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_exp();
226 bool inline_pow();
227 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
228 bool inline_min_max(vmIntrinsics::ID id);
229 bool inline_notify(vmIntrinsics::ID id);
230 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
231 // This returns Type::AnyPtr, RawPtr, or OopPtr.
232 int classify_unsafe_addr(Node* &base, Node* &offset);
233 Node* make_unsafe_address(Node* base, Node* offset);
234 // Helper for inline_unsafe_access.
235 // Generates the guards that check whether the result of
236 // Unsafe.getObject should be recorded in an SATB log buffer.
237 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
238 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
239 static bool klass_needs_init_guard(Node* kls);
240 bool inline_unsafe_allocate();
241 bool inline_unsafe_copyMemory();
242 bool inline_native_currentThread();
243 #ifdef TRACE_HAVE_INTRINSICS
244 bool inline_native_classID();
245 bool inline_native_threadID();
246 #endif
247 bool inline_native_time_funcs(address method, const char* funcName);
248 bool inline_native_isInterrupted();
249 bool inline_native_Class_query(vmIntrinsics::ID id);
250 bool inline_native_subtype_check();
251
252 bool inline_native_newArray();
253 bool inline_native_getLength();
254 bool inline_array_copyOf(bool is_copyOfRange);
255 bool inline_array_equals();
256 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
257 bool inline_native_clone(bool is_virtual);
258 bool inline_native_Reflection_getCallerClass();
259 // Helper function for inlining native object hash method
260 bool inline_native_hashcode(bool is_virtual, bool is_static);
261 bool inline_native_getClass();
262
263 // Helper functions for inlining arraycopy
264 bool inline_arraycopy();
265 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
266 RegionNode* slow_region);
267 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
268 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
269
270 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
271 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
272 bool inline_unsafe_ordered_store(BasicType type);
273 bool inline_unsafe_fence(vmIntrinsics::ID id);
274 bool inline_fp_conversions(vmIntrinsics::ID id);
275 bool inline_number_methods(vmIntrinsics::ID id);
276 bool inline_reference_get();
277 bool inline_Class_cast();
278 bool inline_aescrypt_Block(vmIntrinsics::ID id);
279 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
280 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
281 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
282 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
283 bool inline_ghash_processBlocks();
284 bool inline_sha_implCompress(vmIntrinsics::ID id);
285 bool inline_digestBase_implCompressMB(int predicate);
286 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
287 bool long_state, address stubAddr, const char *stubName,
288 Node* src_start, Node* ofs, Node* limit);
289 Node* get_state_from_sha_object(Node *sha_object);
290 Node* get_state_from_sha5_object(Node *sha_object);
291 Node* inline_digestBase_implCompressMB_predicate(int predicate);
292 bool inline_encodeISOArray();
293 bool inline_updateCRC32();
294 bool inline_updateBytesCRC32();
295 bool inline_updateByteBufferCRC32();
296 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
297 bool inline_updateBytesCRC32C();
298 bool inline_updateDirectByteBufferCRC32C();
299 bool inline_updateBytesAdler32();
300 bool inline_updateByteBufferAdler32();
301 bool inline_multiplyToLen();
302 bool inline_squareToLen();
303 bool inline_mulAdd();
304 bool inline_montgomeryMultiply();
305 bool inline_montgomerySquare();
306
307 bool inline_profileBoolean();
308 bool inline_isCompileConstant();
309 };
310
311 //---------------------------make_vm_intrinsic----------------------------
312 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
313 vmIntrinsics::ID id = m->intrinsic_id();
314 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
315
316 if (!m->is_loaded()) {
317 // Do not attempt to inline unloaded methods.
318 return NULL;
319 }
320
321 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
322 bool is_available = false;
323
324 {
325 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
326 // the compiler must transition to '_thread_in_vm' state because both
327 // methods access VM-internal data.
328 VM_ENTRY_MARK;
329 methodHandle mh(THREAD, m->get_Method());
330 methodHandle ct(THREAD, method()->get_Method());
331 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
332 !vmIntrinsics::is_disabled_by_flags(mh, ct);
333 }
334
335 if (is_available) {
336 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
337 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
338 return new LibraryIntrinsic(m, is_virtual,
339 vmIntrinsics::predicates_needed(id),
340 vmIntrinsics::does_virtual_dispatch(id),
341 (vmIntrinsics::ID) id);
342 } else {
343 return NULL;
344 }
345 }
346
347 //----------------------register_library_intrinsics-----------------------
348 // Initialize this file's data structures, for each Compile instance.
349 void Compile::register_library_intrinsics() {
350 // Nothing to do here.
351 }
352
353 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
354 LibraryCallKit kit(jvms, this);
355 Compile* C = kit.C;
356 int nodes = C->unique();
357 #ifndef PRODUCT
358 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
359 char buf[1000];
360 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
361 tty->print_cr("Intrinsic %s", str);
362 }
363 #endif
364 ciMethod* callee = kit.callee();
365 const int bci = kit.bci();
366
367 // Try to inline the intrinsic.
368 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
369 kit.try_to_inline(_last_predicate)) {
370 if (C->print_intrinsics() || C->print_inlining()) {
371 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
372 }
373 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
374 if (C->log()) {
375 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
376 vmIntrinsics::name_at(intrinsic_id()),
377 (is_virtual() ? " virtual='1'" : ""),
378 C->unique() - nodes);
379 }
380 // Push the result from the inlined method onto the stack.
381 kit.push_result();
382 C->print_inlining_update(this);
383 return kit.transfer_exceptions_into_jvms();
384 }
385
386 // The intrinsic bailed out
387 if (C->print_intrinsics() || C->print_inlining()) {
388 if (jvms->has_method()) {
389 // Not a root compile.
390 const char* msg;
391 if (callee->intrinsic_candidate()) {
392 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
393 } else {
394 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
395 : "failed to inline (intrinsic), method not annotated";
396 }
397 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
398 } else {
399 // Root compile
400 tty->print("Did not generate intrinsic %s%s at bci:%d in",
401 vmIntrinsics::name_at(intrinsic_id()),
402 (is_virtual() ? " (virtual)" : ""), bci);
403 }
404 }
405 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
406 C->print_inlining_update(this);
407 return NULL;
408 }
409
410 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
411 LibraryCallKit kit(jvms, this);
412 Compile* C = kit.C;
413 int nodes = C->unique();
414 _last_predicate = predicate;
415 #ifndef PRODUCT
416 assert(is_predicated() && predicate < predicates_count(), "sanity");
417 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
418 char buf[1000];
419 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
420 tty->print_cr("Predicate for intrinsic %s", str);
421 }
422 #endif
423 ciMethod* callee = kit.callee();
424 const int bci = kit.bci();
425
426 Node* slow_ctl = kit.try_to_predicate(predicate);
427 if (!kit.failing()) {
428 if (C->print_intrinsics() || C->print_inlining()) {
429 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
430 }
431 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
432 if (C->log()) {
433 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
434 vmIntrinsics::name_at(intrinsic_id()),
435 (is_virtual() ? " virtual='1'" : ""),
436 C->unique() - nodes);
437 }
438 return slow_ctl; // Could be NULL if the check folds.
439 }
440
441 // The intrinsic bailed out
442 if (C->print_intrinsics() || C->print_inlining()) {
443 if (jvms->has_method()) {
444 // Not a root compile.
445 const char* msg = "failed to generate predicate for intrinsic";
446 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
447 } else {
448 // Root compile
449 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
450 vmIntrinsics::name_at(intrinsic_id()),
451 (is_virtual() ? " (virtual)" : ""), bci);
452 }
453 }
454 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
455 return NULL;
456 }
457
458 bool LibraryCallKit::try_to_inline(int predicate) {
459 // Handle symbolic names for otherwise undistinguished boolean switches:
460 const bool is_store = true;
461 const bool is_native_ptr = true;
462 const bool is_static = true;
463 const bool is_volatile = true;
464
465 if (!jvms()->has_method()) {
466 // Root JVMState has a null method.
467 assert(map()->memory()->Opcode() == Op_Parm, "");
468 // Insert the memory aliasing node
469 set_all_memory(reset_memory());
470 }
471 assert(merged_memory(), "");
472
473
474 switch (intrinsic_id()) {
475 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
476 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
477 case vmIntrinsics::_getClass: return inline_native_getClass();
478
479 case vmIntrinsics::_dsin:
480 case vmIntrinsics::_dcos:
481 case vmIntrinsics::_dtan:
482 case vmIntrinsics::_dabs:
483 case vmIntrinsics::_datan2:
484 case vmIntrinsics::_dsqrt:
485 case vmIntrinsics::_dexp:
486 case vmIntrinsics::_dlog:
487 case vmIntrinsics::_dlog10:
488 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
489
490 case vmIntrinsics::_min:
491 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
492
493 case vmIntrinsics::_notify:
494 case vmIntrinsics::_notifyAll:
495 if (InlineNotify) {
496 return inline_notify(intrinsic_id());
497 }
498 return false;
499
500 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
501 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
502 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
503 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
504 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
505 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
506 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
507 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
508 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
509 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
510 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
511 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
512
513 case vmIntrinsics::_arraycopy: return inline_arraycopy();
514
515 case vmIntrinsics::_compareTo: return inline_string_compareTo();
516 case vmIntrinsics::_indexOf: return inline_string_indexOf();
517 case vmIntrinsics::_equals: return inline_string_equals();
518
519 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
520 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
521 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
522 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
523 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
524 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
525 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
526 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
527 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
528 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
529 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
530 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
531 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
532 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
533 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
534 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
535 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
536 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
537
538 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
539 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
540 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
541 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
542 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
543 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
544 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
545 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
546
547 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
548 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
549 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
550 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
551 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
552 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
553 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
554 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
555
556 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
557 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
558 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
559 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
560 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
561 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
562 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
563 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
564 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
565
566 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
567 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
568 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
569 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
570 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
571 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
572 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
573 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
574 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
575
576 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
577 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
578 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
579 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
580
581 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
582 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
583 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
584 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
585
586 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
587 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
588 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
589
590 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
591 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
592 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
593
594 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
595 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
596 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
597 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
598 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
599
600 case vmIntrinsics::_loadFence:
601 case vmIntrinsics::_storeFence:
602 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
603
604 case vmIntrinsics::_currentThread: return inline_native_currentThread();
605 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
606
607 #ifdef TRACE_HAVE_INTRINSICS
608 case vmIntrinsics::_classID: return inline_native_classID();
609 case vmIntrinsics::_threadID: return inline_native_threadID();
610 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
611 #endif
612 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
613 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
614 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
615 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
616 case vmIntrinsics::_newArray: return inline_native_newArray();
617 case vmIntrinsics::_getLength: return inline_native_getLength();
618 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
619 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
620 case vmIntrinsics::_equalsC: return inline_array_equals();
621 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
622
623 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
624
625 case vmIntrinsics::_isInstance:
626 case vmIntrinsics::_getModifiers:
627 case vmIntrinsics::_isInterface:
628 case vmIntrinsics::_isArray:
629 case vmIntrinsics::_isPrimitive:
630 case vmIntrinsics::_getSuperclass:
631 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
632
633 case vmIntrinsics::_floatToRawIntBits:
634 case vmIntrinsics::_floatToIntBits:
635 case vmIntrinsics::_intBitsToFloat:
636 case vmIntrinsics::_doubleToRawLongBits:
637 case vmIntrinsics::_doubleToLongBits:
638 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
639
640 case vmIntrinsics::_numberOfLeadingZeros_i:
641 case vmIntrinsics::_numberOfLeadingZeros_l:
642 case vmIntrinsics::_numberOfTrailingZeros_i:
643 case vmIntrinsics::_numberOfTrailingZeros_l:
644 case vmIntrinsics::_bitCount_i:
645 case vmIntrinsics::_bitCount_l:
646 case vmIntrinsics::_reverseBytes_i:
647 case vmIntrinsics::_reverseBytes_l:
648 case vmIntrinsics::_reverseBytes_s:
649 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
650
651 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
652
653 case vmIntrinsics::_Reference_get: return inline_reference_get();
654
655 case vmIntrinsics::_Class_cast: return inline_Class_cast();
656
657 case vmIntrinsics::_aescrypt_encryptBlock:
658 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
659
660 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
661 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
662 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
663
664 case vmIntrinsics::_sha_implCompress:
665 case vmIntrinsics::_sha2_implCompress:
666 case vmIntrinsics::_sha5_implCompress:
667 return inline_sha_implCompress(intrinsic_id());
668
669 case vmIntrinsics::_digestBase_implCompressMB:
670 return inline_digestBase_implCompressMB(predicate);
671
672 case vmIntrinsics::_multiplyToLen:
673 return inline_multiplyToLen();
674
675 case vmIntrinsics::_squareToLen:
676 return inline_squareToLen();
677
678 case vmIntrinsics::_mulAdd:
679 return inline_mulAdd();
680
681 case vmIntrinsics::_montgomeryMultiply:
682 return inline_montgomeryMultiply();
683 case vmIntrinsics::_montgomerySquare:
684 return inline_montgomerySquare();
685
686 case vmIntrinsics::_ghash_processBlocks:
687 return inline_ghash_processBlocks();
688
689 case vmIntrinsics::_encodeISOArray:
690 return inline_encodeISOArray();
691
692 case vmIntrinsics::_updateCRC32:
693 return inline_updateCRC32();
694 case vmIntrinsics::_updateBytesCRC32:
695 return inline_updateBytesCRC32();
696 case vmIntrinsics::_updateByteBufferCRC32:
697 return inline_updateByteBufferCRC32();
698
699 case vmIntrinsics::_updateBytesCRC32C:
700 return inline_updateBytesCRC32C();
701 case vmIntrinsics::_updateDirectByteBufferCRC32C:
702 return inline_updateDirectByteBufferCRC32C();
703
704 case vmIntrinsics::_updateBytesAdler32:
705 return inline_updateBytesAdler32();
706 case vmIntrinsics::_updateByteBufferAdler32:
707 return inline_updateByteBufferAdler32();
708
709 case vmIntrinsics::_profileBoolean:
710 return inline_profileBoolean();
711 case vmIntrinsics::_isCompileConstant:
712 return inline_isCompileConstant();
713
714 default:
715 // If you get here, it may be that someone has added a new intrinsic
716 // to the list in vmSymbols.hpp without implementing it here.
717 #ifndef PRODUCT
718 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
719 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
720 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
721 }
722 #endif
723 return false;
724 }
725 }
726
727 Node* LibraryCallKit::try_to_predicate(int predicate) {
728 if (!jvms()->has_method()) {
729 // Root JVMState has a null method.
730 assert(map()->memory()->Opcode() == Op_Parm, "");
731 // Insert the memory aliasing node
732 set_all_memory(reset_memory());
733 }
734 assert(merged_memory(), "");
735
736 switch (intrinsic_id()) {
737 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
738 return inline_cipherBlockChaining_AESCrypt_predicate(false);
739 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
740 return inline_cipherBlockChaining_AESCrypt_predicate(true);
741 case vmIntrinsics::_digestBase_implCompressMB:
742 return inline_digestBase_implCompressMB_predicate(predicate);
743
744 default:
745 // If you get here, it may be that someone has added a new intrinsic
746 // to the list in vmSymbols.hpp without implementing it here.
747 #ifndef PRODUCT
748 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
749 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
750 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
751 }
752 #endif
753 Node* slow_ctl = control();
754 set_control(top()); // No fast path instrinsic
755 return slow_ctl;
756 }
757 }
758
759 //------------------------------set_result-------------------------------
760 // Helper function for finishing intrinsics.
761 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
762 record_for_igvn(region);
763 set_control(_gvn.transform(region));
764 set_result( _gvn.transform(value));
765 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
766 }
767
768 //------------------------------generate_guard---------------------------
769 // Helper function for generating guarded fast-slow graph structures.
770 // The given 'test', if true, guards a slow path. If the test fails
771 // then a fast path can be taken. (We generally hope it fails.)
772 // In all cases, GraphKit::control() is updated to the fast path.
773 // The returned value represents the control for the slow path.
774 // The return value is never 'top'; it is either a valid control
775 // or NULL if it is obvious that the slow path can never be taken.
776 // Also, if region and the slow control are not NULL, the slow edge
777 // is appended to the region.
778 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
779 if (stopped()) {
780 // Already short circuited.
781 return NULL;
782 }
783
784 // Build an if node and its projections.
785 // If test is true we take the slow path, which we assume is uncommon.
786 if (_gvn.type(test) == TypeInt::ZERO) {
787 // The slow branch is never taken. No need to build this guard.
788 return NULL;
789 }
790
791 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
792
793 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
794 if (if_slow == top()) {
795 // The slow branch is never taken. No need to build this guard.
796 return NULL;
797 }
798
799 if (region != NULL)
800 region->add_req(if_slow);
801
802 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
803 set_control(if_fast);
804
805 return if_slow;
806 }
807
808 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
809 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
810 }
811 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
812 return generate_guard(test, region, PROB_FAIR);
813 }
814
815 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
816 Node* *pos_index) {
817 if (stopped())
818 return NULL; // already stopped
819 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
820 return NULL; // index is already adequately typed
821 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
822 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
823 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
824 if (is_neg != NULL && pos_index != NULL) {
825 // Emulate effect of Parse::adjust_map_after_if.
826 Node* ccast = new CastIINode(index, TypeInt::POS);
827 ccast->set_req(0, control());
828 (*pos_index) = _gvn.transform(ccast);
829 }
830 return is_neg;
831 }
832
833 // Make sure that 'position' is a valid limit index, in [0..length].
834 // There are two equivalent plans for checking this:
835 // A. (offset + copyLength) unsigned<= arrayLength
836 // B. offset <= (arrayLength - copyLength)
837 // We require that all of the values above, except for the sum and
838 // difference, are already known to be non-negative.
839 // Plan A is robust in the face of overflow, if offset and copyLength
840 // are both hugely positive.
841 //
842 // Plan B is less direct and intuitive, but it does not overflow at
843 // all, since the difference of two non-negatives is always
844 // representable. Whenever Java methods must perform the equivalent
845 // check they generally use Plan B instead of Plan A.
846 // For the moment we use Plan A.
847 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
848 Node* subseq_length,
849 Node* array_length,
850 RegionNode* region) {
851 if (stopped())
852 return NULL; // already stopped
853 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
854 if (zero_offset && subseq_length->eqv_uncast(array_length))
855 return NULL; // common case of whole-array copy
856 Node* last = subseq_length;
857 if (!zero_offset) // last += offset
858 last = _gvn.transform(new AddINode(last, offset));
859 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
860 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
861 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
862 return is_over;
863 }
864
865
866 //--------------------------generate_current_thread--------------------
867 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
868 ciKlass* thread_klass = env()->Thread_klass();
869 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
870 Node* thread = _gvn.transform(new ThreadLocalNode());
871 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
872 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
873 tls_output = thread;
874 return threadObj;
875 }
876
877
878 //------------------------------make_string_method_node------------------------
879 // Helper method for String intrinsic functions. This version is called
880 // with str1 and str2 pointing to String object nodes.
881 //
882 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
883 Node* no_ctrl = NULL;
884
885 // Get start addr of string
886 Node* str1_value = load_String_value(no_ctrl, str1);
887 Node* str1_offset = load_String_offset(no_ctrl, str1);
888 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
889
890 // Get length of string 1
891 Node* str1_len = load_String_length(no_ctrl, str1);
892
893 Node* str2_value = load_String_value(no_ctrl, str2);
894 Node* str2_offset = load_String_offset(no_ctrl, str2);
895 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
896
897 Node* str2_len = NULL;
898 Node* result = NULL;
899
900 switch (opcode) {
901 case Op_StrIndexOf:
902 // Get length of string 2
903 str2_len = load_String_length(no_ctrl, str2);
904
905 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
906 str1_start, str1_len, str2_start, str2_len);
907 break;
908 case Op_StrComp:
909 // Get length of string 2
910 str2_len = load_String_length(no_ctrl, str2);
911
912 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
913 str1_start, str1_len, str2_start, str2_len);
914 break;
915 case Op_StrEquals:
916 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
917 str1_start, str2_start, str1_len);
918 break;
919 default:
920 ShouldNotReachHere();
921 return NULL;
922 }
923
924 // All these intrinsics have checks.
925 C->set_has_split_ifs(true); // Has chance for split-if optimization
926
927 return _gvn.transform(result);
928 }
929
930 // Helper method for String intrinsic functions. This version is called
931 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
932 // to Int nodes containing the lenghts of str1 and str2.
933 //
934 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
935 Node* result = NULL;
936 switch (opcode) {
937 case Op_StrIndexOf:
938 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
939 str1_start, cnt1, str2_start, cnt2);
940 break;
941 case Op_StrComp:
942 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
943 str1_start, cnt1, str2_start, cnt2);
944 break;
945 case Op_StrEquals:
946 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
947 str1_start, str2_start, cnt1);
948 break;
949 default:
950 ShouldNotReachHere();
951 return NULL;
952 }
953
954 // All these intrinsics have checks.
955 C->set_has_split_ifs(true); // Has chance for split-if optimization
956
957 return _gvn.transform(result);
958 }
959
960 //------------------------------inline_string_compareTo------------------------
961 // public int java.lang.String.compareTo(String anotherString);
962 bool LibraryCallKit::inline_string_compareTo() {
963 Node* receiver = null_check(argument(0));
964 Node* arg = null_check(argument(1));
965 if (stopped()) {
966 return true;
967 }
968 set_result(make_string_method_node(Op_StrComp, receiver, arg));
969 return true;
970 }
971
972 //------------------------------inline_string_equals------------------------
973 bool LibraryCallKit::inline_string_equals() {
974 Node* receiver = null_check_receiver();
975 // NOTE: Do not null check argument for String.equals() because spec
976 // allows to specify NULL as argument.
977 Node* argument = this->argument(1);
978 if (stopped()) {
979 return true;
980 }
981
982 // paths (plus control) merge
983 RegionNode* region = new RegionNode(5);
984 Node* phi = new PhiNode(region, TypeInt::BOOL);
985
986 // does source == target string?
987 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
988 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
989
990 Node* if_eq = generate_slow_guard(bol, NULL);
991 if (if_eq != NULL) {
992 // receiver == argument
993 phi->init_req(2, intcon(1));
994 region->init_req(2, if_eq);
995 }
996
997 // get String klass for instanceOf
998 ciInstanceKlass* klass = env()->String_klass();
999
1000 if (!stopped()) {
1001 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1002 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1003 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1004
1005 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1006 //instanceOf == true, fallthrough
1007
1008 if (inst_false != NULL) {
1009 phi->init_req(3, intcon(0));
1010 region->init_req(3, inst_false);
1011 }
1012 }
1013
1014 if (!stopped()) {
1015 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1016
1017 // Properly cast the argument to String
1018 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1019 // This path is taken only when argument's type is String:NotNull.
1020 argument = cast_not_null(argument, false);
1021
1022 Node* no_ctrl = NULL;
1023
1024 // Get start addr of receiver
1025 Node* receiver_val = load_String_value(no_ctrl, receiver);
1026 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1027 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1028
1029 // Get length of receiver
1030 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1031
1032 // Get start addr of argument
1033 Node* argument_val = load_String_value(no_ctrl, argument);
1034 Node* argument_offset = load_String_offset(no_ctrl, argument);
1035 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1036
1037 // Get length of argument
1038 Node* argument_cnt = load_String_length(no_ctrl, argument);
1039
1040 // Check for receiver count != argument count
1041 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1042 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1043 Node* if_ne = generate_slow_guard(bol, NULL);
1044 if (if_ne != NULL) {
1045 phi->init_req(4, intcon(0));
1046 region->init_req(4, if_ne);
1047 }
1048
1049 // Check for count == 0 is done by assembler code for StrEquals.
1050
1051 if (!stopped()) {
1052 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1053 phi->init_req(1, equals);
1054 region->init_req(1, control());
1055 }
1056 }
1057
1058 // post merge
1059 set_control(_gvn.transform(region));
1060 record_for_igvn(region);
1061
1062 set_result(_gvn.transform(phi));
1063 return true;
1064 }
1065
1066 //------------------------------inline_array_equals----------------------------
1067 bool LibraryCallKit::inline_array_equals() {
1068 Node* arg1 = argument(0);
1069 Node* arg2 = argument(1);
1070 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1071 return true;
1072 }
1073
1074 // Java version of String.indexOf(constant string)
1075 // class StringDecl {
1076 // StringDecl(char[] ca) {
1077 // offset = 0;
1078 // count = ca.length;
1079 // value = ca;
1080 // }
1081 // int offset;
1082 // int count;
1083 // char[] value;
1084 // }
1085 //
1086 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1087 // int targetOffset, int cache_i, int md2) {
1088 // int cache = cache_i;
1089 // int sourceOffset = string_object.offset;
1090 // int sourceCount = string_object.count;
1091 // int targetCount = target_object.length;
1092 //
1093 // int targetCountLess1 = targetCount - 1;
1094 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1095 //
1096 // char[] source = string_object.value;
1097 // char[] target = target_object;
1098 // int lastChar = target[targetCountLess1];
1099 //
1100 // outer_loop:
1101 // for (int i = sourceOffset; i < sourceEnd; ) {
1102 // int src = source[i + targetCountLess1];
1103 // if (src == lastChar) {
1104 // // With random strings and a 4-character alphabet,
1105 // // reverse matching at this point sets up 0.8% fewer
1106 // // frames, but (paradoxically) makes 0.3% more probes.
1107 // // Since those probes are nearer the lastChar probe,
1108 // // there is may be a net D$ win with reverse matching.
1109 // // But, reversing loop inhibits unroll of inner loop
1110 // // for unknown reason. So, does running outer loop from
1111 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1112 // for (int j = 0; j < targetCountLess1; j++) {
1113 // if (target[targetOffset + j] != source[i+j]) {
1114 // if ((cache & (1 << source[i+j])) == 0) {
1115 // if (md2 < j+1) {
1116 // i += j+1;
1117 // continue outer_loop;
1118 // }
1119 // }
1120 // i += md2;
1121 // continue outer_loop;
1122 // }
1123 // }
1124 // return i - sourceOffset;
1125 // }
1126 // if ((cache & (1 << src)) == 0) {
1127 // i += targetCountLess1;
1128 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1129 // i++;
1130 // }
1131 // return -1;
1132 // }
1133
1134 //------------------------------string_indexOf------------------------
1135 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1136 jint cache_i, jint md2_i) {
1137
1138 Node* no_ctrl = NULL;
1139 float likely = PROB_LIKELY(0.9);
1140 float unlikely = PROB_UNLIKELY(0.9);
1141
1142 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1143
1144 Node* source = load_String_value(no_ctrl, string_object);
1145 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1146 Node* sourceCount = load_String_length(no_ctrl, string_object);
1147
1148 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1149 jint target_length = target_array->length();
1150 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1151 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1152
1153 // String.value field is known to be @Stable.
1154 if (UseImplicitStableValues) {
1155 target = cast_array_to_stable(target, target_type);
1156 }
1157
1158 IdealKit kit(this, false, true);
1159 #define __ kit.
1160 Node* zero = __ ConI(0);
1161 Node* one = __ ConI(1);
1162 Node* cache = __ ConI(cache_i);
1163 Node* md2 = __ ConI(md2_i);
1164 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1165 Node* targetCountLess1 = __ ConI(target_length - 1);
1166 Node* targetOffset = __ ConI(targetOffset_i);
1167 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1168
1169 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1170 Node* outer_loop = __ make_label(2 /* goto */);
1171 Node* return_ = __ make_label(1);
1172
1173 __ set(rtn,__ ConI(-1));
1174 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1175 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1176 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1177 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1178 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1179 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1180 Node* tpj = __ AddI(targetOffset, __ value(j));
1181 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1182 Node* ipj = __ AddI(__ value(i), __ value(j));
1183 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1184 __ if_then(targ, BoolTest::ne, src2); {
1185 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1186 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1187 __ increment(i, __ AddI(__ value(j), one));
1188 __ goto_(outer_loop);
1189 } __ end_if(); __ dead(j);
1190 }__ end_if(); __ dead(j);
1191 __ increment(i, md2);
1192 __ goto_(outer_loop);
1193 }__ end_if();
1194 __ increment(j, one);
1195 }__ end_loop(); __ dead(j);
1196 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1197 __ goto_(return_);
1198 }__ end_if();
1199 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1200 __ increment(i, targetCountLess1);
1201 }__ end_if();
1202 __ increment(i, one);
1203 __ bind(outer_loop);
1204 }__ end_loop(); __ dead(i);
1205 __ bind(return_);
1206
1207 // Final sync IdealKit and GraphKit.
1208 final_sync(kit);
1209 Node* result = __ value(rtn);
1210 #undef __
1211 C->set_has_loops(true);
1212 return result;
1213 }
1214
1215 //------------------------------inline_string_indexOf------------------------
1216 bool LibraryCallKit::inline_string_indexOf() {
1217 Node* receiver = argument(0);
1218 Node* arg = argument(1);
1219
1220 Node* result;
1221 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1222 UseSSE42Intrinsics) {
1223 // Generate SSE4.2 version of indexOf
1224 // We currently only have match rules that use SSE4.2
1225
1226 receiver = null_check(receiver);
1227 arg = null_check(arg);
1228 if (stopped()) {
1229 return true;
1230 }
1231
1232 // Make the merge point
1233 RegionNode* result_rgn = new RegionNode(4);
1234 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1235 Node* no_ctrl = NULL;
1236
1237 // Get start addr of source string
1238 Node* source = load_String_value(no_ctrl, receiver);
1239 Node* source_offset = load_String_offset(no_ctrl, receiver);
1240 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1241
1242 // Get length of source string
1243 Node* source_cnt = load_String_length(no_ctrl, receiver);
1244
1245 // Get start addr of substring
1246 Node* substr = load_String_value(no_ctrl, arg);
1247 Node* substr_offset = load_String_offset(no_ctrl, arg);
1248 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1249
1250 // Get length of source string
1251 Node* substr_cnt = load_String_length(no_ctrl, arg);
1252
1253 // Check for substr count > string count
1254 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1255 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1256 Node* if_gt = generate_slow_guard(bol, NULL);
1257 if (if_gt != NULL) {
1258 result_phi->init_req(2, intcon(-1));
1259 result_rgn->init_req(2, if_gt);
1260 }
1261
1262 if (!stopped()) {
1263 // Check for substr count == 0
1264 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1265 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1266 Node* if_zero = generate_slow_guard(bol, NULL);
1267 if (if_zero != NULL) {
1268 result_phi->init_req(3, intcon(0));
1269 result_rgn->init_req(3, if_zero);
1270 }
1271 }
1272
1273 if (!stopped()) {
1274 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1275 result_phi->init_req(1, result);
1276 result_rgn->init_req(1, control());
1277 }
1278 set_control(_gvn.transform(result_rgn));
1279 record_for_igvn(result_rgn);
1280 result = _gvn.transform(result_phi);
1281
1282 } else { // Use LibraryCallKit::string_indexOf
1283 // don't intrinsify if argument isn't a constant string.
1284 if (!arg->is_Con()) {
1285 return false;
1286 }
1287 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1288 if (str_type == NULL) {
1289 return false;
1290 }
1291 ciInstanceKlass* klass = env()->String_klass();
1292 ciObject* str_const = str_type->const_oop();
1293 if (str_const == NULL || str_const->klass() != klass) {
1294 return false;
1295 }
1296 ciInstance* str = str_const->as_instance();
1297 assert(str != NULL, "must be instance");
1298
1299 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1300 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1301
1302 int o;
1303 int c;
1304 if (java_lang_String::has_offset_field()) {
1305 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1306 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1307 } else {
1308 o = 0;
1309 c = pat->length();
1310 }
1311
1312 // constant strings have no offset and count == length which
1313 // simplifies the resulting code somewhat so lets optimize for that.
1314 if (o != 0 || c != pat->length()) {
1315 return false;
1316 }
1317
1318 receiver = null_check(receiver, T_OBJECT);
1319 // NOTE: No null check on the argument is needed since it's a constant String oop.
1320 if (stopped()) {
1321 return true;
1322 }
1323
1324 // The null string as a pattern always returns 0 (match at beginning of string)
1325 if (c == 0) {
1326 set_result(intcon(0));
1327 return true;
1328 }
1329
1330 // Generate default indexOf
1331 jchar lastChar = pat->char_at(o + (c - 1));
1332 int cache = 0;
1333 int i;
1334 for (i = 0; i < c - 1; i++) {
1335 assert(i < pat->length(), "out of range");
1336 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1337 }
1338
1339 int md2 = c;
1340 for (i = 0; i < c - 1; i++) {
1341 assert(i < pat->length(), "out of range");
1342 if (pat->char_at(o + i) == lastChar) {
1343 md2 = (c - 1) - i;
1344 }
1345 }
1346
1347 result = string_indexOf(receiver, pat, o, cache, md2);
1348 }
1349 set_result(result);
1350 return true;
1351 }
1352
1353 //--------------------------round_double_node--------------------------------
1354 // Round a double node if necessary.
1355 Node* LibraryCallKit::round_double_node(Node* n) {
1356 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1357 n = _gvn.transform(new RoundDoubleNode(0, n));
1358 return n;
1359 }
1360
1361 //------------------------------inline_math-----------------------------------
1362 // public static double Math.abs(double)
1363 // public static double Math.sqrt(double)
1364 // public static double Math.log(double)
1365 // public static double Math.log10(double)
1366 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1367 Node* arg = round_double_node(argument(0));
1368 Node* n;
1369 switch (id) {
1370 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1371 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1372 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1373 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1374 default: fatal_unexpected_iid(id); break;
1375 }
1376 set_result(_gvn.transform(n));
1377 return true;
1378 }
1379
1380 //------------------------------inline_trig----------------------------------
1381 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1382 // argument reduction which will turn into a fast/slow diamond.
1383 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1384 Node* arg = round_double_node(argument(0));
1385 Node* n = NULL;
1386
1387 switch (id) {
1388 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1389 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1390 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1391 default: fatal_unexpected_iid(id); break;
1392 }
1393 n = _gvn.transform(n);
1394
1395 // Rounding required? Check for argument reduction!
1396 if (Matcher::strict_fp_requires_explicit_rounding) {
1397 static const double pi_4 = 0.7853981633974483;
1398 static const double neg_pi_4 = -0.7853981633974483;
1399 // pi/2 in 80-bit extended precision
1400 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1401 // -pi/2 in 80-bit extended precision
1402 // 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};
1403 // Cutoff value for using this argument reduction technique
1404 //static const double pi_2_minus_epsilon = 1.564660403643354;
1405 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1406
1407 // Pseudocode for sin:
1408 // if (x <= Math.PI / 4.0) {
1409 // if (x >= -Math.PI / 4.0) return fsin(x);
1410 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1411 // } else {
1412 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1413 // }
1414 // return StrictMath.sin(x);
1415
1416 // Pseudocode for cos:
1417 // if (x <= Math.PI / 4.0) {
1418 // if (x >= -Math.PI / 4.0) return fcos(x);
1419 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1420 // } else {
1421 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1422 // }
1423 // return StrictMath.cos(x);
1424
1425 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1426 // requires a special machine instruction to load it. Instead we'll try
1427 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1428 // probably do the math inside the SIN encoding.
1429
1430 // Make the merge point
1431 RegionNode* r = new RegionNode(3);
1432 Node* phi = new PhiNode(r, Type::DOUBLE);
1433
1434 // Flatten arg so we need only 1 test
1435 Node *abs = _gvn.transform(new AbsDNode(arg));
1436 // Node for PI/4 constant
1437 Node *pi4 = makecon(TypeD::make(pi_4));
1438 // Check PI/4 : abs(arg)
1439 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1440 // Check: If PI/4 < abs(arg) then go slow
1441 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1442 // Branch either way
1443 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1444 set_control(opt_iff(r,iff));
1445
1446 // Set fast path result
1447 phi->init_req(2, n);
1448
1449 // Slow path - non-blocking leaf call
1450 Node* call = NULL;
1451 switch (id) {
1452 case vmIntrinsics::_dsin:
1453 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1454 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1455 "Sin", NULL, arg, top());
1456 break;
1457 case vmIntrinsics::_dcos:
1458 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1459 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1460 "Cos", NULL, arg, top());
1461 break;
1462 case vmIntrinsics::_dtan:
1463 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1464 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1465 "Tan", NULL, arg, top());
1466 break;
1467 }
1468 assert(control()->in(0) == call, "");
1469 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1470 r->init_req(1, control());
1471 phi->init_req(1, slow_result);
1472
1473 // Post-merge
1474 set_control(_gvn.transform(r));
1475 record_for_igvn(r);
1476 n = _gvn.transform(phi);
1477
1478 C->set_has_split_ifs(true); // Has chance for split-if optimization
1479 }
1480 set_result(n);
1481 return true;
1482 }
1483
1484 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1485 //-------------------
1486 //result=(result.isNaN())? funcAddr():result;
1487 // Check: If isNaN() by checking result!=result? then either trap
1488 // or go to runtime
1489 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1490 // Build the boolean node
1491 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1492
1493 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1494 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1495 // The pow or exp intrinsic returned a NaN, which requires a call
1496 // to the runtime. Recompile with the runtime call.
1497 uncommon_trap(Deoptimization::Reason_intrinsic,
1498 Deoptimization::Action_make_not_entrant);
1499 }
1500 return result;
1501 } else {
1502 // If this inlining ever returned NaN in the past, we compile a call
1503 // to the runtime to properly handle corner cases
1504
1505 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1506 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1507 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1508
1509 if (!if_slow->is_top()) {
1510 RegionNode* result_region = new RegionNode(3);
1511 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1512
1513 result_region->init_req(1, if_fast);
1514 result_val->init_req(1, result);
1515
1516 set_control(if_slow);
1517
1518 const TypePtr* no_memory_effects = NULL;
1519 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1520 no_memory_effects,
1521 x, top(), y, y ? top() : NULL);
1522 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1523 #ifdef ASSERT
1524 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1525 assert(value_top == top(), "second value must be top");
1526 #endif
1527
1528 result_region->init_req(2, control());
1529 result_val->init_req(2, value);
1530 set_control(_gvn.transform(result_region));
1531 return _gvn.transform(result_val);
1532 } else {
1533 return result;
1534 }
1535 }
1536 }
1537
1538 //------------------------------inline_exp-------------------------------------
1539 // Inline exp instructions, if possible. The Intel hardware only misses
1540 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1541 bool LibraryCallKit::inline_exp() {
1542 Node* arg = round_double_node(argument(0));
1543 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1544
1545 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1546 set_result(n);
1547
1548 C->set_has_split_ifs(true); // Has chance for split-if optimization
1549 return true;
1550 }
1551
1552 //------------------------------inline_pow-------------------------------------
1553 // Inline power instructions, if possible.
1554 bool LibraryCallKit::inline_pow() {
1555 // Pseudocode for pow
1556 // if (y == 2) {
1557 // return x * x;
1558 // } else {
1559 // if (x <= 0.0) {
1560 // long longy = (long)y;
1561 // if ((double)longy == y) { // if y is long
1562 // if (y + 1 == y) longy = 0; // huge number: even
1563 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1564 // } else {
1565 // result = NaN;
1566 // }
1567 // } else {
1568 // result = DPow(x,y);
1569 // }
1570 // if (result != result)? {
1571 // result = uncommon_trap() or runtime_call();
1572 // }
1573 // return result;
1574 // }
1575
1576 Node* x = round_double_node(argument(0));
1577 Node* y = round_double_node(argument(2));
1578
1579 Node* result = NULL;
1580
1581 Node* const_two_node = makecon(TypeD::make(2.0));
1582 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1583 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1584 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1585 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1586 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1587
1588 RegionNode* region_node = new RegionNode(3);
1589 region_node->init_req(1, if_true);
1590
1591 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1592 // special case for x^y where y == 2, we can convert it to x * x
1593 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1594
1595 // set control to if_false since we will now process the false branch
1596 set_control(if_false);
1597
1598 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1599 // Short form: skip the fancy tests and just check for NaN result.
1600 result = _gvn.transform(new PowDNode(C, control(), x, y));
1601 } else {
1602 // If this inlining ever returned NaN in the past, include all
1603 // checks + call to the runtime.
1604
1605 // Set the merge point for If node with condition of (x <= 0.0)
1606 // There are four possible paths to region node and phi node
1607 RegionNode *r = new RegionNode(4);
1608 Node *phi = new PhiNode(r, Type::DOUBLE);
1609
1610 // Build the first if node: if (x <= 0.0)
1611 // Node for 0 constant
1612 Node *zeronode = makecon(TypeD::ZERO);
1613 // Check x:0
1614 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1615 // Check: If (x<=0) then go complex path
1616 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1617 // Branch either way
1618 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1619 // Fast path taken; set region slot 3
1620 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1621 r->init_req(3,fast_taken); // Capture fast-control
1622
1623 // Fast path not-taken, i.e. slow path
1624 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1625
1626 // Set fast path result
1627 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1628 phi->init_req(3, fast_result);
1629
1630 // Complex path
1631 // Build the second if node (if y is long)
1632 // Node for (long)y
1633 Node *longy = _gvn.transform(new ConvD2LNode(y));
1634 // Node for (double)((long) y)
1635 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1636 // Check (double)((long) y) : y
1637 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1638 // Check if (y isn't long) then go to slow path
1639
1640 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1641 // Branch either way
1642 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1643 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1644
1645 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1646
1647 // Calculate DPow(abs(x), y)*(1 & (long)y)
1648 // Node for constant 1
1649 Node *conone = longcon(1);
1650 // 1& (long)y
1651 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1652
1653 // A huge number is always even. Detect a huge number by checking
1654 // if y + 1 == y and set integer to be tested for parity to 0.
1655 // Required for corner case:
1656 // (long)9.223372036854776E18 = max_jlong
1657 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1658 // max_jlong is odd but 9.223372036854776E18 is even
1659 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1660 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1661 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1662 Node* correctedsign = NULL;
1663 if (ConditionalMoveLimit != 0) {
1664 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1665 } else {
1666 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1667 RegionNode *r = new RegionNode(3);
1668 Node *phi = new PhiNode(r, TypeLong::LONG);
1669 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1670 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1671 phi->init_req(1, signnode);
1672 phi->init_req(2, longcon(0));
1673 correctedsign = _gvn.transform(phi);
1674 ylong_path = _gvn.transform(r);
1675 record_for_igvn(r);
1676 }
1677
1678 // zero node
1679 Node *conzero = longcon(0);
1680 // Check (1&(long)y)==0?
1681 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1682 // Check if (1&(long)y)!=0?, if so the result is negative
1683 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1684 // abs(x)
1685 Node *absx=_gvn.transform(new AbsDNode(x));
1686 // abs(x)^y
1687 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1688 // -abs(x)^y
1689 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1690 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1691 Node *signresult = NULL;
1692 if (ConditionalMoveLimit != 0) {
1693 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1694 } else {
1695 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1696 RegionNode *r = new RegionNode(3);
1697 Node *phi = new PhiNode(r, Type::DOUBLE);
1698 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1699 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1700 phi->init_req(1, absxpowy);
1701 phi->init_req(2, negabsxpowy);
1702 signresult = _gvn.transform(phi);
1703 ylong_path = _gvn.transform(r);
1704 record_for_igvn(r);
1705 }
1706 // Set complex path fast result
1707 r->init_req(2, ylong_path);
1708 phi->init_req(2, signresult);
1709
1710 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1711 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1712 r->init_req(1,slow_path);
1713 phi->init_req(1,slow_result);
1714
1715 // Post merge
1716 set_control(_gvn.transform(r));
1717 record_for_igvn(r);
1718 result = _gvn.transform(phi);
1719 }
1720
1721 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1722
1723 // control from finish_pow_exp is now input to the region node
1724 region_node->set_req(2, control());
1725 // the result from finish_pow_exp is now input to the phi node
1726 phi_node->init_req(2, result);
1727 set_control(_gvn.transform(region_node));
1728 record_for_igvn(region_node);
1729 set_result(_gvn.transform(phi_node));
1730
1731 C->set_has_split_ifs(true); // Has chance for split-if optimization
1732 return true;
1733 }
1734
1735 //------------------------------runtime_math-----------------------------
1736 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1737 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1738 "must be (DD)D or (D)D type");
1739
1740 // Inputs
1741 Node* a = round_double_node(argument(0));
1742 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1743
1744 const TypePtr* no_memory_effects = NULL;
1745 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1746 no_memory_effects,
1747 a, top(), b, b ? top() : NULL);
1748 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1749 #ifdef ASSERT
1750 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1751 assert(value_top == top(), "second value must be top");
1752 #endif
1753
1754 set_result(value);
1755 return true;
1756 }
1757
1758 //------------------------------inline_math_native-----------------------------
1759 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1760 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1761 switch (id) {
1762 // These intrinsics are not properly supported on all hardware
1763 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1764 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1765 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1766 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1767 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1768 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1769
1770 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1771 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1772 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1773 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1774
1775 // These intrinsics are supported on all hardware
1776 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1777 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1778
1779 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1780 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1781 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1782 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1783 #undef FN_PTR
1784
1785 // These intrinsics are not yet correctly implemented
1786 case vmIntrinsics::_datan2:
1787 return false;
1788
1789 default:
1790 fatal_unexpected_iid(id);
1791 return false;
1792 }
1793 }
1794
1795 static bool is_simple_name(Node* n) {
1796 return (n->req() == 1 // constant
1797 || (n->is_Type() && n->as_Type()->type()->singleton())
1798 || n->is_Proj() // parameter or return value
1799 || n->is_Phi() // local of some sort
1800 );
1801 }
1802
1803 //----------------------------inline_notify-----------------------------------*
1804 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1805 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1806 address func;
1807 if (id == vmIntrinsics::_notify) {
1808 func = OptoRuntime::monitor_notify_Java();
1809 } else {
1810 func = OptoRuntime::monitor_notifyAll_Java();
1811 }
1812 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1813 make_slow_call_ex(call, env()->Throwable_klass(), false);
1814 return true;
1815 }
1816
1817
1818 //----------------------------inline_min_max-----------------------------------
1819 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1820 set_result(generate_min_max(id, argument(0), argument(1)));
1821 return true;
1822 }
1823
1824 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1825 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1826 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1827 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1828 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1829
1830 {
1831 PreserveJVMState pjvms(this);
1832 PreserveReexecuteState preexecs(this);
1833 jvms()->set_should_reexecute(true);
1834
1835 set_control(slow_path);
1836 set_i_o(i_o());
1837
1838 uncommon_trap(Deoptimization::Reason_intrinsic,
1839 Deoptimization::Action_none);
1840 }
1841
1842 set_control(fast_path);
1843 set_result(math);
1844 }
1845
1846 template <typename OverflowOp>
1847 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1848 typedef typename OverflowOp::MathOp MathOp;
1849
1850 MathOp* mathOp = new MathOp(arg1, arg2);
1851 Node* operation = _gvn.transform( mathOp );
1852 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1853 inline_math_mathExact(operation, ofcheck);
1854 return true;
1855 }
1856
1857 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1858 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1859 }
1860
1861 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1862 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1863 }
1864
1865 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1866 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1867 }
1868
1869 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1870 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1871 }
1872
1873 bool LibraryCallKit::inline_math_negateExactI() {
1874 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1875 }
1876
1877 bool LibraryCallKit::inline_math_negateExactL() {
1878 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1879 }
1880
1881 bool LibraryCallKit::inline_math_multiplyExactI() {
1882 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1883 }
1884
1885 bool LibraryCallKit::inline_math_multiplyExactL() {
1886 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1887 }
1888
1889 Node*
1890 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1891 // These are the candidate return value:
1892 Node* xvalue = x0;
1893 Node* yvalue = y0;
1894
1895 if (xvalue == yvalue) {
1896 return xvalue;
1897 }
1898
1899 bool want_max = (id == vmIntrinsics::_max);
1900
1901 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1902 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1903 if (txvalue == NULL || tyvalue == NULL) return top();
1904 // This is not really necessary, but it is consistent with a
1905 // hypothetical MaxINode::Value method:
1906 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1907
1908 // %%% This folding logic should (ideally) be in a different place.
1909 // Some should be inside IfNode, and there to be a more reliable
1910 // transformation of ?: style patterns into cmoves. We also want
1911 // more powerful optimizations around cmove and min/max.
1912
1913 // Try to find a dominating comparison of these guys.
1914 // It can simplify the index computation for Arrays.copyOf
1915 // and similar uses of System.arraycopy.
1916 // First, compute the normalized version of CmpI(x, y).
1917 int cmp_op = Op_CmpI;
1918 Node* xkey = xvalue;
1919 Node* ykey = yvalue;
1920 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1921 if (ideal_cmpxy->is_Cmp()) {
1922 // E.g., if we have CmpI(length - offset, count),
1923 // it might idealize to CmpI(length, count + offset)
1924 cmp_op = ideal_cmpxy->Opcode();
1925 xkey = ideal_cmpxy->in(1);
1926 ykey = ideal_cmpxy->in(2);
1927 }
1928
1929 // Start by locating any relevant comparisons.
1930 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1931 Node* cmpxy = NULL;
1932 Node* cmpyx = NULL;
1933 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1934 Node* cmp = start_from->fast_out(k);
1935 if (cmp->outcnt() > 0 && // must have prior uses
1936 cmp->in(0) == NULL && // must be context-independent
1937 cmp->Opcode() == cmp_op) { // right kind of compare
1938 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1939 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1940 }
1941 }
1942
1943 const int NCMPS = 2;
1944 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1945 int cmpn;
1946 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1947 if (cmps[cmpn] != NULL) break; // find a result
1948 }
1949 if (cmpn < NCMPS) {
1950 // Look for a dominating test that tells us the min and max.
1951 int depth = 0; // Limit search depth for speed
1952 Node* dom = control();
1953 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1954 if (++depth >= 100) break;
1955 Node* ifproj = dom;
1956 if (!ifproj->is_Proj()) continue;
1957 Node* iff = ifproj->in(0);
1958 if (!iff->is_If()) continue;
1959 Node* bol = iff->in(1);
1960 if (!bol->is_Bool()) continue;
1961 Node* cmp = bol->in(1);
1962 if (cmp == NULL) continue;
1963 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1964 if (cmps[cmpn] == cmp) break;
1965 if (cmpn == NCMPS) continue;
1966 BoolTest::mask btest = bol->as_Bool()->_test._test;
1967 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1968 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1969 // At this point, we know that 'x btest y' is true.
1970 switch (btest) {
1971 case BoolTest::eq:
1972 // They are proven equal, so we can collapse the min/max.
1973 // Either value is the answer. Choose the simpler.
1974 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1975 return yvalue;
1976 return xvalue;
1977 case BoolTest::lt: // x < y
1978 case BoolTest::le: // x <= y
1979 return (want_max ? yvalue : xvalue);
1980 case BoolTest::gt: // x > y
1981 case BoolTest::ge: // x >= y
1982 return (want_max ? xvalue : yvalue);
1983 }
1984 }
1985 }
1986
1987 // We failed to find a dominating test.
1988 // Let's pick a test that might GVN with prior tests.
1989 Node* best_bol = NULL;
1990 BoolTest::mask best_btest = BoolTest::illegal;
1991 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1992 Node* cmp = cmps[cmpn];
1993 if (cmp == NULL) continue;
1994 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1995 Node* bol = cmp->fast_out(j);
1996 if (!bol->is_Bool()) continue;
1997 BoolTest::mask btest = bol->as_Bool()->_test._test;
1998 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1999 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2000 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2001 best_bol = bol->as_Bool();
2002 best_btest = btest;
2003 }
2004 }
2005 }
2006
2007 Node* answer_if_true = NULL;
2008 Node* answer_if_false = NULL;
2009 switch (best_btest) {
2010 default:
2011 if (cmpxy == NULL)
2012 cmpxy = ideal_cmpxy;
2013 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2014 // and fall through:
2015 case BoolTest::lt: // x < y
2016 case BoolTest::le: // x <= y
2017 answer_if_true = (want_max ? yvalue : xvalue);
2018 answer_if_false = (want_max ? xvalue : yvalue);
2019 break;
2020 case BoolTest::gt: // x > y
2021 case BoolTest::ge: // x >= y
2022 answer_if_true = (want_max ? xvalue : yvalue);
2023 answer_if_false = (want_max ? yvalue : xvalue);
2024 break;
2025 }
2026
2027 jint hi, lo;
2028 if (want_max) {
2029 // We can sharpen the minimum.
2030 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2031 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2032 } else {
2033 // We can sharpen the maximum.
2034 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2035 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2036 }
2037
2038 // Use a flow-free graph structure, to avoid creating excess control edges
2039 // which could hinder other optimizations.
2040 // Since Math.min/max is often used with arraycopy, we want
2041 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2042 Node* cmov = CMoveNode::make(NULL, best_bol,
2043 answer_if_false, answer_if_true,
2044 TypeInt::make(lo, hi, widen));
2045
2046 return _gvn.transform(cmov);
2047
2048 /*
2049 // This is not as desirable as it may seem, since Min and Max
2050 // nodes do not have a full set of optimizations.
2051 // And they would interfere, anyway, with 'if' optimizations
2052 // and with CMoveI canonical forms.
2053 switch (id) {
2054 case vmIntrinsics::_min:
2055 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2056 case vmIntrinsics::_max:
2057 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2058 default:
2059 ShouldNotReachHere();
2060 }
2061 */
2062 }
2063
2064 inline int
2065 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2066 const TypePtr* base_type = TypePtr::NULL_PTR;
2067 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2068 if (base_type == NULL) {
2069 // Unknown type.
2070 return Type::AnyPtr;
2071 } else if (base_type == TypePtr::NULL_PTR) {
2072 // Since this is a NULL+long form, we have to switch to a rawptr.
2073 base = _gvn.transform(new CastX2PNode(offset));
2074 offset = MakeConX(0);
2075 return Type::RawPtr;
2076 } else if (base_type->base() == Type::RawPtr) {
2077 return Type::RawPtr;
2078 } else if (base_type->isa_oopptr()) {
2079 // Base is never null => always a heap address.
2080 if (base_type->ptr() == TypePtr::NotNull) {
2081 return Type::OopPtr;
2082 }
2083 // Offset is small => always a heap address.
2084 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2085 if (offset_type != NULL &&
2086 base_type->offset() == 0 && // (should always be?)
2087 offset_type->_lo >= 0 &&
2088 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2089 return Type::OopPtr;
2090 }
2091 // Otherwise, it might either be oop+off or NULL+addr.
2092 return Type::AnyPtr;
2093 } else {
2094 // No information:
2095 return Type::AnyPtr;
2096 }
2097 }
2098
2099 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2100 int kind = classify_unsafe_addr(base, offset);
2101 if (kind == Type::RawPtr) {
2102 return basic_plus_adr(top(), base, offset);
2103 } else {
2104 return basic_plus_adr(base, offset);
2105 }
2106 }
2107
2108 //--------------------------inline_number_methods-----------------------------
2109 // inline int Integer.numberOfLeadingZeros(int)
2110 // inline int Long.numberOfLeadingZeros(long)
2111 //
2112 // inline int Integer.numberOfTrailingZeros(int)
2113 // inline int Long.numberOfTrailingZeros(long)
2114 //
2115 // inline int Integer.bitCount(int)
2116 // inline int Long.bitCount(long)
2117 //
2118 // inline char Character.reverseBytes(char)
2119 // inline short Short.reverseBytes(short)
2120 // inline int Integer.reverseBytes(int)
2121 // inline long Long.reverseBytes(long)
2122 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2123 Node* arg = argument(0);
2124 Node* n;
2125 switch (id) {
2126 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2127 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2128 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2129 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2130 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2131 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2132 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2133 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2134 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2135 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2136 default: fatal_unexpected_iid(id); break;
2137 }
2138 set_result(_gvn.transform(n));
2139 return true;
2140 }
2141
2142 //----------------------------inline_unsafe_access----------------------------
2143
2144 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2145
2146 // Helper that guards and inserts a pre-barrier.
2147 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2148 Node* pre_val, bool need_mem_bar) {
2149 // We could be accessing the referent field of a reference object. If so, when G1
2150 // is enabled, we need to log the value in the referent field in an SATB buffer.
2151 // This routine performs some compile time filters and generates suitable
2152 // runtime filters that guard the pre-barrier code.
2153 // Also add memory barrier for non volatile load from the referent field
2154 // to prevent commoning of loads across safepoint.
2155 if (!UseG1GC && !need_mem_bar)
2156 return;
2157
2158 // Some compile time checks.
2159
2160 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2161 const TypeX* otype = offset->find_intptr_t_type();
2162 if (otype != NULL && otype->is_con() &&
2163 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2164 // Constant offset but not the reference_offset so just return
2165 return;
2166 }
2167
2168 // We only need to generate the runtime guards for instances.
2169 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2170 if (btype != NULL) {
2171 if (btype->isa_aryptr()) {
2172 // Array type so nothing to do
2173 return;
2174 }
2175
2176 const TypeInstPtr* itype = btype->isa_instptr();
2177 if (itype != NULL) {
2178 // Can the klass of base_oop be statically determined to be
2179 // _not_ a sub-class of Reference and _not_ Object?
2180 ciKlass* klass = itype->klass();
2181 if ( klass->is_loaded() &&
2182 !klass->is_subtype_of(env()->Reference_klass()) &&
2183 !env()->Object_klass()->is_subtype_of(klass)) {
2184 return;
2185 }
2186 }
2187 }
2188
2189 // The compile time filters did not reject base_oop/offset so
2190 // we need to generate the following runtime filters
2191 //
2192 // if (offset == java_lang_ref_Reference::_reference_offset) {
2193 // if (instance_of(base, java.lang.ref.Reference)) {
2194 // pre_barrier(_, pre_val, ...);
2195 // }
2196 // }
2197
2198 float likely = PROB_LIKELY( 0.999);
2199 float unlikely = PROB_UNLIKELY(0.999);
2200
2201 IdealKit ideal(this);
2202 #define __ ideal.
2203
2204 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2205
2206 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2207 // Update graphKit memory and control from IdealKit.
2208 sync_kit(ideal);
2209
2210 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2211 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2212
2213 // Update IdealKit memory and control from graphKit.
2214 __ sync_kit(this);
2215
2216 Node* one = __ ConI(1);
2217 // is_instof == 0 if base_oop == NULL
2218 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2219
2220 // Update graphKit from IdeakKit.
2221 sync_kit(ideal);
2222
2223 // Use the pre-barrier to record the value in the referent field
2224 pre_barrier(false /* do_load */,
2225 __ ctrl(),
2226 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2227 pre_val /* pre_val */,
2228 T_OBJECT);
2229 if (need_mem_bar) {
2230 // Add memory barrier to prevent commoning reads from this field
2231 // across safepoint since GC can change its value.
2232 insert_mem_bar(Op_MemBarCPUOrder);
2233 }
2234 // Update IdealKit from graphKit.
2235 __ sync_kit(this);
2236
2237 } __ end_if(); // _ref_type != ref_none
2238 } __ end_if(); // offset == referent_offset
2239
2240 // Final sync IdealKit and GraphKit.
2241 final_sync(ideal);
2242 #undef __
2243 }
2244
2245
2246 // Interpret Unsafe.fieldOffset cookies correctly:
2247 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2248
2249 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2250 // Attempt to infer a sharper value type from the offset and base type.
2251 ciKlass* sharpened_klass = NULL;
2252
2253 // See if it is an instance field, with an object type.
2254 if (alias_type->field() != NULL) {
2255 assert(!is_native_ptr, "native pointer op cannot use a java address");
2256 if (alias_type->field()->type()->is_klass()) {
2257 sharpened_klass = alias_type->field()->type()->as_klass();
2258 }
2259 }
2260
2261 // See if it is a narrow oop array.
2262 if (adr_type->isa_aryptr()) {
2263 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2264 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2265 if (elem_type != NULL) {
2266 sharpened_klass = elem_type->klass();
2267 }
2268 }
2269 }
2270
2271 // The sharpened class might be unloaded if there is no class loader
2272 // contraint in place.
2273 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2274 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2275
2276 #ifndef PRODUCT
2277 if (C->print_intrinsics() || C->print_inlining()) {
2278 tty->print(" from base type: "); adr_type->dump();
2279 tty->print(" sharpened value: "); tjp->dump();
2280 }
2281 #endif
2282 // Sharpen the value type.
2283 return tjp;
2284 }
2285 return NULL;
2286 }
2287
2288 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2289 if (callee()->is_static()) return false; // caller must have the capability!
2290
2291 #ifndef PRODUCT
2292 {
2293 ResourceMark rm;
2294 // Check the signatures.
2295 ciSignature* sig = callee()->signature();
2296 #ifdef ASSERT
2297 if (!is_store) {
2298 // Object getObject(Object base, int/long offset), etc.
2299 BasicType rtype = sig->return_type()->basic_type();
2300 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2301 rtype = T_ADDRESS; // it is really a C void*
2302 assert(rtype == type, "getter must return the expected value");
2303 if (!is_native_ptr) {
2304 assert(sig->count() == 2, "oop getter has 2 arguments");
2305 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2306 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2307 } else {
2308 assert(sig->count() == 1, "native getter has 1 argument");
2309 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2310 }
2311 } else {
2312 // void putObject(Object base, int/long offset, Object x), etc.
2313 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2314 if (!is_native_ptr) {
2315 assert(sig->count() == 3, "oop putter has 3 arguments");
2316 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2317 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2318 } else {
2319 assert(sig->count() == 2, "native putter has 2 arguments");
2320 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2321 }
2322 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2323 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2324 vtype = T_ADDRESS; // it is really a C void*
2325 assert(vtype == type, "putter must accept the expected value");
2326 }
2327 #endif // ASSERT
2328 }
2329 #endif //PRODUCT
2330
2331 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2332
2333 Node* receiver = argument(0); // type: oop
2334
2335 // Build address expression.
2336 Node* adr;
2337 Node* heap_base_oop = top();
2338 Node* offset = top();
2339 Node* val;
2340
2341 if (!is_native_ptr) {
2342 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2343 Node* base = argument(1); // type: oop
2344 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2345 offset = argument(2); // type: long
2346 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2347 // to be plain byte offsets, which are also the same as those accepted
2348 // by oopDesc::field_base.
2349 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2350 "fieldOffset must be byte-scaled");
2351 // 32-bit machines ignore the high half!
2352 offset = ConvL2X(offset);
2353 adr = make_unsafe_address(base, offset);
2354 heap_base_oop = base;
2355 val = is_store ? argument(4) : NULL;
2356 } else {
2357 Node* ptr = argument(1); // type: long
2358 ptr = ConvL2X(ptr); // adjust Java long to machine word
2359 adr = make_unsafe_address(NULL, ptr);
2360 val = is_store ? argument(3) : NULL;
2361 }
2362
2363 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2364
2365 // First guess at the value type.
2366 const Type *value_type = Type::get_const_basic_type(type);
2367
2368 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2369 // there was not enough information to nail it down.
2370 Compile::AliasType* alias_type = C->alias_type(adr_type);
2371 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2372
2373 // We will need memory barriers unless we can determine a unique
2374 // alias category for this reference. (Note: If for some reason
2375 // the barriers get omitted and the unsafe reference begins to "pollute"
2376 // the alias analysis of the rest of the graph, either Compile::can_alias
2377 // or Compile::must_alias will throw a diagnostic assert.)
2378 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2379
2380 // If we are reading the value of the referent field of a Reference
2381 // object (either by using Unsafe directly or through reflection)
2382 // then, if G1 is enabled, we need to record the referent in an
2383 // SATB log buffer using the pre-barrier mechanism.
2384 // Also we need to add memory barrier to prevent commoning reads
2385 // from this field across safepoint since GC can change its value.
2386 bool need_read_barrier = !is_native_ptr && !is_store &&
2387 offset != top() && heap_base_oop != top();
2388
2389 if (!is_store && type == T_OBJECT) {
2390 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2391 if (tjp != NULL) {
2392 value_type = tjp;
2393 }
2394 }
2395
2396 receiver = null_check(receiver);
2397 if (stopped()) {
2398 return true;
2399 }
2400 // Heap pointers get a null-check from the interpreter,
2401 // as a courtesy. However, this is not guaranteed by Unsafe,
2402 // and it is not possible to fully distinguish unintended nulls
2403 // from intended ones in this API.
2404
2405 if (is_volatile) {
2406 // We need to emit leading and trailing CPU membars (see below) in
2407 // addition to memory membars when is_volatile. This is a little
2408 // too strong, but avoids the need to insert per-alias-type
2409 // volatile membars (for stores; compare Parse::do_put_xxx), which
2410 // we cannot do effectively here because we probably only have a
2411 // rough approximation of type.
2412 need_mem_bar = true;
2413 // For Stores, place a memory ordering barrier now.
2414 if (is_store) {
2415 insert_mem_bar(Op_MemBarRelease);
2416 } else {
2417 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2418 insert_mem_bar(Op_MemBarVolatile);
2419 }
2420 }
2421 }
2422
2423 // Memory barrier to prevent normal and 'unsafe' accesses from
2424 // bypassing each other. Happens after null checks, so the
2425 // exception paths do not take memory state from the memory barrier,
2426 // so there's no problems making a strong assert about mixing users
2427 // of safe & unsafe memory.
2428 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2429
2430 if (!is_store) {
2431 Node* p = NULL;
2432 // Try to constant fold a load from a constant field
2433 ciField* field = alias_type->field();
2434 if (heap_base_oop != top() &&
2435 field != NULL && field->is_constant() && field->layout_type() == type) {
2436 // final or stable field
2437 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2438 if (con_type != NULL) {
2439 p = makecon(con_type);
2440 }
2441 }
2442 if (p == NULL) {
2443 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2444 // To be valid, unsafe loads may depend on other conditions than
2445 // the one that guards them: pin the Load node
2446 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2447 // load value
2448 switch (type) {
2449 case T_BOOLEAN:
2450 case T_CHAR:
2451 case T_BYTE:
2452 case T_SHORT:
2453 case T_INT:
2454 case T_LONG:
2455 case T_FLOAT:
2456 case T_DOUBLE:
2457 break;
2458 case T_OBJECT:
2459 if (need_read_barrier) {
2460 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2461 }
2462 break;
2463 case T_ADDRESS:
2464 // Cast to an int type.
2465 p = _gvn.transform(new CastP2XNode(NULL, p));
2466 p = ConvX2UL(p);
2467 break;
2468 default:
2469 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2470 break;
2471 }
2472 }
2473 // The load node has the control of the preceding MemBarCPUOrder. All
2474 // following nodes will have the control of the MemBarCPUOrder inserted at
2475 // the end of this method. So, pushing the load onto the stack at a later
2476 // point is fine.
2477 set_result(p);
2478 } else {
2479 // place effect of store into memory
2480 switch (type) {
2481 case T_DOUBLE:
2482 val = dstore_rounding(val);
2483 break;
2484 case T_ADDRESS:
2485 // Repackage the long as a pointer.
2486 val = ConvL2X(val);
2487 val = _gvn.transform(new CastX2PNode(val));
2488 break;
2489 }
2490
2491 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2492 if (type != T_OBJECT ) {
2493 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2494 } else {
2495 // Possibly an oop being stored to Java heap or native memory
2496 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2497 // oop to Java heap.
2498 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2499 } else {
2500 // We can't tell at compile time if we are storing in the Java heap or outside
2501 // of it. So we need to emit code to conditionally do the proper type of
2502 // store.
2503
2504 IdealKit ideal(this);
2505 #define __ ideal.
2506 // QQQ who knows what probability is here??
2507 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2508 // Sync IdealKit and graphKit.
2509 sync_kit(ideal);
2510 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2511 // Update IdealKit memory.
2512 __ sync_kit(this);
2513 } __ else_(); {
2514 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2515 } __ end_if();
2516 // Final sync IdealKit and GraphKit.
2517 final_sync(ideal);
2518 #undef __
2519 }
2520 }
2521 }
2522
2523 if (is_volatile) {
2524 if (!is_store) {
2525 insert_mem_bar(Op_MemBarAcquire);
2526 } else {
2527 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2528 insert_mem_bar(Op_MemBarVolatile);
2529 }
2530 }
2531 }
2532
2533 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2534
2535 return true;
2536 }
2537
2538 //----------------------------inline_unsafe_load_store----------------------------
2539 // This method serves a couple of different customers (depending on LoadStoreKind):
2540 //
2541 // LS_cmpxchg:
2542 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2543 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2544 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2545 //
2546 // LS_xadd:
2547 // public int getAndAddInt( Object o, long offset, int delta)
2548 // public long getAndAddLong(Object o, long offset, long delta)
2549 //
2550 // LS_xchg:
2551 // int getAndSet(Object o, long offset, int newValue)
2552 // long getAndSet(Object o, long offset, long newValue)
2553 // Object getAndSet(Object o, long offset, Object newValue)
2554 //
2555 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2556 // This basic scheme here is the same as inline_unsafe_access, but
2557 // differs in enough details that combining them would make the code
2558 // overly confusing. (This is a true fact! I originally combined
2559 // them, but even I was confused by it!) As much code/comments as
2560 // possible are retained from inline_unsafe_access though to make
2561 // the correspondences clearer. - dl
2562
2563 if (callee()->is_static()) return false; // caller must have the capability!
2564
2565 #ifndef PRODUCT
2566 BasicType rtype;
2567 {
2568 ResourceMark rm;
2569 // Check the signatures.
2570 ciSignature* sig = callee()->signature();
2571 rtype = sig->return_type()->basic_type();
2572 if (kind == LS_xadd || kind == LS_xchg) {
2573 // Check the signatures.
2574 #ifdef ASSERT
2575 assert(rtype == type, "get and set must return the expected type");
2576 assert(sig->count() == 3, "get and set has 3 arguments");
2577 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2578 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2579 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2580 #endif // ASSERT
2581 } else if (kind == LS_cmpxchg) {
2582 // Check the signatures.
2583 #ifdef ASSERT
2584 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2585 assert(sig->count() == 4, "CAS has 4 arguments");
2586 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2587 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2588 #endif // ASSERT
2589 } else {
2590 ShouldNotReachHere();
2591 }
2592 }
2593 #endif //PRODUCT
2594
2595 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2596
2597 // Get arguments:
2598 Node* receiver = NULL;
2599 Node* base = NULL;
2600 Node* offset = NULL;
2601 Node* oldval = NULL;
2602 Node* newval = NULL;
2603 if (kind == LS_cmpxchg) {
2604 const bool two_slot_type = type2size[type] == 2;
2605 receiver = argument(0); // type: oop
2606 base = argument(1); // type: oop
2607 offset = argument(2); // type: long
2608 oldval = argument(4); // type: oop, int, or long
2609 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2610 } else if (kind == LS_xadd || kind == LS_xchg){
2611 receiver = argument(0); // type: oop
2612 base = argument(1); // type: oop
2613 offset = argument(2); // type: long
2614 oldval = NULL;
2615 newval = argument(4); // type: oop, int, or long
2616 }
2617
2618 // Null check receiver.
2619 receiver = null_check(receiver);
2620 if (stopped()) {
2621 return true;
2622 }
2623
2624 // Build field offset expression.
2625 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2626 // to be plain byte offsets, which are also the same as those accepted
2627 // by oopDesc::field_base.
2628 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2629 // 32-bit machines ignore the high half of long offsets
2630 offset = ConvL2X(offset);
2631 Node* adr = make_unsafe_address(base, offset);
2632 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2633
2634 // For CAS, unlike inline_unsafe_access, there seems no point in
2635 // trying to refine types. Just use the coarse types here.
2636 const Type *value_type = Type::get_const_basic_type(type);
2637 Compile::AliasType* alias_type = C->alias_type(adr_type);
2638 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2639
2640 if (kind == LS_xchg && type == T_OBJECT) {
2641 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2642 if (tjp != NULL) {
2643 value_type = tjp;
2644 }
2645 }
2646
2647 int alias_idx = C->get_alias_index(adr_type);
2648
2649 // Memory-model-wise, a LoadStore acts like a little synchronized
2650 // block, so needs barriers on each side. These don't translate
2651 // into actual barriers on most machines, but we still need rest of
2652 // compiler to respect ordering.
2653
2654 insert_mem_bar(Op_MemBarRelease);
2655 insert_mem_bar(Op_MemBarCPUOrder);
2656
2657 // 4984716: MemBars must be inserted before this
2658 // memory node in order to avoid a false
2659 // dependency which will confuse the scheduler.
2660 Node *mem = memory(alias_idx);
2661
2662 // For now, we handle only those cases that actually exist: ints,
2663 // longs, and Object. Adding others should be straightforward.
2664 Node* load_store;
2665 switch(type) {
2666 case T_INT:
2667 if (kind == LS_xadd) {
2668 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2669 } else if (kind == LS_xchg) {
2670 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2671 } else if (kind == LS_cmpxchg) {
2672 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2673 } else {
2674 ShouldNotReachHere();
2675 }
2676 break;
2677 case T_LONG:
2678 if (kind == LS_xadd) {
2679 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2680 } else if (kind == LS_xchg) {
2681 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2682 } else if (kind == LS_cmpxchg) {
2683 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2684 } else {
2685 ShouldNotReachHere();
2686 }
2687 break;
2688 case T_OBJECT:
2689 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2690 // could be delayed during Parse (for example, in adjust_map_after_if()).
2691 // Execute transformation here to avoid barrier generation in such case.
2692 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2693 newval = _gvn.makecon(TypePtr::NULL_PTR);
2694
2695 // Reference stores need a store barrier.
2696 if (kind == LS_xchg) {
2697 // If pre-barrier must execute before the oop store, old value will require do_load here.
2698 if (!can_move_pre_barrier()) {
2699 pre_barrier(true /* do_load*/,
2700 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2701 NULL /* pre_val*/,
2702 T_OBJECT);
2703 } // Else move pre_barrier to use load_store value, see below.
2704 } else if (kind == LS_cmpxchg) {
2705 // Same as for newval above:
2706 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2707 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2708 }
2709 // The only known value which might get overwritten is oldval.
2710 pre_barrier(false /* do_load */,
2711 control(), NULL, NULL, max_juint, NULL, NULL,
2712 oldval /* pre_val */,
2713 T_OBJECT);
2714 } else {
2715 ShouldNotReachHere();
2716 }
2717
2718 #ifdef _LP64
2719 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2720 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2721 if (kind == LS_xchg) {
2722 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2723 newval_enc, adr_type, value_type->make_narrowoop()));
2724 } else {
2725 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2726 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2727 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2728 newval_enc, oldval_enc));
2729 }
2730 } else
2731 #endif
2732 {
2733 if (kind == LS_xchg) {
2734 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2735 } else {
2736 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2737 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2738 }
2739 }
2740 if (kind == LS_cmpxchg) {
2741 // Emit the post barrier only when the actual store happened.
2742 // This makes sense to check only for compareAndSet that can fail to set the value.
2743 // CAS success path is marked more likely since we anticipate this is a performance
2744 // critical path, while CAS failure path can use the penalty for going through unlikely
2745 // path as backoff. Which is still better than doing a store barrier there.
2746 IdealKit ideal(this);
2747 ideal.if_then(load_store, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2748 sync_kit(ideal);
2749 post_barrier(ideal.ctrl(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2750 ideal.sync_kit(this);
2751 } ideal.end_if();
2752 final_sync(ideal);
2753 } else {
2754 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2755 }
2756 break;
2757 default:
2758 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2759 break;
2760 }
2761
2762 // SCMemProjNodes represent the memory state of a LoadStore. Their
2763 // main role is to prevent LoadStore nodes from being optimized away
2764 // when their results aren't used.
2765 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2766 set_memory(proj, alias_idx);
2767
2768 if (type == T_OBJECT && kind == LS_xchg) {
2769 #ifdef _LP64
2770 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2771 load_store = _gvn.transform(new DecodeNNode(load_store, load_store->get_ptr_type()));
2772 }
2773 #endif
2774 if (can_move_pre_barrier()) {
2775 // Don't need to load pre_val. The old value is returned by load_store.
2776 // The pre_barrier can execute after the xchg as long as no safepoint
2777 // gets inserted between them.
2778 pre_barrier(false /* do_load */,
2779 control(), NULL, NULL, max_juint, NULL, NULL,
2780 load_store /* pre_val */,
2781 T_OBJECT);
2782 }
2783 }
2784
2785 // Add the trailing membar surrounding the access
2786 insert_mem_bar(Op_MemBarCPUOrder);
2787 insert_mem_bar(Op_MemBarAcquire);
2788
2789 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2790 set_result(load_store);
2791 return true;
2792 }
2793
2794 //----------------------------inline_unsafe_ordered_store----------------------
2795 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2796 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2797 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2798 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2799 // This is another variant of inline_unsafe_access, differing in
2800 // that it always issues store-store ("release") barrier and ensures
2801 // store-atomicity (which only matters for "long").
2802
2803 if (callee()->is_static()) return false; // caller must have the capability!
2804
2805 #ifndef PRODUCT
2806 {
2807 ResourceMark rm;
2808 // Check the signatures.
2809 ciSignature* sig = callee()->signature();
2810 #ifdef ASSERT
2811 BasicType rtype = sig->return_type()->basic_type();
2812 assert(rtype == T_VOID, "must return void");
2813 assert(sig->count() == 3, "has 3 arguments");
2814 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2815 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2816 #endif // ASSERT
2817 }
2818 #endif //PRODUCT
2819
2820 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2821
2822 // Get arguments:
2823 Node* receiver = argument(0); // type: oop
2824 Node* base = argument(1); // type: oop
2825 Node* offset = argument(2); // type: long
2826 Node* val = argument(4); // type: oop, int, or long
2827
2828 // Null check receiver.
2829 receiver = null_check(receiver);
2830 if (stopped()) {
2831 return true;
2832 }
2833
2834 // Build field offset expression.
2835 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2836 // 32-bit machines ignore the high half of long offsets
2837 offset = ConvL2X(offset);
2838 Node* adr = make_unsafe_address(base, offset);
2839 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2840 const Type *value_type = Type::get_const_basic_type(type);
2841 Compile::AliasType* alias_type = C->alias_type(adr_type);
2842
2843 insert_mem_bar(Op_MemBarRelease);
2844 insert_mem_bar(Op_MemBarCPUOrder);
2845 // Ensure that the store is atomic for longs:
2846 const bool require_atomic_access = true;
2847 Node* store;
2848 if (type == T_OBJECT) // reference stores need a store barrier.
2849 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2850 else {
2851 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2852 }
2853 insert_mem_bar(Op_MemBarCPUOrder);
2854 return true;
2855 }
2856
2857 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2858 // Regardless of form, don't allow previous ld/st to move down,
2859 // then issue acquire, release, or volatile mem_bar.
2860 insert_mem_bar(Op_MemBarCPUOrder);
2861 switch(id) {
2862 case vmIntrinsics::_loadFence:
2863 insert_mem_bar(Op_LoadFence);
2864 return true;
2865 case vmIntrinsics::_storeFence:
2866 insert_mem_bar(Op_StoreFence);
2867 return true;
2868 case vmIntrinsics::_fullFence:
2869 insert_mem_bar(Op_MemBarVolatile);
2870 return true;
2871 default:
2872 fatal_unexpected_iid(id);
2873 return false;
2874 }
2875 }
2876
2877 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2878 if (!kls->is_Con()) {
2879 return true;
2880 }
2881 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2882 if (klsptr == NULL) {
2883 return true;
2884 }
2885 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2886 // don't need a guard for a klass that is already initialized
2887 return !ik->is_initialized();
2888 }
2889
2890 //----------------------------inline_unsafe_allocate---------------------------
2891 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2892 bool LibraryCallKit::inline_unsafe_allocate() {
2893 if (callee()->is_static()) return false; // caller must have the capability!
2894
2895 null_check_receiver(); // null-check, then ignore
2896 Node* cls = null_check(argument(1));
2897 if (stopped()) return true;
2898
2899 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2900 kls = null_check(kls);
2901 if (stopped()) return true; // argument was like int.class
2902
2903 Node* test = NULL;
2904 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2905 // Note: The argument might still be an illegal value like
2906 // Serializable.class or Object[].class. The runtime will handle it.
2907 // But we must make an explicit check for initialization.
2908 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2909 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2910 // can generate code to load it as unsigned byte.
2911 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
2912 Node* bits = intcon(InstanceKlass::fully_initialized);
2913 test = _gvn.transform(new SubINode(inst, bits));
2914 // The 'test' is non-zero if we need to take a slow path.
2915 }
2916
2917 Node* obj = new_instance(kls, test);
2918 set_result(obj);
2919 return true;
2920 }
2921
2922 #ifdef TRACE_HAVE_INTRINSICS
2923 /*
2924 * oop -> myklass
2925 * myklass->trace_id |= USED
2926 * return myklass->trace_id & ~0x3
2927 */
2928 bool LibraryCallKit::inline_native_classID() {
2929 null_check_receiver(); // null-check, then ignore
2930 Node* cls = null_check(argument(1), T_OBJECT);
2931 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2932 kls = null_check(kls, T_OBJECT);
2933 ByteSize offset = TRACE_ID_OFFSET;
2934 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2935 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
2936 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2937 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
2938 Node* clsused = longcon(0x01l); // set the class bit
2939 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
2940
2941 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2942 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
2943 set_result(andl);
2944 return true;
2945 }
2946
2947 bool LibraryCallKit::inline_native_threadID() {
2948 Node* tls_ptr = NULL;
2949 Node* cur_thr = generate_current_thread(tls_ptr);
2950 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2951 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
2952 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2953
2954 Node* threadid = NULL;
2955 size_t thread_id_size = OSThread::thread_id_size();
2956 if (thread_id_size == (size_t) BytesPerLong) {
2957 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
2958 } else if (thread_id_size == (size_t) BytesPerInt) {
2959 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
2960 } else {
2961 ShouldNotReachHere();
2962 }
2963 set_result(threadid);
2964 return true;
2965 }
2966 #endif
2967
2968 //------------------------inline_native_time_funcs--------------
2969 // inline code for System.currentTimeMillis() and System.nanoTime()
2970 // these have the same type and signature
2971 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2972 const TypeFunc* tf = OptoRuntime::void_long_Type();
2973 const TypePtr* no_memory_effects = NULL;
2974 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2975 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
2976 #ifdef ASSERT
2977 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
2978 assert(value_top == top(), "second value must be top");
2979 #endif
2980 set_result(value);
2981 return true;
2982 }
2983
2984 //------------------------inline_native_currentThread------------------
2985 bool LibraryCallKit::inline_native_currentThread() {
2986 Node* junk = NULL;
2987 set_result(generate_current_thread(junk));
2988 return true;
2989 }
2990
2991 //------------------------inline_native_isInterrupted------------------
2992 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2993 bool LibraryCallKit::inline_native_isInterrupted() {
2994 // Add a fast path to t.isInterrupted(clear_int):
2995 // (t == Thread.current() &&
2996 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
2997 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2998 // So, in the common case that the interrupt bit is false,
2999 // we avoid making a call into the VM. Even if the interrupt bit
3000 // is true, if the clear_int argument is false, we avoid the VM call.
3001 // However, if the receiver is not currentThread, we must call the VM,
3002 // because there must be some locking done around the operation.
3003
3004 // We only go to the fast case code if we pass two guards.
3005 // Paths which do not pass are accumulated in the slow_region.
3006
3007 enum {
3008 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3009 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3010 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3011 PATH_LIMIT
3012 };
3013
3014 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3015 // out of the function.
3016 insert_mem_bar(Op_MemBarCPUOrder);
3017
3018 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3019 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3020
3021 RegionNode* slow_region = new RegionNode(1);
3022 record_for_igvn(slow_region);
3023
3024 // (a) Receiving thread must be the current thread.
3025 Node* rec_thr = argument(0);
3026 Node* tls_ptr = NULL;
3027 Node* cur_thr = generate_current_thread(tls_ptr);
3028 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3029 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3030
3031 generate_slow_guard(bol_thr, slow_region);
3032
3033 // (b) Interrupt bit on TLS must be false.
3034 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3035 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3036 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3037
3038 // Set the control input on the field _interrupted read to prevent it floating up.
3039 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3040 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3041 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3042
3043 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3044
3045 // First fast path: if (!TLS._interrupted) return false;
3046 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3047 result_rgn->init_req(no_int_result_path, false_bit);
3048 result_val->init_req(no_int_result_path, intcon(0));
3049
3050 // drop through to next case
3051 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3052
3053 #ifndef TARGET_OS_FAMILY_windows
3054 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3055 Node* clr_arg = argument(1);
3056 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3057 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3058 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3059
3060 // Second fast path: ... else if (!clear_int) return true;
3061 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3062 result_rgn->init_req(no_clear_result_path, false_arg);
3063 result_val->init_req(no_clear_result_path, intcon(1));
3064
3065 // drop through to next case
3066 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3067 #else
3068 // To return true on Windows you must read the _interrupted field
3069 // and check the the event state i.e. take the slow path.
3070 #endif // TARGET_OS_FAMILY_windows
3071
3072 // (d) Otherwise, go to the slow path.
3073 slow_region->add_req(control());
3074 set_control( _gvn.transform(slow_region));
3075
3076 if (stopped()) {
3077 // There is no slow path.
3078 result_rgn->init_req(slow_result_path, top());
3079 result_val->init_req(slow_result_path, top());
3080 } else {
3081 // non-virtual because it is a private non-static
3082 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3083
3084 Node* slow_val = set_results_for_java_call(slow_call);
3085 // this->control() comes from set_results_for_java_call
3086
3087 Node* fast_io = slow_call->in(TypeFunc::I_O);
3088 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3089
3090 // These two phis are pre-filled with copies of of the fast IO and Memory
3091 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3092 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3093
3094 result_rgn->init_req(slow_result_path, control());
3095 result_io ->init_req(slow_result_path, i_o());
3096 result_mem->init_req(slow_result_path, reset_memory());
3097 result_val->init_req(slow_result_path, slow_val);
3098
3099 set_all_memory(_gvn.transform(result_mem));
3100 set_i_o( _gvn.transform(result_io));
3101 }
3102
3103 C->set_has_split_ifs(true); // Has chance for split-if optimization
3104 set_result(result_rgn, result_val);
3105 return true;
3106 }
3107
3108 //---------------------------load_mirror_from_klass----------------------------
3109 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3110 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3111 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3112 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3113 }
3114
3115 //-----------------------load_klass_from_mirror_common-------------------------
3116 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3117 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3118 // and branch to the given path on the region.
3119 // If never_see_null, take an uncommon trap on null, so we can optimistically
3120 // compile for the non-null case.
3121 // If the region is NULL, force never_see_null = true.
3122 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3123 bool never_see_null,
3124 RegionNode* region,
3125 int null_path,
3126 int offset) {
3127 if (region == NULL) never_see_null = true;
3128 Node* p = basic_plus_adr(mirror, offset);
3129 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3130 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3131 Node* null_ctl = top();
3132 kls = null_check_oop(kls, &null_ctl, never_see_null);
3133 if (region != NULL) {
3134 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3135 region->init_req(null_path, null_ctl);
3136 } else {
3137 assert(null_ctl == top(), "no loose ends");
3138 }
3139 return kls;
3140 }
3141
3142 //--------------------(inline_native_Class_query helpers)---------------------
3143 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3144 // Fall through if (mods & mask) == bits, take the guard otherwise.
3145 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3146 // Branch around if the given klass has the given modifier bit set.
3147 // Like generate_guard, adds a new path onto the region.
3148 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3149 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3150 Node* mask = intcon(modifier_mask);
3151 Node* bits = intcon(modifier_bits);
3152 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3153 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3154 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3155 return generate_fair_guard(bol, region);
3156 }
3157 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3158 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3159 }
3160
3161 //-------------------------inline_native_Class_query-------------------
3162 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3163 const Type* return_type = TypeInt::BOOL;
3164 Node* prim_return_value = top(); // what happens if it's a primitive class?
3165 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3166 bool expect_prim = false; // most of these guys expect to work on refs
3167
3168 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3169
3170 Node* mirror = argument(0);
3171 Node* obj = top();
3172
3173 switch (id) {
3174 case vmIntrinsics::_isInstance:
3175 // nothing is an instance of a primitive type
3176 prim_return_value = intcon(0);
3177 obj = argument(1);
3178 break;
3179 case vmIntrinsics::_getModifiers:
3180 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3181 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3182 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3183 break;
3184 case vmIntrinsics::_isInterface:
3185 prim_return_value = intcon(0);
3186 break;
3187 case vmIntrinsics::_isArray:
3188 prim_return_value = intcon(0);
3189 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3190 break;
3191 case vmIntrinsics::_isPrimitive:
3192 prim_return_value = intcon(1);
3193 expect_prim = true; // obviously
3194 break;
3195 case vmIntrinsics::_getSuperclass:
3196 prim_return_value = null();
3197 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3198 break;
3199 case vmIntrinsics::_getClassAccessFlags:
3200 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3201 return_type = TypeInt::INT; // not bool! 6297094
3202 break;
3203 default:
3204 fatal_unexpected_iid(id);
3205 break;
3206 }
3207
3208 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3209 if (mirror_con == NULL) return false; // cannot happen?
3210
3211 #ifndef PRODUCT
3212 if (C->print_intrinsics() || C->print_inlining()) {
3213 ciType* k = mirror_con->java_mirror_type();
3214 if (k) {
3215 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3216 k->print_name();
3217 tty->cr();
3218 }
3219 }
3220 #endif
3221
3222 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3223 RegionNode* region = new RegionNode(PATH_LIMIT);
3224 record_for_igvn(region);
3225 PhiNode* phi = new PhiNode(region, return_type);
3226
3227 // The mirror will never be null of Reflection.getClassAccessFlags, however
3228 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3229 // if it is. See bug 4774291.
3230
3231 // For Reflection.getClassAccessFlags(), the null check occurs in
3232 // the wrong place; see inline_unsafe_access(), above, for a similar
3233 // situation.
3234 mirror = null_check(mirror);
3235 // If mirror or obj is dead, only null-path is taken.
3236 if (stopped()) return true;
3237
3238 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3239
3240 // Now load the mirror's klass metaobject, and null-check it.
3241 // Side-effects region with the control path if the klass is null.
3242 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3243 // If kls is null, we have a primitive mirror.
3244 phi->init_req(_prim_path, prim_return_value);
3245 if (stopped()) { set_result(region, phi); return true; }
3246 bool safe_for_replace = (region->in(_prim_path) == top());
3247
3248 Node* p; // handy temp
3249 Node* null_ctl;
3250
3251 // Now that we have the non-null klass, we can perform the real query.
3252 // For constant classes, the query will constant-fold in LoadNode::Value.
3253 Node* query_value = top();
3254 switch (id) {
3255 case vmIntrinsics::_isInstance:
3256 // nothing is an instance of a primitive type
3257 query_value = gen_instanceof(obj, kls, safe_for_replace);
3258 break;
3259
3260 case vmIntrinsics::_getModifiers:
3261 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3262 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3263 break;
3264
3265 case vmIntrinsics::_isInterface:
3266 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3267 if (generate_interface_guard(kls, region) != NULL)
3268 // A guard was added. If the guard is taken, it was an interface.
3269 phi->add_req(intcon(1));
3270 // If we fall through, it's a plain class.
3271 query_value = intcon(0);
3272 break;
3273
3274 case vmIntrinsics::_isArray:
3275 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3276 if (generate_array_guard(kls, region) != NULL)
3277 // A guard was added. If the guard is taken, it was an array.
3278 phi->add_req(intcon(1));
3279 // If we fall through, it's a plain class.
3280 query_value = intcon(0);
3281 break;
3282
3283 case vmIntrinsics::_isPrimitive:
3284 query_value = intcon(0); // "normal" path produces false
3285 break;
3286
3287 case vmIntrinsics::_getSuperclass:
3288 // The rules here are somewhat unfortunate, but we can still do better
3289 // with random logic than with a JNI call.
3290 // Interfaces store null or Object as _super, but must report null.
3291 // Arrays store an intermediate super as _super, but must report Object.
3292 // Other types can report the actual _super.
3293 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3294 if (generate_interface_guard(kls, region) != NULL)
3295 // A guard was added. If the guard is taken, it was an interface.
3296 phi->add_req(null());
3297 if (generate_array_guard(kls, region) != NULL)
3298 // A guard was added. If the guard is taken, it was an array.
3299 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3300 // If we fall through, it's a plain class. Get its _super.
3301 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3302 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3303 null_ctl = top();
3304 kls = null_check_oop(kls, &null_ctl);
3305 if (null_ctl != top()) {
3306 // If the guard is taken, Object.superClass is null (both klass and mirror).
3307 region->add_req(null_ctl);
3308 phi ->add_req(null());
3309 }
3310 if (!stopped()) {
3311 query_value = load_mirror_from_klass(kls);
3312 }
3313 break;
3314
3315 case vmIntrinsics::_getClassAccessFlags:
3316 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3317 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3318 break;
3319
3320 default:
3321 fatal_unexpected_iid(id);
3322 break;
3323 }
3324
3325 // Fall-through is the normal case of a query to a real class.
3326 phi->init_req(1, query_value);
3327 region->init_req(1, control());
3328
3329 C->set_has_split_ifs(true); // Has chance for split-if optimization
3330 set_result(region, phi);
3331 return true;
3332 }
3333
3334 //-------------------------inline_Class_cast-------------------
3335 bool LibraryCallKit::inline_Class_cast() {
3336 Node* mirror = argument(0); // Class
3337 Node* obj = argument(1);
3338 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3339 if (mirror_con == NULL) {
3340 return false; // dead path (mirror->is_top()).
3341 }
3342 if (obj == NULL || obj->is_top()) {
3343 return false; // dead path
3344 }
3345 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3346
3347 // First, see if Class.cast() can be folded statically.
3348 // java_mirror_type() returns non-null for compile-time Class constants.
3349 ciType* tm = mirror_con->java_mirror_type();
3350 if (tm != NULL && tm->is_klass() &&
3351 tp != NULL && tp->klass() != NULL) {
3352 if (!tp->klass()->is_loaded()) {
3353 // Don't use intrinsic when class is not loaded.
3354 return false;
3355 } else {
3356 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3357 if (static_res == Compile::SSC_always_true) {
3358 // isInstance() is true - fold the code.
3359 set_result(obj);
3360 return true;
3361 } else if (static_res == Compile::SSC_always_false) {
3362 // Don't use intrinsic, have to throw ClassCastException.
3363 // If the reference is null, the non-intrinsic bytecode will
3364 // be optimized appropriately.
3365 return false;
3366 }
3367 }
3368 }
3369
3370 // Bailout intrinsic and do normal inlining if exception path is frequent.
3371 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3372 return false;
3373 }
3374
3375 // Generate dynamic checks.
3376 // Class.cast() is java implementation of _checkcast bytecode.
3377 // Do checkcast (Parse::do_checkcast()) optimizations here.
3378
3379 mirror = null_check(mirror);
3380 // If mirror is dead, only null-path is taken.
3381 if (stopped()) {
3382 return true;
3383 }
3384
3385 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3386 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3387 RegionNode* region = new RegionNode(PATH_LIMIT);
3388 record_for_igvn(region);
3389
3390 // Now load the mirror's klass metaobject, and null-check it.
3391 // If kls is null, we have a primitive mirror and
3392 // nothing is an instance of a primitive type.
3393 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3394
3395 Node* res = top();
3396 if (!stopped()) {
3397 Node* bad_type_ctrl = top();
3398 // Do checkcast optimizations.
3399 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3400 region->init_req(_bad_type_path, bad_type_ctrl);
3401 }
3402 if (region->in(_prim_path) != top() ||
3403 region->in(_bad_type_path) != top()) {
3404 // Let Interpreter throw ClassCastException.
3405 PreserveJVMState pjvms(this);
3406 set_control(_gvn.transform(region));
3407 uncommon_trap(Deoptimization::Reason_intrinsic,
3408 Deoptimization::Action_maybe_recompile);
3409 }
3410 if (!stopped()) {
3411 set_result(res);
3412 }
3413 return true;
3414 }
3415
3416
3417 //--------------------------inline_native_subtype_check------------------------
3418 // This intrinsic takes the JNI calls out of the heart of
3419 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3420 bool LibraryCallKit::inline_native_subtype_check() {
3421 // Pull both arguments off the stack.
3422 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3423 args[0] = argument(0);
3424 args[1] = argument(1);
3425 Node* klasses[2]; // corresponding Klasses: superk, subk
3426 klasses[0] = klasses[1] = top();
3427
3428 enum {
3429 // A full decision tree on {superc is prim, subc is prim}:
3430 _prim_0_path = 1, // {P,N} => false
3431 // {P,P} & superc!=subc => false
3432 _prim_same_path, // {P,P} & superc==subc => true
3433 _prim_1_path, // {N,P} => false
3434 _ref_subtype_path, // {N,N} & subtype check wins => true
3435 _both_ref_path, // {N,N} & subtype check loses => false
3436 PATH_LIMIT
3437 };
3438
3439 RegionNode* region = new RegionNode(PATH_LIMIT);
3440 Node* phi = new PhiNode(region, TypeInt::BOOL);
3441 record_for_igvn(region);
3442
3443 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3444 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3445 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3446
3447 // First null-check both mirrors and load each mirror's klass metaobject.
3448 int which_arg;
3449 for (which_arg = 0; which_arg <= 1; which_arg++) {
3450 Node* arg = args[which_arg];
3451 arg = null_check(arg);
3452 if (stopped()) break;
3453 args[which_arg] = arg;
3454
3455 Node* p = basic_plus_adr(arg, class_klass_offset);
3456 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3457 klasses[which_arg] = _gvn.transform(kls);
3458 }
3459
3460 // Having loaded both klasses, test each for null.
3461 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3462 for (which_arg = 0; which_arg <= 1; which_arg++) {
3463 Node* kls = klasses[which_arg];
3464 Node* null_ctl = top();
3465 kls = null_check_oop(kls, &null_ctl, never_see_null);
3466 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3467 region->init_req(prim_path, null_ctl);
3468 if (stopped()) break;
3469 klasses[which_arg] = kls;
3470 }
3471
3472 if (!stopped()) {
3473 // now we have two reference types, in klasses[0..1]
3474 Node* subk = klasses[1]; // the argument to isAssignableFrom
3475 Node* superk = klasses[0]; // the receiver
3476 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3477 // now we have a successful reference subtype check
3478 region->set_req(_ref_subtype_path, control());
3479 }
3480
3481 // If both operands are primitive (both klasses null), then
3482 // we must return true when they are identical primitives.
3483 // It is convenient to test this after the first null klass check.
3484 set_control(region->in(_prim_0_path)); // go back to first null check
3485 if (!stopped()) {
3486 // Since superc is primitive, make a guard for the superc==subc case.
3487 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3488 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3489 generate_guard(bol_eq, region, PROB_FAIR);
3490 if (region->req() == PATH_LIMIT+1) {
3491 // A guard was added. If the added guard is taken, superc==subc.
3492 region->swap_edges(PATH_LIMIT, _prim_same_path);
3493 region->del_req(PATH_LIMIT);
3494 }
3495 region->set_req(_prim_0_path, control()); // Not equal after all.
3496 }
3497
3498 // these are the only paths that produce 'true':
3499 phi->set_req(_prim_same_path, intcon(1));
3500 phi->set_req(_ref_subtype_path, intcon(1));
3501
3502 // pull together the cases:
3503 assert(region->req() == PATH_LIMIT, "sane region");
3504 for (uint i = 1; i < region->req(); i++) {
3505 Node* ctl = region->in(i);
3506 if (ctl == NULL || ctl == top()) {
3507 region->set_req(i, top());
3508 phi ->set_req(i, top());
3509 } else if (phi->in(i) == NULL) {
3510 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3511 }
3512 }
3513
3514 set_control(_gvn.transform(region));
3515 set_result(_gvn.transform(phi));
3516 return true;
3517 }
3518
3519 //---------------------generate_array_guard_common------------------------
3520 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3521 bool obj_array, bool not_array) {
3522
3523 if (stopped()) {
3524 return NULL;
3525 }
3526
3527 // If obj_array/non_array==false/false:
3528 // Branch around if the given klass is in fact an array (either obj or prim).
3529 // If obj_array/non_array==false/true:
3530 // Branch around if the given klass is not an array klass of any kind.
3531 // If obj_array/non_array==true/true:
3532 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3533 // If obj_array/non_array==true/false:
3534 // Branch around if the kls is an oop array (Object[] or subtype)
3535 //
3536 // Like generate_guard, adds a new path onto the region.
3537 jint layout_con = 0;
3538 Node* layout_val = get_layout_helper(kls, layout_con);
3539 if (layout_val == NULL) {
3540 bool query = (obj_array
3541 ? Klass::layout_helper_is_objArray(layout_con)
3542 : Klass::layout_helper_is_array(layout_con));
3543 if (query == not_array) {
3544 return NULL; // never a branch
3545 } else { // always a branch
3546 Node* always_branch = control();
3547 if (region != NULL)
3548 region->add_req(always_branch);
3549 set_control(top());
3550 return always_branch;
3551 }
3552 }
3553 // Now test the correct condition.
3554 jint nval = (obj_array
3555 ? ((jint)Klass::_lh_array_tag_type_value
3556 << Klass::_lh_array_tag_shift)
3557 : Klass::_lh_neutral_value);
3558 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3559 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3560 // invert the test if we are looking for a non-array
3561 if (not_array) btest = BoolTest(btest).negate();
3562 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3563 return generate_fair_guard(bol, region);
3564 }
3565
3566
3567 //-----------------------inline_native_newArray--------------------------
3568 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3569 bool LibraryCallKit::inline_native_newArray() {
3570 Node* mirror = argument(0);
3571 Node* count_val = argument(1);
3572
3573 mirror = null_check(mirror);
3574 // If mirror or obj is dead, only null-path is taken.
3575 if (stopped()) return true;
3576
3577 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3578 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3579 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3580 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3581 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3582
3583 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3584 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3585 result_reg, _slow_path);
3586 Node* normal_ctl = control();
3587 Node* no_array_ctl = result_reg->in(_slow_path);
3588
3589 // Generate code for the slow case. We make a call to newArray().
3590 set_control(no_array_ctl);
3591 if (!stopped()) {
3592 // Either the input type is void.class, or else the
3593 // array klass has not yet been cached. Either the
3594 // ensuing call will throw an exception, or else it
3595 // will cache the array klass for next time.
3596 PreserveJVMState pjvms(this);
3597 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3598 Node* slow_result = set_results_for_java_call(slow_call);
3599 // this->control() comes from set_results_for_java_call
3600 result_reg->set_req(_slow_path, control());
3601 result_val->set_req(_slow_path, slow_result);
3602 result_io ->set_req(_slow_path, i_o());
3603 result_mem->set_req(_slow_path, reset_memory());
3604 }
3605
3606 set_control(normal_ctl);
3607 if (!stopped()) {
3608 // Normal case: The array type has been cached in the java.lang.Class.
3609 // The following call works fine even if the array type is polymorphic.
3610 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3611 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3612 result_reg->init_req(_normal_path, control());
3613 result_val->init_req(_normal_path, obj);
3614 result_io ->init_req(_normal_path, i_o());
3615 result_mem->init_req(_normal_path, reset_memory());
3616 }
3617
3618 // Return the combined state.
3619 set_i_o( _gvn.transform(result_io) );
3620 set_all_memory( _gvn.transform(result_mem));
3621
3622 C->set_has_split_ifs(true); // Has chance for split-if optimization
3623 set_result(result_reg, result_val);
3624 return true;
3625 }
3626
3627 //----------------------inline_native_getLength--------------------------
3628 // public static native int java.lang.reflect.Array.getLength(Object array);
3629 bool LibraryCallKit::inline_native_getLength() {
3630 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3631
3632 Node* array = null_check(argument(0));
3633 // If array is dead, only null-path is taken.
3634 if (stopped()) return true;
3635
3636 // Deoptimize if it is a non-array.
3637 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3638
3639 if (non_array != NULL) {
3640 PreserveJVMState pjvms(this);
3641 set_control(non_array);
3642 uncommon_trap(Deoptimization::Reason_intrinsic,
3643 Deoptimization::Action_maybe_recompile);
3644 }
3645
3646 // If control is dead, only non-array-path is taken.
3647 if (stopped()) return true;
3648
3649 // The works fine even if the array type is polymorphic.
3650 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3651 Node* result = load_array_length(array);
3652
3653 C->set_has_split_ifs(true); // Has chance for split-if optimization
3654 set_result(result);
3655 return true;
3656 }
3657
3658 //------------------------inline_array_copyOf----------------------------
3659 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3660 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3661 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3662 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3663
3664 // Get the arguments.
3665 Node* original = argument(0);
3666 Node* start = is_copyOfRange? argument(1): intcon(0);
3667 Node* end = is_copyOfRange? argument(2): argument(1);
3668 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3669
3670 Node* newcopy;
3671
3672 // Set the original stack and the reexecute bit for the interpreter to reexecute
3673 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3674 { PreserveReexecuteState preexecs(this);
3675 jvms()->set_should_reexecute(true);
3676
3677 array_type_mirror = null_check(array_type_mirror);
3678 original = null_check(original);
3679
3680 // Check if a null path was taken unconditionally.
3681 if (stopped()) return true;
3682
3683 Node* orig_length = load_array_length(original);
3684
3685 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3686 klass_node = null_check(klass_node);
3687
3688 RegionNode* bailout = new RegionNode(1);
3689 record_for_igvn(bailout);
3690
3691 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3692 // Bail out if that is so.
3693 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3694 if (not_objArray != NULL) {
3695 // Improve the klass node's type from the new optimistic assumption:
3696 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3697 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3698 Node* cast = new CastPPNode(klass_node, akls);
3699 cast->init_req(0, control());
3700 klass_node = _gvn.transform(cast);
3701 }
3702
3703 // Bail out if either start or end is negative.
3704 generate_negative_guard(start, bailout, &start);
3705 generate_negative_guard(end, bailout, &end);
3706
3707 Node* length = end;
3708 if (_gvn.type(start) != TypeInt::ZERO) {
3709 length = _gvn.transform(new SubINode(end, start));
3710 }
3711
3712 // Bail out if length is negative.
3713 // Without this the new_array would throw
3714 // NegativeArraySizeException but IllegalArgumentException is what
3715 // should be thrown
3716 generate_negative_guard(length, bailout, &length);
3717
3718 if (bailout->req() > 1) {
3719 PreserveJVMState pjvms(this);
3720 set_control(_gvn.transform(bailout));
3721 uncommon_trap(Deoptimization::Reason_intrinsic,
3722 Deoptimization::Action_maybe_recompile);
3723 }
3724
3725 if (!stopped()) {
3726 // How many elements will we copy from the original?
3727 // The answer is MinI(orig_length - start, length).
3728 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3729 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3730
3731 // Generate a direct call to the right arraycopy function(s).
3732 // We know the copy is disjoint but we might not know if the
3733 // oop stores need checking.
3734 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3735 // This will fail a store-check if x contains any non-nulls.
3736
3737 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3738 // loads/stores but it is legal only if we're sure the
3739 // Arrays.copyOf would succeed. So we need all input arguments
3740 // to the copyOf to be validated, including that the copy to the
3741 // new array won't trigger an ArrayStoreException. That subtype
3742 // check can be optimized if we know something on the type of
3743 // the input array from type speculation.
3744 if (_gvn.type(klass_node)->singleton()) {
3745 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3746 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3747
3748 int test = C->static_subtype_check(superk, subk);
3749 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3750 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3751 if (t_original->speculative_type() != NULL) {
3752 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3753 }
3754 }
3755 }
3756
3757 bool validated = false;
3758 // Reason_class_check rather than Reason_intrinsic because we
3759 // want to intrinsify even if this traps.
3760 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3761 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3762 klass_node);
3763
3764 if (not_subtype_ctrl != top()) {
3765 PreserveJVMState pjvms(this);
3766 set_control(not_subtype_ctrl);
3767 uncommon_trap(Deoptimization::Reason_class_check,
3768 Deoptimization::Action_make_not_entrant);
3769 assert(stopped(), "Should be stopped");
3770 }
3771 validated = true;
3772 }
3773
3774 if (!stopped()) {
3775 newcopy = new_array(klass_node, length, 0); // no arguments to push
3776
3777 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3778 load_object_klass(original), klass_node);
3779 if (!is_copyOfRange) {
3780 ac->set_copyof(validated);
3781 } else {
3782 ac->set_copyofrange(validated);
3783 }
3784 Node* n = _gvn.transform(ac);
3785 if (n == ac) {
3786 ac->connect_outputs(this);
3787 } else {
3788 assert(validated, "shouldn't transform if all arguments not validated");
3789 set_all_memory(n);
3790 }
3791 }
3792 }
3793 } // original reexecute is set back here
3794
3795 C->set_has_split_ifs(true); // Has chance for split-if optimization
3796 if (!stopped()) {
3797 set_result(newcopy);
3798 }
3799 return true;
3800 }
3801
3802
3803 //----------------------generate_virtual_guard---------------------------
3804 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3805 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3806 RegionNode* slow_region) {
3807 ciMethod* method = callee();
3808 int vtable_index = method->vtable_index();
3809 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3810 err_msg_res("bad index %d", vtable_index));
3811 // Get the Method* out of the appropriate vtable entry.
3812 int entry_offset = (InstanceKlass::vtable_start_offset() +
3813 vtable_index*vtableEntry::size()) * wordSize +
3814 vtableEntry::method_offset_in_bytes();
3815 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3816 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3817
3818 // Compare the target method with the expected method (e.g., Object.hashCode).
3819 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3820
3821 Node* native_call = makecon(native_call_addr);
3822 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3823 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3824
3825 return generate_slow_guard(test_native, slow_region);
3826 }
3827
3828 //-----------------------generate_method_call----------------------------
3829 // Use generate_method_call to make a slow-call to the real
3830 // method if the fast path fails. An alternative would be to
3831 // use a stub like OptoRuntime::slow_arraycopy_Java.
3832 // This only works for expanding the current library call,
3833 // not another intrinsic. (E.g., don't use this for making an
3834 // arraycopy call inside of the copyOf intrinsic.)
3835 CallJavaNode*
3836 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3837 // When compiling the intrinsic method itself, do not use this technique.
3838 guarantee(callee() != C->method(), "cannot make slow-call to self");
3839
3840 ciMethod* method = callee();
3841 // ensure the JVMS we have will be correct for this call
3842 guarantee(method_id == method->intrinsic_id(), "must match");
3843
3844 const TypeFunc* tf = TypeFunc::make(method);
3845 CallJavaNode* slow_call;
3846 if (is_static) {
3847 assert(!is_virtual, "");
3848 slow_call = new CallStaticJavaNode(C, tf,
3849 SharedRuntime::get_resolve_static_call_stub(),
3850 method, bci());
3851 } else if (is_virtual) {
3852 null_check_receiver();
3853 int vtable_index = Method::invalid_vtable_index;
3854 if (UseInlineCaches) {
3855 // Suppress the vtable call
3856 } else {
3857 // hashCode and clone are not a miranda methods,
3858 // so the vtable index is fixed.
3859 // No need to use the linkResolver to get it.
3860 vtable_index = method->vtable_index();
3861 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3862 err_msg_res("bad index %d", vtable_index));
3863 }
3864 slow_call = new CallDynamicJavaNode(tf,
3865 SharedRuntime::get_resolve_virtual_call_stub(),
3866 method, vtable_index, bci());
3867 } else { // neither virtual nor static: opt_virtual
3868 null_check_receiver();
3869 slow_call = new CallStaticJavaNode(C, tf,
3870 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3871 method, bci());
3872 slow_call->set_optimized_virtual(true);
3873 }
3874 set_arguments_for_java_call(slow_call);
3875 set_edges_for_java_call(slow_call);
3876 return slow_call;
3877 }
3878
3879
3880 /**
3881 * Build special case code for calls to hashCode on an object. This call may
3882 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3883 * slightly different code.
3884 */
3885 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3886 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3887 assert(!(is_virtual && is_static), "either virtual, special, or static");
3888
3889 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3890
3891 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3892 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
3893 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3894 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3895 Node* obj = NULL;
3896 if (!is_static) {
3897 // Check for hashing null object
3898 obj = null_check_receiver();
3899 if (stopped()) return true; // unconditionally null
3900 result_reg->init_req(_null_path, top());
3901 result_val->init_req(_null_path, top());
3902 } else {
3903 // Do a null check, and return zero if null.
3904 // System.identityHashCode(null) == 0
3905 obj = argument(0);
3906 Node* null_ctl = top();
3907 obj = null_check_oop(obj, &null_ctl);
3908 result_reg->init_req(_null_path, null_ctl);
3909 result_val->init_req(_null_path, _gvn.intcon(0));
3910 }
3911
3912 // Unconditionally null? Then return right away.
3913 if (stopped()) {
3914 set_control( result_reg->in(_null_path));
3915 if (!stopped())
3916 set_result(result_val->in(_null_path));
3917 return true;
3918 }
3919
3920 // We only go to the fast case code if we pass a number of guards. The
3921 // paths which do not pass are accumulated in the slow_region.
3922 RegionNode* slow_region = new RegionNode(1);
3923 record_for_igvn(slow_region);
3924
3925 // If this is a virtual call, we generate a funny guard. We pull out
3926 // the vtable entry corresponding to hashCode() from the target object.
3927 // If the target method which we are calling happens to be the native
3928 // Object hashCode() method, we pass the guard. We do not need this
3929 // guard for non-virtual calls -- the caller is known to be the native
3930 // Object hashCode().
3931 if (is_virtual) {
3932 // After null check, get the object's klass.
3933 Node* obj_klass = load_object_klass(obj);
3934 generate_virtual_guard(obj_klass, slow_region);
3935 }
3936
3937 // Get the header out of the object, use LoadMarkNode when available
3938 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3939 // The control of the load must be NULL. Otherwise, the load can move before
3940 // the null check after castPP removal.
3941 Node* no_ctrl = NULL;
3942 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3943
3944 // Test the header to see if it is unlocked.
3945 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3946 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
3947 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3948 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
3949 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
3950
3951 generate_slow_guard(test_unlocked, slow_region);
3952
3953 // Get the hash value and check to see that it has been properly assigned.
3954 // We depend on hash_mask being at most 32 bits and avoid the use of
3955 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3956 // vm: see markOop.hpp.
3957 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3958 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3959 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
3960 // This hack lets the hash bits live anywhere in the mark object now, as long
3961 // as the shift drops the relevant bits into the low 32 bits. Note that
3962 // Java spec says that HashCode is an int so there's no point in capturing
3963 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3964 hshifted_header = ConvX2I(hshifted_header);
3965 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
3966
3967 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3968 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
3969 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
3970
3971 generate_slow_guard(test_assigned, slow_region);
3972
3973 Node* init_mem = reset_memory();
3974 // fill in the rest of the null path:
3975 result_io ->init_req(_null_path, i_o());
3976 result_mem->init_req(_null_path, init_mem);
3977
3978 result_val->init_req(_fast_path, hash_val);
3979 result_reg->init_req(_fast_path, control());
3980 result_io ->init_req(_fast_path, i_o());
3981 result_mem->init_req(_fast_path, init_mem);
3982
3983 // Generate code for the slow case. We make a call to hashCode().
3984 set_control(_gvn.transform(slow_region));
3985 if (!stopped()) {
3986 // No need for PreserveJVMState, because we're using up the present state.
3987 set_all_memory(init_mem);
3988 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3989 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3990 Node* slow_result = set_results_for_java_call(slow_call);
3991 // this->control() comes from set_results_for_java_call
3992 result_reg->init_req(_slow_path, control());
3993 result_val->init_req(_slow_path, slow_result);
3994 result_io ->set_req(_slow_path, i_o());
3995 result_mem ->set_req(_slow_path, reset_memory());
3996 }
3997
3998 // Return the combined state.
3999 set_i_o( _gvn.transform(result_io) );
4000 set_all_memory( _gvn.transform(result_mem));
4001
4002 set_result(result_reg, result_val);
4003 return true;
4004 }
4005
4006 //---------------------------inline_native_getClass----------------------------
4007 // public final native Class<?> java.lang.Object.getClass();
4008 //
4009 // Build special case code for calls to getClass on an object.
4010 bool LibraryCallKit::inline_native_getClass() {
4011 Node* obj = null_check_receiver();
4012 if (stopped()) return true;
4013 set_result(load_mirror_from_klass(load_object_klass(obj)));
4014 return true;
4015 }
4016
4017 //-----------------inline_native_Reflection_getCallerClass---------------------
4018 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4019 //
4020 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4021 //
4022 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4023 // in that it must skip particular security frames and checks for
4024 // caller sensitive methods.
4025 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4026 #ifndef PRODUCT
4027 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4028 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4029 }
4030 #endif
4031
4032 if (!jvms()->has_method()) {
4033 #ifndef PRODUCT
4034 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4035 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4036 }
4037 #endif
4038 return false;
4039 }
4040
4041 // Walk back up the JVM state to find the caller at the required
4042 // depth.
4043 JVMState* caller_jvms = jvms();
4044
4045 // Cf. JVM_GetCallerClass
4046 // NOTE: Start the loop at depth 1 because the current JVM state does
4047 // not include the Reflection.getCallerClass() frame.
4048 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4049 ciMethod* m = caller_jvms->method();
4050 switch (n) {
4051 case 0:
4052 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4053 break;
4054 case 1:
4055 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4056 if (!m->caller_sensitive()) {
4057 #ifndef PRODUCT
4058 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4059 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4060 }
4061 #endif
4062 return false; // bail-out; let JVM_GetCallerClass do the work
4063 }
4064 break;
4065 default:
4066 if (!m->is_ignored_by_security_stack_walk()) {
4067 // We have reached the desired frame; return the holder class.
4068 // Acquire method holder as java.lang.Class and push as constant.
4069 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4070 ciInstance* caller_mirror = caller_klass->java_mirror();
4071 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4072
4073 #ifndef PRODUCT
4074 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4075 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());
4076 tty->print_cr(" JVM state at this point:");
4077 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4078 ciMethod* m = jvms()->of_depth(i)->method();
4079 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4080 }
4081 }
4082 #endif
4083 return true;
4084 }
4085 break;
4086 }
4087 }
4088
4089 #ifndef PRODUCT
4090 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4091 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4092 tty->print_cr(" JVM state at this point:");
4093 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4094 ciMethod* m = jvms()->of_depth(i)->method();
4095 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4096 }
4097 }
4098 #endif
4099
4100 return false; // bail-out; let JVM_GetCallerClass do the work
4101 }
4102
4103 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4104 Node* arg = argument(0);
4105 Node* result;
4106
4107 switch (id) {
4108 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4109 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4110 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4111 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4112
4113 case vmIntrinsics::_doubleToLongBits: {
4114 // two paths (plus control) merge in a wood
4115 RegionNode *r = new RegionNode(3);
4116 Node *phi = new PhiNode(r, TypeLong::LONG);
4117
4118 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4119 // Build the boolean node
4120 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4121
4122 // Branch either way.
4123 // NaN case is less traveled, which makes all the difference.
4124 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4125 Node *opt_isnan = _gvn.transform(ifisnan);
4126 assert( opt_isnan->is_If(), "Expect an IfNode");
4127 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4128 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4129
4130 set_control(iftrue);
4131
4132 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4133 Node *slow_result = longcon(nan_bits); // return NaN
4134 phi->init_req(1, _gvn.transform( slow_result ));
4135 r->init_req(1, iftrue);
4136
4137 // Else fall through
4138 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4139 set_control(iffalse);
4140
4141 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4142 r->init_req(2, iffalse);
4143
4144 // Post merge
4145 set_control(_gvn.transform(r));
4146 record_for_igvn(r);
4147
4148 C->set_has_split_ifs(true); // Has chance for split-if optimization
4149 result = phi;
4150 assert(result->bottom_type()->isa_long(), "must be");
4151 break;
4152 }
4153
4154 case vmIntrinsics::_floatToIntBits: {
4155 // two paths (plus control) merge in a wood
4156 RegionNode *r = new RegionNode(3);
4157 Node *phi = new PhiNode(r, TypeInt::INT);
4158
4159 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4160 // Build the boolean node
4161 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4162
4163 // Branch either way.
4164 // NaN case is less traveled, which makes all the difference.
4165 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4166 Node *opt_isnan = _gvn.transform(ifisnan);
4167 assert( opt_isnan->is_If(), "Expect an IfNode");
4168 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4169 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4170
4171 set_control(iftrue);
4172
4173 static const jint nan_bits = 0x7fc00000;
4174 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4175 phi->init_req(1, _gvn.transform( slow_result ));
4176 r->init_req(1, iftrue);
4177
4178 // Else fall through
4179 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4180 set_control(iffalse);
4181
4182 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4183 r->init_req(2, iffalse);
4184
4185 // Post merge
4186 set_control(_gvn.transform(r));
4187 record_for_igvn(r);
4188
4189 C->set_has_split_ifs(true); // Has chance for split-if optimization
4190 result = phi;
4191 assert(result->bottom_type()->isa_int(), "must be");
4192 break;
4193 }
4194
4195 default:
4196 fatal_unexpected_iid(id);
4197 break;
4198 }
4199 set_result(_gvn.transform(result));
4200 return true;
4201 }
4202
4203 #ifdef _LP64
4204 #define XTOP ,top() /*additional argument*/
4205 #else //_LP64
4206 #define XTOP /*no additional argument*/
4207 #endif //_LP64
4208
4209 //----------------------inline_unsafe_copyMemory-------------------------
4210 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4211 bool LibraryCallKit::inline_unsafe_copyMemory() {
4212 if (callee()->is_static()) return false; // caller must have the capability!
4213 null_check_receiver(); // null-check receiver
4214 if (stopped()) return true;
4215
4216 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4217
4218 Node* src_ptr = argument(1); // type: oop
4219 Node* src_off = ConvL2X(argument(2)); // type: long
4220 Node* dst_ptr = argument(4); // type: oop
4221 Node* dst_off = ConvL2X(argument(5)); // type: long
4222 Node* size = ConvL2X(argument(7)); // type: long
4223
4224 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4225 "fieldOffset must be byte-scaled");
4226
4227 Node* src = make_unsafe_address(src_ptr, src_off);
4228 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4229
4230 // Conservatively insert a memory barrier on all memory slices.
4231 // Do not let writes of the copy source or destination float below the copy.
4232 insert_mem_bar(Op_MemBarCPUOrder);
4233
4234 // Call it. Note that the length argument is not scaled.
4235 make_runtime_call(RC_LEAF|RC_NO_FP,
4236 OptoRuntime::fast_arraycopy_Type(),
4237 StubRoutines::unsafe_arraycopy(),
4238 "unsafe_arraycopy",
4239 TypeRawPtr::BOTTOM,
4240 src, dst, size XTOP);
4241
4242 // Do not let reads of the copy destination float above the copy.
4243 insert_mem_bar(Op_MemBarCPUOrder);
4244
4245 return true;
4246 }
4247
4248 //------------------------clone_coping-----------------------------------
4249 // Helper function for inline_native_clone.
4250 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4251 assert(obj_size != NULL, "");
4252 Node* raw_obj = alloc_obj->in(1);
4253 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4254
4255 AllocateNode* alloc = NULL;
4256 if (ReduceBulkZeroing) {
4257 // We will be completely responsible for initializing this object -
4258 // mark Initialize node as complete.
4259 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4260 // The object was just allocated - there should be no any stores!
4261 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4262 // Mark as complete_with_arraycopy so that on AllocateNode
4263 // expansion, we know this AllocateNode is initialized by an array
4264 // copy and a StoreStore barrier exists after the array copy.
4265 alloc->initialization()->set_complete_with_arraycopy();
4266 }
4267
4268 // Copy the fastest available way.
4269 // TODO: generate fields copies for small objects instead.
4270 Node* src = obj;
4271 Node* dest = alloc_obj;
4272 Node* size = _gvn.transform(obj_size);
4273
4274 // Exclude the header but include array length to copy by 8 bytes words.
4275 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4276 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4277 instanceOopDesc::base_offset_in_bytes();
4278 // base_off:
4279 // 8 - 32-bit VM
4280 // 12 - 64-bit VM, compressed klass
4281 // 16 - 64-bit VM, normal klass
4282 if (base_off % BytesPerLong != 0) {
4283 assert(UseCompressedClassPointers, "");
4284 if (is_array) {
4285 // Exclude length to copy by 8 bytes words.
4286 base_off += sizeof(int);
4287 } else {
4288 // Include klass to copy by 8 bytes words.
4289 base_off = instanceOopDesc::klass_offset_in_bytes();
4290 }
4291 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4292 }
4293 src = basic_plus_adr(src, base_off);
4294 dest = basic_plus_adr(dest, base_off);
4295
4296 // Compute the length also, if needed:
4297 Node* countx = size;
4298 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4299 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4300
4301 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4302
4303 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4304 ac->set_clonebasic();
4305 Node* n = _gvn.transform(ac);
4306 if (n == ac) {
4307 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4308 } else {
4309 set_all_memory(n);
4310 }
4311
4312 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4313 if (card_mark) {
4314 assert(!is_array, "");
4315 // Put in store barrier for any and all oops we are sticking
4316 // into this object. (We could avoid this if we could prove
4317 // that the object type contains no oop fields at all.)
4318 Node* no_particular_value = NULL;
4319 Node* no_particular_field = NULL;
4320 int raw_adr_idx = Compile::AliasIdxRaw;
4321 post_barrier(control(),
4322 memory(raw_adr_type),
4323 alloc_obj,
4324 no_particular_field,
4325 raw_adr_idx,
4326 no_particular_value,
4327 T_OBJECT,
4328 false);
4329 }
4330
4331 // Do not let reads from the cloned object float above the arraycopy.
4332 if (alloc != NULL) {
4333 // Do not let stores that initialize this object be reordered with
4334 // a subsequent store that would make this object accessible by
4335 // other threads.
4336 // Record what AllocateNode this StoreStore protects so that
4337 // escape analysis can go from the MemBarStoreStoreNode to the
4338 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4339 // based on the escape status of the AllocateNode.
4340 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4341 } else {
4342 insert_mem_bar(Op_MemBarCPUOrder);
4343 }
4344 }
4345
4346 //------------------------inline_native_clone----------------------------
4347 // protected native Object java.lang.Object.clone();
4348 //
4349 // Here are the simple edge cases:
4350 // null receiver => normal trap
4351 // virtual and clone was overridden => slow path to out-of-line clone
4352 // not cloneable or finalizer => slow path to out-of-line Object.clone
4353 //
4354 // The general case has two steps, allocation and copying.
4355 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4356 //
4357 // Copying also has two cases, oop arrays and everything else.
4358 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4359 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4360 //
4361 // These steps fold up nicely if and when the cloned object's klass
4362 // can be sharply typed as an object array, a type array, or an instance.
4363 //
4364 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4365 PhiNode* result_val;
4366
4367 // Set the reexecute bit for the interpreter to reexecute
4368 // the bytecode that invokes Object.clone if deoptimization happens.
4369 { PreserveReexecuteState preexecs(this);
4370 jvms()->set_should_reexecute(true);
4371
4372 Node* obj = null_check_receiver();
4373 if (stopped()) return true;
4374
4375 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4376
4377 // If we are going to clone an instance, we need its exact type to
4378 // know the number and types of fields to convert the clone to
4379 // loads/stores. Maybe a speculative type can help us.
4380 if (!obj_type->klass_is_exact() &&
4381 obj_type->speculative_type() != NULL &&
4382 obj_type->speculative_type()->is_instance_klass()) {
4383 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4384 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4385 !spec_ik->has_injected_fields()) {
4386 ciKlass* k = obj_type->klass();
4387 if (!k->is_instance_klass() ||
4388 k->as_instance_klass()->is_interface() ||
4389 k->as_instance_klass()->has_subklass()) {
4390 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4391 }
4392 }
4393 }
4394
4395 Node* obj_klass = load_object_klass(obj);
4396 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4397 const TypeOopPtr* toop = ((tklass != NULL)
4398 ? tklass->as_instance_type()
4399 : TypeInstPtr::NOTNULL);
4400
4401 // Conservatively insert a memory barrier on all memory slices.
4402 // Do not let writes into the original float below the clone.
4403 insert_mem_bar(Op_MemBarCPUOrder);
4404
4405 // paths into result_reg:
4406 enum {
4407 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4408 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4409 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4410 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4411 PATH_LIMIT
4412 };
4413 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4414 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4415 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4416 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4417 record_for_igvn(result_reg);
4418
4419 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4420 int raw_adr_idx = Compile::AliasIdxRaw;
4421
4422 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4423 if (array_ctl != NULL) {
4424 // It's an array.
4425 PreserveJVMState pjvms(this);
4426 set_control(array_ctl);
4427 Node* obj_length = load_array_length(obj);
4428 Node* obj_size = NULL;
4429 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4430
4431 if (!use_ReduceInitialCardMarks()) {
4432 // If it is an oop array, it requires very special treatment,
4433 // because card marking is required on each card of the array.
4434 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4435 if (is_obja != NULL) {
4436 PreserveJVMState pjvms2(this);
4437 set_control(is_obja);
4438 // Generate a direct call to the right arraycopy function(s).
4439 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4440 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4441 ac->set_cloneoop();
4442 Node* n = _gvn.transform(ac);
4443 assert(n == ac, "cannot disappear");
4444 ac->connect_outputs(this);
4445
4446 result_reg->init_req(_objArray_path, control());
4447 result_val->init_req(_objArray_path, alloc_obj);
4448 result_i_o ->set_req(_objArray_path, i_o());
4449 result_mem ->set_req(_objArray_path, reset_memory());
4450 }
4451 }
4452 // Otherwise, there are no card marks to worry about.
4453 // (We can dispense with card marks if we know the allocation
4454 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4455 // causes the non-eden paths to take compensating steps to
4456 // simulate a fresh allocation, so that no further
4457 // card marks are required in compiled code to initialize
4458 // the object.)
4459
4460 if (!stopped()) {
4461 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4462
4463 // Present the results of the copy.
4464 result_reg->init_req(_array_path, control());
4465 result_val->init_req(_array_path, alloc_obj);
4466 result_i_o ->set_req(_array_path, i_o());
4467 result_mem ->set_req(_array_path, reset_memory());
4468 }
4469 }
4470
4471 // We only go to the instance fast case code if we pass a number of guards.
4472 // The paths which do not pass are accumulated in the slow_region.
4473 RegionNode* slow_region = new RegionNode(1);
4474 record_for_igvn(slow_region);
4475 if (!stopped()) {
4476 // It's an instance (we did array above). Make the slow-path tests.
4477 // If this is a virtual call, we generate a funny guard. We grab
4478 // the vtable entry corresponding to clone() from the target object.
4479 // If the target method which we are calling happens to be the
4480 // Object clone() method, we pass the guard. We do not need this
4481 // guard for non-virtual calls; the caller is known to be the native
4482 // Object clone().
4483 if (is_virtual) {
4484 generate_virtual_guard(obj_klass, slow_region);
4485 }
4486
4487 // The object must be cloneable and must not have a finalizer.
4488 // Both of these conditions may be checked in a single test.
4489 // We could optimize the cloneable test further, but we don't care.
4490 generate_access_flags_guard(obj_klass,
4491 // Test both conditions:
4492 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4493 // Must be cloneable but not finalizer:
4494 JVM_ACC_IS_CLONEABLE,
4495 slow_region);
4496 }
4497
4498 if (!stopped()) {
4499 // It's an instance, and it passed the slow-path tests.
4500 PreserveJVMState pjvms(this);
4501 Node* obj_size = NULL;
4502 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4503 // is reexecuted if deoptimization occurs and there could be problems when merging
4504 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4505 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4506
4507 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4508
4509 // Present the results of the slow call.
4510 result_reg->init_req(_instance_path, control());
4511 result_val->init_req(_instance_path, alloc_obj);
4512 result_i_o ->set_req(_instance_path, i_o());
4513 result_mem ->set_req(_instance_path, reset_memory());
4514 }
4515
4516 // Generate code for the slow case. We make a call to clone().
4517 set_control(_gvn.transform(slow_region));
4518 if (!stopped()) {
4519 PreserveJVMState pjvms(this);
4520 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4521 Node* slow_result = set_results_for_java_call(slow_call);
4522 // this->control() comes from set_results_for_java_call
4523 result_reg->init_req(_slow_path, control());
4524 result_val->init_req(_slow_path, slow_result);
4525 result_i_o ->set_req(_slow_path, i_o());
4526 result_mem ->set_req(_slow_path, reset_memory());
4527 }
4528
4529 // Return the combined state.
4530 set_control( _gvn.transform(result_reg));
4531 set_i_o( _gvn.transform(result_i_o));
4532 set_all_memory( _gvn.transform(result_mem));
4533 } // original reexecute is set back here
4534
4535 set_result(_gvn.transform(result_val));
4536 return true;
4537 }
4538
4539 // If we have a tighly coupled allocation, the arraycopy may take care
4540 // of the array initialization. If one of the guards we insert between
4541 // the allocation and the arraycopy causes a deoptimization, an
4542 // unitialized array will escape the compiled method. To prevent that
4543 // we set the JVM state for uncommon traps between the allocation and
4544 // the arraycopy to the state before the allocation so, in case of
4545 // deoptimization, we'll reexecute the allocation and the
4546 // initialization.
4547 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4548 if (alloc != NULL) {
4549 ciMethod* trap_method = alloc->jvms()->method();
4550 int trap_bci = alloc->jvms()->bci();
4551
4552 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4553 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4554 // Make sure there's no store between the allocation and the
4555 // arraycopy otherwise visible side effects could be rexecuted
4556 // in case of deoptimization and cause incorrect execution.
4557 bool no_interfering_store = true;
4558 Node* mem = alloc->in(TypeFunc::Memory);
4559 if (mem->is_MergeMem()) {
4560 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4561 Node* n = mms.memory();
4562 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4563 assert(n->is_Store(), "what else?");
4564 no_interfering_store = false;
4565 break;
4566 }
4567 }
4568 } else {
4569 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4570 Node* n = mms.memory();
4571 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4572 assert(n->is_Store(), "what else?");
4573 no_interfering_store = false;
4574 break;
4575 }
4576 }
4577 }
4578
4579 if (no_interfering_store) {
4580 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4581 uint size = alloc->req();
4582 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4583 old_jvms->set_map(sfpt);
4584 for (uint i = 0; i < size; i++) {
4585 sfpt->init_req(i, alloc->in(i));
4586 }
4587 // re-push array length for deoptimization
4588 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4589 old_jvms->set_sp(old_jvms->sp()+1);
4590 old_jvms->set_monoff(old_jvms->monoff()+1);
4591 old_jvms->set_scloff(old_jvms->scloff()+1);
4592 old_jvms->set_endoff(old_jvms->endoff()+1);
4593 old_jvms->set_should_reexecute(true);
4594
4595 sfpt->set_i_o(map()->i_o());
4596 sfpt->set_memory(map()->memory());
4597 sfpt->set_control(map()->control());
4598
4599 JVMState* saved_jvms = jvms();
4600 saved_reexecute_sp = _reexecute_sp;
4601
4602 set_jvms(sfpt->jvms());
4603 _reexecute_sp = jvms()->sp();
4604
4605 return saved_jvms;
4606 }
4607 }
4608 }
4609 return NULL;
4610 }
4611
4612 // In case of a deoptimization, we restart execution at the
4613 // allocation, allocating a new array. We would leave an uninitialized
4614 // array in the heap that GCs wouldn't expect. Move the allocation
4615 // after the traps so we don't allocate the array if we
4616 // deoptimize. This is possible because tightly_coupled_allocation()
4617 // guarantees there's no observer of the allocated array at this point
4618 // and the control flow is simple enough.
4619 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4620 if (saved_jvms != NULL && !stopped()) {
4621 assert(alloc != NULL, "only with a tightly coupled allocation");
4622 // restore JVM state to the state at the arraycopy
4623 saved_jvms->map()->set_control(map()->control());
4624 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4625 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4626 // If we've improved the types of some nodes (null check) while
4627 // emitting the guards, propagate them to the current state
4628 map()->replaced_nodes().apply(saved_jvms->map());
4629 set_jvms(saved_jvms);
4630 _reexecute_sp = saved_reexecute_sp;
4631
4632 // Remove the allocation from above the guards
4633 CallProjections callprojs;
4634 alloc->extract_projections(&callprojs, true);
4635 InitializeNode* init = alloc->initialization();
4636 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4637 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4638 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4639 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4640
4641 // move the allocation here (after the guards)
4642 _gvn.hash_delete(alloc);
4643 alloc->set_req(TypeFunc::Control, control());
4644 alloc->set_req(TypeFunc::I_O, i_o());
4645 Node *mem = reset_memory();
4646 set_all_memory(mem);
4647 alloc->set_req(TypeFunc::Memory, mem);
4648 set_control(init->proj_out(TypeFunc::Control));
4649 set_i_o(callprojs.fallthrough_ioproj);
4650
4651 // Update memory as done in GraphKit::set_output_for_allocation()
4652 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4653 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4654 if (ary_type->isa_aryptr() && length_type != NULL) {
4655 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4656 }
4657 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4658 int elemidx = C->get_alias_index(telemref);
4659 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4660 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4661
4662 Node* allocx = _gvn.transform(alloc);
4663 assert(allocx == alloc, "where has the allocation gone?");
4664 assert(dest->is_CheckCastPP(), "not an allocation result?");
4665
4666 _gvn.hash_delete(dest);
4667 dest->set_req(0, control());
4668 Node* destx = _gvn.transform(dest);
4669 assert(destx == dest, "where has the allocation result gone?");
4670 }
4671 }
4672
4673
4674 //------------------------------inline_arraycopy-----------------------
4675 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4676 // Object dest, int destPos,
4677 // int length);
4678 bool LibraryCallKit::inline_arraycopy() {
4679 // Get the arguments.
4680 Node* src = argument(0); // type: oop
4681 Node* src_offset = argument(1); // type: int
4682 Node* dest = argument(2); // type: oop
4683 Node* dest_offset = argument(3); // type: int
4684 Node* length = argument(4); // type: int
4685
4686
4687 // Check for allocation before we add nodes that would confuse
4688 // tightly_coupled_allocation()
4689 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4690
4691 int saved_reexecute_sp = -1;
4692 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4693 // See arraycopy_restore_alloc_state() comment
4694 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4695 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4696 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4697 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4698
4699 // The following tests must be performed
4700 // (1) src and dest are arrays.
4701 // (2) src and dest arrays must have elements of the same BasicType
4702 // (3) src and dest must not be null.
4703 // (4) src_offset must not be negative.
4704 // (5) dest_offset must not be negative.
4705 // (6) length must not be negative.
4706 // (7) src_offset + length must not exceed length of src.
4707 // (8) dest_offset + length must not exceed length of dest.
4708 // (9) each element of an oop array must be assignable
4709
4710 // (3) src and dest must not be null.
4711 // always do this here because we need the JVM state for uncommon traps
4712 Node* null_ctl = top();
4713 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4714 assert(null_ctl->is_top(), "no null control here");
4715 dest = null_check(dest, T_ARRAY);
4716
4717 if (!can_emit_guards) {
4718 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4719 // guards but the arraycopy node could still take advantage of a
4720 // tightly allocated allocation. tightly_coupled_allocation() is
4721 // called again to make sure it takes the null check above into
4722 // account: the null check is mandatory and if it caused an
4723 // uncommon trap to be emitted then the allocation can't be
4724 // considered tightly coupled in this context.
4725 alloc = tightly_coupled_allocation(dest, NULL);
4726 }
4727
4728 bool validated = false;
4729
4730 const Type* src_type = _gvn.type(src);
4731 const Type* dest_type = _gvn.type(dest);
4732 const TypeAryPtr* top_src = src_type->isa_aryptr();
4733 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4734
4735 // Do we have the type of src?
4736 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4737 // Do we have the type of dest?
4738 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4739 // Is the type for src from speculation?
4740 bool src_spec = false;
4741 // Is the type for dest from speculation?
4742 bool dest_spec = false;
4743
4744 if ((!has_src || !has_dest) && can_emit_guards) {
4745 // We don't have sufficient type information, let's see if
4746 // speculative types can help. We need to have types for both src
4747 // and dest so that it pays off.
4748
4749 // Do we already have or could we have type information for src
4750 bool could_have_src = has_src;
4751 // Do we already have or could we have type information for dest
4752 bool could_have_dest = has_dest;
4753
4754 ciKlass* src_k = NULL;
4755 if (!has_src) {
4756 src_k = src_type->speculative_type_not_null();
4757 if (src_k != NULL && src_k->is_array_klass()) {
4758 could_have_src = true;
4759 }
4760 }
4761
4762 ciKlass* dest_k = NULL;
4763 if (!has_dest) {
4764 dest_k = dest_type->speculative_type_not_null();
4765 if (dest_k != NULL && dest_k->is_array_klass()) {
4766 could_have_dest = true;
4767 }
4768 }
4769
4770 if (could_have_src && could_have_dest) {
4771 // This is going to pay off so emit the required guards
4772 if (!has_src) {
4773 src = maybe_cast_profiled_obj(src, src_k, true);
4774 src_type = _gvn.type(src);
4775 top_src = src_type->isa_aryptr();
4776 has_src = (top_src != NULL && top_src->klass() != NULL);
4777 src_spec = true;
4778 }
4779 if (!has_dest) {
4780 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4781 dest_type = _gvn.type(dest);
4782 top_dest = dest_type->isa_aryptr();
4783 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4784 dest_spec = true;
4785 }
4786 }
4787 }
4788
4789 if (has_src && has_dest && can_emit_guards) {
4790 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4791 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4792 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4793 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4794
4795 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4796 // If both arrays are object arrays then having the exact types
4797 // for both will remove the need for a subtype check at runtime
4798 // before the call and may make it possible to pick a faster copy
4799 // routine (without a subtype check on every element)
4800 // Do we have the exact type of src?
4801 bool could_have_src = src_spec;
4802 // Do we have the exact type of dest?
4803 bool could_have_dest = dest_spec;
4804 ciKlass* src_k = top_src->klass();
4805 ciKlass* dest_k = top_dest->klass();
4806 if (!src_spec) {
4807 src_k = src_type->speculative_type_not_null();
4808 if (src_k != NULL && src_k->is_array_klass()) {
4809 could_have_src = true;
4810 }
4811 }
4812 if (!dest_spec) {
4813 dest_k = dest_type->speculative_type_not_null();
4814 if (dest_k != NULL && dest_k->is_array_klass()) {
4815 could_have_dest = true;
4816 }
4817 }
4818 if (could_have_src && could_have_dest) {
4819 // If we can have both exact types, emit the missing guards
4820 if (could_have_src && !src_spec) {
4821 src = maybe_cast_profiled_obj(src, src_k, true);
4822 }
4823 if (could_have_dest && !dest_spec) {
4824 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4825 }
4826 }
4827 }
4828 }
4829
4830 ciMethod* trap_method = method();
4831 int trap_bci = bci();
4832 if (saved_jvms != NULL) {
4833 trap_method = alloc->jvms()->method();
4834 trap_bci = alloc->jvms()->bci();
4835 }
4836
4837 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4838 can_emit_guards &&
4839 !src->is_top() && !dest->is_top()) {
4840 // validate arguments: enables transformation the ArrayCopyNode
4841 validated = true;
4842
4843 RegionNode* slow_region = new RegionNode(1);
4844 record_for_igvn(slow_region);
4845
4846 // (1) src and dest are arrays.
4847 generate_non_array_guard(load_object_klass(src), slow_region);
4848 generate_non_array_guard(load_object_klass(dest), slow_region);
4849
4850 // (2) src and dest arrays must have elements of the same BasicType
4851 // done at macro expansion or at Ideal transformation time
4852
4853 // (4) src_offset must not be negative.
4854 generate_negative_guard(src_offset, slow_region);
4855
4856 // (5) dest_offset must not be negative.
4857 generate_negative_guard(dest_offset, slow_region);
4858
4859 // (7) src_offset + length must not exceed length of src.
4860 generate_limit_guard(src_offset, length,
4861 load_array_length(src),
4862 slow_region);
4863
4864 // (8) dest_offset + length must not exceed length of dest.
4865 generate_limit_guard(dest_offset, length,
4866 load_array_length(dest),
4867 slow_region);
4868
4869 // (9) each element of an oop array must be assignable
4870 Node* src_klass = load_object_klass(src);
4871 Node* dest_klass = load_object_klass(dest);
4872 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4873
4874 if (not_subtype_ctrl != top()) {
4875 PreserveJVMState pjvms(this);
4876 set_control(not_subtype_ctrl);
4877 uncommon_trap(Deoptimization::Reason_intrinsic,
4878 Deoptimization::Action_make_not_entrant);
4879 assert(stopped(), "Should be stopped");
4880 }
4881 {
4882 PreserveJVMState pjvms(this);
4883 set_control(_gvn.transform(slow_region));
4884 uncommon_trap(Deoptimization::Reason_intrinsic,
4885 Deoptimization::Action_make_not_entrant);
4886 assert(stopped(), "Should be stopped");
4887 }
4888 }
4889
4890 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
4891
4892 if (stopped()) {
4893 return true;
4894 }
4895
4896 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
4897 // Create LoadRange and LoadKlass nodes for use during macro expansion here
4898 // so the compiler has a chance to eliminate them: during macro expansion,
4899 // we have to set their control (CastPP nodes are eliminated).
4900 load_object_klass(src), load_object_klass(dest),
4901 load_array_length(src), load_array_length(dest));
4902
4903 ac->set_arraycopy(validated);
4904
4905 Node* n = _gvn.transform(ac);
4906 if (n == ac) {
4907 ac->connect_outputs(this);
4908 } else {
4909 assert(validated, "shouldn't transform if all arguments not validated");
4910 set_all_memory(n);
4911 }
4912
4913 return true;
4914 }
4915
4916
4917 // Helper function which determines if an arraycopy immediately follows
4918 // an allocation, with no intervening tests or other escapes for the object.
4919 AllocateArrayNode*
4920 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4921 RegionNode* slow_region) {
4922 if (stopped()) return NULL; // no fast path
4923 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4924
4925 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4926 if (alloc == NULL) return NULL;
4927
4928 Node* rawmem = memory(Compile::AliasIdxRaw);
4929 // Is the allocation's memory state untouched?
4930 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4931 // Bail out if there have been raw-memory effects since the allocation.
4932 // (Example: There might have been a call or safepoint.)
4933 return NULL;
4934 }
4935 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4936 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4937 return NULL;
4938 }
4939
4940 // There must be no unexpected observers of this allocation.
4941 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4942 Node* obs = ptr->fast_out(i);
4943 if (obs != this->map()) {
4944 return NULL;
4945 }
4946 }
4947
4948 // This arraycopy must unconditionally follow the allocation of the ptr.
4949 Node* alloc_ctl = ptr->in(0);
4950 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4951
4952 Node* ctl = control();
4953 while (ctl != alloc_ctl) {
4954 // There may be guards which feed into the slow_region.
4955 // Any other control flow means that we might not get a chance
4956 // to finish initializing the allocated object.
4957 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4958 IfNode* iff = ctl->in(0)->as_If();
4959 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4960 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4961 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4962 ctl = iff->in(0); // This test feeds the known slow_region.
4963 continue;
4964 }
4965 // One more try: Various low-level checks bottom out in
4966 // uncommon traps. If the debug-info of the trap omits
4967 // any reference to the allocation, as we've already
4968 // observed, then there can be no objection to the trap.
4969 bool found_trap = false;
4970 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
4971 Node* obs = not_ctl->fast_out(j);
4972 if (obs->in(0) == not_ctl && obs->is_Call() &&
4973 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
4974 found_trap = true; break;
4975 }
4976 }
4977 if (found_trap) {
4978 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
4979 continue;
4980 }
4981 }
4982 return NULL;
4983 }
4984
4985 // If we get this far, we have an allocation which immediately
4986 // precedes the arraycopy, and we can take over zeroing the new object.
4987 // The arraycopy will finish the initialization, and provide
4988 // a new control state to which we will anchor the destination pointer.
4989
4990 return alloc;
4991 }
4992
4993 //-------------inline_encodeISOArray-----------------------------------
4994 // encode char[] to byte[] in ISO_8859_1
4995 bool LibraryCallKit::inline_encodeISOArray() {
4996 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
4997 // no receiver since it is static method
4998 Node *src = argument(0);
4999 Node *src_offset = argument(1);
5000 Node *dst = argument(2);
5001 Node *dst_offset = argument(3);
5002 Node *length = argument(4);
5003
5004 const Type* src_type = src->Value(&_gvn);
5005 const Type* dst_type = dst->Value(&_gvn);
5006 const TypeAryPtr* top_src = src_type->isa_aryptr();
5007 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5008 if (top_src == NULL || top_src->klass() == NULL ||
5009 top_dest == NULL || top_dest->klass() == NULL) {
5010 // failed array check
5011 return false;
5012 }
5013
5014 // Figure out the size and type of the elements we will be copying.
5015 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5016 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5017 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5018 return false;
5019 }
5020 Node* src_start = array_element_address(src, src_offset, src_elem);
5021 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5022 // 'src_start' points to src array + scaled offset
5023 // 'dst_start' points to dst array + scaled offset
5024
5025 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5026 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5027 enc = _gvn.transform(enc);
5028 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5029 set_memory(res_mem, mtype);
5030 set_result(enc);
5031 return true;
5032 }
5033
5034 //-------------inline_multiplyToLen-----------------------------------
5035 bool LibraryCallKit::inline_multiplyToLen() {
5036 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5037
5038 address stubAddr = StubRoutines::multiplyToLen();
5039 if (stubAddr == NULL) {
5040 return false; // Intrinsic's stub is not implemented on this platform
5041 }
5042 const char* stubName = "multiplyToLen";
5043
5044 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5045
5046 // no receiver because it is a static method
5047 Node* x = argument(0);
5048 Node* xlen = argument(1);
5049 Node* y = argument(2);
5050 Node* ylen = argument(3);
5051 Node* z = argument(4);
5052
5053 const Type* x_type = x->Value(&_gvn);
5054 const Type* y_type = y->Value(&_gvn);
5055 const TypeAryPtr* top_x = x_type->isa_aryptr();
5056 const TypeAryPtr* top_y = y_type->isa_aryptr();
5057 if (top_x == NULL || top_x->klass() == NULL ||
5058 top_y == NULL || top_y->klass() == NULL) {
5059 // failed array check
5060 return false;
5061 }
5062
5063 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5064 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5065 if (x_elem != T_INT || y_elem != T_INT) {
5066 return false;
5067 }
5068
5069 // Set the original stack and the reexecute bit for the interpreter to reexecute
5070 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5071 // on the return from z array allocation in runtime.
5072 { PreserveReexecuteState preexecs(this);
5073 jvms()->set_should_reexecute(true);
5074
5075 Node* x_start = array_element_address(x, intcon(0), x_elem);
5076 Node* y_start = array_element_address(y, intcon(0), y_elem);
5077 // 'x_start' points to x array + scaled xlen
5078 // 'y_start' points to y array + scaled ylen
5079
5080 // Allocate the result array
5081 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5082 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5083 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5084
5085 IdealKit ideal(this);
5086
5087 #define __ ideal.
5088 Node* one = __ ConI(1);
5089 Node* zero = __ ConI(0);
5090 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5091 __ set(need_alloc, zero);
5092 __ set(z_alloc, z);
5093 __ if_then(z, BoolTest::eq, null()); {
5094 __ increment (need_alloc, one);
5095 } __ else_(); {
5096 // Update graphKit memory and control from IdealKit.
5097 sync_kit(ideal);
5098 Node* zlen_arg = load_array_length(z);
5099 // Update IdealKit memory and control from graphKit.
5100 __ sync_kit(this);
5101 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5102 __ increment (need_alloc, one);
5103 } __ end_if();
5104 } __ end_if();
5105
5106 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5107 // Update graphKit memory and control from IdealKit.
5108 sync_kit(ideal);
5109 Node * narr = new_array(klass_node, zlen, 1);
5110 // Update IdealKit memory and control from graphKit.
5111 __ sync_kit(this);
5112 __ set(z_alloc, narr);
5113 } __ end_if();
5114
5115 sync_kit(ideal);
5116 z = __ value(z_alloc);
5117 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5118 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5119 // Final sync IdealKit and GraphKit.
5120 final_sync(ideal);
5121 #undef __
5122
5123 Node* z_start = array_element_address(z, intcon(0), T_INT);
5124
5125 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5126 OptoRuntime::multiplyToLen_Type(),
5127 stubAddr, stubName, TypePtr::BOTTOM,
5128 x_start, xlen, y_start, ylen, z_start, zlen);
5129 } // original reexecute is set back here
5130
5131 C->set_has_split_ifs(true); // Has chance for split-if optimization
5132 set_result(z);
5133 return true;
5134 }
5135
5136 //-------------inline_squareToLen------------------------------------
5137 bool LibraryCallKit::inline_squareToLen() {
5138 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5139
5140 address stubAddr = StubRoutines::squareToLen();
5141 if (stubAddr == NULL) {
5142 return false; // Intrinsic's stub is not implemented on this platform
5143 }
5144 const char* stubName = "squareToLen";
5145
5146 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5147
5148 Node* x = argument(0);
5149 Node* len = argument(1);
5150 Node* z = argument(2);
5151 Node* zlen = argument(3);
5152
5153 const Type* x_type = x->Value(&_gvn);
5154 const Type* z_type = z->Value(&_gvn);
5155 const TypeAryPtr* top_x = x_type->isa_aryptr();
5156 const TypeAryPtr* top_z = z_type->isa_aryptr();
5157 if (top_x == NULL || top_x->klass() == NULL ||
5158 top_z == NULL || top_z->klass() == NULL) {
5159 // failed array check
5160 return false;
5161 }
5162
5163 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5164 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5165 if (x_elem != T_INT || z_elem != T_INT) {
5166 return false;
5167 }
5168
5169
5170 Node* x_start = array_element_address(x, intcon(0), x_elem);
5171 Node* z_start = array_element_address(z, intcon(0), z_elem);
5172
5173 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5174 OptoRuntime::squareToLen_Type(),
5175 stubAddr, stubName, TypePtr::BOTTOM,
5176 x_start, len, z_start, zlen);
5177
5178 set_result(z);
5179 return true;
5180 }
5181
5182 //-------------inline_mulAdd------------------------------------------
5183 bool LibraryCallKit::inline_mulAdd() {
5184 assert(UseMulAddIntrinsic, "not implementated on this platform");
5185
5186 address stubAddr = StubRoutines::mulAdd();
5187 if (stubAddr == NULL) {
5188 return false; // Intrinsic's stub is not implemented on this platform
5189 }
5190 const char* stubName = "mulAdd";
5191
5192 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5193
5194 Node* out = argument(0);
5195 Node* in = argument(1);
5196 Node* offset = argument(2);
5197 Node* len = argument(3);
5198 Node* k = argument(4);
5199
5200 const Type* out_type = out->Value(&_gvn);
5201 const Type* in_type = in->Value(&_gvn);
5202 const TypeAryPtr* top_out = out_type->isa_aryptr();
5203 const TypeAryPtr* top_in = in_type->isa_aryptr();
5204 if (top_out == NULL || top_out->klass() == NULL ||
5205 top_in == NULL || top_in->klass() == NULL) {
5206 // failed array check
5207 return false;
5208 }
5209
5210 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5211 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5212 if (out_elem != T_INT || in_elem != T_INT) {
5213 return false;
5214 }
5215
5216 Node* outlen = load_array_length(out);
5217 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5218 Node* out_start = array_element_address(out, intcon(0), out_elem);
5219 Node* in_start = array_element_address(in, intcon(0), in_elem);
5220
5221 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5222 OptoRuntime::mulAdd_Type(),
5223 stubAddr, stubName, TypePtr::BOTTOM,
5224 out_start,in_start, new_offset, len, k);
5225 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5226 set_result(result);
5227 return true;
5228 }
5229
5230 //-------------inline_montgomeryMultiply-----------------------------------
5231 bool LibraryCallKit::inline_montgomeryMultiply() {
5232 address stubAddr = StubRoutines::montgomeryMultiply();
5233 if (stubAddr == NULL) {
5234 return false; // Intrinsic's stub is not implemented on this platform
5235 }
5236
5237 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5238 const char* stubName = "montgomery_square";
5239
5240 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5241
5242 Node* a = argument(0);
5243 Node* b = argument(1);
5244 Node* n = argument(2);
5245 Node* len = argument(3);
5246 Node* inv = argument(4);
5247 Node* m = argument(6);
5248
5249 const Type* a_type = a->Value(&_gvn);
5250 const TypeAryPtr* top_a = a_type->isa_aryptr();
5251 const Type* b_type = b->Value(&_gvn);
5252 const TypeAryPtr* top_b = b_type->isa_aryptr();
5253 const Type* n_type = a->Value(&_gvn);
5254 const TypeAryPtr* top_n = n_type->isa_aryptr();
5255 const Type* m_type = a->Value(&_gvn);
5256 const TypeAryPtr* top_m = m_type->isa_aryptr();
5257 if (top_a == NULL || top_a->klass() == NULL ||
5258 top_b == NULL || top_b->klass() == NULL ||
5259 top_n == NULL || top_n->klass() == NULL ||
5260 top_m == NULL || top_m->klass() == NULL) {
5261 // failed array check
5262 return false;
5263 }
5264
5265 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5266 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5267 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5268 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5269 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5270 return false;
5271 }
5272
5273 // Make the call
5274 {
5275 Node* a_start = array_element_address(a, intcon(0), a_elem);
5276 Node* b_start = array_element_address(b, intcon(0), b_elem);
5277 Node* n_start = array_element_address(n, intcon(0), n_elem);
5278 Node* m_start = array_element_address(m, intcon(0), m_elem);
5279
5280 Node* call = make_runtime_call(RC_LEAF,
5281 OptoRuntime::montgomeryMultiply_Type(),
5282 stubAddr, stubName, TypePtr::BOTTOM,
5283 a_start, b_start, n_start, len, inv, top(),
5284 m_start);
5285 set_result(m);
5286 }
5287
5288 return true;
5289 }
5290
5291 bool LibraryCallKit::inline_montgomerySquare() {
5292 address stubAddr = StubRoutines::montgomerySquare();
5293 if (stubAddr == NULL) {
5294 return false; // Intrinsic's stub is not implemented on this platform
5295 }
5296
5297 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5298 const char* stubName = "montgomery_square";
5299
5300 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5301
5302 Node* a = argument(0);
5303 Node* n = argument(1);
5304 Node* len = argument(2);
5305 Node* inv = argument(3);
5306 Node* m = argument(5);
5307
5308 const Type* a_type = a->Value(&_gvn);
5309 const TypeAryPtr* top_a = a_type->isa_aryptr();
5310 const Type* n_type = a->Value(&_gvn);
5311 const TypeAryPtr* top_n = n_type->isa_aryptr();
5312 const Type* m_type = a->Value(&_gvn);
5313 const TypeAryPtr* top_m = m_type->isa_aryptr();
5314 if (top_a == NULL || top_a->klass() == NULL ||
5315 top_n == NULL || top_n->klass() == NULL ||
5316 top_m == NULL || top_m->klass() == NULL) {
5317 // failed array check
5318 return false;
5319 }
5320
5321 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5322 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5323 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5324 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5325 return false;
5326 }
5327
5328 // Make the call
5329 {
5330 Node* a_start = array_element_address(a, intcon(0), a_elem);
5331 Node* n_start = array_element_address(n, intcon(0), n_elem);
5332 Node* m_start = array_element_address(m, intcon(0), m_elem);
5333
5334 Node* call = make_runtime_call(RC_LEAF,
5335 OptoRuntime::montgomerySquare_Type(),
5336 stubAddr, stubName, TypePtr::BOTTOM,
5337 a_start, n_start, len, inv, top(),
5338 m_start);
5339 set_result(m);
5340 }
5341
5342 return true;
5343 }
5344
5345
5346 /**
5347 * Calculate CRC32 for byte.
5348 * int java.util.zip.CRC32.update(int crc, int b)
5349 */
5350 bool LibraryCallKit::inline_updateCRC32() {
5351 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5352 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5353 // no receiver since it is static method
5354 Node* crc = argument(0); // type: int
5355 Node* b = argument(1); // type: int
5356
5357 /*
5358 * int c = ~ crc;
5359 * b = timesXtoThe32[(b ^ c) & 0xFF];
5360 * b = b ^ (c >>> 8);
5361 * crc = ~b;
5362 */
5363
5364 Node* M1 = intcon(-1);
5365 crc = _gvn.transform(new XorINode(crc, M1));
5366 Node* result = _gvn.transform(new XorINode(crc, b));
5367 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5368
5369 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5370 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5371 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5372 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5373
5374 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5375 result = _gvn.transform(new XorINode(crc, result));
5376 result = _gvn.transform(new XorINode(result, M1));
5377 set_result(result);
5378 return true;
5379 }
5380
5381 /**
5382 * Calculate CRC32 for byte[] array.
5383 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5384 */
5385 bool LibraryCallKit::inline_updateBytesCRC32() {
5386 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5387 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5388 // no receiver since it is static method
5389 Node* crc = argument(0); // type: int
5390 Node* src = argument(1); // type: oop
5391 Node* offset = argument(2); // type: int
5392 Node* length = argument(3); // type: int
5393
5394 const Type* src_type = src->Value(&_gvn);
5395 const TypeAryPtr* top_src = src_type->isa_aryptr();
5396 if (top_src == NULL || top_src->klass() == NULL) {
5397 // failed array check
5398 return false;
5399 }
5400
5401 // Figure out the size and type of the elements we will be copying.
5402 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5403 if (src_elem != T_BYTE) {
5404 return false;
5405 }
5406
5407 // 'src_start' points to src array + scaled offset
5408 Node* src_start = array_element_address(src, offset, src_elem);
5409
5410 // We assume that range check is done by caller.
5411 // TODO: generate range check (offset+length < src.length) in debug VM.
5412
5413 // Call the stub.
5414 address stubAddr = StubRoutines::updateBytesCRC32();
5415 const char *stubName = "updateBytesCRC32";
5416
5417 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5418 stubAddr, stubName, TypePtr::BOTTOM,
5419 crc, src_start, length);
5420 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5421 set_result(result);
5422 return true;
5423 }
5424
5425 /**
5426 * Calculate CRC32 for ByteBuffer.
5427 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5428 */
5429 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5430 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5431 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5432 // no receiver since it is static method
5433 Node* crc = argument(0); // type: int
5434 Node* src = argument(1); // type: long
5435 Node* offset = argument(3); // type: int
5436 Node* length = argument(4); // type: int
5437
5438 src = ConvL2X(src); // adjust Java long to machine word
5439 Node* base = _gvn.transform(new CastX2PNode(src));
5440 offset = ConvI2X(offset);
5441
5442 // 'src_start' points to src array + scaled offset
5443 Node* src_start = basic_plus_adr(top(), base, offset);
5444
5445 // Call the stub.
5446 address stubAddr = StubRoutines::updateBytesCRC32();
5447 const char *stubName = "updateBytesCRC32";
5448
5449 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5450 stubAddr, stubName, TypePtr::BOTTOM,
5451 crc, src_start, length);
5452 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5453 set_result(result);
5454 return true;
5455 }
5456
5457 //------------------------------get_table_from_crc32c_class-----------------------
5458 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5459 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5460 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5461
5462 return table;
5463 }
5464
5465 //------------------------------inline_updateBytesCRC32C-----------------------
5466 //
5467 // Calculate CRC32C for byte[] array.
5468 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5469 //
5470 bool LibraryCallKit::inline_updateBytesCRC32C() {
5471 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5472 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5473 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5474 // no receiver since it is a static method
5475 Node* crc = argument(0); // type: int
5476 Node* src = argument(1); // type: oop
5477 Node* offset = argument(2); // type: int
5478 Node* end = argument(3); // type: int
5479
5480 Node* length = _gvn.transform(new SubINode(end, offset));
5481
5482 const Type* src_type = src->Value(&_gvn);
5483 const TypeAryPtr* top_src = src_type->isa_aryptr();
5484 if (top_src == NULL || top_src->klass() == NULL) {
5485 // failed array check
5486 return false;
5487 }
5488
5489 // Figure out the size and type of the elements we will be copying.
5490 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5491 if (src_elem != T_BYTE) {
5492 return false;
5493 }
5494
5495 // 'src_start' points to src array + scaled offset
5496 Node* src_start = array_element_address(src, offset, src_elem);
5497
5498 // static final int[] byteTable in class CRC32C
5499 Node* table = get_table_from_crc32c_class(callee()->holder());
5500 Node* table_start = array_element_address(table, intcon(0), T_INT);
5501
5502 // We assume that range check is done by caller.
5503 // TODO: generate range check (offset+length < src.length) in debug VM.
5504
5505 // Call the stub.
5506 address stubAddr = StubRoutines::updateBytesCRC32C();
5507 const char *stubName = "updateBytesCRC32C";
5508
5509 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5510 stubAddr, stubName, TypePtr::BOTTOM,
5511 crc, src_start, length, table_start);
5512 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5513 set_result(result);
5514 return true;
5515 }
5516
5517 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5518 //
5519 // Calculate CRC32C for DirectByteBuffer.
5520 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5521 //
5522 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5523 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5524 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5525 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5526 // no receiver since it is a static method
5527 Node* crc = argument(0); // type: int
5528 Node* src = argument(1); // type: long
5529 Node* offset = argument(3); // type: int
5530 Node* end = argument(4); // type: int
5531
5532 Node* length = _gvn.transform(new SubINode(end, offset));
5533
5534 src = ConvL2X(src); // adjust Java long to machine word
5535 Node* base = _gvn.transform(new CastX2PNode(src));
5536 offset = ConvI2X(offset);
5537
5538 // 'src_start' points to src array + scaled offset
5539 Node* src_start = basic_plus_adr(top(), base, offset);
5540
5541 // static final int[] byteTable in class CRC32C
5542 Node* table = get_table_from_crc32c_class(callee()->holder());
5543 Node* table_start = array_element_address(table, intcon(0), T_INT);
5544
5545 // Call the stub.
5546 address stubAddr = StubRoutines::updateBytesCRC32C();
5547 const char *stubName = "updateBytesCRC32C";
5548
5549 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5550 stubAddr, stubName, TypePtr::BOTTOM,
5551 crc, src_start, length, table_start);
5552 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5553 set_result(result);
5554 return true;
5555 }
5556
5557 //------------------------------inline_updateBytesAdler32----------------------
5558 //
5559 // Calculate Adler32 checksum for byte[] array.
5560 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5561 //
5562 bool LibraryCallKit::inline_updateBytesAdler32() {
5563 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5564 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5565 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5566 // no receiver since it is static method
5567 Node* crc = argument(0); // type: int
5568 Node* src = argument(1); // type: oop
5569 Node* offset = argument(2); // type: int
5570 Node* length = argument(3); // type: int
5571
5572 const Type* src_type = src->Value(&_gvn);
5573 const TypeAryPtr* top_src = src_type->isa_aryptr();
5574 if (top_src == NULL || top_src->klass() == NULL) {
5575 // failed array check
5576 return false;
5577 }
5578
5579 // Figure out the size and type of the elements we will be copying.
5580 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5581 if (src_elem != T_BYTE) {
5582 return false;
5583 }
5584
5585 // 'src_start' points to src array + scaled offset
5586 Node* src_start = array_element_address(src, offset, src_elem);
5587
5588 // We assume that range check is done by caller.
5589 // TODO: generate range check (offset+length < src.length) in debug VM.
5590
5591 // Call the stub.
5592 address stubAddr = StubRoutines::updateBytesAdler32();
5593 const char *stubName = "updateBytesAdler32";
5594
5595 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5596 stubAddr, stubName, TypePtr::BOTTOM,
5597 crc, src_start, length);
5598 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5599 set_result(result);
5600 return true;
5601 }
5602
5603 //------------------------------inline_updateByteBufferAdler32---------------
5604 //
5605 // Calculate Adler32 checksum for DirectByteBuffer.
5606 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5607 //
5608 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5609 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5610 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5611 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5612 // no receiver since it is static method
5613 Node* crc = argument(0); // type: int
5614 Node* src = argument(1); // type: long
5615 Node* offset = argument(3); // type: int
5616 Node* length = argument(4); // type: int
5617
5618 src = ConvL2X(src); // adjust Java long to machine word
5619 Node* base = _gvn.transform(new CastX2PNode(src));
5620 offset = ConvI2X(offset);
5621
5622 // 'src_start' points to src array + scaled offset
5623 Node* src_start = basic_plus_adr(top(), base, offset);
5624
5625 // Call the stub.
5626 address stubAddr = StubRoutines::updateBytesAdler32();
5627 const char *stubName = "updateBytesAdler32";
5628
5629 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5630 stubAddr, stubName, TypePtr::BOTTOM,
5631 crc, src_start, length);
5632
5633 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5634 set_result(result);
5635 return true;
5636 }
5637
5638 //----------------------------inline_reference_get----------------------------
5639 // public T java.lang.ref.Reference.get();
5640 bool LibraryCallKit::inline_reference_get() {
5641 const int referent_offset = java_lang_ref_Reference::referent_offset;
5642 guarantee(referent_offset > 0, "should have already been set");
5643
5644 // Get the argument:
5645 Node* reference_obj = null_check_receiver();
5646 if (stopped()) return true;
5647
5648 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5649
5650 ciInstanceKlass* klass = env()->Object_klass();
5651 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5652
5653 Node* no_ctrl = NULL;
5654 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5655
5656 // Use the pre-barrier to record the value in the referent field
5657 pre_barrier(false /* do_load */,
5658 control(),
5659 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5660 result /* pre_val */,
5661 T_OBJECT);
5662
5663 // Add memory barrier to prevent commoning reads from this field
5664 // across safepoint since GC can change its value.
5665 insert_mem_bar(Op_MemBarCPUOrder);
5666
5667 set_result(result);
5668 return true;
5669 }
5670
5671
5672 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5673 bool is_exact=true, bool is_static=false,
5674 ciInstanceKlass * fromKls=NULL) {
5675 if (fromKls == NULL) {
5676 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5677 assert(tinst != NULL, "obj is null");
5678 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5679 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5680 fromKls = tinst->klass()->as_instance_klass();
5681 } else {
5682 assert(is_static, "only for static field access");
5683 }
5684 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5685 ciSymbol::make(fieldTypeString),
5686 is_static);
5687
5688 assert (field != NULL, "undefined field");
5689 if (field == NULL) return (Node *) NULL;
5690
5691 if (is_static) {
5692 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5693 fromObj = makecon(tip);
5694 }
5695
5696 // Next code copied from Parse::do_get_xxx():
5697
5698 // Compute address and memory type.
5699 int offset = field->offset_in_bytes();
5700 bool is_vol = field->is_volatile();
5701 ciType* field_klass = field->type();
5702 assert(field_klass->is_loaded(), "should be loaded");
5703 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5704 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5705 BasicType bt = field->layout_type();
5706
5707 // Build the resultant type of the load
5708 const Type *type;
5709 if (bt == T_OBJECT) {
5710 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5711 } else {
5712 type = Type::get_const_basic_type(bt);
5713 }
5714
5715 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5716 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5717 }
5718 // Build the load.
5719 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5720 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5721 // If reference is volatile, prevent following memory ops from
5722 // floating up past the volatile read. Also prevents commoning
5723 // another volatile read.
5724 if (is_vol) {
5725 // Memory barrier includes bogus read of value to force load BEFORE membar
5726 insert_mem_bar(Op_MemBarAcquire, loadedField);
5727 }
5728 return loadedField;
5729 }
5730
5731
5732 //------------------------------inline_aescrypt_Block-----------------------
5733 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5734 address stubAddr;
5735 const char *stubName;
5736 assert(UseAES, "need AES instruction support");
5737
5738 switch(id) {
5739 case vmIntrinsics::_aescrypt_encryptBlock:
5740 stubAddr = StubRoutines::aescrypt_encryptBlock();
5741 stubName = "aescrypt_encryptBlock";
5742 break;
5743 case vmIntrinsics::_aescrypt_decryptBlock:
5744 stubAddr = StubRoutines::aescrypt_decryptBlock();
5745 stubName = "aescrypt_decryptBlock";
5746 break;
5747 }
5748 if (stubAddr == NULL) return false;
5749
5750 Node* aescrypt_object = argument(0);
5751 Node* src = argument(1);
5752 Node* src_offset = argument(2);
5753 Node* dest = argument(3);
5754 Node* dest_offset = argument(4);
5755
5756 // (1) src and dest are arrays.
5757 const Type* src_type = src->Value(&_gvn);
5758 const Type* dest_type = dest->Value(&_gvn);
5759 const TypeAryPtr* top_src = src_type->isa_aryptr();
5760 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5761 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5762
5763 // for the quick and dirty code we will skip all the checks.
5764 // we are just trying to get the call to be generated.
5765 Node* src_start = src;
5766 Node* dest_start = dest;
5767 if (src_offset != NULL || dest_offset != NULL) {
5768 assert(src_offset != NULL && dest_offset != NULL, "");
5769 src_start = array_element_address(src, src_offset, T_BYTE);
5770 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5771 }
5772
5773 // now need to get the start of its expanded key array
5774 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5775 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5776 if (k_start == NULL) return false;
5777
5778 if (Matcher::pass_original_key_for_aes()) {
5779 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5780 // compatibility issues between Java key expansion and SPARC crypto instructions
5781 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5782 if (original_k_start == NULL) return false;
5783
5784 // Call the stub.
5785 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5786 stubAddr, stubName, TypePtr::BOTTOM,
5787 src_start, dest_start, k_start, original_k_start);
5788 } else {
5789 // Call the stub.
5790 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5791 stubAddr, stubName, TypePtr::BOTTOM,
5792 src_start, dest_start, k_start);
5793 }
5794
5795 return true;
5796 }
5797
5798 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5799 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5800 address stubAddr;
5801 const char *stubName;
5802
5803 assert(UseAES, "need AES instruction support");
5804
5805 switch(id) {
5806 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5807 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5808 stubName = "cipherBlockChaining_encryptAESCrypt";
5809 break;
5810 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5811 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5812 stubName = "cipherBlockChaining_decryptAESCrypt";
5813 break;
5814 }
5815 if (stubAddr == NULL) return false;
5816
5817 Node* cipherBlockChaining_object = argument(0);
5818 Node* src = argument(1);
5819 Node* src_offset = argument(2);
5820 Node* len = argument(3);
5821 Node* dest = argument(4);
5822 Node* dest_offset = argument(5);
5823
5824 // (1) src and dest are arrays.
5825 const Type* src_type = src->Value(&_gvn);
5826 const Type* dest_type = dest->Value(&_gvn);
5827 const TypeAryPtr* top_src = src_type->isa_aryptr();
5828 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5829 assert (top_src != NULL && top_src->klass() != NULL
5830 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5831
5832 // checks are the responsibility of the caller
5833 Node* src_start = src;
5834 Node* dest_start = dest;
5835 if (src_offset != NULL || dest_offset != NULL) {
5836 assert(src_offset != NULL && dest_offset != NULL, "");
5837 src_start = array_element_address(src, src_offset, T_BYTE);
5838 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5839 }
5840
5841 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5842 // (because of the predicated logic executed earlier).
5843 // so we cast it here safely.
5844 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5845
5846 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5847 if (embeddedCipherObj == NULL) return false;
5848
5849 // cast it to what we know it will be at runtime
5850 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5851 assert(tinst != NULL, "CBC obj is null");
5852 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5853 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5854 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
5855
5856 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5857 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5858 const TypeOopPtr* xtype = aklass->as_instance_type();
5859 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
5860 aescrypt_object = _gvn.transform(aescrypt_object);
5861
5862 // we need to get the start of the aescrypt_object's expanded key array
5863 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5864 if (k_start == NULL) return false;
5865
5866 // similarly, get the start address of the r vector
5867 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5868 if (objRvec == NULL) return false;
5869 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5870
5871 Node* cbcCrypt;
5872 if (Matcher::pass_original_key_for_aes()) {
5873 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5874 // compatibility issues between Java key expansion and SPARC crypto instructions
5875 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5876 if (original_k_start == NULL) return false;
5877
5878 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5879 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5880 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5881 stubAddr, stubName, TypePtr::BOTTOM,
5882 src_start, dest_start, k_start, r_start, len, original_k_start);
5883 } else {
5884 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5885 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5886 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5887 stubAddr, stubName, TypePtr::BOTTOM,
5888 src_start, dest_start, k_start, r_start, len);
5889 }
5890
5891 // return cipher length (int)
5892 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
5893 set_result(retvalue);
5894 return true;
5895 }
5896
5897 //------------------------------get_key_start_from_aescrypt_object-----------------------
5898 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5899 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5900 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5901 if (objAESCryptKey == NULL) return (Node *) NULL;
5902
5903 // now have the array, need to get the start address of the K array
5904 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5905 return k_start;
5906 }
5907
5908 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
5909 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
5910 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
5911 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5912 if (objAESCryptKey == NULL) return (Node *) NULL;
5913
5914 // now have the array, need to get the start address of the lastKey array
5915 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
5916 return original_k_start;
5917 }
5918
5919 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5920 // Return node representing slow path of predicate check.
5921 // the pseudo code we want to emulate with this predicate is:
5922 // for encryption:
5923 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5924 // for decryption:
5925 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5926 // note cipher==plain is more conservative than the original java code but that's OK
5927 //
5928 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5929 // The receiver was checked for NULL already.
5930 Node* objCBC = argument(0);
5931
5932 // Load embeddedCipher field of CipherBlockChaining object.
5933 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5934
5935 // get AESCrypt klass for instanceOf check
5936 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5937 // will have same classloader as CipherBlockChaining object
5938 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5939 assert(tinst != NULL, "CBCobj is null");
5940 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5941
5942 // we want to do an instanceof comparison against the AESCrypt class
5943 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5944 if (!klass_AESCrypt->is_loaded()) {
5945 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5946 Node* ctrl = control();
5947 set_control(top()); // no regular fast path
5948 return ctrl;
5949 }
5950 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5951
5952 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5953 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
5954 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
5955
5956 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5957
5958 // for encryption, we are done
5959 if (!decrypting)
5960 return instof_false; // even if it is NULL
5961
5962 // for decryption, we need to add a further check to avoid
5963 // taking the intrinsic path when cipher and plain are the same
5964 // see the original java code for why.
5965 RegionNode* region = new RegionNode(3);
5966 region->init_req(1, instof_false);
5967 Node* src = argument(1);
5968 Node* dest = argument(4);
5969 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
5970 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
5971 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5972 region->init_req(2, src_dest_conjoint);
5973
5974 record_for_igvn(region);
5975 return _gvn.transform(region);
5976 }
5977
5978 //------------------------------inline_ghash_processBlocks
5979 bool LibraryCallKit::inline_ghash_processBlocks() {
5980 address stubAddr;
5981 const char *stubName;
5982 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
5983
5984 stubAddr = StubRoutines::ghash_processBlocks();
5985 stubName = "ghash_processBlocks";
5986
5987 Node* data = argument(0);
5988 Node* offset = argument(1);
5989 Node* len = argument(2);
5990 Node* state = argument(3);
5991 Node* subkeyH = argument(4);
5992
5993 Node* state_start = array_element_address(state, intcon(0), T_LONG);
5994 assert(state_start, "state is NULL");
5995 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
5996 assert(subkeyH_start, "subkeyH is NULL");
5997 Node* data_start = array_element_address(data, offset, T_BYTE);
5998 assert(data_start, "data is NULL");
5999
6000 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6001 OptoRuntime::ghash_processBlocks_Type(),
6002 stubAddr, stubName, TypePtr::BOTTOM,
6003 state_start, subkeyH_start, data_start, len);
6004 return true;
6005 }
6006
6007 //------------------------------inline_sha_implCompress-----------------------
6008 //
6009 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6010 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6011 //
6012 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6013 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6014 //
6015 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6016 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6017 //
6018 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6019 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6020
6021 Node* sha_obj = argument(0);
6022 Node* src = argument(1); // type oop
6023 Node* ofs = argument(2); // type int
6024
6025 const Type* src_type = src->Value(&_gvn);
6026 const TypeAryPtr* top_src = src_type->isa_aryptr();
6027 if (top_src == NULL || top_src->klass() == NULL) {
6028 // failed array check
6029 return false;
6030 }
6031 // Figure out the size and type of the elements we will be copying.
6032 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6033 if (src_elem != T_BYTE) {
6034 return false;
6035 }
6036 // 'src_start' points to src array + offset
6037 Node* src_start = array_element_address(src, ofs, src_elem);
6038 Node* state = NULL;
6039 address stubAddr;
6040 const char *stubName;
6041
6042 switch(id) {
6043 case vmIntrinsics::_sha_implCompress:
6044 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6045 state = get_state_from_sha_object(sha_obj);
6046 stubAddr = StubRoutines::sha1_implCompress();
6047 stubName = "sha1_implCompress";
6048 break;
6049 case vmIntrinsics::_sha2_implCompress:
6050 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6051 state = get_state_from_sha_object(sha_obj);
6052 stubAddr = StubRoutines::sha256_implCompress();
6053 stubName = "sha256_implCompress";
6054 break;
6055 case vmIntrinsics::_sha5_implCompress:
6056 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6057 state = get_state_from_sha5_object(sha_obj);
6058 stubAddr = StubRoutines::sha512_implCompress();
6059 stubName = "sha512_implCompress";
6060 break;
6061 default:
6062 fatal_unexpected_iid(id);
6063 return false;
6064 }
6065 if (state == NULL) return false;
6066
6067 // Call the stub.
6068 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6069 stubAddr, stubName, TypePtr::BOTTOM,
6070 src_start, state);
6071
6072 return true;
6073 }
6074
6075 //------------------------------inline_digestBase_implCompressMB-----------------------
6076 //
6077 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6078 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6079 //
6080 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6081 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6082 "need SHA1/SHA256/SHA512 instruction support");
6083 assert((uint)predicate < 3, "sanity");
6084 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6085
6086 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6087 Node* src = argument(1); // byte[] array
6088 Node* ofs = argument(2); // type int
6089 Node* limit = argument(3); // type int
6090
6091 const Type* src_type = src->Value(&_gvn);
6092 const TypeAryPtr* top_src = src_type->isa_aryptr();
6093 if (top_src == NULL || top_src->klass() == NULL) {
6094 // failed array check
6095 return false;
6096 }
6097 // Figure out the size and type of the elements we will be copying.
6098 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6099 if (src_elem != T_BYTE) {
6100 return false;
6101 }
6102 // 'src_start' points to src array + offset
6103 Node* src_start = array_element_address(src, ofs, src_elem);
6104
6105 const char* klass_SHA_name = NULL;
6106 const char* stub_name = NULL;
6107 address stub_addr = NULL;
6108 bool long_state = false;
6109
6110 switch (predicate) {
6111 case 0:
6112 if (UseSHA1Intrinsics) {
6113 klass_SHA_name = "sun/security/provider/SHA";
6114 stub_name = "sha1_implCompressMB";
6115 stub_addr = StubRoutines::sha1_implCompressMB();
6116 }
6117 break;
6118 case 1:
6119 if (UseSHA256Intrinsics) {
6120 klass_SHA_name = "sun/security/provider/SHA2";
6121 stub_name = "sha256_implCompressMB";
6122 stub_addr = StubRoutines::sha256_implCompressMB();
6123 }
6124 break;
6125 case 2:
6126 if (UseSHA512Intrinsics) {
6127 klass_SHA_name = "sun/security/provider/SHA5";
6128 stub_name = "sha512_implCompressMB";
6129 stub_addr = StubRoutines::sha512_implCompressMB();
6130 long_state = true;
6131 }
6132 break;
6133 default:
6134 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6135 }
6136 if (klass_SHA_name != NULL) {
6137 // get DigestBase klass to lookup for SHA klass
6138 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6139 assert(tinst != NULL, "digestBase_obj is not instance???");
6140 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6141
6142 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6143 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6144 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6145 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6146 }
6147 return false;
6148 }
6149 //------------------------------inline_sha_implCompressMB-----------------------
6150 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6151 bool long_state, address stubAddr, const char *stubName,
6152 Node* src_start, Node* ofs, Node* limit) {
6153 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6154 const TypeOopPtr* xtype = aklass->as_instance_type();
6155 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6156 sha_obj = _gvn.transform(sha_obj);
6157
6158 Node* state;
6159 if (long_state) {
6160 state = get_state_from_sha5_object(sha_obj);
6161 } else {
6162 state = get_state_from_sha_object(sha_obj);
6163 }
6164 if (state == NULL) return false;
6165
6166 // Call the stub.
6167 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6168 OptoRuntime::digestBase_implCompressMB_Type(),
6169 stubAddr, stubName, TypePtr::BOTTOM,
6170 src_start, state, ofs, limit);
6171 // return ofs (int)
6172 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6173 set_result(result);
6174
6175 return true;
6176 }
6177
6178 //------------------------------get_state_from_sha_object-----------------------
6179 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6180 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6181 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6182 if (sha_state == NULL) return (Node *) NULL;
6183
6184 // now have the array, need to get the start address of the state array
6185 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6186 return state;
6187 }
6188
6189 //------------------------------get_state_from_sha5_object-----------------------
6190 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6191 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6192 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6193 if (sha_state == NULL) return (Node *) NULL;
6194
6195 // now have the array, need to get the start address of the state array
6196 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6197 return state;
6198 }
6199
6200 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6201 // Return node representing slow path of predicate check.
6202 // the pseudo code we want to emulate with this predicate is:
6203 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6204 //
6205 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6206 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6207 "need SHA1/SHA256/SHA512 instruction support");
6208 assert((uint)predicate < 3, "sanity");
6209
6210 // The receiver was checked for NULL already.
6211 Node* digestBaseObj = argument(0);
6212
6213 // get DigestBase klass for instanceOf check
6214 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6215 assert(tinst != NULL, "digestBaseObj is null");
6216 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6217
6218 const char* klass_SHA_name = NULL;
6219 switch (predicate) {
6220 case 0:
6221 if (UseSHA1Intrinsics) {
6222 // we want to do an instanceof comparison against the SHA class
6223 klass_SHA_name = "sun/security/provider/SHA";
6224 }
6225 break;
6226 case 1:
6227 if (UseSHA256Intrinsics) {
6228 // we want to do an instanceof comparison against the SHA2 class
6229 klass_SHA_name = "sun/security/provider/SHA2";
6230 }
6231 break;
6232 case 2:
6233 if (UseSHA512Intrinsics) {
6234 // we want to do an instanceof comparison against the SHA5 class
6235 klass_SHA_name = "sun/security/provider/SHA5";
6236 }
6237 break;
6238 default:
6239 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6240 }
6241
6242 ciKlass* klass_SHA = NULL;
6243 if (klass_SHA_name != NULL) {
6244 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6245 }
6246 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6247 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6248 Node* ctrl = control();
6249 set_control(top()); // no intrinsic path
6250 return ctrl;
6251 }
6252 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6253
6254 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6255 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6256 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6257 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6258
6259 return instof_false; // even if it is NULL
6260 }
6261
6262 bool LibraryCallKit::inline_profileBoolean() {
6263 Node* counts = argument(1);
6264 const TypeAryPtr* ary = NULL;
6265 ciArray* aobj = NULL;
6266 if (counts->is_Con()
6267 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6268 && (aobj = ary->const_oop()->as_array()) != NULL
6269 && (aobj->length() == 2)) {
6270 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6271 jint false_cnt = aobj->element_value(0).as_int();
6272 jint true_cnt = aobj->element_value(1).as_int();
6273
6274 if (C->log() != NULL) {
6275 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6276 false_cnt, true_cnt);
6277 }
6278
6279 if (false_cnt + true_cnt == 0) {
6280 // According to profile, never executed.
6281 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6282 Deoptimization::Action_reinterpret);
6283 return true;
6284 }
6285
6286 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6287 // is a number of each value occurrences.
6288 Node* result = argument(0);
6289 if (false_cnt == 0 || true_cnt == 0) {
6290 // According to profile, one value has been never seen.
6291 int expected_val = (false_cnt == 0) ? 1 : 0;
6292
6293 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6294 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6295
6296 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6297 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6298 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6299
6300 { // Slow path: uncommon trap for never seen value and then reexecute
6301 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6302 // the value has been seen at least once.
6303 PreserveJVMState pjvms(this);
6304 PreserveReexecuteState preexecs(this);
6305 jvms()->set_should_reexecute(true);
6306
6307 set_control(slow_path);
6308 set_i_o(i_o());
6309
6310 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6311 Deoptimization::Action_reinterpret);
6312 }
6313 // The guard for never seen value enables sharpening of the result and
6314 // returning a constant. It allows to eliminate branches on the same value
6315 // later on.
6316 set_control(fast_path);
6317 result = intcon(expected_val);
6318 }
6319 // Stop profiling.
6320 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6321 // By replacing method body with profile data (represented as ProfileBooleanNode
6322 // on IR level) we effectively disable profiling.
6323 // It enables full speed execution once optimized code is generated.
6324 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6325 C->record_for_igvn(profile);
6326 set_result(profile);
6327 return true;
6328 } else {
6329 // Continue profiling.
6330 // Profile data isn't available at the moment. So, execute method's bytecode version.
6331 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6332 // is compiled and counters aren't available since corresponding MethodHandle
6333 // isn't a compile-time constant.
6334 return false;
6335 }
6336 }
6337
6338 bool LibraryCallKit::inline_isCompileConstant() {
6339 Node* n = argument(0);
6340 set_result(n->is_Con() ? intcon(1) : intcon(0));
6341 return true;
6342 }