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(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
137 }
138
139 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
140 void set_result(RegionNode* region, PhiNode* value);
141 Node* result() { return _result; }
142
143 virtual int reexecute_sp() { return _reexecute_sp; }
144
145 // Helper functions to inline natives
146 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
147 Node* generate_slow_guard(Node* test, RegionNode* region);
148 Node* generate_fair_guard(Node* test, RegionNode* region);
149 Node* generate_negative_guard(Node* index, RegionNode* region,
150 // resulting CastII of index:
151 Node* *pos_index = NULL);
152 Node* generate_limit_guard(Node* offset, Node* subseq_length,
153 Node* array_length,
154 RegionNode* region);
155 Node* generate_current_thread(Node* &tls_output);
156 Node* load_mirror_from_klass(Node* klass);
157 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
158 RegionNode* region, int null_path,
159 int offset);
160 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
161 RegionNode* region, int null_path) {
162 int offset = java_lang_Class::klass_offset_in_bytes();
163 return load_klass_from_mirror_common(mirror, never_see_null,
164 region, null_path,
165 offset);
166 }
167 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
168 RegionNode* region, int null_path) {
169 int offset = java_lang_Class::array_klass_offset_in_bytes();
170 return load_klass_from_mirror_common(mirror, never_see_null,
171 region, null_path,
172 offset);
173 }
174 Node* generate_access_flags_guard(Node* kls,
175 int modifier_mask, int modifier_bits,
176 RegionNode* region);
177 Node* generate_interface_guard(Node* kls, RegionNode* region);
178 Node* generate_array_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, false, false);
180 }
181 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
182 return generate_array_guard_common(kls, region, false, true);
183 }
184 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
185 return generate_array_guard_common(kls, region, true, false);
186 }
187 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
188 return generate_array_guard_common(kls, region, true, true);
189 }
190 Node* generate_array_guard_common(Node* kls, RegionNode* region,
191 bool obj_array, bool not_array);
192 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
193 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
194 bool is_virtual = false, bool is_static = false);
195 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
196 return generate_method_call(method_id, false, true);
197 }
198 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
199 return generate_method_call(method_id, true, false);
200 }
201 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
202
203 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
204 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
205 bool inline_string_compareTo();
206 bool inline_string_indexOf();
207 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
208 bool inline_string_equals();
209 Node* round_double_node(Node* n);
210 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
211 bool inline_math_native(vmIntrinsics::ID id);
212 bool inline_trig(vmIntrinsics::ID id);
213 bool inline_math(vmIntrinsics::ID id);
214 template <typename OverflowOp>
215 bool inline_math_overflow(Node* arg1, Node* arg2);
216 void inline_math_mathExact(Node* math, Node* test);
217 bool inline_math_addExactI(bool is_increment);
218 bool inline_math_addExactL(bool is_increment);
219 bool inline_math_multiplyExactI();
220 bool inline_math_multiplyExactL();
221 bool inline_math_negateExactI();
222 bool inline_math_negateExactL();
223 bool inline_math_subtractExactI(bool is_decrement);
224 bool inline_math_subtractExactL(bool is_decrement);
225 bool inline_pow();
226 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
227 bool inline_min_max(vmIntrinsics::ID id);
228 bool inline_notify(vmIntrinsics::ID id);
229 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
230 // This returns Type::AnyPtr, RawPtr, or OopPtr.
231 int classify_unsafe_addr(Node* &base, Node* &offset);
232 Node* make_unsafe_address(Node* base, Node* offset);
233 // Helper for inline_unsafe_access.
234 // Generates the guards that check whether the result of
235 // Unsafe.getObject should be recorded in an SATB log buffer.
236 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
237 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
238 static bool klass_needs_init_guard(Node* kls);
239 bool inline_unsafe_allocate();
240 bool inline_unsafe_copyMemory();
241 bool inline_native_currentThread();
242 #ifdef TRACE_HAVE_INTRINSICS
243 bool inline_native_classID();
244 bool inline_native_threadID();
245 #endif
246 bool inline_native_time_funcs(address method, const char* funcName);
247 bool inline_native_isInterrupted();
248 bool inline_native_Class_query(vmIntrinsics::ID id);
249 bool inline_native_subtype_check();
250
251 bool inline_native_newArray();
252 bool inline_native_getLength();
253 bool inline_array_copyOf(bool is_copyOfRange);
254 bool inline_array_equals();
255 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
256 bool inline_native_clone(bool is_virtual);
257 bool inline_native_Reflection_getCallerClass();
258 // Helper function for inlining native object hash method
259 bool inline_native_hashcode(bool is_virtual, bool is_static);
260 bool inline_native_getClass();
261
262 // Helper functions for inlining arraycopy
263 bool inline_arraycopy();
264 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
265 RegionNode* slow_region);
266 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
267 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
268
269 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
270 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
271 bool inline_unsafe_ordered_store(BasicType type);
272 bool inline_unsafe_fence(vmIntrinsics::ID id);
273 bool inline_fp_conversions(vmIntrinsics::ID id);
274 bool inline_number_methods(vmIntrinsics::ID id);
275 bool inline_reference_get();
276 bool inline_Class_cast();
277 bool inline_aescrypt_Block(vmIntrinsics::ID id);
278 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
279 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
280 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
281 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
282 bool inline_ghash_processBlocks();
283 bool inline_sha_implCompress(vmIntrinsics::ID id);
284 bool inline_digestBase_implCompressMB(int predicate);
285 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
286 bool long_state, address stubAddr, const char *stubName,
287 Node* src_start, Node* ofs, Node* limit);
288 Node* get_state_from_sha_object(Node *sha_object);
289 Node* get_state_from_sha5_object(Node *sha_object);
290 Node* inline_digestBase_implCompressMB_predicate(int predicate);
291 bool inline_encodeISOArray();
292 bool inline_updateCRC32();
293 bool inline_updateBytesCRC32();
294 bool inline_updateByteBufferCRC32();
295 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
296 bool inline_updateBytesCRC32C();
297 bool inline_updateDirectByteBufferCRC32C();
298 bool inline_updateBytesAdler32();
299 bool inline_updateByteBufferAdler32();
300 bool inline_multiplyToLen();
301 bool inline_squareToLen();
302 bool inline_mulAdd();
303 bool inline_montgomeryMultiply();
304 bool inline_montgomerySquare();
305
306 bool inline_profileBoolean();
307 bool inline_isCompileConstant();
308 };
309
310 //---------------------------make_vm_intrinsic----------------------------
311 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
312 vmIntrinsics::ID id = m->intrinsic_id();
313 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
314
315 if (!m->is_loaded()) {
316 // Do not attempt to inline unloaded methods.
317 return NULL;
318 }
319
320 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
321 bool is_available = false;
322
323 {
324 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
325 // the compiler must transition to '_thread_in_vm' state because both
326 // methods access VM-internal data.
327 VM_ENTRY_MARK;
328 methodHandle mh(THREAD, m->get_Method());
329 methodHandle ct(THREAD, method()->get_Method());
330 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
331 !vmIntrinsics::is_disabled_by_flags(mh, ct);
332 }
333
334 if (is_available) {
335 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
336 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
337 return new LibraryIntrinsic(m, is_virtual,
338 vmIntrinsics::predicates_needed(id),
339 vmIntrinsics::does_virtual_dispatch(id),
340 (vmIntrinsics::ID) id);
341 } else {
342 return NULL;
343 }
344 }
345
346 //----------------------register_library_intrinsics-----------------------
347 // Initialize this file's data structures, for each Compile instance.
348 void Compile::register_library_intrinsics() {
349 // Nothing to do here.
350 }
351
352 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
353 LibraryCallKit kit(jvms, this);
354 Compile* C = kit.C;
355 int nodes = C->unique();
356 #ifndef PRODUCT
357 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
358 char buf[1000];
359 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
360 tty->print_cr("Intrinsic %s", str);
361 }
362 #endif
363 ciMethod* callee = kit.callee();
364 const int bci = kit.bci();
365
366 // Try to inline the intrinsic.
367 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
368 kit.try_to_inline(_last_predicate)) {
369 if (C->print_intrinsics() || C->print_inlining()) {
370 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
371 }
372 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
373 if (C->log()) {
374 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
375 vmIntrinsics::name_at(intrinsic_id()),
376 (is_virtual() ? " virtual='1'" : ""),
377 C->unique() - nodes);
378 }
379 // Push the result from the inlined method onto the stack.
380 kit.push_result();
381 C->print_inlining_update(this);
382 return kit.transfer_exceptions_into_jvms();
383 }
384
385 // The intrinsic bailed out
386 if (C->print_intrinsics() || C->print_inlining()) {
387 if (jvms->has_method()) {
388 // Not a root compile.
389 const char* msg;
390 if (callee->intrinsic_candidate()) {
391 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
392 } else {
393 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
394 : "failed to inline (intrinsic), method not annotated";
395 }
396 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
397 } else {
398 // Root compile
399 tty->print("Did not generate intrinsic %s%s at bci:%d in",
400 vmIntrinsics::name_at(intrinsic_id()),
401 (is_virtual() ? " (virtual)" : ""), bci);
402 }
403 }
404 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
405 C->print_inlining_update(this);
406 return NULL;
407 }
408
409 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
410 LibraryCallKit kit(jvms, this);
411 Compile* C = kit.C;
412 int nodes = C->unique();
413 _last_predicate = predicate;
414 #ifndef PRODUCT
415 assert(is_predicated() && predicate < predicates_count(), "sanity");
416 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
417 char buf[1000];
418 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
419 tty->print_cr("Predicate for intrinsic %s", str);
420 }
421 #endif
422 ciMethod* callee = kit.callee();
423 const int bci = kit.bci();
424
425 Node* slow_ctl = kit.try_to_predicate(predicate);
426 if (!kit.failing()) {
427 if (C->print_intrinsics() || C->print_inlining()) {
428 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
429 }
430 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
431 if (C->log()) {
432 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
433 vmIntrinsics::name_at(intrinsic_id()),
434 (is_virtual() ? " virtual='1'" : ""),
435 C->unique() - nodes);
436 }
437 return slow_ctl; // Could be NULL if the check folds.
438 }
439
440 // The intrinsic bailed out
441 if (C->print_intrinsics() || C->print_inlining()) {
442 if (jvms->has_method()) {
443 // Not a root compile.
444 const char* msg = "failed to generate predicate for intrinsic";
445 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
446 } else {
447 // Root compile
448 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
449 vmIntrinsics::name_at(intrinsic_id()),
450 (is_virtual() ? " (virtual)" : ""), bci);
451 }
452 }
453 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
454 return NULL;
455 }
456
457 bool LibraryCallKit::try_to_inline(int predicate) {
458 // Handle symbolic names for otherwise undistinguished boolean switches:
459 const bool is_store = true;
460 const bool is_native_ptr = true;
461 const bool is_static = true;
462 const bool is_volatile = true;
463
464 if (!jvms()->has_method()) {
465 // Root JVMState has a null method.
466 assert(map()->memory()->Opcode() == Op_Parm, "");
467 // Insert the memory aliasing node
468 set_all_memory(reset_memory());
469 }
470 assert(merged_memory(), "");
471
472
473 switch (intrinsic_id()) {
474 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
475 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
476 case vmIntrinsics::_getClass: return inline_native_getClass();
477
478 case vmIntrinsics::_dsin:
479 case vmIntrinsics::_dcos:
480 case vmIntrinsics::_dtan:
481 case vmIntrinsics::_dabs:
482 case vmIntrinsics::_datan2:
483 case vmIntrinsics::_dsqrt:
484 case vmIntrinsics::_dexp:
485 case vmIntrinsics::_dlog:
486 case vmIntrinsics::_dlog10:
487 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
488
489 case vmIntrinsics::_min:
490 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
491
492 case vmIntrinsics::_notify:
493 case vmIntrinsics::_notifyAll:
494 if (InlineNotify) {
495 return inline_notify(intrinsic_id());
496 }
497 return false;
498
499 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
500 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
501 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
502 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
503 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
504 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
505 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
506 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
507 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
508 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
509 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
510 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
511
512 case vmIntrinsics::_arraycopy: return inline_arraycopy();
513
514 case vmIntrinsics::_compareTo: return inline_string_compareTo();
515 case vmIntrinsics::_indexOf: return inline_string_indexOf();
516 case vmIntrinsics::_equals: return inline_string_equals();
517
518 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
519 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
520 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
521 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
522 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
523 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
524 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
525 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
526 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
527 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
528 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
529 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
530 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
531 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
532 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
533 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
534 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
535 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
536
537 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
538 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
539 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
540 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
541 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
542 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
543 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
544 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
545
546 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
547 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
548 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
549 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
550 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
551 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
552 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
553 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
554
555 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
556 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
557 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
558 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
559 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
560 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
561 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
562 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
563 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
564
565 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
566 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
567 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
568 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
569 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
570 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
571 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
572 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
573 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
574
575 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
576 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
577 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
578 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
579
580 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
581 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
582 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
583 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
584
585 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
586 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
587 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
588
589 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
590 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
591 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
592
593 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
594 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
595 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
596 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
597 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
598
599 case vmIntrinsics::_loadFence:
600 case vmIntrinsics::_storeFence:
601 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
602
603 case vmIntrinsics::_currentThread: return inline_native_currentThread();
604 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
605
606 #ifdef TRACE_HAVE_INTRINSICS
607 case vmIntrinsics::_classID: return inline_native_classID();
608 case vmIntrinsics::_threadID: return inline_native_threadID();
609 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
610 #endif
611 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
612 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
613 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
614 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
615 case vmIntrinsics::_newArray: return inline_native_newArray();
616 case vmIntrinsics::_getLength: return inline_native_getLength();
617 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
618 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
619 case vmIntrinsics::_equalsC: return inline_array_equals();
620 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
621
622 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
623
624 case vmIntrinsics::_isInstance:
625 case vmIntrinsics::_getModifiers:
626 case vmIntrinsics::_isInterface:
627 case vmIntrinsics::_isArray:
628 case vmIntrinsics::_isPrimitive:
629 case vmIntrinsics::_getSuperclass:
630 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
631
632 case vmIntrinsics::_floatToRawIntBits:
633 case vmIntrinsics::_floatToIntBits:
634 case vmIntrinsics::_intBitsToFloat:
635 case vmIntrinsics::_doubleToRawLongBits:
636 case vmIntrinsics::_doubleToLongBits:
637 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
638
639 case vmIntrinsics::_numberOfLeadingZeros_i:
640 case vmIntrinsics::_numberOfLeadingZeros_l:
641 case vmIntrinsics::_numberOfTrailingZeros_i:
642 case vmIntrinsics::_numberOfTrailingZeros_l:
643 case vmIntrinsics::_bitCount_i:
644 case vmIntrinsics::_bitCount_l:
645 case vmIntrinsics::_reverseBytes_i:
646 case vmIntrinsics::_reverseBytes_l:
647 case vmIntrinsics::_reverseBytes_s:
648 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
649
650 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
651
652 case vmIntrinsics::_Reference_get: return inline_reference_get();
653
654 case vmIntrinsics::_Class_cast: return inline_Class_cast();
655
656 case vmIntrinsics::_aescrypt_encryptBlock:
657 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
658
659 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
660 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
661 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
662
663 case vmIntrinsics::_sha_implCompress:
664 case vmIntrinsics::_sha2_implCompress:
665 case vmIntrinsics::_sha5_implCompress:
666 return inline_sha_implCompress(intrinsic_id());
667
668 case vmIntrinsics::_digestBase_implCompressMB:
669 return inline_digestBase_implCompressMB(predicate);
670
671 case vmIntrinsics::_multiplyToLen:
672 return inline_multiplyToLen();
673
674 case vmIntrinsics::_squareToLen:
675 return inline_squareToLen();
676
677 case vmIntrinsics::_mulAdd:
678 return inline_mulAdd();
679
680 case vmIntrinsics::_montgomeryMultiply:
681 return inline_montgomeryMultiply();
682 case vmIntrinsics::_montgomerySquare:
683 return inline_montgomerySquare();
684
685 case vmIntrinsics::_ghash_processBlocks:
686 return inline_ghash_processBlocks();
687
688 case vmIntrinsics::_encodeISOArray:
689 return inline_encodeISOArray();
690
691 case vmIntrinsics::_updateCRC32:
692 return inline_updateCRC32();
693 case vmIntrinsics::_updateBytesCRC32:
694 return inline_updateBytesCRC32();
695 case vmIntrinsics::_updateByteBufferCRC32:
696 return inline_updateByteBufferCRC32();
697
698 case vmIntrinsics::_updateBytesCRC32C:
699 return inline_updateBytesCRC32C();
700 case vmIntrinsics::_updateDirectByteBufferCRC32C:
701 return inline_updateDirectByteBufferCRC32C();
702
703 case vmIntrinsics::_updateBytesAdler32:
704 return inline_updateBytesAdler32();
705 case vmIntrinsics::_updateByteBufferAdler32:
706 return inline_updateByteBufferAdler32();
707
708 case vmIntrinsics::_profileBoolean:
709 return inline_profileBoolean();
710 case vmIntrinsics::_isCompileConstant:
711 return inline_isCompileConstant();
712
713 default:
714 // If you get here, it may be that someone has added a new intrinsic
715 // to the list in vmSymbols.hpp without implementing it here.
716 #ifndef PRODUCT
717 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
718 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
719 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
720 }
721 #endif
722 return false;
723 }
724 }
725
726 Node* LibraryCallKit::try_to_predicate(int predicate) {
727 if (!jvms()->has_method()) {
728 // Root JVMState has a null method.
729 assert(map()->memory()->Opcode() == Op_Parm, "");
730 // Insert the memory aliasing node
731 set_all_memory(reset_memory());
732 }
733 assert(merged_memory(), "");
734
735 switch (intrinsic_id()) {
736 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
737 return inline_cipherBlockChaining_AESCrypt_predicate(false);
738 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
739 return inline_cipherBlockChaining_AESCrypt_predicate(true);
740 case vmIntrinsics::_digestBase_implCompressMB:
741 return inline_digestBase_implCompressMB_predicate(predicate);
742
743 default:
744 // If you get here, it may be that someone has added a new intrinsic
745 // to the list in vmSymbols.hpp without implementing it here.
746 #ifndef PRODUCT
747 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
748 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
749 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
750 }
751 #endif
752 Node* slow_ctl = control();
753 set_control(top()); // No fast path instrinsic
754 return slow_ctl;
755 }
756 }
757
758 //------------------------------set_result-------------------------------
759 // Helper function for finishing intrinsics.
760 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
761 record_for_igvn(region);
762 set_control(_gvn.transform(region));
763 set_result( _gvn.transform(value));
764 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
765 }
766
767 //------------------------------generate_guard---------------------------
768 // Helper function for generating guarded fast-slow graph structures.
769 // The given 'test', if true, guards a slow path. If the test fails
770 // then a fast path can be taken. (We generally hope it fails.)
771 // In all cases, GraphKit::control() is updated to the fast path.
772 // The returned value represents the control for the slow path.
773 // The return value is never 'top'; it is either a valid control
774 // or NULL if it is obvious that the slow path can never be taken.
775 // Also, if region and the slow control are not NULL, the slow edge
776 // is appended to the region.
777 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
778 if (stopped()) {
779 // Already short circuited.
780 return NULL;
781 }
782
783 // Build an if node and its projections.
784 // If test is true we take the slow path, which we assume is uncommon.
785 if (_gvn.type(test) == TypeInt::ZERO) {
786 // The slow branch is never taken. No need to build this guard.
787 return NULL;
788 }
789
790 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
791
792 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
793 if (if_slow == top()) {
794 // The slow branch is never taken. No need to build this guard.
795 return NULL;
796 }
797
798 if (region != NULL)
799 region->add_req(if_slow);
800
801 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
802 set_control(if_fast);
803
804 return if_slow;
805 }
806
807 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
808 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
809 }
810 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
811 return generate_guard(test, region, PROB_FAIR);
812 }
813
814 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
815 Node* *pos_index) {
816 if (stopped())
817 return NULL; // already stopped
818 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
819 return NULL; // index is already adequately typed
820 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
821 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
822 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
823 if (is_neg != NULL && pos_index != NULL) {
824 // Emulate effect of Parse::adjust_map_after_if.
825 Node* ccast = new CastIINode(index, TypeInt::POS);
826 ccast->set_req(0, control());
827 (*pos_index) = _gvn.transform(ccast);
828 }
829 return is_neg;
830 }
831
832 // Make sure that 'position' is a valid limit index, in [0..length].
833 // There are two equivalent plans for checking this:
834 // A. (offset + copyLength) unsigned<= arrayLength
835 // B. offset <= (arrayLength - copyLength)
836 // We require that all of the values above, except for the sum and
837 // difference, are already known to be non-negative.
838 // Plan A is robust in the face of overflow, if offset and copyLength
839 // are both hugely positive.
840 //
841 // Plan B is less direct and intuitive, but it does not overflow at
842 // all, since the difference of two non-negatives is always
843 // representable. Whenever Java methods must perform the equivalent
844 // check they generally use Plan B instead of Plan A.
845 // For the moment we use Plan A.
846 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
847 Node* subseq_length,
848 Node* array_length,
849 RegionNode* region) {
850 if (stopped())
851 return NULL; // already stopped
852 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
853 if (zero_offset && subseq_length->eqv_uncast(array_length))
854 return NULL; // common case of whole-array copy
855 Node* last = subseq_length;
856 if (!zero_offset) // last += offset
857 last = _gvn.transform(new AddINode(last, offset));
858 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
859 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
860 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
861 return is_over;
862 }
863
864
865 //--------------------------generate_current_thread--------------------
866 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
867 ciKlass* thread_klass = env()->Thread_klass();
868 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
869 Node* thread = _gvn.transform(new ThreadLocalNode());
870 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
871 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
872 tls_output = thread;
873 return threadObj;
874 }
875
876
877 //------------------------------make_string_method_node------------------------
878 // Helper method for String intrinsic functions. This version is called
879 // with str1 and str2 pointing to String object nodes.
880 //
881 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
882 Node* no_ctrl = NULL;
883
884 // Get start addr of string
885 Node* str1_value = load_String_value(no_ctrl, str1);
886 Node* str1_offset = load_String_offset(no_ctrl, str1);
887 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
888
889 // Get length of string 1
890 Node* str1_len = load_String_length(no_ctrl, str1);
891
892 Node* str2_value = load_String_value(no_ctrl, str2);
893 Node* str2_offset = load_String_offset(no_ctrl, str2);
894 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
895
896 Node* str2_len = NULL;
897 Node* result = NULL;
898
899 switch (opcode) {
900 case Op_StrIndexOf:
901 // Get length of string 2
902 str2_len = load_String_length(no_ctrl, str2);
903
904 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
905 str1_start, str1_len, str2_start, str2_len);
906 break;
907 case Op_StrComp:
908 // Get length of string 2
909 str2_len = load_String_length(no_ctrl, str2);
910
911 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
912 str1_start, str1_len, str2_start, str2_len);
913 break;
914 case Op_StrEquals:
915 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
916 str1_start, str2_start, str1_len);
917 break;
918 default:
919 ShouldNotReachHere();
920 return NULL;
921 }
922
923 // All these intrinsics have checks.
924 C->set_has_split_ifs(true); // Has chance for split-if optimization
925
926 return _gvn.transform(result);
927 }
928
929 // Helper method for String intrinsic functions. This version is called
930 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
931 // to Int nodes containing the lenghts of str1 and str2.
932 //
933 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
934 Node* result = NULL;
935 switch (opcode) {
936 case Op_StrIndexOf:
937 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
938 str1_start, cnt1, str2_start, cnt2);
939 break;
940 case Op_StrComp:
941 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
942 str1_start, cnt1, str2_start, cnt2);
943 break;
944 case Op_StrEquals:
945 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
946 str1_start, str2_start, cnt1);
947 break;
948 default:
949 ShouldNotReachHere();
950 return NULL;
951 }
952
953 // All these intrinsics have checks.
954 C->set_has_split_ifs(true); // Has chance for split-if optimization
955
956 return _gvn.transform(result);
957 }
958
959 //------------------------------inline_string_compareTo------------------------
960 // public int java.lang.String.compareTo(String anotherString);
961 bool LibraryCallKit::inline_string_compareTo() {
962 Node* receiver = null_check(argument(0));
963 Node* arg = null_check(argument(1));
964 if (stopped()) {
965 return true;
966 }
967 set_result(make_string_method_node(Op_StrComp, receiver, arg));
968 return true;
969 }
970
971 //------------------------------inline_string_equals------------------------
972 bool LibraryCallKit::inline_string_equals() {
973 Node* receiver = null_check_receiver();
974 // NOTE: Do not null check argument for String.equals() because spec
975 // allows to specify NULL as argument.
976 Node* argument = this->argument(1);
977 if (stopped()) {
978 return true;
979 }
980
981 // paths (plus control) merge
982 RegionNode* region = new RegionNode(5);
983 Node* phi = new PhiNode(region, TypeInt::BOOL);
984
985 // does source == target string?
986 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
987 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
988
989 Node* if_eq = generate_slow_guard(bol, NULL);
990 if (if_eq != NULL) {
991 // receiver == argument
992 phi->init_req(2, intcon(1));
993 region->init_req(2, if_eq);
994 }
995
996 // get String klass for instanceOf
997 ciInstanceKlass* klass = env()->String_klass();
998
999 if (!stopped()) {
1000 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1001 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1002 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1003
1004 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1005 //instanceOf == true, fallthrough
1006
1007 if (inst_false != NULL) {
1008 phi->init_req(3, intcon(0));
1009 region->init_req(3, inst_false);
1010 }
1011 }
1012
1013 if (!stopped()) {
1014 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1015
1016 // Properly cast the argument to String
1017 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1018 // This path is taken only when argument's type is String:NotNull.
1019 argument = cast_not_null(argument, false);
1020
1021 Node* no_ctrl = NULL;
1022
1023 // Get start addr of receiver
1024 Node* receiver_val = load_String_value(no_ctrl, receiver);
1025 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1026 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1027
1028 // Get length of receiver
1029 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1030
1031 // Get start addr of argument
1032 Node* argument_val = load_String_value(no_ctrl, argument);
1033 Node* argument_offset = load_String_offset(no_ctrl, argument);
1034 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1035
1036 // Get length of argument
1037 Node* argument_cnt = load_String_length(no_ctrl, argument);
1038
1039 // Check for receiver count != argument count
1040 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1041 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1042 Node* if_ne = generate_slow_guard(bol, NULL);
1043 if (if_ne != NULL) {
1044 phi->init_req(4, intcon(0));
1045 region->init_req(4, if_ne);
1046 }
1047
1048 // Check for count == 0 is done by assembler code for StrEquals.
1049
1050 if (!stopped()) {
1051 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1052 phi->init_req(1, equals);
1053 region->init_req(1, control());
1054 }
1055 }
1056
1057 // post merge
1058 set_control(_gvn.transform(region));
1059 record_for_igvn(region);
1060
1061 set_result(_gvn.transform(phi));
1062 return true;
1063 }
1064
1065 //------------------------------inline_array_equals----------------------------
1066 bool LibraryCallKit::inline_array_equals() {
1067 Node* arg1 = argument(0);
1068 Node* arg2 = argument(1);
1069 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1070 return true;
1071 }
1072
1073 // Java version of String.indexOf(constant string)
1074 // class StringDecl {
1075 // StringDecl(char[] ca) {
1076 // offset = 0;
1077 // count = ca.length;
1078 // value = ca;
1079 // }
1080 // int offset;
1081 // int count;
1082 // char[] value;
1083 // }
1084 //
1085 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1086 // int targetOffset, int cache_i, int md2) {
1087 // int cache = cache_i;
1088 // int sourceOffset = string_object.offset;
1089 // int sourceCount = string_object.count;
1090 // int targetCount = target_object.length;
1091 //
1092 // int targetCountLess1 = targetCount - 1;
1093 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1094 //
1095 // char[] source = string_object.value;
1096 // char[] target = target_object;
1097 // int lastChar = target[targetCountLess1];
1098 //
1099 // outer_loop:
1100 // for (int i = sourceOffset; i < sourceEnd; ) {
1101 // int src = source[i + targetCountLess1];
1102 // if (src == lastChar) {
1103 // // With random strings and a 4-character alphabet,
1104 // // reverse matching at this point sets up 0.8% fewer
1105 // // frames, but (paradoxically) makes 0.3% more probes.
1106 // // Since those probes are nearer the lastChar probe,
1107 // // there is may be a net D$ win with reverse matching.
1108 // // But, reversing loop inhibits unroll of inner loop
1109 // // for unknown reason. So, does running outer loop from
1110 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1111 // for (int j = 0; j < targetCountLess1; j++) {
1112 // if (target[targetOffset + j] != source[i+j]) {
1113 // if ((cache & (1 << source[i+j])) == 0) {
1114 // if (md2 < j+1) {
1115 // i += j+1;
1116 // continue outer_loop;
1117 // }
1118 // }
1119 // i += md2;
1120 // continue outer_loop;
1121 // }
1122 // }
1123 // return i - sourceOffset;
1124 // }
1125 // if ((cache & (1 << src)) == 0) {
1126 // i += targetCountLess1;
1127 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1128 // i++;
1129 // }
1130 // return -1;
1131 // }
1132
1133 //------------------------------string_indexOf------------------------
1134 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1135 jint cache_i, jint md2_i) {
1136
1137 Node* no_ctrl = NULL;
1138 float likely = PROB_LIKELY(0.9);
1139 float unlikely = PROB_UNLIKELY(0.9);
1140
1141 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1142
1143 Node* source = load_String_value(no_ctrl, string_object);
1144 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1145 Node* sourceCount = load_String_length(no_ctrl, string_object);
1146
1147 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1148 jint target_length = target_array->length();
1149 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1150 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1151
1152 // String.value field is known to be @Stable.
1153 if (UseImplicitStableValues) {
1154 target = cast_array_to_stable(target, target_type);
1155 }
1156
1157 IdealKit kit(this, false, true);
1158 #define __ kit.
1159 Node* zero = __ ConI(0);
1160 Node* one = __ ConI(1);
1161 Node* cache = __ ConI(cache_i);
1162 Node* md2 = __ ConI(md2_i);
1163 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1164 Node* targetCountLess1 = __ ConI(target_length - 1);
1165 Node* targetOffset = __ ConI(targetOffset_i);
1166 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1167
1168 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1169 Node* outer_loop = __ make_label(2 /* goto */);
1170 Node* return_ = __ make_label(1);
1171
1172 __ set(rtn,__ ConI(-1));
1173 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1174 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1175 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1176 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1177 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1178 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1179 Node* tpj = __ AddI(targetOffset, __ value(j));
1180 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1181 Node* ipj = __ AddI(__ value(i), __ value(j));
1182 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1183 __ if_then(targ, BoolTest::ne, src2); {
1184 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1185 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1186 __ increment(i, __ AddI(__ value(j), one));
1187 __ goto_(outer_loop);
1188 } __ end_if(); __ dead(j);
1189 }__ end_if(); __ dead(j);
1190 __ increment(i, md2);
1191 __ goto_(outer_loop);
1192 }__ end_if();
1193 __ increment(j, one);
1194 }__ end_loop(); __ dead(j);
1195 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1196 __ goto_(return_);
1197 }__ end_if();
1198 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1199 __ increment(i, targetCountLess1);
1200 }__ end_if();
1201 __ increment(i, one);
1202 __ bind(outer_loop);
1203 }__ end_loop(); __ dead(i);
1204 __ bind(return_);
1205
1206 // Final sync IdealKit and GraphKit.
1207 final_sync(kit);
1208 Node* result = __ value(rtn);
1209 #undef __
1210 C->set_has_loops(true);
1211 return result;
1212 }
1213
1214 //------------------------------inline_string_indexOf------------------------
1215 bool LibraryCallKit::inline_string_indexOf() {
1216 Node* receiver = argument(0);
1217 Node* arg = argument(1);
1218
1219 Node* result;
1220 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1221 UseSSE42Intrinsics) {
1222 // Generate SSE4.2 version of indexOf
1223 // We currently only have match rules that use SSE4.2
1224
1225 receiver = null_check(receiver);
1226 arg = null_check(arg);
1227 if (stopped()) {
1228 return true;
1229 }
1230
1231 // Make the merge point
1232 RegionNode* result_rgn = new RegionNode(4);
1233 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1234 Node* no_ctrl = NULL;
1235
1236 // Get start addr of source string
1237 Node* source = load_String_value(no_ctrl, receiver);
1238 Node* source_offset = load_String_offset(no_ctrl, receiver);
1239 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1240
1241 // Get length of source string
1242 Node* source_cnt = load_String_length(no_ctrl, receiver);
1243
1244 // Get start addr of substring
1245 Node* substr = load_String_value(no_ctrl, arg);
1246 Node* substr_offset = load_String_offset(no_ctrl, arg);
1247 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1248
1249 // Get length of source string
1250 Node* substr_cnt = load_String_length(no_ctrl, arg);
1251
1252 // Check for substr count > string count
1253 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1254 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1255 Node* if_gt = generate_slow_guard(bol, NULL);
1256 if (if_gt != NULL) {
1257 result_phi->init_req(2, intcon(-1));
1258 result_rgn->init_req(2, if_gt);
1259 }
1260
1261 if (!stopped()) {
1262 // Check for substr count == 0
1263 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1264 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1265 Node* if_zero = generate_slow_guard(bol, NULL);
1266 if (if_zero != NULL) {
1267 result_phi->init_req(3, intcon(0));
1268 result_rgn->init_req(3, if_zero);
1269 }
1270 }
1271
1272 if (!stopped()) {
1273 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1274 result_phi->init_req(1, result);
1275 result_rgn->init_req(1, control());
1276 }
1277 set_control(_gvn.transform(result_rgn));
1278 record_for_igvn(result_rgn);
1279 result = _gvn.transform(result_phi);
1280
1281 } else { // Use LibraryCallKit::string_indexOf
1282 // don't intrinsify if argument isn't a constant string.
1283 if (!arg->is_Con()) {
1284 return false;
1285 }
1286 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1287 if (str_type == NULL) {
1288 return false;
1289 }
1290 ciInstanceKlass* klass = env()->String_klass();
1291 ciObject* str_const = str_type->const_oop();
1292 if (str_const == NULL || str_const->klass() != klass) {
1293 return false;
1294 }
1295 ciInstance* str = str_const->as_instance();
1296 assert(str != NULL, "must be instance");
1297
1298 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1299 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1300
1301 int o;
1302 int c;
1303 if (java_lang_String::has_offset_field()) {
1304 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1305 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1306 } else {
1307 o = 0;
1308 c = pat->length();
1309 }
1310
1311 // constant strings have no offset and count == length which
1312 // simplifies the resulting code somewhat so lets optimize for that.
1313 if (o != 0 || c != pat->length()) {
1314 return false;
1315 }
1316
1317 receiver = null_check(receiver, T_OBJECT);
1318 // NOTE: No null check on the argument is needed since it's a constant String oop.
1319 if (stopped()) {
1320 return true;
1321 }
1322
1323 // The null string as a pattern always returns 0 (match at beginning of string)
1324 if (c == 0) {
1325 set_result(intcon(0));
1326 return true;
1327 }
1328
1329 // Generate default indexOf
1330 jchar lastChar = pat->char_at(o + (c - 1));
1331 int cache = 0;
1332 int i;
1333 for (i = 0; i < c - 1; i++) {
1334 assert(i < pat->length(), "out of range");
1335 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1336 }
1337
1338 int md2 = c;
1339 for (i = 0; i < c - 1; i++) {
1340 assert(i < pat->length(), "out of range");
1341 if (pat->char_at(o + i) == lastChar) {
1342 md2 = (c - 1) - i;
1343 }
1344 }
1345
1346 result = string_indexOf(receiver, pat, o, cache, md2);
1347 }
1348 set_result(result);
1349 return true;
1350 }
1351
1352 //--------------------------round_double_node--------------------------------
1353 // Round a double node if necessary.
1354 Node* LibraryCallKit::round_double_node(Node* n) {
1355 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1356 n = _gvn.transform(new RoundDoubleNode(0, n));
1357 return n;
1358 }
1359
1360 //------------------------------inline_math-----------------------------------
1361 // public static double Math.abs(double)
1362 // public static double Math.sqrt(double)
1363 // public static double Math.log(double)
1364 // public static double Math.log10(double)
1365 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1366 Node* arg = round_double_node(argument(0));
1367 Node* n;
1368 switch (id) {
1369 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1370 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1371 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1372 default: fatal_unexpected_iid(id); break;
1373 }
1374 set_result(_gvn.transform(n));
1375 return true;
1376 }
1377
1378 //------------------------------inline_trig----------------------------------
1379 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1380 // argument reduction which will turn into a fast/slow diamond.
1381 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1382 Node* arg = round_double_node(argument(0));
1383 Node* n = NULL;
1384
1385 switch (id) {
1386 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1387 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1388 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1389 default: fatal_unexpected_iid(id); break;
1390 }
1391 n = _gvn.transform(n);
1392
1393 // Rounding required? Check for argument reduction!
1394 if (Matcher::strict_fp_requires_explicit_rounding) {
1395 static const double pi_4 = 0.7853981633974483;
1396 static const double neg_pi_4 = -0.7853981633974483;
1397 // pi/2 in 80-bit extended precision
1398 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1399 // -pi/2 in 80-bit extended precision
1400 // 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};
1401 // Cutoff value for using this argument reduction technique
1402 //static const double pi_2_minus_epsilon = 1.564660403643354;
1403 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1404
1405 // Pseudocode for sin:
1406 // if (x <= Math.PI / 4.0) {
1407 // if (x >= -Math.PI / 4.0) return fsin(x);
1408 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1409 // } else {
1410 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1411 // }
1412 // return StrictMath.sin(x);
1413
1414 // Pseudocode for cos:
1415 // if (x <= Math.PI / 4.0) {
1416 // if (x >= -Math.PI / 4.0) return fcos(x);
1417 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1418 // } else {
1419 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1420 // }
1421 // return StrictMath.cos(x);
1422
1423 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1424 // requires a special machine instruction to load it. Instead we'll try
1425 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1426 // probably do the math inside the SIN encoding.
1427
1428 // Make the merge point
1429 RegionNode* r = new RegionNode(3);
1430 Node* phi = new PhiNode(r, Type::DOUBLE);
1431
1432 // Flatten arg so we need only 1 test
1433 Node *abs = _gvn.transform(new AbsDNode(arg));
1434 // Node for PI/4 constant
1435 Node *pi4 = makecon(TypeD::make(pi_4));
1436 // Check PI/4 : abs(arg)
1437 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1438 // Check: If PI/4 < abs(arg) then go slow
1439 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1440 // Branch either way
1441 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1442 set_control(opt_iff(r,iff));
1443
1444 // Set fast path result
1445 phi->init_req(2, n);
1446
1447 // Slow path - non-blocking leaf call
1448 Node* call = NULL;
1449 switch (id) {
1450 case vmIntrinsics::_dsin:
1451 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1452 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1453 "Sin", NULL, arg, top());
1454 break;
1455 case vmIntrinsics::_dcos:
1456 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1457 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1458 "Cos", NULL, arg, top());
1459 break;
1460 case vmIntrinsics::_dtan:
1461 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1462 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1463 "Tan", NULL, arg, top());
1464 break;
1465 }
1466 assert(control()->in(0) == call, "");
1467 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1468 r->init_req(1, control());
1469 phi->init_req(1, slow_result);
1470
1471 // Post-merge
1472 set_control(_gvn.transform(r));
1473 record_for_igvn(r);
1474 n = _gvn.transform(phi);
1475
1476 C->set_has_split_ifs(true); // Has chance for split-if optimization
1477 }
1478 set_result(n);
1479 return true;
1480 }
1481
1482 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1483 //-------------------
1484 //result=(result.isNaN())? funcAddr():result;
1485 // Check: If isNaN() by checking result!=result? then either trap
1486 // or go to runtime
1487 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1488 // Build the boolean node
1489 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1490
1491 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1492 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1493 // The pow or exp intrinsic returned a NaN, which requires a call
1494 // to the runtime. Recompile with the runtime call.
1495 uncommon_trap(Deoptimization::Reason_intrinsic,
1496 Deoptimization::Action_make_not_entrant);
1497 }
1498 return result;
1499 } else {
1500 // If this inlining ever returned NaN in the past, we compile a call
1501 // to the runtime to properly handle corner cases
1502
1503 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1504 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1505 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1506
1507 if (!if_slow->is_top()) {
1508 RegionNode* result_region = new RegionNode(3);
1509 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1510
1511 result_region->init_req(1, if_fast);
1512 result_val->init_req(1, result);
1513
1514 set_control(if_slow);
1515
1516 const TypePtr* no_memory_effects = NULL;
1517 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1518 no_memory_effects,
1519 x, top(), y, y ? top() : NULL);
1520 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1521 #ifdef ASSERT
1522 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1523 assert(value_top == top(), "second value must be top");
1524 #endif
1525
1526 result_region->init_req(2, control());
1527 result_val->init_req(2, value);
1528 set_control(_gvn.transform(result_region));
1529 return _gvn.transform(result_val);
1530 } else {
1531 return result;
1532 }
1533 }
1534 }
1535
1536 //------------------------------inline_pow-------------------------------------
1537 // Inline power instructions, if possible.
1538 bool LibraryCallKit::inline_pow() {
1539 // Pseudocode for pow
1540 // if (y == 2) {
1541 // return x * x;
1542 // } else {
1543 // if (x <= 0.0) {
1544 // long longy = (long)y;
1545 // if ((double)longy == y) { // if y is long
1546 // if (y + 1 == y) longy = 0; // huge number: even
1547 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1548 // } else {
1549 // result = NaN;
1550 // }
1551 // } else {
1552 // result = DPow(x,y);
1553 // }
1554 // if (result != result)? {
1555 // result = uncommon_trap() or runtime_call();
1556 // }
1557 // return result;
1558 // }
1559
1560 Node* x = round_double_node(argument(0));
1561 Node* y = round_double_node(argument(2));
1562
1563 Node* result = NULL;
1564
1565 Node* const_two_node = makecon(TypeD::make(2.0));
1566 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1567 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1568 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1569 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1570 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1571
1572 RegionNode* region_node = new RegionNode(3);
1573 region_node->init_req(1, if_true);
1574
1575 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1576 // special case for x^y where y == 2, we can convert it to x * x
1577 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1578
1579 // set control to if_false since we will now process the false branch
1580 set_control(if_false);
1581
1582 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1583 // Short form: skip the fancy tests and just check for NaN result.
1584 result = _gvn.transform(new PowDNode(C, control(), x, y));
1585 } else {
1586 // If this inlining ever returned NaN in the past, include all
1587 // checks + call to the runtime.
1588
1589 // Set the merge point for If node with condition of (x <= 0.0)
1590 // There are four possible paths to region node and phi node
1591 RegionNode *r = new RegionNode(4);
1592 Node *phi = new PhiNode(r, Type::DOUBLE);
1593
1594 // Build the first if node: if (x <= 0.0)
1595 // Node for 0 constant
1596 Node *zeronode = makecon(TypeD::ZERO);
1597 // Check x:0
1598 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1599 // Check: If (x<=0) then go complex path
1600 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1601 // Branch either way
1602 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1603 // Fast path taken; set region slot 3
1604 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1605 r->init_req(3,fast_taken); // Capture fast-control
1606
1607 // Fast path not-taken, i.e. slow path
1608 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1609
1610 // Set fast path result
1611 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1612 phi->init_req(3, fast_result);
1613
1614 // Complex path
1615 // Build the second if node (if y is long)
1616 // Node for (long)y
1617 Node *longy = _gvn.transform(new ConvD2LNode(y));
1618 // Node for (double)((long) y)
1619 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1620 // Check (double)((long) y) : y
1621 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1622 // Check if (y isn't long) then go to slow path
1623
1624 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1625 // Branch either way
1626 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1627 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1628
1629 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1630
1631 // Calculate DPow(abs(x), y)*(1 & (long)y)
1632 // Node for constant 1
1633 Node *conone = longcon(1);
1634 // 1& (long)y
1635 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1636
1637 // A huge number is always even. Detect a huge number by checking
1638 // if y + 1 == y and set integer to be tested for parity to 0.
1639 // Required for corner case:
1640 // (long)9.223372036854776E18 = max_jlong
1641 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1642 // max_jlong is odd but 9.223372036854776E18 is even
1643 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1644 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1645 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1646 Node* correctedsign = NULL;
1647 if (ConditionalMoveLimit != 0) {
1648 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1649 } else {
1650 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1651 RegionNode *r = new RegionNode(3);
1652 Node *phi = new PhiNode(r, TypeLong::LONG);
1653 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1654 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1655 phi->init_req(1, signnode);
1656 phi->init_req(2, longcon(0));
1657 correctedsign = _gvn.transform(phi);
1658 ylong_path = _gvn.transform(r);
1659 record_for_igvn(r);
1660 }
1661
1662 // zero node
1663 Node *conzero = longcon(0);
1664 // Check (1&(long)y)==0?
1665 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1666 // Check if (1&(long)y)!=0?, if so the result is negative
1667 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1668 // abs(x)
1669 Node *absx=_gvn.transform(new AbsDNode(x));
1670 // abs(x)^y
1671 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1672 // -abs(x)^y
1673 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1674 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1675 Node *signresult = NULL;
1676 if (ConditionalMoveLimit != 0) {
1677 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1678 } else {
1679 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1680 RegionNode *r = new RegionNode(3);
1681 Node *phi = new PhiNode(r, Type::DOUBLE);
1682 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1683 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1684 phi->init_req(1, absxpowy);
1685 phi->init_req(2, negabsxpowy);
1686 signresult = _gvn.transform(phi);
1687 ylong_path = _gvn.transform(r);
1688 record_for_igvn(r);
1689 }
1690 // Set complex path fast result
1691 r->init_req(2, ylong_path);
1692 phi->init_req(2, signresult);
1693
1694 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1695 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1696 r->init_req(1,slow_path);
1697 phi->init_req(1,slow_result);
1698
1699 // Post merge
1700 set_control(_gvn.transform(r));
1701 record_for_igvn(r);
1702 result = _gvn.transform(phi);
1703 }
1704
1705 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1706
1707 // control from finish_pow_exp is now input to the region node
1708 region_node->set_req(2, control());
1709 // the result from finish_pow_exp is now input to the phi node
1710 phi_node->init_req(2, result);
1711 set_control(_gvn.transform(region_node));
1712 record_for_igvn(region_node);
1713 set_result(_gvn.transform(phi_node));
1714
1715 C->set_has_split_ifs(true); // Has chance for split-if optimization
1716 return true;
1717 }
1718
1719 //------------------------------runtime_math-----------------------------
1720 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1721 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1722 "must be (DD)D or (D)D type");
1723
1724 // Inputs
1725 Node* a = round_double_node(argument(0));
1726 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1727
1728 const TypePtr* no_memory_effects = NULL;
1729 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1730 no_memory_effects,
1731 a, top(), b, b ? top() : NULL);
1732 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1733 #ifdef ASSERT
1734 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1735 assert(value_top == top(), "second value must be top");
1736 #endif
1737
1738 set_result(value);
1739 return true;
1740 }
1741
1742 //------------------------------inline_math_native-----------------------------
1743 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1744 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1745 switch (id) {
1746 // These intrinsics are not properly supported on all hardware
1747 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1748 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1749 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1750 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1751 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1752 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1753
1754 case vmIntrinsics::_dlog:
1755 return StubRoutines::dlog() != NULL ?
1756 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dlog(), "dlog") :
1757 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1758 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1759 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1760
1761 // These intrinsics are supported on all hardware
1762 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1763 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1764
1765 case vmIntrinsics::_dexp:
1766 return StubRoutines::dexp() != NULL ?
1767 runtime_math(OptoRuntime::Math_D_D_Type(), StubRoutines::dexp(), "dexp") :
1768 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1769 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1770 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1771 #undef FN_PTR
1772
1773 // These intrinsics are not yet correctly implemented
1774 case vmIntrinsics::_datan2:
1775 return false;
1776
1777 default:
1778 fatal_unexpected_iid(id);
1779 return false;
1780 }
1781 }
1782
1783 static bool is_simple_name(Node* n) {
1784 return (n->req() == 1 // constant
1785 || (n->is_Type() && n->as_Type()->type()->singleton())
1786 || n->is_Proj() // parameter or return value
1787 || n->is_Phi() // local of some sort
1788 );
1789 }
1790
1791 //----------------------------inline_notify-----------------------------------*
1792 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1793 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1794 address func;
1795 if (id == vmIntrinsics::_notify) {
1796 func = OptoRuntime::monitor_notify_Java();
1797 } else {
1798 func = OptoRuntime::monitor_notifyAll_Java();
1799 }
1800 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1801 make_slow_call_ex(call, env()->Throwable_klass(), false);
1802 return true;
1803 }
1804
1805
1806 //----------------------------inline_min_max-----------------------------------
1807 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1808 set_result(generate_min_max(id, argument(0), argument(1)));
1809 return true;
1810 }
1811
1812 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1813 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1814 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1815 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1816 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1817
1818 {
1819 PreserveJVMState pjvms(this);
1820 PreserveReexecuteState preexecs(this);
1821 jvms()->set_should_reexecute(true);
1822
1823 set_control(slow_path);
1824 set_i_o(i_o());
1825
1826 uncommon_trap(Deoptimization::Reason_intrinsic,
1827 Deoptimization::Action_none);
1828 }
1829
1830 set_control(fast_path);
1831 set_result(math);
1832 }
1833
1834 template <typename OverflowOp>
1835 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1836 typedef typename OverflowOp::MathOp MathOp;
1837
1838 MathOp* mathOp = new MathOp(arg1, arg2);
1839 Node* operation = _gvn.transform( mathOp );
1840 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1841 inline_math_mathExact(operation, ofcheck);
1842 return true;
1843 }
1844
1845 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1846 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1847 }
1848
1849 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1850 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1851 }
1852
1853 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1854 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1855 }
1856
1857 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1858 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1859 }
1860
1861 bool LibraryCallKit::inline_math_negateExactI() {
1862 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1863 }
1864
1865 bool LibraryCallKit::inline_math_negateExactL() {
1866 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1867 }
1868
1869 bool LibraryCallKit::inline_math_multiplyExactI() {
1870 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1871 }
1872
1873 bool LibraryCallKit::inline_math_multiplyExactL() {
1874 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1875 }
1876
1877 Node*
1878 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1879 // These are the candidate return value:
1880 Node* xvalue = x0;
1881 Node* yvalue = y0;
1882
1883 if (xvalue == yvalue) {
1884 return xvalue;
1885 }
1886
1887 bool want_max = (id == vmIntrinsics::_max);
1888
1889 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1890 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1891 if (txvalue == NULL || tyvalue == NULL) return top();
1892 // This is not really necessary, but it is consistent with a
1893 // hypothetical MaxINode::Value method:
1894 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1895
1896 // %%% This folding logic should (ideally) be in a different place.
1897 // Some should be inside IfNode, and there to be a more reliable
1898 // transformation of ?: style patterns into cmoves. We also want
1899 // more powerful optimizations around cmove and min/max.
1900
1901 // Try to find a dominating comparison of these guys.
1902 // It can simplify the index computation for Arrays.copyOf
1903 // and similar uses of System.arraycopy.
1904 // First, compute the normalized version of CmpI(x, y).
1905 int cmp_op = Op_CmpI;
1906 Node* xkey = xvalue;
1907 Node* ykey = yvalue;
1908 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1909 if (ideal_cmpxy->is_Cmp()) {
1910 // E.g., if we have CmpI(length - offset, count),
1911 // it might idealize to CmpI(length, count + offset)
1912 cmp_op = ideal_cmpxy->Opcode();
1913 xkey = ideal_cmpxy->in(1);
1914 ykey = ideal_cmpxy->in(2);
1915 }
1916
1917 // Start by locating any relevant comparisons.
1918 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1919 Node* cmpxy = NULL;
1920 Node* cmpyx = NULL;
1921 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1922 Node* cmp = start_from->fast_out(k);
1923 if (cmp->outcnt() > 0 && // must have prior uses
1924 cmp->in(0) == NULL && // must be context-independent
1925 cmp->Opcode() == cmp_op) { // right kind of compare
1926 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1927 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1928 }
1929 }
1930
1931 const int NCMPS = 2;
1932 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1933 int cmpn;
1934 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1935 if (cmps[cmpn] != NULL) break; // find a result
1936 }
1937 if (cmpn < NCMPS) {
1938 // Look for a dominating test that tells us the min and max.
1939 int depth = 0; // Limit search depth for speed
1940 Node* dom = control();
1941 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1942 if (++depth >= 100) break;
1943 Node* ifproj = dom;
1944 if (!ifproj->is_Proj()) continue;
1945 Node* iff = ifproj->in(0);
1946 if (!iff->is_If()) continue;
1947 Node* bol = iff->in(1);
1948 if (!bol->is_Bool()) continue;
1949 Node* cmp = bol->in(1);
1950 if (cmp == NULL) continue;
1951 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1952 if (cmps[cmpn] == cmp) break;
1953 if (cmpn == NCMPS) continue;
1954 BoolTest::mask btest = bol->as_Bool()->_test._test;
1955 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1956 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1957 // At this point, we know that 'x btest y' is true.
1958 switch (btest) {
1959 case BoolTest::eq:
1960 // They are proven equal, so we can collapse the min/max.
1961 // Either value is the answer. Choose the simpler.
1962 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1963 return yvalue;
1964 return xvalue;
1965 case BoolTest::lt: // x < y
1966 case BoolTest::le: // x <= y
1967 return (want_max ? yvalue : xvalue);
1968 case BoolTest::gt: // x > y
1969 case BoolTest::ge: // x >= y
1970 return (want_max ? xvalue : yvalue);
1971 }
1972 }
1973 }
1974
1975 // We failed to find a dominating test.
1976 // Let's pick a test that might GVN with prior tests.
1977 Node* best_bol = NULL;
1978 BoolTest::mask best_btest = BoolTest::illegal;
1979 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1980 Node* cmp = cmps[cmpn];
1981 if (cmp == NULL) continue;
1982 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1983 Node* bol = cmp->fast_out(j);
1984 if (!bol->is_Bool()) continue;
1985 BoolTest::mask btest = bol->as_Bool()->_test._test;
1986 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1987 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1988 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1989 best_bol = bol->as_Bool();
1990 best_btest = btest;
1991 }
1992 }
1993 }
1994
1995 Node* answer_if_true = NULL;
1996 Node* answer_if_false = NULL;
1997 switch (best_btest) {
1998 default:
1999 if (cmpxy == NULL)
2000 cmpxy = ideal_cmpxy;
2001 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2002 // and fall through:
2003 case BoolTest::lt: // x < y
2004 case BoolTest::le: // x <= y
2005 answer_if_true = (want_max ? yvalue : xvalue);
2006 answer_if_false = (want_max ? xvalue : yvalue);
2007 break;
2008 case BoolTest::gt: // x > y
2009 case BoolTest::ge: // x >= y
2010 answer_if_true = (want_max ? xvalue : yvalue);
2011 answer_if_false = (want_max ? yvalue : xvalue);
2012 break;
2013 }
2014
2015 jint hi, lo;
2016 if (want_max) {
2017 // We can sharpen the minimum.
2018 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2019 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2020 } else {
2021 // We can sharpen the maximum.
2022 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2023 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2024 }
2025
2026 // Use a flow-free graph structure, to avoid creating excess control edges
2027 // which could hinder other optimizations.
2028 // Since Math.min/max is often used with arraycopy, we want
2029 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2030 Node* cmov = CMoveNode::make(NULL, best_bol,
2031 answer_if_false, answer_if_true,
2032 TypeInt::make(lo, hi, widen));
2033
2034 return _gvn.transform(cmov);
2035
2036 /*
2037 // This is not as desirable as it may seem, since Min and Max
2038 // nodes do not have a full set of optimizations.
2039 // And they would interfere, anyway, with 'if' optimizations
2040 // and with CMoveI canonical forms.
2041 switch (id) {
2042 case vmIntrinsics::_min:
2043 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2044 case vmIntrinsics::_max:
2045 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2046 default:
2047 ShouldNotReachHere();
2048 }
2049 */
2050 }
2051
2052 inline int
2053 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2054 const TypePtr* base_type = TypePtr::NULL_PTR;
2055 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2056 if (base_type == NULL) {
2057 // Unknown type.
2058 return Type::AnyPtr;
2059 } else if (base_type == TypePtr::NULL_PTR) {
2060 // Since this is a NULL+long form, we have to switch to a rawptr.
2061 base = _gvn.transform(new CastX2PNode(offset));
2062 offset = MakeConX(0);
2063 return Type::RawPtr;
2064 } else if (base_type->base() == Type::RawPtr) {
2065 return Type::RawPtr;
2066 } else if (base_type->isa_oopptr()) {
2067 // Base is never null => always a heap address.
2068 if (base_type->ptr() == TypePtr::NotNull) {
2069 return Type::OopPtr;
2070 }
2071 // Offset is small => always a heap address.
2072 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2073 if (offset_type != NULL &&
2074 base_type->offset() == 0 && // (should always be?)
2075 offset_type->_lo >= 0 &&
2076 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2077 return Type::OopPtr;
2078 }
2079 // Otherwise, it might either be oop+off or NULL+addr.
2080 return Type::AnyPtr;
2081 } else {
2082 // No information:
2083 return Type::AnyPtr;
2084 }
2085 }
2086
2087 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2088 int kind = classify_unsafe_addr(base, offset);
2089 if (kind == Type::RawPtr) {
2090 return basic_plus_adr(top(), base, offset);
2091 } else {
2092 return basic_plus_adr(base, offset);
2093 }
2094 }
2095
2096 //--------------------------inline_number_methods-----------------------------
2097 // inline int Integer.numberOfLeadingZeros(int)
2098 // inline int Long.numberOfLeadingZeros(long)
2099 //
2100 // inline int Integer.numberOfTrailingZeros(int)
2101 // inline int Long.numberOfTrailingZeros(long)
2102 //
2103 // inline int Integer.bitCount(int)
2104 // inline int Long.bitCount(long)
2105 //
2106 // inline char Character.reverseBytes(char)
2107 // inline short Short.reverseBytes(short)
2108 // inline int Integer.reverseBytes(int)
2109 // inline long Long.reverseBytes(long)
2110 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2111 Node* arg = argument(0);
2112 Node* n;
2113 switch (id) {
2114 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2115 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2116 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2117 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2118 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2119 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2120 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2121 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2122 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2123 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2124 default: fatal_unexpected_iid(id); break;
2125 }
2126 set_result(_gvn.transform(n));
2127 return true;
2128 }
2129
2130 //----------------------------inline_unsafe_access----------------------------
2131
2132 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2133
2134 // Helper that guards and inserts a pre-barrier.
2135 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2136 Node* pre_val, bool need_mem_bar) {
2137 // We could be accessing the referent field of a reference object. If so, when G1
2138 // is enabled, we need to log the value in the referent field in an SATB buffer.
2139 // This routine performs some compile time filters and generates suitable
2140 // runtime filters that guard the pre-barrier code.
2141 // Also add memory barrier for non volatile load from the referent field
2142 // to prevent commoning of loads across safepoint.
2143 if (!UseG1GC && !need_mem_bar)
2144 return;
2145
2146 // Some compile time checks.
2147
2148 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2149 const TypeX* otype = offset->find_intptr_t_type();
2150 if (otype != NULL && otype->is_con() &&
2151 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2152 // Constant offset but not the reference_offset so just return
2153 return;
2154 }
2155
2156 // We only need to generate the runtime guards for instances.
2157 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2158 if (btype != NULL) {
2159 if (btype->isa_aryptr()) {
2160 // Array type so nothing to do
2161 return;
2162 }
2163
2164 const TypeInstPtr* itype = btype->isa_instptr();
2165 if (itype != NULL) {
2166 // Can the klass of base_oop be statically determined to be
2167 // _not_ a sub-class of Reference and _not_ Object?
2168 ciKlass* klass = itype->klass();
2169 if ( klass->is_loaded() &&
2170 !klass->is_subtype_of(env()->Reference_klass()) &&
2171 !env()->Object_klass()->is_subtype_of(klass)) {
2172 return;
2173 }
2174 }
2175 }
2176
2177 // The compile time filters did not reject base_oop/offset so
2178 // we need to generate the following runtime filters
2179 //
2180 // if (offset == java_lang_ref_Reference::_reference_offset) {
2181 // if (instance_of(base, java.lang.ref.Reference)) {
2182 // pre_barrier(_, pre_val, ...);
2183 // }
2184 // }
2185
2186 float likely = PROB_LIKELY( 0.999);
2187 float unlikely = PROB_UNLIKELY(0.999);
2188
2189 IdealKit ideal(this);
2190 #define __ ideal.
2191
2192 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2193
2194 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2195 // Update graphKit memory and control from IdealKit.
2196 sync_kit(ideal);
2197
2198 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2199 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2200
2201 // Update IdealKit memory and control from graphKit.
2202 __ sync_kit(this);
2203
2204 Node* one = __ ConI(1);
2205 // is_instof == 0 if base_oop == NULL
2206 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2207
2208 // Update graphKit from IdeakKit.
2209 sync_kit(ideal);
2210
2211 // Use the pre-barrier to record the value in the referent field
2212 pre_barrier(false /* do_load */,
2213 __ ctrl(),
2214 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2215 pre_val /* pre_val */,
2216 T_OBJECT);
2217 if (need_mem_bar) {
2218 // Add memory barrier to prevent commoning reads from this field
2219 // across safepoint since GC can change its value.
2220 insert_mem_bar(Op_MemBarCPUOrder);
2221 }
2222 // Update IdealKit from graphKit.
2223 __ sync_kit(this);
2224
2225 } __ end_if(); // _ref_type != ref_none
2226 } __ end_if(); // offset == referent_offset
2227
2228 // Final sync IdealKit and GraphKit.
2229 final_sync(ideal);
2230 #undef __
2231 }
2232
2233
2234 // Interpret Unsafe.fieldOffset cookies correctly:
2235 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2236
2237 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2238 // Attempt to infer a sharper value type from the offset and base type.
2239 ciKlass* sharpened_klass = NULL;
2240
2241 // See if it is an instance field, with an object type.
2242 if (alias_type->field() != NULL) {
2243 assert(!is_native_ptr, "native pointer op cannot use a java address");
2244 if (alias_type->field()->type()->is_klass()) {
2245 sharpened_klass = alias_type->field()->type()->as_klass();
2246 }
2247 }
2248
2249 // See if it is a narrow oop array.
2250 if (adr_type->isa_aryptr()) {
2251 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2252 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2253 if (elem_type != NULL) {
2254 sharpened_klass = elem_type->klass();
2255 }
2256 }
2257 }
2258
2259 // The sharpened class might be unloaded if there is no class loader
2260 // contraint in place.
2261 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2262 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2263
2264 #ifndef PRODUCT
2265 if (C->print_intrinsics() || C->print_inlining()) {
2266 tty->print(" from base type: "); adr_type->dump();
2267 tty->print(" sharpened value: "); tjp->dump();
2268 }
2269 #endif
2270 // Sharpen the value type.
2271 return tjp;
2272 }
2273 return NULL;
2274 }
2275
2276 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2277 if (callee()->is_static()) return false; // caller must have the capability!
2278
2279 #ifndef PRODUCT
2280 {
2281 ResourceMark rm;
2282 // Check the signatures.
2283 ciSignature* sig = callee()->signature();
2284 #ifdef ASSERT
2285 if (!is_store) {
2286 // Object getObject(Object base, int/long offset), etc.
2287 BasicType rtype = sig->return_type()->basic_type();
2288 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2289 rtype = T_ADDRESS; // it is really a C void*
2290 assert(rtype == type, "getter must return the expected value");
2291 if (!is_native_ptr) {
2292 assert(sig->count() == 2, "oop getter has 2 arguments");
2293 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2294 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2295 } else {
2296 assert(sig->count() == 1, "native getter has 1 argument");
2297 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2298 }
2299 } else {
2300 // void putObject(Object base, int/long offset, Object x), etc.
2301 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2302 if (!is_native_ptr) {
2303 assert(sig->count() == 3, "oop putter has 3 arguments");
2304 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2305 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2306 } else {
2307 assert(sig->count() == 2, "native putter has 2 arguments");
2308 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2309 }
2310 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2311 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2312 vtype = T_ADDRESS; // it is really a C void*
2313 assert(vtype == type, "putter must accept the expected value");
2314 }
2315 #endif // ASSERT
2316 }
2317 #endif //PRODUCT
2318
2319 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2320
2321 Node* receiver = argument(0); // type: oop
2322
2323 // Build address expression.
2324 Node* adr;
2325 Node* heap_base_oop = top();
2326 Node* offset = top();
2327 Node* val;
2328
2329 if (!is_native_ptr) {
2330 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2331 Node* base = argument(1); // type: oop
2332 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2333 offset = argument(2); // type: long
2334 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2335 // to be plain byte offsets, which are also the same as those accepted
2336 // by oopDesc::field_base.
2337 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2338 "fieldOffset must be byte-scaled");
2339 // 32-bit machines ignore the high half!
2340 offset = ConvL2X(offset);
2341 adr = make_unsafe_address(base, offset);
2342 heap_base_oop = base;
2343 val = is_store ? argument(4) : NULL;
2344 } else {
2345 Node* ptr = argument(1); // type: long
2346 ptr = ConvL2X(ptr); // adjust Java long to machine word
2347 adr = make_unsafe_address(NULL, ptr);
2348 val = is_store ? argument(3) : NULL;
2349 }
2350
2351 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2352
2353 // First guess at the value type.
2354 const Type *value_type = Type::get_const_basic_type(type);
2355
2356 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2357 // there was not enough information to nail it down.
2358 Compile::AliasType* alias_type = C->alias_type(adr_type);
2359 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2360
2361 // We will need memory barriers unless we can determine a unique
2362 // alias category for this reference. (Note: If for some reason
2363 // the barriers get omitted and the unsafe reference begins to "pollute"
2364 // the alias analysis of the rest of the graph, either Compile::can_alias
2365 // or Compile::must_alias will throw a diagnostic assert.)
2366 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2367
2368 // If we are reading the value of the referent field of a Reference
2369 // object (either by using Unsafe directly or through reflection)
2370 // then, if G1 is enabled, we need to record the referent in an
2371 // SATB log buffer using the pre-barrier mechanism.
2372 // Also we need to add memory barrier to prevent commoning reads
2373 // from this field across safepoint since GC can change its value.
2374 bool need_read_barrier = !is_native_ptr && !is_store &&
2375 offset != top() && heap_base_oop != top();
2376
2377 if (!is_store && type == T_OBJECT) {
2378 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2379 if (tjp != NULL) {
2380 value_type = tjp;
2381 }
2382 }
2383
2384 receiver = null_check(receiver);
2385 if (stopped()) {
2386 return true;
2387 }
2388 // Heap pointers get a null-check from the interpreter,
2389 // as a courtesy. However, this is not guaranteed by Unsafe,
2390 // and it is not possible to fully distinguish unintended nulls
2391 // from intended ones in this API.
2392
2393 if (is_volatile) {
2394 // We need to emit leading and trailing CPU membars (see below) in
2395 // addition to memory membars when is_volatile. This is a little
2396 // too strong, but avoids the need to insert per-alias-type
2397 // volatile membars (for stores; compare Parse::do_put_xxx), which
2398 // we cannot do effectively here because we probably only have a
2399 // rough approximation of type.
2400 need_mem_bar = true;
2401 // For Stores, place a memory ordering barrier now.
2402 if (is_store) {
2403 insert_mem_bar(Op_MemBarRelease);
2404 } else {
2405 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2406 insert_mem_bar(Op_MemBarVolatile);
2407 }
2408 }
2409 }
2410
2411 // Memory barrier to prevent normal and 'unsafe' accesses from
2412 // bypassing each other. Happens after null checks, so the
2413 // exception paths do not take memory state from the memory barrier,
2414 // so there's no problems making a strong assert about mixing users
2415 // of safe & unsafe memory.
2416 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2417
2418 if (!is_store) {
2419 Node* p = NULL;
2420 // Try to constant fold a load from a constant field
2421 ciField* field = alias_type->field();
2422 if (heap_base_oop != top() &&
2423 field != NULL && field->is_constant() && field->layout_type() == type) {
2424 // final or stable field
2425 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2426 if (con_type != NULL) {
2427 p = makecon(con_type);
2428 }
2429 }
2430 if (p == NULL) {
2431 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2432 // To be valid, unsafe loads may depend on other conditions than
2433 // the one that guards them: pin the Load node
2434 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2435 // load value
2436 switch (type) {
2437 case T_BOOLEAN:
2438 case T_CHAR:
2439 case T_BYTE:
2440 case T_SHORT:
2441 case T_INT:
2442 case T_LONG:
2443 case T_FLOAT:
2444 case T_DOUBLE:
2445 break;
2446 case T_OBJECT:
2447 if (need_read_barrier) {
2448 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2449 }
2450 break;
2451 case T_ADDRESS:
2452 // Cast to an int type.
2453 p = _gvn.transform(new CastP2XNode(NULL, p));
2454 p = ConvX2UL(p);
2455 break;
2456 default:
2457 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2458 break;
2459 }
2460 }
2461 // The load node has the control of the preceding MemBarCPUOrder. All
2462 // following nodes will have the control of the MemBarCPUOrder inserted at
2463 // the end of this method. So, pushing the load onto the stack at a later
2464 // point is fine.
2465 set_result(p);
2466 } else {
2467 // place effect of store into memory
2468 switch (type) {
2469 case T_DOUBLE:
2470 val = dstore_rounding(val);
2471 break;
2472 case T_ADDRESS:
2473 // Repackage the long as a pointer.
2474 val = ConvL2X(val);
2475 val = _gvn.transform(new CastX2PNode(val));
2476 break;
2477 }
2478
2479 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2480 if (type != T_OBJECT ) {
2481 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2482 } else {
2483 // Possibly an oop being stored to Java heap or native memory
2484 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2485 // oop to Java heap.
2486 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2487 } else {
2488 // We can't tell at compile time if we are storing in the Java heap or outside
2489 // of it. So we need to emit code to conditionally do the proper type of
2490 // store.
2491
2492 IdealKit ideal(this);
2493 #define __ ideal.
2494 // QQQ who knows what probability is here??
2495 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2496 // Sync IdealKit and graphKit.
2497 sync_kit(ideal);
2498 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2499 // Update IdealKit memory.
2500 __ sync_kit(this);
2501 } __ else_(); {
2502 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2503 } __ end_if();
2504 // Final sync IdealKit and GraphKit.
2505 final_sync(ideal);
2506 #undef __
2507 }
2508 }
2509 }
2510
2511 if (is_volatile) {
2512 if (!is_store) {
2513 insert_mem_bar(Op_MemBarAcquire);
2514 } else {
2515 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2516 insert_mem_bar(Op_MemBarVolatile);
2517 }
2518 }
2519 }
2520
2521 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2522
2523 return true;
2524 }
2525
2526 //----------------------------inline_unsafe_load_store----------------------------
2527 // This method serves a couple of different customers (depending on LoadStoreKind):
2528 //
2529 // LS_cmpxchg:
2530 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2531 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2532 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2533 //
2534 // LS_xadd:
2535 // public int getAndAddInt( Object o, long offset, int delta)
2536 // public long getAndAddLong(Object o, long offset, long delta)
2537 //
2538 // LS_xchg:
2539 // int getAndSet(Object o, long offset, int newValue)
2540 // long getAndSet(Object o, long offset, long newValue)
2541 // Object getAndSet(Object o, long offset, Object newValue)
2542 //
2543 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2544 // This basic scheme here is the same as inline_unsafe_access, but
2545 // differs in enough details that combining them would make the code
2546 // overly confusing. (This is a true fact! I originally combined
2547 // them, but even I was confused by it!) As much code/comments as
2548 // possible are retained from inline_unsafe_access though to make
2549 // the correspondences clearer. - dl
2550
2551 if (callee()->is_static()) return false; // caller must have the capability!
2552
2553 #ifndef PRODUCT
2554 BasicType rtype;
2555 {
2556 ResourceMark rm;
2557 // Check the signatures.
2558 ciSignature* sig = callee()->signature();
2559 rtype = sig->return_type()->basic_type();
2560 if (kind == LS_xadd || kind == LS_xchg) {
2561 // Check the signatures.
2562 #ifdef ASSERT
2563 assert(rtype == type, "get and set must return the expected type");
2564 assert(sig->count() == 3, "get and set has 3 arguments");
2565 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2566 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2567 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2568 #endif // ASSERT
2569 } else if (kind == LS_cmpxchg) {
2570 // Check the signatures.
2571 #ifdef ASSERT
2572 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2573 assert(sig->count() == 4, "CAS has 4 arguments");
2574 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2575 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2576 #endif // ASSERT
2577 } else {
2578 ShouldNotReachHere();
2579 }
2580 }
2581 #endif //PRODUCT
2582
2583 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2584
2585 // Get arguments:
2586 Node* receiver = NULL;
2587 Node* base = NULL;
2588 Node* offset = NULL;
2589 Node* oldval = NULL;
2590 Node* newval = NULL;
2591 if (kind == LS_cmpxchg) {
2592 const bool two_slot_type = type2size[type] == 2;
2593 receiver = argument(0); // type: oop
2594 base = argument(1); // type: oop
2595 offset = argument(2); // type: long
2596 oldval = argument(4); // type: oop, int, or long
2597 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2598 } else if (kind == LS_xadd || kind == LS_xchg){
2599 receiver = argument(0); // type: oop
2600 base = argument(1); // type: oop
2601 offset = argument(2); // type: long
2602 oldval = NULL;
2603 newval = argument(4); // type: oop, int, or long
2604 }
2605
2606 // Null check receiver.
2607 receiver = null_check(receiver);
2608 if (stopped()) {
2609 return true;
2610 }
2611
2612 // Build field offset expression.
2613 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2614 // to be plain byte offsets, which are also the same as those accepted
2615 // by oopDesc::field_base.
2616 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2617 // 32-bit machines ignore the high half of long offsets
2618 offset = ConvL2X(offset);
2619 Node* adr = make_unsafe_address(base, offset);
2620 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2621
2622 // For CAS, unlike inline_unsafe_access, there seems no point in
2623 // trying to refine types. Just use the coarse types here.
2624 const Type *value_type = Type::get_const_basic_type(type);
2625 Compile::AliasType* alias_type = C->alias_type(adr_type);
2626 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2627
2628 if (kind == LS_xchg && type == T_OBJECT) {
2629 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2630 if (tjp != NULL) {
2631 value_type = tjp;
2632 }
2633 }
2634
2635 int alias_idx = C->get_alias_index(adr_type);
2636
2637 // Memory-model-wise, a LoadStore acts like a little synchronized
2638 // block, so needs barriers on each side. These don't translate
2639 // into actual barriers on most machines, but we still need rest of
2640 // compiler to respect ordering.
2641
2642 insert_mem_bar(Op_MemBarRelease);
2643 insert_mem_bar(Op_MemBarCPUOrder);
2644
2645 // 4984716: MemBars must be inserted before this
2646 // memory node in order to avoid a false
2647 // dependency which will confuse the scheduler.
2648 Node *mem = memory(alias_idx);
2649
2650 // For now, we handle only those cases that actually exist: ints,
2651 // longs, and Object. Adding others should be straightforward.
2652 Node* load_store;
2653 switch(type) {
2654 case T_INT:
2655 if (kind == LS_xadd) {
2656 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2657 } else if (kind == LS_xchg) {
2658 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2659 } else if (kind == LS_cmpxchg) {
2660 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2661 } else {
2662 ShouldNotReachHere();
2663 }
2664 break;
2665 case T_LONG:
2666 if (kind == LS_xadd) {
2667 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2668 } else if (kind == LS_xchg) {
2669 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2670 } else if (kind == LS_cmpxchg) {
2671 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2672 } else {
2673 ShouldNotReachHere();
2674 }
2675 break;
2676 case T_OBJECT:
2677 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2678 // could be delayed during Parse (for example, in adjust_map_after_if()).
2679 // Execute transformation here to avoid barrier generation in such case.
2680 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2681 newval = _gvn.makecon(TypePtr::NULL_PTR);
2682
2683 // Reference stores need a store barrier.
2684 if (kind == LS_xchg) {
2685 // If pre-barrier must execute before the oop store, old value will require do_load here.
2686 if (!can_move_pre_barrier()) {
2687 pre_barrier(true /* do_load*/,
2688 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2689 NULL /* pre_val*/,
2690 T_OBJECT);
2691 } // Else move pre_barrier to use load_store value, see below.
2692 } else if (kind == LS_cmpxchg) {
2693 // Same as for newval above:
2694 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2695 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2696 }
2697 // The only known value which might get overwritten is oldval.
2698 pre_barrier(false /* do_load */,
2699 control(), NULL, NULL, max_juint, NULL, NULL,
2700 oldval /* pre_val */,
2701 T_OBJECT);
2702 } else {
2703 ShouldNotReachHere();
2704 }
2705
2706 #ifdef _LP64
2707 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2708 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2709 if (kind == LS_xchg) {
2710 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2711 newval_enc, adr_type, value_type->make_narrowoop()));
2712 } else {
2713 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2714 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2715 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2716 newval_enc, oldval_enc));
2717 }
2718 } else
2719 #endif
2720 {
2721 if (kind == LS_xchg) {
2722 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2723 } else {
2724 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2725 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2726 }
2727 }
2728 if (kind == LS_cmpxchg) {
2729 // Emit the post barrier only when the actual store happened.
2730 // This makes sense to check only for compareAndSet that can fail to set the value.
2731 // CAS success path is marked more likely since we anticipate this is a performance
2732 // critical path, while CAS failure path can use the penalty for going through unlikely
2733 // path as backoff. Which is still better than doing a store barrier there.
2734 IdealKit ideal(this);
2735 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2736 sync_kit(ideal);
2737 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2738 ideal.sync_kit(this);
2739 } ideal.end_if();
2740 final_sync(ideal);
2741 } else {
2742 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2743 }
2744 break;
2745 default:
2746 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2747 break;
2748 }
2749
2750 // SCMemProjNodes represent the memory state of a LoadStore. Their
2751 // main role is to prevent LoadStore nodes from being optimized away
2752 // when their results aren't used.
2753 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2754 set_memory(proj, alias_idx);
2755
2756 if (type == T_OBJECT && kind == LS_xchg) {
2757 #ifdef _LP64
2758 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2759 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2760 }
2761 #endif
2762 if (can_move_pre_barrier()) {
2763 // Don't need to load pre_val. The old value is returned by load_store.
2764 // The pre_barrier can execute after the xchg as long as no safepoint
2765 // gets inserted between them.
2766 pre_barrier(false /* do_load */,
2767 control(), NULL, NULL, max_juint, NULL, NULL,
2768 load_store /* pre_val */,
2769 T_OBJECT);
2770 }
2771 }
2772
2773 // Add the trailing membar surrounding the access
2774 insert_mem_bar(Op_MemBarCPUOrder);
2775 insert_mem_bar(Op_MemBarAcquire);
2776
2777 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2778 set_result(load_store);
2779 return true;
2780 }
2781
2782 //----------------------------inline_unsafe_ordered_store----------------------
2783 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2784 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2785 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2786 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2787 // This is another variant of inline_unsafe_access, differing in
2788 // that it always issues store-store ("release") barrier and ensures
2789 // store-atomicity (which only matters for "long").
2790
2791 if (callee()->is_static()) return false; // caller must have the capability!
2792
2793 #ifndef PRODUCT
2794 {
2795 ResourceMark rm;
2796 // Check the signatures.
2797 ciSignature* sig = callee()->signature();
2798 #ifdef ASSERT
2799 BasicType rtype = sig->return_type()->basic_type();
2800 assert(rtype == T_VOID, "must return void");
2801 assert(sig->count() == 3, "has 3 arguments");
2802 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2803 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2804 #endif // ASSERT
2805 }
2806 #endif //PRODUCT
2807
2808 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2809
2810 // Get arguments:
2811 Node* receiver = argument(0); // type: oop
2812 Node* base = argument(1); // type: oop
2813 Node* offset = argument(2); // type: long
2814 Node* val = argument(4); // type: oop, int, or long
2815
2816 // Null check receiver.
2817 receiver = null_check(receiver);
2818 if (stopped()) {
2819 return true;
2820 }
2821
2822 // Build field offset expression.
2823 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2824 // 32-bit machines ignore the high half of long offsets
2825 offset = ConvL2X(offset);
2826 Node* adr = make_unsafe_address(base, offset);
2827 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2828 const Type *value_type = Type::get_const_basic_type(type);
2829 Compile::AliasType* alias_type = C->alias_type(adr_type);
2830
2831 insert_mem_bar(Op_MemBarRelease);
2832 insert_mem_bar(Op_MemBarCPUOrder);
2833 // Ensure that the store is atomic for longs:
2834 const bool require_atomic_access = true;
2835 Node* store;
2836 if (type == T_OBJECT) // reference stores need a store barrier.
2837 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2838 else {
2839 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2840 }
2841 insert_mem_bar(Op_MemBarCPUOrder);
2842 return true;
2843 }
2844
2845 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2846 // Regardless of form, don't allow previous ld/st to move down,
2847 // then issue acquire, release, or volatile mem_bar.
2848 insert_mem_bar(Op_MemBarCPUOrder);
2849 switch(id) {
2850 case vmIntrinsics::_loadFence:
2851 insert_mem_bar(Op_LoadFence);
2852 return true;
2853 case vmIntrinsics::_storeFence:
2854 insert_mem_bar(Op_StoreFence);
2855 return true;
2856 case vmIntrinsics::_fullFence:
2857 insert_mem_bar(Op_MemBarVolatile);
2858 return true;
2859 default:
2860 fatal_unexpected_iid(id);
2861 return false;
2862 }
2863 }
2864
2865 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2866 if (!kls->is_Con()) {
2867 return true;
2868 }
2869 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2870 if (klsptr == NULL) {
2871 return true;
2872 }
2873 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2874 // don't need a guard for a klass that is already initialized
2875 return !ik->is_initialized();
2876 }
2877
2878 //----------------------------inline_unsafe_allocate---------------------------
2879 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2880 bool LibraryCallKit::inline_unsafe_allocate() {
2881 if (callee()->is_static()) return false; // caller must have the capability!
2882
2883 null_check_receiver(); // null-check, then ignore
2884 Node* cls = null_check(argument(1));
2885 if (stopped()) return true;
2886
2887 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2888 kls = null_check(kls);
2889 if (stopped()) return true; // argument was like int.class
2890
2891 Node* test = NULL;
2892 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2893 // Note: The argument might still be an illegal value like
2894 // Serializable.class or Object[].class. The runtime will handle it.
2895 // But we must make an explicit check for initialization.
2896 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2897 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2898 // can generate code to load it as unsigned byte.
2899 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2900 Node* bits = intcon(InstanceKlass::fully_initialized);
2901 test = _gvn.transform(new SubINode(inst, bits));
2902 // The 'test' is non-zero if we need to take a slow path.
2903 }
2904
2905 Node* obj = new_instance(kls, test);
2906 set_result(obj);
2907 return true;
2908 }
2909
2910 #ifdef TRACE_HAVE_INTRINSICS
2911 /*
2912 * oop -> myklass
2913 * myklass->trace_id |= USED
2914 * return myklass->trace_id & ~0x3
2915 */
2916 bool LibraryCallKit::inline_native_classID() {
2917 null_check_receiver(); // null-check, then ignore
2918 Node* cls = null_check(argument(1), T_OBJECT);
2919 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2920 kls = null_check(kls, T_OBJECT);
2921 ByteSize offset = TRACE_ID_OFFSET;
2922 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2923 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2924 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2925 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2926 Node* clsused = longcon(0x01l); // set the class bit
2927 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2928
2929 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2930 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2931 set_result(andl);
2932 return true;
2933 }
2934
2935 bool LibraryCallKit::inline_native_threadID() {
2936 Node* tls_ptr = NULL;
2937 Node* cur_thr = generate_current_thread(tls_ptr);
2938 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2939 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2940 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2941
2942 Node* threadid = NULL;
2943 size_t thread_id_size = OSThread::thread_id_size();
2944 if (thread_id_size == (size_t) BytesPerLong) {
2945 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2946 } else if (thread_id_size == (size_t) BytesPerInt) {
2947 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2948 } else {
2949 ShouldNotReachHere();
2950 }
2951 set_result(threadid);
2952 return true;
2953 }
2954 #endif
2955
2956 //------------------------inline_native_time_funcs--------------
2957 // inline code for System.currentTimeMillis() and System.nanoTime()
2958 // these have the same type and signature
2959 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2960 const TypeFunc* tf = OptoRuntime::void_long_Type();
2961 const TypePtr* no_memory_effects = NULL;
2962 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2963 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2964 #ifdef ASSERT
2965 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2966 assert(value_top == top(), "second value must be top");
2967 #endif
2968 set_result(value);
2969 return true;
2970 }
2971
2972 //------------------------inline_native_currentThread------------------
2973 bool LibraryCallKit::inline_native_currentThread() {
2974 Node* junk = NULL;
2975 set_result(generate_current_thread(junk));
2976 return true;
2977 }
2978
2979 //------------------------inline_native_isInterrupted------------------
2980 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2981 bool LibraryCallKit::inline_native_isInterrupted() {
2982 // Add a fast path to t.isInterrupted(clear_int):
2983 // (t == Thread.current() &&
2984 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2985 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2986 // So, in the common case that the interrupt bit is false,
2987 // we avoid making a call into the VM. Even if the interrupt bit
2988 // is true, if the clear_int argument is false, we avoid the VM call.
2989 // However, if the receiver is not currentThread, we must call the VM,
2990 // because there must be some locking done around the operation.
2991
2992 // We only go to the fast case code if we pass two guards.
2993 // Paths which do not pass are accumulated in the slow_region.
2994
2995 enum {
2996 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2997 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2998 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2999 PATH_LIMIT
3000 };
3001
3002 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3003 // out of the function.
3004 insert_mem_bar(Op_MemBarCPUOrder);
3005
3006 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3007 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3008
3009 RegionNode* slow_region = new RegionNode(1);
3010 record_for_igvn(slow_region);
3011
3012 // (a) Receiving thread must be the current thread.
3013 Node* rec_thr = argument(0);
3014 Node* tls_ptr = NULL;
3015 Node* cur_thr = generate_current_thread(tls_ptr);
3016 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3017 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3018
3019 generate_slow_guard(bol_thr, slow_region);
3020
3021 // (b) Interrupt bit on TLS must be false.
3022 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3023 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3024 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3025
3026 // Set the control input on the field _interrupted read to prevent it floating up.
3027 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3028 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3029 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3030
3031 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3032
3033 // First fast path: if (!TLS._interrupted) return false;
3034 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3035 result_rgn->init_req(no_int_result_path, false_bit);
3036 result_val->init_req(no_int_result_path, intcon(0));
3037
3038 // drop through to next case
3039 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3040
3041 #ifndef TARGET_OS_FAMILY_windows
3042 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3043 Node* clr_arg = argument(1);
3044 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3045 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3046 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3047
3048 // Second fast path: ... else if (!clear_int) return true;
3049 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3050 result_rgn->init_req(no_clear_result_path, false_arg);
3051 result_val->init_req(no_clear_result_path, intcon(1));
3052
3053 // drop through to next case
3054 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3055 #else
3056 // To return true on Windows you must read the _interrupted field
3057 // and check the the event state i.e. take the slow path.
3058 #endif // TARGET_OS_FAMILY_windows
3059
3060 // (d) Otherwise, go to the slow path.
3061 slow_region->add_req(control());
3062 set_control( _gvn.transform(slow_region));
3063
3064 if (stopped()) {
3065 // There is no slow path.
3066 result_rgn->init_req(slow_result_path, top());
3067 result_val->init_req(slow_result_path, top());
3068 } else {
3069 // non-virtual because it is a private non-static
3070 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3071
3072 Node* slow_val = set_results_for_java_call(slow_call);
3073 // this->control() comes from set_results_for_java_call
3074
3075 Node* fast_io = slow_call->in(TypeFunc::I_O);
3076 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3077
3078 // These two phis are pre-filled with copies of of the fast IO and Memory
3079 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3080 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3081
3082 result_rgn->init_req(slow_result_path, control());
3083 result_io ->init_req(slow_result_path, i_o());
3084 result_mem->init_req(slow_result_path, reset_memory());
3085 result_val->init_req(slow_result_path, slow_val);
3086
3087 set_all_memory(_gvn.transform(result_mem));
3088 set_i_o( _gvn.transform(result_io));
3089 }
3090
3091 C->set_has_split_ifs(true); // Has chance for split-if optimization
3092 set_result(result_rgn, result_val);
3093 return true;
3094 }
3095
3096 //---------------------------load_mirror_from_klass----------------------------
3097 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3098 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3099 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3100 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3101 }
3102
3103 //-----------------------load_klass_from_mirror_common-------------------------
3104 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3105 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3106 // and branch to the given path on the region.
3107 // If never_see_null, take an uncommon trap on null, so we can optimistically
3108 // compile for the non-null case.
3109 // If the region is NULL, force never_see_null = true.
3110 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3111 bool never_see_null,
3112 RegionNode* region,
3113 int null_path,
3114 int offset) {
3115 if (region == NULL) never_see_null = true;
3116 Node* p = basic_plus_adr(mirror, offset);
3117 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3118 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3119 Node* null_ctl = top();
3120 kls = null_check_oop(kls, &null_ctl, never_see_null);
3121 if (region != NULL) {
3122 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3123 region->init_req(null_path, null_ctl);
3124 } else {
3125 assert(null_ctl == top(), "no loose ends");
3126 }
3127 return kls;
3128 }
3129
3130 //--------------------(inline_native_Class_query helpers)---------------------
3131 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3132 // Fall through if (mods & mask) == bits, take the guard otherwise.
3133 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3134 // Branch around if the given klass has the given modifier bit set.
3135 // Like generate_guard, adds a new path onto the region.
3136 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3137 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3138 Node* mask = intcon(modifier_mask);
3139 Node* bits = intcon(modifier_bits);
3140 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3141 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3142 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3143 return generate_fair_guard(bol, region);
3144 }
3145 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3146 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3147 }
3148
3149 //-------------------------inline_native_Class_query-------------------
3150 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3151 const Type* return_type = TypeInt::BOOL;
3152 Node* prim_return_value = top(); // what happens if it's a primitive class?
3153 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3154 bool expect_prim = false; // most of these guys expect to work on refs
3155
3156 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3157
3158 Node* mirror = argument(0);
3159 Node* obj = top();
3160
3161 switch (id) {
3162 case vmIntrinsics::_isInstance:
3163 // nothing is an instance of a primitive type
3164 prim_return_value = intcon(0);
3165 obj = argument(1);
3166 break;
3167 case vmIntrinsics::_getModifiers:
3168 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3169 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3170 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3171 break;
3172 case vmIntrinsics::_isInterface:
3173 prim_return_value = intcon(0);
3174 break;
3175 case vmIntrinsics::_isArray:
3176 prim_return_value = intcon(0);
3177 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3178 break;
3179 case vmIntrinsics::_isPrimitive:
3180 prim_return_value = intcon(1);
3181 expect_prim = true; // obviously
3182 break;
3183 case vmIntrinsics::_getSuperclass:
3184 prim_return_value = null();
3185 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3186 break;
3187 case vmIntrinsics::_getClassAccessFlags:
3188 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3189 return_type = TypeInt::INT; // not bool! 6297094
3190 break;
3191 default:
3192 fatal_unexpected_iid(id);
3193 break;
3194 }
3195
3196 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3197 if (mirror_con == NULL) return false; // cannot happen?
3198
3199 #ifndef PRODUCT
3200 if (C->print_intrinsics() || C->print_inlining()) {
3201 ciType* k = mirror_con->java_mirror_type();
3202 if (k) {
3203 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3204 k->print_name();
3205 tty->cr();
3206 }
3207 }
3208 #endif
3209
3210 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3211 RegionNode* region = new RegionNode(PATH_LIMIT);
3212 record_for_igvn(region);
3213 PhiNode* phi = new PhiNode(region, return_type);
3214
3215 // The mirror will never be null of Reflection.getClassAccessFlags, however
3216 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3217 // if it is. See bug 4774291.
3218
3219 // For Reflection.getClassAccessFlags(), the null check occurs in
3220 // the wrong place; see inline_unsafe_access(), above, for a similar
3221 // situation.
3222 mirror = null_check(mirror);
3223 // If mirror or obj is dead, only null-path is taken.
3224 if (stopped()) return true;
3225
3226 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3227
3228 // Now load the mirror's klass metaobject, and null-check it.
3229 // Side-effects region with the control path if the klass is null.
3230 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3231 // If kls is null, we have a primitive mirror.
3232 phi->init_req(_prim_path, prim_return_value);
3233 if (stopped()) { set_result(region, phi); return true; }
3234 bool safe_for_replace = (region->in(_prim_path) == top());
3235
3236 Node* p; // handy temp
3237 Node* null_ctl;
3238
3239 // Now that we have the non-null klass, we can perform the real query.
3240 // For constant classes, the query will constant-fold in LoadNode::Value.
3241 Node* query_value = top();
3242 switch (id) {
3243 case vmIntrinsics::_isInstance:
3244 // nothing is an instance of a primitive type
3245 query_value = gen_instanceof(obj, kls, safe_for_replace);
3246 break;
3247
3248 case vmIntrinsics::_getModifiers:
3249 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3250 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3251 break;
3252
3253 case vmIntrinsics::_isInterface:
3254 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3255 if (generate_interface_guard(kls, region) != NULL)
3256 // A guard was added. If the guard is taken, it was an interface.
3257 phi->add_req(intcon(1));
3258 // If we fall through, it's a plain class.
3259 query_value = intcon(0);
3260 break;
3261
3262 case vmIntrinsics::_isArray:
3263 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3264 if (generate_array_guard(kls, region) != NULL)
3265 // A guard was added. If the guard is taken, it was an array.
3266 phi->add_req(intcon(1));
3267 // If we fall through, it's a plain class.
3268 query_value = intcon(0);
3269 break;
3270
3271 case vmIntrinsics::_isPrimitive:
3272 query_value = intcon(0); // "normal" path produces false
3273 break;
3274
3275 case vmIntrinsics::_getSuperclass:
3276 // The rules here are somewhat unfortunate, but we can still do better
3277 // with random logic than with a JNI call.
3278 // Interfaces store null or Object as _super, but must report null.
3279 // Arrays store an intermediate super as _super, but must report Object.
3280 // Other types can report the actual _super.
3281 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3282 if (generate_interface_guard(kls, region) != NULL)
3283 // A guard was added. If the guard is taken, it was an interface.
3284 phi->add_req(null());
3285 if (generate_array_guard(kls, region) != NULL)
3286 // A guard was added. If the guard is taken, it was an array.
3287 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3288 // If we fall through, it's a plain class. Get its _super.
3289 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3290 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3291 null_ctl = top();
3292 kls = null_check_oop(kls, &null_ctl);
3293 if (null_ctl != top()) {
3294 // If the guard is taken, Object.superClass is null (both klass and mirror).
3295 region->add_req(null_ctl);
3296 phi ->add_req(null());
3297 }
3298 if (!stopped()) {
3299 query_value = load_mirror_from_klass(kls);
3300 }
3301 break;
3302
3303 case vmIntrinsics::_getClassAccessFlags:
3304 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3305 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3306 break;
3307
3308 default:
3309 fatal_unexpected_iid(id);
3310 break;
3311 }
3312
3313 // Fall-through is the normal case of a query to a real class.
3314 phi->init_req(1, query_value);
3315 region->init_req(1, control());
3316
3317 C->set_has_split_ifs(true); // Has chance for split-if optimization
3318 set_result(region, phi);
3319 return true;
3320 }
3321
3322 //-------------------------inline_Class_cast-------------------
3323 bool LibraryCallKit::inline_Class_cast() {
3324 Node* mirror = argument(0); // Class
3325 Node* obj = argument(1);
3326 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3327 if (mirror_con == NULL) {
3328 return false; // dead path (mirror->is_top()).
3329 }
3330 if (obj == NULL || obj->is_top()) {
3331 return false; // dead path
3332 }
3333 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3334
3335 // First, see if Class.cast() can be folded statically.
3336 // java_mirror_type() returns non-null for compile-time Class constants.
3337 ciType* tm = mirror_con->java_mirror_type();
3338 if (tm != NULL && tm->is_klass() &&
3339 tp != NULL && tp->klass() != NULL) {
3340 if (!tp->klass()->is_loaded()) {
3341 // Don't use intrinsic when class is not loaded.
3342 return false;
3343 } else {
3344 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3345 if (static_res == Compile::SSC_always_true) {
3346 // isInstance() is true - fold the code.
3347 set_result(obj);
3348 return true;
3349 } else if (static_res == Compile::SSC_always_false) {
3350 // Don't use intrinsic, have to throw ClassCastException.
3351 // If the reference is null, the non-intrinsic bytecode will
3352 // be optimized appropriately.
3353 return false;
3354 }
3355 }
3356 }
3357
3358 // Bailout intrinsic and do normal inlining if exception path is frequent.
3359 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3360 return false;
3361 }
3362
3363 // Generate dynamic checks.
3364 // Class.cast() is java implementation of _checkcast bytecode.
3365 // Do checkcast (Parse::do_checkcast()) optimizations here.
3366
3367 mirror = null_check(mirror);
3368 // If mirror is dead, only null-path is taken.
3369 if (stopped()) {
3370 return true;
3371 }
3372
3373 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3374 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3375 RegionNode* region = new RegionNode(PATH_LIMIT);
3376 record_for_igvn(region);
3377
3378 // Now load the mirror's klass metaobject, and null-check it.
3379 // If kls is null, we have a primitive mirror and
3380 // nothing is an instance of a primitive type.
3381 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3382
3383 Node* res = top();
3384 if (!stopped()) {
3385 Node* bad_type_ctrl = top();
3386 // Do checkcast optimizations.
3387 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3388 region->init_req(_bad_type_path, bad_type_ctrl);
3389 }
3390 if (region->in(_prim_path) != top() ||
3391 region->in(_bad_type_path) != top()) {
3392 // Let Interpreter throw ClassCastException.
3393 PreserveJVMState pjvms(this);
3394 set_control(_gvn.transform(region));
3395 uncommon_trap(Deoptimization::Reason_intrinsic,
3396 Deoptimization::Action_maybe_recompile);
3397 }
3398 if (!stopped()) {
3399 set_result(res);
3400 }
3401 return true;
3402 }
3403
3404
3405 //--------------------------inline_native_subtype_check------------------------
3406 // This intrinsic takes the JNI calls out of the heart of
3407 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3408 bool LibraryCallKit::inline_native_subtype_check() {
3409 // Pull both arguments off the stack.
3410 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3411 args[0] = argument(0);
3412 args[1] = argument(1);
3413 Node* klasses[2]; // corresponding Klasses: superk, subk
3414 klasses[0] = klasses[1] = top();
3415
3416 enum {
3417 // A full decision tree on {superc is prim, subc is prim}:
3418 _prim_0_path = 1, // {P,N} => false
3419 // {P,P} & superc!=subc => false
3420 _prim_same_path, // {P,P} & superc==subc => true
3421 _prim_1_path, // {N,P} => false
3422 _ref_subtype_path, // {N,N} & subtype check wins => true
3423 _both_ref_path, // {N,N} & subtype check loses => false
3424 PATH_LIMIT
3425 };
3426
3427 RegionNode* region = new RegionNode(PATH_LIMIT);
3428 Node* phi = new PhiNode(region, TypeInt::BOOL);
3429 record_for_igvn(region);
3430
3431 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3432 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3433 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3434
3435 // First null-check both mirrors and load each mirror's klass metaobject.
3436 int which_arg;
3437 for (which_arg = 0; which_arg <= 1; which_arg++) {
3438 Node* arg = args[which_arg];
3439 arg = null_check(arg);
3440 if (stopped()) break;
3441 args[which_arg] = arg;
3442
3443 Node* p = basic_plus_adr(arg, class_klass_offset);
3444 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3445 klasses[which_arg] = _gvn.transform(kls);
3446 }
3447
3448 // Having loaded both klasses, test each for null.
3449 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3450 for (which_arg = 0; which_arg <= 1; which_arg++) {
3451 Node* kls = klasses[which_arg];
3452 Node* null_ctl = top();
3453 kls = null_check_oop(kls, &null_ctl, never_see_null);
3454 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3455 region->init_req(prim_path, null_ctl);
3456 if (stopped()) break;
3457 klasses[which_arg] = kls;
3458 }
3459
3460 if (!stopped()) {
3461 // now we have two reference types, in klasses[0..1]
3462 Node* subk = klasses[1]; // the argument to isAssignableFrom
3463 Node* superk = klasses[0]; // the receiver
3464 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3465 // now we have a successful reference subtype check
3466 region->set_req(_ref_subtype_path, control());
3467 }
3468
3469 // If both operands are primitive (both klasses null), then
3470 // we must return true when they are identical primitives.
3471 // It is convenient to test this after the first null klass check.
3472 set_control(region->in(_prim_0_path)); // go back to first null check
3473 if (!stopped()) {
3474 // Since superc is primitive, make a guard for the superc==subc case.
3475 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3476 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3477 generate_guard(bol_eq, region, PROB_FAIR);
3478 if (region->req() == PATH_LIMIT+1) {
3479 // A guard was added. If the added guard is taken, superc==subc.
3480 region->swap_edges(PATH_LIMIT, _prim_same_path);
3481 region->del_req(PATH_LIMIT);
3482 }
3483 region->set_req(_prim_0_path, control()); // Not equal after all.
3484 }
3485
3486 // these are the only paths that produce 'true':
3487 phi->set_req(_prim_same_path, intcon(1));
3488 phi->set_req(_ref_subtype_path, intcon(1));
3489
3490 // pull together the cases:
3491 assert(region->req() == PATH_LIMIT, "sane region");
3492 for (uint i = 1; i < region->req(); i++) {
3493 Node* ctl = region->in(i);
3494 if (ctl == NULL || ctl == top()) {
3495 region->set_req(i, top());
3496 phi ->set_req(i, top());
3497 } else if (phi->in(i) == NULL) {
3498 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3499 }
3500 }
3501
3502 set_control(_gvn.transform(region));
3503 set_result(_gvn.transform(phi));
3504 return true;
3505 }
3506
3507 //---------------------generate_array_guard_common------------------------
3508 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3509 bool obj_array, bool not_array) {
3510
3511 if (stopped()) {
3512 return NULL;
3513 }
3514
3515 // If obj_array/non_array==false/false:
3516 // Branch around if the given klass is in fact an array (either obj or prim).
3517 // If obj_array/non_array==false/true:
3518 // Branch around if the given klass is not an array klass of any kind.
3519 // If obj_array/non_array==true/true:
3520 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3521 // If obj_array/non_array==true/false:
3522 // Branch around if the kls is an oop array (Object[] or subtype)
3523 //
3524 // Like generate_guard, adds a new path onto the region.
3525 jint layout_con = 0;
3526 Node* layout_val = get_layout_helper(kls, layout_con);
3527 if (layout_val == NULL) {
3528 bool query = (obj_array
3529 ? Klass::layout_helper_is_objArray(layout_con)
3530 : Klass::layout_helper_is_array(layout_con));
3531 if (query == not_array) {
3532 return NULL; // never a branch
3533 } else { // always a branch
3534 Node* always_branch = control();
3535 if (region != NULL)
3536 region->add_req(always_branch);
3537 set_control(top());
3538 return always_branch;
3539 }
3540 }
3541 // Now test the correct condition.
3542 jint nval = (obj_array
3543 ? ((jint)Klass::_lh_array_tag_type_value
3544 << Klass::_lh_array_tag_shift)
3545 : Klass::_lh_neutral_value);
3546 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3547 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3548 // invert the test if we are looking for a non-array
3549 if (not_array) btest = BoolTest(btest).negate();
3550 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3551 return generate_fair_guard(bol, region);
3552 }
3553
3554
3555 //-----------------------inline_native_newArray--------------------------
3556 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3557 bool LibraryCallKit::inline_native_newArray() {
3558 Node* mirror = argument(0);
3559 Node* count_val = argument(1);
3560
3561 mirror = null_check(mirror);
3562 // If mirror or obj is dead, only null-path is taken.
3563 if (stopped()) return true;
3564
3565 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3566 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3567 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3568 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3569 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3570
3571 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3572 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3573 result_reg, _slow_path);
3574 Node* normal_ctl = control();
3575 Node* no_array_ctl = result_reg->in(_slow_path);
3576
3577 // Generate code for the slow case. We make a call to newArray().
3578 set_control(no_array_ctl);
3579 if (!stopped()) {
3580 // Either the input type is void.class, or else the
3581 // array klass has not yet been cached. Either the
3582 // ensuing call will throw an exception, or else it
3583 // will cache the array klass for next time.
3584 PreserveJVMState pjvms(this);
3585 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3586 Node* slow_result = set_results_for_java_call(slow_call);
3587 // this->control() comes from set_results_for_java_call
3588 result_reg->set_req(_slow_path, control());
3589 result_val->set_req(_slow_path, slow_result);
3590 result_io ->set_req(_slow_path, i_o());
3591 result_mem->set_req(_slow_path, reset_memory());
3592 }
3593
3594 set_control(normal_ctl);
3595 if (!stopped()) {
3596 // Normal case: The array type has been cached in the java.lang.Class.
3597 // The following call works fine even if the array type is polymorphic.
3598 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3599 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3600 result_reg->init_req(_normal_path, control());
3601 result_val->init_req(_normal_path, obj);
3602 result_io ->init_req(_normal_path, i_o());
3603 result_mem->init_req(_normal_path, reset_memory());
3604 }
3605
3606 // Return the combined state.
3607 set_i_o( _gvn.transform(result_io) );
3608 set_all_memory( _gvn.transform(result_mem));
3609
3610 C->set_has_split_ifs(true); // Has chance for split-if optimization
3611 set_result(result_reg, result_val);
3612 return true;
3613 }
3614
3615 //----------------------inline_native_getLength--------------------------
3616 // public static native int java.lang.reflect.Array.getLength(Object array);
3617 bool LibraryCallKit::inline_native_getLength() {
3618 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3619
3620 Node* array = null_check(argument(0));
3621 // If array is dead, only null-path is taken.
3622 if (stopped()) return true;
3623
3624 // Deoptimize if it is a non-array.
3625 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3626
3627 if (non_array != NULL) {
3628 PreserveJVMState pjvms(this);
3629 set_control(non_array);
3630 uncommon_trap(Deoptimization::Reason_intrinsic,
3631 Deoptimization::Action_maybe_recompile);
3632 }
3633
3634 // If control is dead, only non-array-path is taken.
3635 if (stopped()) return true;
3636
3637 // The works fine even if the array type is polymorphic.
3638 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3639 Node* result = load_array_length(array);
3640
3641 C->set_has_split_ifs(true); // Has chance for split-if optimization
3642 set_result(result);
3643 return true;
3644 }
3645
3646 //------------------------inline_array_copyOf----------------------------
3647 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3648 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3649 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3650 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3651
3652 // Get the arguments.
3653 Node* original = argument(0);
3654 Node* start = is_copyOfRange? argument(1): intcon(0);
3655 Node* end = is_copyOfRange? argument(2): argument(1);
3656 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3657
3658 Node* newcopy;
3659
3660 // Set the original stack and the reexecute bit for the interpreter to reexecute
3661 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3662 { PreserveReexecuteState preexecs(this);
3663 jvms()->set_should_reexecute(true);
3664
3665 array_type_mirror = null_check(array_type_mirror);
3666 original = null_check(original);
3667
3668 // Check if a null path was taken unconditionally.
3669 if (stopped()) return true;
3670
3671 Node* orig_length = load_array_length(original);
3672
3673 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3674 klass_node = null_check(klass_node);
3675
3676 RegionNode* bailout = new RegionNode(1);
3677 record_for_igvn(bailout);
3678
3679 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3680 // Bail out if that is so.
3681 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3682 if (not_objArray != NULL) {
3683 // Improve the klass node's type from the new optimistic assumption:
3684 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3685 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3686 Node* cast = new CastPPNode(klass_node, akls);
3687 cast->init_req(0, control());
3688 klass_node = _gvn.transform(cast);
3689 }
3690
3691 // Bail out if either start or end is negative.
3692 generate_negative_guard(start, bailout, &start);
3693 generate_negative_guard(end, bailout, &end);
3694
3695 Node* length = end;
3696 if (_gvn.type(start) != TypeInt::ZERO) {
3697 length = _gvn.transform(new SubINode(end, start));
3698 }
3699
3700 // Bail out if length is negative.
3701 // Without this the new_array would throw
3702 // NegativeArraySizeException but IllegalArgumentException is what
3703 // should be thrown
3704 generate_negative_guard(length, bailout, &length);
3705
3706 if (bailout->req() > 1) {
3707 PreserveJVMState pjvms(this);
3708 set_control(_gvn.transform(bailout));
3709 uncommon_trap(Deoptimization::Reason_intrinsic,
3710 Deoptimization::Action_maybe_recompile);
3711 }
3712
3713 if (!stopped()) {
3714 // How many elements will we copy from the original?
3715 // The answer is MinI(orig_length - start, length).
3716 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3717 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3718
3719 // Generate a direct call to the right arraycopy function(s).
3720 // We know the copy is disjoint but we might not know if the
3721 // oop stores need checking.
3722 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3723 // This will fail a store-check if x contains any non-nulls.
3724
3725 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3726 // loads/stores but it is legal only if we're sure the
3727 // Arrays.copyOf would succeed. So we need all input arguments
3728 // to the copyOf to be validated, including that the copy to the
3729 // new array won't trigger an ArrayStoreException. That subtype
3730 // check can be optimized if we know something on the type of
3731 // the input array from type speculation.
3732 if (_gvn.type(klass_node)->singleton()) {
3733 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3734 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3735
3736 int test = C->static_subtype_check(superk, subk);
3737 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3738 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3739 if (t_original->speculative_type() != NULL) {
3740 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3741 }
3742 }
3743 }
3744
3745 bool validated = false;
3746 // Reason_class_check rather than Reason_intrinsic because we
3747 // want to intrinsify even if this traps.
3748 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3749 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3750 klass_node);
3751
3752 if (not_subtype_ctrl != top()) {
3753 PreserveJVMState pjvms(this);
3754 set_control(not_subtype_ctrl);
3755 uncommon_trap(Deoptimization::Reason_class_check,
3756 Deoptimization::Action_make_not_entrant);
3757 assert(stopped(), "Should be stopped");
3758 }
3759 validated = true;
3760 }
3761
3762 if (!stopped()) {
3763 newcopy = new_array(klass_node, length, 0); // no arguments to push
3764
3765 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3766 load_object_klass(original), klass_node);
3767 if (!is_copyOfRange) {
3768 ac->set_copyof(validated);
3769 } else {
3770 ac->set_copyofrange(validated);
3771 }
3772 Node* n = _gvn.transform(ac);
3773 if (n == ac) {
3774 ac->connect_outputs(this);
3775 } else {
3776 assert(validated, "shouldn't transform if all arguments not validated");
3777 set_all_memory(n);
3778 }
3779 }
3780 }
3781 } // original reexecute is set back here
3782
3783 C->set_has_split_ifs(true); // Has chance for split-if optimization
3784 if (!stopped()) {
3785 set_result(newcopy);
3786 }
3787 return true;
3788 }
3789
3790
3791 //----------------------generate_virtual_guard---------------------------
3792 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3793 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3794 RegionNode* slow_region) {
3795 ciMethod* method = callee();
3796 int vtable_index = method->vtable_index();
3797 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3798 err_msg_res("bad index %d", vtable_index));
3799 // Get the Method* out of the appropriate vtable entry.
3800 int entry_offset = (InstanceKlass::vtable_start_offset() +
3801 vtable_index*vtableEntry::size()) * wordSize +
3802 vtableEntry::method_offset_in_bytes();
3803 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3804 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3805
3806 // Compare the target method with the expected method (e.g., Object.hashCode).
3807 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3808
3809 Node* native_call = makecon(native_call_addr);
3810 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3811 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3812
3813 return generate_slow_guard(test_native, slow_region);
3814 }
3815
3816 //-----------------------generate_method_call----------------------------
3817 // Use generate_method_call to make a slow-call to the real
3818 // method if the fast path fails. An alternative would be to
3819 // use a stub like OptoRuntime::slow_arraycopy_Java.
3820 // This only works for expanding the current library call,
3821 // not another intrinsic. (E.g., don't use this for making an
3822 // arraycopy call inside of the copyOf intrinsic.)
3823 CallJavaNode*
3824 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3825 // When compiling the intrinsic method itself, do not use this technique.
3826 guarantee(callee() != C->method(), "cannot make slow-call to self");
3827
3828 ciMethod* method = callee();
3829 // ensure the JVMS we have will be correct for this call
3830 guarantee(method_id == method->intrinsic_id(), "must match");
3831
3832 const TypeFunc* tf = TypeFunc::make(method);
3833 CallJavaNode* slow_call;
3834 if (is_static) {
3835 assert(!is_virtual, "");
3836 slow_call = new CallStaticJavaNode(C, tf,
3837 SharedRuntime::get_resolve_static_call_stub(),
3838 method, bci());
3839 } else if (is_virtual) {
3840 null_check_receiver();
3841 int vtable_index = Method::invalid_vtable_index;
3842 if (UseInlineCaches) {
3843 // Suppress the vtable call
3844 } else {
3845 // hashCode and clone are not a miranda methods,
3846 // so the vtable index is fixed.
3847 // No need to use the linkResolver to get it.
3848 vtable_index = method->vtable_index();
3849 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3850 err_msg_res("bad index %d", vtable_index));
3851 }
3852 slow_call = new CallDynamicJavaNode(tf,
3853 SharedRuntime::get_resolve_virtual_call_stub(),
3854 method, vtable_index, bci());
3855 } else { // neither virtual nor static: opt_virtual
3856 null_check_receiver();
3857 slow_call = new CallStaticJavaNode(C, tf,
3858 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3859 method, bci());
3860 slow_call->set_optimized_virtual(true);
3861 }
3862 set_arguments_for_java_call(slow_call);
3863 set_edges_for_java_call(slow_call);
3864 return slow_call;
3865 }
3866
3867
3868 /**
3869 * Build special case code for calls to hashCode on an object. This call may
3870 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3871 * slightly different code.
3872 */
3873 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3874 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3875 assert(!(is_virtual && is_static), "either virtual, special, or static");
3876
3877 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3878
3879 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3880 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3881 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3882 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3883 Node* obj = NULL;
3884 if (!is_static) {
3885 // Check for hashing null object
3886 obj = null_check_receiver();
3887 if (stopped()) return true; // unconditionally null
3888 result_reg->init_req(_null_path, top());
3889 result_val->init_req(_null_path, top());
3890 } else {
3891 // Do a null check, and return zero if null.
3892 // System.identityHashCode(null) == 0
3893 obj = argument(0);
3894 Node* null_ctl = top();
3895 obj = null_check_oop(obj, &null_ctl);
3896 result_reg->init_req(_null_path, null_ctl);
3897 result_val->init_req(_null_path, _gvn.intcon(0));
3898 }
3899
3900 // Unconditionally null? Then return right away.
3901 if (stopped()) {
3902 set_control( result_reg->in(_null_path));
3903 if (!stopped())
3904 set_result(result_val->in(_null_path));
3905 return true;
3906 }
3907
3908 // We only go to the fast case code if we pass a number of guards. The
3909 // paths which do not pass are accumulated in the slow_region.
3910 RegionNode* slow_region = new RegionNode(1);
3911 record_for_igvn(slow_region);
3912
3913 // If this is a virtual call, we generate a funny guard. We pull out
3914 // the vtable entry corresponding to hashCode() from the target object.
3915 // If the target method which we are calling happens to be the native
3916 // Object hashCode() method, we pass the guard. We do not need this
3917 // guard for non-virtual calls -- the caller is known to be the native
3918 // Object hashCode().
3919 if (is_virtual) {
3920 // After null check, get the object's klass.
3921 Node* obj_klass = load_object_klass(obj);
3922 generate_virtual_guard(obj_klass, slow_region);
3923 }
3924
3925 // Get the header out of the object, use LoadMarkNode when available
3926 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3927 // The control of the load must be NULL. Otherwise, the load can move before
3928 // the null check after castPP removal.
3929 Node* no_ctrl = NULL;
3930 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3931
3932 // Test the header to see if it is unlocked.
3933 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3934 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3935 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3936 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3937 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3938
3939 generate_slow_guard(test_unlocked, slow_region);
3940
3941 // Get the hash value and check to see that it has been properly assigned.
3942 // We depend on hash_mask being at most 32 bits and avoid the use of
3943 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3944 // vm: see markOop.hpp.
3945 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3946 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3947 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3948 // This hack lets the hash bits live anywhere in the mark object now, as long
3949 // as the shift drops the relevant bits into the low 32 bits. Note that
3950 // Java spec says that HashCode is an int so there's no point in capturing
3951 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3952 hshifted_header = ConvX2I(hshifted_header);
3953 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3954
3955 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3956 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3957 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3958
3959 generate_slow_guard(test_assigned, slow_region);
3960
3961 Node* init_mem = reset_memory();
3962 // fill in the rest of the null path:
3963 result_io ->init_req(_null_path, i_o());
3964 result_mem->init_req(_null_path, init_mem);
3965
3966 result_val->init_req(_fast_path, hash_val);
3967 result_reg->init_req(_fast_path, control());
3968 result_io ->init_req(_fast_path, i_o());
3969 result_mem->init_req(_fast_path, init_mem);
3970
3971 // Generate code for the slow case. We make a call to hashCode().
3972 set_control(_gvn.transform(slow_region));
3973 if (!stopped()) {
3974 // No need for PreserveJVMState, because we're using up the present state.
3975 set_all_memory(init_mem);
3976 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3977 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3978 Node* slow_result = set_results_for_java_call(slow_call);
3979 // this->control() comes from set_results_for_java_call
3980 result_reg->init_req(_slow_path, control());
3981 result_val->init_req(_slow_path, slow_result);
3982 result_io ->set_req(_slow_path, i_o());
3983 result_mem ->set_req(_slow_path, reset_memory());
3984 }
3985
3986 // Return the combined state.
3987 set_i_o( _gvn.transform(result_io) );
3988 set_all_memory( _gvn.transform(result_mem));
3989
3990 set_result(result_reg, result_val);
3991 return true;
3992 }
3993
3994 //---------------------------inline_native_getClass----------------------------
3995 // public final native Class<?> java.lang.Object.getClass();
3996 //
3997 // Build special case code for calls to getClass on an object.
3998 bool LibraryCallKit::inline_native_getClass() {
3999 Node* obj = null_check_receiver();
4000 if (stopped()) return true;
4001 set_result(load_mirror_from_klass(load_object_klass(obj)));
4002 return true;
4003 }
4004
4005 //-----------------inline_native_Reflection_getCallerClass---------------------
4006 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4007 //
4008 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4009 //
4010 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4011 // in that it must skip particular security frames and checks for
4012 // caller sensitive methods.
4013 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4014 #ifndef PRODUCT
4015 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4016 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4017 }
4018 #endif
4019
4020 if (!jvms()->has_method()) {
4021 #ifndef PRODUCT
4022 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4023 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4024 }
4025 #endif
4026 return false;
4027 }
4028
4029 // Walk back up the JVM state to find the caller at the required
4030 // depth.
4031 JVMState* caller_jvms = jvms();
4032
4033 // Cf. JVM_GetCallerClass
4034 // NOTE: Start the loop at depth 1 because the current JVM state does
4035 // not include the Reflection.getCallerClass() frame.
4036 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4037 ciMethod* m = caller_jvms->method();
4038 switch (n) {
4039 case 0:
4040 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4041 break;
4042 case 1:
4043 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4044 if (!m->caller_sensitive()) {
4045 #ifndef PRODUCT
4046 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4047 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4048 }
4049 #endif
4050 return false; // bail-out; let JVM_GetCallerClass do the work
4051 }
4052 break;
4053 default:
4054 if (!m->is_ignored_by_security_stack_walk()) {
4055 // We have reached the desired frame; return the holder class.
4056 // Acquire method holder as java.lang.Class and push as constant.
4057 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4058 ciInstance* caller_mirror = caller_klass->java_mirror();
4059 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4060
4061 #ifndef PRODUCT
4062 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4063 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());
4064 tty->print_cr(" JVM state at this point:");
4065 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4066 ciMethod* m = jvms()->of_depth(i)->method();
4067 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4068 }
4069 }
4070 #endif
4071 return true;
4072 }
4073 break;
4074 }
4075 }
4076
4077 #ifndef PRODUCT
4078 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4079 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4080 tty->print_cr(" JVM state at this point:");
4081 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4082 ciMethod* m = jvms()->of_depth(i)->method();
4083 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4084 }
4085 }
4086 #endif
4087
4088 return false; // bail-out; let JVM_GetCallerClass do the work
4089 }
4090
4091 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4092 Node* arg = argument(0);
4093 Node* result;
4094
4095 switch (id) {
4096 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4097 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4098 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4099 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4100
4101 case vmIntrinsics::_doubleToLongBits: {
4102 // two paths (plus control) merge in a wood
4103 RegionNode *r = new RegionNode(3);
4104 Node *phi = new PhiNode(r, TypeLong::LONG);
4105
4106 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4107 // Build the boolean node
4108 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4109
4110 // Branch either way.
4111 // NaN case is less traveled, which makes all the difference.
4112 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4113 Node *opt_isnan = _gvn.transform(ifisnan);
4114 assert( opt_isnan->is_If(), "Expect an IfNode");
4115 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4116 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4117
4118 set_control(iftrue);
4119
4120 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4121 Node *slow_result = longcon(nan_bits); // return NaN
4122 phi->init_req(1, _gvn.transform( slow_result ));
4123 r->init_req(1, iftrue);
4124
4125 // Else fall through
4126 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4127 set_control(iffalse);
4128
4129 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4130 r->init_req(2, iffalse);
4131
4132 // Post merge
4133 set_control(_gvn.transform(r));
4134 record_for_igvn(r);
4135
4136 C->set_has_split_ifs(true); // Has chance for split-if optimization
4137 result = phi;
4138 assert(result->bottom_type()->isa_long(), "must be");
4139 break;
4140 }
4141
4142 case vmIntrinsics::_floatToIntBits: {
4143 // two paths (plus control) merge in a wood
4144 RegionNode *r = new RegionNode(3);
4145 Node *phi = new PhiNode(r, TypeInt::INT);
4146
4147 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4148 // Build the boolean node
4149 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4150
4151 // Branch either way.
4152 // NaN case is less traveled, which makes all the difference.
4153 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4154 Node *opt_isnan = _gvn.transform(ifisnan);
4155 assert( opt_isnan->is_If(), "Expect an IfNode");
4156 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4157 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4158
4159 set_control(iftrue);
4160
4161 static const jint nan_bits = 0x7fc00000;
4162 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4163 phi->init_req(1, _gvn.transform( slow_result ));
4164 r->init_req(1, iftrue);
4165
4166 // Else fall through
4167 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4168 set_control(iffalse);
4169
4170 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4171 r->init_req(2, iffalse);
4172
4173 // Post merge
4174 set_control(_gvn.transform(r));
4175 record_for_igvn(r);
4176
4177 C->set_has_split_ifs(true); // Has chance for split-if optimization
4178 result = phi;
4179 assert(result->bottom_type()->isa_int(), "must be");
4180 break;
4181 }
4182
4183 default:
4184 fatal_unexpected_iid(id);
4185 break;
4186 }
4187 set_result(_gvn.transform(result));
4188 return true;
4189 }
4190
4191 #ifdef _LP64
4192 #define XTOP ,top() /*additional argument*/
4193 #else //_LP64
4194 #define XTOP /*no additional argument*/
4195 #endif //_LP64
4196
4197 //----------------------inline_unsafe_copyMemory-------------------------
4198 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4199 bool LibraryCallKit::inline_unsafe_copyMemory() {
4200 if (callee()->is_static()) return false; // caller must have the capability!
4201 null_check_receiver(); // null-check receiver
4202 if (stopped()) return true;
4203
4204 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4205
4206 Node* src_ptr = argument(1); // type: oop
4207 Node* src_off = ConvL2X(argument(2)); // type: long
4208 Node* dst_ptr = argument(4); // type: oop
4209 Node* dst_off = ConvL2X(argument(5)); // type: long
4210 Node* size = ConvL2X(argument(7)); // type: long
4211
4212 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4213 "fieldOffset must be byte-scaled");
4214
4215 Node* src = make_unsafe_address(src_ptr, src_off);
4216 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4217
4218 // Conservatively insert a memory barrier on all memory slices.
4219 // Do not let writes of the copy source or destination float below the copy.
4220 insert_mem_bar(Op_MemBarCPUOrder);
4221
4222 // Call it. Note that the length argument is not scaled.
4223 make_runtime_call(RC_LEAF|RC_NO_FP,
4224 OptoRuntime::fast_arraycopy_Type(),
4225 StubRoutines::unsafe_arraycopy(),
4226 "unsafe_arraycopy",
4227 TypeRawPtr::BOTTOM,
4228 src, dst, size XTOP);
4229
4230 // Do not let reads of the copy destination float above the copy.
4231 insert_mem_bar(Op_MemBarCPUOrder);
4232
4233 return true;
4234 }
4235
4236 //------------------------clone_coping-----------------------------------
4237 // Helper function for inline_native_clone.
4238 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4239 assert(obj_size != NULL, "");
4240 Node* raw_obj = alloc_obj->in(1);
4241 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4242
4243 AllocateNode* alloc = NULL;
4244 if (ReduceBulkZeroing) {
4245 // We will be completely responsible for initializing this object -
4246 // mark Initialize node as complete.
4247 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4248 // The object was just allocated - there should be no any stores!
4249 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4250 // Mark as complete_with_arraycopy so that on AllocateNode
4251 // expansion, we know this AllocateNode is initialized by an array
4252 // copy and a StoreStore barrier exists after the array copy.
4253 alloc->initialization()->set_complete_with_arraycopy();
4254 }
4255
4256 // Copy the fastest available way.
4257 // TODO: generate fields copies for small objects instead.
4258 Node* src = obj;
4259 Node* dest = alloc_obj;
4260 Node* size = _gvn.transform(obj_size);
4261
4262 // Exclude the header but include array length to copy by 8 bytes words.
4263 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4264 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4265 instanceOopDesc::base_offset_in_bytes();
4266 // base_off:
4267 // 8 - 32-bit VM
4268 // 12 - 64-bit VM, compressed klass
4269 // 16 - 64-bit VM, normal klass
4270 if (base_off % BytesPerLong != 0) {
4271 assert(UseCompressedClassPointers, "");
4272 if (is_array) {
4273 // Exclude length to copy by 8 bytes words.
4274 base_off += sizeof(int);
4275 } else {
4276 // Include klass to copy by 8 bytes words.
4277 base_off = instanceOopDesc::klass_offset_in_bytes();
4278 }
4279 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4280 }
4281 src = basic_plus_adr(src, base_off);
4282 dest = basic_plus_adr(dest, base_off);
4283
4284 // Compute the length also, if needed:
4285 Node* countx = size;
4286 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4287 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4288
4289 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4290
4291 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4292 ac->set_clonebasic();
4293 Node* n = _gvn.transform(ac);
4294 if (n == ac) {
4295 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4296 } else {
4297 set_all_memory(n);
4298 }
4299
4300 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4301 if (card_mark) {
4302 assert(!is_array, "");
4303 // Put in store barrier for any and all oops we are sticking
4304 // into this object. (We could avoid this if we could prove
4305 // that the object type contains no oop fields at all.)
4306 Node* no_particular_value = NULL;
4307 Node* no_particular_field = NULL;
4308 int raw_adr_idx = Compile::AliasIdxRaw;
4309 post_barrier(control(),
4310 memory(raw_adr_type),
4311 alloc_obj,
4312 no_particular_field,
4313 raw_adr_idx,
4314 no_particular_value,
4315 T_OBJECT,
4316 false);
4317 }
4318
4319 // Do not let reads from the cloned object float above the arraycopy.
4320 if (alloc != NULL) {
4321 // Do not let stores that initialize this object be reordered with
4322 // a subsequent store that would make this object accessible by
4323 // other threads.
4324 // Record what AllocateNode this StoreStore protects so that
4325 // escape analysis can go from the MemBarStoreStoreNode to the
4326 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4327 // based on the escape status of the AllocateNode.
4328 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4329 } else {
4330 insert_mem_bar(Op_MemBarCPUOrder);
4331 }
4332 }
4333
4334 //------------------------inline_native_clone----------------------------
4335 // protected native Object java.lang.Object.clone();
4336 //
4337 // Here are the simple edge cases:
4338 // null receiver => normal trap
4339 // virtual and clone was overridden => slow path to out-of-line clone
4340 // not cloneable or finalizer => slow path to out-of-line Object.clone
4341 //
4342 // The general case has two steps, allocation and copying.
4343 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4344 //
4345 // Copying also has two cases, oop arrays and everything else.
4346 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4347 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4348 //
4349 // These steps fold up nicely if and when the cloned object's klass
4350 // can be sharply typed as an object array, a type array, or an instance.
4351 //
4352 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4353 PhiNode* result_val;
4354
4355 // Set the reexecute bit for the interpreter to reexecute
4356 // the bytecode that invokes Object.clone if deoptimization happens.
4357 { PreserveReexecuteState preexecs(this);
4358 jvms()->set_should_reexecute(true);
4359
4360 Node* obj = null_check_receiver();
4361 if (stopped()) return true;
4362
4363 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4364
4365 // If we are going to clone an instance, we need its exact type to
4366 // know the number and types of fields to convert the clone to
4367 // loads/stores. Maybe a speculative type can help us.
4368 if (!obj_type->klass_is_exact() &&
4369 obj_type->speculative_type() != NULL &&
4370 obj_type->speculative_type()->is_instance_klass()) {
4371 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4372 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4373 !spec_ik->has_injected_fields()) {
4374 ciKlass* k = obj_type->klass();
4375 if (!k->is_instance_klass() ||
4376 k->as_instance_klass()->is_interface() ||
4377 k->as_instance_klass()->has_subklass()) {
4378 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4379 }
4380 }
4381 }
4382
4383 Node* obj_klass = load_object_klass(obj);
4384 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4385 const TypeOopPtr* toop = ((tklass != NULL)
4386 ? tklass->as_instance_type()
4387 : TypeInstPtr::NOTNULL);
4388
4389 // Conservatively insert a memory barrier on all memory slices.
4390 // Do not let writes into the original float below the clone.
4391 insert_mem_bar(Op_MemBarCPUOrder);
4392
4393 // paths into result_reg:
4394 enum {
4395 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4396 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4397 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4398 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4399 PATH_LIMIT
4400 };
4401 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4402 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4403 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4404 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4405 record_for_igvn(result_reg);
4406
4407 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4408 int raw_adr_idx = Compile::AliasIdxRaw;
4409
4410 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4411 if (array_ctl != NULL) {
4412 // It's an array.
4413 PreserveJVMState pjvms(this);
4414 set_control(array_ctl);
4415 Node* obj_length = load_array_length(obj);
4416 Node* obj_size = NULL;
4417 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4418
4419 if (!use_ReduceInitialCardMarks()) {
4420 // If it is an oop array, it requires very special treatment,
4421 // because card marking is required on each card of the array.
4422 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4423 if (is_obja != NULL) {
4424 PreserveJVMState pjvms2(this);
4425 set_control(is_obja);
4426 // Generate a direct call to the right arraycopy function(s).
4427 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4428 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4429 ac->set_cloneoop();
4430 Node* n = _gvn.transform(ac);
4431 assert(n == ac, "cannot disappear");
4432 ac->connect_outputs(this);
4433
4434 result_reg->init_req(_objArray_path, control());
4435 result_val->init_req(_objArray_path, alloc_obj);
4436 result_i_o ->set_req(_objArray_path, i_o());
4437 result_mem ->set_req(_objArray_path, reset_memory());
4438 }
4439 }
4440 // Otherwise, there are no card marks to worry about.
4441 // (We can dispense with card marks if we know the allocation
4442 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4443 // causes the non-eden paths to take compensating steps to
4444 // simulate a fresh allocation, so that no further
4445 // card marks are required in compiled code to initialize
4446 // the object.)
4447
4448 if (!stopped()) {
4449 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4450
4451 // Present the results of the copy.
4452 result_reg->init_req(_array_path, control());
4453 result_val->init_req(_array_path, alloc_obj);
4454 result_i_o ->set_req(_array_path, i_o());
4455 result_mem ->set_req(_array_path, reset_memory());
4456 }
4457 }
4458
4459 // We only go to the instance fast case code if we pass a number of guards.
4460 // The paths which do not pass are accumulated in the slow_region.
4461 RegionNode* slow_region = new RegionNode(1);
4462 record_for_igvn(slow_region);
4463 if (!stopped()) {
4464 // It's an instance (we did array above). Make the slow-path tests.
4465 // If this is a virtual call, we generate a funny guard. We grab
4466 // the vtable entry corresponding to clone() from the target object.
4467 // If the target method which we are calling happens to be the
4468 // Object clone() method, we pass the guard. We do not need this
4469 // guard for non-virtual calls; the caller is known to be the native
4470 // Object clone().
4471 if (is_virtual) {
4472 generate_virtual_guard(obj_klass, slow_region);
4473 }
4474
4475 // The object must be cloneable and must not have a finalizer.
4476 // Both of these conditions may be checked in a single test.
4477 // We could optimize the cloneable test further, but we don't care.
4478 generate_access_flags_guard(obj_klass,
4479 // Test both conditions:
4480 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4481 // Must be cloneable but not finalizer:
4482 JVM_ACC_IS_CLONEABLE,
4483 slow_region);
4484 }
4485
4486 if (!stopped()) {
4487 // It's an instance, and it passed the slow-path tests.
4488 PreserveJVMState pjvms(this);
4489 Node* obj_size = NULL;
4490 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4491 // is reexecuted if deoptimization occurs and there could be problems when merging
4492 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4493 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4494
4495 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4496
4497 // Present the results of the slow call.
4498 result_reg->init_req(_instance_path, control());
4499 result_val->init_req(_instance_path, alloc_obj);
4500 result_i_o ->set_req(_instance_path, i_o());
4501 result_mem ->set_req(_instance_path, reset_memory());
4502 }
4503
4504 // Generate code for the slow case. We make a call to clone().
4505 set_control(_gvn.transform(slow_region));
4506 if (!stopped()) {
4507 PreserveJVMState pjvms(this);
4508 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4509 Node* slow_result = set_results_for_java_call(slow_call);
4510 // this->control() comes from set_results_for_java_call
4511 result_reg->init_req(_slow_path, control());
4512 result_val->init_req(_slow_path, slow_result);
4513 result_i_o ->set_req(_slow_path, i_o());
4514 result_mem ->set_req(_slow_path, reset_memory());
4515 }
4516
4517 // Return the combined state.
4518 set_control( _gvn.transform(result_reg));
4519 set_i_o( _gvn.transform(result_i_o));
4520 set_all_memory( _gvn.transform(result_mem));
4521 } // original reexecute is set back here
4522
4523 set_result(_gvn.transform(result_val));
4524 return true;
4525 }
4526
4527 // If we have a tighly coupled allocation, the arraycopy may take care
4528 // of the array initialization. If one of the guards we insert between
4529 // the allocation and the arraycopy causes a deoptimization, an
4530 // unitialized array will escape the compiled method. To prevent that
4531 // we set the JVM state for uncommon traps between the allocation and
4532 // the arraycopy to the state before the allocation so, in case of
4533 // deoptimization, we'll reexecute the allocation and the
4534 // initialization.
4535 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4536 if (alloc != NULL) {
4537 ciMethod* trap_method = alloc->jvms()->method();
4538 int trap_bci = alloc->jvms()->bci();
4539
4540 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4541 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4542 // Make sure there's no store between the allocation and the
4543 // arraycopy otherwise visible side effects could be rexecuted
4544 // in case of deoptimization and cause incorrect execution.
4545 bool no_interfering_store = true;
4546 Node* mem = alloc->in(TypeFunc::Memory);
4547 if (mem->is_MergeMem()) {
4548 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4549 Node* n = mms.memory();
4550 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4551 assert(n->is_Store(), "what else?");
4552 no_interfering_store = false;
4553 break;
4554 }
4555 }
4556 } else {
4557 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4558 Node* n = mms.memory();
4559 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4560 assert(n->is_Store(), "what else?");
4561 no_interfering_store = false;
4562 break;
4563 }
4564 }
4565 }
4566
4567 if (no_interfering_store) {
4568 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4569 uint size = alloc->req();
4570 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4571 old_jvms->set_map(sfpt);
4572 for (uint i = 0; i < size; i++) {
4573 sfpt->init_req(i, alloc->in(i));
4574 }
4575 // re-push array length for deoptimization
4576 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4577 old_jvms->set_sp(old_jvms->sp()+1);
4578 old_jvms->set_monoff(old_jvms->monoff()+1);
4579 old_jvms->set_scloff(old_jvms->scloff()+1);
4580 old_jvms->set_endoff(old_jvms->endoff()+1);
4581 old_jvms->set_should_reexecute(true);
4582
4583 sfpt->set_i_o(map()->i_o());
4584 sfpt->set_memory(map()->memory());
4585 sfpt->set_control(map()->control());
4586
4587 JVMState* saved_jvms = jvms();
4588 saved_reexecute_sp = _reexecute_sp;
4589
4590 set_jvms(sfpt->jvms());
4591 _reexecute_sp = jvms()->sp();
4592
4593 return saved_jvms;
4594 }
4595 }
4596 }
4597 return NULL;
4598 }
4599
4600 // In case of a deoptimization, we restart execution at the
4601 // allocation, allocating a new array. We would leave an uninitialized
4602 // array in the heap that GCs wouldn't expect. Move the allocation
4603 // after the traps so we don't allocate the array if we
4604 // deoptimize. This is possible because tightly_coupled_allocation()
4605 // guarantees there's no observer of the allocated array at this point
4606 // and the control flow is simple enough.
4607 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4608 if (saved_jvms != NULL && !stopped()) {
4609 assert(alloc != NULL, "only with a tightly coupled allocation");
4610 // restore JVM state to the state at the arraycopy
4611 saved_jvms->map()->set_control(map()->control());
4612 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4613 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4614 // If we've improved the types of some nodes (null check) while
4615 // emitting the guards, propagate them to the current state
4616 map()->replaced_nodes().apply(saved_jvms->map());
4617 set_jvms(saved_jvms);
4618 _reexecute_sp = saved_reexecute_sp;
4619
4620 // Remove the allocation from above the guards
4621 CallProjections callprojs;
4622 alloc->extract_projections(&callprojs, true);
4623 InitializeNode* init = alloc->initialization();
4624 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4625 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4626 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4627 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4628
4629 // move the allocation here (after the guards)
4630 _gvn.hash_delete(alloc);
4631 alloc->set_req(TypeFunc::Control, control());
4632 alloc->set_req(TypeFunc::I_O, i_o());
4633 Node *mem = reset_memory();
4634 set_all_memory(mem);
4635 alloc->set_req(TypeFunc::Memory, mem);
4636 set_control(init->proj_out(TypeFunc::Control));
4637 set_i_o(callprojs.fallthrough_ioproj);
4638
4639 // Update memory as done in GraphKit::set_output_for_allocation()
4640 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4641 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4642 if (ary_type->isa_aryptr() && length_type != NULL) {
4643 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4644 }
4645 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4646 int elemidx = C->get_alias_index(telemref);
4647 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4648 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4649
4650 Node* allocx = _gvn.transform(alloc);
4651 assert(allocx == alloc, "where has the allocation gone?");
4652 assert(dest->is_CheckCastPP(), "not an allocation result?");
4653
4654 _gvn.hash_delete(dest);
4655 dest->set_req(0, control());
4656 Node* destx = _gvn.transform(dest);
4657 assert(destx == dest, "where has the allocation result gone?");
4658 }
4659 }
4660
4661
4662 //------------------------------inline_arraycopy-----------------------
4663 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4664 // Object dest, int destPos,
4665 // int length);
4666 bool LibraryCallKit::inline_arraycopy() {
4667 // Get the arguments.
4668 Node* src = argument(0); // type: oop
4669 Node* src_offset = argument(1); // type: int
4670 Node* dest = argument(2); // type: oop
4671 Node* dest_offset = argument(3); // type: int
4672 Node* length = argument(4); // type: int
4673
4674
4675 // Check for allocation before we add nodes that would confuse
4676 // tightly_coupled_allocation()
4677 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4678
4679 int saved_reexecute_sp = -1;
4680 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4681 // See arraycopy_restore_alloc_state() comment
4682 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4683 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4684 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4685 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4686
4687 // The following tests must be performed
4688 // (1) src and dest are arrays.
4689 // (2) src and dest arrays must have elements of the same BasicType
4690 // (3) src and dest must not be null.
4691 // (4) src_offset must not be negative.
4692 // (5) dest_offset must not be negative.
4693 // (6) length must not be negative.
4694 // (7) src_offset + length must not exceed length of src.
4695 // (8) dest_offset + length must not exceed length of dest.
4696 // (9) each element of an oop array must be assignable
4697
4698 // (3) src and dest must not be null.
4699 // always do this here because we need the JVM state for uncommon traps
4700 Node* null_ctl = top();
4701 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4702 assert(null_ctl->is_top(), "no null control here");
4703 dest = null_check(dest, T_ARRAY);
4704
4705 if (!can_emit_guards) {
4706 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4707 // guards but the arraycopy node could still take advantage of a
4708 // tightly allocated allocation. tightly_coupled_allocation() is
4709 // called again to make sure it takes the null check above into
4710 // account: the null check is mandatory and if it caused an
4711 // uncommon trap to be emitted then the allocation can't be
4712 // considered tightly coupled in this context.
4713 alloc = tightly_coupled_allocation(dest, NULL);
4714 }
4715
4716 bool validated = false;
4717
4718 const Type* src_type = _gvn.type(src);
4719 const Type* dest_type = _gvn.type(dest);
4720 const TypeAryPtr* top_src = src_type->isa_aryptr();
4721 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4722
4723 // Do we have the type of src?
4724 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4725 // Do we have the type of dest?
4726 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4727 // Is the type for src from speculation?
4728 bool src_spec = false;
4729 // Is the type for dest from speculation?
4730 bool dest_spec = false;
4731
4732 if ((!has_src || !has_dest) && can_emit_guards) {
4733 // We don't have sufficient type information, let's see if
4734 // speculative types can help. We need to have types for both src
4735 // and dest so that it pays off.
4736
4737 // Do we already have or could we have type information for src
4738 bool could_have_src = has_src;
4739 // Do we already have or could we have type information for dest
4740 bool could_have_dest = has_dest;
4741
4742 ciKlass* src_k = NULL;
4743 if (!has_src) {
4744 src_k = src_type->speculative_type_not_null();
4745 if (src_k != NULL && src_k->is_array_klass()) {
4746 could_have_src = true;
4747 }
4748 }
4749
4750 ciKlass* dest_k = NULL;
4751 if (!has_dest) {
4752 dest_k = dest_type->speculative_type_not_null();
4753 if (dest_k != NULL && dest_k->is_array_klass()) {
4754 could_have_dest = true;
4755 }
4756 }
4757
4758 if (could_have_src && could_have_dest) {
4759 // This is going to pay off so emit the required guards
4760 if (!has_src) {
4761 src = maybe_cast_profiled_obj(src, src_k, true);
4762 src_type = _gvn.type(src);
4763 top_src = src_type->isa_aryptr();
4764 has_src = (top_src != NULL && top_src->klass() != NULL);
4765 src_spec = true;
4766 }
4767 if (!has_dest) {
4768 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4769 dest_type = _gvn.type(dest);
4770 top_dest = dest_type->isa_aryptr();
4771 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4772 dest_spec = true;
4773 }
4774 }
4775 }
4776
4777 if (has_src && has_dest && can_emit_guards) {
4778 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4779 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4780 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4781 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4782
4783 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4784 // If both arrays are object arrays then having the exact types
4785 // for both will remove the need for a subtype check at runtime
4786 // before the call and may make it possible to pick a faster copy
4787 // routine (without a subtype check on every element)
4788 // Do we have the exact type of src?
4789 bool could_have_src = src_spec;
4790 // Do we have the exact type of dest?
4791 bool could_have_dest = dest_spec;
4792 ciKlass* src_k = top_src->klass();
4793 ciKlass* dest_k = top_dest->klass();
4794 if (!src_spec) {
4795 src_k = src_type->speculative_type_not_null();
4796 if (src_k != NULL && src_k->is_array_klass()) {
4797 could_have_src = true;
4798 }
4799 }
4800 if (!dest_spec) {
4801 dest_k = dest_type->speculative_type_not_null();
4802 if (dest_k != NULL && dest_k->is_array_klass()) {
4803 could_have_dest = true;
4804 }
4805 }
4806 if (could_have_src && could_have_dest) {
4807 // If we can have both exact types, emit the missing guards
4808 if (could_have_src && !src_spec) {
4809 src = maybe_cast_profiled_obj(src, src_k, true);
4810 }
4811 if (could_have_dest && !dest_spec) {
4812 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4813 }
4814 }
4815 }
4816 }
4817
4818 ciMethod* trap_method = method();
4819 int trap_bci = bci();
4820 if (saved_jvms != NULL) {
4821 trap_method = alloc->jvms()->method();
4822 trap_bci = alloc->jvms()->bci();
4823 }
4824
4825 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4826 can_emit_guards &&
4827 !src->is_top() && !dest->is_top()) {
4828 // validate arguments: enables transformation the ArrayCopyNode
4829 validated = true;
4830
4831 RegionNode* slow_region = new RegionNode(1);
4832 record_for_igvn(slow_region);
4833
4834 // (1) src and dest are arrays.
4835 generate_non_array_guard(load_object_klass(src), slow_region);
4836 generate_non_array_guard(load_object_klass(dest), slow_region);
4837
4838 // (2) src and dest arrays must have elements of the same BasicType
4839 // done at macro expansion or at Ideal transformation time
4840
4841 // (4) src_offset must not be negative.
4842 generate_negative_guard(src_offset, slow_region);
4843
4844 // (5) dest_offset must not be negative.
4845 generate_negative_guard(dest_offset, slow_region);
4846
4847 // (7) src_offset + length must not exceed length of src.
4848 generate_limit_guard(src_offset, length,
4849 load_array_length(src),
4850 slow_region);
4851
4852 // (8) dest_offset + length must not exceed length of dest.
4853 generate_limit_guard(dest_offset, length,
4854 load_array_length(dest),
4855 slow_region);
4856
4857 // (9) each element of an oop array must be assignable
4858 Node* src_klass = load_object_klass(src);
4859 Node* dest_klass = load_object_klass(dest);
4860 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4861
4862 if (not_subtype_ctrl != top()) {
4863 PreserveJVMState pjvms(this);
4864 set_control(not_subtype_ctrl);
4865 uncommon_trap(Deoptimization::Reason_intrinsic,
4866 Deoptimization::Action_make_not_entrant);
4867 assert(stopped(), "Should be stopped");
4868 }
4869 {
4870 PreserveJVMState pjvms(this);
4871 set_control(_gvn.transform(slow_region));
4872 uncommon_trap(Deoptimization::Reason_intrinsic,
4873 Deoptimization::Action_make_not_entrant);
4874 assert(stopped(), "Should be stopped");
4875 }
4876 }
4877
4878 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4879
4880 if (stopped()) {
4881 return true;
4882 }
4883
4884 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4885 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4886 // so the compiler has a chance to eliminate them: during macro expansion,
4887 // we have to set their control (CastPP nodes are eliminated).
4888 load_object_klass(src), load_object_klass(dest),
4889 load_array_length(src), load_array_length(dest));
4890
4891 ac->set_arraycopy(validated);
4892
4893 Node* n = _gvn.transform(ac);
4894 if (n == ac) {
4895 ac->connect_outputs(this);
4896 } else {
4897 assert(validated, "shouldn't transform if all arguments not validated");
4898 set_all_memory(n);
4899 }
4900
4901 return true;
4902 }
4903
4904
4905 // Helper function which determines if an arraycopy immediately follows
4906 // an allocation, with no intervening tests or other escapes for the object.
4907 AllocateArrayNode*
4908 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4909 RegionNode* slow_region) {
4910 if (stopped()) return NULL; // no fast path
4911 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4912
4913 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4914 if (alloc == NULL) return NULL;
4915
4916 Node* rawmem = memory(Compile::AliasIdxRaw);
4917 // Is the allocation's memory state untouched?
4918 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4919 // Bail out if there have been raw-memory effects since the allocation.
4920 // (Example: There might have been a call or safepoint.)
4921 return NULL;
4922 }
4923 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4924 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4925 return NULL;
4926 }
4927
4928 // There must be no unexpected observers of this allocation.
4929 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4930 Node* obs = ptr->fast_out(i);
4931 if (obs != this->map()) {
4932 return NULL;
4933 }
4934 }
4935
4936 // This arraycopy must unconditionally follow the allocation of the ptr.
4937 Node* alloc_ctl = ptr->in(0);
4938 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4939
4940 Node* ctl = control();
4941 while (ctl != alloc_ctl) {
4942 // There may be guards which feed into the slow_region.
4943 // Any other control flow means that we might not get a chance
4944 // to finish initializing the allocated object.
4945 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4946 IfNode* iff = ctl->in(0)->as_If();
4947 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4948 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4949 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4950 ctl = iff->in(0); // This test feeds the known slow_region.
4951 continue;
4952 }
4953 // One more try: Various low-level checks bottom out in
4954 // uncommon traps. If the debug-info of the trap omits
4955 // any reference to the allocation, as we've already
4956 // observed, then there can be no objection to the trap.
4957 bool found_trap = false;
4958 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4959 Node* obs = not_ctl->fast_out(j);
4960 if (obs->in(0) == not_ctl && obs->is_Call() &&
4961 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4962 found_trap = true; break;
4963 }
4964 }
4965 if (found_trap) {
4966 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4967 continue;
4968 }
4969 }
4970 return NULL;
4971 }
4972
4973 // If we get this far, we have an allocation which immediately
4974 // precedes the arraycopy, and we can take over zeroing the new object.
4975 // The arraycopy will finish the initialization, and provide
4976 // a new control state to which we will anchor the destination pointer.
4977
4978 return alloc;
4979 }
4980
4981 //-------------inline_encodeISOArray-----------------------------------
4982 // encode char[] to byte[] in ISO_8859_1
4983 bool LibraryCallKit::inline_encodeISOArray() {
4984 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4985 // no receiver since it is static method
4986 Node *src = argument(0);
4987 Node *src_offset = argument(1);
4988 Node *dst = argument(2);
4989 Node *dst_offset = argument(3);
4990 Node *length = argument(4);
4991
4992 const Type* src_type = src->Value(&_gvn);
4993 const Type* dst_type = dst->Value(&_gvn);
4994 const TypeAryPtr* top_src = src_type->isa_aryptr();
4995 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
4996 if (top_src == NULL || top_src->klass() == NULL ||
4997 top_dest == NULL || top_dest->klass() == NULL) {
4998 // failed array check
4999 return false;
5000 }
5001
5002 // Figure out the size and type of the elements we will be copying.
5003 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5004 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5005 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5006 return false;
5007 }
5008 Node* src_start = array_element_address(src, src_offset, src_elem);
5009 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5010 // 'src_start' points to src array + scaled offset
5011 // 'dst_start' points to dst array + scaled offset
5012
5013 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5014 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5015 enc = _gvn.transform(enc);
5016 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5017 set_memory(res_mem, mtype);
5018 set_result(enc);
5019 return true;
5020 }
5021
5022 //-------------inline_multiplyToLen-----------------------------------
5023 bool LibraryCallKit::inline_multiplyToLen() {
5024 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5025
5026 address stubAddr = StubRoutines::multiplyToLen();
5027 if (stubAddr == NULL) {
5028 return false; // Intrinsic's stub is not implemented on this platform
5029 }
5030 const char* stubName = "multiplyToLen";
5031
5032 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5033
5034 // no receiver because it is a static method
5035 Node* x = argument(0);
5036 Node* xlen = argument(1);
5037 Node* y = argument(2);
5038 Node* ylen = argument(3);
5039 Node* z = argument(4);
5040
5041 const Type* x_type = x->Value(&_gvn);
5042 const Type* y_type = y->Value(&_gvn);
5043 const TypeAryPtr* top_x = x_type->isa_aryptr();
5044 const TypeAryPtr* top_y = y_type->isa_aryptr();
5045 if (top_x == NULL || top_x->klass() == NULL ||
5046 top_y == NULL || top_y->klass() == NULL) {
5047 // failed array check
5048 return false;
5049 }
5050
5051 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5052 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5053 if (x_elem != T_INT || y_elem != T_INT) {
5054 return false;
5055 }
5056
5057 // Set the original stack and the reexecute bit for the interpreter to reexecute
5058 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5059 // on the return from z array allocation in runtime.
5060 { PreserveReexecuteState preexecs(this);
5061 jvms()->set_should_reexecute(true);
5062
5063 Node* x_start = array_element_address(x, intcon(0), x_elem);
5064 Node* y_start = array_element_address(y, intcon(0), y_elem);
5065 // 'x_start' points to x array + scaled xlen
5066 // 'y_start' points to y array + scaled ylen
5067
5068 // Allocate the result array
5069 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5070 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5071 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5072
5073 IdealKit ideal(this);
5074
5075 #define __ ideal.
5076 Node* one = __ ConI(1);
5077 Node* zero = __ ConI(0);
5078 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5079 __ set(need_alloc, zero);
5080 __ set(z_alloc, z);
5081 __ if_then(z, BoolTest::eq, null()); {
5082 __ increment (need_alloc, one);
5083 } __ else_(); {
5084 // Update graphKit memory and control from IdealKit.
5085 sync_kit(ideal);
5086 Node* zlen_arg = load_array_length(z);
5087 // Update IdealKit memory and control from graphKit.
5088 __ sync_kit(this);
5089 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5090 __ increment (need_alloc, one);
5091 } __ end_if();
5092 } __ end_if();
5093
5094 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5095 // Update graphKit memory and control from IdealKit.
5096 sync_kit(ideal);
5097 Node * narr = new_array(klass_node, zlen, 1);
5098 // Update IdealKit memory and control from graphKit.
5099 __ sync_kit(this);
5100 __ set(z_alloc, narr);
5101 } __ end_if();
5102
5103 sync_kit(ideal);
5104 z = __ value(z_alloc);
5105 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5106 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5107 // Final sync IdealKit and GraphKit.
5108 final_sync(ideal);
5109 #undef __
5110
5111 Node* z_start = array_element_address(z, intcon(0), T_INT);
5112
5113 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5114 OptoRuntime::multiplyToLen_Type(),
5115 stubAddr, stubName, TypePtr::BOTTOM,
5116 x_start, xlen, y_start, ylen, z_start, zlen);
5117 } // original reexecute is set back here
5118
5119 C->set_has_split_ifs(true); // Has chance for split-if optimization
5120 set_result(z);
5121 return true;
5122 }
5123
5124 //-------------inline_squareToLen------------------------------------
5125 bool LibraryCallKit::inline_squareToLen() {
5126 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5127
5128 address stubAddr = StubRoutines::squareToLen();
5129 if (stubAddr == NULL) {
5130 return false; // Intrinsic's stub is not implemented on this platform
5131 }
5132 const char* stubName = "squareToLen";
5133
5134 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5135
5136 Node* x = argument(0);
5137 Node* len = argument(1);
5138 Node* z = argument(2);
5139 Node* zlen = argument(3);
5140
5141 const Type* x_type = x->Value(&_gvn);
5142 const Type* z_type = z->Value(&_gvn);
5143 const TypeAryPtr* top_x = x_type->isa_aryptr();
5144 const TypeAryPtr* top_z = z_type->isa_aryptr();
5145 if (top_x == NULL || top_x->klass() == NULL ||
5146 top_z == NULL || top_z->klass() == NULL) {
5147 // failed array check
5148 return false;
5149 }
5150
5151 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5152 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5153 if (x_elem != T_INT || z_elem != T_INT) {
5154 return false;
5155 }
5156
5157
5158 Node* x_start = array_element_address(x, intcon(0), x_elem);
5159 Node* z_start = array_element_address(z, intcon(0), z_elem);
5160
5161 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5162 OptoRuntime::squareToLen_Type(),
5163 stubAddr, stubName, TypePtr::BOTTOM,
5164 x_start, len, z_start, zlen);
5165
5166 set_result(z);
5167 return true;
5168 }
5169
5170 //-------------inline_mulAdd------------------------------------------
5171 bool LibraryCallKit::inline_mulAdd() {
5172 assert(UseMulAddIntrinsic, "not implementated on this platform");
5173
5174 address stubAddr = StubRoutines::mulAdd();
5175 if (stubAddr == NULL) {
5176 return false; // Intrinsic's stub is not implemented on this platform
5177 }
5178 const char* stubName = "mulAdd";
5179
5180 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5181
5182 Node* out = argument(0);
5183 Node* in = argument(1);
5184 Node* offset = argument(2);
5185 Node* len = argument(3);
5186 Node* k = argument(4);
5187
5188 const Type* out_type = out->Value(&_gvn);
5189 const Type* in_type = in->Value(&_gvn);
5190 const TypeAryPtr* top_out = out_type->isa_aryptr();
5191 const TypeAryPtr* top_in = in_type->isa_aryptr();
5192 if (top_out == NULL || top_out->klass() == NULL ||
5193 top_in == NULL || top_in->klass() == NULL) {
5194 // failed array check
5195 return false;
5196 }
5197
5198 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5199 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5200 if (out_elem != T_INT || in_elem != T_INT) {
5201 return false;
5202 }
5203
5204 Node* outlen = load_array_length(out);
5205 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5206 Node* out_start = array_element_address(out, intcon(0), out_elem);
5207 Node* in_start = array_element_address(in, intcon(0), in_elem);
5208
5209 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5210 OptoRuntime::mulAdd_Type(),
5211 stubAddr, stubName, TypePtr::BOTTOM,
5212 out_start,in_start, new_offset, len, k);
5213 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5214 set_result(result);
5215 return true;
5216 }
5217
5218 //-------------inline_montgomeryMultiply-----------------------------------
5219 bool LibraryCallKit::inline_montgomeryMultiply() {
5220 address stubAddr = StubRoutines::montgomeryMultiply();
5221 if (stubAddr == NULL) {
5222 return false; // Intrinsic's stub is not implemented on this platform
5223 }
5224
5225 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5226 const char* stubName = "montgomery_square";
5227
5228 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5229
5230 Node* a = argument(0);
5231 Node* b = argument(1);
5232 Node* n = argument(2);
5233 Node* len = argument(3);
5234 Node* inv = argument(4);
5235 Node* m = argument(6);
5236
5237 const Type* a_type = a->Value(&_gvn);
5238 const TypeAryPtr* top_a = a_type->isa_aryptr();
5239 const Type* b_type = b->Value(&_gvn);
5240 const TypeAryPtr* top_b = b_type->isa_aryptr();
5241 const Type* n_type = a->Value(&_gvn);
5242 const TypeAryPtr* top_n = n_type->isa_aryptr();
5243 const Type* m_type = a->Value(&_gvn);
5244 const TypeAryPtr* top_m = m_type->isa_aryptr();
5245 if (top_a == NULL || top_a->klass() == NULL ||
5246 top_b == NULL || top_b->klass() == NULL ||
5247 top_n == NULL || top_n->klass() == NULL ||
5248 top_m == NULL || top_m->klass() == NULL) {
5249 // failed array check
5250 return false;
5251 }
5252
5253 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5254 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5255 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5256 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5257 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5258 return false;
5259 }
5260
5261 // Make the call
5262 {
5263 Node* a_start = array_element_address(a, intcon(0), a_elem);
5264 Node* b_start = array_element_address(b, intcon(0), b_elem);
5265 Node* n_start = array_element_address(n, intcon(0), n_elem);
5266 Node* m_start = array_element_address(m, intcon(0), m_elem);
5267
5268 Node* call = make_runtime_call(RC_LEAF,
5269 OptoRuntime::montgomeryMultiply_Type(),
5270 stubAddr, stubName, TypePtr::BOTTOM,
5271 a_start, b_start, n_start, len, inv, top(),
5272 m_start);
5273 set_result(m);
5274 }
5275
5276 return true;
5277 }
5278
5279 bool LibraryCallKit::inline_montgomerySquare() {
5280 address stubAddr = StubRoutines::montgomerySquare();
5281 if (stubAddr == NULL) {
5282 return false; // Intrinsic's stub is not implemented on this platform
5283 }
5284
5285 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5286 const char* stubName = "montgomery_square";
5287
5288 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5289
5290 Node* a = argument(0);
5291 Node* n = argument(1);
5292 Node* len = argument(2);
5293 Node* inv = argument(3);
5294 Node* m = argument(5);
5295
5296 const Type* a_type = a->Value(&_gvn);
5297 const TypeAryPtr* top_a = a_type->isa_aryptr();
5298 const Type* n_type = a->Value(&_gvn);
5299 const TypeAryPtr* top_n = n_type->isa_aryptr();
5300 const Type* m_type = a->Value(&_gvn);
5301 const TypeAryPtr* top_m = m_type->isa_aryptr();
5302 if (top_a == NULL || top_a->klass() == NULL ||
5303 top_n == NULL || top_n->klass() == NULL ||
5304 top_m == NULL || top_m->klass() == NULL) {
5305 // failed array check
5306 return false;
5307 }
5308
5309 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5310 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5311 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5312 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5313 return false;
5314 }
5315
5316 // Make the call
5317 {
5318 Node* a_start = array_element_address(a, intcon(0), a_elem);
5319 Node* n_start = array_element_address(n, intcon(0), n_elem);
5320 Node* m_start = array_element_address(m, intcon(0), m_elem);
5321
5322 Node* call = make_runtime_call(RC_LEAF,
5323 OptoRuntime::montgomerySquare_Type(),
5324 stubAddr, stubName, TypePtr::BOTTOM,
5325 a_start, n_start, len, inv, top(),
5326 m_start);
5327 set_result(m);
5328 }
5329
5330 return true;
5331 }
5332
5333
5334 /**
5335 * Calculate CRC32 for byte.
5336 * int java.util.zip.CRC32.update(int crc, int b)
5337 */
5338 bool LibraryCallKit::inline_updateCRC32() {
5339 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5340 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5341 // no receiver since it is static method
5342 Node* crc = argument(0); // type: int
5343 Node* b = argument(1); // type: int
5344
5345 /*
5346 * int c = ~ crc;
5347 * b = timesXtoThe32[(b ^ c) & 0xFF];
5348 * b = b ^ (c >>> 8);
5349 * crc = ~b;
5350 */
5351
5352 Node* M1 = intcon(-1);
5353 crc = _gvn.transform(new XorINode(crc, M1));
5354 Node* result = _gvn.transform(new XorINode(crc, b));
5355 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5356
5357 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5358 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5359 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5360 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5361
5362 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5363 result = _gvn.transform(new XorINode(crc, result));
5364 result = _gvn.transform(new XorINode(result, M1));
5365 set_result(result);
5366 return true;
5367 }
5368
5369 /**
5370 * Calculate CRC32 for byte[] array.
5371 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5372 */
5373 bool LibraryCallKit::inline_updateBytesCRC32() {
5374 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5375 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5376 // no receiver since it is static method
5377 Node* crc = argument(0); // type: int
5378 Node* src = argument(1); // type: oop
5379 Node* offset = argument(2); // type: int
5380 Node* length = argument(3); // type: int
5381
5382 const Type* src_type = src->Value(&_gvn);
5383 const TypeAryPtr* top_src = src_type->isa_aryptr();
5384 if (top_src == NULL || top_src->klass() == NULL) {
5385 // failed array check
5386 return false;
5387 }
5388
5389 // Figure out the size and type of the elements we will be copying.
5390 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5391 if (src_elem != T_BYTE) {
5392 return false;
5393 }
5394
5395 // 'src_start' points to src array + scaled offset
5396 Node* src_start = array_element_address(src, offset, src_elem);
5397
5398 // We assume that range check is done by caller.
5399 // TODO: generate range check (offset+length < src.length) in debug VM.
5400
5401 // Call the stub.
5402 address stubAddr = StubRoutines::updateBytesCRC32();
5403 const char *stubName = "updateBytesCRC32";
5404
5405 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5406 stubAddr, stubName, TypePtr::BOTTOM,
5407 crc, src_start, length);
5408 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5409 set_result(result);
5410 return true;
5411 }
5412
5413 /**
5414 * Calculate CRC32 for ByteBuffer.
5415 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5416 */
5417 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5418 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5419 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5420 // no receiver since it is static method
5421 Node* crc = argument(0); // type: int
5422 Node* src = argument(1); // type: long
5423 Node* offset = argument(3); // type: int
5424 Node* length = argument(4); // type: int
5425
5426 src = ConvL2X(src); // adjust Java long to machine word
5427 Node* base = _gvn.transform(new CastX2PNode(src));
5428 offset = ConvI2X(offset);
5429
5430 // 'src_start' points to src array + scaled offset
5431 Node* src_start = basic_plus_adr(top(), base, offset);
5432
5433 // Call the stub.
5434 address stubAddr = StubRoutines::updateBytesCRC32();
5435 const char *stubName = "updateBytesCRC32";
5436
5437 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5438 stubAddr, stubName, TypePtr::BOTTOM,
5439 crc, src_start, length);
5440 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5441 set_result(result);
5442 return true;
5443 }
5444
5445 //------------------------------get_table_from_crc32c_class-----------------------
5446 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5447 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5448 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5449
5450 return table;
5451 }
5452
5453 //------------------------------inline_updateBytesCRC32C-----------------------
5454 //
5455 // Calculate CRC32C for byte[] array.
5456 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5457 //
5458 bool LibraryCallKit::inline_updateBytesCRC32C() {
5459 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5460 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5461 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5462 // no receiver since it is a static method
5463 Node* crc = argument(0); // type: int
5464 Node* src = argument(1); // type: oop
5465 Node* offset = argument(2); // type: int
5466 Node* end = argument(3); // type: int
5467
5468 Node* length = _gvn.transform(new SubINode(end, offset));
5469
5470 const Type* src_type = src->Value(&_gvn);
5471 const TypeAryPtr* top_src = src_type->isa_aryptr();
5472 if (top_src == NULL || top_src->klass() == NULL) {
5473 // failed array check
5474 return false;
5475 }
5476
5477 // Figure out the size and type of the elements we will be copying.
5478 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5479 if (src_elem != T_BYTE) {
5480 return false;
5481 }
5482
5483 // 'src_start' points to src array + scaled offset
5484 Node* src_start = array_element_address(src, offset, src_elem);
5485
5486 // static final int[] byteTable in class CRC32C
5487 Node* table = get_table_from_crc32c_class(callee()->holder());
5488 Node* table_start = array_element_address(table, intcon(0), T_INT);
5489
5490 // We assume that range check is done by caller.
5491 // TODO: generate range check (offset+length < src.length) in debug VM.
5492
5493 // Call the stub.
5494 address stubAddr = StubRoutines::updateBytesCRC32C();
5495 const char *stubName = "updateBytesCRC32C";
5496
5497 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5498 stubAddr, stubName, TypePtr::BOTTOM,
5499 crc, src_start, length, table_start);
5500 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5501 set_result(result);
5502 return true;
5503 }
5504
5505 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5506 //
5507 // Calculate CRC32C for DirectByteBuffer.
5508 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5509 //
5510 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5511 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5512 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5513 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5514 // no receiver since it is a static method
5515 Node* crc = argument(0); // type: int
5516 Node* src = argument(1); // type: long
5517 Node* offset = argument(3); // type: int
5518 Node* end = argument(4); // type: int
5519
5520 Node* length = _gvn.transform(new SubINode(end, offset));
5521
5522 src = ConvL2X(src); // adjust Java long to machine word
5523 Node* base = _gvn.transform(new CastX2PNode(src));
5524 offset = ConvI2X(offset);
5525
5526 // 'src_start' points to src array + scaled offset
5527 Node* src_start = basic_plus_adr(top(), base, offset);
5528
5529 // static final int[] byteTable in class CRC32C
5530 Node* table = get_table_from_crc32c_class(callee()->holder());
5531 Node* table_start = array_element_address(table, intcon(0), T_INT);
5532
5533 // Call the stub.
5534 address stubAddr = StubRoutines::updateBytesCRC32C();
5535 const char *stubName = "updateBytesCRC32C";
5536
5537 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5538 stubAddr, stubName, TypePtr::BOTTOM,
5539 crc, src_start, length, table_start);
5540 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5541 set_result(result);
5542 return true;
5543 }
5544
5545 //------------------------------inline_updateBytesAdler32----------------------
5546 //
5547 // Calculate Adler32 checksum for byte[] array.
5548 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5549 //
5550 bool LibraryCallKit::inline_updateBytesAdler32() {
5551 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5552 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5553 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5554 // no receiver since it is static method
5555 Node* crc = argument(0); // type: int
5556 Node* src = argument(1); // type: oop
5557 Node* offset = argument(2); // type: int
5558 Node* length = argument(3); // type: int
5559
5560 const Type* src_type = src->Value(&_gvn);
5561 const TypeAryPtr* top_src = src_type->isa_aryptr();
5562 if (top_src == NULL || top_src->klass() == NULL) {
5563 // failed array check
5564 return false;
5565 }
5566
5567 // Figure out the size and type of the elements we will be copying.
5568 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5569 if (src_elem != T_BYTE) {
5570 return false;
5571 }
5572
5573 // 'src_start' points to src array + scaled offset
5574 Node* src_start = array_element_address(src, offset, src_elem);
5575
5576 // We assume that range check is done by caller.
5577 // TODO: generate range check (offset+length < src.length) in debug VM.
5578
5579 // Call the stub.
5580 address stubAddr = StubRoutines::updateBytesAdler32();
5581 const char *stubName = "updateBytesAdler32";
5582
5583 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5584 stubAddr, stubName, TypePtr::BOTTOM,
5585 crc, src_start, length);
5586 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5587 set_result(result);
5588 return true;
5589 }
5590
5591 //------------------------------inline_updateByteBufferAdler32---------------
5592 //
5593 // Calculate Adler32 checksum for DirectByteBuffer.
5594 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5595 //
5596 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5597 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5598 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5599 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5600 // no receiver since it is static method
5601 Node* crc = argument(0); // type: int
5602 Node* src = argument(1); // type: long
5603 Node* offset = argument(3); // type: int
5604 Node* length = argument(4); // type: int
5605
5606 src = ConvL2X(src); // adjust Java long to machine word
5607 Node* base = _gvn.transform(new CastX2PNode(src));
5608 offset = ConvI2X(offset);
5609
5610 // 'src_start' points to src array + scaled offset
5611 Node* src_start = basic_plus_adr(top(), base, offset);
5612
5613 // Call the stub.
5614 address stubAddr = StubRoutines::updateBytesAdler32();
5615 const char *stubName = "updateBytesAdler32";
5616
5617 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5618 stubAddr, stubName, TypePtr::BOTTOM,
5619 crc, src_start, length);
5620
5621 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5622 set_result(result);
5623 return true;
5624 }
5625
5626 //----------------------------inline_reference_get----------------------------
5627 // public T java.lang.ref.Reference.get();
5628 bool LibraryCallKit::inline_reference_get() {
5629 const int referent_offset = java_lang_ref_Reference::referent_offset;
5630 guarantee(referent_offset > 0, "should have already been set");
5631
5632 // Get the argument:
5633 Node* reference_obj = null_check_receiver();
5634 if (stopped()) return true;
5635
5636 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5637
5638 ciInstanceKlass* klass = env()->Object_klass();
5639 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5640
5641 Node* no_ctrl = NULL;
5642 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5643
5644 // Use the pre-barrier to record the value in the referent field
5645 pre_barrier(false /* do_load */,
5646 control(),
5647 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5648 result /* pre_val */,
5649 T_OBJECT);
5650
5651 // Add memory barrier to prevent commoning reads from this field
5652 // across safepoint since GC can change its value.
5653 insert_mem_bar(Op_MemBarCPUOrder);
5654
5655 set_result(result);
5656 return true;
5657 }
5658
5659
5660 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5661 bool is_exact=true, bool is_static=false,
5662 ciInstanceKlass * fromKls=NULL) {
5663 if (fromKls == NULL) {
5664 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5665 assert(tinst != NULL, "obj is null");
5666 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5667 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5668 fromKls = tinst->klass()->as_instance_klass();
5669 } else {
5670 assert(is_static, "only for static field access");
5671 }
5672 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5673 ciSymbol::make(fieldTypeString),
5674 is_static);
5675
5676 assert (field != NULL, "undefined field");
5677 if (field == NULL) return (Node *) NULL;
5678
5679 if (is_static) {
5680 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5681 fromObj = makecon(tip);
5682 }
5683
5684 // Next code copied from Parse::do_get_xxx():
5685
5686 // Compute address and memory type.
5687 int offset = field->offset_in_bytes();
5688 bool is_vol = field->is_volatile();
5689 ciType* field_klass = field->type();
5690 assert(field_klass->is_loaded(), "should be loaded");
5691 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5692 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5693 BasicType bt = field->layout_type();
5694
5695 // Build the resultant type of the load
5696 const Type *type;
5697 if (bt == T_OBJECT) {
5698 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5699 } else {
5700 type = Type::get_const_basic_type(bt);
5701 }
5702
5703 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5704 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5705 }
5706 // Build the load.
5707 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5708 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5709 // If reference is volatile, prevent following memory ops from
5710 // floating up past the volatile read. Also prevents commoning
5711 // another volatile read.
5712 if (is_vol) {
5713 // Memory barrier includes bogus read of value to force load BEFORE membar
5714 insert_mem_bar(Op_MemBarAcquire, loadedField);
5715 }
5716 return loadedField;
5717 }
5718
5719
5720 //------------------------------inline_aescrypt_Block-----------------------
5721 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5722 address stubAddr;
5723 const char *stubName;
5724 assert(UseAES, "need AES instruction support");
5725
5726 switch(id) {
5727 case vmIntrinsics::_aescrypt_encryptBlock:
5728 stubAddr = StubRoutines::aescrypt_encryptBlock();
5729 stubName = "aescrypt_encryptBlock";
5730 break;
5731 case vmIntrinsics::_aescrypt_decryptBlock:
5732 stubAddr = StubRoutines::aescrypt_decryptBlock();
5733 stubName = "aescrypt_decryptBlock";
5734 break;
5735 }
5736 if (stubAddr == NULL) return false;
5737
5738 Node* aescrypt_object = argument(0);
5739 Node* src = argument(1);
5740 Node* src_offset = argument(2);
5741 Node* dest = argument(3);
5742 Node* dest_offset = argument(4);
5743
5744 // (1) src and dest are arrays.
5745 const Type* src_type = src->Value(&_gvn);
5746 const Type* dest_type = dest->Value(&_gvn);
5747 const TypeAryPtr* top_src = src_type->isa_aryptr();
5748 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5749 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5750
5751 // for the quick and dirty code we will skip all the checks.
5752 // we are just trying to get the call to be generated.
5753 Node* src_start = src;
5754 Node* dest_start = dest;
5755 if (src_offset != NULL || dest_offset != NULL) {
5756 assert(src_offset != NULL && dest_offset != NULL, "");
5757 src_start = array_element_address(src, src_offset, T_BYTE);
5758 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5759 }
5760
5761 // now need to get the start of its expanded key array
5762 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5763 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5764 if (k_start == NULL) return false;
5765
5766 if (Matcher::pass_original_key_for_aes()) {
5767 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5768 // compatibility issues between Java key expansion and SPARC crypto instructions
5769 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5770 if (original_k_start == NULL) return false;
5771
5772 // Call the stub.
5773 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5774 stubAddr, stubName, TypePtr::BOTTOM,
5775 src_start, dest_start, k_start, original_k_start);
5776 } else {
5777 // Call the stub.
5778 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5779 stubAddr, stubName, TypePtr::BOTTOM,
5780 src_start, dest_start, k_start);
5781 }
5782
5783 return true;
5784 }
5785
5786 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5787 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5788 address stubAddr;
5789 const char *stubName;
5790
5791 assert(UseAES, "need AES instruction support");
5792
5793 switch(id) {
5794 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5795 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5796 stubName = "cipherBlockChaining_encryptAESCrypt";
5797 break;
5798 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5799 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5800 stubName = "cipherBlockChaining_decryptAESCrypt";
5801 break;
5802 }
5803 if (stubAddr == NULL) return false;
5804
5805 Node* cipherBlockChaining_object = argument(0);
5806 Node* src = argument(1);
5807 Node* src_offset = argument(2);
5808 Node* len = argument(3);
5809 Node* dest = argument(4);
5810 Node* dest_offset = argument(5);
5811
5812 // (1) src and dest are arrays.
5813 const Type* src_type = src->Value(&_gvn);
5814 const Type* dest_type = dest->Value(&_gvn);
5815 const TypeAryPtr* top_src = src_type->isa_aryptr();
5816 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5817 assert (top_src != NULL && top_src->klass() != NULL
5818 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5819
5820 // checks are the responsibility of the caller
5821 Node* src_start = src;
5822 Node* dest_start = dest;
5823 if (src_offset != NULL || dest_offset != NULL) {
5824 assert(src_offset != NULL && dest_offset != NULL, "");
5825 src_start = array_element_address(src, src_offset, T_BYTE);
5826 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5827 }
5828
5829 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5830 // (because of the predicated logic executed earlier).
5831 // so we cast it here safely.
5832 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5833
5834 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5835 if (embeddedCipherObj == NULL) return false;
5836
5837 // cast it to what we know it will be at runtime
5838 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5839 assert(tinst != NULL, "CBC obj is null");
5840 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5841 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5842 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5843
5844 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5845 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5846 const TypeOopPtr* xtype = aklass->as_instance_type();
5847 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5848 aescrypt_object = _gvn.transform(aescrypt_object);
5849
5850 // we need to get the start of the aescrypt_object's expanded key array
5851 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5852 if (k_start == NULL) return false;
5853
5854 // similarly, get the start address of the r vector
5855 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5856 if (objRvec == NULL) return false;
5857 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5858
5859 Node* cbcCrypt;
5860 if (Matcher::pass_original_key_for_aes()) {
5861 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5862 // compatibility issues between Java key expansion and SPARC crypto instructions
5863 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5864 if (original_k_start == NULL) return false;
5865
5866 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5867 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5868 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5869 stubAddr, stubName, TypePtr::BOTTOM,
5870 src_start, dest_start, k_start, r_start, len, original_k_start);
5871 } else {
5872 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
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);
5877 }
5878
5879 // return cipher length (int)
5880 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5881 set_result(retvalue);
5882 return true;
5883 }
5884
5885 //------------------------------get_key_start_from_aescrypt_object-----------------------
5886 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5887 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5888 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5889 if (objAESCryptKey == NULL) return (Node *) NULL;
5890
5891 // now have the array, need to get the start address of the K array
5892 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5893 return k_start;
5894 }
5895
5896 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5897 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5898 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5899 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5900 if (objAESCryptKey == NULL) return (Node *) NULL;
5901
5902 // now have the array, need to get the start address of the lastKey array
5903 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5904 return original_k_start;
5905 }
5906
5907 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5908 // Return node representing slow path of predicate check.
5909 // the pseudo code we want to emulate with this predicate is:
5910 // for encryption:
5911 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5912 // for decryption:
5913 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5914 // note cipher==plain is more conservative than the original java code but that's OK
5915 //
5916 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5917 // The receiver was checked for NULL already.
5918 Node* objCBC = argument(0);
5919
5920 // Load embeddedCipher field of CipherBlockChaining object.
5921 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5922
5923 // get AESCrypt klass for instanceOf check
5924 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5925 // will have same classloader as CipherBlockChaining object
5926 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5927 assert(tinst != NULL, "CBCobj is null");
5928 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5929
5930 // we want to do an instanceof comparison against the AESCrypt class
5931 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5932 if (!klass_AESCrypt->is_loaded()) {
5933 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5934 Node* ctrl = control();
5935 set_control(top()); // no regular fast path
5936 return ctrl;
5937 }
5938 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5939
5940 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5941 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5942 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5943
5944 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5945
5946 // for encryption, we are done
5947 if (!decrypting)
5948 return instof_false; // even if it is NULL
5949
5950 // for decryption, we need to add a further check to avoid
5951 // taking the intrinsic path when cipher and plain are the same
5952 // see the original java code for why.
5953 RegionNode* region = new RegionNode(3);
5954 region->init_req(1, instof_false);
5955 Node* src = argument(1);
5956 Node* dest = argument(4);
5957 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5958 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5959 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5960 region->init_req(2, src_dest_conjoint);
5961
5962 record_for_igvn(region);
5963 return _gvn.transform(region);
5964 }
5965
5966 //------------------------------inline_ghash_processBlocks
5967 bool LibraryCallKit::inline_ghash_processBlocks() {
5968 address stubAddr;
5969 const char *stubName;
5970 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5971
5972 stubAddr = StubRoutines::ghash_processBlocks();
5973 stubName = "ghash_processBlocks";
5974
5975 Node* data = argument(0);
5976 Node* offset = argument(1);
5977 Node* len = argument(2);
5978 Node* state = argument(3);
5979 Node* subkeyH = argument(4);
5980
5981 Node* state_start = array_element_address(state, intcon(0), T_LONG);
5982 assert(state_start, "state is NULL");
5983 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
5984 assert(subkeyH_start, "subkeyH is NULL");
5985 Node* data_start = array_element_address(data, offset, T_BYTE);
5986 assert(data_start, "data is NULL");
5987
5988 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
5989 OptoRuntime::ghash_processBlocks_Type(),
5990 stubAddr, stubName, TypePtr::BOTTOM,
5991 state_start, subkeyH_start, data_start, len);
5992 return true;
5993 }
5994
5995 //------------------------------inline_sha_implCompress-----------------------
5996 //
5997 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
5998 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
5999 //
6000 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6001 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6002 //
6003 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6004 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6005 //
6006 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6007 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6008
6009 Node* sha_obj = argument(0);
6010 Node* src = argument(1); // type oop
6011 Node* ofs = argument(2); // type int
6012
6013 const Type* src_type = src->Value(&_gvn);
6014 const TypeAryPtr* top_src = src_type->isa_aryptr();
6015 if (top_src == NULL || top_src->klass() == NULL) {
6016 // failed array check
6017 return false;
6018 }
6019 // Figure out the size and type of the elements we will be copying.
6020 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6021 if (src_elem != T_BYTE) {
6022 return false;
6023 }
6024 // 'src_start' points to src array + offset
6025 Node* src_start = array_element_address(src, ofs, src_elem);
6026 Node* state = NULL;
6027 address stubAddr;
6028 const char *stubName;
6029
6030 switch(id) {
6031 case vmIntrinsics::_sha_implCompress:
6032 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6033 state = get_state_from_sha_object(sha_obj);
6034 stubAddr = StubRoutines::sha1_implCompress();
6035 stubName = "sha1_implCompress";
6036 break;
6037 case vmIntrinsics::_sha2_implCompress:
6038 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6039 state = get_state_from_sha_object(sha_obj);
6040 stubAddr = StubRoutines::sha256_implCompress();
6041 stubName = "sha256_implCompress";
6042 break;
6043 case vmIntrinsics::_sha5_implCompress:
6044 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6045 state = get_state_from_sha5_object(sha_obj);
6046 stubAddr = StubRoutines::sha512_implCompress();
6047 stubName = "sha512_implCompress";
6048 break;
6049 default:
6050 fatal_unexpected_iid(id);
6051 return false;
6052 }
6053 if (state == NULL) return false;
6054
6055 // Call the stub.
6056 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6057 stubAddr, stubName, TypePtr::BOTTOM,
6058 src_start, state);
6059
6060 return true;
6061 }
6062
6063 //------------------------------inline_digestBase_implCompressMB-----------------------
6064 //
6065 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6066 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6067 //
6068 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6069 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6070 "need SHA1/SHA256/SHA512 instruction support");
6071 assert((uint)predicate < 3, "sanity");
6072 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6073
6074 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6075 Node* src = argument(1); // byte[] array
6076 Node* ofs = argument(2); // type int
6077 Node* limit = argument(3); // type int
6078
6079 const Type* src_type = src->Value(&_gvn);
6080 const TypeAryPtr* top_src = src_type->isa_aryptr();
6081 if (top_src == NULL || top_src->klass() == NULL) {
6082 // failed array check
6083 return false;
6084 }
6085 // Figure out the size and type of the elements we will be copying.
6086 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6087 if (src_elem != T_BYTE) {
6088 return false;
6089 }
6090 // 'src_start' points to src array + offset
6091 Node* src_start = array_element_address(src, ofs, src_elem);
6092
6093 const char* klass_SHA_name = NULL;
6094 const char* stub_name = NULL;
6095 address stub_addr = NULL;
6096 bool long_state = false;
6097
6098 switch (predicate) {
6099 case 0:
6100 if (UseSHA1Intrinsics) {
6101 klass_SHA_name = "sun/security/provider/SHA";
6102 stub_name = "sha1_implCompressMB";
6103 stub_addr = StubRoutines::sha1_implCompressMB();
6104 }
6105 break;
6106 case 1:
6107 if (UseSHA256Intrinsics) {
6108 klass_SHA_name = "sun/security/provider/SHA2";
6109 stub_name = "sha256_implCompressMB";
6110 stub_addr = StubRoutines::sha256_implCompressMB();
6111 }
6112 break;
6113 case 2:
6114 if (UseSHA512Intrinsics) {
6115 klass_SHA_name = "sun/security/provider/SHA5";
6116 stub_name = "sha512_implCompressMB";
6117 stub_addr = StubRoutines::sha512_implCompressMB();
6118 long_state = true;
6119 }
6120 break;
6121 default:
6122 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6123 }
6124 if (klass_SHA_name != NULL) {
6125 // get DigestBase klass to lookup for SHA klass
6126 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6127 assert(tinst != NULL, "digestBase_obj is not instance???");
6128 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6129
6130 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6131 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6132 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6133 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6134 }
6135 return false;
6136 }
6137 //------------------------------inline_sha_implCompressMB-----------------------
6138 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6139 bool long_state, address stubAddr, const char *stubName,
6140 Node* src_start, Node* ofs, Node* limit) {
6141 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6142 const TypeOopPtr* xtype = aklass->as_instance_type();
6143 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6144 sha_obj = _gvn.transform(sha_obj);
6145
6146 Node* state;
6147 if (long_state) {
6148 state = get_state_from_sha5_object(sha_obj);
6149 } else {
6150 state = get_state_from_sha_object(sha_obj);
6151 }
6152 if (state == NULL) return false;
6153
6154 // Call the stub.
6155 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6156 OptoRuntime::digestBase_implCompressMB_Type(),
6157 stubAddr, stubName, TypePtr::BOTTOM,
6158 src_start, state, ofs, limit);
6159 // return ofs (int)
6160 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6161 set_result(result);
6162
6163 return true;
6164 }
6165
6166 //------------------------------get_state_from_sha_object-----------------------
6167 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6168 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6169 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6170 if (sha_state == NULL) return (Node *) NULL;
6171
6172 // now have the array, need to get the start address of the state array
6173 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6174 return state;
6175 }
6176
6177 //------------------------------get_state_from_sha5_object-----------------------
6178 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6179 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6180 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6181 if (sha_state == NULL) return (Node *) NULL;
6182
6183 // now have the array, need to get the start address of the state array
6184 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6185 return state;
6186 }
6187
6188 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6189 // Return node representing slow path of predicate check.
6190 // the pseudo code we want to emulate with this predicate is:
6191 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6192 //
6193 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6194 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6195 "need SHA1/SHA256/SHA512 instruction support");
6196 assert((uint)predicate < 3, "sanity");
6197
6198 // The receiver was checked for NULL already.
6199 Node* digestBaseObj = argument(0);
6200
6201 // get DigestBase klass for instanceOf check
6202 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6203 assert(tinst != NULL, "digestBaseObj is null");
6204 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6205
6206 const char* klass_SHA_name = NULL;
6207 switch (predicate) {
6208 case 0:
6209 if (UseSHA1Intrinsics) {
6210 // we want to do an instanceof comparison against the SHA class
6211 klass_SHA_name = "sun/security/provider/SHA";
6212 }
6213 break;
6214 case 1:
6215 if (UseSHA256Intrinsics) {
6216 // we want to do an instanceof comparison against the SHA2 class
6217 klass_SHA_name = "sun/security/provider/SHA2";
6218 }
6219 break;
6220 case 2:
6221 if (UseSHA512Intrinsics) {
6222 // we want to do an instanceof comparison against the SHA5 class
6223 klass_SHA_name = "sun/security/provider/SHA5";
6224 }
6225 break;
6226 default:
6227 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6228 }
6229
6230 ciKlass* klass_SHA = NULL;
6231 if (klass_SHA_name != NULL) {
6232 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6233 }
6234 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6235 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6236 Node* ctrl = control();
6237 set_control(top()); // no intrinsic path
6238 return ctrl;
6239 }
6240 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6241
6242 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6243 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6244 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6245 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6246
6247 return instof_false; // even if it is NULL
6248 }
6249
6250 bool LibraryCallKit::inline_profileBoolean() {
6251 Node* counts = argument(1);
6252 const TypeAryPtr* ary = NULL;
6253 ciArray* aobj = NULL;
6254 if (counts->is_Con()
6255 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6256 && (aobj = ary->const_oop()->as_array()) != NULL
6257 && (aobj->length() == 2)) {
6258 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6259 jint false_cnt = aobj->element_value(0).as_int();
6260 jint true_cnt = aobj->element_value(1).as_int();
6261
6262 if (C->log() != NULL) {
6263 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6264 false_cnt, true_cnt);
6265 }
6266
6267 if (false_cnt + true_cnt == 0) {
6268 // According to profile, never executed.
6269 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6270 Deoptimization::Action_reinterpret);
6271 return true;
6272 }
6273
6274 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6275 // is a number of each value occurrences.
6276 Node* result = argument(0);
6277 if (false_cnt == 0 || true_cnt == 0) {
6278 // According to profile, one value has been never seen.
6279 int expected_val = (false_cnt == 0) ? 1 : 0;
6280
6281 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6282 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6283
6284 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6285 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6286 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6287
6288 { // Slow path: uncommon trap for never seen value and then reexecute
6289 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6290 // the value has been seen at least once.
6291 PreserveJVMState pjvms(this);
6292 PreserveReexecuteState preexecs(this);
6293 jvms()->set_should_reexecute(true);
6294
6295 set_control(slow_path);
6296 set_i_o(i_o());
6297
6298 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6299 Deoptimization::Action_reinterpret);
6300 }
6301 // The guard for never seen value enables sharpening of the result and
6302 // returning a constant. It allows to eliminate branches on the same value
6303 // later on.
6304 set_control(fast_path);
6305 result = intcon(expected_val);
6306 }
6307 // Stop profiling.
6308 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6309 // By replacing method body with profile data (represented as ProfileBooleanNode
6310 // on IR level) we effectively disable profiling.
6311 // It enables full speed execution once optimized code is generated.
6312 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6313 C->record_for_igvn(profile);
6314 set_result(profile);
6315 return true;
6316 } else {
6317 // Continue profiling.
6318 // Profile data isn't available at the moment. So, execute method's bytecode version.
6319 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6320 // is compiled and counters aren't available since corresponding MethodHandle
6321 // isn't a compile-time constant.
6322 return false;
6323 }
6324 }
6325
6326 bool LibraryCallKit::inline_isCompileConstant() {
6327 Node* n = argument(0);
6328 set_result(n->is_Con() ? intcon(1) : intcon(0));
6329 return true;
6330 }