compile for the non-null case.
3115 // If the region is NULL, force never_see_null = true.
3116 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3117 bool never_see_null,
3118 RegionNode* region,
3119 int null_path,
3120 int offset) {
3121 if (region == NULL) never_see_null = true;
3122 Node* p = basic_plus_adr(mirror, offset);
3123 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3124 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3125 Node* null_ctl = top();
3126 kls = null_check_oop(kls, &null_ctl, never_see_null);
3127 if (region != NULL) {
3128 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3129 region->init_req(null_path, null_ctl);
3130 } else {
3131 assert(null_ctl == top(), "no loose ends");
3132 }
3133 return kls;
3134 }
3135
3136 //--------------------(inline_native_Class_query helpers)---------------------
3137 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3138 // Fall through if (mods & mask) == bits, take the guard otherwise.
3139 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3140 // Branch around if the given klass has the given modifier bit set.
3141 // Like generate_guard, adds a new path onto the region.
3142 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3143 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3144 Node* mask = intcon(modifier_mask);
3145 Node* bits = intcon(modifier_bits);
3146 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3147 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3148 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3149 return generate_fair_guard(bol, region);
3150 }
3151 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3152 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3153 }
3154
3155 //-------------------------inline_native_Class_query-------------------
3156 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3157 const Type* return_type = TypeInt::BOOL;
3158 Node* prim_return_value = top(); // what happens if it's a primitive class?
3159 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3160 bool expect_prim = false; // most of these guys expect to work on refs
3161
3162 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3163
3164 Node* mirror = argument(0);
3165 Node* obj = top();
3166
3167 switch (id) {
3168 case vmIntrinsics::_isInstance:
3169 // nothing is an instance of a primitive type
3170 prim_return_value = intcon(0);
3171 obj = argument(1);
3172 break;
3173 case vmIntrinsics::_getModifiers:
3174 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3175 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3176 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3177 break;
3178 case vmIntrinsics::_isInterface:
3179 prim_return_value = intcon(0);
3180 break;
3181 case vmIntrinsics::_isArray:
3182 prim_return_value = intcon(0);
3183 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3184 break;
3185 case vmIntrinsics::_isPrimitive:
3186 prim_return_value = intcon(1);
3187 expect_prim = true; // obviously
3188 break;
3189 case vmIntrinsics::_getSuperclass:
3190 prim_return_value = null();
3191 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3192 break;
3193 case vmIntrinsics::_getClassAccessFlags:
3194 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3195 return_type = TypeInt::INT; // not bool! 6297094
3196 break;
3197 default:
3198 fatal_unexpected_iid(id);
3199 break;
3200 }
3201
3202 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3203 if (mirror_con == NULL) return false; // cannot happen?
3204
3205 #ifndef PRODUCT
3206 if (C->print_intrinsics() || C->print_inlining()) {
3207 ciType* k = mirror_con->java_mirror_type();
3208 if (k) {
3209 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3210 k->print_name();
3211 tty->cr();
3212 }
3213 }
3214 #endif
3215
3216 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3217 RegionNode* region = new RegionNode(PATH_LIMIT);
3218 record_for_igvn(region);
3219 PhiNode* phi = new PhiNode(region, return_type);
3220
3221 // The mirror will never be null of Reflection.getClassAccessFlags, however
3222 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3223 // if it is. See bug 4774291.
3224
3225 // For Reflection.getClassAccessFlags(), the null check occurs in
3226 // the wrong place; see inline_unsafe_access(), above, for a similar
3227 // situation.
3228 mirror = null_check(mirror);
3229 // If mirror or obj is dead, only null-path is taken.
3230 if (stopped()) return true;
3231
3232 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3233
3234 // Now load the mirror's klass metaobject, and null-check it.
3235 // Side-effects region with the control path if the klass is null.
3236 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3237 // If kls is null, we have a primitive mirror.
3238 phi->init_req(_prim_path, prim_return_value);
3239 if (stopped()) { set_result(region, phi); return true; }
3240 bool safe_for_replace = (region->in(_prim_path) == top());
3241
3242 Node* p; // handy temp
3243 Node* null_ctl;
3244
3245 // Now that we have the non-null klass, we can perform the real query.
3246 // For constant classes, the query will constant-fold in LoadNode::Value.
3247 Node* query_value = top();
3248 switch (id) {
3249 case vmIntrinsics::_isInstance:
3250 // nothing is an instance of a primitive type
3251 query_value = gen_instanceof(obj, kls, safe_for_replace);
3252 break;
3253
3254 case vmIntrinsics::_getModifiers:
3255 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3256 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3257 break;
3258
3259 case vmIntrinsics::_isInterface:
3260 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3261 if (generate_interface_guard(kls, region) != NULL)
3262 // A guard was added. If the guard is taken, it was an interface.
3263 phi->add_req(intcon(1));
3264 // If we fall through, it's a plain class.
3265 query_value = intcon(0);
3266 break;
3267
3268 case vmIntrinsics::_isArray:
3269 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3270 if (generate_array_guard(kls, region) != NULL)
3271 // A guard was added. If the guard is taken, it was an array.
3272 phi->add_req(intcon(1));
3273 // If we fall through, it's a plain class.
3274 query_value = intcon(0);
3275 break;
3276
3277 case vmIntrinsics::_isPrimitive:
3278 query_value = intcon(0); // "normal" path produces false
3279 break;
3280
3281 case vmIntrinsics::_getSuperclass:
3282 // The rules here are somewhat unfortunate, but we can still do better
3283 // with random logic than with a JNI call.
3284 // Interfaces store null or Object as _super, but must report null.
3285 // Arrays store an intermediate super as _super, but must report Object.
3286 // Other types can report the actual _super.
3287 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3288 if (generate_interface_guard(kls, region) != NULL)
3289 // A guard was added. If the guard is taken, it was an interface.
3290 phi->add_req(null());
3291 if (generate_array_guard(kls, region) != NULL)
3292 // A guard was added. If the guard is taken, it was an array.
3293 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3294 // If we fall through, it's a plain class. Get its _super.
3295 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3296 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3297 null_ctl = top();
3298 kls = null_check_oop(kls, &null_ctl);
3299 if (null_ctl != top()) {
3300 // If the guard is taken, Object.superClass is null (both klass and mirror).
3301 region->add_req(null_ctl);
3302 phi ->add_req(null());
3303 }
3304 if (!stopped()) {
3305 query_value = load_mirror_from_klass(kls);
3306 }
3307 break;
3308
3309 case vmIntrinsics::_getClassAccessFlags:
3310 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3311 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3312 break;
3313
3314 default:
3315 fatal_unexpected_iid(id);
3316 break;
3317 }
3318
3319 // Fall-through is the normal case of a query to a real class.
3320 phi->init_req(1, query_value);
3321 region->init_req(1, control());
3322
3323 C->set_has_split_ifs(true); // Has chance for split-if optimization
3324 set_result(region, phi);
3325 return true;
3326 }
3327
3328 //-------------------------inline_Class_cast-------------------
3329 bool LibraryCallKit::inline_Class_cast() {
3330 Node* mirror = argument(0); // Class
3331 Node* obj = argument(1);
3332 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3333 if (mirror_con == NULL) {
3334 return false; // dead path (mirror->is_top()).
3335 }
3336 if (obj == NULL || obj->is_top()) {
3337 return false; // dead path
3338 }
3339 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3340
3341 // First, see if Class.cast() can be folded statically.
3342 // java_mirror_type() returns non-null for compile-time Class constants.
3343 ciType* tm = mirror_con->java_mirror_type();
3344 if (tm != NULL && tm->is_klass() &&
3345 tp != NULL && tp->klass() != NULL) {
3346 if (!tp->klass()->is_loaded()) {
3347 // Don't use intrinsic when class is not loaded.
3348 return false;
3349 } else {
3350 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3351 if (static_res == Compile::SSC_always_true) {
3352 // isInstance() is true - fold the code.
3353 set_result(obj);
3354 return true;
3355 } else if (static_res == Compile::SSC_always_false) {
3356 // Don't use intrinsic, have to throw ClassCastException.
3357 // If the reference is null, the non-intrinsic bytecode will
3358 // be optimized appropriately.
3359 return false;
3360 }
3361 }
3362 }
3363
3364 // Bailout intrinsic and do normal inlining if exception path is frequent.
3365 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3366 return false;
3367 }
3368
3369 // Generate dynamic checks.
3370 // Class.cast() is java implementation of _checkcast bytecode.
3371 // Do checkcast (Parse::do_checkcast()) optimizations here.
3372
3373 mirror = null_check(mirror);
3374 // If mirror is dead, only null-path is taken.
3375 if (stopped()) {
3376 return true;
3377 }
3378
3379 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3380 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3381 RegionNode* region = new RegionNode(PATH_LIMIT);
3382 record_for_igvn(region);
3383
3384 // Now load the mirror's klass metaobject, and null-check it.
3385 // If kls is null, we have a primitive mirror and
3386 // nothing is an instance of a primitive type.
3387 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3388
3389 Node* res = top();
3390 if (!stopped()) {
3391 Node* bad_type_ctrl = top();
3392 // Do checkcast optimizations.
3393 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3394 region->init_req(_bad_type_path, bad_type_ctrl);
3395 }
3396 if (region->in(_prim_path) != top() ||
3397 region->in(_bad_type_path) != top()) {
3398 // Let Interpreter throw ClassCastException.
3399 PreserveJVMState pjvms(this);
3400 set_control(_gvn.transform(region));
3401 uncommon_trap(Deoptimization::Reason_intrinsic,
3402 Deoptimization::Action_maybe_recompile);
3403 }
3404 if (!stopped()) {
3405 set_result(res);
3406 }
3407 return true;
3408 }
3409
3410
3411 //--------------------------inline_native_subtype_check------------------------
3412 // This intrinsic takes the JNI calls out of the heart of
3413 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3414 bool LibraryCallKit::inline_native_subtype_check() {
3415 // Pull both arguments off the stack.
3416 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3417 args[0] = argument(0);
3418 args[1] = argument(1);
3419 Node* klasses[2]; // corresponding Klasses: superk, subk
3420 klasses[0] = klasses[1] = top();
3421
3422 enum {
3423 // A full decision tree on {superc is prim, subc is prim}:
3424 _prim_0_path = 1, // {P,N} => false
3425 // {P,P} & superc!=subc => false
3426 _prim_same_path, // {P,P} & superc==subc => true
3427 _prim_1_path, // {N,P} => false
3428 _ref_subtype_path, // {N,N} & subtype check wins => true
3429 _both_ref_path, // {N,N} & subtype check loses => false
3430 PATH_LIMIT
3431 };
3432
3433 RegionNode* region = new RegionNode(PATH_LIMIT);
3434 Node* phi = new PhiNode(region, TypeInt::BOOL);
3435 record_for_igvn(region);
3436
3437 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3438 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3439 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3440
3441 // First null-check both mirrors and load each mirror's klass metaobject.
3442 int which_arg;
3443 for (which_arg = 0; which_arg <= 1; which_arg++) {
3444 Node* arg = args[which_arg];
3445 arg = null_check(arg);
3446 if (stopped()) break;
3447 args[which_arg] = arg;
3448
3449 Node* p = basic_plus_adr(arg, class_klass_offset);
3450 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3451 klasses[which_arg] = _gvn.transform(kls);
3452 }
3453
3454 // Having loaded both klasses, test each for null.
3455 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3456 for (which_arg = 0; which_arg <= 1; which_arg++) {
3457 Node* kls = klasses[which_arg];
3458 Node* null_ctl = top();
3459 kls = null_check_oop(kls, &null_ctl, never_see_null);
3460 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3461 region->init_req(prim_path, null_ctl);
3462 if (stopped()) break;
3463 klasses[which_arg] = kls;
3464 }
3465
3466 if (!stopped()) {
3467 // now we have two reference types, in klasses[0..1]
3468 Node* subk = klasses[1]; // the argument to isAssignableFrom
3469 Node* superk = klasses[0]; // the receiver
3470 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3471 // now we have a successful reference subtype check
3472 region->set_req(_ref_subtype_path, control());
3473 }
3474
3475 // If both operands are primitive (both klasses null), then
3476 // we must return true when they are identical primitives.
3477 // It is convenient to test this after the first null klass check.
3478 set_control(region->in(_prim_0_path)); // go back to first null check
3479 if (!stopped()) {
3480 // Since superc is primitive, make a guard for the superc==subc case.
3481 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3482 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3483 generate_guard(bol_eq, region, PROB_FAIR);
3484 if (region->req() == PATH_LIMIT+1) {
3485 // A guard was added. If the added guard is taken, superc==subc.
3486 region->swap_edges(PATH_LIMIT, _prim_same_path);
3487 region->del_req(PATH_LIMIT);
3488 }
3489 region->set_req(_prim_0_path, control()); // Not equal after all.
3490 }
3491
3492 // these are the only paths that produce 'true':
3493 phi->set_req(_prim_same_path, intcon(1));
3494 phi->set_req(_ref_subtype_path, intcon(1));
3495
3496 // pull together the cases:
3497 assert(region->req() == PATH_LIMIT, "sane region");
3498 for (uint i = 1; i < region->req(); i++) {
3499 Node* ctl = region->in(i);
3500 if (ctl == NULL || ctl == top()) {
3501 region->set_req(i, top());
3502 phi ->set_req(i, top());
3503 } else if (phi->in(i) == NULL) {
3504 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3505 }
3506 }
3507
3508 set_control(_gvn.transform(region));
3509 set_result(_gvn.transform(phi));
3510 return true;
3511 }
3512
3513 //---------------------generate_array_guard_common------------------------
3514 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3515 bool obj_array, bool not_array) {
3516
3517 if (stopped()) {
3518 return NULL;
3519 }
3520
3521 // If obj_array/non_array==false/false:
3522 // Branch around if the given klass is in fact an array (either obj or prim).
3523 // If obj_array/non_array==false/true:
3524 // Branch around if the given klass is not an array klass of any kind.
3525 // If obj_array/non_array==true/true:
3526 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3527 // If obj_array/non_array==true/false:
3528 // Branch around if the kls is an oop array (Object[] or subtype)
3529 //
3530 // Like generate_guard, adds a new path onto the region.
3531 jint layout_con = 0;
3532 Node* layout_val = get_layout_helper(kls, layout_con);
3533 if (layout_val == NULL) {
3534 bool query = (obj_array
3535 ? Klass::layout_helper_is_objArray(layout_con)
3536 : Klass::layout_helper_is_array(layout_con));
3537 if (query == not_array) {
3538 return NULL; // never a branch
3539 } else { // always a branch
3540 Node* always_branch = control();
3541 if (region != NULL)
3542 region->add_req(always_branch);
3543 set_control(top());
3544 return always_branch;
3545 }
3546 }
3547 // Now test the correct condition.
3548 jint nval = (obj_array
3549 ? ((jint)Klass::_lh_array_tag_type_value
3550 << Klass::_lh_array_tag_shift)
3551 : Klass::_lh_neutral_value);
3552 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3553 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3554 // invert the test if we are looking for a non-array
3555 if (not_array) btest = BoolTest(btest).negate();
3556 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3557 return generate_fair_guard(bol, region);
3558 }
3559
3560
3561 //-----------------------inline_native_newArray--------------------------
3562 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3563 bool LibraryCallKit::inline_native_newArray() {
3564 Node* mirror = argument(0);
3565 Node* count_val = argument(1);
3566
3567 mirror = null_check(mirror);
3568 // If mirror or obj is dead, only null-path is taken.
3569 if (stopped()) return true;
3570
3571 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3572 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3573 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3574 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3575 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3576
3577 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3578 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3579 result_reg, _slow_path);
3580 Node* normal_ctl = control();
3581 Node* no_array_ctl = result_reg->in(_slow_path);
3582
3583 // Generate code for the slow case. We make a call to newArray().
3584 set_control(no_array_ctl);
3585 if (!stopped()) {
3586 // Either the input type is void.class, or else the
3587 // array klass has not yet been cached. Either the
3588 // ensuing call will throw an exception, or else it
3589 // will cache the array klass for next time.
3590 PreserveJVMState pjvms(this);
3591 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3592 Node* slow_result = set_results_for_java_call(slow_call);
3593 // this->control() comes from set_results_for_java_call
3594 result_reg->set_req(_slow_path, control());
3595 result_val->set_req(_slow_path, slow_result);
3596 result_io ->set_req(_slow_path, i_o());
3597 result_mem->set_req(_slow_path, reset_memory());
3598 }
3599
3600 set_control(normal_ctl);
3601 if (!stopped()) {
3602 // Normal case: The array type has been cached in the java.lang.Class.
3603 // The following call works fine even if the array type is polymorphic.
3604 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3605 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3606 result_reg->init_req(_normal_path, control());
3607 result_val->init_req(_normal_path, obj);
3608 result_io ->init_req(_normal_path, i_o());
3609 result_mem->init_req(_normal_path, reset_memory());
3610 }
3611
3612 // Return the combined state.
3613 set_i_o( _gvn.transform(result_io) );
3614 set_all_memory( _gvn.transform(result_mem));
3615
3616 C->set_has_split_ifs(true); // Has chance for split-if optimization
3617 set_result(result_reg, result_val);
3618 return true;
3619 }
3620
3621 //----------------------inline_native_getLength--------------------------
3622 // public static native int java.lang.reflect.Array.getLength(Object array);
3623 bool LibraryCallKit::inline_native_getLength() {
3624 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3625
3626 Node* array = null_check(argument(0));
3627 // If array is dead, only null-path is taken.
3628 if (stopped()) return true;
3629
3630 // Deoptimize if it is a non-array.
3631 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3632
3633 if (non_array != NULL) {
3634 PreserveJVMState pjvms(this);
3635 set_control(non_array);
3636 uncommon_trap(Deoptimization::Reason_intrinsic,
3637 Deoptimization::Action_maybe_recompile);
3638 }
3639
3640 // If control is dead, only non-array-path is taken.
3641 if (stopped()) return true;
3642
3643 // The works fine even if the array type is polymorphic.
3644 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3645 Node* result = load_array_length(array);
3646
3647 C->set_has_split_ifs(true); // Has chance for split-if optimization
3648 set_result(result);
3649 return true;
3650 }
3651
3652 //------------------------inline_array_copyOf----------------------------
3653 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3654 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3655 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3656 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3657
3658 // Get the arguments.
3659 Node* original = argument(0);
3660 Node* start = is_copyOfRange? argument(1): intcon(0);
3661 Node* end = is_copyOfRange? argument(2): argument(1);
3662 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3663
3664 Node* newcopy;
3665
3666 // Set the original stack and the reexecute bit for the interpreter to reexecute
3667 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3668 { PreserveReexecuteState preexecs(this);
3669 jvms()->set_should_reexecute(true);
3670
3671 array_type_mirror = null_check(array_type_mirror);
3672 original = null_check(original);
3673
3674 // Check if a null path was taken unconditionally.
3675 if (stopped()) return true;
3676
3677 Node* orig_length = load_array_length(original);
3678
3679 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3680 klass_node = null_check(klass_node);
3681
3682 RegionNode* bailout = new RegionNode(1);
3683 record_for_igvn(bailout);
3684
3685 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3686 // Bail out if that is so.
3687 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3688 if (not_objArray != NULL) {
3689 // Improve the klass node's type from the new optimistic assumption:
3690 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3691 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3692 Node* cast = new CastPPNode(klass_node, akls);
3693 cast->init_req(0, control());
3694 klass_node = _gvn.transform(cast);
3695 }
3696
3697 // Bail out if either start or end is negative.
3698 generate_negative_guard(start, bailout, &start);
3699 generate_negative_guard(end, bailout, &end);
3700
3701 Node* length = end;
3702 if (_gvn.type(start) != TypeInt::ZERO) {
3703 length = _gvn.transform(new SubINode(end, start));
3704 }
3705
3706 // Bail out if length is negative.
3707 // Without this the new_array would throw
3708 // NegativeArraySizeException but IllegalArgumentException is what
3709 // should be thrown
3710 generate_negative_guard(length, bailout, &length);
3711
3712 if (bailout->req() > 1) {
3713 PreserveJVMState pjvms(this);
3714 set_control(_gvn.transform(bailout));
3715 uncommon_trap(Deoptimization::Reason_intrinsic,
3716 Deoptimization::Action_maybe_recompile);
3717 }
3718
3719 if (!stopped()) {
3720 // How many elements will we copy from the original?
3721 // The answer is MinI(orig_length - start, length).
3722 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3723 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3724
3725 // Generate a direct call to the right arraycopy function(s).
3726 // We know the copy is disjoint but we might not know if the
3727 // oop stores need checking.
3728 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3729 // This will fail a store-check if x contains any non-nulls.
3730
3731 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3732 // loads/stores but it is legal only if we're sure the
3733 // Arrays.copyOf would succeed. So we need all input arguments
3734 // to the copyOf to be validated, including that the copy to the
3735 // new array won't trigger an ArrayStoreException. That subtype
3736 // check can be optimized if we know something on the type of
3737 // the input array from type speculation.
3738 if (_gvn.type(klass_node)->singleton()) {
3739 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3740 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3741
3742 int test = C->static_subtype_check(superk, subk);
3743 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3744 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3745 if (t_original->speculative_type() != NULL) {
3746 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3747 }
3748 }
3749 }
3750
3751 bool validated = false;
3752 // Reason_class_check rather than Reason_intrinsic because we
3753 // want to intrinsify even if this traps.
3754 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3755 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3756 klass_node);
3757
3758 if (not_subtype_ctrl != top()) {
3759 PreserveJVMState pjvms(this);
3760 set_control(not_subtype_ctrl);
3761 uncommon_trap(Deoptimization::Reason_class_check,
3762 Deoptimization::Action_make_not_entrant);
3763 assert(stopped(), "Should be stopped");
3764 }
3765 validated = true;
3766 }
3767
3768 if (!stopped()) {
3769 newcopy = new_array(klass_node, length, 0); // no arguments to push
3770
3771 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3772 load_object_klass(original), klass_node);
3773 if (!is_copyOfRange) {
3774 ac->set_copyof(validated);
3775 } else {
3776 ac->set_copyofrange(validated);
3777 }
3778 Node* n = _gvn.transform(ac);
3779 if (n == ac) {
3780 ac->connect_outputs(this);
3781 } else {
3782 assert(validated, "shouldn't transform if all arguments not validated");
3783 set_all_memory(n);
3784 }
3785 }
3786 }
3787 } // original reexecute is set back here
3788
3789 C->set_has_split_ifs(true); // Has chance for split-if optimization
3790 if (!stopped()) {
3791 set_result(newcopy);
3792 }
3793 return true;
3794 }
3795
3796
3797 //----------------------generate_virtual_guard---------------------------
3798 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3799 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3800 RegionNode* slow_region) {
3801 ciMethod* method = callee();
3802 int vtable_index = method->vtable_index();
3803 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3804 "bad index %d", vtable_index);
3805 // Get the Method* out of the appropriate vtable entry.
3806 int entry_offset = (InstanceKlass::vtable_start_offset() +
3807 vtable_index*vtableEntry::size()) * wordSize +
3808 vtableEntry::method_offset_in_bytes();
3809 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3810 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3811
3812 // Compare the target method with the expected method (e.g., Object.hashCode).
3813 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3814
3815 Node* native_call = makecon(native_call_addr);
3816 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3817 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3818
3819 return generate_slow_guard(test_native, slow_region);
3820 }
3821
3822 //-----------------------generate_method_call----------------------------
3823 // Use generate_method_call to make a slow-call to the real
3824 // method if the fast path fails. An alternative would be to
3825 // use a stub like OptoRuntime::slow_arraycopy_Java.
3826 // This only works for expanding the current library call,
3827 // not another intrinsic. (E.g., don't use this for making an
3828 // arraycopy call inside of the copyOf intrinsic.)
3829 CallJavaNode*
3830 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3831 // When compiling the intrinsic method itself, do not use this technique.
3832 guarantee(callee() != C->method(), "cannot make slow-call to self");
3833
3834 ciMethod* method = callee();
3835 // ensure the JVMS we have will be correct for this call
3836 guarantee(method_id == method->intrinsic_id(), "must match");
3837
3838 const TypeFunc* tf = TypeFunc::make(method);
3839 CallJavaNode* slow_call;
3840 if (is_static) {
3841 assert(!is_virtual, "");
3842 slow_call = new CallStaticJavaNode(C, tf,
3843 SharedRuntime::get_resolve_static_call_stub(),
3844 method, bci());
3845 } else if (is_virtual) {
3846 null_check_receiver();
3847 int vtable_index = Method::invalid_vtable_index;
3848 if (UseInlineCaches) {
3849 // Suppress the vtable call
3850 } else {
3851 // hashCode and clone are not a miranda methods,
3852 // so the vtable index is fixed.
3853 // No need to use the linkResolver to get it.
3854 vtable_index = method->vtable_index();
3855 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3856 "bad index %d", vtable_index);
3857 }
3858 slow_call = new CallDynamicJavaNode(tf,
3859 SharedRuntime::get_resolve_virtual_call_stub(),
3860 method, vtable_index, bci());
3861 } else { // neither virtual nor static: opt_virtual
3862 null_check_receiver();
3863 slow_call = new CallStaticJavaNode(C, tf,
3864 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3865 method, bci());
3866 slow_call->set_optimized_virtual(true);
3867 }
3868 set_arguments_for_java_call(slow_call);
3869 set_edges_for_java_call(slow_call);
3870 return slow_call;
3871 }
3872
3873
3874 /**
3875 * Build special case code for calls to hashCode on an object. This call may
3876 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3877 * slightly different code.
3878 */
3879 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3880 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3881 assert(!(is_virtual && is_static), "either virtual, special, or static");
3882
3883 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3884
3885 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3886 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3887 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3888 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3889 Node* obj = NULL;
3890 if (!is_static) {
3891 // Check for hashing null object
3892 obj = null_check_receiver();
3893 if (stopped()) return true; // unconditionally null
3894 result_reg->init_req(_null_path, top());
3895 result_val->init_req(_null_path, top());
3896 } else {
3897 // Do a null check, and return zero if null.
3898 // System.identityHashCode(null) == 0
3899 obj = argument(0);
3900 Node* null_ctl = top();
3901 obj = null_check_oop(obj, &null_ctl);
3902 result_reg->init_req(_null_path, null_ctl);
3903 result_val->init_req(_null_path, _gvn.intcon(0));
3904 }
3905
3906 // Unconditionally null? Then return right away.
3907 if (stopped()) {
3908 set_control( result_reg->in(_null_path));
3909 if (!stopped())
3910 set_result(result_val->in(_null_path));
3911 return true;
3912 }
3913
3914 // We only go to the fast case code if we pass a number of guards. The
3915 // paths which do not pass are accumulated in the slow_region.
3916 RegionNode* slow_region = new RegionNode(1);
3917 record_for_igvn(slow_region);
3918
3919 // If this is a virtual call, we generate a funny guard. We pull out
3920 // the vtable entry corresponding to hashCode() from the target object.
3921 // If the target method which we are calling happens to be the native
3922 // Object hashCode() method, we pass the guard. We do not need this
3923 // guard for non-virtual calls -- the caller is known to be the native
3924 // Object hashCode().
3925 if (is_virtual) {
3926 // After null check, get the object's klass.
3927 Node* obj_klass = load_object_klass(obj);
3928 generate_virtual_guard(obj_klass, slow_region);
3929 }
3930
3931 // Get the header out of the object, use LoadMarkNode when available
3932 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3933 // The control of the load must be NULL. Otherwise, the load can move before
3934 // the null check after castPP removal.
3935 Node* no_ctrl = NULL;
3936 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3937
3938 // Test the header to see if it is unlocked.
3939 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3940 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3941 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3942 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3943 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3944
3945 generate_slow_guard(test_unlocked, slow_region);
3946
3947 // Get the hash value and check to see that it has been properly assigned.
3948 // We depend on hash_mask being at most 32 bits and avoid the use of
3949 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3950 // vm: see markOop.hpp.
3951 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3952 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3953 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3954 // This hack lets the hash bits live anywhere in the mark object now, as long
3955 // as the shift drops the relevant bits into the low 32 bits. Note that
3956 // Java spec says that HashCode is an int so there's no point in capturing
3957 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3958 hshifted_header = ConvX2I(hshifted_header);
3959 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3960
3961 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3962 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3963 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3964
3965 generate_slow_guard(test_assigned, slow_region);
3966
3967 Node* init_mem = reset_memory();
3968 // fill in the rest of the null path:
3969 result_io ->init_req(_null_path, i_o());
3970 result_mem->init_req(_null_path, init_mem);
3971
3972 result_val->init_req(_fast_path, hash_val);
3973 result_reg->init_req(_fast_path, control());
3974 result_io ->init_req(_fast_path, i_o());
3975 result_mem->init_req(_fast_path, init_mem);
3976
3977 // Generate code for the slow case. We make a call to hashCode().
3978 set_control(_gvn.transform(slow_region));
3979 if (!stopped()) {
3980 // No need for PreserveJVMState, because we're using up the present state.
3981 set_all_memory(init_mem);
3982 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3983 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3984 Node* slow_result = set_results_for_java_call(slow_call);
3985 // this->control() comes from set_results_for_java_call
3986 result_reg->init_req(_slow_path, control());
3987 result_val->init_req(_slow_path, slow_result);
3988 result_io ->set_req(_slow_path, i_o());
3989 result_mem ->set_req(_slow_path, reset_memory());
3990 }
3991
3992 // Return the combined state.
3993 set_i_o( _gvn.transform(result_io) );
3994 set_all_memory( _gvn.transform(result_mem));
3995
3996 set_result(result_reg, result_val);
3997 return true;
3998 }
3999
4000 //---------------------------inline_native_getClass----------------------------
4001 // public final native Class<?> java.lang.Object.getClass();
4002 //
4003 // Build special case code for calls to getClass on an object.
4004 bool LibraryCallKit::inline_native_getClass() {
4005 Node* obj = null_check_receiver();
4006 if (stopped()) return true;
4007 set_result(load_mirror_from_klass(load_object_klass(obj)));
4008 return true;
4009 }
4010
4011 //-----------------inline_native_Reflection_getCallerClass---------------------
4012 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4013 //
4014 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4015 //
4016 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4017 // in that it must skip particular security frames and checks for
4018 // caller sensitive methods.
4019 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4020 #ifndef PRODUCT
4021 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4022 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4023 }
4024 #endif
4025
4026 if (!jvms()->has_method()) {
4027 #ifndef PRODUCT
4028 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4029 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4030 }
4031 #endif
4032 return false;
4033 }
4034
4035 // Walk back up the JVM state to find the caller at the required
4036 // depth.
4037 JVMState* caller_jvms = jvms();
4038
4039 // Cf. JVM_GetCallerClass
4040 // NOTE: Start the loop at depth 1 because the current JVM state does
4041 // not include the Reflection.getCallerClass() frame.
4042 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4043 ciMethod* m = caller_jvms->method();
4044 switch (n) {
4045 case 0:
4046 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4047 break;
4048 case 1:
4049 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4050 if (!m->caller_sensitive()) {
4051 #ifndef PRODUCT
4052 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4053 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4054 }
4055 #endif
4056 return false; // bail-out; let JVM_GetCallerClass do the work
4057 }
4058 break;
4059 default:
4060 if (!m->is_ignored_by_security_stack_walk()) {
4061 // We have reached the desired frame; return the holder class.
4062 // Acquire method holder as java.lang.Class and push as constant.
4063 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4064 ciInstance* caller_mirror = caller_klass->java_mirror();
4065 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4066
4067 #ifndef PRODUCT
4068 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4069 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());
4070 tty->print_cr(" JVM state at this point:");
4071 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4072 ciMethod* m = jvms()->of_depth(i)->method();
4073 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4074 }
4075 }
4076 #endif
4077 return true;
4078 }
4079 break;
4080 }
4081 }
4082
4083 #ifndef PRODUCT
4084 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4085 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4086 tty->print_cr(" JVM state at this point:");
4087 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4088 ciMethod* m = jvms()->of_depth(i)->method();
4089 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4090 }
4091 }
4092 #endif
4093
4094 return false; // bail-out; let JVM_GetCallerClass do the work
4095 }
4096
4097 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4098 Node* arg = argument(0);
4099 Node* result;
4100
4101 switch (id) {
4102 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4103 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4104 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4105 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4106
4107 case vmIntrinsics::_doubleToLongBits: {
4108 // two paths (plus control) merge in a wood
4109 RegionNode *r = new RegionNode(3);
4110 Node *phi = new PhiNode(r, TypeLong::LONG);
4111
4112 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4113 // Build the boolean node
4114 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4115
4116 // Branch either way.
4117 // NaN case is less traveled, which makes all the difference.
4118 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4119 Node *opt_isnan = _gvn.transform(ifisnan);
4120 assert( opt_isnan->is_If(), "Expect an IfNode");
4121 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4122 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4123
4124 set_control(iftrue);
4125
4126 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4127 Node *slow_result = longcon(nan_bits); // return NaN
4128 phi->init_req(1, _gvn.transform( slow_result ));
4129 r->init_req(1, iftrue);
4130
4131 // Else fall through
4132 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4133 set_control(iffalse);
4134
4135 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4136 r->init_req(2, iffalse);
4137
4138 // Post merge
4139 set_control(_gvn.transform(r));
4140 record_for_igvn(r);
4141
4142 C->set_has_split_ifs(true); // Has chance for split-if optimization
4143 result = phi;
4144 assert(result->bottom_type()->isa_long(), "must be");
4145 break;
4146 }
4147
4148 case vmIntrinsics::_floatToIntBits: {
4149 // two paths (plus control) merge in a wood
4150 RegionNode *r = new RegionNode(3);
4151 Node *phi = new PhiNode(r, TypeInt::INT);
4152
4153 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4154 // Build the boolean node
4155 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4156
4157 // Branch either way.
4158 // NaN case is less traveled, which makes all the difference.
4159 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4160 Node *opt_isnan = _gvn.transform(ifisnan);
4161 assert( opt_isnan->is_If(), "Expect an IfNode");
4162 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4163 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4164
4165 set_control(iftrue);
4166
4167 static const jint nan_bits = 0x7fc00000;
4168 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4169 phi->init_req(1, _gvn.transform( slow_result ));
4170 r->init_req(1, iftrue);
4171
4172 // Else fall through
4173 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4174 set_control(iffalse);
4175
4176 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4177 r->init_req(2, iffalse);
4178
4179 // Post merge
4180 set_control(_gvn.transform(r));
4181 record_for_igvn(r);
4182
4183 C->set_has_split_ifs(true); // Has chance for split-if optimization
4184 result = phi;
4185 assert(result->bottom_type()->isa_int(), "must be");
4186 break;
4187 }
4188
4189 default:
4190 fatal_unexpected_iid(id);
4191 break;
4192 }
4193 set_result(_gvn.transform(result));
4194 return true;
4195 }
4196
4197 #ifdef _LP64
4198 #define XTOP ,top() /*additional argument*/
4199 #else //_LP64
4200 #define XTOP /*no additional argument*/
4201 #endif //_LP64
4202
4203 //----------------------inline_unsafe_copyMemory-------------------------
4204 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4205 bool LibraryCallKit::inline_unsafe_copyMemory() {
4206 if (callee()->is_static()) return false; // caller must have the capability!
4207 null_check_receiver(); // null-check receiver
4208 if (stopped()) return true;
4209
4210 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4211
4212 Node* src_ptr = argument(1); // type: oop
4213 Node* src_off = ConvL2X(argument(2)); // type: long
4214 Node* dst_ptr = argument(4); // type: oop
4215 Node* dst_off = ConvL2X(argument(5)); // type: long
4216 Node* size = ConvL2X(argument(7)); // type: long
4217
4218 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4219 "fieldOffset must be byte-scaled");
4220
4221 Node* src = make_unsafe_address(src_ptr, src_off);
4222 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4223
4224 // Conservatively insert a memory barrier on all memory slices.
4225 // Do not let writes of the copy source or destination float below the copy.
4226 insert_mem_bar(Op_MemBarCPUOrder);
4227
4228 // Call it. Note that the length argument is not scaled.
4229 make_runtime_call(RC_LEAF|RC_NO_FP,
4230 OptoRuntime::fast_arraycopy_Type(),
4231 StubRoutines::unsafe_arraycopy(),
4232 "unsafe_arraycopy",
4233 TypeRawPtr::BOTTOM,
4234 src, dst, size XTOP);
4235
4236 // Do not let reads of the copy destination float above the copy.
4237 insert_mem_bar(Op_MemBarCPUOrder);
4238
4239 return true;
4240 }
4241
4242 //------------------------clone_coping-----------------------------------
4243 // Helper function for inline_native_clone.
4244 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4245 assert(obj_size != NULL, "");
4246 Node* raw_obj = alloc_obj->in(1);
4247 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4248
4249 AllocateNode* alloc = NULL;
4250 if (ReduceBulkZeroing) {
4251 // We will be completely responsible for initializing this object -
4252 // mark Initialize node as complete.
4253 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4254 // The object was just allocated - there should be no any stores!
4255 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4256 // Mark as complete_with_arraycopy so that on AllocateNode
4257 // expansion, we know this AllocateNode is initialized by an array
4258 // copy and a StoreStore barrier exists after the array copy.
4259 alloc->initialization()->set_complete_with_arraycopy();
4260 }
4261
4262 // Copy the fastest available way.
4263 // TODO: generate fields copies for small objects instead.
4264 Node* src = obj;
4265 Node* dest = alloc_obj;
4266 Node* size = _gvn.transform(obj_size);
4267
4268 // Exclude the header but include array length to copy by 8 bytes words.
4269 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4270 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4271 instanceOopDesc::base_offset_in_bytes();
4272 // base_off:
4273 // 8 - 32-bit VM
4274 // 12 - 64-bit VM, compressed klass
4275 // 16 - 64-bit VM, normal klass
4276 if (base_off % BytesPerLong != 0) {
4277 assert(UseCompressedClassPointers, "");
4278 if (is_array) {
4279 // Exclude length to copy by 8 bytes words.
4280 base_off += sizeof(int);
4281 } else {
4282 // Include klass to copy by 8 bytes words.
4283 base_off = instanceOopDesc::klass_offset_in_bytes();
4284 }
4285 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4286 }
4287 src = basic_plus_adr(src, base_off);
4288 dest = basic_plus_adr(dest, base_off);
4289
4290 // Compute the length also, if needed:
4291 Node* countx = size;
4292 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4293 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4294
4295 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4296
4297 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4298 ac->set_clonebasic();
4299 Node* n = _gvn.transform(ac);
4300 if (n == ac) {
4301 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4302 } else {
4303 set_all_memory(n);
4304 }
4305
4306 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4307 if (card_mark) {
4308 assert(!is_array, "");
4309 // Put in store barrier for any and all oops we are sticking
4310 // into this object. (We could avoid this if we could prove
4311 // that the object type contains no oop fields at all.)
4312 Node* no_particular_value = NULL;
4313 Node* no_particular_field = NULL;
4314 int raw_adr_idx = Compile::AliasIdxRaw;
4315 post_barrier(control(),
4316 memory(raw_adr_type),
4317 alloc_obj,
4318 no_particular_field,
4319 raw_adr_idx,
4320 no_particular_value,
4321 T_OBJECT,
4322 false);
4323 }
4324
4325 // Do not let reads from the cloned object float above the arraycopy.
4326 if (alloc != NULL) {
4327 // Do not let stores that initialize this object be reordered with
4328 // a subsequent store that would make this object accessible by
4329 // other threads.
4330 // Record what AllocateNode this StoreStore protects so that
4331 // escape analysis can go from the MemBarStoreStoreNode to the
4332 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4333 // based on the escape status of the AllocateNode.
4334 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4335 } else {
4336 insert_mem_bar(Op_MemBarCPUOrder);
4337 }
4338 }
4339
4340 //------------------------inline_native_clone----------------------------
4341 // protected native Object java.lang.Object.clone();
4342 //
4343 // Here are the simple edge cases:
4344 // null receiver => normal trap
4345 // virtual and clone was overridden => slow path to out-of-line clone
4346 // not cloneable or finalizer => slow path to out-of-line Object.clone
4347 //
4348 // The general case has two steps, allocation and copying.
4349 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4350 //
4351 // Copying also has two cases, oop arrays and everything else.
4352 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4353 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4354 //
4355 // These steps fold up nicely if and when the cloned object's klass
4356 // can be sharply typed as an object array, a type array, or an instance.
4357 //
4358 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4359 PhiNode* result_val;
4360
4361 // Set the reexecute bit for the interpreter to reexecute
4362 // the bytecode that invokes Object.clone if deoptimization happens.
4363 { PreserveReexecuteState preexecs(this);
4364 jvms()->set_should_reexecute(true);
4365
4366 Node* obj = null_check_receiver();
4367 if (stopped()) return true;
4368
4369 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4370
4371 // If we are going to clone an instance, we need its exact type to
4372 // know the number and types of fields to convert the clone to
4373 // loads/stores. Maybe a speculative type can help us.
4374 if (!obj_type->klass_is_exact() &&
4375 obj_type->speculative_type() != NULL &&
4376 obj_type->speculative_type()->is_instance_klass()) {
4377 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4378 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4379 !spec_ik->has_injected_fields()) {
4380 ciKlass* k = obj_type->klass();
4381 if (!k->is_instance_klass() ||
4382 k->as_instance_klass()->is_interface() ||
4383 k->as_instance_klass()->has_subklass()) {
4384 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4385 }
4386 }
4387 }
4388
4389 Node* obj_klass = load_object_klass(obj);
4390 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4391 const TypeOopPtr* toop = ((tklass != NULL)
4392 ? tklass->as_instance_type()
4393 : TypeInstPtr::NOTNULL);
4394
4395 // Conservatively insert a memory barrier on all memory slices.
4396 // Do not let writes into the original float below the clone.
4397 insert_mem_bar(Op_MemBarCPUOrder);
4398
4399 // paths into result_reg:
4400 enum {
4401 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4402 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4403 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4404 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4405 PATH_LIMIT
4406 };
4407 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4408 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4409 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4410 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4411 record_for_igvn(result_reg);
4412
4413 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4414 int raw_adr_idx = Compile::AliasIdxRaw;
4415
4416 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4417 if (array_ctl != NULL) {
4418 // It's an array.
4419 PreserveJVMState pjvms(this);
4420 set_control(array_ctl);
4421 Node* obj_length = load_array_length(obj);
4422 Node* obj_size = NULL;
4423 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4424
4425 if (!use_ReduceInitialCardMarks()) {
4426 // If it is an oop array, it requires very special treatment,
4427 // because card marking is required on each card of the array.
4428 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4429 if (is_obja != NULL) {
4430 PreserveJVMState pjvms2(this);
4431 set_control(is_obja);
4432 // Generate a direct call to the right arraycopy function(s).
4433 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4434 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4435 ac->set_cloneoop();
4436 Node* n = _gvn.transform(ac);
4437 assert(n == ac, "cannot disappear");
4438 ac->connect_outputs(this);
4439
4440 result_reg->init_req(_objArray_path, control());
4441 result_val->init_req(_objArray_path, alloc_obj);
4442 result_i_o ->set_req(_objArray_path, i_o());
4443 result_mem ->set_req(_objArray_path, reset_memory());
4444 }
4445 }
4446 // Otherwise, there are no card marks to worry about.
4447 // (We can dispense with card marks if we know the allocation
4448 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4449 // causes the non-eden paths to take compensating steps to
4450 // simulate a fresh allocation, so that no further
4451 // card marks are required in compiled code to initialize
4452 // the object.)
4453
4454 if (!stopped()) {
4455 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4456
4457 // Present the results of the copy.
4458 result_reg->init_req(_array_path, control());
4459 result_val->init_req(_array_path, alloc_obj);
4460 result_i_o ->set_req(_array_path, i_o());
4461 result_mem ->set_req(_array_path, reset_memory());
4462 }
4463 }
4464
4465 // We only go to the instance fast case code if we pass a number of guards.
4466 // The paths which do not pass are accumulated in the slow_region.
4467 RegionNode* slow_region = new RegionNode(1);
4468 record_for_igvn(slow_region);
4469 if (!stopped()) {
4470 // It's an instance (we did array above). Make the slow-path tests.
4471 // If this is a virtual call, we generate a funny guard. We grab
4472 // the vtable entry corresponding to clone() from the target object.
4473 // If the target method which we are calling happens to be the
4474 // Object clone() method, we pass the guard. We do not need this
4475 // guard for non-virtual calls; the caller is known to be the native
4476 // Object clone().
4477 if (is_virtual) {
4478 generate_virtual_guard(obj_klass, slow_region);
4479 }
4480
4481 // The object must be cloneable and must not have a finalizer.
4482 // Both of these conditions may be checked in a single test.
4483 // We could optimize the cloneable test further, but we don't care.
4484 generate_access_flags_guard(obj_klass,
4485 // Test both conditions:
4486 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4487 // Must be cloneable but not finalizer:
4488 JVM_ACC_IS_CLONEABLE,
4489 slow_region);
4490 }
4491
4492 if (!stopped()) {
4493 // It's an instance, and it passed the slow-path tests.
4494 PreserveJVMState pjvms(this);
4495 Node* obj_size = NULL;
4496 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4497 // is reexecuted if deoptimization occurs and there could be problems when merging
4498 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4499 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4500
4501 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4502
4503 // Present the results of the slow call.
4504 result_reg->init_req(_instance_path, control());
4505 result_val->init_req(_instance_path, alloc_obj);
4506 result_i_o ->set_req(_instance_path, i_o());
4507 result_mem ->set_req(_instance_path, reset_memory());
4508 }
4509
4510 // Generate code for the slow case. We make a call to clone().
4511 set_control(_gvn.transform(slow_region));
4512 if (!stopped()) {
4513 PreserveJVMState pjvms(this);
4514 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4515 Node* slow_result = set_results_for_java_call(slow_call);
4516 // this->control() comes from set_results_for_java_call
4517 result_reg->init_req(_slow_path, control());
4518 result_val->init_req(_slow_path, slow_result);
4519 result_i_o ->set_req(_slow_path, i_o());
4520 result_mem ->set_req(_slow_path, reset_memory());
4521 }
4522
4523 // Return the combined state.
4524 set_control( _gvn.transform(result_reg));
4525 set_i_o( _gvn.transform(result_i_o));
4526 set_all_memory( _gvn.transform(result_mem));
4527 } // original reexecute is set back here
4528
4529 set_result(_gvn.transform(result_val));
4530 return true;
4531 }
4532
4533 // If we have a tighly coupled allocation, the arraycopy may take care
4534 // of the array initialization. If one of the guards we insert between
4535 // the allocation and the arraycopy causes a deoptimization, an
4536 // unitialized array will escape the compiled method. To prevent that
4537 // we set the JVM state for uncommon traps between the allocation and
4538 // the arraycopy to the state before the allocation so, in case of
4539 // deoptimization, we'll reexecute the allocation and the
4540 // initialization.
4541 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4542 if (alloc != NULL) {
4543 ciMethod* trap_method = alloc->jvms()->method();
4544 int trap_bci = alloc->jvms()->bci();
4545
4546 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4547 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4548 // Make sure there's no store between the allocation and the
4549 // arraycopy otherwise visible side effects could be rexecuted
4550 // in case of deoptimization and cause incorrect execution.
4551 bool no_interfering_store = true;
4552 Node* mem = alloc->in(TypeFunc::Memory);
4553 if (mem->is_MergeMem()) {
4554 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4555 Node* n = mms.memory();
4556 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4557 assert(n->is_Store(), "what else?");
4558 no_interfering_store = false;
4559 break;
4560 }
4561 }
4562 } else {
4563 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4564 Node* n = mms.memory();
4565 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4566 assert(n->is_Store(), "what else?");
4567 no_interfering_store = false;
4568 break;
4569 }
4570 }
4571 }
4572
4573 if (no_interfering_store) {
4574 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4575 uint size = alloc->req();
4576 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4577 old_jvms->set_map(sfpt);
4578 for (uint i = 0; i < size; i++) {
4579 sfpt->init_req(i, alloc->in(i));
4580 }
4581 // re-push array length for deoptimization
4582 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4583 old_jvms->set_sp(old_jvms->sp()+1);
4584 old_jvms->set_monoff(old_jvms->monoff()+1);
4585 old_jvms->set_scloff(old_jvms->scloff()+1);
4586 old_jvms->set_endoff(old_jvms->endoff()+1);
4587 old_jvms->set_should_reexecute(true);
4588
4589 sfpt->set_i_o(map()->i_o());
4590 sfpt->set_memory(map()->memory());
4591 sfpt->set_control(map()->control());
4592
4593 JVMState* saved_jvms = jvms();
4594 saved_reexecute_sp = _reexecute_sp;
4595
4596 set_jvms(sfpt->jvms());
4597 _reexecute_sp = jvms()->sp();
4598
4599 return saved_jvms;
4600 }
4601 }
4602 }
4603 return NULL;
4604 }
4605
4606 // In case of a deoptimization, we restart execution at the
4607 // allocation, allocating a new array. We would leave an uninitialized
4608 // array in the heap that GCs wouldn't expect. Move the allocation
4609 // after the traps so we don't allocate the array if we
4610 // deoptimize. This is possible because tightly_coupled_allocation()
4611 // guarantees there's no observer of the allocated array at this point
4612 // and the control flow is simple enough.
4613 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4614 if (saved_jvms != NULL && !stopped()) {
4615 assert(alloc != NULL, "only with a tightly coupled allocation");
4616 // restore JVM state to the state at the arraycopy
4617 saved_jvms->map()->set_control(map()->control());
4618 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4619 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4620 // If we've improved the types of some nodes (null check) while
4621 // emitting the guards, propagate them to the current state
4622 map()->replaced_nodes().apply(saved_jvms->map());
4623 set_jvms(saved_jvms);
4624 _reexecute_sp = saved_reexecute_sp;
4625
4626 // Remove the allocation from above the guards
4627 CallProjections callprojs;
4628 alloc->extract_projections(&callprojs, true);
4629 InitializeNode* init = alloc->initialization();
4630 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4631 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4632 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4633 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4634
4635 // move the allocation here (after the guards)
4636 _gvn.hash_delete(alloc);
4637 alloc->set_req(TypeFunc::Control, control());
4638 alloc->set_req(TypeFunc::I_O, i_o());
4639 Node *mem = reset_memory();
4640 set_all_memory(mem);
4641 alloc->set_req(TypeFunc::Memory, mem);
4642 set_control(init->proj_out(TypeFunc::Control));
4643 set_i_o(callprojs.fallthrough_ioproj);
4644
4645 // Update memory as done in GraphKit::set_output_for_allocation()
4646 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4647 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4648 if (ary_type->isa_aryptr() && length_type != NULL) {
4649 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4650 }
4651 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4652 int elemidx = C->get_alias_index(telemref);
4653 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4654 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4655
4656 Node* allocx = _gvn.transform(alloc);
4657 assert(allocx == alloc, "where has the allocation gone?");
4658 assert(dest->is_CheckCastPP(), "not an allocation result?");
4659
4660 _gvn.hash_delete(dest);
4661 dest->set_req(0, control());
4662 Node* destx = _gvn.transform(dest);
4663 assert(destx == dest, "where has the allocation result gone?");
4664 }
4665 }
4666
4667
4668 //------------------------------inline_arraycopy-----------------------
4669 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4670 // Object dest, int destPos,
4671 // int length);
4672 bool LibraryCallKit::inline_arraycopy() {
4673 // Get the arguments.
4674 Node* src = argument(0); // type: oop
4675 Node* src_offset = argument(1); // type: int
4676 Node* dest = argument(2); // type: oop
4677 Node* dest_offset = argument(3); // type: int
4678 Node* length = argument(4); // type: int
4679
4680
4681 // Check for allocation before we add nodes that would confuse
4682 // tightly_coupled_allocation()
4683 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4684
4685 int saved_reexecute_sp = -1;
4686 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4687 // See arraycopy_restore_alloc_state() comment
4688 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4689 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4690 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4691 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4692
4693 // The following tests must be performed
4694 // (1) src and dest are arrays.
4695 // (2) src and dest arrays must have elements of the same BasicType
4696 // (3) src and dest must not be null.
4697 // (4) src_offset must not be negative.
4698 // (5) dest_offset must not be negative.
4699 // (6) length must not be negative.
4700 // (7) src_offset + length must not exceed length of src.
4701 // (8) dest_offset + length must not exceed length of dest.
4702 // (9) each element of an oop array must be assignable
4703
4704 // (3) src and dest must not be null.
4705 // always do this here because we need the JVM state for uncommon traps
4706 Node* null_ctl = top();
4707 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4708 assert(null_ctl->is_top(), "no null control here");
4709 dest = null_check(dest, T_ARRAY);
4710
4711 if (!can_emit_guards) {
4712 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4713 // guards but the arraycopy node could still take advantage of a
4714 // tightly allocated allocation. tightly_coupled_allocation() is
4715 // called again to make sure it takes the null check above into
4716 // account: the null check is mandatory and if it caused an
4717 // uncommon trap to be emitted then the allocation can't be
4718 // considered tightly coupled in this context.
4719 alloc = tightly_coupled_allocation(dest, NULL);
4720 }
4721
4722 bool validated = false;
4723
4724 const Type* src_type = _gvn.type(src);
4725 const Type* dest_type = _gvn.type(dest);
4726 const TypeAryPtr* top_src = src_type->isa_aryptr();
4727 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4728
4729 // Do we have the type of src?
4730 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4731 // Do we have the type of dest?
4732 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4733 // Is the type for src from speculation?
4734 bool src_spec = false;
4735 // Is the type for dest from speculation?
4736 bool dest_spec = false;
4737
4738 if ((!has_src || !has_dest) && can_emit_guards) {
4739 // We don't have sufficient type information, let's see if
4740 // speculative types can help. We need to have types for both src
4741 // and dest so that it pays off.
4742
4743 // Do we already have or could we have type information for src
4744 bool could_have_src = has_src;
4745 // Do we already have or could we have type information for dest
4746 bool could_have_dest = has_dest;
4747
4748 ciKlass* src_k = NULL;
4749 if (!has_src) {
4750 src_k = src_type->speculative_type_not_null();
4751 if (src_k != NULL && src_k->is_array_klass()) {
4752 could_have_src = true;
4753 }
4754 }
4755
4756 ciKlass* dest_k = NULL;
4757 if (!has_dest) {
4758 dest_k = dest_type->speculative_type_not_null();
4759 if (dest_k != NULL && dest_k->is_array_klass()) {
4760 could_have_dest = true;
4761 }
4762 }
4763
4764 if (could_have_src && could_have_dest) {
4765 // This is going to pay off so emit the required guards
4766 if (!has_src) {
4767 src = maybe_cast_profiled_obj(src, src_k, true);
4768 src_type = _gvn.type(src);
4769 top_src = src_type->isa_aryptr();
4770 has_src = (top_src != NULL && top_src->klass() != NULL);
4771 src_spec = true;
4772 }
4773 if (!has_dest) {
4774 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4775 dest_type = _gvn.type(dest);
4776 top_dest = dest_type->isa_aryptr();
4777 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4778 dest_spec = true;
4779 }
4780 }
4781 }
4782
4783 if (has_src && has_dest && can_emit_guards) {
4784 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4785 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4786 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4787 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4788
4789 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4790 // If both arrays are object arrays then having the exact types
4791 // for both will remove the need for a subtype check at runtime
4792 // before the call and may make it possible to pick a faster copy
4793 // routine (without a subtype check on every element)
4794 // Do we have the exact type of src?
4795 bool could_have_src = src_spec;
4796 // Do we have the exact type of dest?
4797 bool could_have_dest = dest_spec;
4798 ciKlass* src_k = top_src->klass();
4799 ciKlass* dest_k = top_dest->klass();
4800 if (!src_spec) {
4801 src_k = src_type->speculative_type_not_null();
4802 if (src_k != NULL && src_k->is_array_klass()) {
4803 could_have_src = true;
4804 }
4805 }
4806 if (!dest_spec) {
4807 dest_k = dest_type->speculative_type_not_null();
4808 if (dest_k != NULL && dest_k->is_array_klass()) {
4809 could_have_dest = true;
4810 }
4811 }
4812 if (could_have_src && could_have_dest) {
4813 // If we can have both exact types, emit the missing guards
4814 if (could_have_src && !src_spec) {
4815 src = maybe_cast_profiled_obj(src, src_k, true);
4816 }
4817 if (could_have_dest && !dest_spec) {
4818 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4819 }
4820 }
4821 }
4822 }
4823
4824 ciMethod* trap_method = method();
4825 int trap_bci = bci();
4826 if (saved_jvms != NULL) {
4827 trap_method = alloc->jvms()->method();
4828 trap_bci = alloc->jvms()->bci();
4829 }
4830
4831 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4832 can_emit_guards &&
4833 !src->is_top() && !dest->is_top()) {
4834 // validate arguments: enables transformation the ArrayCopyNode
4835 validated = true;
4836
4837 RegionNode* slow_region = new RegionNode(1);
4838 record_for_igvn(slow_region);
4839
4840 // (1) src and dest are arrays.
4841 generate_non_array_guard(load_object_klass(src), slow_region);
4842 generate_non_array_guard(load_object_klass(dest), slow_region);
4843
4844 // (2) src and dest arrays must have elements of the same BasicType
4845 // done at macro expansion or at Ideal transformation time
4846
4847 // (4) src_offset must not be negative.
4848 generate_negative_guard(src_offset, slow_region);
4849
4850 // (5) dest_offset must not be negative.
4851 generate_negative_guard(dest_offset, slow_region);
4852
4853 // (7) src_offset + length must not exceed length of src.
4854 generate_limit_guard(src_offset, length,
4855 load_array_length(src),
4856 slow_region);
4857
4858 // (8) dest_offset + length must not exceed length of dest.
4859 generate_limit_guard(dest_offset, length,
4860 load_array_length(dest),
4861 slow_region);
4862
4863 // (9) each element of an oop array must be assignable
4864 Node* src_klass = load_object_klass(src);
4865 Node* dest_klass = load_object_klass(dest);
4866 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4867
4868 if (not_subtype_ctrl != top()) {
4869 PreserveJVMState pjvms(this);
4870 set_control(not_subtype_ctrl);
4871 uncommon_trap(Deoptimization::Reason_intrinsic,
4872 Deoptimization::Action_make_not_entrant);
4873 assert(stopped(), "Should be stopped");
4874 }
4875 {
4876 PreserveJVMState pjvms(this);
4877 set_control(_gvn.transform(slow_region));
4878 uncommon_trap(Deoptimization::Reason_intrinsic,
4879 Deoptimization::Action_make_not_entrant);
4880 assert(stopped(), "Should be stopped");
4881 }
4882 }
4883
4884 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4885
4886 if (stopped()) {
4887 return true;
4888 }
4889
4890 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4891 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4892 // so the compiler has a chance to eliminate them: during macro expansion,
4893 // we have to set their control (CastPP nodes are eliminated).
4894 load_object_klass(src), load_object_klass(dest),
4895 load_array_length(src), load_array_length(dest));
4896
4897 ac->set_arraycopy(validated);
4898
4899 Node* n = _gvn.transform(ac);
4900 if (n == ac) {
4901 ac->connect_outputs(this);
4902 } else {
4903 assert(validated, "shouldn't transform if all arguments not validated");
4904 set_all_memory(n);
4905 }
4906
4907 return true;
4908 }
4909
4910
4911 // Helper function which determines if an arraycopy immediately follows
4912 // an allocation, with no intervening tests or other escapes for the object.
4913 AllocateArrayNode*
4914 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4915 RegionNode* slow_region) {
4916 if (stopped()) return NULL; // no fast path
4917 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4918
4919 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4920 if (alloc == NULL) return NULL;
4921
4922 Node* rawmem = memory(Compile::AliasIdxRaw);
4923 // Is the allocation's memory state untouched?
4924 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4925 // Bail out if there have been raw-memory effects since the allocation.
4926 // (Example: There might have been a call or safepoint.)
4927 return NULL;
4928 }
4929 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4930 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4931 return NULL;
4932 }
4933
4934 // There must be no unexpected observers of this allocation.
4935 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4936 Node* obs = ptr->fast_out(i);
4937 if (obs != this->map()) {
4938 return NULL;
4939 }
4940 }
4941
4942 // This arraycopy must unconditionally follow the allocation of the ptr.
4943 Node* alloc_ctl = ptr->in(0);
4944 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4945
4946 Node* ctl = control();
4947 while (ctl != alloc_ctl) {
4948 // There may be guards which feed into the slow_region.
4949 // Any other control flow means that we might not get a chance
4950 // to finish initializing the allocated object.
4951 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4952 IfNode* iff = ctl->in(0)->as_If();
4953 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4954 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4955 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4956 ctl = iff->in(0); // This test feeds the known slow_region.
4957 continue;
4958 }
4959 // One more try: Various low-level checks bottom out in
4960 // uncommon traps. If the debug-info of the trap omits
4961 // any reference to the allocation, as we've already
4962 // observed, then there can be no objection to the trap.
4963 bool found_trap = false;
4964 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4965 Node* obs = not_ctl->fast_out(j);
4966 if (obs->in(0) == not_ctl && obs->is_Call() &&
4967 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4968 found_trap = true; break;
4969 }
4970 }
4971 if (found_trap) {
4972 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4973 continue;
4974 }
4975 }
4976 return NULL;
4977 }
4978
4979 // If we get this far, we have an allocation which immediately
4980 // precedes the arraycopy, and we can take over zeroing the new object.
4981 // The arraycopy will finish the initialization, and provide
4982 // a new control state to which we will anchor the destination pointer.
4983
4984 return alloc;
4985 }
4986
4987 //-------------inline_encodeISOArray-----------------------------------
4988 // encode char[] to byte[] in ISO_8859_1
4989 bool LibraryCallKit::inline_encodeISOArray() {
4990 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4991 // no receiver since it is static method
4992 Node *src = argument(0);
4993 Node *src_offset = argument(1);
4994 Node *dst = argument(2);
4995 Node *dst_offset = argument(3);
4996 Node *length = argument(4);
4997
4998 const Type* src_type = src->Value(&_gvn);
4999 const Type* dst_type = dst->Value(&_gvn);
5000 const TypeAryPtr* top_src = src_type->isa_aryptr();
5001 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5002 if (top_src == NULL || top_src->klass() == NULL ||
5003 top_dest == NULL || top_dest->klass() == NULL) {
5004 // failed array check
5005 return false;
5006 }
5007
5008 // Figure out the size and type of the elements we will be copying.
5009 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5010 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5011 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5012 return false;
5013 }
5014 Node* src_start = array_element_address(src, src_offset, src_elem);
5015 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5016 // 'src_start' points to src array + scaled offset
5017 // 'dst_start' points to dst array + scaled offset
5018
5019 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5020 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5021 enc = _gvn.transform(enc);
5022 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5023 set_memory(res_mem, mtype);
5024 set_result(enc);
5025 return true;
5026 }
5027
5028 //-------------inline_multiplyToLen-----------------------------------
5029 bool LibraryCallKit::inline_multiplyToLen() {
5030 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5031
5032 address stubAddr = StubRoutines::multiplyToLen();
5033 if (stubAddr == NULL) {
5034 return false; // Intrinsic's stub is not implemented on this platform
5035 }
5036 const char* stubName = "multiplyToLen";
5037
5038 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5039
5040 // no receiver because it is a static method
5041 Node* x = argument(0);
5042 Node* xlen = argument(1);
5043 Node* y = argument(2);
5044 Node* ylen = argument(3);
5045 Node* z = argument(4);
5046
5047 const Type* x_type = x->Value(&_gvn);
5048 const Type* y_type = y->Value(&_gvn);
5049 const TypeAryPtr* top_x = x_type->isa_aryptr();
5050 const TypeAryPtr* top_y = y_type->isa_aryptr();
5051 if (top_x == NULL || top_x->klass() == NULL ||
5052 top_y == NULL || top_y->klass() == NULL) {
5053 // failed array check
5054 return false;
5055 }
5056
5057 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5058 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5059 if (x_elem != T_INT || y_elem != T_INT) {
5060 return false;
5061 }
5062
5063 // Set the original stack and the reexecute bit for the interpreter to reexecute
5064 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5065 // on the return from z array allocation in runtime.
5066 { PreserveReexecuteState preexecs(this);
5067 jvms()->set_should_reexecute(true);
5068
5069 Node* x_start = array_element_address(x, intcon(0), x_elem);
5070 Node* y_start = array_element_address(y, intcon(0), y_elem);
5071 // 'x_start' points to x array + scaled xlen
5072 // 'y_start' points to y array + scaled ylen
5073
5074 // Allocate the result array
5075 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5076 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5077 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5078
5079 IdealKit ideal(this);
5080
5081 #define __ ideal.
5082 Node* one = __ ConI(1);
5083 Node* zero = __ ConI(0);
5084 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5085 __ set(need_alloc, zero);
5086 __ set(z_alloc, z);
5087 __ if_then(z, BoolTest::eq, null()); {
5088 __ increment (need_alloc, one);
5089 } __ else_(); {
5090 // Update graphKit memory and control from IdealKit.
5091 sync_kit(ideal);
5092 Node* zlen_arg = load_array_length(z);
5093 // Update IdealKit memory and control from graphKit.
5094 __ sync_kit(this);
5095 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5096 __ increment (need_alloc, one);
5097 } __ end_if();
5098 } __ end_if();
5099
5100 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5101 // Update graphKit memory and control from IdealKit.
5102 sync_kit(ideal);
5103 Node * narr = new_array(klass_node, zlen, 1);
5104 // Update IdealKit memory and control from graphKit.
5105 __ sync_kit(this);
5106 __ set(z_alloc, narr);
5107 } __ end_if();
5108
5109 sync_kit(ideal);
5110 z = __ value(z_alloc);
5111 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5112 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5113 // Final sync IdealKit and GraphKit.
5114 final_sync(ideal);
5115 #undef __
5116
5117 Node* z_start = array_element_address(z, intcon(0), T_INT);
5118
5119 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5120 OptoRuntime::multiplyToLen_Type(),
5121 stubAddr, stubName, TypePtr::BOTTOM,
5122 x_start, xlen, y_start, ylen, z_start, zlen);
5123 } // original reexecute is set back here
5124
5125 C->set_has_split_ifs(true); // Has chance for split-if optimization
5126 set_result(z);
5127 return true;
5128 }
5129
5130 //-------------inline_squareToLen------------------------------------
5131 bool LibraryCallKit::inline_squareToLen() {
5132 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5133
5134 address stubAddr = StubRoutines::squareToLen();
5135 if (stubAddr == NULL) {
5136 return false; // Intrinsic's stub is not implemented on this platform
5137 }
5138 const char* stubName = "squareToLen";
5139
5140 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5141
5142 Node* x = argument(0);
5143 Node* len = argument(1);
5144 Node* z = argument(2);
5145 Node* zlen = argument(3);
5146
5147 const Type* x_type = x->Value(&_gvn);
5148 const Type* z_type = z->Value(&_gvn);
5149 const TypeAryPtr* top_x = x_type->isa_aryptr();
5150 const TypeAryPtr* top_z = z_type->isa_aryptr();
5151 if (top_x == NULL || top_x->klass() == NULL ||
5152 top_z == NULL || top_z->klass() == NULL) {
5153 // failed array check
5154 return false;
5155 }
5156
5157 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5158 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5159 if (x_elem != T_INT || z_elem != T_INT) {
5160 return false;
5161 }
5162
5163
5164 Node* x_start = array_element_address(x, intcon(0), x_elem);
5165 Node* z_start = array_element_address(z, intcon(0), z_elem);
5166
5167 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5168 OptoRuntime::squareToLen_Type(),
5169 stubAddr, stubName, TypePtr::BOTTOM,
5170 x_start, len, z_start, zlen);
5171
5172 set_result(z);
5173 return true;
5174 }
5175
5176 //-------------inline_mulAdd------------------------------------------
5177 bool LibraryCallKit::inline_mulAdd() {
5178 assert(UseMulAddIntrinsic, "not implementated on this platform");
5179
5180 address stubAddr = StubRoutines::mulAdd();
5181 if (stubAddr == NULL) {
5182 return false; // Intrinsic's stub is not implemented on this platform
5183 }
5184 const char* stubName = "mulAdd";
5185
5186 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5187
5188 Node* out = argument(0);
5189 Node* in = argument(1);
5190 Node* offset = argument(2);
5191 Node* len = argument(3);
5192 Node* k = argument(4);
5193
5194 const Type* out_type = out->Value(&_gvn);
5195 const Type* in_type = in->Value(&_gvn);
5196 const TypeAryPtr* top_out = out_type->isa_aryptr();
5197 const TypeAryPtr* top_in = in_type->isa_aryptr();
5198 if (top_out == NULL || top_out->klass() == NULL ||
5199 top_in == NULL || top_in->klass() == NULL) {
5200 // failed array check
5201 return false;
5202 }
5203
5204 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5205 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5206 if (out_elem != T_INT || in_elem != T_INT) {
5207 return false;
5208 }
5209
5210 Node* outlen = load_array_length(out);
5211 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5212 Node* out_start = array_element_address(out, intcon(0), out_elem);
5213 Node* in_start = array_element_address(in, intcon(0), in_elem);
5214
5215 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5216 OptoRuntime::mulAdd_Type(),
5217 stubAddr, stubName, TypePtr::BOTTOM,
5218 out_start,in_start, new_offset, len, k);
5219 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5220 set_result(result);
5221 return true;
5222 }
5223
5224 //-------------inline_montgomeryMultiply-----------------------------------
5225 bool LibraryCallKit::inline_montgomeryMultiply() {
5226 address stubAddr = StubRoutines::montgomeryMultiply();
5227 if (stubAddr == NULL) {
5228 return false; // Intrinsic's stub is not implemented on this platform
5229 }
5230
5231 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5232 const char* stubName = "montgomery_square";
5233
5234 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5235
5236 Node* a = argument(0);
5237 Node* b = argument(1);
5238 Node* n = argument(2);
5239 Node* len = argument(3);
5240 Node* inv = argument(4);
5241 Node* m = argument(6);
5242
5243 const Type* a_type = a->Value(&_gvn);
5244 const TypeAryPtr* top_a = a_type->isa_aryptr();
5245 const Type* b_type = b->Value(&_gvn);
5246 const TypeAryPtr* top_b = b_type->isa_aryptr();
5247 const Type* n_type = a->Value(&_gvn);
5248 const TypeAryPtr* top_n = n_type->isa_aryptr();
5249 const Type* m_type = a->Value(&_gvn);
5250 const TypeAryPtr* top_m = m_type->isa_aryptr();
5251 if (top_a == NULL || top_a->klass() == NULL ||
5252 top_b == NULL || top_b->klass() == NULL ||
5253 top_n == NULL || top_n->klass() == NULL ||
5254 top_m == NULL || top_m->klass() == NULL) {
5255 // failed array check
5256 return false;
5257 }
5258
5259 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5260 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5261 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5262 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5263 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5264 return false;
5265 }
5266
5267 // Make the call
5268 {
5269 Node* a_start = array_element_address(a, intcon(0), a_elem);
5270 Node* b_start = array_element_address(b, intcon(0), b_elem);
5271 Node* n_start = array_element_address(n, intcon(0), n_elem);
5272 Node* m_start = array_element_address(m, intcon(0), m_elem);
5273
5274 Node* call = make_runtime_call(RC_LEAF,
5275 OptoRuntime::montgomeryMultiply_Type(),
5276 stubAddr, stubName, TypePtr::BOTTOM,
5277 a_start, b_start, n_start, len, inv, top(),
5278 m_start);
5279 set_result(m);
5280 }
5281
5282 return true;
5283 }
5284
5285 bool LibraryCallKit::inline_montgomerySquare() {
5286 address stubAddr = StubRoutines::montgomerySquare();
5287 if (stubAddr == NULL) {
5288 return false; // Intrinsic's stub is not implemented on this platform
5289 }
5290
5291 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5292 const char* stubName = "montgomery_square";
5293
5294 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5295
5296 Node* a = argument(0);
5297 Node* n = argument(1);
5298 Node* len = argument(2);
5299 Node* inv = argument(3);
5300 Node* m = argument(5);
5301
5302 const Type* a_type = a->Value(&_gvn);
5303 const TypeAryPtr* top_a = a_type->isa_aryptr();
5304 const Type* n_type = a->Value(&_gvn);
5305 const TypeAryPtr* top_n = n_type->isa_aryptr();
5306 const Type* m_type = a->Value(&_gvn);
5307 const TypeAryPtr* top_m = m_type->isa_aryptr();
5308 if (top_a == NULL || top_a->klass() == NULL ||
5309 top_n == NULL || top_n->klass() == NULL ||
5310 top_m == NULL || top_m->klass() == NULL) {
5311 // failed array check
5312 return false;
5313 }
5314
5315 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5316 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5317 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5318 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5319 return false;
5320 }
5321
5322 // Make the call
5323 {
5324 Node* a_start = array_element_address(a, intcon(0), a_elem);
5325 Node* n_start = array_element_address(n, intcon(0), n_elem);
5326 Node* m_start = array_element_address(m, intcon(0), m_elem);
5327
5328 Node* call = make_runtime_call(RC_LEAF,
5329 OptoRuntime::montgomerySquare_Type(),
5330 stubAddr, stubName, TypePtr::BOTTOM,
5331 a_start, n_start, len, inv, top(),
5332 m_start);
5333 set_result(m);
5334 }
5335
5336 return true;
5337 }
5338
5339
5340 /**
5341 * Calculate CRC32 for byte.
5342 * int java.util.zip.CRC32.update(int crc, int b)
5343 */
5344 bool LibraryCallKit::inline_updateCRC32() {
5345 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5346 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5347 // no receiver since it is static method
5348 Node* crc = argument(0); // type: int
5349 Node* b = argument(1); // type: int
5350
5351 /*
5352 * int c = ~ crc;
5353 * b = timesXtoThe32[(b ^ c) & 0xFF];
5354 * b = b ^ (c >>> 8);
5355 * crc = ~b;
5356 */
5357
5358 Node* M1 = intcon(-1);
5359 crc = _gvn.transform(new XorINode(crc, M1));
5360 Node* result = _gvn.transform(new XorINode(crc, b));
5361 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5362
5363 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5364 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5365 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5366 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5367
5368 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5369 result = _gvn.transform(new XorINode(crc, result));
5370 result = _gvn.transform(new XorINode(result, M1));
5371 set_result(result);
5372 return true;
5373 }
5374
5375 /**
5376 * Calculate CRC32 for byte[] array.
5377 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5378 */
5379 bool LibraryCallKit::inline_updateBytesCRC32() {
5380 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5381 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5382 // no receiver since it is static method
5383 Node* crc = argument(0); // type: int
5384 Node* src = argument(1); // type: oop
5385 Node* offset = argument(2); // type: int
5386 Node* length = argument(3); // type: int
5387
5388 const Type* src_type = src->Value(&_gvn);
5389 const TypeAryPtr* top_src = src_type->isa_aryptr();
5390 if (top_src == NULL || top_src->klass() == NULL) {
5391 // failed array check
5392 return false;
5393 }
5394
5395 // Figure out the size and type of the elements we will be copying.
5396 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5397 if (src_elem != T_BYTE) {
5398 return false;
5399 }
5400
5401 // 'src_start' points to src array + scaled offset
5402 Node* src_start = array_element_address(src, offset, src_elem);
5403
5404 // We assume that range check is done by caller.
5405 // TODO: generate range check (offset+length < src.length) in debug VM.
5406
5407 // Call the stub.
5408 address stubAddr = StubRoutines::updateBytesCRC32();
5409 const char *stubName = "updateBytesCRC32";
5410
5411 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5412 stubAddr, stubName, TypePtr::BOTTOM,
5413 crc, src_start, length);
5414 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5415 set_result(result);
5416 return true;
5417 }
5418
5419 /**
5420 * Calculate CRC32 for ByteBuffer.
5421 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5422 */
5423 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5424 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5425 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5426 // no receiver since it is static method
5427 Node* crc = argument(0); // type: int
5428 Node* src = argument(1); // type: long
5429 Node* offset = argument(3); // type: int
5430 Node* length = argument(4); // type: int
5431
5432 src = ConvL2X(src); // adjust Java long to machine word
5433 Node* base = _gvn.transform(new CastX2PNode(src));
5434 offset = ConvI2X(offset);
5435
5436 // 'src_start' points to src array + scaled offset
5437 Node* src_start = basic_plus_adr(top(), base, offset);
5438
5439 // Call the stub.
5440 address stubAddr = StubRoutines::updateBytesCRC32();
5441 const char *stubName = "updateBytesCRC32";
5442
5443 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5444 stubAddr, stubName, TypePtr::BOTTOM,
5445 crc, src_start, length);
5446 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5447 set_result(result);
5448 return true;
5449 }
5450
5451 //------------------------------get_table_from_crc32c_class-----------------------
5452 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5453 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5454 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5455
5456 return table;
5457 }
5458
5459 //------------------------------inline_updateBytesCRC32C-----------------------
5460 //
5461 // Calculate CRC32C for byte[] array.
5462 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5463 //
5464 bool LibraryCallKit::inline_updateBytesCRC32C() {
5465 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5466 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5467 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5468 // no receiver since it is a static method
5469 Node* crc = argument(0); // type: int
5470 Node* src = argument(1); // type: oop
5471 Node* offset = argument(2); // type: int
5472 Node* end = argument(3); // type: int
5473
5474 Node* length = _gvn.transform(new SubINode(end, offset));
5475
5476 const Type* src_type = src->Value(&_gvn);
5477 const TypeAryPtr* top_src = src_type->isa_aryptr();
5478 if (top_src == NULL || top_src->klass() == NULL) {
5479 // failed array check
5480 return false;
5481 }
5482
5483 // Figure out the size and type of the elements we will be copying.
5484 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5485 if (src_elem != T_BYTE) {
5486 return false;
5487 }
5488
5489 // 'src_start' points to src array + scaled offset
5490 Node* src_start = array_element_address(src, offset, src_elem);
5491
5492 // static final int[] byteTable in class CRC32C
5493 Node* table = get_table_from_crc32c_class(callee()->holder());
5494 Node* table_start = array_element_address(table, intcon(0), T_INT);
5495
5496 // We assume that range check is done by caller.
5497 // TODO: generate range check (offset+length < src.length) in debug VM.
5498
5499 // Call the stub.
5500 address stubAddr = StubRoutines::updateBytesCRC32C();
5501 const char *stubName = "updateBytesCRC32C";
5502
5503 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5504 stubAddr, stubName, TypePtr::BOTTOM,
5505 crc, src_start, length, table_start);
5506 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5507 set_result(result);
5508 return true;
5509 }
5510
5511 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5512 //
5513 // Calculate CRC32C for DirectByteBuffer.
5514 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5515 //
5516 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5517 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5518 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5519 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5520 // no receiver since it is a static method
5521 Node* crc = argument(0); // type: int
5522 Node* src = argument(1); // type: long
5523 Node* offset = argument(3); // type: int
5524 Node* end = argument(4); // type: int
5525
5526 Node* length = _gvn.transform(new SubINode(end, offset));
5527
5528 src = ConvL2X(src); // adjust Java long to machine word
5529 Node* base = _gvn.transform(new CastX2PNode(src));
5530 offset = ConvI2X(offset);
5531
5532 // 'src_start' points to src array + scaled offset
5533 Node* src_start = basic_plus_adr(top(), base, offset);
5534
5535 // static final int[] byteTable in class CRC32C
5536 Node* table = get_table_from_crc32c_class(callee()->holder());
5537 Node* table_start = array_element_address(table, intcon(0), T_INT);
5538
5539 // Call the stub.
5540 address stubAddr = StubRoutines::updateBytesCRC32C();
5541 const char *stubName = "updateBytesCRC32C";
5542
5543 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5544 stubAddr, stubName, TypePtr::BOTTOM,
5545 crc, src_start, length, table_start);
5546 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5547 set_result(result);
5548 return true;
5549 }
5550
5551 //------------------------------inline_updateBytesAdler32----------------------
5552 //
5553 // Calculate Adler32 checksum for byte[] array.
5554 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5555 //
5556 bool LibraryCallKit::inline_updateBytesAdler32() {
5557 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5558 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5559 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5560 // no receiver since it is static method
5561 Node* crc = argument(0); // type: int
5562 Node* src = argument(1); // type: oop
5563 Node* offset = argument(2); // type: int
5564 Node* length = argument(3); // type: int
5565
5566 const Type* src_type = src->Value(&_gvn);
5567 const TypeAryPtr* top_src = src_type->isa_aryptr();
5568 if (top_src == NULL || top_src->klass() == NULL) {
5569 // failed array check
5570 return false;
5571 }
5572
5573 // Figure out the size and type of the elements we will be copying.
5574 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5575 if (src_elem != T_BYTE) {
5576 return false;
5577 }
5578
5579 // 'src_start' points to src array + scaled offset
5580 Node* src_start = array_element_address(src, offset, src_elem);
5581
5582 // We assume that range check is done by caller.
5583 // TODO: generate range check (offset+length < src.length) in debug VM.
5584
5585 // Call the stub.
5586 address stubAddr = StubRoutines::updateBytesAdler32();
5587 const char *stubName = "updateBytesAdler32";
5588
5589 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5590 stubAddr, stubName, TypePtr::BOTTOM,
5591 crc, src_start, length);
5592 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5593 set_result(result);
5594 return true;
5595 }
5596
5597 //------------------------------inline_updateByteBufferAdler32---------------
5598 //
5599 // Calculate Adler32 checksum for DirectByteBuffer.
5600 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5601 //
5602 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5603 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5604 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5605 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5606 // no receiver since it is static method
5607 Node* crc = argument(0); // type: int
5608 Node* src = argument(1); // type: long
5609 Node* offset = argument(3); // type: int
5610 Node* length = argument(4); // type: int
5611
5612 src = ConvL2X(src); // adjust Java long to machine word
5613 Node* base = _gvn.transform(new CastX2PNode(src));
5614 offset = ConvI2X(offset);
5615
5616 // 'src_start' points to src array + scaled offset
5617 Node* src_start = basic_plus_adr(top(), base, offset);
5618
5619 // Call the stub.
5620 address stubAddr = StubRoutines::updateBytesAdler32();
5621 const char *stubName = "updateBytesAdler32";
5622
5623 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5624 stubAddr, stubName, TypePtr::BOTTOM,
5625 crc, src_start, length);
5626
5627 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5628 set_result(result);
5629 return true;
5630 }
5631
5632 //----------------------------inline_reference_get----------------------------
5633 // public T java.lang.ref.Reference.get();
5634 bool LibraryCallKit::inline_reference_get() {
5635 const int referent_offset = java_lang_ref_Reference::referent_offset;
5636 guarantee(referent_offset > 0, "should have already been set");
5637
5638 // Get the argument:
5639 Node* reference_obj = null_check_receiver();
5640 if (stopped()) return true;
5641
5642 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5643
5644 ciInstanceKlass* klass = env()->Object_klass();
5645 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5646
5647 Node* no_ctrl = NULL;
5648 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5649
5650 // Use the pre-barrier to record the value in the referent field
5651 pre_barrier(false /* do_load */,
5652 control(),
5653 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5654 result /* pre_val */,
5655 T_OBJECT);
5656
5657 // Add memory barrier to prevent commoning reads from this field
5658 // across safepoint since GC can change its value.
5659 insert_mem_bar(Op_MemBarCPUOrder);
5660
5661 set_result(result);
5662 return true;
5663 }
5664
5665
5666 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5667 bool is_exact=true, bool is_static=false,
5668 ciInstanceKlass * fromKls=NULL) {
5669 if (fromKls == NULL) {
5670 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5671 assert(tinst != NULL, "obj is null");
5672 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5673 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5674 fromKls = tinst->klass()->as_instance_klass();
5675 } else {
5676 assert(is_static, "only for static field access");
5677 }
5678 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5679 ciSymbol::make(fieldTypeString),
5680 is_static);
5681
5682 assert (field != NULL, "undefined field");
5683 if (field == NULL) return (Node *) NULL;
5684
5685 if (is_static) {
5686 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5687 fromObj = makecon(tip);
5688 }
5689
5690 // Next code copied from Parse::do_get_xxx():
5691
5692 // Compute address and memory type.
5693 int offset = field->offset_in_bytes();
5694 bool is_vol = field->is_volatile();
5695 ciType* field_klass = field->type();
5696 assert(field_klass->is_loaded(), "should be loaded");
5697 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5698 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5699 BasicType bt = field->layout_type();
5700
5701 // Build the resultant type of the load
5702 const Type *type;
5703 if (bt == T_OBJECT) {
5704 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5705 } else {
5706 type = Type::get_const_basic_type(bt);
5707 }
5708
5709 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5710 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5711 }
5712 // Build the load.
5713 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5714 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5715 // If reference is volatile, prevent following memory ops from
5716 // floating up past the volatile read. Also prevents commoning
5717 // another volatile read.
5718 if (is_vol) {
5719 // Memory barrier includes bogus read of value to force load BEFORE membar
5720 insert_mem_bar(Op_MemBarAcquire, loadedField);
5721 }
5722 return loadedField;
5723 }
5724
5725
5726 //------------------------------inline_aescrypt_Block-----------------------
5727 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5728 address stubAddr;
5729 const char *stubName;
5730 assert(UseAES, "need AES instruction support");
5731
5732 switch(id) {
5733 case vmIntrinsics::_aescrypt_encryptBlock:
5734 stubAddr = StubRoutines::aescrypt_encryptBlock();
5735 stubName = "aescrypt_encryptBlock";
5736 break;
5737 case vmIntrinsics::_aescrypt_decryptBlock:
5738 stubAddr = StubRoutines::aescrypt_decryptBlock();
5739 stubName = "aescrypt_decryptBlock";
5740 break;
5741 }
5742 if (stubAddr == NULL) return false;
5743
5744 Node* aescrypt_object = argument(0);
5745 Node* src = argument(1);
5746 Node* src_offset = argument(2);
5747 Node* dest = argument(3);
5748 Node* dest_offset = argument(4);
5749
5750 // (1) src and dest are arrays.
5751 const Type* src_type = src->Value(&_gvn);
5752 const Type* dest_type = dest->Value(&_gvn);
5753 const TypeAryPtr* top_src = src_type->isa_aryptr();
5754 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5755 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5756
5757 // for the quick and dirty code we will skip all the checks.
5758 // we are just trying to get the call to be generated.
5759 Node* src_start = src;
5760 Node* dest_start = dest;
5761 if (src_offset != NULL || dest_offset != NULL) {
5762 assert(src_offset != NULL && dest_offset != NULL, "");
5763 src_start = array_element_address(src, src_offset, T_BYTE);
5764 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5765 }
5766
5767 // now need to get the start of its expanded key array
5768 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5769 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5770 if (k_start == NULL) return false;
5771
5772 if (Matcher::pass_original_key_for_aes()) {
5773 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5774 // compatibility issues between Java key expansion and SPARC crypto instructions
5775 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5776 if (original_k_start == NULL) return false;
5777
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, original_k_start);
5782 } else {
5783 // Call the stub.
5784 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5785 stubAddr, stubName, TypePtr::BOTTOM,
5786 src_start, dest_start, k_start);
5787 }
5788
5789 return true;
5790 }
5791
5792 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5793 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5794 address stubAddr;
5795 const char *stubName;
5796
5797 assert(UseAES, "need AES instruction support");
5798
5799 switch(id) {
5800 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5801 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5802 stubName = "cipherBlockChaining_encryptAESCrypt";
5803 break;
5804 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5805 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5806 stubName = "cipherBlockChaining_decryptAESCrypt";
5807 break;
5808 }
5809 if (stubAddr == NULL) return false;
5810
5811 Node* cipherBlockChaining_object = argument(0);
5812 Node* src = argument(1);
5813 Node* src_offset = argument(2);
5814 Node* len = argument(3);
5815 Node* dest = argument(4);
5816 Node* dest_offset = argument(5);
5817
5818 // (1) src and dest are arrays.
5819 const Type* src_type = src->Value(&_gvn);
5820 const Type* dest_type = dest->Value(&_gvn);
5821 const TypeAryPtr* top_src = src_type->isa_aryptr();
5822 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5823 assert (top_src != NULL && top_src->klass() != NULL
5824 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5825
5826 // checks are the responsibility of the caller
5827 Node* src_start = src;
5828 Node* dest_start = dest;
5829 if (src_offset != NULL || dest_offset != NULL) {
5830 assert(src_offset != NULL && dest_offset != NULL, "");
5831 src_start = array_element_address(src, src_offset, T_BYTE);
5832 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5833 }
5834
5835 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5836 // (because of the predicated logic executed earlier).
5837 // so we cast it here safely.
5838 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5839
5840 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5841 if (embeddedCipherObj == NULL) return false;
5842
5843 // cast it to what we know it will be at runtime
5844 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5845 assert(tinst != NULL, "CBC obj is null");
5846 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5847 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5848 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5849
5850 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5851 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5852 const TypeOopPtr* xtype = aklass->as_instance_type();
5853 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5854 aescrypt_object = _gvn.transform(aescrypt_object);
5855
5856 // we need to get the start of the aescrypt_object's expanded key array
5857 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5858 if (k_start == NULL) return false;
5859
5860 // similarly, get the start address of the r vector
5861 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5862 if (objRvec == NULL) return false;
5863 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5864
5865 Node* cbcCrypt;
5866 if (Matcher::pass_original_key_for_aes()) {
5867 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5868 // compatibility issues between Java key expansion and SPARC crypto instructions
5869 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5870 if (original_k_start == NULL) return false;
5871
5872 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5873 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5874 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5875 stubAddr, stubName, TypePtr::BOTTOM,
5876 src_start, dest_start, k_start, r_start, len, original_k_start);
5877 } else {
5878 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5879 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5880 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5881 stubAddr, stubName, TypePtr::BOTTOM,
5882 src_start, dest_start, k_start, r_start, len);
5883 }
5884
5885 // return cipher length (int)
5886 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5887 set_result(retvalue);
5888 return true;
5889 }
5890
5891 //------------------------------get_key_start_from_aescrypt_object-----------------------
5892 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5893 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5894 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5895 if (objAESCryptKey == NULL) return (Node *) NULL;
5896
5897 // now have the array, need to get the start address of the K array
5898 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5899 return k_start;
5900 }
5901
5902 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5903 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5904 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5905 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5906 if (objAESCryptKey == NULL) return (Node *) NULL;
5907
5908 // now have the array, need to get the start address of the lastKey array
5909 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5910 return original_k_start;
5911 }
5912
5913 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5914 // Return node representing slow path of predicate check.
5915 // the pseudo code we want to emulate with this predicate is:
5916 // for encryption:
5917 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5918 // for decryption:
5919 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5920 // note cipher==plain is more conservative than the original java code but that's OK
5921 //
5922 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5923 // The receiver was checked for NULL already.
5924 Node* objCBC = argument(0);
5925
5926 // Load embeddedCipher field of CipherBlockChaining object.
5927 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5928
5929 // get AESCrypt klass for instanceOf check
5930 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5931 // will have same classloader as CipherBlockChaining object
5932 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5933 assert(tinst != NULL, "CBCobj is null");
5934 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5935
5936 // we want to do an instanceof comparison against the AESCrypt class
5937 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5938 if (!klass_AESCrypt->is_loaded()) {
5939 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5940 Node* ctrl = control();
5941 set_control(top()); // no regular fast path
5942 return ctrl;
5943 }
5944 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5945
5946 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5947 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5948 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5949
5950 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5951
5952 // for encryption, we are done
5953 if (!decrypting)
5954 return instof_false; // even if it is NULL
5955
5956 // for decryption, we need to add a further check to avoid
5957 // taking the intrinsic path when cipher and plain are the same
5958 // see the original java code for why.
5959 RegionNode* region = new RegionNode(3);
5960 region->init_req(1, instof_false);
5961 Node* src = argument(1);
5962 Node* dest = argument(4);
5963 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5964 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5965 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5966 region->init_req(2, src_dest_conjoint);
5967
5968 record_for_igvn(region);
5969 return _gvn.transform(region);
5970 }
5971
5972 //------------------------------inline_ghash_processBlocks
5973 bool LibraryCallKit::inline_ghash_processBlocks() {
5974 address stubAddr;
5975 const char *stubName;
5976 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5977
5978 stubAddr = StubRoutines::ghash_processBlocks();
5979 stubName = "ghash_processBlocks";
5980
5981 Node* data = argument(0);
5982 Node* offset = argument(1);
5983 Node* len = argument(2);
5984 Node* state = argument(3);
5985 Node* subkeyH = argument(4);
5986
5987 Node* state_start = array_element_address(state, intcon(0), T_LONG);
5988 assert(state_start, "state is NULL");
5989 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
5990 assert(subkeyH_start, "subkeyH is NULL");
5991 Node* data_start = array_element_address(data, offset, T_BYTE);
5992 assert(data_start, "data is NULL");
5993
5994 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
5995 OptoRuntime::ghash_processBlocks_Type(),
5996 stubAddr, stubName, TypePtr::BOTTOM,
5997 state_start, subkeyH_start, data_start, len);
5998 return true;
5999 }
6000
6001 //------------------------------inline_sha_implCompress-----------------------
6002 //
6003 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6004 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6005 //
6006 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6007 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6008 //
6009 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6010 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6011 //
6012 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6013 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6014
6015 Node* sha_obj = argument(0);
6016 Node* src = argument(1); // type oop
6017 Node* ofs = argument(2); // type int
6018
6019 const Type* src_type = src->Value(&_gvn);
6020 const TypeAryPtr* top_src = src_type->isa_aryptr();
6021 if (top_src == NULL || top_src->klass() == NULL) {
6022 // failed array check
6023 return false;
6024 }
6025 // Figure out the size and type of the elements we will be copying.
6026 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6027 if (src_elem != T_BYTE) {
6028 return false;
6029 }
6030 // 'src_start' points to src array + offset
6031 Node* src_start = array_element_address(src, ofs, src_elem);
6032 Node* state = NULL;
6033 address stubAddr;
6034 const char *stubName;
6035
6036 switch(id) {
6037 case vmIntrinsics::_sha_implCompress:
6038 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6039 state = get_state_from_sha_object(sha_obj);
6040 stubAddr = StubRoutines::sha1_implCompress();
6041 stubName = "sha1_implCompress";
6042 break;
6043 case vmIntrinsics::_sha2_implCompress:
6044 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6045 state = get_state_from_sha_object(sha_obj);
6046 stubAddr = StubRoutines::sha256_implCompress();
6047 stubName = "sha256_implCompress";
6048 break;
6049 case vmIntrinsics::_sha5_implCompress:
6050 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6051 state = get_state_from_sha5_object(sha_obj);
6052 stubAddr = StubRoutines::sha512_implCompress();
6053 stubName = "sha512_implCompress";
6054 break;
6055 default:
6056 fatal_unexpected_iid(id);
6057 return false;
6058 }
6059 if (state == NULL) return false;
6060
6061 // Call the stub.
6062 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6063 stubAddr, stubName, TypePtr::BOTTOM,
6064 src_start, state);
6065
6066 return true;
6067 }
6068
6069 //------------------------------inline_digestBase_implCompressMB-----------------------
6070 //
6071 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6072 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6073 //
6074 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6075 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6076 "need SHA1/SHA256/SHA512 instruction support");
6077 assert((uint)predicate < 3, "sanity");
6078 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6079
6080 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6081 Node* src = argument(1); // byte[] array
6082 Node* ofs = argument(2); // type int
6083 Node* limit = argument(3); // type int
6084
6085 const Type* src_type = src->Value(&_gvn);
6086 const TypeAryPtr* top_src = src_type->isa_aryptr();
6087 if (top_src == NULL || top_src->klass() == NULL) {
6088 // failed array check
6089 return false;
6090 }
6091 // Figure out the size and type of the elements we will be copying.
6092 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6093 if (src_elem != T_BYTE) {
6094 return false;
6095 }
6096 // 'src_start' points to src array + offset
6097 Node* src_start = array_element_address(src, ofs, src_elem);
6098
6099 const char* klass_SHA_name = NULL;
6100 const char* stub_name = NULL;
6101 address stub_addr = NULL;
6102 bool long_state = false;
6103
6104 switch (predicate) {
6105 case 0:
6106 if (UseSHA1Intrinsics) {
6107 klass_SHA_name = "sun/security/provider/SHA";
6108 stub_name = "sha1_implCompressMB";
6109 stub_addr = StubRoutines::sha1_implCompressMB();
6110 }
6111 break;
6112 case 1:
6113 if (UseSHA256Intrinsics) {
6114 klass_SHA_name = "sun/security/provider/SHA2";
6115 stub_name = "sha256_implCompressMB";
6116 stub_addr = StubRoutines::sha256_implCompressMB();
6117 }
6118 break;
6119 case 2:
6120 if (UseSHA512Intrinsics) {
6121 klass_SHA_name = "sun/security/provider/SHA5";
6122 stub_name = "sha512_implCompressMB";
6123 stub_addr = StubRoutines::sha512_implCompressMB();
6124 long_state = true;
6125 }
6126 break;
6127 default:
6128 fatal("unknown SHA intrinsic predicate: %d", predicate);
6129 }
6130 if (klass_SHA_name != NULL) {
6131 // get DigestBase klass to lookup for SHA klass
6132 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6133 assert(tinst != NULL, "digestBase_obj is not instance???");
6134 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6135
6136 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6137 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6138 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6139 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6140 }
6141 return false;
6142 }
6143 //------------------------------inline_sha_implCompressMB-----------------------
6144 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6145 bool long_state, address stubAddr, const char *stubName,
6146 Node* src_start, Node* ofs, Node* limit) {
6147 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6148 const TypeOopPtr* xtype = aklass->as_instance_type();
6149 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6150 sha_obj = _gvn.transform(sha_obj);
6151
6152 Node* state;
6153 if (long_state) {
6154 state = get_state_from_sha5_object(sha_obj);
6155 } else {
6156 state = get_state_from_sha_object(sha_obj);
6157 }
6158 if (state == NULL) return false;
6159
6160 // Call the stub.
6161 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6162 OptoRuntime::digestBase_implCompressMB_Type(),
6163 stubAddr, stubName, TypePtr::BOTTOM,
6164 src_start, state, ofs, limit);
6165 // return ofs (int)
6166 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6167 set_result(result);
6168
6169 return true;
6170 }
6171
6172 //------------------------------get_state_from_sha_object-----------------------
6173 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6174 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6175 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6176 if (sha_state == NULL) return (Node *) NULL;
6177
6178 // now have the array, need to get the start address of the state array
6179 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6180 return state;
6181 }
6182
6183 //------------------------------get_state_from_sha5_object-----------------------
6184 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6185 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6186 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6187 if (sha_state == NULL) return (Node *) NULL;
6188
6189 // now have the array, need to get the start address of the state array
6190 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6191 return state;
6192 }
6193
6194 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6195 // Return node representing slow path of predicate check.
6196 // the pseudo code we want to emulate with this predicate is:
6197 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6198 //
6199 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6200 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6201 "need SHA1/SHA256/SHA512 instruction support");
6202 assert((uint)predicate < 3, "sanity");
6203
6204 // The receiver was checked for NULL already.
6205 Node* digestBaseObj = argument(0);
6206
6207 // get DigestBase klass for instanceOf check
6208 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6209 assert(tinst != NULL, "digestBaseObj is null");
6210 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6211
6212 const char* klass_SHA_name = NULL;
6213 switch (predicate) {
6214 case 0:
6215 if (UseSHA1Intrinsics) {
6216 // we want to do an instanceof comparison against the SHA class
6217 klass_SHA_name = "sun/security/provider/SHA";
6218 }
6219 break;
6220 case 1:
6221 if (UseSHA256Intrinsics) {
6222 // we want to do an instanceof comparison against the SHA2 class
6223 klass_SHA_name = "sun/security/provider/SHA2";
6224 }
6225 break;
6226 case 2:
6227 if (UseSHA512Intrinsics) {
6228 // we want to do an instanceof comparison against the SHA5 class
6229 klass_SHA_name = "sun/security/provider/SHA5";
6230 }
6231 break;
6232 default:
6233 fatal("unknown SHA intrinsic predicate: %d", predicate);
6234 }
6235
6236 ciKlass* klass_SHA = NULL;
6237 if (klass_SHA_name != NULL) {
6238 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6239 }
6240 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6241 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6242 Node* ctrl = control();
6243 set_control(top()); // no intrinsic path
6244 return ctrl;
6245 }
6246 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6247
6248 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6249 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6250 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6251 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6252
6253 return instof_false; // even if it is NULL
6254 }
6255
6256 bool LibraryCallKit::inline_profileBoolean() {
6257 Node* counts = argument(1);
6258 const TypeAryPtr* ary = NULL;
6259 ciArray* aobj = NULL;
6260 if (counts->is_Con()
6261 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6262 && (aobj = ary->const_oop()->as_array()) != NULL
6263 && (aobj->length() == 2)) {
6264 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6265 jint false_cnt = aobj->element_value(0).as_int();
6266 jint true_cnt = aobj->element_value(1).as_int();
6267
6268 if (C->log() != NULL) {
6269 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6270 false_cnt, true_cnt);
6271 }
6272
6273 if (false_cnt + true_cnt == 0) {
6274 // According to profile, never executed.
6275 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6276 Deoptimization::Action_reinterpret);
6277 return true;
6278 }
6279
6280 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6281 // is a number of each value occurrences.
6282 Node* result = argument(0);
6283 if (false_cnt == 0 || true_cnt == 0) {
6284 // According to profile, one value has been never seen.
6285 int expected_val = (false_cnt == 0) ? 1 : 0;
6286
6287 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6288 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6289
6290 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6291 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6292 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6293
6294 { // Slow path: uncommon trap for never seen value and then reexecute
6295 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6296 // the value has been seen at least once.
6297 PreserveJVMState pjvms(this);
6298 PreserveReexecuteState preexecs(this);
6299 jvms()->set_should_reexecute(true);
6300
6301 set_control(slow_path);
6302 set_i_o(i_o());
6303
6304 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6305 Deoptimization::Action_reinterpret);
6306 }
6307 // The guard for never seen value enables sharpening of the result and
6308 // returning a constant. It allows to eliminate branches on the same value
6309 // later on.
6310 set_control(fast_path);
6311 result = intcon(expected_val);
6312 }
6313 // Stop profiling.
6314 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6315 // By replacing method body with profile data (represented as ProfileBooleanNode
6316 // on IR level) we effectively disable profiling.
6317 // It enables full speed execution once optimized code is generated.
6318 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6319 C->record_for_igvn(profile);
6320 set_result(profile);
6321 return true;
6322 } else {
6323 // Continue profiling.
6324 // Profile data isn't available at the moment. So, execute method's bytecode version.
6325 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6326 // is compiled and counters aren't available since corresponding MethodHandle
6327 // isn't a compile-time constant.
6328 return false;
6329 }
6330 }
6331
6332 bool LibraryCallKit::inline_isCompileConstant() {
6333 Node* n = argument(0);
6334 set_result(n->is_Con() ? intcon(1) : intcon(0));
6335 return true;
6336 }
6337
6338 bool LibraryCallKit::inline_deoptimize() {
6339 assert(WhiteBoxAPI, "");
6340 PreserveReexecuteState preexecs(this);
6341 jvms()->set_should_reexecute(false);
6342 uncommon_trap(Deoptimization::Reason_intrinsic,
6343 Deoptimization::Action_none);
6344 return true;
6345 }