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 "gc/shenandoah/shenandoahRuntime.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "opto/addnode.hpp"
34 #include "opto/arraycopynode.hpp"
35 #include "opto/c2compiler.hpp"
36 #include "opto/callGenerator.hpp"
37 #include "opto/castnode.hpp"
38 #include "opto/cfgnode.hpp"
39 #include "opto/convertnode.hpp"
40 #include "opto/countbitsnode.hpp"
41 #include "opto/intrinsicnode.hpp"
42 #include "opto/idealKit.hpp"
43 #include "opto/mathexactnode.hpp"
44 #include "opto/movenode.hpp"
45 #include "opto/mulnode.hpp"
46 #include "opto/narrowptrnode.hpp"
47 #include "opto/opaquenode.hpp"
48 #include "opto/parse.hpp"
49 #include "opto/runtime.hpp"
50 #include "opto/shenandoahSupport.hpp"
51 #include "opto/subnode.hpp"
52 #include "prims/nativeLookup.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "trace/traceMacros.hpp"
55
56 class LibraryIntrinsic : public InlineCallGenerator {
57 // Extend the set of intrinsics known to the runtime:
58 public:
59 private:
60 bool _is_virtual;
61 bool _does_virtual_dispatch;
62 int8_t _predicates_count; // Intrinsic is predicated by several conditions
63 int8_t _last_predicate; // Last generated predicate
64 vmIntrinsics::ID _intrinsic_id;
65
66 public:
67 LibraryIntrinsic(ciMethod* m, bool is_virtual, int predicates_count, bool does_virtual_dispatch, vmIntrinsics::ID id)
68 : InlineCallGenerator(m),
69 _is_virtual(is_virtual),
70 _does_virtual_dispatch(does_virtual_dispatch),
71 _predicates_count((int8_t)predicates_count),
72 _last_predicate((int8_t)-1),
73 _intrinsic_id(id)
74 {
75 }
76 virtual bool is_intrinsic() const { return true; }
77 virtual bool is_virtual() const { return _is_virtual; }
78 virtual bool is_predicated() const { return _predicates_count > 0; }
79 virtual int predicates_count() const { return _predicates_count; }
80 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
81 virtual JVMState* generate(JVMState* jvms);
82 virtual Node* generate_predicate(JVMState* jvms, int predicate);
83 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
84 };
85
86
87 // Local helper class for LibraryIntrinsic:
88 class LibraryCallKit : public GraphKit {
89 private:
90 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
91 Node* _result; // the result node, if any
92 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
93
94 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
95
96 public:
97 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
98 : GraphKit(jvms),
99 _intrinsic(intrinsic),
100 _result(NULL)
101 {
102 // Check if this is a root compile. In that case we don't have a caller.
103 if (!jvms->has_method()) {
104 _reexecute_sp = sp();
105 } else {
106 // Find out how many arguments the interpreter needs when deoptimizing
107 // and save the stack pointer value so it can used by uncommon_trap.
108 // We find the argument count by looking at the declared signature.
109 bool ignored_will_link;
110 ciSignature* declared_signature = NULL;
111 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
112 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
113 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
114 }
115 }
116
117 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
118
119 ciMethod* caller() const { return jvms()->method(); }
120 int bci() const { return jvms()->bci(); }
121 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
122 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
123 ciMethod* callee() const { return _intrinsic->method(); }
124
125 bool try_to_inline(int predicate);
126 Node* try_to_predicate(int predicate);
127
128 void push_result() {
129 // Push the result onto the stack.
130 if (!stopped() && result() != NULL) {
131 BasicType bt = result()->bottom_type()->basic_type();
132 push_node(bt, result());
133 }
134 }
135
136 private:
137 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
138 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
139 }
140
141 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
142 void set_result(RegionNode* region, PhiNode* value);
143 Node* result() { return _result; }
144
145 virtual int reexecute_sp() { return _reexecute_sp; }
146
147 // Helper functions to inline natives
148 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
149 Node* generate_slow_guard(Node* test, RegionNode* region);
150 Node* generate_fair_guard(Node* test, RegionNode* region);
151 Node* generate_negative_guard(Node* index, RegionNode* region,
152 // resulting CastII of index:
153 Node* *pos_index = NULL);
154 Node* generate_limit_guard(Node* offset, Node* subseq_length,
155 Node* array_length,
156 RegionNode* region);
157 Node* generate_current_thread(Node* &tls_output);
158 Node* load_mirror_from_klass(Node* klass);
159 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path,
161 int offset);
162 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
163 RegionNode* region, int null_path) {
164 int offset = java_lang_Class::klass_offset_in_bytes();
165 return load_klass_from_mirror_common(mirror, never_see_null,
166 region, null_path,
167 offset);
168 }
169 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
170 RegionNode* region, int null_path) {
171 int offset = java_lang_Class::array_klass_offset_in_bytes();
172 return load_klass_from_mirror_common(mirror, never_see_null,
173 region, null_path,
174 offset);
175 }
176 Node* generate_access_flags_guard(Node* kls,
177 int modifier_mask, int modifier_bits,
178 RegionNode* region);
179 Node* generate_interface_guard(Node* kls, RegionNode* region);
180 Node* generate_array_guard(Node* kls, RegionNode* region) {
181 return generate_array_guard_common(kls, region, false, false);
182 }
183 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
184 return generate_array_guard_common(kls, region, false, true);
185 }
186 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
187 return generate_array_guard_common(kls, region, true, false);
188 }
189 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
190 return generate_array_guard_common(kls, region, true, true);
191 }
192 Node* generate_array_guard_common(Node* kls, RegionNode* region,
193 bool obj_array, bool not_array);
194 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
195 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
196 bool is_virtual = false, bool is_static = false);
197 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
198 return generate_method_call(method_id, false, true);
199 }
200 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
201 return generate_method_call(method_id, true, false);
202 }
203 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static, ciInstanceKlass * fromKls);
204
205 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
206 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
207 bool inline_string_compareTo();
208 bool inline_string_indexOf();
209 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
210 bool inline_string_equals();
211 Node* round_double_node(Node* n);
212 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
213 bool inline_math_native(vmIntrinsics::ID id);
214 bool inline_trig(vmIntrinsics::ID id);
215 bool inline_math(vmIntrinsics::ID id);
216 template <typename OverflowOp>
217 bool inline_math_overflow(Node* arg1, Node* arg2);
218 void inline_math_mathExact(Node* math, Node* test);
219 bool inline_math_addExactI(bool is_increment);
220 bool inline_math_addExactL(bool is_increment);
221 bool inline_math_multiplyExactI();
222 bool inline_math_multiplyExactL();
223 bool inline_math_negateExactI();
224 bool inline_math_negateExactL();
225 bool inline_math_subtractExactI(bool is_decrement);
226 bool inline_math_subtractExactL(bool is_decrement);
227 bool inline_exp();
228 bool inline_pow();
229 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
230 bool inline_min_max(vmIntrinsics::ID id);
231 bool inline_notify(vmIntrinsics::ID id);
232 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
233 // This returns Type::AnyPtr, RawPtr, or OopPtr.
234 int classify_unsafe_addr(Node* &base, Node* &offset);
235 Node* make_unsafe_address(Node* base, Node* offset);
236 // Helper for inline_unsafe_access.
237 // Generates the guards that check whether the result of
238 // Unsafe.getObject should be recorded in an SATB log buffer.
239 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
240 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
241 static bool klass_needs_init_guard(Node* kls);
242 bool inline_unsafe_allocate();
243 bool inline_unsafe_copyMemory();
244 bool inline_native_currentThread();
245 #ifdef TRACE_HAVE_INTRINSICS
246 bool inline_native_classID();
247 bool inline_native_threadID();
248 #endif
249 bool inline_native_time_funcs(address method, const char* funcName);
250 bool inline_native_isInterrupted();
251 bool inline_native_Class_query(vmIntrinsics::ID id);
252 bool inline_native_subtype_check();
253
254 bool inline_native_newArray();
255 bool inline_native_getLength();
256 bool inline_array_copyOf(bool is_copyOfRange);
257 bool inline_array_equals();
258 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
259 bool inline_native_clone(bool is_virtual);
260 bool inline_native_Reflection_getCallerClass();
261 // Helper function for inlining native object hash method
262 bool inline_native_hashcode(bool is_virtual, bool is_static);
263 bool inline_native_getClass();
264
265 // Helper functions for inlining arraycopy
266 bool inline_arraycopy();
267 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
268 RegionNode* slow_region);
269 JVMState* arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp);
270 void arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp);
271
272 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
273 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
274 bool inline_unsafe_ordered_store(BasicType type);
275 bool inline_unsafe_fence(vmIntrinsics::ID id);
276 bool inline_fp_conversions(vmIntrinsics::ID id);
277 bool inline_number_methods(vmIntrinsics::ID id);
278 bool inline_reference_get();
279 bool inline_Class_cast();
280 bool inline_aescrypt_Block(vmIntrinsics::ID id);
281 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
282 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
283 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
284 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
285 bool inline_ghash_processBlocks();
286 bool inline_sha_implCompress(vmIntrinsics::ID id);
287 bool inline_digestBase_implCompressMB(int predicate);
288 bool inline_sha_implCompressMB(Node* digestBaseObj, ciInstanceKlass* instklass_SHA,
289 bool long_state, address stubAddr, const char *stubName,
290 Node* src_start, Node* ofs, Node* limit);
291 Node* get_state_from_sha_object(Node *sha_object);
292 Node* get_state_from_sha5_object(Node *sha_object);
293 Node* inline_digestBase_implCompressMB_predicate(int predicate);
294 bool inline_encodeISOArray();
295 bool inline_updateCRC32();
296 bool inline_updateBytesCRC32();
297 bool inline_updateByteBufferCRC32();
298 Node* get_table_from_crc32c_class(ciInstanceKlass *crc32c_class);
299 bool inline_updateBytesCRC32C();
300 bool inline_updateDirectByteBufferCRC32C();
301 bool inline_updateBytesAdler32();
302 bool inline_updateByteBufferAdler32();
303 bool inline_multiplyToLen();
304 bool inline_squareToLen();
305 bool inline_mulAdd();
306 bool inline_montgomeryMultiply();
307 bool inline_montgomerySquare();
308
309 bool inline_profileBoolean();
310 bool inline_isCompileConstant();
311 };
312
313 //---------------------------make_vm_intrinsic----------------------------
314 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
315 vmIntrinsics::ID id = m->intrinsic_id();
316 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
317
318 if (!m->is_loaded()) {
319 // Do not attempt to inline unloaded methods.
320 return NULL;
321 }
322
323 C2Compiler* compiler = (C2Compiler*)CompileBroker::compiler(CompLevel_full_optimization);
324 bool is_available = false;
325
326 {
327 // For calling is_intrinsic_supported and is_intrinsic_disabled_by_flag
328 // the compiler must transition to '_thread_in_vm' state because both
329 // methods access VM-internal data.
330 VM_ENTRY_MARK;
331 methodHandle mh(THREAD, m->get_Method());
332 methodHandle ct(THREAD, method()->get_Method());
333 is_available = compiler->is_intrinsic_supported(mh, is_virtual) &&
334 !vmIntrinsics::is_disabled_by_flags(mh, ct);
335 }
336
337 if (is_available) {
338 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
339 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
340 return new LibraryIntrinsic(m, is_virtual,
341 vmIntrinsics::predicates_needed(id),
342 vmIntrinsics::does_virtual_dispatch(id),
343 (vmIntrinsics::ID) id);
344 } else {
345 return NULL;
346 }
347 }
348
349 //----------------------register_library_intrinsics-----------------------
350 // Initialize this file's data structures, for each Compile instance.
351 void Compile::register_library_intrinsics() {
352 // Nothing to do here.
353 }
354
355 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
356 LibraryCallKit kit(jvms, this);
357 Compile* C = kit.C;
358 int nodes = C->unique();
359 #ifndef PRODUCT
360 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
361 char buf[1000];
362 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
363 tty->print_cr("Intrinsic %s", str);
364 }
365 #endif
366 ciMethod* callee = kit.callee();
367 const int bci = kit.bci();
368
369 // Try to inline the intrinsic.
370 if ((CheckIntrinsics ? callee->intrinsic_candidate() : true) &&
371 kit.try_to_inline(_last_predicate)) {
372 if (C->print_intrinsics() || C->print_inlining()) {
373 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
374 }
375 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
376 if (C->log()) {
377 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
378 vmIntrinsics::name_at(intrinsic_id()),
379 (is_virtual() ? " virtual='1'" : ""),
380 C->unique() - nodes);
381 }
382 // Push the result from the inlined method onto the stack.
383 kit.push_result();
384 C->print_inlining_update(this);
385 return kit.transfer_exceptions_into_jvms();
386 }
387
388 // The intrinsic bailed out
389 if (C->print_intrinsics() || C->print_inlining()) {
390 if (jvms->has_method()) {
391 // Not a root compile.
392 const char* msg;
393 if (callee->intrinsic_candidate()) {
394 msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
395 } else {
396 msg = is_virtual() ? "failed to inline (intrinsic, virtual), method not annotated"
397 : "failed to inline (intrinsic), method not annotated";
398 }
399 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
400 } else {
401 // Root compile
402 tty->print("Did not generate intrinsic %s%s at bci:%d in",
403 vmIntrinsics::name_at(intrinsic_id()),
404 (is_virtual() ? " (virtual)" : ""), bci);
405 }
406 }
407 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
408 C->print_inlining_update(this);
409 return NULL;
410 }
411
412 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms, int predicate) {
413 LibraryCallKit kit(jvms, this);
414 Compile* C = kit.C;
415 int nodes = C->unique();
416 _last_predicate = predicate;
417 #ifndef PRODUCT
418 assert(is_predicated() && predicate < predicates_count(), "sanity");
419 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
420 char buf[1000];
421 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
422 tty->print_cr("Predicate for intrinsic %s", str);
423 }
424 #endif
425 ciMethod* callee = kit.callee();
426 const int bci = kit.bci();
427
428 Node* slow_ctl = kit.try_to_predicate(predicate);
429 if (!kit.failing()) {
430 if (C->print_intrinsics() || C->print_inlining()) {
431 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual, predicate)" : "(intrinsic, predicate)");
432 }
433 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
434 if (C->log()) {
435 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
436 vmIntrinsics::name_at(intrinsic_id()),
437 (is_virtual() ? " virtual='1'" : ""),
438 C->unique() - nodes);
439 }
440 return slow_ctl; // Could be NULL if the check folds.
441 }
442
443 // The intrinsic bailed out
444 if (C->print_intrinsics() || C->print_inlining()) {
445 if (jvms->has_method()) {
446 // Not a root compile.
447 const char* msg = "failed to generate predicate for intrinsic";
448 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
449 } else {
450 // Root compile
451 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
452 vmIntrinsics::name_at(intrinsic_id()),
453 (is_virtual() ? " (virtual)" : ""), bci);
454 }
455 }
456 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
457 return NULL;
458 }
459
460 bool LibraryCallKit::try_to_inline(int predicate) {
461 // Handle symbolic names for otherwise undistinguished boolean switches:
462 const bool is_store = true;
463 const bool is_native_ptr = true;
464 const bool is_static = true;
465 const bool is_volatile = true;
466
467 if (!jvms()->has_method()) {
468 // Root JVMState has a null method.
469 assert(map()->memory()->Opcode() == Op_Parm, "");
470 // Insert the memory aliasing node
471 set_all_memory(reset_memory());
472 }
473 assert(merged_memory(), "");
474
475
476 switch (intrinsic_id()) {
477 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
478 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
479 case vmIntrinsics::_getClass: return inline_native_getClass();
480
481 case vmIntrinsics::_dsin:
482 case vmIntrinsics::_dcos:
483 case vmIntrinsics::_dtan:
484 case vmIntrinsics::_dabs:
485 case vmIntrinsics::_datan2:
486 case vmIntrinsics::_dsqrt:
487 case vmIntrinsics::_dexp:
488 case vmIntrinsics::_dlog:
489 case vmIntrinsics::_dlog10:
490 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
491
492 case vmIntrinsics::_min:
493 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
494
495 case vmIntrinsics::_notify:
496 case vmIntrinsics::_notifyAll:
497 if (InlineNotify) {
498 return inline_notify(intrinsic_id());
499 }
500 return false;
501
502 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
503 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
504 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
505 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
506 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
507 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
508 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
509 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
510 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
511 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
512 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
513 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
514
515 case vmIntrinsics::_arraycopy: return inline_arraycopy();
516
517 case vmIntrinsics::_compareTo: return inline_string_compareTo();
518 case vmIntrinsics::_indexOf: return inline_string_indexOf();
519 case vmIntrinsics::_equals: return inline_string_equals();
520
521 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
522 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
523 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
524 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
525 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
526 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
527 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
528 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
529 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
530 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
531 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
532 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
533 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
534 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
535 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
536 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
537 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
538 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
539
540 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
541 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
542 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
543 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
544 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
545 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
546 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
547 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
548
549 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
550 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
551 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
552 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
553 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
554 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
555 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
556 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
557
558 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
559 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
560 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
561 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
562 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
563 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
564 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
565 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
566 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
567
568 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
569 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
570 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
571 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
572 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
573 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
574 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
575 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
576 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
577
578 case vmIntrinsics::_getShortUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
579 case vmIntrinsics::_getCharUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
580 case vmIntrinsics::_getIntUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
581 case vmIntrinsics::_getLongUnaligned: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
582
583 case vmIntrinsics::_putShortUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
584 case vmIntrinsics::_putCharUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
585 case vmIntrinsics::_putIntUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
586 case vmIntrinsics::_putLongUnaligned: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
587
588 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
589 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
590 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
591
592 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
593 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
594 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
595
596 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
597 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
598 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
599 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
600 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
601
602 case vmIntrinsics::_loadFence:
603 case vmIntrinsics::_storeFence:
604 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
605
606 case vmIntrinsics::_currentThread: return inline_native_currentThread();
607 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
608
609 #ifdef TRACE_HAVE_INTRINSICS
610 case vmIntrinsics::_classID: return inline_native_classID();
611 case vmIntrinsics::_threadID: return inline_native_threadID();
612 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
613 #endif
614 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
615 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
616 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
617 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
618 case vmIntrinsics::_newArray: return inline_native_newArray();
619 case vmIntrinsics::_getLength: return inline_native_getLength();
620 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
621 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
622 case vmIntrinsics::_equalsC: return inline_array_equals();
623 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
624
625 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
626
627 case vmIntrinsics::_isInstance:
628 case vmIntrinsics::_getModifiers:
629 case vmIntrinsics::_isInterface:
630 case vmIntrinsics::_isArray:
631 case vmIntrinsics::_isPrimitive:
632 case vmIntrinsics::_getSuperclass:
633 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
634
635 case vmIntrinsics::_floatToRawIntBits:
636 case vmIntrinsics::_floatToIntBits:
637 case vmIntrinsics::_intBitsToFloat:
638 case vmIntrinsics::_doubleToRawLongBits:
639 case vmIntrinsics::_doubleToLongBits:
640 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
641
642 case vmIntrinsics::_numberOfLeadingZeros_i:
643 case vmIntrinsics::_numberOfLeadingZeros_l:
644 case vmIntrinsics::_numberOfTrailingZeros_i:
645 case vmIntrinsics::_numberOfTrailingZeros_l:
646 case vmIntrinsics::_bitCount_i:
647 case vmIntrinsics::_bitCount_l:
648 case vmIntrinsics::_reverseBytes_i:
649 case vmIntrinsics::_reverseBytes_l:
650 case vmIntrinsics::_reverseBytes_s:
651 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
652
653 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
654
655 case vmIntrinsics::_Reference_get: return inline_reference_get();
656
657 case vmIntrinsics::_Class_cast: return inline_Class_cast();
658
659 case vmIntrinsics::_aescrypt_encryptBlock:
660 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
661
662 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
663 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
664 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
665
666 case vmIntrinsics::_sha_implCompress:
667 case vmIntrinsics::_sha2_implCompress:
668 case vmIntrinsics::_sha5_implCompress:
669 return inline_sha_implCompress(intrinsic_id());
670
671 case vmIntrinsics::_digestBase_implCompressMB:
672 return inline_digestBase_implCompressMB(predicate);
673
674 case vmIntrinsics::_multiplyToLen:
675 return inline_multiplyToLen();
676
677 case vmIntrinsics::_squareToLen:
678 return inline_squareToLen();
679
680 case vmIntrinsics::_mulAdd:
681 return inline_mulAdd();
682
683 case vmIntrinsics::_montgomeryMultiply:
684 return inline_montgomeryMultiply();
685 case vmIntrinsics::_montgomerySquare:
686 return inline_montgomerySquare();
687
688 case vmIntrinsics::_ghash_processBlocks:
689 return inline_ghash_processBlocks();
690
691 case vmIntrinsics::_encodeISOArray:
692 return inline_encodeISOArray();
693
694 case vmIntrinsics::_updateCRC32:
695 return inline_updateCRC32();
696 case vmIntrinsics::_updateBytesCRC32:
697 return inline_updateBytesCRC32();
698 case vmIntrinsics::_updateByteBufferCRC32:
699 return inline_updateByteBufferCRC32();
700
701 case vmIntrinsics::_updateBytesCRC32C:
702 return inline_updateBytesCRC32C();
703 case vmIntrinsics::_updateDirectByteBufferCRC32C:
704 return inline_updateDirectByteBufferCRC32C();
705
706 case vmIntrinsics::_updateBytesAdler32:
707 return inline_updateBytesAdler32();
708 case vmIntrinsics::_updateByteBufferAdler32:
709 return inline_updateByteBufferAdler32();
710
711 case vmIntrinsics::_profileBoolean:
712 return inline_profileBoolean();
713 case vmIntrinsics::_isCompileConstant:
714 return inline_isCompileConstant();
715
716 default:
717 // If you get here, it may be that someone has added a new intrinsic
718 // to the list in vmSymbols.hpp without implementing it here.
719 #ifndef PRODUCT
720 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
721 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
722 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
723 }
724 #endif
725 return false;
726 }
727 }
728
729 Node* LibraryCallKit::try_to_predicate(int predicate) {
730 if (!jvms()->has_method()) {
731 // Root JVMState has a null method.
732 assert(map()->memory()->Opcode() == Op_Parm, "");
733 // Insert the memory aliasing node
734 set_all_memory(reset_memory());
735 }
736 assert(merged_memory(), "");
737
738 switch (intrinsic_id()) {
739 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
740 return inline_cipherBlockChaining_AESCrypt_predicate(false);
741 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
742 return inline_cipherBlockChaining_AESCrypt_predicate(true);
743 case vmIntrinsics::_digestBase_implCompressMB:
744 return inline_digestBase_implCompressMB_predicate(predicate);
745
746 default:
747 // If you get here, it may be that someone has added a new intrinsic
748 // to the list in vmSymbols.hpp without implementing it here.
749 #ifndef PRODUCT
750 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
751 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
752 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
753 }
754 #endif
755 Node* slow_ctl = control();
756 set_control(top()); // No fast path instrinsic
757 return slow_ctl;
758 }
759 }
760
761 //------------------------------set_result-------------------------------
762 // Helper function for finishing intrinsics.
763 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
764 record_for_igvn(region);
765 set_control(_gvn.transform(region));
766 set_result( _gvn.transform(value));
767 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
768 }
769
770 //------------------------------generate_guard---------------------------
771 // Helper function for generating guarded fast-slow graph structures.
772 // The given 'test', if true, guards a slow path. If the test fails
773 // then a fast path can be taken. (We generally hope it fails.)
774 // In all cases, GraphKit::control() is updated to the fast path.
775 // The returned value represents the control for the slow path.
776 // The return value is never 'top'; it is either a valid control
777 // or NULL if it is obvious that the slow path can never be taken.
778 // Also, if region and the slow control are not NULL, the slow edge
779 // is appended to the region.
780 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
781 if (stopped()) {
782 // Already short circuited.
783 return NULL;
784 }
785
786 // Build an if node and its projections.
787 // If test is true we take the slow path, which we assume is uncommon.
788 if (_gvn.type(test) == TypeInt::ZERO) {
789 // The slow branch is never taken. No need to build this guard.
790 return NULL;
791 }
792
793 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
794
795 Node* if_slow = _gvn.transform(new IfTrueNode(iff));
796 if (if_slow == top()) {
797 // The slow branch is never taken. No need to build this guard.
798 return NULL;
799 }
800
801 if (region != NULL)
802 region->add_req(if_slow);
803
804 Node* if_fast = _gvn.transform(new IfFalseNode(iff));
805 set_control(if_fast);
806
807 return if_slow;
808 }
809
810 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
811 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
812 }
813 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
814 return generate_guard(test, region, PROB_FAIR);
815 }
816
817 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
818 Node* *pos_index) {
819 if (stopped())
820 return NULL; // already stopped
821 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
822 return NULL; // index is already adequately typed
823 Node* cmp_lt = _gvn.transform(new CmpINode(index, intcon(0)));
824 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
825 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
826 if (is_neg != NULL && pos_index != NULL) {
827 // Emulate effect of Parse::adjust_map_after_if.
828 Node* ccast = new CastIINode(index, TypeInt::POS);
829 ccast->set_req(0, control());
830 (*pos_index) = _gvn.transform(ccast);
831 }
832 return is_neg;
833 }
834
835 // Make sure that 'position' is a valid limit index, in [0..length].
836 // There are two equivalent plans for checking this:
837 // A. (offset + copyLength) unsigned<= arrayLength
838 // B. offset <= (arrayLength - copyLength)
839 // We require that all of the values above, except for the sum and
840 // difference, are already known to be non-negative.
841 // Plan A is robust in the face of overflow, if offset and copyLength
842 // are both hugely positive.
843 //
844 // Plan B is less direct and intuitive, but it does not overflow at
845 // all, since the difference of two non-negatives is always
846 // representable. Whenever Java methods must perform the equivalent
847 // check they generally use Plan B instead of Plan A.
848 // For the moment we use Plan A.
849 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
850 Node* subseq_length,
851 Node* array_length,
852 RegionNode* region) {
853 if (stopped())
854 return NULL; // already stopped
855 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
856 if (zero_offset && subseq_length->eqv_uncast(array_length))
857 return NULL; // common case of whole-array copy
858 Node* last = subseq_length;
859 if (!zero_offset) // last += offset
860 last = _gvn.transform(new AddINode(last, offset));
861 Node* cmp_lt = _gvn.transform(new CmpUNode(array_length, last));
862 Node* bol_lt = _gvn.transform(new BoolNode(cmp_lt, BoolTest::lt));
863 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
864 return is_over;
865 }
866
867
868 //--------------------------generate_current_thread--------------------
869 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
870 ciKlass* thread_klass = env()->Thread_klass();
871 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
872 Node* thread = _gvn.transform(new ThreadLocalNode());
873 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
874 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
875 tls_output = thread;
876 return threadObj;
877 }
878
879
880 //------------------------------make_string_method_node------------------------
881 // Helper method for String intrinsic functions. This version is called
882 // with str1 and str2 pointing to String object nodes.
883 //
884 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
885 Node* no_ctrl = NULL;
886
887 // Get start addr of string
888 Node* str1_value = load_String_value(no_ctrl, str1);
889 Node* str1_offset = load_String_offset(no_ctrl, str1);
890 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
891
892 // Get length of string 1
893 Node* str1_len = load_String_length(no_ctrl, str1);
894
895 Node* str2_value = load_String_value(no_ctrl, str2);
896 Node* str2_offset = load_String_offset(no_ctrl, str2);
897 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
898
899 Node* str2_len = NULL;
900 Node* result = NULL;
901
902 switch (opcode) {
903 case Op_StrIndexOf:
904 // Get length of string 2
905 str2_len = load_String_length(no_ctrl, str2);
906
907 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
908 str1_start, str1_len, str2_start, str2_len);
909 break;
910 case Op_StrComp:
911 // Get length of string 2
912 str2_len = load_String_length(no_ctrl, str2);
913
914 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
915 str1_start, str1_len, str2_start, str2_len);
916 break;
917 case Op_StrEquals:
918 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
919 str1_start, str2_start, str1_len);
920 break;
921 default:
922 ShouldNotReachHere();
923 return NULL;
924 }
925
926 // All these intrinsics have checks.
927 C->set_has_split_ifs(true); // Has chance for split-if optimization
928
929 return _gvn.transform(result);
930 }
931
932 // Helper method for String intrinsic functions. This version is called
933 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
934 // to Int nodes containing the lenghts of str1 and str2.
935 //
936 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
937 Node* result = NULL;
938 switch (opcode) {
939 case Op_StrIndexOf:
940 result = new StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
941 str1_start, cnt1, str2_start, cnt2);
942 break;
943 case Op_StrComp:
944 result = new StrCompNode(control(), memory(TypeAryPtr::CHARS),
945 str1_start, cnt1, str2_start, cnt2);
946 break;
947 case Op_StrEquals:
948 result = new StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
949 str1_start, str2_start, cnt1);
950 break;
951 default:
952 ShouldNotReachHere();
953 return NULL;
954 }
955
956 // All these intrinsics have checks.
957 C->set_has_split_ifs(true); // Has chance for split-if optimization
958
959 return _gvn.transform(result);
960 }
961
962 //------------------------------inline_string_compareTo------------------------
963 // public int java.lang.String.compareTo(String anotherString);
964 bool LibraryCallKit::inline_string_compareTo() {
965 Node* receiver = null_check(argument(0));
966 Node* arg = null_check(argument(1));
967 if (stopped()) {
968 return true;
969 }
970 set_result(make_string_method_node(Op_StrComp, receiver, arg));
971 return true;
972 }
973
974 //------------------------------inline_string_equals------------------------
975 bool LibraryCallKit::inline_string_equals() {
976 Node* receiver = null_check_receiver();
977
978 if (ShenandoahVerifyReadsToFromSpace) {
979 receiver = shenandoah_read_barrier(receiver);
980 }
981
982 // NOTE: Do not null check argument for String.equals() because spec
983 // allows to specify NULL as argument.
984 Node* argument = this->argument(1);
985
986 if (ShenandoahVerifyReadsToFromSpace) {
987 argument = shenandoah_read_barrier(argument);
988 }
989
990 if (stopped()) {
991 return true;
992 }
993
994 // paths (plus control) merge
995 RegionNode* region = new RegionNode(5);
996 Node* phi = new PhiNode(region, TypeInt::BOOL);
997
998 // does source == target string?
999 Node* cmp = _gvn.transform(new CmpPNode(receiver, argument));
1000 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1001
1002 Node* if_eq = generate_slow_guard(bol, NULL);
1003 if (if_eq != NULL) {
1004 // receiver == argument
1005 phi->init_req(2, intcon(1));
1006 region->init_req(2, if_eq);
1007 }
1008
1009 // get String klass for instanceOf
1010 ciInstanceKlass* klass = env()->String_klass();
1011
1012 if (!stopped()) {
1013 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1014 Node* cmp = _gvn.transform(new CmpINode(inst, intcon(1)));
1015 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1016
1017 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1018 //instanceOf == true, fallthrough
1019
1020 if (inst_false != NULL) {
1021 phi->init_req(3, intcon(0));
1022 region->init_req(3, inst_false);
1023 }
1024 }
1025
1026 if (!stopped()) {
1027 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1028
1029 // Properly cast the argument to String
1030 argument = _gvn.transform(new CheckCastPPNode(control(), argument, string_type));
1031 // This path is taken only when argument's type is String:NotNull.
1032 argument = cast_not_null(argument, false);
1033
1034 Node* no_ctrl = NULL;
1035
1036 // Get start addr of receiver
1037 Node* receiver_val = load_String_value(no_ctrl, receiver);
1038
1039 if (ShenandoahVerifyReadsToFromSpace) {
1040 receiver_val = shenandoah_read_barrier(receiver_val);
1041 }
1042
1043 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1044 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1045
1046 // Get length of receiver
1047 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1048
1049 // Get start addr of argument
1050 Node* argument_val = load_String_value(no_ctrl, argument);
1051
1052 if (ShenandoahVerifyReadsToFromSpace) {
1053 argument_val = shenandoah_read_barrier(argument_val);
1054 }
1055
1056 Node* argument_offset = load_String_offset(no_ctrl, argument);
1057 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1058
1059 // Get length of argument
1060 Node* argument_cnt = load_String_length(no_ctrl, argument);
1061
1062 // Check for receiver count != argument count
1063 Node* cmp = _gvn.transform(new CmpINode(receiver_cnt, argument_cnt));
1064 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
1065 Node* if_ne = generate_slow_guard(bol, NULL);
1066 if (if_ne != NULL) {
1067 phi->init_req(4, intcon(0));
1068 region->init_req(4, if_ne);
1069 }
1070
1071 // Check for count == 0 is done by assembler code for StrEquals.
1072
1073 if (!stopped()) {
1074 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1075 phi->init_req(1, equals);
1076 region->init_req(1, control());
1077 }
1078 }
1079
1080 // post merge
1081 set_control(_gvn.transform(region));
1082 record_for_igvn(region);
1083
1084 set_result(_gvn.transform(phi));
1085 return true;
1086 }
1087
1088 //------------------------------inline_array_equals----------------------------
1089 bool LibraryCallKit::inline_array_equals() {
1090 Node* arg1 = argument(0);
1091 Node* arg2 = argument(1);
1092
1093 arg1 = shenandoah_read_barrier(arg1);
1094 arg2 = shenandoah_read_barrier(arg2);
1095
1096 set_result(_gvn.transform(new AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1097 return true;
1098 }
1099
1100 // Java version of String.indexOf(constant string)
1101 // class StringDecl {
1102 // StringDecl(char[] ca) {
1103 // offset = 0;
1104 // count = ca.length;
1105 // value = ca;
1106 // }
1107 // int offset;
1108 // int count;
1109 // char[] value;
1110 // }
1111 //
1112 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1113 // int targetOffset, int cache_i, int md2) {
1114 // int cache = cache_i;
1115 // int sourceOffset = string_object.offset;
1116 // int sourceCount = string_object.count;
1117 // int targetCount = target_object.length;
1118 //
1119 // int targetCountLess1 = targetCount - 1;
1120 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1121 //
1122 // char[] source = string_object.value;
1123 // char[] target = target_object;
1124 // int lastChar = target[targetCountLess1];
1125 //
1126 // outer_loop:
1127 // for (int i = sourceOffset; i < sourceEnd; ) {
1128 // int src = source[i + targetCountLess1];
1129 // if (src == lastChar) {
1130 // // With random strings and a 4-character alphabet,
1131 // // reverse matching at this point sets up 0.8% fewer
1132 // // frames, but (paradoxically) makes 0.3% more probes.
1133 // // Since those probes are nearer the lastChar probe,
1134 // // there is may be a net D$ win with reverse matching.
1135 // // But, reversing loop inhibits unroll of inner loop
1136 // // for unknown reason. So, does running outer loop from
1137 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1138 // for (int j = 0; j < targetCountLess1; j++) {
1139 // if (target[targetOffset + j] != source[i+j]) {
1140 // if ((cache & (1 << source[i+j])) == 0) {
1141 // if (md2 < j+1) {
1142 // i += j+1;
1143 // continue outer_loop;
1144 // }
1145 // }
1146 // i += md2;
1147 // continue outer_loop;
1148 // }
1149 // }
1150 // return i - sourceOffset;
1151 // }
1152 // if ((cache & (1 << src)) == 0) {
1153 // i += targetCountLess1;
1154 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1155 // i++;
1156 // }
1157 // return -1;
1158 // }
1159
1160 //------------------------------string_indexOf------------------------
1161 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1162 jint cache_i, jint md2_i) {
1163
1164 Node* no_ctrl = NULL;
1165 float likely = PROB_LIKELY(0.9);
1166 float unlikely = PROB_UNLIKELY(0.9);
1167
1168 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1169
1170 Node* source = load_String_value(no_ctrl, string_object);
1171 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1172 Node* sourceCount = load_String_length(no_ctrl, string_object);
1173
1174 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1175 jint target_length = target_array->length();
1176 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1177 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1178
1179 // String.value field is known to be @Stable.
1180 if (UseImplicitStableValues) {
1181 target = cast_array_to_stable(target, target_type);
1182 }
1183
1184 IdealKit kit(this, false, true);
1185 #define __ kit.
1186 Node* zero = __ ConI(0);
1187 Node* one = __ ConI(1);
1188 Node* cache = __ ConI(cache_i);
1189 Node* md2 = __ ConI(md2_i);
1190 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1191 Node* targetCountLess1 = __ ConI(target_length - 1);
1192 Node* targetOffset = __ ConI(targetOffset_i);
1193 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1194
1195 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1196 Node* outer_loop = __ make_label(2 /* goto */);
1197 Node* return_ = __ make_label(1);
1198
1199 __ set(rtn,__ ConI(-1));
1200 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1201 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1202 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1203 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1204 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1205 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1206 Node* tpj = __ AddI(targetOffset, __ value(j));
1207 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1208 Node* ipj = __ AddI(__ value(i), __ value(j));
1209 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1210 __ if_then(targ, BoolTest::ne, src2); {
1211 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1212 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1213 __ increment(i, __ AddI(__ value(j), one));
1214 __ goto_(outer_loop);
1215 } __ end_if(); __ dead(j);
1216 }__ end_if(); __ dead(j);
1217 __ increment(i, md2);
1218 __ goto_(outer_loop);
1219 }__ end_if();
1220 __ increment(j, one);
1221 }__ end_loop(); __ dead(j);
1222 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1223 __ goto_(return_);
1224 }__ end_if();
1225 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1226 __ increment(i, targetCountLess1);
1227 }__ end_if();
1228 __ increment(i, one);
1229 __ bind(outer_loop);
1230 }__ end_loop(); __ dead(i);
1231 __ bind(return_);
1232
1233 // Final sync IdealKit and GraphKit.
1234 final_sync(kit);
1235 Node* result = __ value(rtn);
1236 #undef __
1237 C->set_has_loops(true);
1238 return result;
1239 }
1240
1241 //------------------------------inline_string_indexOf------------------------
1242 bool LibraryCallKit::inline_string_indexOf() {
1243 Node* receiver = argument(0);
1244 Node* arg = argument(1);
1245
1246 Node* result;
1247 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1248 UseSSE42Intrinsics) {
1249 // Generate SSE4.2 version of indexOf
1250 // We currently only have match rules that use SSE4.2
1251
1252 receiver = null_check(receiver);
1253 arg = null_check(arg);
1254 if (stopped()) {
1255 return true;
1256 }
1257
1258 // Make the merge point
1259 RegionNode* result_rgn = new RegionNode(4);
1260 Node* result_phi = new PhiNode(result_rgn, TypeInt::INT);
1261 Node* no_ctrl = NULL;
1262
1263 // Get start addr of source string
1264 Node* source = load_String_value(no_ctrl, receiver);
1265 Node* source_offset = load_String_offset(no_ctrl, receiver);
1266 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1267
1268 // Get length of source string
1269 Node* source_cnt = load_String_length(no_ctrl, receiver);
1270
1271 // Get start addr of substring
1272 Node* substr = load_String_value(no_ctrl, arg);
1273 Node* substr_offset = load_String_offset(no_ctrl, arg);
1274 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1275
1276 // Get length of source string
1277 Node* substr_cnt = load_String_length(no_ctrl, arg);
1278
1279 // Check for substr count > string count
1280 Node* cmp = _gvn.transform(new CmpINode(substr_cnt, source_cnt));
1281 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::gt));
1282 Node* if_gt = generate_slow_guard(bol, NULL);
1283 if (if_gt != NULL) {
1284 result_phi->init_req(2, intcon(-1));
1285 result_rgn->init_req(2, if_gt);
1286 }
1287
1288 if (!stopped()) {
1289 // Check for substr count == 0
1290 cmp = _gvn.transform(new CmpINode(substr_cnt, intcon(0)));
1291 bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
1292 Node* if_zero = generate_slow_guard(bol, NULL);
1293 if (if_zero != NULL) {
1294 result_phi->init_req(3, intcon(0));
1295 result_rgn->init_req(3, if_zero);
1296 }
1297 }
1298
1299 if (!stopped()) {
1300 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1301 result_phi->init_req(1, result);
1302 result_rgn->init_req(1, control());
1303 }
1304 set_control(_gvn.transform(result_rgn));
1305 record_for_igvn(result_rgn);
1306 result = _gvn.transform(result_phi);
1307
1308 } else { // Use LibraryCallKit::string_indexOf
1309 // don't intrinsify if argument isn't a constant string.
1310 if (!arg->is_Con()) {
1311 return false;
1312 }
1313 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1314 if (str_type == NULL) {
1315 return false;
1316 }
1317 ciInstanceKlass* klass = env()->String_klass();
1318 ciObject* str_const = str_type->const_oop();
1319 if (str_const == NULL || str_const->klass() != klass) {
1320 return false;
1321 }
1322 ciInstance* str = str_const->as_instance();
1323 assert(str != NULL, "must be instance");
1324
1325 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1326 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1327
1328 int o;
1329 int c;
1330 if (java_lang_String::has_offset_field()) {
1331 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1332 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1333 } else {
1334 o = 0;
1335 c = pat->length();
1336 }
1337
1338 // constant strings have no offset and count == length which
1339 // simplifies the resulting code somewhat so lets optimize for that.
1340 if (o != 0 || c != pat->length()) {
1341 return false;
1342 }
1343
1344 receiver = null_check(receiver, T_OBJECT);
1345 // NOTE: No null check on the argument is needed since it's a constant String oop.
1346 if (stopped()) {
1347 return true;
1348 }
1349
1350 // The null string as a pattern always returns 0 (match at beginning of string)
1351 if (c == 0) {
1352 set_result(intcon(0));
1353 return true;
1354 }
1355
1356 // Generate default indexOf
1357 jchar lastChar = pat->char_at(o + (c - 1));
1358 int cache = 0;
1359 int i;
1360 for (i = 0; i < c - 1; i++) {
1361 assert(i < pat->length(), "out of range");
1362 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1363 }
1364
1365 int md2 = c;
1366 for (i = 0; i < c - 1; i++) {
1367 assert(i < pat->length(), "out of range");
1368 if (pat->char_at(o + i) == lastChar) {
1369 md2 = (c - 1) - i;
1370 }
1371 }
1372
1373 result = string_indexOf(receiver, pat, o, cache, md2);
1374 }
1375 set_result(result);
1376 return true;
1377 }
1378
1379 //--------------------------round_double_node--------------------------------
1380 // Round a double node if necessary.
1381 Node* LibraryCallKit::round_double_node(Node* n) {
1382 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1383 n = _gvn.transform(new RoundDoubleNode(0, n));
1384 return n;
1385 }
1386
1387 //------------------------------inline_math-----------------------------------
1388 // public static double Math.abs(double)
1389 // public static double Math.sqrt(double)
1390 // public static double Math.log(double)
1391 // public static double Math.log10(double)
1392 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1393 Node* arg = round_double_node(argument(0));
1394 Node* n;
1395 switch (id) {
1396 case vmIntrinsics::_dabs: n = new AbsDNode( arg); break;
1397 case vmIntrinsics::_dsqrt: n = new SqrtDNode(C, control(), arg); break;
1398 case vmIntrinsics::_dlog: n = new LogDNode(C, control(), arg); break;
1399 case vmIntrinsics::_dlog10: n = new Log10DNode(C, control(), arg); break;
1400 default: fatal_unexpected_iid(id); break;
1401 }
1402 set_result(_gvn.transform(n));
1403 return true;
1404 }
1405
1406 //------------------------------inline_trig----------------------------------
1407 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1408 // argument reduction which will turn into a fast/slow diamond.
1409 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1410 Node* arg = round_double_node(argument(0));
1411 Node* n = NULL;
1412
1413 switch (id) {
1414 case vmIntrinsics::_dsin: n = new SinDNode(C, control(), arg); break;
1415 case vmIntrinsics::_dcos: n = new CosDNode(C, control(), arg); break;
1416 case vmIntrinsics::_dtan: n = new TanDNode(C, control(), arg); break;
1417 default: fatal_unexpected_iid(id); break;
1418 }
1419 n = _gvn.transform(n);
1420
1421 // Rounding required? Check for argument reduction!
1422 if (Matcher::strict_fp_requires_explicit_rounding) {
1423 static const double pi_4 = 0.7853981633974483;
1424 static const double neg_pi_4 = -0.7853981633974483;
1425 // pi/2 in 80-bit extended precision
1426 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1427 // -pi/2 in 80-bit extended precision
1428 // 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};
1429 // Cutoff value for using this argument reduction technique
1430 //static const double pi_2_minus_epsilon = 1.564660403643354;
1431 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1432
1433 // Pseudocode for sin:
1434 // if (x <= Math.PI / 4.0) {
1435 // if (x >= -Math.PI / 4.0) return fsin(x);
1436 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1437 // } else {
1438 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1439 // }
1440 // return StrictMath.sin(x);
1441
1442 // Pseudocode for cos:
1443 // if (x <= Math.PI / 4.0) {
1444 // if (x >= -Math.PI / 4.0) return fcos(x);
1445 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1446 // } else {
1447 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1448 // }
1449 // return StrictMath.cos(x);
1450
1451 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1452 // requires a special machine instruction to load it. Instead we'll try
1453 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1454 // probably do the math inside the SIN encoding.
1455
1456 // Make the merge point
1457 RegionNode* r = new RegionNode(3);
1458 Node* phi = new PhiNode(r, Type::DOUBLE);
1459
1460 // Flatten arg so we need only 1 test
1461 Node *abs = _gvn.transform(new AbsDNode(arg));
1462 // Node for PI/4 constant
1463 Node *pi4 = makecon(TypeD::make(pi_4));
1464 // Check PI/4 : abs(arg)
1465 Node *cmp = _gvn.transform(new CmpDNode(pi4,abs));
1466 // Check: If PI/4 < abs(arg) then go slow
1467 Node *bol = _gvn.transform(new BoolNode( cmp, BoolTest::lt ));
1468 // Branch either way
1469 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1470 set_control(opt_iff(r,iff));
1471
1472 // Set fast path result
1473 phi->init_req(2, n);
1474
1475 // Slow path - non-blocking leaf call
1476 Node* call = NULL;
1477 switch (id) {
1478 case vmIntrinsics::_dsin:
1479 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1480 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1481 "Sin", NULL, arg, top());
1482 break;
1483 case vmIntrinsics::_dcos:
1484 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1485 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1486 "Cos", NULL, arg, top());
1487 break;
1488 case vmIntrinsics::_dtan:
1489 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1490 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1491 "Tan", NULL, arg, top());
1492 break;
1493 }
1494 assert(control()->in(0) == call, "");
1495 Node* slow_result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1496 r->init_req(1, control());
1497 phi->init_req(1, slow_result);
1498
1499 // Post-merge
1500 set_control(_gvn.transform(r));
1501 record_for_igvn(r);
1502 n = _gvn.transform(phi);
1503
1504 C->set_has_split_ifs(true); // Has chance for split-if optimization
1505 }
1506 set_result(n);
1507 return true;
1508 }
1509
1510 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1511 //-------------------
1512 //result=(result.isNaN())? funcAddr():result;
1513 // Check: If isNaN() by checking result!=result? then either trap
1514 // or go to runtime
1515 Node* cmpisnan = _gvn.transform(new CmpDNode(result, result));
1516 // Build the boolean node
1517 Node* bolisnum = _gvn.transform(new BoolNode(cmpisnan, BoolTest::eq));
1518
1519 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1520 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1521 // The pow or exp intrinsic returned a NaN, which requires a call
1522 // to the runtime. Recompile with the runtime call.
1523 uncommon_trap(Deoptimization::Reason_intrinsic,
1524 Deoptimization::Action_make_not_entrant);
1525 }
1526 return result;
1527 } else {
1528 // If this inlining ever returned NaN in the past, we compile a call
1529 // to the runtime to properly handle corner cases
1530
1531 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1532 Node* if_slow = _gvn.transform(new IfFalseNode(iff));
1533 Node* if_fast = _gvn.transform(new IfTrueNode(iff));
1534
1535 if (!if_slow->is_top()) {
1536 RegionNode* result_region = new RegionNode(3);
1537 PhiNode* result_val = new PhiNode(result_region, Type::DOUBLE);
1538
1539 result_region->init_req(1, if_fast);
1540 result_val->init_req(1, result);
1541
1542 set_control(if_slow);
1543
1544 const TypePtr* no_memory_effects = NULL;
1545 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1546 no_memory_effects,
1547 x, top(), y, y ? top() : NULL);
1548 Node* value = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+0));
1549 #ifdef ASSERT
1550 Node* value_top = _gvn.transform(new ProjNode(rt, TypeFunc::Parms+1));
1551 assert(value_top == top(), "second value must be top");
1552 #endif
1553
1554 result_region->init_req(2, control());
1555 result_val->init_req(2, value);
1556 set_control(_gvn.transform(result_region));
1557 return _gvn.transform(result_val);
1558 } else {
1559 return result;
1560 }
1561 }
1562 }
1563
1564 //------------------------------inline_exp-------------------------------------
1565 // Inline exp instructions, if possible. The Intel hardware only misses
1566 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1567 bool LibraryCallKit::inline_exp() {
1568 Node* arg = round_double_node(argument(0));
1569 Node* n = _gvn.transform(new ExpDNode(C, control(), arg));
1570
1571 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1572 set_result(n);
1573
1574 C->set_has_split_ifs(true); // Has chance for split-if optimization
1575 return true;
1576 }
1577
1578 //------------------------------inline_pow-------------------------------------
1579 // Inline power instructions, if possible.
1580 bool LibraryCallKit::inline_pow() {
1581 // Pseudocode for pow
1582 // if (y == 2) {
1583 // return x * x;
1584 // } else {
1585 // if (x <= 0.0) {
1586 // long longy = (long)y;
1587 // if ((double)longy == y) { // if y is long
1588 // if (y + 1 == y) longy = 0; // huge number: even
1589 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1590 // } else {
1591 // result = NaN;
1592 // }
1593 // } else {
1594 // result = DPow(x,y);
1595 // }
1596 // if (result != result)? {
1597 // result = uncommon_trap() or runtime_call();
1598 // }
1599 // return result;
1600 // }
1601
1602 Node* x = round_double_node(argument(0));
1603 Node* y = round_double_node(argument(2));
1604
1605 Node* result = NULL;
1606
1607 Node* const_two_node = makecon(TypeD::make(2.0));
1608 Node* cmp_node = _gvn.transform(new CmpDNode(y, const_two_node));
1609 Node* bool_node = _gvn.transform(new BoolNode(cmp_node, BoolTest::eq));
1610 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1611 Node* if_true = _gvn.transform(new IfTrueNode(if_node));
1612 Node* if_false = _gvn.transform(new IfFalseNode(if_node));
1613
1614 RegionNode* region_node = new RegionNode(3);
1615 region_node->init_req(1, if_true);
1616
1617 Node* phi_node = new PhiNode(region_node, Type::DOUBLE);
1618 // special case for x^y where y == 2, we can convert it to x * x
1619 phi_node->init_req(1, _gvn.transform(new MulDNode(x, x)));
1620
1621 // set control to if_false since we will now process the false branch
1622 set_control(if_false);
1623
1624 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1625 // Short form: skip the fancy tests and just check for NaN result.
1626 result = _gvn.transform(new PowDNode(C, control(), x, y));
1627 } else {
1628 // If this inlining ever returned NaN in the past, include all
1629 // checks + call to the runtime.
1630
1631 // Set the merge point for If node with condition of (x <= 0.0)
1632 // There are four possible paths to region node and phi node
1633 RegionNode *r = new RegionNode(4);
1634 Node *phi = new PhiNode(r, Type::DOUBLE);
1635
1636 // Build the first if node: if (x <= 0.0)
1637 // Node for 0 constant
1638 Node *zeronode = makecon(TypeD::ZERO);
1639 // Check x:0
1640 Node *cmp = _gvn.transform(new CmpDNode(x, zeronode));
1641 // Check: If (x<=0) then go complex path
1642 Node *bol1 = _gvn.transform(new BoolNode( cmp, BoolTest::le ));
1643 // Branch either way
1644 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1645 // Fast path taken; set region slot 3
1646 Node *fast_taken = _gvn.transform(new IfFalseNode(if1));
1647 r->init_req(3,fast_taken); // Capture fast-control
1648
1649 // Fast path not-taken, i.e. slow path
1650 Node *complex_path = _gvn.transform(new IfTrueNode(if1));
1651
1652 // Set fast path result
1653 Node *fast_result = _gvn.transform(new PowDNode(C, control(), x, y));
1654 phi->init_req(3, fast_result);
1655
1656 // Complex path
1657 // Build the second if node (if y is long)
1658 // Node for (long)y
1659 Node *longy = _gvn.transform(new ConvD2LNode(y));
1660 // Node for (double)((long) y)
1661 Node *doublelongy= _gvn.transform(new ConvL2DNode(longy));
1662 // Check (double)((long) y) : y
1663 Node *cmplongy= _gvn.transform(new CmpDNode(doublelongy, y));
1664 // Check if (y isn't long) then go to slow path
1665
1666 Node *bol2 = _gvn.transform(new BoolNode( cmplongy, BoolTest::ne ));
1667 // Branch either way
1668 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1669 Node* ylong_path = _gvn.transform(new IfFalseNode(if2));
1670
1671 Node *slow_path = _gvn.transform(new IfTrueNode(if2));
1672
1673 // Calculate DPow(abs(x), y)*(1 & (long)y)
1674 // Node for constant 1
1675 Node *conone = longcon(1);
1676 // 1& (long)y
1677 Node *signnode= _gvn.transform(new AndLNode(conone, longy));
1678
1679 // A huge number is always even. Detect a huge number by checking
1680 // if y + 1 == y and set integer to be tested for parity to 0.
1681 // Required for corner case:
1682 // (long)9.223372036854776E18 = max_jlong
1683 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1684 // max_jlong is odd but 9.223372036854776E18 is even
1685 Node* yplus1 = _gvn.transform(new AddDNode(y, makecon(TypeD::make(1))));
1686 Node *cmpyplus1= _gvn.transform(new CmpDNode(yplus1, y));
1687 Node *bolyplus1 = _gvn.transform(new BoolNode( cmpyplus1, BoolTest::eq ));
1688 Node* correctedsign = NULL;
1689 if (ConditionalMoveLimit != 0) {
1690 correctedsign = _gvn.transform(CMoveNode::make(NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1691 } else {
1692 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1693 RegionNode *r = new RegionNode(3);
1694 Node *phi = new PhiNode(r, TypeLong::LONG);
1695 r->init_req(1, _gvn.transform(new IfFalseNode(ifyplus1)));
1696 r->init_req(2, _gvn.transform(new IfTrueNode(ifyplus1)));
1697 phi->init_req(1, signnode);
1698 phi->init_req(2, longcon(0));
1699 correctedsign = _gvn.transform(phi);
1700 ylong_path = _gvn.transform(r);
1701 record_for_igvn(r);
1702 }
1703
1704 // zero node
1705 Node *conzero = longcon(0);
1706 // Check (1&(long)y)==0?
1707 Node *cmpeq1 = _gvn.transform(new CmpLNode(correctedsign, conzero));
1708 // Check if (1&(long)y)!=0?, if so the result is negative
1709 Node *bol3 = _gvn.transform(new BoolNode( cmpeq1, BoolTest::ne ));
1710 // abs(x)
1711 Node *absx=_gvn.transform(new AbsDNode(x));
1712 // abs(x)^y
1713 Node *absxpowy = _gvn.transform(new PowDNode(C, control(), absx, y));
1714 // -abs(x)^y
1715 Node *negabsxpowy = _gvn.transform(new NegDNode (absxpowy));
1716 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1717 Node *signresult = NULL;
1718 if (ConditionalMoveLimit != 0) {
1719 signresult = _gvn.transform(CMoveNode::make(NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1720 } else {
1721 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1722 RegionNode *r = new RegionNode(3);
1723 Node *phi = new PhiNode(r, Type::DOUBLE);
1724 r->init_req(1, _gvn.transform(new IfFalseNode(ifyeven)));
1725 r->init_req(2, _gvn.transform(new IfTrueNode(ifyeven)));
1726 phi->init_req(1, absxpowy);
1727 phi->init_req(2, negabsxpowy);
1728 signresult = _gvn.transform(phi);
1729 ylong_path = _gvn.transform(r);
1730 record_for_igvn(r);
1731 }
1732 // Set complex path fast result
1733 r->init_req(2, ylong_path);
1734 phi->init_req(2, signresult);
1735
1736 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1737 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1738 r->init_req(1,slow_path);
1739 phi->init_req(1,slow_result);
1740
1741 // Post merge
1742 set_control(_gvn.transform(r));
1743 record_for_igvn(r);
1744 result = _gvn.transform(phi);
1745 }
1746
1747 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1748
1749 // control from finish_pow_exp is now input to the region node
1750 region_node->set_req(2, control());
1751 // the result from finish_pow_exp is now input to the phi node
1752 phi_node->init_req(2, result);
1753 set_control(_gvn.transform(region_node));
1754 record_for_igvn(region_node);
1755 set_result(_gvn.transform(phi_node));
1756
1757 C->set_has_split_ifs(true); // Has chance for split-if optimization
1758 return true;
1759 }
1760
1761 //------------------------------runtime_math-----------------------------
1762 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1763 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1764 "must be (DD)D or (D)D type");
1765
1766 // Inputs
1767 Node* a = round_double_node(argument(0));
1768 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1769
1770 const TypePtr* no_memory_effects = NULL;
1771 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1772 no_memory_effects,
1773 a, top(), b, b ? top() : NULL);
1774 Node* value = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+0));
1775 #ifdef ASSERT
1776 Node* value_top = _gvn.transform(new ProjNode(trig, TypeFunc::Parms+1));
1777 assert(value_top == top(), "second value must be top");
1778 #endif
1779
1780 set_result(value);
1781 return true;
1782 }
1783
1784 //------------------------------inline_math_native-----------------------------
1785 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1786 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1787 switch (id) {
1788 // These intrinsics are not properly supported on all hardware
1789 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1790 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1791 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1792 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1793 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1794 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1795
1796 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1797 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1798 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1799 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1800
1801 // These intrinsics are supported on all hardware
1802 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1803 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1804
1805 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1806 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1807 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1808 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1809 #undef FN_PTR
1810
1811 // These intrinsics are not yet correctly implemented
1812 case vmIntrinsics::_datan2:
1813 return false;
1814
1815 default:
1816 fatal_unexpected_iid(id);
1817 return false;
1818 }
1819 }
1820
1821 static bool is_simple_name(Node* n) {
1822 return (n->req() == 1 // constant
1823 || (n->is_Type() && n->as_Type()->type()->singleton())
1824 || n->is_Proj() // parameter or return value
1825 || n->is_Phi() // local of some sort
1826 );
1827 }
1828
1829 //----------------------------inline_notify-----------------------------------*
1830 bool LibraryCallKit::inline_notify(vmIntrinsics::ID id) {
1831 const TypeFunc* ftype = OptoRuntime::monitor_notify_Type();
1832 address func;
1833 if (id == vmIntrinsics::_notify) {
1834 func = OptoRuntime::monitor_notify_Java();
1835 } else {
1836 func = OptoRuntime::monitor_notifyAll_Java();
1837 }
1838 Node* call = make_runtime_call(RC_NO_LEAF, ftype, func, NULL, TypeRawPtr::BOTTOM, argument(0));
1839 make_slow_call_ex(call, env()->Throwable_klass(), false);
1840 return true;
1841 }
1842
1843
1844 //----------------------------inline_min_max-----------------------------------
1845 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1846 set_result(generate_min_max(id, argument(0), argument(1)));
1847 return true;
1848 }
1849
1850 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1851 Node* bol = _gvn.transform( new BoolNode(test, BoolTest::overflow) );
1852 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1853 Node* fast_path = _gvn.transform( new IfFalseNode(check));
1854 Node* slow_path = _gvn.transform( new IfTrueNode(check) );
1855
1856 {
1857 PreserveJVMState pjvms(this);
1858 PreserveReexecuteState preexecs(this);
1859 jvms()->set_should_reexecute(true);
1860
1861 set_control(slow_path);
1862 set_i_o(i_o());
1863
1864 uncommon_trap(Deoptimization::Reason_intrinsic,
1865 Deoptimization::Action_none);
1866 }
1867
1868 set_control(fast_path);
1869 set_result(math);
1870 }
1871
1872 template <typename OverflowOp>
1873 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1874 typedef typename OverflowOp::MathOp MathOp;
1875
1876 MathOp* mathOp = new MathOp(arg1, arg2);
1877 Node* operation = _gvn.transform( mathOp );
1878 Node* ofcheck = _gvn.transform( new OverflowOp(arg1, arg2) );
1879 inline_math_mathExact(operation, ofcheck);
1880 return true;
1881 }
1882
1883 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
1884 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
1885 }
1886
1887 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
1888 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
1889 }
1890
1891 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
1892 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
1893 }
1894
1895 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
1896 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
1897 }
1898
1899 bool LibraryCallKit::inline_math_negateExactI() {
1900 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
1901 }
1902
1903 bool LibraryCallKit::inline_math_negateExactL() {
1904 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
1905 }
1906
1907 bool LibraryCallKit::inline_math_multiplyExactI() {
1908 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
1909 }
1910
1911 bool LibraryCallKit::inline_math_multiplyExactL() {
1912 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
1913 }
1914
1915 Node*
1916 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1917 // These are the candidate return value:
1918 Node* xvalue = x0;
1919 Node* yvalue = y0;
1920
1921 if (xvalue == yvalue) {
1922 return xvalue;
1923 }
1924
1925 bool want_max = (id == vmIntrinsics::_max);
1926
1927 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1928 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1929 if (txvalue == NULL || tyvalue == NULL) return top();
1930 // This is not really necessary, but it is consistent with a
1931 // hypothetical MaxINode::Value method:
1932 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1933
1934 // %%% This folding logic should (ideally) be in a different place.
1935 // Some should be inside IfNode, and there to be a more reliable
1936 // transformation of ?: style patterns into cmoves. We also want
1937 // more powerful optimizations around cmove and min/max.
1938
1939 // Try to find a dominating comparison of these guys.
1940 // It can simplify the index computation for Arrays.copyOf
1941 // and similar uses of System.arraycopy.
1942 // First, compute the normalized version of CmpI(x, y).
1943 int cmp_op = Op_CmpI;
1944 Node* xkey = xvalue;
1945 Node* ykey = yvalue;
1946 Node* ideal_cmpxy = _gvn.transform(new CmpINode(xkey, ykey));
1947 if (ideal_cmpxy->is_Cmp()) {
1948 // E.g., if we have CmpI(length - offset, count),
1949 // it might idealize to CmpI(length, count + offset)
1950 cmp_op = ideal_cmpxy->Opcode();
1951 xkey = ideal_cmpxy->in(1);
1952 ykey = ideal_cmpxy->in(2);
1953 }
1954
1955 // Start by locating any relevant comparisons.
1956 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1957 Node* cmpxy = NULL;
1958 Node* cmpyx = NULL;
1959 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1960 Node* cmp = start_from->fast_out(k);
1961 if (cmp->outcnt() > 0 && // must have prior uses
1962 cmp->in(0) == NULL && // must be context-independent
1963 cmp->Opcode() == cmp_op) { // right kind of compare
1964 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1965 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1966 }
1967 }
1968
1969 const int NCMPS = 2;
1970 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1971 int cmpn;
1972 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1973 if (cmps[cmpn] != NULL) break; // find a result
1974 }
1975 if (cmpn < NCMPS) {
1976 // Look for a dominating test that tells us the min and max.
1977 int depth = 0; // Limit search depth for speed
1978 Node* dom = control();
1979 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1980 if (++depth >= 100) break;
1981 Node* ifproj = dom;
1982 if (!ifproj->is_Proj()) continue;
1983 Node* iff = ifproj->in(0);
1984 if (!iff->is_If()) continue;
1985 Node* bol = iff->in(1);
1986 if (!bol->is_Bool()) continue;
1987 Node* cmp = bol->in(1);
1988 if (cmp == NULL) continue;
1989 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1990 if (cmps[cmpn] == cmp) break;
1991 if (cmpn == NCMPS) continue;
1992 BoolTest::mask btest = bol->as_Bool()->_test._test;
1993 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1994 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1995 // At this point, we know that 'x btest y' is true.
1996 switch (btest) {
1997 case BoolTest::eq:
1998 // They are proven equal, so we can collapse the min/max.
1999 // Either value is the answer. Choose the simpler.
2000 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2001 return yvalue;
2002 return xvalue;
2003 case BoolTest::lt: // x < y
2004 case BoolTest::le: // x <= y
2005 return (want_max ? yvalue : xvalue);
2006 case BoolTest::gt: // x > y
2007 case BoolTest::ge: // x >= y
2008 return (want_max ? xvalue : yvalue);
2009 }
2010 }
2011 }
2012
2013 // We failed to find a dominating test.
2014 // Let's pick a test that might GVN with prior tests.
2015 Node* best_bol = NULL;
2016 BoolTest::mask best_btest = BoolTest::illegal;
2017 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2018 Node* cmp = cmps[cmpn];
2019 if (cmp == NULL) continue;
2020 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2021 Node* bol = cmp->fast_out(j);
2022 if (!bol->is_Bool()) continue;
2023 BoolTest::mask btest = bol->as_Bool()->_test._test;
2024 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2025 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2026 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2027 best_bol = bol->as_Bool();
2028 best_btest = btest;
2029 }
2030 }
2031 }
2032
2033 Node* answer_if_true = NULL;
2034 Node* answer_if_false = NULL;
2035 switch (best_btest) {
2036 default:
2037 if (cmpxy == NULL)
2038 cmpxy = ideal_cmpxy;
2039 best_bol = _gvn.transform(new BoolNode(cmpxy, BoolTest::lt));
2040 // and fall through:
2041 case BoolTest::lt: // x < y
2042 case BoolTest::le: // x <= y
2043 answer_if_true = (want_max ? yvalue : xvalue);
2044 answer_if_false = (want_max ? xvalue : yvalue);
2045 break;
2046 case BoolTest::gt: // x > y
2047 case BoolTest::ge: // x >= y
2048 answer_if_true = (want_max ? xvalue : yvalue);
2049 answer_if_false = (want_max ? yvalue : xvalue);
2050 break;
2051 }
2052
2053 jint hi, lo;
2054 if (want_max) {
2055 // We can sharpen the minimum.
2056 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2057 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2058 } else {
2059 // We can sharpen the maximum.
2060 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2061 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2062 }
2063
2064 // Use a flow-free graph structure, to avoid creating excess control edges
2065 // which could hinder other optimizations.
2066 // Since Math.min/max is often used with arraycopy, we want
2067 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2068 Node* cmov = CMoveNode::make(NULL, best_bol,
2069 answer_if_false, answer_if_true,
2070 TypeInt::make(lo, hi, widen));
2071
2072 return _gvn.transform(cmov);
2073
2074 /*
2075 // This is not as desirable as it may seem, since Min and Max
2076 // nodes do not have a full set of optimizations.
2077 // And they would interfere, anyway, with 'if' optimizations
2078 // and with CMoveI canonical forms.
2079 switch (id) {
2080 case vmIntrinsics::_min:
2081 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2082 case vmIntrinsics::_max:
2083 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2084 default:
2085 ShouldNotReachHere();
2086 }
2087 */
2088 }
2089
2090 inline int
2091 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2092 const TypePtr* base_type = TypePtr::NULL_PTR;
2093 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2094 if (base_type == NULL) {
2095 // Unknown type.
2096 return Type::AnyPtr;
2097 } else if (base_type == TypePtr::NULL_PTR) {
2098 // Since this is a NULL+long form, we have to switch to a rawptr.
2099 base = _gvn.transform(new CastX2PNode(offset));
2100 offset = MakeConX(0);
2101 return Type::RawPtr;
2102 } else if (base_type->base() == Type::RawPtr) {
2103 return Type::RawPtr;
2104 } else if (base_type->isa_oopptr()) {
2105 // Base is never null => always a heap address.
2106 if (base_type->ptr() == TypePtr::NotNull) {
2107 return Type::OopPtr;
2108 }
2109 // Offset is small => always a heap address.
2110 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2111 if (offset_type != NULL &&
2112 base_type->offset() == 0 && // (should always be?)
2113 offset_type->_lo >= 0 &&
2114 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2115 return Type::OopPtr;
2116 }
2117 // Otherwise, it might either be oop+off or NULL+addr.
2118 return Type::AnyPtr;
2119 } else {
2120 // No information:
2121 return Type::AnyPtr;
2122 }
2123 }
2124
2125 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2126 int kind = classify_unsafe_addr(base, offset);
2127 if (kind == Type::RawPtr) {
2128 return basic_plus_adr(top(), base, offset);
2129 } else {
2130 return basic_plus_adr(base, offset);
2131 }
2132 }
2133
2134 //--------------------------inline_number_methods-----------------------------
2135 // inline int Integer.numberOfLeadingZeros(int)
2136 // inline int Long.numberOfLeadingZeros(long)
2137 //
2138 // inline int Integer.numberOfTrailingZeros(int)
2139 // inline int Long.numberOfTrailingZeros(long)
2140 //
2141 // inline int Integer.bitCount(int)
2142 // inline int Long.bitCount(long)
2143 //
2144 // inline char Character.reverseBytes(char)
2145 // inline short Short.reverseBytes(short)
2146 // inline int Integer.reverseBytes(int)
2147 // inline long Long.reverseBytes(long)
2148 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2149 Node* arg = argument(0);
2150 Node* n;
2151 switch (id) {
2152 case vmIntrinsics::_numberOfLeadingZeros_i: n = new CountLeadingZerosINode( arg); break;
2153 case vmIntrinsics::_numberOfLeadingZeros_l: n = new CountLeadingZerosLNode( arg); break;
2154 case vmIntrinsics::_numberOfTrailingZeros_i: n = new CountTrailingZerosINode(arg); break;
2155 case vmIntrinsics::_numberOfTrailingZeros_l: n = new CountTrailingZerosLNode(arg); break;
2156 case vmIntrinsics::_bitCount_i: n = new PopCountINode( arg); break;
2157 case vmIntrinsics::_bitCount_l: n = new PopCountLNode( arg); break;
2158 case vmIntrinsics::_reverseBytes_c: n = new ReverseBytesUSNode(0, arg); break;
2159 case vmIntrinsics::_reverseBytes_s: n = new ReverseBytesSNode( 0, arg); break;
2160 case vmIntrinsics::_reverseBytes_i: n = new ReverseBytesINode( 0, arg); break;
2161 case vmIntrinsics::_reverseBytes_l: n = new ReverseBytesLNode( 0, arg); break;
2162 default: fatal_unexpected_iid(id); break;
2163 }
2164 set_result(_gvn.transform(n));
2165 return true;
2166 }
2167
2168 //----------------------------inline_unsafe_access----------------------------
2169
2170 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2171
2172 // Helper that guards and inserts a pre-barrier.
2173 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2174 Node* pre_val, bool need_mem_bar) {
2175 // We could be accessing the referent field of a reference object. If so, when G1
2176 // is enabled, we need to log the value in the referent field in an SATB buffer.
2177 // This routine performs some compile time filters and generates suitable
2178 // runtime filters that guard the pre-barrier code.
2179 // Also add memory barrier for non volatile load from the referent field
2180 // to prevent commoning of loads across safepoint.
2181 if (!(UseG1GC || UseShenandoahGC) && !need_mem_bar)
2182 return;
2183
2184 // Some compile time checks.
2185
2186 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2187 const TypeX* otype = offset->find_intptr_t_type();
2188 if (otype != NULL && otype->is_con() &&
2189 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2190 // Constant offset but not the reference_offset so just return
2191 return;
2192 }
2193
2194 // We only need to generate the runtime guards for instances.
2195 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2196 if (btype != NULL) {
2197 if (btype->isa_aryptr()) {
2198 // Array type so nothing to do
2199 return;
2200 }
2201
2202 const TypeInstPtr* itype = btype->isa_instptr();
2203 if (itype != NULL) {
2204 // Can the klass of base_oop be statically determined to be
2205 // _not_ a sub-class of Reference and _not_ Object?
2206 ciKlass* klass = itype->klass();
2207 if ( klass->is_loaded() &&
2208 !klass->is_subtype_of(env()->Reference_klass()) &&
2209 !env()->Object_klass()->is_subtype_of(klass)) {
2210 return;
2211 }
2212 }
2213 }
2214
2215 // The compile time filters did not reject base_oop/offset so
2216 // we need to generate the following runtime filters
2217 //
2218 // if (offset == java_lang_ref_Reference::_reference_offset) {
2219 // if (instance_of(base, java.lang.ref.Reference)) {
2220 // pre_barrier(_, pre_val, ...);
2221 // }
2222 // }
2223
2224 float likely = PROB_LIKELY( 0.999);
2225 float unlikely = PROB_UNLIKELY(0.999);
2226
2227 IdealKit ideal(this);
2228 #define __ ideal.
2229
2230 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2231
2232 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2233 // Update graphKit memory and control from IdealKit.
2234 sync_kit(ideal);
2235
2236 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2237 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2238
2239 // Update IdealKit memory and control from graphKit.
2240 __ sync_kit(this);
2241
2242 Node* one = __ ConI(1);
2243 // is_instof == 0 if base_oop == NULL
2244 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2245
2246 // Update graphKit from IdeakKit.
2247 sync_kit(ideal);
2248
2249 // Use the pre-barrier to record the value in the referent field
2250 pre_barrier(false /* do_load */,
2251 __ ctrl(),
2252 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2253 pre_val /* pre_val */,
2254 T_OBJECT);
2255 if (need_mem_bar) {
2256 // Add memory barrier to prevent commoning reads from this field
2257 // across safepoint since GC can change its value.
2258 insert_mem_bar(Op_MemBarCPUOrder);
2259 }
2260 // Update IdealKit from graphKit.
2261 __ sync_kit(this);
2262
2263 } __ end_if(); // _ref_type != ref_none
2264 } __ end_if(); // offset == referent_offset
2265
2266 // Final sync IdealKit and GraphKit.
2267 final_sync(ideal);
2268 #undef __
2269 }
2270
2271
2272 // Interpret Unsafe.fieldOffset cookies correctly:
2273 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2274
2275 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2276 // Attempt to infer a sharper value type from the offset and base type.
2277 ciKlass* sharpened_klass = NULL;
2278
2279 // See if it is an instance field, with an object type.
2280 if (alias_type->field() != NULL) {
2281 assert(!is_native_ptr, "native pointer op cannot use a java address");
2282 if (alias_type->field()->type()->is_klass()) {
2283 sharpened_klass = alias_type->field()->type()->as_klass();
2284 }
2285 }
2286
2287 // See if it is a narrow oop array.
2288 if (adr_type->isa_aryptr()) {
2289 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2290 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2291 if (elem_type != NULL) {
2292 sharpened_klass = elem_type->klass();
2293 }
2294 }
2295 }
2296
2297 // The sharpened class might be unloaded if there is no class loader
2298 // contraint in place.
2299 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2300 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2301
2302 #ifndef PRODUCT
2303 if (C->print_intrinsics() || C->print_inlining()) {
2304 tty->print(" from base type: "); adr_type->dump();
2305 tty->print(" sharpened value: "); tjp->dump();
2306 }
2307 #endif
2308 // Sharpen the value type.
2309 return tjp;
2310 }
2311 return NULL;
2312 }
2313
2314 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2315 if (callee()->is_static()) return false; // caller must have the capability!
2316
2317 #ifndef PRODUCT
2318 {
2319 ResourceMark rm;
2320 // Check the signatures.
2321 ciSignature* sig = callee()->signature();
2322 #ifdef ASSERT
2323 if (!is_store) {
2324 // Object getObject(Object base, int/long offset), etc.
2325 BasicType rtype = sig->return_type()->basic_type();
2326 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2327 rtype = T_ADDRESS; // it is really a C void*
2328 assert(rtype == type, "getter must return the expected value");
2329 if (!is_native_ptr) {
2330 assert(sig->count() == 2, "oop getter has 2 arguments");
2331 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2332 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2333 } else {
2334 assert(sig->count() == 1, "native getter has 1 argument");
2335 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2336 }
2337 } else {
2338 // void putObject(Object base, int/long offset, Object x), etc.
2339 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2340 if (!is_native_ptr) {
2341 assert(sig->count() == 3, "oop putter has 3 arguments");
2342 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2343 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2344 } else {
2345 assert(sig->count() == 2, "native putter has 2 arguments");
2346 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2347 }
2348 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2349 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2350 vtype = T_ADDRESS; // it is really a C void*
2351 assert(vtype == type, "putter must accept the expected value");
2352 }
2353 #endif // ASSERT
2354 }
2355 #endif //PRODUCT
2356
2357 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2358
2359 Node* receiver = argument(0); // type: oop
2360
2361 // Build address expression.
2362 Node* adr;
2363 Node* heap_base_oop = top();
2364 Node* offset = top();
2365 Node* val;
2366
2367 if (!is_native_ptr) {
2368 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2369 Node* base = argument(1); // type: oop
2370 if (is_store) {
2371 base = shenandoah_write_barrier(base);
2372 } else {
2373 base = shenandoah_read_barrier(base);
2374 }
2375 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2376 offset = argument(2); // type: long
2377 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2378 // to be plain byte offsets, which are also the same as those accepted
2379 // by oopDesc::field_base.
2380 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2381 "fieldOffset must be byte-scaled");
2382 // 32-bit machines ignore the high half!
2383 offset = ConvL2X(offset);
2384 adr = make_unsafe_address(base, offset);
2385 heap_base_oop = base;
2386 val = is_store ? argument(4) : NULL;
2387 } else {
2388 Node* ptr = argument(1); // type: long
2389 ptr = ConvL2X(ptr); // adjust Java long to machine word
2390 adr = make_unsafe_address(NULL, ptr);
2391 val = is_store ? argument(3) : NULL;
2392 }
2393
2394 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2395
2396 // First guess at the value type.
2397 const Type *value_type = Type::get_const_basic_type(type);
2398
2399 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2400 // there was not enough information to nail it down.
2401 Compile::AliasType* alias_type = C->alias_type(adr_type);
2402 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2403
2404 // We will need memory barriers unless we can determine a unique
2405 // alias category for this reference. (Note: If for some reason
2406 // the barriers get omitted and the unsafe reference begins to "pollute"
2407 // the alias analysis of the rest of the graph, either Compile::can_alias
2408 // or Compile::must_alias will throw a diagnostic assert.)
2409 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2410
2411 // If we are reading the value of the referent field of a Reference
2412 // object (either by using Unsafe directly or through reflection)
2413 // then, if G1 is enabled, we need to record the referent in an
2414 // SATB log buffer using the pre-barrier mechanism.
2415 // Also we need to add memory barrier to prevent commoning reads
2416 // from this field across safepoint since GC can change its value.
2417 bool need_read_barrier = !is_native_ptr && !is_store &&
2418 offset != top() && heap_base_oop != top();
2419
2420 if (!is_store && type == T_OBJECT) {
2421 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2422 if (tjp != NULL) {
2423 value_type = tjp;
2424 }
2425 }
2426
2427 receiver = null_check(receiver);
2428 if (stopped()) {
2429 return true;
2430 }
2431 // Heap pointers get a null-check from the interpreter,
2432 // as a courtesy. However, this is not guaranteed by Unsafe,
2433 // and it is not possible to fully distinguish unintended nulls
2434 // from intended ones in this API.
2435
2436 if (is_volatile) {
2437 // We need to emit leading and trailing CPU membars (see below) in
2438 // addition to memory membars when is_volatile. This is a little
2439 // too strong, but avoids the need to insert per-alias-type
2440 // volatile membars (for stores; compare Parse::do_put_xxx), which
2441 // we cannot do effectively here because we probably only have a
2442 // rough approximation of type.
2443 need_mem_bar = true;
2444 // For Stores, place a memory ordering barrier now.
2445 if (is_store) {
2446 insert_mem_bar(Op_MemBarRelease);
2447 } else {
2448 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2449 insert_mem_bar(Op_MemBarVolatile);
2450 }
2451 }
2452 }
2453
2454 // Memory barrier to prevent normal and 'unsafe' accesses from
2455 // bypassing each other. Happens after null checks, so the
2456 // exception paths do not take memory state from the memory barrier,
2457 // so there's no problems making a strong assert about mixing users
2458 // of safe & unsafe memory.
2459 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2460
2461 if (!is_store) {
2462 Node* p = NULL;
2463 // Try to constant fold a load from a constant field
2464 ciField* field = alias_type->field();
2465 if (heap_base_oop != top() &&
2466 field != NULL && field->is_constant() && field->layout_type() == type) {
2467 // final or stable field
2468 const Type* con_type = Type::make_constant(alias_type->field(), heap_base_oop);
2469 if (con_type != NULL) {
2470 p = makecon(con_type);
2471 }
2472 }
2473 if (p == NULL) {
2474 MemNode::MemOrd mo = is_volatile ? MemNode::acquire : MemNode::unordered;
2475 // To be valid, unsafe loads may depend on other conditions than
2476 // the one that guards them: pin the Load node
2477 p = make_load(control(), adr, value_type, type, adr_type, mo, LoadNode::Pinned, is_volatile);
2478 // load value
2479 switch (type) {
2480 case T_BOOLEAN:
2481 case T_CHAR:
2482 case T_BYTE:
2483 case T_SHORT:
2484 case T_INT:
2485 case T_LONG:
2486 case T_FLOAT:
2487 case T_DOUBLE:
2488 break;
2489 case T_OBJECT:
2490 if (need_read_barrier) {
2491 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2492 }
2493 break;
2494 case T_ADDRESS:
2495 // Cast to an int type.
2496 p = _gvn.transform(new CastP2XNode(NULL, p));
2497 p = ConvX2UL(p);
2498 break;
2499 default:
2500 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2501 break;
2502 }
2503 }
2504 // The load node has the control of the preceding MemBarCPUOrder. All
2505 // following nodes will have the control of the MemBarCPUOrder inserted at
2506 // the end of this method. So, pushing the load onto the stack at a later
2507 // point is fine.
2508 set_result(p);
2509 } else {
2510 // place effect of store into memory
2511 switch (type) {
2512 case T_DOUBLE:
2513 val = dstore_rounding(val);
2514 break;
2515 case T_ADDRESS:
2516 // Repackage the long as a pointer.
2517 val = ConvL2X(val);
2518 val = _gvn.transform(new CastX2PNode(val));
2519 break;
2520 }
2521
2522 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2523 if (type != T_OBJECT ) {
2524 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2525 } else {
2526 val = shenandoah_read_barrier_nomem(val);
2527 // Possibly an oop being stored to Java heap or native memory
2528 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2529 // oop to Java heap.
2530 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2531 } else {
2532 // We can't tell at compile time if we are storing in the Java heap or outside
2533 // of it. So we need to emit code to conditionally do the proper type of
2534 // store.
2535
2536 IdealKit ideal(this);
2537 #define __ ideal.
2538 // QQQ who knows what probability is here??
2539 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2540 // Sync IdealKit and graphKit.
2541 sync_kit(ideal);
2542 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2543 // Update IdealKit memory.
2544 __ sync_kit(this);
2545 } __ else_(); {
2546 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2547 } __ end_if();
2548 // Final sync IdealKit and GraphKit.
2549 final_sync(ideal);
2550 #undef __
2551 }
2552 }
2553 }
2554
2555 if (is_volatile) {
2556 if (!is_store) {
2557 insert_mem_bar(Op_MemBarAcquire);
2558 } else {
2559 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2560 insert_mem_bar(Op_MemBarVolatile);
2561 }
2562 }
2563 }
2564
2565 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2566
2567 return true;
2568 }
2569
2570 //----------------------------inline_unsafe_load_store----------------------------
2571 // This method serves a couple of different customers (depending on LoadStoreKind):
2572 //
2573 // LS_cmpxchg:
2574 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2575 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2576 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2577 //
2578 // LS_xadd:
2579 // public int getAndAddInt( Object o, long offset, int delta)
2580 // public long getAndAddLong(Object o, long offset, long delta)
2581 //
2582 // LS_xchg:
2583 // int getAndSet(Object o, long offset, int newValue)
2584 // long getAndSet(Object o, long offset, long newValue)
2585 // Object getAndSet(Object o, long offset, Object newValue)
2586 //
2587 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2588 // This basic scheme here is the same as inline_unsafe_access, but
2589 // differs in enough details that combining them would make the code
2590 // overly confusing. (This is a true fact! I originally combined
2591 // them, but even I was confused by it!) As much code/comments as
2592 // possible are retained from inline_unsafe_access though to make
2593 // the correspondences clearer. - dl
2594
2595 if (callee()->is_static()) return false; // caller must have the capability!
2596
2597 #ifndef PRODUCT
2598 BasicType rtype;
2599 {
2600 ResourceMark rm;
2601 // Check the signatures.
2602 ciSignature* sig = callee()->signature();
2603 rtype = sig->return_type()->basic_type();
2604 if (kind == LS_xadd || kind == LS_xchg) {
2605 // Check the signatures.
2606 #ifdef ASSERT
2607 assert(rtype == type, "get and set must return the expected type");
2608 assert(sig->count() == 3, "get and set has 3 arguments");
2609 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2610 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2611 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2612 #endif // ASSERT
2613 } else if (kind == LS_cmpxchg) {
2614 // Check the signatures.
2615 #ifdef ASSERT
2616 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2617 assert(sig->count() == 4, "CAS has 4 arguments");
2618 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2619 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2620 #endif // ASSERT
2621 } else {
2622 ShouldNotReachHere();
2623 }
2624 }
2625 #endif //PRODUCT
2626
2627 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2628
2629 // Get arguments:
2630 Node* receiver = NULL;
2631 Node* base = NULL;
2632 Node* offset = NULL;
2633 Node* oldval = NULL;
2634 Node* newval = NULL;
2635 if (kind == LS_cmpxchg) {
2636 const bool two_slot_type = type2size[type] == 2;
2637 receiver = argument(0); // type: oop
2638 base = argument(1); // type: oop
2639 offset = argument(2); // type: long
2640 oldval = argument(4); // type: oop, int, or long
2641 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2642 } else if (kind == LS_xadd || kind == LS_xchg){
2643 receiver = argument(0); // type: oop
2644 base = argument(1); // type: oop
2645 offset = argument(2); // type: long
2646 oldval = NULL;
2647 newval = argument(4); // type: oop, int, or long
2648 }
2649
2650 // Null check receiver.
2651 receiver = null_check(receiver);
2652 if (stopped()) {
2653 return true;
2654 }
2655
2656 base = shenandoah_write_barrier(base);
2657
2658 // Build field offset expression.
2659 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2660 // to be plain byte offsets, which are also the same as those accepted
2661 // by oopDesc::field_base.
2662 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2663 // 32-bit machines ignore the high half of long offsets
2664 offset = ConvL2X(offset);
2665 Node* adr = make_unsafe_address(base, offset);
2666 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2667
2668 // For CAS, unlike inline_unsafe_access, there seems no point in
2669 // trying to refine types. Just use the coarse types here.
2670 const Type *value_type = Type::get_const_basic_type(type);
2671 Compile::AliasType* alias_type = C->alias_type(adr_type);
2672 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2673
2674 if (kind == LS_xchg && type == T_OBJECT) {
2675 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2676 if (tjp != NULL) {
2677 value_type = tjp;
2678 }
2679 }
2680
2681 int alias_idx = C->get_alias_index(adr_type);
2682
2683 // Memory-model-wise, a LoadStore acts like a little synchronized
2684 // block, so needs barriers on each side. These don't translate
2685 // into actual barriers on most machines, but we still need rest of
2686 // compiler to respect ordering.
2687
2688 insert_mem_bar(Op_MemBarRelease);
2689 insert_mem_bar(Op_MemBarCPUOrder);
2690
2691 // 4984716: MemBars must be inserted before this
2692 // memory node in order to avoid a false
2693 // dependency which will confuse the scheduler.
2694 Node *mem = memory(alias_idx);
2695
2696 // For now, we handle only those cases that actually exist: ints,
2697 // longs, and Object. Adding others should be straightforward.
2698 Node* load_store;
2699 Node* result;
2700 switch(type) {
2701 case T_INT:
2702 if (kind == LS_xadd) {
2703 load_store = _gvn.transform(new GetAndAddINode(control(), mem, adr, newval, adr_type));
2704 } else if (kind == LS_xchg) {
2705 load_store = _gvn.transform(new GetAndSetINode(control(), mem, adr, newval, adr_type));
2706 } else if (kind == LS_cmpxchg) {
2707 load_store = _gvn.transform(new CompareAndSwapINode(control(), mem, adr, newval, oldval));
2708 } else {
2709 ShouldNotReachHere();
2710 }
2711 result = load_store;
2712 break;
2713 case T_LONG:
2714 if (kind == LS_xadd) {
2715 load_store = _gvn.transform(new GetAndAddLNode(control(), mem, adr, newval, adr_type));
2716 } else if (kind == LS_xchg) {
2717 load_store = _gvn.transform(new GetAndSetLNode(control(), mem, adr, newval, adr_type));
2718 } else if (kind == LS_cmpxchg) {
2719 load_store = _gvn.transform(new CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2720 } else {
2721 ShouldNotReachHere();
2722 }
2723 result = load_store;
2724 break;
2725 case T_OBJECT:
2726 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2727 // could be delayed during Parse (for example, in adjust_map_after_if()).
2728 // Execute transformation here to avoid barrier generation in such case.
2729 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2730 newval = _gvn.makecon(TypePtr::NULL_PTR);
2731
2732 newval = shenandoah_read_barrier_nomem(newval);
2733
2734 // Reference stores need a store barrier.
2735 if (kind == LS_xchg) {
2736 // If pre-barrier must execute before the oop store, old value will require do_load here.
2737 if (!can_move_pre_barrier()) {
2738 pre_barrier(true /* do_load*/,
2739 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2740 NULL /* pre_val*/,
2741 T_OBJECT);
2742 } // Else move pre_barrier to use load_store value, see below.
2743 } else if (kind == LS_cmpxchg) {
2744 // Same as for newval above:
2745 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2746 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2747 }
2748 // The only known value which might get overwritten is oldval.
2749 pre_barrier(false /* do_load */,
2750 control(), NULL, NULL, max_juint, NULL, NULL,
2751 oldval /* pre_val */,
2752 T_OBJECT);
2753 } else {
2754 ShouldNotReachHere();
2755 }
2756
2757 #ifdef _LP64
2758 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2759 Node *newval_enc = _gvn.transform(new EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2760 if (kind == LS_xchg) {
2761 load_store = _gvn.transform(new GetAndSetNNode(control(), mem, adr,
2762 newval_enc, adr_type, value_type->make_narrowoop()));
2763 } else {
2764 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2765 Node *oldval_enc = _gvn.transform(new EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2766 load_store = _gvn.transform(new CompareAndSwapNNode(control(), mem, adr,
2767 newval_enc, oldval_enc));
2768 }
2769 result = load_store;
2770 } else
2771 #endif
2772 {
2773 if (kind == LS_xchg) {
2774 load_store = _gvn.transform(new GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2775 result = load_store;
2776 } else {
2777 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2778 load_store = _gvn.transform(new CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2779 result = load_store;
2780
2781 if (UseShenandoahGC) {
2782 // if (! success)
2783 Node* cmp_true = _gvn.transform(new CmpINode(load_store, intcon(1)));
2784 Node* tst_true = _gvn.transform(new BoolNode(cmp_true, BoolTest::eq));
2785 IfNode* iff = create_and_map_if(control(), tst_true, PROB_LIKELY_MAG(2), COUNT_UNKNOWN);
2786 Node* iftrue = _gvn.transform(new IfTrueNode(iff));
2787 Node* iffalse = _gvn.transform(new IfFalseNode(iff));
2788
2789 enum { _success_path = 1, _fail_path, _shenandoah_path, PATH_LIMIT };
2790 RegionNode* region = new RegionNode(PATH_LIMIT);
2791 Node* phi = new PhiNode(region, TypeInt::BOOL);
2792 // success -> return result of CAS1.
2793 region->init_req(_success_path, iftrue);
2794 phi ->init_req(_success_path, load_store);
2795
2796 // failure
2797 set_control(iffalse);
2798
2799 // if (read_barrier(expected) == read_barrier(old)
2800 oldval = shenandoah_read_barrier(oldval);
2801
2802 // Load old value from memory. We shuold really use what we get back from the CAS,
2803 // if we can.
2804 Node* current = make_load(control(), adr, TypeInstPtr::BOTTOM, type, MemNode::unordered);
2805 // read_barrier(old)
2806 Node* new_current = shenandoah_read_barrier(current);
2807
2808 Node* chk = _gvn.transform(new CmpPNode(new_current, oldval));
2809 Node* test = _gvn.transform(new BoolNode(chk, BoolTest::eq));
2810
2811 IfNode* iff2 = create_and_map_if(control(), test, PROB_UNLIKELY_MAG(2), COUNT_UNKNOWN);
2812 Node* iftrue2 = _gvn.transform(new IfTrueNode(iff2));
2813 Node* iffalse2 = _gvn.transform(new IfFalseNode(iff2));
2814
2815 // If they are not equal, it's a legitimate failure and we return the result of CAS1.
2816 region->init_req(_fail_path, iffalse2);
2817 phi ->init_req(_fail_path, load_store);
2818
2819 // Otherwise we retry with old.
2820 set_control(iftrue2);
2821
2822 Node *call = make_runtime_call(RC_LEAF | RC_NO_IO,
2823 OptoRuntime::shenandoah_cas_obj_Type(),
2824 CAST_FROM_FN_PTR(address, ShenandoahRuntime::compare_and_swap_object),
2825 "shenandoah_cas_obj",
2826 NULL,
2827 adr, newval, current);
2828
2829 Node* retval = _gvn.transform(new ProjNode(call, TypeFunc::Parms + 0));
2830
2831 region->init_req(_shenandoah_path, control());
2832 phi ->init_req(_shenandoah_path, retval);
2833
2834 set_control(_gvn.transform(region));
2835 record_for_igvn(region);
2836 phi = _gvn.transform(phi);
2837 result = phi;
2838 }
2839
2840 }
2841 }
2842 if (kind == LS_cmpxchg) {
2843 // Emit the post barrier only when the actual store happened.
2844 // This makes sense to check only for compareAndSet that can fail to set the value.
2845 // CAS success path is marked more likely since we anticipate this is a performance
2846 // critical path, while CAS failure path can use the penalty for going through unlikely
2847 // path as backoff. Which is still better than doing a store barrier there.
2848 IdealKit ideal(this);
2849 ideal.if_then(result, BoolTest::ne, ideal.ConI(0), PROB_STATIC_FREQUENT); {
2850 sync_kit(ideal);
2851 post_barrier(ideal.ctrl(), result, base, adr, alias_idx, newval, T_OBJECT, true);
2852 ideal.sync_kit(this);
2853 } ideal.end_if();
2854 final_sync(ideal);
2855 } else {
2856 post_barrier(control(), result, base, adr, alias_idx, newval, T_OBJECT, true);
2857 }
2858 break;
2859 default:
2860 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2861 break;
2862 }
2863
2864 // SCMemProjNodes represent the memory state of a LoadStore. Their
2865 // main role is to prevent LoadStore nodes from being optimized away
2866 // when their results aren't used.
2867 Node* proj = _gvn.transform(new SCMemProjNode(load_store));
2868 set_memory(proj, alias_idx);
2869
2870 if (type == T_OBJECT && kind == LS_xchg) {
2871 #ifdef _LP64
2872 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2873 result = _gvn.transform(new DecodeNNode(result, result->get_ptr_type()));
2874 }
2875 #endif
2876 if (can_move_pre_barrier()) {
2877 // Don't need to load pre_val. The old value is returned by load_store.
2878 // The pre_barrier can execute after the xchg as long as no safepoint
2879 // gets inserted between them.
2880 pre_barrier(false /* do_load */,
2881 control(), NULL, NULL, max_juint, NULL, NULL,
2882 result /* pre_val */,
2883 T_OBJECT);
2884 }
2885 }
2886
2887 // Add the trailing membar surrounding the access
2888 insert_mem_bar(Op_MemBarCPUOrder);
2889 insert_mem_bar(Op_MemBarAcquire);
2890
2891 assert(type2size[result->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2892 set_result(result);
2893 return true;
2894 }
2895
2896 //----------------------------inline_unsafe_ordered_store----------------------
2897 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2898 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2899 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2900 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2901 // This is another variant of inline_unsafe_access, differing in
2902 // that it always issues store-store ("release") barrier and ensures
2903 // store-atomicity (which only matters for "long").
2904
2905 if (callee()->is_static()) return false; // caller must have the capability!
2906
2907 #ifndef PRODUCT
2908 {
2909 ResourceMark rm;
2910 // Check the signatures.
2911 ciSignature* sig = callee()->signature();
2912 #ifdef ASSERT
2913 BasicType rtype = sig->return_type()->basic_type();
2914 assert(rtype == T_VOID, "must return void");
2915 assert(sig->count() == 3, "has 3 arguments");
2916 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2917 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2918 #endif // ASSERT
2919 }
2920 #endif //PRODUCT
2921
2922 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2923
2924 // Get arguments:
2925 Node* receiver = argument(0); // type: oop
2926 Node* base = argument(1); // type: oop
2927 Node* offset = argument(2); // type: long
2928 Node* val = argument(4); // type: oop, int, or long
2929
2930 // Null check receiver.
2931 receiver = null_check(receiver);
2932 if (stopped()) {
2933 return true;
2934 }
2935
2936 base = shenandoah_write_barrier(base);
2937
2938 // Build field offset expression.
2939 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2940 // 32-bit machines ignore the high half of long offsets
2941 offset = ConvL2X(offset);
2942 Node* adr = make_unsafe_address(base, offset);
2943 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2944 const Type *value_type = Type::get_const_basic_type(type);
2945 Compile::AliasType* alias_type = C->alias_type(adr_type);
2946
2947 insert_mem_bar(Op_MemBarRelease);
2948 insert_mem_bar(Op_MemBarCPUOrder);
2949 // Ensure that the store is atomic for longs:
2950 const bool require_atomic_access = true;
2951 Node* store;
2952 if (type == T_OBJECT) { // reference stores need a store barrier.
2953 val = shenandoah_read_barrier_nomem(val);
2954 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
2955 }
2956 else {
2957 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
2958 }
2959 insert_mem_bar(Op_MemBarCPUOrder);
2960 return true;
2961 }
2962
2963 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2964 // Regardless of form, don't allow previous ld/st to move down,
2965 // then issue acquire, release, or volatile mem_bar.
2966 insert_mem_bar(Op_MemBarCPUOrder);
2967 switch(id) {
2968 case vmIntrinsics::_loadFence:
2969 insert_mem_bar(Op_LoadFence);
2970 return true;
2971 case vmIntrinsics::_storeFence:
2972 insert_mem_bar(Op_StoreFence);
2973 return true;
2974 case vmIntrinsics::_fullFence:
2975 insert_mem_bar(Op_MemBarVolatile);
2976 return true;
2977 default:
2978 fatal_unexpected_iid(id);
2979 return false;
2980 }
2981 }
2982
2983 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2984 if (!kls->is_Con()) {
2985 return true;
2986 }
2987 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2988 if (klsptr == NULL) {
2989 return true;
2990 }
2991 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2992 // don't need a guard for a klass that is already initialized
2993 return !ik->is_initialized();
2994 }
2995
2996 //----------------------------inline_unsafe_allocate---------------------------
2997 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2998 bool LibraryCallKit::inline_unsafe_allocate() {
2999 if (callee()->is_static()) return false; // caller must have the capability!
3000
3001 null_check_receiver(); // null-check, then ignore
3002 Node* cls = null_check(argument(1));
3003 if (stopped()) return true;
3004
3005 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3006 kls = null_check(kls);
3007 if (stopped()) return true; // argument was like int.class
3008
3009 Node* test = NULL;
3010 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3011 // Note: The argument might still be an illegal value like
3012 // Serializable.class or Object[].class. The runtime will handle it.
3013 // But we must make an explicit check for initialization.
3014 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3015 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3016 // can generate code to load it as unsigned byte.
3017 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3018 Node* bits = intcon(InstanceKlass::fully_initialized);
3019 test = _gvn.transform(new SubINode(inst, bits));
3020 // The 'test' is non-zero if we need to take a slow path.
3021 }
3022
3023 Node* obj = new_instance(kls, test);
3024 set_result(obj);
3025 return true;
3026 }
3027
3028 #ifdef TRACE_HAVE_INTRINSICS
3029 /*
3030 * oop -> myklass
3031 * myklass->trace_id |= USED
3032 * return myklass->trace_id & ~0x3
3033 */
3034 bool LibraryCallKit::inline_native_classID() {
3035 null_check_receiver(); // null-check, then ignore
3036 Node* cls = null_check(argument(1), T_OBJECT);
3037 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3038 kls = null_check(kls, T_OBJECT);
3039 ByteSize offset = TRACE_ID_OFFSET;
3040 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3041 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3042 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3043 Node* andl = _gvn.transform(new AndLNode(tvalue, bits));
3044 Node* clsused = longcon(0x01l); // set the class bit
3045 Node* orl = _gvn.transform(new OrLNode(tvalue, clsused));
3046
3047 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3048 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3049 set_result(andl);
3050 return true;
3051 }
3052
3053 bool LibraryCallKit::inline_native_threadID() {
3054 Node* tls_ptr = NULL;
3055 Node* cur_thr = generate_current_thread(tls_ptr);
3056 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3057 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3058 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3059
3060 Node* threadid = NULL;
3061 size_t thread_id_size = OSThread::thread_id_size();
3062 if (thread_id_size == (size_t) BytesPerLong) {
3063 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3064 } else if (thread_id_size == (size_t) BytesPerInt) {
3065 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3066 } else {
3067 ShouldNotReachHere();
3068 }
3069 set_result(threadid);
3070 return true;
3071 }
3072 #endif
3073
3074 //------------------------inline_native_time_funcs--------------
3075 // inline code for System.currentTimeMillis() and System.nanoTime()
3076 // these have the same type and signature
3077 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3078 const TypeFunc* tf = OptoRuntime::void_long_Type();
3079 const TypePtr* no_memory_effects = NULL;
3080 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3081 Node* value = _gvn.transform(new ProjNode(time, TypeFunc::Parms+0));
3082 #ifdef ASSERT
3083 Node* value_top = _gvn.transform(new ProjNode(time, TypeFunc::Parms+1));
3084 assert(value_top == top(), "second value must be top");
3085 #endif
3086 set_result(value);
3087 return true;
3088 }
3089
3090 //------------------------inline_native_currentThread------------------
3091 bool LibraryCallKit::inline_native_currentThread() {
3092 Node* junk = NULL;
3093 set_result(generate_current_thread(junk));
3094 return true;
3095 }
3096
3097 //------------------------inline_native_isInterrupted------------------
3098 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3099 bool LibraryCallKit::inline_native_isInterrupted() {
3100 // Add a fast path to t.isInterrupted(clear_int):
3101 // (t == Thread.current() &&
3102 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3103 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3104 // So, in the common case that the interrupt bit is false,
3105 // we avoid making a call into the VM. Even if the interrupt bit
3106 // is true, if the clear_int argument is false, we avoid the VM call.
3107 // However, if the receiver is not currentThread, we must call the VM,
3108 // because there must be some locking done around the operation.
3109
3110 // We only go to the fast case code if we pass two guards.
3111 // Paths which do not pass are accumulated in the slow_region.
3112
3113 enum {
3114 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3115 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3116 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3117 PATH_LIMIT
3118 };
3119
3120 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3121 // out of the function.
3122 insert_mem_bar(Op_MemBarCPUOrder);
3123
3124 RegionNode* result_rgn = new RegionNode(PATH_LIMIT);
3125 PhiNode* result_val = new PhiNode(result_rgn, TypeInt::BOOL);
3126
3127 RegionNode* slow_region = new RegionNode(1);
3128 record_for_igvn(slow_region);
3129
3130 // (a) Receiving thread must be the current thread.
3131 Node* rec_thr = argument(0);
3132 Node* tls_ptr = NULL;
3133 Node* cur_thr = generate_current_thread(tls_ptr);
3134 Node* cmp_thr = _gvn.transform(new CmpPNode(cur_thr, rec_thr));
3135 Node* bol_thr = _gvn.transform(new BoolNode(cmp_thr, BoolTest::ne));
3136
3137 generate_slow_guard(bol_thr, slow_region);
3138
3139 // (b) Interrupt bit on TLS must be false.
3140 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3141 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3142 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3143
3144 // Set the control input on the field _interrupted read to prevent it floating up.
3145 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3146 Node* cmp_bit = _gvn.transform(new CmpINode(int_bit, intcon(0)));
3147 Node* bol_bit = _gvn.transform(new BoolNode(cmp_bit, BoolTest::ne));
3148
3149 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3150
3151 // First fast path: if (!TLS._interrupted) return false;
3152 Node* false_bit = _gvn.transform(new IfFalseNode(iff_bit));
3153 result_rgn->init_req(no_int_result_path, false_bit);
3154 result_val->init_req(no_int_result_path, intcon(0));
3155
3156 // drop through to next case
3157 set_control( _gvn.transform(new IfTrueNode(iff_bit)));
3158
3159 #ifndef TARGET_OS_FAMILY_windows
3160 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3161 Node* clr_arg = argument(1);
3162 Node* cmp_arg = _gvn.transform(new CmpINode(clr_arg, intcon(0)));
3163 Node* bol_arg = _gvn.transform(new BoolNode(cmp_arg, BoolTest::ne));
3164 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3165
3166 // Second fast path: ... else if (!clear_int) return true;
3167 Node* false_arg = _gvn.transform(new IfFalseNode(iff_arg));
3168 result_rgn->init_req(no_clear_result_path, false_arg);
3169 result_val->init_req(no_clear_result_path, intcon(1));
3170
3171 // drop through to next case
3172 set_control( _gvn.transform(new IfTrueNode(iff_arg)));
3173 #else
3174 // To return true on Windows you must read the _interrupted field
3175 // and check the the event state i.e. take the slow path.
3176 #endif // TARGET_OS_FAMILY_windows
3177
3178 // (d) Otherwise, go to the slow path.
3179 slow_region->add_req(control());
3180 set_control( _gvn.transform(slow_region));
3181
3182 if (stopped()) {
3183 // There is no slow path.
3184 result_rgn->init_req(slow_result_path, top());
3185 result_val->init_req(slow_result_path, top());
3186 } else {
3187 // non-virtual because it is a private non-static
3188 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3189
3190 Node* slow_val = set_results_for_java_call(slow_call);
3191 // this->control() comes from set_results_for_java_call
3192
3193 Node* fast_io = slow_call->in(TypeFunc::I_O);
3194 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3195
3196 // These two phis are pre-filled with copies of of the fast IO and Memory
3197 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3198 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3199
3200 result_rgn->init_req(slow_result_path, control());
3201 result_io ->init_req(slow_result_path, i_o());
3202 result_mem->init_req(slow_result_path, reset_memory());
3203 result_val->init_req(slow_result_path, slow_val);
3204
3205 set_all_memory(_gvn.transform(result_mem));
3206 set_i_o( _gvn.transform(result_io));
3207 }
3208
3209 C->set_has_split_ifs(true); // Has chance for split-if optimization
3210 set_result(result_rgn, result_val);
3211 return true;
3212 }
3213
3214 //---------------------------load_mirror_from_klass----------------------------
3215 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3216 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3217 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3218 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3219 }
3220
3221 //-----------------------load_klass_from_mirror_common-------------------------
3222 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3223 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3224 // and branch to the given path on the region.
3225 // If never_see_null, take an uncommon trap on null, so we can optimistically
3226 // compile for the non-null case.
3227 // If the region is NULL, force never_see_null = true.
3228 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3229 bool never_see_null,
3230 RegionNode* region,
3231 int null_path,
3232 int offset) {
3233 if (region == NULL) never_see_null = true;
3234 Node* p = basic_plus_adr(mirror, offset);
3235 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3236 Node* kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3237 Node* null_ctl = top();
3238 kls = null_check_oop(kls, &null_ctl, never_see_null);
3239 if (region != NULL) {
3240 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3241 region->init_req(null_path, null_ctl);
3242 } else {
3243 assert(null_ctl == top(), "no loose ends");
3244 }
3245 return kls;
3246 }
3247
3248 //--------------------(inline_native_Class_query helpers)---------------------
3249 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3250 // Fall through if (mods & mask) == bits, take the guard otherwise.
3251 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3252 // Branch around if the given klass has the given modifier bit set.
3253 // Like generate_guard, adds a new path onto the region.
3254 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3255 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3256 Node* mask = intcon(modifier_mask);
3257 Node* bits = intcon(modifier_bits);
3258 Node* mbit = _gvn.transform(new AndINode(mods, mask));
3259 Node* cmp = _gvn.transform(new CmpINode(mbit, bits));
3260 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3261 return generate_fair_guard(bol, region);
3262 }
3263 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3264 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3265 }
3266
3267 //-------------------------inline_native_Class_query-------------------
3268 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3269 const Type* return_type = TypeInt::BOOL;
3270 Node* prim_return_value = top(); // what happens if it's a primitive class?
3271 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3272 bool expect_prim = false; // most of these guys expect to work on refs
3273
3274 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3275
3276 Node* mirror = argument(0);
3277
3278 if (ShenandoahVerifyReadsToFromSpace) {
3279 mirror = shenandoah_read_barrier(mirror);
3280 }
3281
3282 Node* obj = top();
3283
3284 switch (id) {
3285 case vmIntrinsics::_isInstance:
3286 // nothing is an instance of a primitive type
3287 prim_return_value = intcon(0);
3288 obj = argument(1);
3289 if (ShenandoahVerifyReadsToFromSpace) {
3290 obj = shenandoah_read_barrier(obj);
3291 }
3292 break;
3293 case vmIntrinsics::_getModifiers:
3294 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3295 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3296 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3297 break;
3298 case vmIntrinsics::_isInterface:
3299 prim_return_value = intcon(0);
3300 break;
3301 case vmIntrinsics::_isArray:
3302 prim_return_value = intcon(0);
3303 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3304 break;
3305 case vmIntrinsics::_isPrimitive:
3306 prim_return_value = intcon(1);
3307 expect_prim = true; // obviously
3308 break;
3309 case vmIntrinsics::_getSuperclass:
3310 prim_return_value = null();
3311 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3312 break;
3313 case vmIntrinsics::_getClassAccessFlags:
3314 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3315 return_type = TypeInt::INT; // not bool! 6297094
3316 break;
3317 default:
3318 fatal_unexpected_iid(id);
3319 break;
3320 }
3321
3322 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3323 if (mirror_con == NULL) return false; // cannot happen?
3324
3325 #ifndef PRODUCT
3326 if (C->print_intrinsics() || C->print_inlining()) {
3327 ciType* k = mirror_con->java_mirror_type();
3328 if (k) {
3329 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3330 k->print_name();
3331 tty->cr();
3332 }
3333 }
3334 #endif
3335
3336 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3337 RegionNode* region = new RegionNode(PATH_LIMIT);
3338 record_for_igvn(region);
3339 PhiNode* phi = new PhiNode(region, return_type);
3340
3341 // The mirror will never be null of Reflection.getClassAccessFlags, however
3342 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3343 // if it is. See bug 4774291.
3344
3345 // For Reflection.getClassAccessFlags(), the null check occurs in
3346 // the wrong place; see inline_unsafe_access(), above, for a similar
3347 // situation.
3348 mirror = null_check(mirror);
3349 // If mirror or obj is dead, only null-path is taken.
3350 if (stopped()) return true;
3351
3352 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3353
3354 // Now load the mirror's klass metaobject, and null-check it.
3355 // Side-effects region with the control path if the klass is null.
3356 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3357 // If kls is null, we have a primitive mirror.
3358 phi->init_req(_prim_path, prim_return_value);
3359 if (stopped()) { set_result(region, phi); return true; }
3360 bool safe_for_replace = (region->in(_prim_path) == top());
3361
3362 Node* p; // handy temp
3363 Node* null_ctl;
3364
3365 // Now that we have the non-null klass, we can perform the real query.
3366 // For constant classes, the query will constant-fold in LoadNode::Value.
3367 Node* query_value = top();
3368 switch (id) {
3369 case vmIntrinsics::_isInstance:
3370 // nothing is an instance of a primitive type
3371 query_value = gen_instanceof(obj, kls, safe_for_replace);
3372 break;
3373
3374 case vmIntrinsics::_getModifiers:
3375 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3376 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3377 break;
3378
3379 case vmIntrinsics::_isInterface:
3380 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3381 if (generate_interface_guard(kls, region) != NULL)
3382 // A guard was added. If the guard is taken, it was an interface.
3383 phi->add_req(intcon(1));
3384 // If we fall through, it's a plain class.
3385 query_value = intcon(0);
3386 break;
3387
3388 case vmIntrinsics::_isArray:
3389 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3390 if (generate_array_guard(kls, region) != NULL)
3391 // A guard was added. If the guard is taken, it was an array.
3392 phi->add_req(intcon(1));
3393 // If we fall through, it's a plain class.
3394 query_value = intcon(0);
3395 break;
3396
3397 case vmIntrinsics::_isPrimitive:
3398 query_value = intcon(0); // "normal" path produces false
3399 break;
3400
3401 case vmIntrinsics::_getSuperclass:
3402 // The rules here are somewhat unfortunate, but we can still do better
3403 // with random logic than with a JNI call.
3404 // Interfaces store null or Object as _super, but must report null.
3405 // Arrays store an intermediate super as _super, but must report Object.
3406 // Other types can report the actual _super.
3407 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3408 if (generate_interface_guard(kls, region) != NULL)
3409 // A guard was added. If the guard is taken, it was an interface.
3410 phi->add_req(null());
3411 if (generate_array_guard(kls, region) != NULL)
3412 // A guard was added. If the guard is taken, it was an array.
3413 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3414 // If we fall through, it's a plain class. Get its _super.
3415 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3416 kls = _gvn.transform(LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3417 null_ctl = top();
3418 kls = null_check_oop(kls, &null_ctl);
3419 if (null_ctl != top()) {
3420 // If the guard is taken, Object.superClass is null (both klass and mirror).
3421 region->add_req(null_ctl);
3422 phi ->add_req(null());
3423 }
3424 if (!stopped()) {
3425 query_value = load_mirror_from_klass(kls);
3426 }
3427 break;
3428
3429 case vmIntrinsics::_getClassAccessFlags:
3430 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3431 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3432 break;
3433
3434 default:
3435 fatal_unexpected_iid(id);
3436 break;
3437 }
3438
3439 // Fall-through is the normal case of a query to a real class.
3440 phi->init_req(1, query_value);
3441 region->init_req(1, control());
3442
3443 C->set_has_split_ifs(true); // Has chance for split-if optimization
3444 set_result(region, phi);
3445 return true;
3446 }
3447
3448 //-------------------------inline_Class_cast-------------------
3449 bool LibraryCallKit::inline_Class_cast() {
3450 Node* mirror = argument(0); // Class
3451 Node* obj = argument(1);
3452 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3453 if (mirror_con == NULL) {
3454 return false; // dead path (mirror->is_top()).
3455 }
3456 if (obj == NULL || obj->is_top()) {
3457 return false; // dead path
3458 }
3459 const TypeOopPtr* tp = _gvn.type(obj)->isa_oopptr();
3460
3461 // First, see if Class.cast() can be folded statically.
3462 // java_mirror_type() returns non-null for compile-time Class constants.
3463 ciType* tm = mirror_con->java_mirror_type();
3464 if (tm != NULL && tm->is_klass() &&
3465 tp != NULL && tp->klass() != NULL) {
3466 if (!tp->klass()->is_loaded()) {
3467 // Don't use intrinsic when class is not loaded.
3468 return false;
3469 } else {
3470 int static_res = C->static_subtype_check(tm->as_klass(), tp->klass());
3471 if (static_res == Compile::SSC_always_true) {
3472 // isInstance() is true - fold the code.
3473 set_result(obj);
3474 return true;
3475 } else if (static_res == Compile::SSC_always_false) {
3476 // Don't use intrinsic, have to throw ClassCastException.
3477 // If the reference is null, the non-intrinsic bytecode will
3478 // be optimized appropriately.
3479 return false;
3480 }
3481 }
3482 }
3483
3484 // Bailout intrinsic and do normal inlining if exception path is frequent.
3485 if (too_many_traps(Deoptimization::Reason_intrinsic)) {
3486 return false;
3487 }
3488
3489 // Generate dynamic checks.
3490 // Class.cast() is java implementation of _checkcast bytecode.
3491 // Do checkcast (Parse::do_checkcast()) optimizations here.
3492
3493 mirror = null_check(mirror);
3494 // If mirror is dead, only null-path is taken.
3495 if (stopped()) {
3496 return true;
3497 }
3498
3499 // Not-subtype or the mirror's klass ptr is NULL (in case it is a primitive).
3500 enum { _bad_type_path = 1, _prim_path = 2, PATH_LIMIT };
3501 RegionNode* region = new RegionNode(PATH_LIMIT);
3502 record_for_igvn(region);
3503
3504 // Now load the mirror's klass metaobject, and null-check it.
3505 // If kls is null, we have a primitive mirror and
3506 // nothing is an instance of a primitive type.
3507 Node* kls = load_klass_from_mirror(mirror, false, region, _prim_path);
3508
3509 Node* res = top();
3510 if (!stopped()) {
3511 Node* bad_type_ctrl = top();
3512 // Do checkcast optimizations.
3513 res = gen_checkcast(obj, kls, &bad_type_ctrl);
3514 region->init_req(_bad_type_path, bad_type_ctrl);
3515 }
3516 if (region->in(_prim_path) != top() ||
3517 region->in(_bad_type_path) != top()) {
3518 // Let Interpreter throw ClassCastException.
3519 PreserveJVMState pjvms(this);
3520 set_control(_gvn.transform(region));
3521 uncommon_trap(Deoptimization::Reason_intrinsic,
3522 Deoptimization::Action_maybe_recompile);
3523 }
3524 if (!stopped()) {
3525 set_result(res);
3526 }
3527 return true;
3528 }
3529
3530
3531 //--------------------------inline_native_subtype_check------------------------
3532 // This intrinsic takes the JNI calls out of the heart of
3533 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3534 bool LibraryCallKit::inline_native_subtype_check() {
3535 // Pull both arguments off the stack.
3536 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3537 args[0] = argument(0);
3538 args[1] = argument(1);
3539
3540 Node* klasses[2]; // corresponding Klasses: superk, subk
3541 klasses[0] = klasses[1] = top();
3542
3543 enum {
3544 // A full decision tree on {superc is prim, subc is prim}:
3545 _prim_0_path = 1, // {P,N} => false
3546 // {P,P} & superc!=subc => false
3547 _prim_same_path, // {P,P} & superc==subc => true
3548 _prim_1_path, // {N,P} => false
3549 _ref_subtype_path, // {N,N} & subtype check wins => true
3550 _both_ref_path, // {N,N} & subtype check loses => false
3551 PATH_LIMIT
3552 };
3553
3554 RegionNode* region = new RegionNode(PATH_LIMIT);
3555 Node* phi = new PhiNode(region, TypeInt::BOOL);
3556 record_for_igvn(region);
3557
3558 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3559 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3560 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3561
3562 // First null-check both mirrors and load each mirror's klass metaobject.
3563 int which_arg;
3564 for (which_arg = 0; which_arg <= 1; which_arg++) {
3565 Node* arg = args[which_arg];
3566 arg = null_check(arg);
3567 if (stopped()) break;
3568 args[which_arg] = arg;
3569
3570 Node* p = basic_plus_adr(arg, class_klass_offset);
3571 Node* kls = LoadKlassNode::make(_gvn, NULL, immutable_memory(), p, adr_type, kls_type);
3572 klasses[which_arg] = _gvn.transform(kls);
3573 }
3574
3575 // Having loaded both klasses, test each for null.
3576 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3577 for (which_arg = 0; which_arg <= 1; which_arg++) {
3578 Node* kls = klasses[which_arg];
3579 Node* null_ctl = top();
3580 kls = null_check_oop(kls, &null_ctl, never_see_null);
3581 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3582 region->init_req(prim_path, null_ctl);
3583 if (stopped()) break;
3584 klasses[which_arg] = kls;
3585 }
3586
3587 if (!stopped()) {
3588 // now we have two reference types, in klasses[0..1]
3589 Node* subk = klasses[1]; // the argument to isAssignableFrom
3590 Node* superk = klasses[0]; // the receiver
3591 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3592 // now we have a successful reference subtype check
3593 region->set_req(_ref_subtype_path, control());
3594 }
3595
3596 // If both operands are primitive (both klasses null), then
3597 // we must return true when they are identical primitives.
3598 // It is convenient to test this after the first null klass check.
3599 set_control(region->in(_prim_0_path)); // go back to first null check
3600 if (!stopped()) {
3601 // Since superc is primitive, make a guard for the superc==subc case.
3602 shenandoah_acmp_barrier(args[0], args[1]);
3603 Node* cmp_eq = _gvn.transform(new CmpPNode(args[0], args[1]));
3604 Node* bol_eq = _gvn.transform(new BoolNode(cmp_eq, BoolTest::eq));
3605 generate_guard(bol_eq, region, PROB_FAIR);
3606 if (region->req() == PATH_LIMIT+1) {
3607 // A guard was added. If the added guard is taken, superc==subc.
3608 region->swap_edges(PATH_LIMIT, _prim_same_path);
3609 region->del_req(PATH_LIMIT);
3610 }
3611 region->set_req(_prim_0_path, control()); // Not equal after all.
3612 }
3613
3614 // these are the only paths that produce 'true':
3615 phi->set_req(_prim_same_path, intcon(1));
3616 phi->set_req(_ref_subtype_path, intcon(1));
3617
3618 // pull together the cases:
3619 assert(region->req() == PATH_LIMIT, "sane region");
3620 for (uint i = 1; i < region->req(); i++) {
3621 Node* ctl = region->in(i);
3622 if (ctl == NULL || ctl == top()) {
3623 region->set_req(i, top());
3624 phi ->set_req(i, top());
3625 } else if (phi->in(i) == NULL) {
3626 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3627 }
3628 }
3629
3630 set_control(_gvn.transform(region));
3631 set_result(_gvn.transform(phi));
3632 return true;
3633 }
3634
3635 //---------------------generate_array_guard_common------------------------
3636 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3637 bool obj_array, bool not_array) {
3638
3639 if (stopped()) {
3640 return NULL;
3641 }
3642
3643 // If obj_array/non_array==false/false:
3644 // Branch around if the given klass is in fact an array (either obj or prim).
3645 // If obj_array/non_array==false/true:
3646 // Branch around if the given klass is not an array klass of any kind.
3647 // If obj_array/non_array==true/true:
3648 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3649 // If obj_array/non_array==true/false:
3650 // Branch around if the kls is an oop array (Object[] or subtype)
3651 //
3652 // Like generate_guard, adds a new path onto the region.
3653 jint layout_con = 0;
3654 Node* layout_val = get_layout_helper(kls, layout_con);
3655 if (layout_val == NULL) {
3656 bool query = (obj_array
3657 ? Klass::layout_helper_is_objArray(layout_con)
3658 : Klass::layout_helper_is_array(layout_con));
3659 if (query == not_array) {
3660 return NULL; // never a branch
3661 } else { // always a branch
3662 Node* always_branch = control();
3663 if (region != NULL)
3664 region->add_req(always_branch);
3665 set_control(top());
3666 return always_branch;
3667 }
3668 }
3669 // Now test the correct condition.
3670 jint nval = (obj_array
3671 ? ((jint)Klass::_lh_array_tag_type_value
3672 << Klass::_lh_array_tag_shift)
3673 : Klass::_lh_neutral_value);
3674 Node* cmp = _gvn.transform(new CmpINode(layout_val, intcon(nval)));
3675 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3676 // invert the test if we are looking for a non-array
3677 if (not_array) btest = BoolTest(btest).negate();
3678 Node* bol = _gvn.transform(new BoolNode(cmp, btest));
3679 return generate_fair_guard(bol, region);
3680 }
3681
3682
3683 //-----------------------inline_native_newArray--------------------------
3684 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3685 bool LibraryCallKit::inline_native_newArray() {
3686 Node* mirror = argument(0);
3687 Node* count_val = argument(1);
3688
3689 mirror = null_check(mirror);
3690 // If mirror or obj is dead, only null-path is taken.
3691 if (stopped()) return true;
3692
3693 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3694 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
3695 PhiNode* result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
3696 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
3697 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3698
3699 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3700 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3701 result_reg, _slow_path);
3702 Node* normal_ctl = control();
3703 Node* no_array_ctl = result_reg->in(_slow_path);
3704
3705 // Generate code for the slow case. We make a call to newArray().
3706 set_control(no_array_ctl);
3707 if (!stopped()) {
3708 // Either the input type is void.class, or else the
3709 // array klass has not yet been cached. Either the
3710 // ensuing call will throw an exception, or else it
3711 // will cache the array klass for next time.
3712 PreserveJVMState pjvms(this);
3713 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3714 Node* slow_result = set_results_for_java_call(slow_call);
3715 // this->control() comes from set_results_for_java_call
3716 result_reg->set_req(_slow_path, control());
3717 result_val->set_req(_slow_path, slow_result);
3718 result_io ->set_req(_slow_path, i_o());
3719 result_mem->set_req(_slow_path, reset_memory());
3720 }
3721
3722 set_control(normal_ctl);
3723 if (!stopped()) {
3724 // Normal case: The array type has been cached in the java.lang.Class.
3725 // The following call works fine even if the array type is polymorphic.
3726 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3727 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3728 result_reg->init_req(_normal_path, control());
3729 result_val->init_req(_normal_path, obj);
3730 result_io ->init_req(_normal_path, i_o());
3731 result_mem->init_req(_normal_path, reset_memory());
3732 }
3733
3734 // Return the combined state.
3735 set_i_o( _gvn.transform(result_io) );
3736 set_all_memory( _gvn.transform(result_mem));
3737
3738 C->set_has_split_ifs(true); // Has chance for split-if optimization
3739 set_result(result_reg, result_val);
3740 return true;
3741 }
3742
3743 //----------------------inline_native_getLength--------------------------
3744 // public static native int java.lang.reflect.Array.getLength(Object array);
3745 bool LibraryCallKit::inline_native_getLength() {
3746 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3747
3748 Node* array = null_check(argument(0));
3749 // If array is dead, only null-path is taken.
3750 if (stopped()) return true;
3751
3752 // Deoptimize if it is a non-array.
3753 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3754
3755 if (non_array != NULL) {
3756 PreserveJVMState pjvms(this);
3757 set_control(non_array);
3758 uncommon_trap(Deoptimization::Reason_intrinsic,
3759 Deoptimization::Action_maybe_recompile);
3760 }
3761
3762 // If control is dead, only non-array-path is taken.
3763 if (stopped()) return true;
3764
3765 // The works fine even if the array type is polymorphic.
3766 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3767 Node* result = load_array_length(array);
3768
3769 C->set_has_split_ifs(true); // Has chance for split-if optimization
3770 set_result(result);
3771 return true;
3772 }
3773
3774 //------------------------inline_array_copyOf----------------------------
3775 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3776 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3777 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3778 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3779
3780 // Get the arguments.
3781 Node* original = argument(0);
3782 Node* start = is_copyOfRange? argument(1): intcon(0);
3783 Node* end = is_copyOfRange? argument(2): argument(1);
3784 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3785
3786 Node* newcopy;
3787
3788 // Set the original stack and the reexecute bit for the interpreter to reexecute
3789 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3790 { PreserveReexecuteState preexecs(this);
3791 jvms()->set_should_reexecute(true);
3792
3793 array_type_mirror = null_check(array_type_mirror);
3794 original = null_check(original);
3795
3796 // Check if a null path was taken unconditionally.
3797 if (stopped()) return true;
3798
3799 Node* orig_length = load_array_length(original);
3800
3801 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3802 klass_node = null_check(klass_node);
3803
3804 RegionNode* bailout = new RegionNode(1);
3805 record_for_igvn(bailout);
3806
3807 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3808 // Bail out if that is so.
3809 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3810 if (not_objArray != NULL) {
3811 // Improve the klass node's type from the new optimistic assumption:
3812 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3813 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3814 Node* cast = new CastPPNode(klass_node, akls);
3815 cast->init_req(0, control());
3816 klass_node = _gvn.transform(cast);
3817 }
3818
3819 // Bail out if either start or end is negative.
3820 generate_negative_guard(start, bailout, &start);
3821 generate_negative_guard(end, bailout, &end);
3822
3823 Node* length = end;
3824 if (_gvn.type(start) != TypeInt::ZERO) {
3825 length = _gvn.transform(new SubINode(end, start));
3826 }
3827
3828 // Bail out if length is negative.
3829 // Without this the new_array would throw
3830 // NegativeArraySizeException but IllegalArgumentException is what
3831 // should be thrown
3832 generate_negative_guard(length, bailout, &length);
3833
3834 if (bailout->req() > 1) {
3835 PreserveJVMState pjvms(this);
3836 set_control(_gvn.transform(bailout));
3837 uncommon_trap(Deoptimization::Reason_intrinsic,
3838 Deoptimization::Action_maybe_recompile);
3839 }
3840
3841 if (!stopped()) {
3842 // How many elements will we copy from the original?
3843 // The answer is MinI(orig_length - start, length).
3844 Node* orig_tail = _gvn.transform(new SubINode(orig_length, start));
3845 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3846
3847 original = shenandoah_read_barrier(original);
3848
3849 // Generate a direct call to the right arraycopy function(s).
3850 // We know the copy is disjoint but we might not know if the
3851 // oop stores need checking.
3852 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3853 // This will fail a store-check if x contains any non-nulls.
3854
3855 // ArrayCopyNode:Ideal may transform the ArrayCopyNode to
3856 // loads/stores but it is legal only if we're sure the
3857 // Arrays.copyOf would succeed. So we need all input arguments
3858 // to the copyOf to be validated, including that the copy to the
3859 // new array won't trigger an ArrayStoreException. That subtype
3860 // check can be optimized if we know something on the type of
3861 // the input array from type speculation.
3862 if (_gvn.type(klass_node)->singleton()) {
3863 ciKlass* subk = _gvn.type(load_object_klass(original))->is_klassptr()->klass();
3864 ciKlass* superk = _gvn.type(klass_node)->is_klassptr()->klass();
3865
3866 int test = C->static_subtype_check(superk, subk);
3867 if (test != Compile::SSC_always_true && test != Compile::SSC_always_false) {
3868 const TypeOopPtr* t_original = _gvn.type(original)->is_oopptr();
3869 if (t_original->speculative_type() != NULL) {
3870 original = maybe_cast_profiled_obj(original, t_original->speculative_type(), true);
3871 }
3872 }
3873 }
3874
3875 bool validated = false;
3876 // Reason_class_check rather than Reason_intrinsic because we
3877 // want to intrinsify even if this traps.
3878 if (!too_many_traps(Deoptimization::Reason_class_check)) {
3879 Node* not_subtype_ctrl = gen_subtype_check(load_object_klass(original),
3880 klass_node);
3881
3882 if (not_subtype_ctrl != top()) {
3883 PreserveJVMState pjvms(this);
3884 set_control(not_subtype_ctrl);
3885 uncommon_trap(Deoptimization::Reason_class_check,
3886 Deoptimization::Action_make_not_entrant);
3887 assert(stopped(), "Should be stopped");
3888 }
3889 validated = true;
3890 }
3891
3892 if (!stopped()) {
3893 newcopy = new_array(klass_node, length, 0); // no arguments to push
3894
3895 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, original, start, newcopy, intcon(0), moved, true,
3896 load_object_klass(original), klass_node);
3897 if (!is_copyOfRange) {
3898 ac->set_copyof(validated);
3899 } else {
3900 ac->set_copyofrange(validated);
3901 }
3902 Node* n = _gvn.transform(ac);
3903 if (n == ac) {
3904 ac->connect_outputs(this);
3905 } else {
3906 assert(validated, "shouldn't transform if all arguments not validated");
3907 set_all_memory(n);
3908 }
3909 }
3910 }
3911 } // original reexecute is set back here
3912
3913 C->set_has_split_ifs(true); // Has chance for split-if optimization
3914 if (!stopped()) {
3915 set_result(newcopy);
3916 }
3917 return true;
3918 }
3919
3920
3921 //----------------------generate_virtual_guard---------------------------
3922 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3923 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3924 RegionNode* slow_region) {
3925 ciMethod* method = callee();
3926 int vtable_index = method->vtable_index();
3927 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3928 err_msg_res("bad index %d", vtable_index));
3929 // Get the Method* out of the appropriate vtable entry.
3930 int entry_offset = (InstanceKlass::vtable_start_offset() +
3931 vtable_index*vtableEntry::size()) * wordSize +
3932 vtableEntry::method_offset_in_bytes();
3933 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3934 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3935
3936 // Compare the target method with the expected method (e.g., Object.hashCode).
3937 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3938
3939 Node* native_call = makecon(native_call_addr);
3940 Node* chk_native = _gvn.transform(new CmpPNode(target_call, native_call));
3941 Node* test_native = _gvn.transform(new BoolNode(chk_native, BoolTest::ne));
3942
3943 return generate_slow_guard(test_native, slow_region);
3944 }
3945
3946 //-----------------------generate_method_call----------------------------
3947 // Use generate_method_call to make a slow-call to the real
3948 // method if the fast path fails. An alternative would be to
3949 // use a stub like OptoRuntime::slow_arraycopy_Java.
3950 // This only works for expanding the current library call,
3951 // not another intrinsic. (E.g., don't use this for making an
3952 // arraycopy call inside of the copyOf intrinsic.)
3953 CallJavaNode*
3954 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3955 // When compiling the intrinsic method itself, do not use this technique.
3956 guarantee(callee() != C->method(), "cannot make slow-call to self");
3957
3958 ciMethod* method = callee();
3959 // ensure the JVMS we have will be correct for this call
3960 guarantee(method_id == method->intrinsic_id(), "must match");
3961
3962 const TypeFunc* tf = TypeFunc::make(method);
3963 CallJavaNode* slow_call;
3964 if (is_static) {
3965 assert(!is_virtual, "");
3966 slow_call = new CallStaticJavaNode(C, tf,
3967 SharedRuntime::get_resolve_static_call_stub(),
3968 method, bci());
3969 } else if (is_virtual) {
3970 null_check_receiver();
3971 int vtable_index = Method::invalid_vtable_index;
3972 if (UseInlineCaches) {
3973 // Suppress the vtable call
3974 } else {
3975 // hashCode and clone are not a miranda methods,
3976 // so the vtable index is fixed.
3977 // No need to use the linkResolver to get it.
3978 vtable_index = method->vtable_index();
3979 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3980 err_msg_res("bad index %d", vtable_index));
3981 }
3982 slow_call = new CallDynamicJavaNode(tf,
3983 SharedRuntime::get_resolve_virtual_call_stub(),
3984 method, vtable_index, bci());
3985 } else { // neither virtual nor static: opt_virtual
3986 null_check_receiver();
3987 slow_call = new CallStaticJavaNode(C, tf,
3988 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3989 method, bci());
3990 slow_call->set_optimized_virtual(true);
3991 }
3992 set_arguments_for_java_call(slow_call);
3993 set_edges_for_java_call(slow_call);
3994 return slow_call;
3995 }
3996
3997
3998 /**
3999 * Build special case code for calls to hashCode on an object. This call may
4000 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
4001 * slightly different code.
4002 */
4003 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4004 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4005 assert(!(is_virtual && is_static), "either virtual, special, or static");
4006
4007 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4008
4009 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4010 PhiNode* result_val = new PhiNode(result_reg, TypeInt::INT);
4011 PhiNode* result_io = new PhiNode(result_reg, Type::ABIO);
4012 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4013 Node* obj = NULL;
4014 if (!is_static) {
4015 // Check for hashing null object
4016 obj = null_check_receiver();
4017 if (stopped()) return true; // unconditionally null
4018 result_reg->init_req(_null_path, top());
4019 result_val->init_req(_null_path, top());
4020 } else {
4021 // Do a null check, and return zero if null.
4022 // System.identityHashCode(null) == 0
4023 obj = argument(0);
4024 Node* null_ctl = top();
4025 obj = null_check_oop(obj, &null_ctl);
4026 result_reg->init_req(_null_path, null_ctl);
4027 result_val->init_req(_null_path, _gvn.intcon(0));
4028 }
4029
4030 if (ShenandoahVerifyReadsToFromSpace) {
4031 obj = shenandoah_read_barrier(obj);
4032 }
4033
4034 // Unconditionally null? Then return right away.
4035 if (stopped()) {
4036 set_control( result_reg->in(_null_path));
4037 if (!stopped())
4038 set_result(result_val->in(_null_path));
4039 return true;
4040 }
4041
4042 // We only go to the fast case code if we pass a number of guards. The
4043 // paths which do not pass are accumulated in the slow_region.
4044 RegionNode* slow_region = new RegionNode(1);
4045 record_for_igvn(slow_region);
4046
4047 // If this is a virtual call, we generate a funny guard. We pull out
4048 // the vtable entry corresponding to hashCode() from the target object.
4049 // If the target method which we are calling happens to be the native
4050 // Object hashCode() method, we pass the guard. We do not need this
4051 // guard for non-virtual calls -- the caller is known to be the native
4052 // Object hashCode().
4053 if (is_virtual) {
4054 // After null check, get the object's klass.
4055 Node* obj_klass = load_object_klass(obj);
4056 generate_virtual_guard(obj_klass, slow_region);
4057 }
4058
4059 // Get the header out of the object, use LoadMarkNode when available
4060 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4061 // The control of the load must be NULL. Otherwise, the load can move before
4062 // the null check after castPP removal.
4063 Node* no_ctrl = NULL;
4064 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4065
4066 // Test the header to see if it is unlocked.
4067 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4068 Node *lmasked_header = _gvn.transform(new AndXNode(header, lock_mask));
4069 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4070 Node *chk_unlocked = _gvn.transform(new CmpXNode( lmasked_header, unlocked_val));
4071 Node *test_unlocked = _gvn.transform(new BoolNode( chk_unlocked, BoolTest::ne));
4072
4073 generate_slow_guard(test_unlocked, slow_region);
4074
4075 // Get the hash value and check to see that it has been properly assigned.
4076 // We depend on hash_mask being at most 32 bits and avoid the use of
4077 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4078 // vm: see markOop.hpp.
4079 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4080 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4081 Node *hshifted_header= _gvn.transform(new URShiftXNode(header, hash_shift));
4082 // This hack lets the hash bits live anywhere in the mark object now, as long
4083 // as the shift drops the relevant bits into the low 32 bits. Note that
4084 // Java spec says that HashCode is an int so there's no point in capturing
4085 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4086 hshifted_header = ConvX2I(hshifted_header);
4087 Node *hash_val = _gvn.transform(new AndINode(hshifted_header, hash_mask));
4088
4089 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4090 Node *chk_assigned = _gvn.transform(new CmpINode( hash_val, no_hash_val));
4091 Node *test_assigned = _gvn.transform(new BoolNode( chk_assigned, BoolTest::eq));
4092
4093 generate_slow_guard(test_assigned, slow_region);
4094
4095 Node* init_mem = reset_memory();
4096 // fill in the rest of the null path:
4097 result_io ->init_req(_null_path, i_o());
4098 result_mem->init_req(_null_path, init_mem);
4099
4100 result_val->init_req(_fast_path, hash_val);
4101 result_reg->init_req(_fast_path, control());
4102 result_io ->init_req(_fast_path, i_o());
4103 result_mem->init_req(_fast_path, init_mem);
4104
4105 // Generate code for the slow case. We make a call to hashCode().
4106 set_control(_gvn.transform(slow_region));
4107 if (!stopped()) {
4108 // No need for PreserveJVMState, because we're using up the present state.
4109 set_all_memory(init_mem);
4110 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4111 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4112 Node* slow_result = set_results_for_java_call(slow_call);
4113 // this->control() comes from set_results_for_java_call
4114 result_reg->init_req(_slow_path, control());
4115 result_val->init_req(_slow_path, slow_result);
4116 result_io ->set_req(_slow_path, i_o());
4117 result_mem ->set_req(_slow_path, reset_memory());
4118 }
4119
4120 // Return the combined state.
4121 set_i_o( _gvn.transform(result_io) );
4122 set_all_memory( _gvn.transform(result_mem));
4123
4124 set_result(result_reg, result_val);
4125 return true;
4126 }
4127
4128 //---------------------------inline_native_getClass----------------------------
4129 // public final native Class<?> java.lang.Object.getClass();
4130 //
4131 // Build special case code for calls to getClass on an object.
4132 bool LibraryCallKit::inline_native_getClass() {
4133 Node* obj = null_check_receiver();
4134 if (stopped()) return true;
4135 set_result(load_mirror_from_klass(load_object_klass(obj)));
4136 return true;
4137 }
4138
4139 //-----------------inline_native_Reflection_getCallerClass---------------------
4140 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4141 //
4142 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4143 //
4144 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4145 // in that it must skip particular security frames and checks for
4146 // caller sensitive methods.
4147 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4148 #ifndef PRODUCT
4149 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4150 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4151 }
4152 #endif
4153
4154 if (!jvms()->has_method()) {
4155 #ifndef PRODUCT
4156 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4157 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4158 }
4159 #endif
4160 return false;
4161 }
4162
4163 // Walk back up the JVM state to find the caller at the required
4164 // depth.
4165 JVMState* caller_jvms = jvms();
4166
4167 // Cf. JVM_GetCallerClass
4168 // NOTE: Start the loop at depth 1 because the current JVM state does
4169 // not include the Reflection.getCallerClass() frame.
4170 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4171 ciMethod* m = caller_jvms->method();
4172 switch (n) {
4173 case 0:
4174 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4175 break;
4176 case 1:
4177 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4178 if (!m->caller_sensitive()) {
4179 #ifndef PRODUCT
4180 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4181 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4182 }
4183 #endif
4184 return false; // bail-out; let JVM_GetCallerClass do the work
4185 }
4186 break;
4187 default:
4188 if (!m->is_ignored_by_security_stack_walk()) {
4189 // We have reached the desired frame; return the holder class.
4190 // Acquire method holder as java.lang.Class and push as constant.
4191 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4192 ciInstance* caller_mirror = caller_klass->java_mirror();
4193 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4194
4195 #ifndef PRODUCT
4196 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4197 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());
4198 tty->print_cr(" JVM state at this point:");
4199 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4200 ciMethod* m = jvms()->of_depth(i)->method();
4201 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4202 }
4203 }
4204 #endif
4205 return true;
4206 }
4207 break;
4208 }
4209 }
4210
4211 #ifndef PRODUCT
4212 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4213 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4214 tty->print_cr(" JVM state at this point:");
4215 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4216 ciMethod* m = jvms()->of_depth(i)->method();
4217 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4218 }
4219 }
4220 #endif
4221
4222 return false; // bail-out; let JVM_GetCallerClass do the work
4223 }
4224
4225 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4226 Node* arg = argument(0);
4227 Node* result;
4228
4229 switch (id) {
4230 case vmIntrinsics::_floatToRawIntBits: result = new MoveF2INode(arg); break;
4231 case vmIntrinsics::_intBitsToFloat: result = new MoveI2FNode(arg); break;
4232 case vmIntrinsics::_doubleToRawLongBits: result = new MoveD2LNode(arg); break;
4233 case vmIntrinsics::_longBitsToDouble: result = new MoveL2DNode(arg); break;
4234
4235 case vmIntrinsics::_doubleToLongBits: {
4236 // two paths (plus control) merge in a wood
4237 RegionNode *r = new RegionNode(3);
4238 Node *phi = new PhiNode(r, TypeLong::LONG);
4239
4240 Node *cmpisnan = _gvn.transform(new CmpDNode(arg, arg));
4241 // Build the boolean node
4242 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4243
4244 // Branch either way.
4245 // NaN case is less traveled, which makes all the difference.
4246 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4247 Node *opt_isnan = _gvn.transform(ifisnan);
4248 assert( opt_isnan->is_If(), "Expect an IfNode");
4249 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4250 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4251
4252 set_control(iftrue);
4253
4254 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4255 Node *slow_result = longcon(nan_bits); // return NaN
4256 phi->init_req(1, _gvn.transform( slow_result ));
4257 r->init_req(1, iftrue);
4258
4259 // Else fall through
4260 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4261 set_control(iffalse);
4262
4263 phi->init_req(2, _gvn.transform(new MoveD2LNode(arg)));
4264 r->init_req(2, iffalse);
4265
4266 // Post merge
4267 set_control(_gvn.transform(r));
4268 record_for_igvn(r);
4269
4270 C->set_has_split_ifs(true); // Has chance for split-if optimization
4271 result = phi;
4272 assert(result->bottom_type()->isa_long(), "must be");
4273 break;
4274 }
4275
4276 case vmIntrinsics::_floatToIntBits: {
4277 // two paths (plus control) merge in a wood
4278 RegionNode *r = new RegionNode(3);
4279 Node *phi = new PhiNode(r, TypeInt::INT);
4280
4281 Node *cmpisnan = _gvn.transform(new CmpFNode(arg, arg));
4282 // Build the boolean node
4283 Node *bolisnan = _gvn.transform(new BoolNode(cmpisnan, BoolTest::ne));
4284
4285 // Branch either way.
4286 // NaN case is less traveled, which makes all the difference.
4287 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4288 Node *opt_isnan = _gvn.transform(ifisnan);
4289 assert( opt_isnan->is_If(), "Expect an IfNode");
4290 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4291 Node *iftrue = _gvn.transform(new IfTrueNode(opt_ifisnan));
4292
4293 set_control(iftrue);
4294
4295 static const jint nan_bits = 0x7fc00000;
4296 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4297 phi->init_req(1, _gvn.transform( slow_result ));
4298 r->init_req(1, iftrue);
4299
4300 // Else fall through
4301 Node *iffalse = _gvn.transform(new IfFalseNode(opt_ifisnan));
4302 set_control(iffalse);
4303
4304 phi->init_req(2, _gvn.transform(new MoveF2INode(arg)));
4305 r->init_req(2, iffalse);
4306
4307 // Post merge
4308 set_control(_gvn.transform(r));
4309 record_for_igvn(r);
4310
4311 C->set_has_split_ifs(true); // Has chance for split-if optimization
4312 result = phi;
4313 assert(result->bottom_type()->isa_int(), "must be");
4314 break;
4315 }
4316
4317 default:
4318 fatal_unexpected_iid(id);
4319 break;
4320 }
4321 set_result(_gvn.transform(result));
4322 return true;
4323 }
4324
4325 #ifdef _LP64
4326 #define XTOP ,top() /*additional argument*/
4327 #else //_LP64
4328 #define XTOP /*no additional argument*/
4329 #endif //_LP64
4330
4331 //----------------------inline_unsafe_copyMemory-------------------------
4332 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4333 bool LibraryCallKit::inline_unsafe_copyMemory() {
4334 if (callee()->is_static()) return false; // caller must have the capability!
4335 null_check_receiver(); // null-check receiver
4336 if (stopped()) return true;
4337
4338 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4339
4340 Node* src_ptr = argument(1); // type: oop
4341 Node* src_off = ConvL2X(argument(2)); // type: long
4342 Node* dst_ptr = argument(4); // type: oop
4343 Node* dst_off = ConvL2X(argument(5)); // type: long
4344 Node* size = ConvL2X(argument(7)); // type: long
4345
4346 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4347 "fieldOffset must be byte-scaled");
4348
4349 src_ptr = shenandoah_read_barrier(src_ptr);
4350 dst_ptr = shenandoah_write_barrier(dst_ptr);
4351
4352 Node* src = make_unsafe_address(src_ptr, src_off);
4353 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4354
4355 // Conservatively insert a memory barrier on all memory slices.
4356 // Do not let writes of the copy source or destination float below the copy.
4357 insert_mem_bar(Op_MemBarCPUOrder);
4358
4359 // Call it. Note that the length argument is not scaled.
4360 make_runtime_call(RC_LEAF|RC_NO_FP,
4361 OptoRuntime::fast_arraycopy_Type(),
4362 StubRoutines::unsafe_arraycopy(),
4363 "unsafe_arraycopy",
4364 TypeRawPtr::BOTTOM,
4365 src, dst, size XTOP);
4366
4367 // Do not let reads of the copy destination float above the copy.
4368 insert_mem_bar(Op_MemBarCPUOrder);
4369
4370 return true;
4371 }
4372
4373 //------------------------clone_coping-----------------------------------
4374 // Helper function for inline_native_clone.
4375 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4376 assert(obj_size != NULL, "");
4377 Node* raw_obj = alloc_obj->in(1);
4378 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4379
4380 obj = shenandoah_read_barrier(obj);
4381
4382 AllocateNode* alloc = NULL;
4383 if (ReduceBulkZeroing) {
4384 // We will be completely responsible for initializing this object -
4385 // mark Initialize node as complete.
4386 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4387 // The object was just allocated - there should be no any stores!
4388 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4389 // Mark as complete_with_arraycopy so that on AllocateNode
4390 // expansion, we know this AllocateNode is initialized by an array
4391 // copy and a StoreStore barrier exists after the array copy.
4392 alloc->initialization()->set_complete_with_arraycopy();
4393 }
4394
4395 // Copy the fastest available way.
4396 // TODO: generate fields copies for small objects instead.
4397 Node* src = obj;
4398 Node* dest = alloc_obj;
4399 Node* size = _gvn.transform(obj_size);
4400
4401 // Exclude the header but include array length to copy by 8 bytes words.
4402 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4403 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4404 instanceOopDesc::base_offset_in_bytes();
4405 // base_off:
4406 // 8 - 32-bit VM
4407 // 12 - 64-bit VM, compressed klass
4408 // 16 - 64-bit VM, normal klass
4409 if (base_off % BytesPerLong != 0) {
4410 assert(UseCompressedClassPointers, "");
4411 if (is_array) {
4412 // Exclude length to copy by 8 bytes words.
4413 base_off += sizeof(int);
4414 } else {
4415 // Include klass to copy by 8 bytes words.
4416 base_off = instanceOopDesc::klass_offset_in_bytes();
4417 }
4418 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4419 }
4420 src = basic_plus_adr(src, base_off);
4421 dest = basic_plus_adr(dest, base_off);
4422
4423 // Compute the length also, if needed:
4424 Node* countx = size;
4425 countx = _gvn.transform(new SubXNode(countx, MakeConX(base_off)));
4426 countx = _gvn.transform(new URShiftXNode(countx, intcon(LogBytesPerLong) ));
4427
4428 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4429
4430 ArrayCopyNode* ac = ArrayCopyNode::make(this, false, src, NULL, dest, NULL, countx, false);
4431 ac->set_clonebasic();
4432 Node* n = _gvn.transform(ac);
4433 if (n == ac) {
4434 set_predefined_output_for_runtime_call(ac, ac->in(TypeFunc::Memory), raw_adr_type);
4435 } else {
4436 set_all_memory(n);
4437 }
4438
4439 if (UseShenandoahGC) {
4440 // Make sure that references in the cloned object are updated for Shenandoah.
4441 make_runtime_call(RC_LEAF|RC_NO_FP,
4442 OptoRuntime::shenandoah_clone_barrier_Type(),
4443 CAST_FROM_FN_PTR(address, SharedRuntime::shenandoah_clone_barrier),
4444 "shenandoah_clone_barrier", TypePtr::BOTTOM,
4445 alloc_obj);
4446 }
4447
4448 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4449 if (card_mark) {
4450 assert(!is_array, "");
4451 // Put in store barrier for any and all oops we are sticking
4452 // into this object. (We could avoid this if we could prove
4453 // that the object type contains no oop fields at all.)
4454 Node* no_particular_value = NULL;
4455 Node* no_particular_field = NULL;
4456 int raw_adr_idx = Compile::AliasIdxRaw;
4457 post_barrier(control(),
4458 memory(raw_adr_type),
4459 alloc_obj,
4460 no_particular_field,
4461 raw_adr_idx,
4462 no_particular_value,
4463 T_OBJECT,
4464 false);
4465 }
4466
4467 // Do not let reads from the cloned object float above the arraycopy.
4468 if (alloc != NULL) {
4469 // Do not let stores that initialize this object be reordered with
4470 // a subsequent store that would make this object accessible by
4471 // other threads.
4472 // Record what AllocateNode this StoreStore protects so that
4473 // escape analysis can go from the MemBarStoreStoreNode to the
4474 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4475 // based on the escape status of the AllocateNode.
4476 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4477 } else {
4478 insert_mem_bar(Op_MemBarCPUOrder);
4479 }
4480 }
4481
4482 //------------------------inline_native_clone----------------------------
4483 // protected native Object java.lang.Object.clone();
4484 //
4485 // Here are the simple edge cases:
4486 // null receiver => normal trap
4487 // virtual and clone was overridden => slow path to out-of-line clone
4488 // not cloneable or finalizer => slow path to out-of-line Object.clone
4489 //
4490 // The general case has two steps, allocation and copying.
4491 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4492 //
4493 // Copying also has two cases, oop arrays and everything else.
4494 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4495 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4496 //
4497 // These steps fold up nicely if and when the cloned object's klass
4498 // can be sharply typed as an object array, a type array, or an instance.
4499 //
4500 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4501 PhiNode* result_val;
4502
4503 // Set the reexecute bit for the interpreter to reexecute
4504 // the bytecode that invokes Object.clone if deoptimization happens.
4505 { PreserveReexecuteState preexecs(this);
4506 jvms()->set_should_reexecute(true);
4507
4508 Node* obj = null_check_receiver();
4509 if (stopped()) return true;
4510
4511 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
4512
4513 // If we are going to clone an instance, we need its exact type to
4514 // know the number and types of fields to convert the clone to
4515 // loads/stores. Maybe a speculative type can help us.
4516 if (!obj_type->klass_is_exact() &&
4517 obj_type->speculative_type() != NULL &&
4518 obj_type->speculative_type()->is_instance_klass()) {
4519 ciInstanceKlass* spec_ik = obj_type->speculative_type()->as_instance_klass();
4520 if (spec_ik->nof_nonstatic_fields() <= ArrayCopyLoadStoreMaxElem &&
4521 !spec_ik->has_injected_fields()) {
4522 ciKlass* k = obj_type->klass();
4523 if (!k->is_instance_klass() ||
4524 k->as_instance_klass()->is_interface() ||
4525 k->as_instance_klass()->has_subklass()) {
4526 obj = maybe_cast_profiled_obj(obj, obj_type->speculative_type(), false);
4527 }
4528 }
4529 }
4530
4531 Node* obj_klass = load_object_klass(obj);
4532 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4533 const TypeOopPtr* toop = ((tklass != NULL)
4534 ? tklass->as_instance_type()
4535 : TypeInstPtr::NOTNULL);
4536
4537 // Conservatively insert a memory barrier on all memory slices.
4538 // Do not let writes into the original float below the clone.
4539 insert_mem_bar(Op_MemBarCPUOrder);
4540
4541 // paths into result_reg:
4542 enum {
4543 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4544 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4545 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4546 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4547 PATH_LIMIT
4548 };
4549 RegionNode* result_reg = new RegionNode(PATH_LIMIT);
4550 result_val = new PhiNode(result_reg, TypeInstPtr::NOTNULL);
4551 PhiNode* result_i_o = new PhiNode(result_reg, Type::ABIO);
4552 PhiNode* result_mem = new PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
4553 record_for_igvn(result_reg);
4554
4555 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4556 int raw_adr_idx = Compile::AliasIdxRaw;
4557
4558 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4559 if (array_ctl != NULL) {
4560 // It's an array.
4561 PreserveJVMState pjvms(this);
4562 set_control(array_ctl);
4563 Node* obj_length = load_array_length(obj);
4564 Node* obj_size = NULL;
4565 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4566
4567 if (!use_ReduceInitialCardMarks()) {
4568 // If it is an oop array, it requires very special treatment,
4569 // because card marking is required on each card of the array.
4570 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4571 if (is_obja != NULL) {
4572 PreserveJVMState pjvms2(this);
4573 set_control(is_obja);
4574
4575 obj = shenandoah_read_barrier(obj);
4576
4577 // Generate a direct call to the right arraycopy function(s).
4578 Node* alloc = tightly_coupled_allocation(alloc_obj, NULL);
4579 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, obj, intcon(0), alloc_obj, intcon(0), obj_length, alloc != NULL);
4580 ac->set_cloneoop();
4581 Node* n = _gvn.transform(ac);
4582 assert(n == ac, "cannot disappear");
4583 ac->connect_outputs(this);
4584
4585 result_reg->init_req(_objArray_path, control());
4586 result_val->init_req(_objArray_path, alloc_obj);
4587 result_i_o ->set_req(_objArray_path, i_o());
4588 result_mem ->set_req(_objArray_path, reset_memory());
4589 }
4590 }
4591 // Otherwise, there are no card marks to worry about.
4592 // (We can dispense with card marks if we know the allocation
4593 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4594 // causes the non-eden paths to take compensating steps to
4595 // simulate a fresh allocation, so that no further
4596 // card marks are required in compiled code to initialize
4597 // the object.)
4598
4599 if (!stopped()) {
4600 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4601
4602 // Present the results of the copy.
4603 result_reg->init_req(_array_path, control());
4604 result_val->init_req(_array_path, alloc_obj);
4605 result_i_o ->set_req(_array_path, i_o());
4606 result_mem ->set_req(_array_path, reset_memory());
4607 }
4608 }
4609
4610 // We only go to the instance fast case code if we pass a number of guards.
4611 // The paths which do not pass are accumulated in the slow_region.
4612 RegionNode* slow_region = new RegionNode(1);
4613 record_for_igvn(slow_region);
4614 if (!stopped()) {
4615 // It's an instance (we did array above). Make the slow-path tests.
4616 // If this is a virtual call, we generate a funny guard. We grab
4617 // the vtable entry corresponding to clone() from the target object.
4618 // If the target method which we are calling happens to be the
4619 // Object clone() method, we pass the guard. We do not need this
4620 // guard for non-virtual calls; the caller is known to be the native
4621 // Object clone().
4622 if (is_virtual) {
4623 generate_virtual_guard(obj_klass, slow_region);
4624 }
4625
4626 // The object must be cloneable and must not have a finalizer.
4627 // Both of these conditions may be checked in a single test.
4628 // We could optimize the cloneable test further, but we don't care.
4629 generate_access_flags_guard(obj_klass,
4630 // Test both conditions:
4631 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4632 // Must be cloneable but not finalizer:
4633 JVM_ACC_IS_CLONEABLE,
4634 slow_region);
4635 }
4636
4637 if (!stopped()) {
4638 // It's an instance, and it passed the slow-path tests.
4639 PreserveJVMState pjvms(this);
4640 Node* obj_size = NULL;
4641 // Need to deoptimize on exception from allocation since Object.clone intrinsic
4642 // is reexecuted if deoptimization occurs and there could be problems when merging
4643 // exception state between multiple Object.clone versions (reexecute=true vs reexecute=false).
4644 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size, /*deoptimize_on_exception=*/true);
4645
4646 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4647
4648 // Present the results of the slow call.
4649 result_reg->init_req(_instance_path, control());
4650 result_val->init_req(_instance_path, alloc_obj);
4651 result_i_o ->set_req(_instance_path, i_o());
4652 result_mem ->set_req(_instance_path, reset_memory());
4653 }
4654
4655 // Generate code for the slow case. We make a call to clone().
4656 set_control(_gvn.transform(slow_region));
4657 if (!stopped()) {
4658 PreserveJVMState pjvms(this);
4659 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4660 Node* slow_result = set_results_for_java_call(slow_call);
4661 // this->control() comes from set_results_for_java_call
4662 result_reg->init_req(_slow_path, control());
4663 result_val->init_req(_slow_path, slow_result);
4664 result_i_o ->set_req(_slow_path, i_o());
4665 result_mem ->set_req(_slow_path, reset_memory());
4666 }
4667
4668 // Return the combined state.
4669 set_control( _gvn.transform(result_reg));
4670 set_i_o( _gvn.transform(result_i_o));
4671 set_all_memory( _gvn.transform(result_mem));
4672 } // original reexecute is set back here
4673
4674 set_result(_gvn.transform(result_val));
4675 return true;
4676 }
4677
4678 // If we have a tighly coupled allocation, the arraycopy may take care
4679 // of the array initialization. If one of the guards we insert between
4680 // the allocation and the arraycopy causes a deoptimization, an
4681 // unitialized array will escape the compiled method. To prevent that
4682 // we set the JVM state for uncommon traps between the allocation and
4683 // the arraycopy to the state before the allocation so, in case of
4684 // deoptimization, we'll reexecute the allocation and the
4685 // initialization.
4686 JVMState* LibraryCallKit::arraycopy_restore_alloc_state(AllocateArrayNode* alloc, int& saved_reexecute_sp) {
4687 if (alloc != NULL) {
4688 ciMethod* trap_method = alloc->jvms()->method();
4689 int trap_bci = alloc->jvms()->bci();
4690
4691 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &
4692 !C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_null_check)) {
4693 // Make sure there's no store between the allocation and the
4694 // arraycopy otherwise visible side effects could be rexecuted
4695 // in case of deoptimization and cause incorrect execution.
4696 bool no_interfering_store = true;
4697 Node* mem = alloc->in(TypeFunc::Memory);
4698 if (mem->is_MergeMem()) {
4699 for (MergeMemStream mms(merged_memory(), mem->as_MergeMem()); mms.next_non_empty2(); ) {
4700 Node* n = mms.memory();
4701 if (n != mms.memory2() && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4702 assert(n->is_Store(), "what else?");
4703 no_interfering_store = false;
4704 break;
4705 }
4706 }
4707 } else {
4708 for (MergeMemStream mms(merged_memory()); mms.next_non_empty(); ) {
4709 Node* n = mms.memory();
4710 if (n != mem && !(n->is_Proj() && n->in(0) == alloc->initialization())) {
4711 assert(n->is_Store(), "what else?");
4712 no_interfering_store = false;
4713 break;
4714 }
4715 }
4716 }
4717
4718 if (no_interfering_store) {
4719 JVMState* old_jvms = alloc->jvms()->clone_shallow(C);
4720 uint size = alloc->req();
4721 SafePointNode* sfpt = new SafePointNode(size, old_jvms);
4722 old_jvms->set_map(sfpt);
4723 for (uint i = 0; i < size; i++) {
4724 sfpt->init_req(i, alloc->in(i));
4725 }
4726 // re-push array length for deoptimization
4727 sfpt->ins_req(old_jvms->stkoff() + old_jvms->sp(), alloc->in(AllocateNode::ALength));
4728 old_jvms->set_sp(old_jvms->sp()+1);
4729 old_jvms->set_monoff(old_jvms->monoff()+1);
4730 old_jvms->set_scloff(old_jvms->scloff()+1);
4731 old_jvms->set_endoff(old_jvms->endoff()+1);
4732 old_jvms->set_should_reexecute(true);
4733
4734 sfpt->set_i_o(map()->i_o());
4735 sfpt->set_memory(map()->memory());
4736 sfpt->set_control(map()->control());
4737
4738 JVMState* saved_jvms = jvms();
4739 saved_reexecute_sp = _reexecute_sp;
4740
4741 set_jvms(sfpt->jvms());
4742 _reexecute_sp = jvms()->sp();
4743
4744 return saved_jvms;
4745 }
4746 }
4747 }
4748 return NULL;
4749 }
4750
4751 // In case of a deoptimization, we restart execution at the
4752 // allocation, allocating a new array. We would leave an uninitialized
4753 // array in the heap that GCs wouldn't expect. Move the allocation
4754 // after the traps so we don't allocate the array if we
4755 // deoptimize. This is possible because tightly_coupled_allocation()
4756 // guarantees there's no observer of the allocated array at this point
4757 // and the control flow is simple enough.
4758 void LibraryCallKit::arraycopy_move_allocation_here(AllocateArrayNode* alloc, Node* dest, JVMState* saved_jvms, int saved_reexecute_sp) {
4759 if (saved_jvms != NULL && !stopped()) {
4760 assert(alloc != NULL, "only with a tightly coupled allocation");
4761 // restore JVM state to the state at the arraycopy
4762 saved_jvms->map()->set_control(map()->control());
4763 assert(saved_jvms->map()->memory() == map()->memory(), "memory state changed?");
4764 assert(saved_jvms->map()->i_o() == map()->i_o(), "IO state changed?");
4765 // If we've improved the types of some nodes (null check) while
4766 // emitting the guards, propagate them to the current state
4767 map()->replaced_nodes().apply(saved_jvms->map());
4768 set_jvms(saved_jvms);
4769 _reexecute_sp = saved_reexecute_sp;
4770
4771 // Remove the allocation from above the guards
4772 CallProjections callprojs;
4773 alloc->extract_projections(&callprojs, true);
4774 InitializeNode* init = alloc->initialization();
4775 Node* alloc_mem = alloc->in(TypeFunc::Memory);
4776 C->gvn_replace_by(callprojs.fallthrough_ioproj, alloc->in(TypeFunc::I_O));
4777 C->gvn_replace_by(init->proj_out(TypeFunc::Memory), alloc_mem);
4778 C->gvn_replace_by(init->proj_out(TypeFunc::Control), alloc->in(0));
4779
4780 // move the allocation here (after the guards)
4781 _gvn.hash_delete(alloc);
4782 alloc->set_req(TypeFunc::Control, control());
4783 alloc->set_req(TypeFunc::I_O, i_o());
4784 Node *mem = reset_memory();
4785 set_all_memory(mem);
4786 alloc->set_req(TypeFunc::Memory, mem);
4787 set_control(init->proj_out(TypeFunc::Control));
4788 set_i_o(callprojs.fallthrough_ioproj);
4789
4790 // Update memory as done in GraphKit::set_output_for_allocation()
4791 const TypeInt* length_type = _gvn.find_int_type(alloc->in(AllocateNode::ALength));
4792 const TypeOopPtr* ary_type = _gvn.type(alloc->in(AllocateNode::KlassNode))->is_klassptr()->as_instance_type();
4793 if (ary_type->isa_aryptr() && length_type != NULL) {
4794 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4795 }
4796 const TypePtr* telemref = ary_type->add_offset(Type::OffsetBot);
4797 int elemidx = C->get_alias_index(telemref);
4798 set_memory(init->proj_out(TypeFunc::Memory), Compile::AliasIdxRaw);
4799 set_memory(init->proj_out(TypeFunc::Memory), elemidx);
4800
4801 Node* allocx = _gvn.transform(alloc);
4802 assert(allocx == alloc, "where has the allocation gone?");
4803 assert(dest->is_CheckCastPP(), "not an allocation result?");
4804
4805 _gvn.hash_delete(dest);
4806 dest->set_req(0, control());
4807 Node* destx = _gvn.transform(dest);
4808 assert(destx == dest, "where has the allocation result gone?");
4809 }
4810 }
4811
4812
4813 //------------------------------inline_arraycopy-----------------------
4814 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4815 // Object dest, int destPos,
4816 // int length);
4817 bool LibraryCallKit::inline_arraycopy() {
4818 // Get the arguments.
4819 Node* src = argument(0); // type: oop
4820 Node* src_offset = argument(1); // type: int
4821 Node* dest = argument(2); // type: oop
4822 Node* dest_offset = argument(3); // type: int
4823 Node* length = argument(4); // type: int
4824
4825 // Check for allocation before we add nodes that would confuse
4826 // tightly_coupled_allocation()
4827 AllocateArrayNode* alloc = tightly_coupled_allocation(dest, NULL);
4828
4829 int saved_reexecute_sp = -1;
4830 JVMState* saved_jvms = arraycopy_restore_alloc_state(alloc, saved_reexecute_sp);
4831 // See arraycopy_restore_alloc_state() comment
4832 // if alloc == NULL we don't have to worry about a tightly coupled allocation so we can emit all needed guards
4833 // if saved_jvms != NULL (then alloc != NULL) then we can handle guards and a tightly coupled allocation
4834 // if saved_jvms == NULL and alloc != NULL, we can’t emit any guards
4835 bool can_emit_guards = (alloc == NULL || saved_jvms != NULL);
4836
4837 // The following tests must be performed
4838 // (1) src and dest are arrays.
4839 // (2) src and dest arrays must have elements of the same BasicType
4840 // (3) src and dest must not be null.
4841 // (4) src_offset must not be negative.
4842 // (5) dest_offset must not be negative.
4843 // (6) length must not be negative.
4844 // (7) src_offset + length must not exceed length of src.
4845 // (8) dest_offset + length must not exceed length of dest.
4846 // (9) each element of an oop array must be assignable
4847
4848 // (3) src and dest must not be null.
4849 // always do this here because we need the JVM state for uncommon traps
4850 Node* null_ctl = top();
4851 src = saved_jvms != NULL ? null_check_oop(src, &null_ctl, true, true) : null_check(src, T_ARRAY);
4852 assert(null_ctl->is_top(), "no null control here");
4853 dest = null_check(dest, T_ARRAY);
4854
4855 if (!can_emit_guards) {
4856 // if saved_jvms == NULL and alloc != NULL, we don't emit any
4857 // guards but the arraycopy node could still take advantage of a
4858 // tightly allocated allocation. tightly_coupled_allocation() is
4859 // called again to make sure it takes the null check above into
4860 // account: the null check is mandatory and if it caused an
4861 // uncommon trap to be emitted then the allocation can't be
4862 // considered tightly coupled in this context.
4863 alloc = tightly_coupled_allocation(dest, NULL);
4864 }
4865
4866 bool validated = false;
4867
4868 const Type* src_type = _gvn.type(src);
4869 const Type* dest_type = _gvn.type(dest);
4870 const TypeAryPtr* top_src = src_type->isa_aryptr();
4871 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4872
4873 // Do we have the type of src?
4874 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4875 // Do we have the type of dest?
4876 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4877 // Is the type for src from speculation?
4878 bool src_spec = false;
4879 // Is the type for dest from speculation?
4880 bool dest_spec = false;
4881
4882 if ((!has_src || !has_dest) && can_emit_guards) {
4883 // We don't have sufficient type information, let's see if
4884 // speculative types can help. We need to have types for both src
4885 // and dest so that it pays off.
4886
4887 // Do we already have or could we have type information for src
4888 bool could_have_src = has_src;
4889 // Do we already have or could we have type information for dest
4890 bool could_have_dest = has_dest;
4891
4892 ciKlass* src_k = NULL;
4893 if (!has_src) {
4894 src_k = src_type->speculative_type_not_null();
4895 if (src_k != NULL && src_k->is_array_klass()) {
4896 could_have_src = true;
4897 }
4898 }
4899
4900 ciKlass* dest_k = NULL;
4901 if (!has_dest) {
4902 dest_k = dest_type->speculative_type_not_null();
4903 if (dest_k != NULL && dest_k->is_array_klass()) {
4904 could_have_dest = true;
4905 }
4906 }
4907
4908 if (could_have_src && could_have_dest) {
4909 // This is going to pay off so emit the required guards
4910 if (!has_src) {
4911 src = maybe_cast_profiled_obj(src, src_k, true);
4912 src_type = _gvn.type(src);
4913 top_src = src_type->isa_aryptr();
4914 has_src = (top_src != NULL && top_src->klass() != NULL);
4915 src_spec = true;
4916 }
4917 if (!has_dest) {
4918 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4919 dest_type = _gvn.type(dest);
4920 top_dest = dest_type->isa_aryptr();
4921 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4922 dest_spec = true;
4923 }
4924 }
4925 }
4926
4927 if (has_src && has_dest && can_emit_guards) {
4928 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4929 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4930 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4931 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4932
4933 if (src_elem == dest_elem && src_elem == T_OBJECT) {
4934 // If both arrays are object arrays then having the exact types
4935 // for both will remove the need for a subtype check at runtime
4936 // before the call and may make it possible to pick a faster copy
4937 // routine (without a subtype check on every element)
4938 // Do we have the exact type of src?
4939 bool could_have_src = src_spec;
4940 // Do we have the exact type of dest?
4941 bool could_have_dest = dest_spec;
4942 ciKlass* src_k = top_src->klass();
4943 ciKlass* dest_k = top_dest->klass();
4944 if (!src_spec) {
4945 src_k = src_type->speculative_type_not_null();
4946 if (src_k != NULL && src_k->is_array_klass()) {
4947 could_have_src = true;
4948 }
4949 }
4950 if (!dest_spec) {
4951 dest_k = dest_type->speculative_type_not_null();
4952 if (dest_k != NULL && dest_k->is_array_klass()) {
4953 could_have_dest = true;
4954 }
4955 }
4956 if (could_have_src && could_have_dest) {
4957 // If we can have both exact types, emit the missing guards
4958 if (could_have_src && !src_spec) {
4959 src = maybe_cast_profiled_obj(src, src_k, true);
4960 }
4961 if (could_have_dest && !dest_spec) {
4962 dest = maybe_cast_profiled_obj(dest, dest_k, true);
4963 }
4964 }
4965 }
4966 }
4967
4968 ciMethod* trap_method = method();
4969 int trap_bci = bci();
4970 if (saved_jvms != NULL) {
4971 trap_method = alloc->jvms()->method();
4972 trap_bci = alloc->jvms()->bci();
4973 }
4974
4975 if (!C->too_many_traps(trap_method, trap_bci, Deoptimization::Reason_intrinsic) &&
4976 can_emit_guards &&
4977 !src->is_top() && !dest->is_top()) {
4978 // validate arguments: enables transformation the ArrayCopyNode
4979 validated = true;
4980
4981 RegionNode* slow_region = new RegionNode(1);
4982 record_for_igvn(slow_region);
4983
4984 // (1) src and dest are arrays.
4985 generate_non_array_guard(load_object_klass(src), slow_region);
4986 generate_non_array_guard(load_object_klass(dest), slow_region);
4987
4988 // (2) src and dest arrays must have elements of the same BasicType
4989 // done at macro expansion or at Ideal transformation time
4990
4991 // (4) src_offset must not be negative.
4992 generate_negative_guard(src_offset, slow_region);
4993
4994 // (5) dest_offset must not be negative.
4995 generate_negative_guard(dest_offset, slow_region);
4996
4997 // (7) src_offset + length must not exceed length of src.
4998 generate_limit_guard(src_offset, length,
4999 load_array_length(src),
5000 slow_region);
5001
5002 // (8) dest_offset + length must not exceed length of dest.
5003 generate_limit_guard(dest_offset, length,
5004 load_array_length(dest),
5005 slow_region);
5006
5007 // (9) each element of an oop array must be assignable
5008 Node* src_klass = load_object_klass(src);
5009 Node* dest_klass = load_object_klass(dest);
5010 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5011
5012 if (not_subtype_ctrl != top()) {
5013 PreserveJVMState pjvms(this);
5014 set_control(not_subtype_ctrl);
5015 uncommon_trap(Deoptimization::Reason_intrinsic,
5016 Deoptimization::Action_make_not_entrant);
5017 assert(stopped(), "Should be stopped");
5018 }
5019 {
5020 PreserveJVMState pjvms(this);
5021 set_control(_gvn.transform(slow_region));
5022 uncommon_trap(Deoptimization::Reason_intrinsic,
5023 Deoptimization::Action_make_not_entrant);
5024 assert(stopped(), "Should be stopped");
5025 }
5026 }
5027
5028 arraycopy_move_allocation_here(alloc, dest, saved_jvms, saved_reexecute_sp);
5029
5030 if (stopped()) {
5031 return true;
5032 }
5033
5034 src = shenandoah_read_barrier(src);
5035 dest = shenandoah_write_barrier(dest);
5036
5037 ArrayCopyNode* ac = ArrayCopyNode::make(this, true, src, src_offset, dest, dest_offset, length, alloc != NULL,
5038 // Create LoadRange and LoadKlass nodes for use during macro expansion here
5039 // so the compiler has a chance to eliminate them: during macro expansion,
5040 // we have to set their control (CastPP nodes are eliminated).
5041 load_object_klass(src), load_object_klass(dest),
5042 load_array_length(src), load_array_length(dest));
5043
5044 ac->set_arraycopy(validated);
5045
5046 Node* n = _gvn.transform(ac);
5047 if (n == ac) {
5048 ac->connect_outputs(this);
5049 } else {
5050 assert(validated, "shouldn't transform if all arguments not validated");
5051 set_all_memory(n);
5052 }
5053
5054 return true;
5055 }
5056
5057
5058 // Helper function which determines if an arraycopy immediately follows
5059 // an allocation, with no intervening tests or other escapes for the object.
5060 AllocateArrayNode*
5061 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5062 RegionNode* slow_region) {
5063 if (stopped()) return NULL; // no fast path
5064 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5065
5066 ptr = ShenandoahBarrierNode::skip_through_barrier(ptr);
5067
5068 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5069 if (alloc == NULL) return NULL;
5070
5071 Node* rawmem = memory(Compile::AliasIdxRaw);
5072 // Is the allocation's memory state untouched?
5073 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5074 // Bail out if there have been raw-memory effects since the allocation.
5075 // (Example: There might have been a call or safepoint.)
5076 return NULL;
5077 }
5078 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5079 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5080 return NULL;
5081 }
5082
5083 // There must be no unexpected observers of this allocation.
5084 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5085 Node* obs = ptr->fast_out(i);
5086 if (obs != this->map()) {
5087 return NULL;
5088 }
5089 }
5090
5091 // This arraycopy must unconditionally follow the allocation of the ptr.
5092 Node* alloc_ctl = ptr->in(0);
5093 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5094
5095 Node* ctl = control();
5096 while (ctl != alloc_ctl) {
5097 // There may be guards which feed into the slow_region.
5098 // Any other control flow means that we might not get a chance
5099 // to finish initializing the allocated object.
5100 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5101 IfNode* iff = ctl->in(0)->as_If();
5102 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5103 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5104 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5105 ctl = iff->in(0); // This test feeds the known slow_region.
5106 continue;
5107 }
5108 // One more try: Various low-level checks bottom out in
5109 // uncommon traps. If the debug-info of the trap omits
5110 // any reference to the allocation, as we've already
5111 // observed, then there can be no objection to the trap.
5112 bool found_trap = false;
5113 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5114 Node* obs = not_ctl->fast_out(j);
5115 if (obs->in(0) == not_ctl && obs->is_Call() &&
5116 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5117 found_trap = true; break;
5118 }
5119 }
5120 if (found_trap) {
5121 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5122 continue;
5123 }
5124 }
5125 return NULL;
5126 }
5127
5128 // If we get this far, we have an allocation which immediately
5129 // precedes the arraycopy, and we can take over zeroing the new object.
5130 // The arraycopy will finish the initialization, and provide
5131 // a new control state to which we will anchor the destination pointer.
5132
5133 return alloc;
5134 }
5135
5136 //-------------inline_encodeISOArray-----------------------------------
5137 // encode char[] to byte[] in ISO_8859_1
5138 bool LibraryCallKit::inline_encodeISOArray() {
5139 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5140 // no receiver since it is static method
5141 Node *src = argument(0);
5142 Node *src_offset = argument(1);
5143 Node *dst = argument(2);
5144 Node *dst_offset = argument(3);
5145 Node *length = argument(4);
5146
5147 src = shenandoah_read_barrier(src);
5148 dst = shenandoah_write_barrier(dst);
5149
5150 const Type* src_type = src->Value(&_gvn);
5151 const Type* dst_type = dst->Value(&_gvn);
5152 const TypeAryPtr* top_src = src_type->isa_aryptr();
5153 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5154 if (top_src == NULL || top_src->klass() == NULL ||
5155 top_dest == NULL || top_dest->klass() == NULL) {
5156 // failed array check
5157 return false;
5158 }
5159
5160 // Figure out the size and type of the elements we will be copying.
5161 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5162 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5163 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5164 return false;
5165 }
5166 Node* src_start = array_element_address(src, src_offset, src_elem);
5167 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5168 // 'src_start' points to src array + scaled offset
5169 // 'dst_start' points to dst array + scaled offset
5170
5171 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5172 Node* enc = new EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5173 enc = _gvn.transform(enc);
5174 Node* res_mem = _gvn.transform(new SCMemProjNode(enc));
5175 set_memory(res_mem, mtype);
5176 set_result(enc);
5177 return true;
5178 }
5179
5180 //-------------inline_multiplyToLen-----------------------------------
5181 bool LibraryCallKit::inline_multiplyToLen() {
5182 assert(UseMultiplyToLenIntrinsic, "not implemented on this platform");
5183
5184 address stubAddr = StubRoutines::multiplyToLen();
5185 if (stubAddr == NULL) {
5186 return false; // Intrinsic's stub is not implemented on this platform
5187 }
5188 const char* stubName = "multiplyToLen";
5189
5190 assert(callee()->signature()->size() == 5, "multiplyToLen has 5 parameters");
5191
5192 // no receiver because it is a static method
5193 Node* x = argument(0);
5194 Node* xlen = argument(1);
5195 Node* y = argument(2);
5196 Node* ylen = argument(3);
5197 Node* z = argument(4);
5198
5199 x = shenandoah_read_barrier(x);
5200 y = shenandoah_read_barrier(y);
5201 z = shenandoah_write_barrier(z);
5202
5203 const Type* x_type = x->Value(&_gvn);
5204 const Type* y_type = y->Value(&_gvn);
5205 const TypeAryPtr* top_x = x_type->isa_aryptr();
5206 const TypeAryPtr* top_y = y_type->isa_aryptr();
5207 if (top_x == NULL || top_x->klass() == NULL ||
5208 top_y == NULL || top_y->klass() == NULL) {
5209 // failed array check
5210 return false;
5211 }
5212
5213 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5214 BasicType y_elem = y_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5215 if (x_elem != T_INT || y_elem != T_INT) {
5216 return false;
5217 }
5218
5219 // Set the original stack and the reexecute bit for the interpreter to reexecute
5220 // the bytecode that invokes BigInteger.multiplyToLen() if deoptimization happens
5221 // on the return from z array allocation in runtime.
5222 { PreserveReexecuteState preexecs(this);
5223 jvms()->set_should_reexecute(true);
5224
5225 Node* x_start = array_element_address(x, intcon(0), x_elem);
5226 Node* y_start = array_element_address(y, intcon(0), y_elem);
5227 // 'x_start' points to x array + scaled xlen
5228 // 'y_start' points to y array + scaled ylen
5229
5230 // Allocate the result array
5231 Node* zlen = _gvn.transform(new AddINode(xlen, ylen));
5232 ciKlass* klass = ciTypeArrayKlass::make(T_INT);
5233 Node* klass_node = makecon(TypeKlassPtr::make(klass));
5234
5235 IdealKit ideal(this);
5236
5237 #define __ ideal.
5238 Node* one = __ ConI(1);
5239 Node* zero = __ ConI(0);
5240 IdealVariable need_alloc(ideal), z_alloc(ideal); __ declarations_done();
5241 __ set(need_alloc, zero);
5242 __ set(z_alloc, z);
5243 __ if_then(z, BoolTest::eq, null()); {
5244 __ increment (need_alloc, one);
5245 } __ else_(); {
5246 // Update graphKit memory and control from IdealKit.
5247 sync_kit(ideal);
5248 Node* zlen_arg = load_array_length(z);
5249 // Update IdealKit memory and control from graphKit.
5250 __ sync_kit(this);
5251 __ if_then(zlen_arg, BoolTest::lt, zlen); {
5252 __ increment (need_alloc, one);
5253 } __ end_if();
5254 } __ end_if();
5255
5256 __ if_then(__ value(need_alloc), BoolTest::ne, zero); {
5257 // Update graphKit memory and control from IdealKit.
5258 sync_kit(ideal);
5259 Node * narr = new_array(klass_node, zlen, 1);
5260 // Update IdealKit memory and control from graphKit.
5261 __ sync_kit(this);
5262 __ set(z_alloc, narr);
5263 } __ end_if();
5264
5265 sync_kit(ideal);
5266 z = __ value(z_alloc);
5267 // Can't use TypeAryPtr::INTS which uses Bottom offset.
5268 _gvn.set_type(z, TypeOopPtr::make_from_klass(klass));
5269 // Final sync IdealKit and GraphKit.
5270 final_sync(ideal);
5271 #undef __
5272
5273 Node* z_start = array_element_address(z, intcon(0), T_INT);
5274
5275 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5276 OptoRuntime::multiplyToLen_Type(),
5277 stubAddr, stubName, TypePtr::BOTTOM,
5278 x_start, xlen, y_start, ylen, z_start, zlen);
5279 } // original reexecute is set back here
5280
5281 C->set_has_split_ifs(true); // Has chance for split-if optimization
5282 set_result(z);
5283 return true;
5284 }
5285
5286 //-------------inline_squareToLen------------------------------------
5287 bool LibraryCallKit::inline_squareToLen() {
5288 assert(UseSquareToLenIntrinsic, "not implementated on this platform");
5289
5290 address stubAddr = StubRoutines::squareToLen();
5291 if (stubAddr == NULL) {
5292 return false; // Intrinsic's stub is not implemented on this platform
5293 }
5294 const char* stubName = "squareToLen";
5295
5296 assert(callee()->signature()->size() == 4, "implSquareToLen has 4 parameters");
5297
5298 Node* x = argument(0);
5299 Node* len = argument(1);
5300 Node* z = argument(2);
5301 Node* zlen = argument(3);
5302
5303 x = shenandoah_read_barrier(x);
5304 z = shenandoah_write_barrier(z);
5305
5306 const Type* x_type = x->Value(&_gvn);
5307 const Type* z_type = z->Value(&_gvn);
5308 const TypeAryPtr* top_x = x_type->isa_aryptr();
5309 const TypeAryPtr* top_z = z_type->isa_aryptr();
5310 if (top_x == NULL || top_x->klass() == NULL ||
5311 top_z == NULL || top_z->klass() == NULL) {
5312 // failed array check
5313 return false;
5314 }
5315
5316 BasicType x_elem = x_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5317 BasicType z_elem = z_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5318 if (x_elem != T_INT || z_elem != T_INT) {
5319 return false;
5320 }
5321
5322
5323 Node* x_start = array_element_address(x, intcon(0), x_elem);
5324 Node* z_start = array_element_address(z, intcon(0), z_elem);
5325
5326 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5327 OptoRuntime::squareToLen_Type(),
5328 stubAddr, stubName, TypePtr::BOTTOM,
5329 x_start, len, z_start, zlen);
5330
5331 set_result(z);
5332 return true;
5333 }
5334
5335 //-------------inline_mulAdd------------------------------------------
5336 bool LibraryCallKit::inline_mulAdd() {
5337 assert(UseMulAddIntrinsic, "not implementated on this platform");
5338
5339 address stubAddr = StubRoutines::mulAdd();
5340 if (stubAddr == NULL) {
5341 return false; // Intrinsic's stub is not implemented on this platform
5342 }
5343 const char* stubName = "mulAdd";
5344
5345 assert(callee()->signature()->size() == 5, "mulAdd has 5 parameters");
5346
5347 Node* out = argument(0);
5348 Node* in = argument(1);
5349 Node* offset = argument(2);
5350 Node* len = argument(3);
5351 Node* k = argument(4);
5352
5353 in = shenandoah_read_barrier(in);
5354 out = shenandoah_write_barrier(out);
5355
5356 const Type* out_type = out->Value(&_gvn);
5357 const Type* in_type = in->Value(&_gvn);
5358 const TypeAryPtr* top_out = out_type->isa_aryptr();
5359 const TypeAryPtr* top_in = in_type->isa_aryptr();
5360 if (top_out == NULL || top_out->klass() == NULL ||
5361 top_in == NULL || top_in->klass() == NULL) {
5362 // failed array check
5363 return false;
5364 }
5365
5366 BasicType out_elem = out_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5367 BasicType in_elem = in_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5368 if (out_elem != T_INT || in_elem != T_INT) {
5369 return false;
5370 }
5371
5372 Node* outlen = load_array_length(out);
5373 Node* new_offset = _gvn.transform(new SubINode(outlen, offset));
5374 Node* out_start = array_element_address(out, intcon(0), out_elem);
5375 Node* in_start = array_element_address(in, intcon(0), in_elem);
5376
5377 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5378 OptoRuntime::mulAdd_Type(),
5379 stubAddr, stubName, TypePtr::BOTTOM,
5380 out_start,in_start, new_offset, len, k);
5381 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5382 set_result(result);
5383 return true;
5384 }
5385
5386 //-------------inline_montgomeryMultiply-----------------------------------
5387 bool LibraryCallKit::inline_montgomeryMultiply() {
5388 address stubAddr = StubRoutines::montgomeryMultiply();
5389 if (stubAddr == NULL) {
5390 return false; // Intrinsic's stub is not implemented on this platform
5391 }
5392
5393 assert(UseMontgomeryMultiplyIntrinsic, "not implemented on this platform");
5394 const char* stubName = "montgomery_square";
5395
5396 assert(callee()->signature()->size() == 7, "montgomeryMultiply has 7 parameters");
5397
5398 Node* a = argument(0);
5399 Node* b = argument(1);
5400 Node* n = argument(2);
5401 Node* len = argument(3);
5402 Node* inv = argument(4);
5403 Node* m = argument(6);
5404
5405 a = shenandoah_read_barrier(a);
5406 b = shenandoah_read_barrier(b);
5407 n = shenandoah_read_barrier(n);
5408 m = shenandoah_write_barrier(m);
5409
5410 const Type* a_type = a->Value(&_gvn);
5411 const TypeAryPtr* top_a = a_type->isa_aryptr();
5412 const Type* b_type = b->Value(&_gvn);
5413 const TypeAryPtr* top_b = b_type->isa_aryptr();
5414 const Type* n_type = a->Value(&_gvn);
5415 const TypeAryPtr* top_n = n_type->isa_aryptr();
5416 const Type* m_type = a->Value(&_gvn);
5417 const TypeAryPtr* top_m = m_type->isa_aryptr();
5418 if (top_a == NULL || top_a->klass() == NULL ||
5419 top_b == NULL || top_b->klass() == NULL ||
5420 top_n == NULL || top_n->klass() == NULL ||
5421 top_m == NULL || top_m->klass() == NULL) {
5422 // failed array check
5423 return false;
5424 }
5425
5426 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5427 BasicType b_elem = b_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5428 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5429 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5430 if (a_elem != T_INT || b_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5431 return false;
5432 }
5433
5434 // Make the call
5435 {
5436 Node* a_start = array_element_address(a, intcon(0), a_elem);
5437 Node* b_start = array_element_address(b, intcon(0), b_elem);
5438 Node* n_start = array_element_address(n, intcon(0), n_elem);
5439 Node* m_start = array_element_address(m, intcon(0), m_elem);
5440
5441 Node* call = make_runtime_call(RC_LEAF,
5442 OptoRuntime::montgomeryMultiply_Type(),
5443 stubAddr, stubName, TypePtr::BOTTOM,
5444 a_start, b_start, n_start, len, inv, top(),
5445 m_start);
5446 set_result(m);
5447 }
5448
5449 return true;
5450 }
5451
5452 bool LibraryCallKit::inline_montgomerySquare() {
5453 address stubAddr = StubRoutines::montgomerySquare();
5454 if (stubAddr == NULL) {
5455 return false; // Intrinsic's stub is not implemented on this platform
5456 }
5457
5458 assert(UseMontgomerySquareIntrinsic, "not implemented on this platform");
5459 const char* stubName = "montgomery_square";
5460
5461 assert(callee()->signature()->size() == 6, "montgomerySquare has 6 parameters");
5462
5463 Node* a = argument(0);
5464 Node* n = argument(1);
5465 Node* len = argument(2);
5466 Node* inv = argument(3);
5467 Node* m = argument(5);
5468
5469 a = shenandoah_read_barrier(a);
5470 n = shenandoah_read_barrier(n);
5471 m = shenandoah_write_barrier(m);
5472
5473 const Type* a_type = a->Value(&_gvn);
5474 const TypeAryPtr* top_a = a_type->isa_aryptr();
5475 const Type* n_type = a->Value(&_gvn);
5476 const TypeAryPtr* top_n = n_type->isa_aryptr();
5477 const Type* m_type = a->Value(&_gvn);
5478 const TypeAryPtr* top_m = m_type->isa_aryptr();
5479 if (top_a == NULL || top_a->klass() == NULL ||
5480 top_n == NULL || top_n->klass() == NULL ||
5481 top_m == NULL || top_m->klass() == NULL) {
5482 // failed array check
5483 return false;
5484 }
5485
5486 BasicType a_elem = a_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5487 BasicType n_elem = n_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5488 BasicType m_elem = m_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5489 if (a_elem != T_INT || n_elem != T_INT || m_elem != T_INT) {
5490 return false;
5491 }
5492
5493 // Make the call
5494 {
5495 Node* a_start = array_element_address(a, intcon(0), a_elem);
5496 Node* n_start = array_element_address(n, intcon(0), n_elem);
5497 Node* m_start = array_element_address(m, intcon(0), m_elem);
5498
5499 Node* call = make_runtime_call(RC_LEAF,
5500 OptoRuntime::montgomerySquare_Type(),
5501 stubAddr, stubName, TypePtr::BOTTOM,
5502 a_start, n_start, len, inv, top(),
5503 m_start);
5504 set_result(m);
5505 }
5506
5507 return true;
5508 }
5509
5510
5511 /**
5512 * Calculate CRC32 for byte.
5513 * int java.util.zip.CRC32.update(int crc, int b)
5514 */
5515 bool LibraryCallKit::inline_updateCRC32() {
5516 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5517 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5518 // no receiver since it is static method
5519 Node* crc = argument(0); // type: int
5520 Node* b = argument(1); // type: int
5521
5522 /*
5523 * int c = ~ crc;
5524 * b = timesXtoThe32[(b ^ c) & 0xFF];
5525 * b = b ^ (c >>> 8);
5526 * crc = ~b;
5527 */
5528
5529 Node* M1 = intcon(-1);
5530 crc = _gvn.transform(new XorINode(crc, M1));
5531 Node* result = _gvn.transform(new XorINode(crc, b));
5532 result = _gvn.transform(new AndINode(result, intcon(0xFF)));
5533
5534 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5535 Node* offset = _gvn.transform(new LShiftINode(result, intcon(0x2)));
5536 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5537 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5538
5539 crc = _gvn.transform(new URShiftINode(crc, intcon(8)));
5540 result = _gvn.transform(new XorINode(crc, result));
5541 result = _gvn.transform(new XorINode(result, M1));
5542 set_result(result);
5543 return true;
5544 }
5545
5546 /**
5547 * Calculate CRC32 for byte[] array.
5548 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5549 */
5550 bool LibraryCallKit::inline_updateBytesCRC32() {
5551 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5552 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5553 // no receiver since it is static method
5554 Node* crc = argument(0); // type: int
5555 Node* src = argument(1); // type: oop
5556 Node* offset = argument(2); // type: int
5557 Node* length = argument(3); // type: int
5558
5559 src = shenandoah_read_barrier(src);
5560
5561 const Type* src_type = src->Value(&_gvn);
5562 const TypeAryPtr* top_src = src_type->isa_aryptr();
5563 if (top_src == NULL || top_src->klass() == NULL) {
5564 // failed array check
5565 return false;
5566 }
5567
5568 // Figure out the size and type of the elements we will be copying.
5569 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5570 if (src_elem != T_BYTE) {
5571 return false;
5572 }
5573
5574 // 'src_start' points to src array + scaled offset
5575 Node* src_start = array_element_address(src, offset, src_elem);
5576
5577 // We assume that range check is done by caller.
5578 // TODO: generate range check (offset+length < src.length) in debug VM.
5579
5580 // Call the stub.
5581 address stubAddr = StubRoutines::updateBytesCRC32();
5582 const char *stubName = "updateBytesCRC32";
5583
5584 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5585 stubAddr, stubName, TypePtr::BOTTOM,
5586 crc, src_start, length);
5587 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5588 set_result(result);
5589 return true;
5590 }
5591
5592 /**
5593 * Calculate CRC32 for ByteBuffer.
5594 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5595 */
5596 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5597 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5598 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5599 // no receiver since it is static method
5600 Node* crc = argument(0); // type: int
5601 Node* src = argument(1); // type: long
5602 Node* offset = argument(3); // type: int
5603 Node* length = argument(4); // type: int
5604
5605 src = ConvL2X(src); // adjust Java long to machine word
5606 Node* base = _gvn.transform(new CastX2PNode(src));
5607 offset = ConvI2X(offset);
5608
5609 // 'src_start' points to src array + scaled offset
5610 Node* src_start = basic_plus_adr(top(), base, offset);
5611
5612 // Call the stub.
5613 address stubAddr = StubRoutines::updateBytesCRC32();
5614 const char *stubName = "updateBytesCRC32";
5615
5616 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5617 stubAddr, stubName, TypePtr::BOTTOM,
5618 crc, src_start, length);
5619 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5620 set_result(result);
5621 return true;
5622 }
5623
5624 //------------------------------get_table_from_crc32c_class-----------------------
5625 Node * LibraryCallKit::get_table_from_crc32c_class(ciInstanceKlass *crc32c_class) {
5626 Node* table = load_field_from_object(NULL, "byteTable", "[I", /*is_exact*/ false, /*is_static*/ true, crc32c_class);
5627 assert (table != NULL, "wrong version of java.util.zip.CRC32C");
5628
5629 return table;
5630 }
5631
5632 //------------------------------inline_updateBytesCRC32C-----------------------
5633 //
5634 // Calculate CRC32C for byte[] array.
5635 // int java.util.zip.CRC32C.updateBytes(int crc, byte[] buf, int off, int end)
5636 //
5637 bool LibraryCallKit::inline_updateBytesCRC32C() {
5638 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5639 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5640 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5641 // no receiver since it is a static method
5642 Node* crc = argument(0); // type: int
5643 Node* src = argument(1); // type: oop
5644 Node* offset = argument(2); // type: int
5645 Node* end = argument(3); // type: int
5646
5647 Node* length = _gvn.transform(new SubINode(end, offset));
5648
5649 const Type* src_type = src->Value(&_gvn);
5650 const TypeAryPtr* top_src = src_type->isa_aryptr();
5651 if (top_src == NULL || top_src->klass() == NULL) {
5652 // failed array check
5653 return false;
5654 }
5655
5656 // Figure out the size and type of the elements we will be copying.
5657 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5658 if (src_elem != T_BYTE) {
5659 return false;
5660 }
5661
5662 // 'src_start' points to src array + scaled offset
5663 src = shenandoah_read_barrier(src);
5664 Node* src_start = array_element_address(src, offset, src_elem);
5665
5666 // static final int[] byteTable in class CRC32C
5667 Node* table = get_table_from_crc32c_class(callee()->holder());
5668 table = shenandoah_read_barrier(table);
5669 Node* table_start = array_element_address(table, intcon(0), T_INT);
5670
5671 // We assume that range check is done by caller.
5672 // TODO: generate range check (offset+length < src.length) in debug VM.
5673
5674 // Call the stub.
5675 address stubAddr = StubRoutines::updateBytesCRC32C();
5676 const char *stubName = "updateBytesCRC32C";
5677
5678 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5679 stubAddr, stubName, TypePtr::BOTTOM,
5680 crc, src_start, length, table_start);
5681 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5682 set_result(result);
5683 return true;
5684 }
5685
5686 //------------------------------inline_updateDirectByteBufferCRC32C-----------------------
5687 //
5688 // Calculate CRC32C for DirectByteBuffer.
5689 // int java.util.zip.CRC32C.updateDirectByteBuffer(int crc, long buf, int off, int end)
5690 //
5691 bool LibraryCallKit::inline_updateDirectByteBufferCRC32C() {
5692 assert(UseCRC32CIntrinsics, "need CRC32C instruction support");
5693 assert(callee()->signature()->size() == 5, "updateDirectByteBuffer has 4 parameters and one is long");
5694 assert(callee()->holder()->is_loaded(), "CRC32C class must be loaded");
5695 // no receiver since it is a static method
5696 Node* crc = argument(0); // type: int
5697 Node* src = argument(1); // type: long
5698 Node* offset = argument(3); // type: int
5699 Node* end = argument(4); // type: int
5700
5701 Node* length = _gvn.transform(new SubINode(end, offset));
5702
5703 src = ConvL2X(src); // adjust Java long to machine word
5704 Node* base = _gvn.transform(new CastX2PNode(src));
5705 offset = ConvI2X(offset);
5706
5707 // 'src_start' points to src array + scaled offset
5708 Node* src_start = basic_plus_adr(top(), base, offset);
5709
5710 // static final int[] byteTable in class CRC32C
5711 Node* table = get_table_from_crc32c_class(callee()->holder());
5712 table = shenandoah_read_barrier(table);
5713 Node* table_start = array_element_address(table, intcon(0), T_INT);
5714
5715 // Call the stub.
5716 address stubAddr = StubRoutines::updateBytesCRC32C();
5717 const char *stubName = "updateBytesCRC32C";
5718
5719 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesCRC32C_Type(),
5720 stubAddr, stubName, TypePtr::BOTTOM,
5721 crc, src_start, length, table_start);
5722 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5723 set_result(result);
5724 return true;
5725 }
5726
5727 //------------------------------inline_updateBytesAdler32----------------------
5728 //
5729 // Calculate Adler32 checksum for byte[] array.
5730 // int java.util.zip.Adler32.updateBytes(int crc, byte[] buf, int off, int len)
5731 //
5732 bool LibraryCallKit::inline_updateBytesAdler32() {
5733 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5734 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5735 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5736 // no receiver since it is static method
5737 Node* crc = argument(0); // type: int
5738 Node* src = argument(1); // type: oop
5739 Node* offset = argument(2); // type: int
5740 Node* length = argument(3); // type: int
5741
5742 const Type* src_type = src->Value(&_gvn);
5743 const TypeAryPtr* top_src = src_type->isa_aryptr();
5744 if (top_src == NULL || top_src->klass() == NULL) {
5745 // failed array check
5746 return false;
5747 }
5748
5749 // Figure out the size and type of the elements we will be copying.
5750 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5751 if (src_elem != T_BYTE) {
5752 return false;
5753 }
5754
5755 // 'src_start' points to src array + scaled offset
5756 src = shenandoah_read_barrier(src);
5757 Node* src_start = array_element_address(src, offset, src_elem);
5758
5759 // We assume that range check is done by caller.
5760 // TODO: generate range check (offset+length < src.length) in debug VM.
5761
5762 // Call the stub.
5763 address stubAddr = StubRoutines::updateBytesAdler32();
5764 const char *stubName = "updateBytesAdler32";
5765
5766 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5767 stubAddr, stubName, TypePtr::BOTTOM,
5768 crc, src_start, length);
5769 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5770 set_result(result);
5771 return true;
5772 }
5773
5774 //------------------------------inline_updateByteBufferAdler32---------------
5775 //
5776 // Calculate Adler32 checksum for DirectByteBuffer.
5777 // int java.util.zip.Adler32.updateByteBuffer(int crc, long buf, int off, int len)
5778 //
5779 bool LibraryCallKit::inline_updateByteBufferAdler32() {
5780 assert(UseAdler32Intrinsics, "Adler32 Instrinsic support need"); // check if we actually need to check this flag or check a different one
5781 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5782 assert(callee()->holder()->is_loaded(), "Adler32 class must be loaded");
5783 // no receiver since it is static method
5784 Node* crc = argument(0); // type: int
5785 Node* src = argument(1); // type: long
5786 Node* offset = argument(3); // type: int
5787 Node* length = argument(4); // type: int
5788
5789 src = ConvL2X(src); // adjust Java long to machine word
5790 Node* base = _gvn.transform(new CastX2PNode(src));
5791 offset = ConvI2X(offset);
5792
5793 // 'src_start' points to src array + scaled offset
5794 Node* src_start = basic_plus_adr(top(), base, offset);
5795
5796 // Call the stub.
5797 address stubAddr = StubRoutines::updateBytesAdler32();
5798 const char *stubName = "updateBytesAdler32";
5799
5800 Node* call = make_runtime_call(RC_LEAF, OptoRuntime::updateBytesAdler32_Type(),
5801 stubAddr, stubName, TypePtr::BOTTOM,
5802 crc, src_start, length);
5803
5804 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
5805 set_result(result);
5806 return true;
5807 }
5808
5809 //----------------------------inline_reference_get----------------------------
5810 // public T java.lang.ref.Reference.get();
5811 bool LibraryCallKit::inline_reference_get() {
5812 const int referent_offset = java_lang_ref_Reference::referent_offset;
5813 guarantee(referent_offset > 0, "should have already been set");
5814
5815 // Get the argument:
5816 Node* reference_obj = null_check_receiver();
5817 if (stopped()) return true;
5818
5819 if (ShenandoahVerifyReadsToFromSpace) {
5820 reference_obj = shenandoah_read_barrier(reference_obj);
5821 }
5822
5823 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5824
5825 ciInstanceKlass* klass = env()->Object_klass();
5826 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5827
5828 Node* no_ctrl = NULL;
5829 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5830
5831 // Use the pre-barrier to record the value in the referent field
5832 pre_barrier(false /* do_load */,
5833 control(),
5834 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5835 result /* pre_val */,
5836 T_OBJECT);
5837
5838 // Add memory barrier to prevent commoning reads from this field
5839 // across safepoint since GC can change its value.
5840 insert_mem_bar(Op_MemBarCPUOrder);
5841
5842 set_result(result);
5843 return true;
5844 }
5845
5846
5847 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5848 bool is_exact=true, bool is_static=false,
5849 ciInstanceKlass * fromKls=NULL) {
5850 if (fromKls == NULL) {
5851 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5852 assert(tinst != NULL, "obj is null");
5853 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5854 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5855 fromKls = tinst->klass()->as_instance_klass();
5856 } else {
5857 assert(is_static, "only for static field access");
5858 }
5859 ciField* field = fromKls->get_field_by_name(ciSymbol::make(fieldName),
5860 ciSymbol::make(fieldTypeString),
5861 is_static);
5862
5863 assert (field != NULL, "undefined field");
5864 if (field == NULL) return (Node *) NULL;
5865
5866 if (is_static) {
5867 const TypeInstPtr* tip = TypeInstPtr::make(fromKls->java_mirror());
5868 fromObj = makecon(tip);
5869 }
5870
5871 fromObj = shenandoah_read_barrier(fromObj);
5872
5873 // Next code copied from Parse::do_get_xxx():
5874
5875 // Compute address and memory type.
5876 int offset = field->offset_in_bytes();
5877 bool is_vol = field->is_volatile();
5878 ciType* field_klass = field->type();
5879 assert(field_klass->is_loaded(), "should be loaded");
5880 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5881 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5882 BasicType bt = field->layout_type();
5883
5884 // Build the resultant type of the load
5885 const Type *type;
5886 if (bt == T_OBJECT) {
5887 type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5888 } else {
5889 type = Type::get_const_basic_type(bt);
5890 }
5891
5892 if (support_IRIW_for_not_multiple_copy_atomic_cpu && is_vol) {
5893 insert_mem_bar(Op_MemBarVolatile); // StoreLoad barrier
5894 }
5895 // Build the load.
5896 MemNode::MemOrd mo = is_vol ? MemNode::acquire : MemNode::unordered;
5897 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, mo, LoadNode::DependsOnlyOnTest, is_vol);
5898 // If reference is volatile, prevent following memory ops from
5899 // floating up past the volatile read. Also prevents commoning
5900 // another volatile read.
5901 if (is_vol) {
5902 // Memory barrier includes bogus read of value to force load BEFORE membar
5903 insert_mem_bar(Op_MemBarAcquire, loadedField);
5904 }
5905 return loadedField;
5906 }
5907
5908
5909 //------------------------------inline_aescrypt_Block-----------------------
5910 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5911 address stubAddr;
5912 const char *stubName;
5913 assert(UseAES, "need AES instruction support");
5914
5915 switch(id) {
5916 case vmIntrinsics::_aescrypt_encryptBlock:
5917 stubAddr = StubRoutines::aescrypt_encryptBlock();
5918 stubName = "aescrypt_encryptBlock";
5919 break;
5920 case vmIntrinsics::_aescrypt_decryptBlock:
5921 stubAddr = StubRoutines::aescrypt_decryptBlock();
5922 stubName = "aescrypt_decryptBlock";
5923 break;
5924 }
5925 if (stubAddr == NULL) return false;
5926
5927 Node* aescrypt_object = argument(0);
5928 Node* src = argument(1);
5929 Node* src_offset = argument(2);
5930 Node* dest = argument(3);
5931 Node* dest_offset = argument(4);
5932
5933 // Resolve src and dest arrays for ShenandoahGC.
5934 src = shenandoah_read_barrier(src);
5935 dest = shenandoah_write_barrier(dest);
5936
5937 // (1) src and dest are arrays.
5938 const Type* src_type = src->Value(&_gvn);
5939 const Type* dest_type = dest->Value(&_gvn);
5940 const TypeAryPtr* top_src = src_type->isa_aryptr();
5941 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5942 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5943
5944 // for the quick and dirty code we will skip all the checks.
5945 // we are just trying to get the call to be generated.
5946 Node* src_start = src;
5947 Node* dest_start = dest;
5948 if (src_offset != NULL || dest_offset != NULL) {
5949 assert(src_offset != NULL && dest_offset != NULL, "");
5950 src_start = array_element_address(src, src_offset, T_BYTE);
5951 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5952 }
5953
5954 // now need to get the start of its expanded key array
5955 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5956 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5957 if (k_start == NULL) return false;
5958
5959 if (Matcher::pass_original_key_for_aes()) {
5960 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5961 // compatibility issues between Java key expansion and SPARC crypto instructions
5962 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5963 if (original_k_start == NULL) return false;
5964
5965 // Call the stub.
5966 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5967 stubAddr, stubName, TypePtr::BOTTOM,
5968 src_start, dest_start, k_start, original_k_start);
5969 } else {
5970 // Call the stub.
5971 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5972 stubAddr, stubName, TypePtr::BOTTOM,
5973 src_start, dest_start, k_start);
5974 }
5975
5976 return true;
5977 }
5978
5979 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5980 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5981 address stubAddr;
5982 const char *stubName;
5983
5984 assert(UseAES, "need AES instruction support");
5985
5986 switch(id) {
5987 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5988 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5989 stubName = "cipherBlockChaining_encryptAESCrypt";
5990 break;
5991 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5992 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5993 stubName = "cipherBlockChaining_decryptAESCrypt";
5994 break;
5995 }
5996 if (stubAddr == NULL) return false;
5997
5998 Node* cipherBlockChaining_object = argument(0);
5999 Node* src = argument(1);
6000 Node* src_offset = argument(2);
6001 Node* len = argument(3);
6002 Node* dest = argument(4);
6003 Node* dest_offset = argument(5);
6004
6005 // Resolve src and dest arrays for ShenandoahGC.
6006 src = shenandoah_read_barrier(src);
6007 dest = shenandoah_write_barrier(dest);
6008
6009 // (1) src and dest are arrays.
6010 const Type* src_type = src->Value(&_gvn);
6011 const Type* dest_type = dest->Value(&_gvn);
6012 const TypeAryPtr* top_src = src_type->isa_aryptr();
6013 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6014 assert (top_src != NULL && top_src->klass() != NULL
6015 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6016
6017 // checks are the responsibility of the caller
6018 Node* src_start = src;
6019 Node* dest_start = dest;
6020 if (src_offset != NULL || dest_offset != NULL) {
6021 assert(src_offset != NULL && dest_offset != NULL, "");
6022 src_start = array_element_address(src, src_offset, T_BYTE);
6023 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6024 }
6025
6026 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6027 // (because of the predicated logic executed earlier).
6028 // so we cast it here safely.
6029 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6030
6031 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6032 if (embeddedCipherObj == NULL) return false;
6033
6034 // cast it to what we know it will be at runtime
6035 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6036 assert(tinst != NULL, "CBC obj is null");
6037 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6038 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6039 assert(klass_AESCrypt->is_loaded(), "predicate checks that this class is loaded");
6040
6041 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6042 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6043 const TypeOopPtr* xtype = aklass->as_instance_type();
6044 Node* aescrypt_object = new CheckCastPPNode(control(), embeddedCipherObj, xtype);
6045 aescrypt_object = _gvn.transform(aescrypt_object);
6046
6047 // we need to get the start of the aescrypt_object's expanded key array
6048 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6049 if (k_start == NULL) return false;
6050
6051 // similarly, get the start address of the r vector
6052 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6053
6054 objRvec = shenandoah_write_barrier(objRvec);
6055
6056 if (objRvec == NULL) return false;
6057 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6058
6059 Node* cbcCrypt;
6060 if (Matcher::pass_original_key_for_aes()) {
6061 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6062 // compatibility issues between Java key expansion and SPARC crypto instructions
6063 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6064 if (original_k_start == NULL) return false;
6065
6066 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6067 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6068 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6069 stubAddr, stubName, TypePtr::BOTTOM,
6070 src_start, dest_start, k_start, r_start, len, original_k_start);
6071 } else {
6072 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6073 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6074 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6075 stubAddr, stubName, TypePtr::BOTTOM,
6076 src_start, dest_start, k_start, r_start, len);
6077 }
6078
6079 // return cipher length (int)
6080 Node* retvalue = _gvn.transform(new ProjNode(cbcCrypt, TypeFunc::Parms));
6081 set_result(retvalue);
6082 return true;
6083 }
6084
6085 //------------------------------get_key_start_from_aescrypt_object-----------------------
6086 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6087 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6088 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6089 if (objAESCryptKey == NULL) return (Node *) NULL;
6090
6091 objAESCryptKey = shenandoah_read_barrier(objAESCryptKey);
6092
6093 // now have the array, need to get the start address of the K array
6094 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6095 return k_start;
6096 }
6097
6098 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6099 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6100 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6101 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6102 if (objAESCryptKey == NULL) return (Node *) NULL;
6103
6104 // now have the array, need to get the start address of the lastKey array
6105 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6106 return original_k_start;
6107 }
6108
6109 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6110 // Return node representing slow path of predicate check.
6111 // the pseudo code we want to emulate with this predicate is:
6112 // for encryption:
6113 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6114 // for decryption:
6115 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6116 // note cipher==plain is more conservative than the original java code but that's OK
6117 //
6118 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6119 // The receiver was checked for NULL already.
6120 Node* objCBC = argument(0);
6121
6122 // Load embeddedCipher field of CipherBlockChaining object.
6123 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6124
6125 // get AESCrypt klass for instanceOf check
6126 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6127 // will have same classloader as CipherBlockChaining object
6128 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6129 assert(tinst != NULL, "CBCobj is null");
6130 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6131
6132 // we want to do an instanceof comparison against the AESCrypt class
6133 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6134 if (!klass_AESCrypt->is_loaded()) {
6135 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6136 Node* ctrl = control();
6137 set_control(top()); // no regular fast path
6138 return ctrl;
6139 }
6140 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6141
6142 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6143 Node* cmp_instof = _gvn.transform(new CmpINode(instof, intcon(1)));
6144 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6145
6146 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6147
6148 // for encryption, we are done
6149 if (!decrypting)
6150 return instof_false; // even if it is NULL
6151
6152 // for decryption, we need to add a further check to avoid
6153 // taking the intrinsic path when cipher and plain are the same
6154 // see the original java code for why.
6155 RegionNode* region = new RegionNode(3);
6156 region->init_req(1, instof_false);
6157 Node* src = argument(1);
6158 Node* dest = argument(4);
6159 Node* cmp_src_dest = _gvn.transform(new CmpPNode(src, dest));
6160 Node* bool_src_dest = _gvn.transform(new BoolNode(cmp_src_dest, BoolTest::eq));
6161 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6162 region->init_req(2, src_dest_conjoint);
6163
6164 record_for_igvn(region);
6165 return _gvn.transform(region);
6166 }
6167
6168 //------------------------------inline_ghash_processBlocks
6169 bool LibraryCallKit::inline_ghash_processBlocks() {
6170 address stubAddr;
6171 const char *stubName;
6172 assert(UseGHASHIntrinsics, "need GHASH intrinsics support");
6173
6174 stubAddr = StubRoutines::ghash_processBlocks();
6175 stubName = "ghash_processBlocks";
6176
6177 Node* data = argument(0);
6178 Node* offset = argument(1);
6179 Node* len = argument(2);
6180 Node* state = argument(3);
6181 Node* subkeyH = argument(4);
6182
6183 Node* state_start = array_element_address(state, intcon(0), T_LONG);
6184 assert(state_start, "state is NULL");
6185 Node* subkeyH_start = array_element_address(subkeyH, intcon(0), T_LONG);
6186 assert(subkeyH_start, "subkeyH is NULL");
6187 Node* data_start = array_element_address(data, offset, T_BYTE);
6188 assert(data_start, "data is NULL");
6189
6190 Node* ghash = make_runtime_call(RC_LEAF|RC_NO_FP,
6191 OptoRuntime::ghash_processBlocks_Type(),
6192 stubAddr, stubName, TypePtr::BOTTOM,
6193 state_start, subkeyH_start, data_start, len);
6194 return true;
6195 }
6196
6197 //------------------------------inline_sha_implCompress-----------------------
6198 //
6199 // Calculate SHA (i.e., SHA-1) for single-block byte[] array.
6200 // void com.sun.security.provider.SHA.implCompress(byte[] buf, int ofs)
6201 //
6202 // Calculate SHA2 (i.e., SHA-244 or SHA-256) for single-block byte[] array.
6203 // void com.sun.security.provider.SHA2.implCompress(byte[] buf, int ofs)
6204 //
6205 // Calculate SHA5 (i.e., SHA-384 or SHA-512) for single-block byte[] array.
6206 // void com.sun.security.provider.SHA5.implCompress(byte[] buf, int ofs)
6207 //
6208 bool LibraryCallKit::inline_sha_implCompress(vmIntrinsics::ID id) {
6209 assert(callee()->signature()->size() == 2, "sha_implCompress has 2 parameters");
6210
6211 Node* sha_obj = argument(0);
6212 Node* src = argument(1); // type oop
6213 Node* ofs = argument(2); // type int
6214
6215 const Type* src_type = src->Value(&_gvn);
6216 const TypeAryPtr* top_src = src_type->isa_aryptr();
6217 if (top_src == NULL || top_src->klass() == NULL) {
6218 // failed array check
6219 return false;
6220 }
6221 // Figure out the size and type of the elements we will be copying.
6222 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6223 if (src_elem != T_BYTE) {
6224 return false;
6225 }
6226 // 'src_start' points to src array + offset
6227 Node* src_start = array_element_address(src, ofs, src_elem);
6228 Node* state = NULL;
6229 address stubAddr;
6230 const char *stubName;
6231
6232 switch(id) {
6233 case vmIntrinsics::_sha_implCompress:
6234 assert(UseSHA1Intrinsics, "need SHA1 instruction support");
6235 state = get_state_from_sha_object(sha_obj);
6236 stubAddr = StubRoutines::sha1_implCompress();
6237 stubName = "sha1_implCompress";
6238 break;
6239 case vmIntrinsics::_sha2_implCompress:
6240 assert(UseSHA256Intrinsics, "need SHA256 instruction support");
6241 state = get_state_from_sha_object(sha_obj);
6242 stubAddr = StubRoutines::sha256_implCompress();
6243 stubName = "sha256_implCompress";
6244 break;
6245 case vmIntrinsics::_sha5_implCompress:
6246 assert(UseSHA512Intrinsics, "need SHA512 instruction support");
6247 state = get_state_from_sha5_object(sha_obj);
6248 stubAddr = StubRoutines::sha512_implCompress();
6249 stubName = "sha512_implCompress";
6250 break;
6251 default:
6252 fatal_unexpected_iid(id);
6253 return false;
6254 }
6255 if (state == NULL) return false;
6256
6257 // Call the stub.
6258 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::sha_implCompress_Type(),
6259 stubAddr, stubName, TypePtr::BOTTOM,
6260 src_start, state);
6261
6262 return true;
6263 }
6264
6265 //------------------------------inline_digestBase_implCompressMB-----------------------
6266 //
6267 // Calculate SHA/SHA2/SHA5 for multi-block byte[] array.
6268 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
6269 //
6270 bool LibraryCallKit::inline_digestBase_implCompressMB(int predicate) {
6271 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6272 "need SHA1/SHA256/SHA512 instruction support");
6273 assert((uint)predicate < 3, "sanity");
6274 assert(callee()->signature()->size() == 3, "digestBase_implCompressMB has 3 parameters");
6275
6276 Node* digestBase_obj = argument(0); // The receiver was checked for NULL already.
6277 Node* src = argument(1); // byte[] array
6278 Node* ofs = argument(2); // type int
6279 Node* limit = argument(3); // type int
6280
6281 const Type* src_type = src->Value(&_gvn);
6282 const TypeAryPtr* top_src = src_type->isa_aryptr();
6283 if (top_src == NULL || top_src->klass() == NULL) {
6284 // failed array check
6285 return false;
6286 }
6287 // Figure out the size and type of the elements we will be copying.
6288 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
6289 if (src_elem != T_BYTE) {
6290 return false;
6291 }
6292 // 'src_start' points to src array + offset
6293 Node* src_start = array_element_address(src, ofs, src_elem);
6294
6295 const char* klass_SHA_name = NULL;
6296 const char* stub_name = NULL;
6297 address stub_addr = NULL;
6298 bool long_state = false;
6299
6300 switch (predicate) {
6301 case 0:
6302 if (UseSHA1Intrinsics) {
6303 klass_SHA_name = "sun/security/provider/SHA";
6304 stub_name = "sha1_implCompressMB";
6305 stub_addr = StubRoutines::sha1_implCompressMB();
6306 }
6307 break;
6308 case 1:
6309 if (UseSHA256Intrinsics) {
6310 klass_SHA_name = "sun/security/provider/SHA2";
6311 stub_name = "sha256_implCompressMB";
6312 stub_addr = StubRoutines::sha256_implCompressMB();
6313 }
6314 break;
6315 case 2:
6316 if (UseSHA512Intrinsics) {
6317 klass_SHA_name = "sun/security/provider/SHA5";
6318 stub_name = "sha512_implCompressMB";
6319 stub_addr = StubRoutines::sha512_implCompressMB();
6320 long_state = true;
6321 }
6322 break;
6323 default:
6324 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6325 }
6326 if (klass_SHA_name != NULL) {
6327 // get DigestBase klass to lookup for SHA klass
6328 const TypeInstPtr* tinst = _gvn.type(digestBase_obj)->isa_instptr();
6329 assert(tinst != NULL, "digestBase_obj is not instance???");
6330 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6331
6332 ciKlass* klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6333 assert(klass_SHA->is_loaded(), "predicate checks that this class is loaded");
6334 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6335 return inline_sha_implCompressMB(digestBase_obj, instklass_SHA, long_state, stub_addr, stub_name, src_start, ofs, limit);
6336 }
6337 return false;
6338 }
6339 //------------------------------inline_sha_implCompressMB-----------------------
6340 bool LibraryCallKit::inline_sha_implCompressMB(Node* digestBase_obj, ciInstanceKlass* instklass_SHA,
6341 bool long_state, address stubAddr, const char *stubName,
6342 Node* src_start, Node* ofs, Node* limit) {
6343 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_SHA);
6344 const TypeOopPtr* xtype = aklass->as_instance_type();
6345 Node* sha_obj = new CheckCastPPNode(control(), digestBase_obj, xtype);
6346 sha_obj = _gvn.transform(sha_obj);
6347
6348 Node* state;
6349 if (long_state) {
6350 state = get_state_from_sha5_object(sha_obj);
6351 } else {
6352 state = get_state_from_sha_object(sha_obj);
6353 }
6354 if (state == NULL) return false;
6355
6356 // Call the stub.
6357 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
6358 OptoRuntime::digestBase_implCompressMB_Type(),
6359 stubAddr, stubName, TypePtr::BOTTOM,
6360 src_start, state, ofs, limit);
6361 // return ofs (int)
6362 Node* result = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
6363 set_result(result);
6364
6365 return true;
6366 }
6367
6368 //------------------------------get_state_from_sha_object-----------------------
6369 Node * LibraryCallKit::get_state_from_sha_object(Node *sha_object) {
6370 Node* sha_state = load_field_from_object(sha_object, "state", "[I", /*is_exact*/ false);
6371 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA/SHA2");
6372 if (sha_state == NULL) return (Node *) NULL;
6373
6374 // now have the array, need to get the start address of the state array
6375 Node* state = array_element_address(sha_state, intcon(0), T_INT);
6376 return state;
6377 }
6378
6379 //------------------------------get_state_from_sha5_object-----------------------
6380 Node * LibraryCallKit::get_state_from_sha5_object(Node *sha_object) {
6381 Node* sha_state = load_field_from_object(sha_object, "state", "[J", /*is_exact*/ false);
6382 assert (sha_state != NULL, "wrong version of sun.security.provider.SHA5");
6383 if (sha_state == NULL) return (Node *) NULL;
6384
6385 // now have the array, need to get the start address of the state array
6386 Node* state = array_element_address(sha_state, intcon(0), T_LONG);
6387 return state;
6388 }
6389
6390 //----------------------------inline_digestBase_implCompressMB_predicate----------------------------
6391 // Return node representing slow path of predicate check.
6392 // the pseudo code we want to emulate with this predicate is:
6393 // if (digestBaseObj instanceof SHA/SHA2/SHA5) do_intrinsic, else do_javapath
6394 //
6395 Node* LibraryCallKit::inline_digestBase_implCompressMB_predicate(int predicate) {
6396 assert(UseSHA1Intrinsics || UseSHA256Intrinsics || UseSHA512Intrinsics,
6397 "need SHA1/SHA256/SHA512 instruction support");
6398 assert((uint)predicate < 3, "sanity");
6399
6400 // The receiver was checked for NULL already.
6401 Node* digestBaseObj = argument(0);
6402
6403 // get DigestBase klass for instanceOf check
6404 const TypeInstPtr* tinst = _gvn.type(digestBaseObj)->isa_instptr();
6405 assert(tinst != NULL, "digestBaseObj is null");
6406 assert(tinst->klass()->is_loaded(), "DigestBase is not loaded");
6407
6408 const char* klass_SHA_name = NULL;
6409 switch (predicate) {
6410 case 0:
6411 if (UseSHA1Intrinsics) {
6412 // we want to do an instanceof comparison against the SHA class
6413 klass_SHA_name = "sun/security/provider/SHA";
6414 }
6415 break;
6416 case 1:
6417 if (UseSHA256Intrinsics) {
6418 // we want to do an instanceof comparison against the SHA2 class
6419 klass_SHA_name = "sun/security/provider/SHA2";
6420 }
6421 break;
6422 case 2:
6423 if (UseSHA512Intrinsics) {
6424 // we want to do an instanceof comparison against the SHA5 class
6425 klass_SHA_name = "sun/security/provider/SHA5";
6426 }
6427 break;
6428 default:
6429 fatal(err_msg_res("unknown SHA intrinsic predicate: %d", predicate));
6430 }
6431
6432 ciKlass* klass_SHA = NULL;
6433 if (klass_SHA_name != NULL) {
6434 klass_SHA = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make(klass_SHA_name));
6435 }
6436 if ((klass_SHA == NULL) || !klass_SHA->is_loaded()) {
6437 // if none of SHA/SHA2/SHA5 is loaded, we never take the intrinsic fast path
6438 Node* ctrl = control();
6439 set_control(top()); // no intrinsic path
6440 return ctrl;
6441 }
6442 ciInstanceKlass* instklass_SHA = klass_SHA->as_instance_klass();
6443
6444 Node* instofSHA = gen_instanceof(digestBaseObj, makecon(TypeKlassPtr::make(instklass_SHA)));
6445 Node* cmp_instof = _gvn.transform(new CmpINode(instofSHA, intcon(1)));
6446 Node* bool_instof = _gvn.transform(new BoolNode(cmp_instof, BoolTest::ne));
6447 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6448
6449 return instof_false; // even if it is NULL
6450 }
6451
6452 bool LibraryCallKit::inline_profileBoolean() {
6453 Node* counts = argument(1);
6454 const TypeAryPtr* ary = NULL;
6455 ciArray* aobj = NULL;
6456 if (counts->is_Con()
6457 && (ary = counts->bottom_type()->isa_aryptr()) != NULL
6458 && (aobj = ary->const_oop()->as_array()) != NULL
6459 && (aobj->length() == 2)) {
6460 // Profile is int[2] where [0] and [1] correspond to false and true value occurrences respectively.
6461 jint false_cnt = aobj->element_value(0).as_int();
6462 jint true_cnt = aobj->element_value(1).as_int();
6463
6464 if (C->log() != NULL) {
6465 C->log()->elem("observe source='profileBoolean' false='%d' true='%d'",
6466 false_cnt, true_cnt);
6467 }
6468
6469 if (false_cnt + true_cnt == 0) {
6470 // According to profile, never executed.
6471 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6472 Deoptimization::Action_reinterpret);
6473 return true;
6474 }
6475
6476 // result is a boolean (0 or 1) and its profile (false_cnt & true_cnt)
6477 // is a number of each value occurrences.
6478 Node* result = argument(0);
6479 if (false_cnt == 0 || true_cnt == 0) {
6480 // According to profile, one value has been never seen.
6481 int expected_val = (false_cnt == 0) ? 1 : 0;
6482
6483 Node* cmp = _gvn.transform(new CmpINode(result, intcon(expected_val)));
6484 Node* test = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
6485
6486 IfNode* check = create_and_map_if(control(), test, PROB_ALWAYS, COUNT_UNKNOWN);
6487 Node* fast_path = _gvn.transform(new IfTrueNode(check));
6488 Node* slow_path = _gvn.transform(new IfFalseNode(check));
6489
6490 { // Slow path: uncommon trap for never seen value and then reexecute
6491 // MethodHandleImpl::profileBoolean() to bump the count, so JIT knows
6492 // the value has been seen at least once.
6493 PreserveJVMState pjvms(this);
6494 PreserveReexecuteState preexecs(this);
6495 jvms()->set_should_reexecute(true);
6496
6497 set_control(slow_path);
6498 set_i_o(i_o());
6499
6500 uncommon_trap_exact(Deoptimization::Reason_intrinsic,
6501 Deoptimization::Action_reinterpret);
6502 }
6503 // The guard for never seen value enables sharpening of the result and
6504 // returning a constant. It allows to eliminate branches on the same value
6505 // later on.
6506 set_control(fast_path);
6507 result = intcon(expected_val);
6508 }
6509 // Stop profiling.
6510 // MethodHandleImpl::profileBoolean() has profiling logic in its bytecode.
6511 // By replacing method body with profile data (represented as ProfileBooleanNode
6512 // on IR level) we effectively disable profiling.
6513 // It enables full speed execution once optimized code is generated.
6514 Node* profile = _gvn.transform(new ProfileBooleanNode(result, false_cnt, true_cnt));
6515 C->record_for_igvn(profile);
6516 set_result(profile);
6517 return true;
6518 } else {
6519 // Continue profiling.
6520 // Profile data isn't available at the moment. So, execute method's bytecode version.
6521 // Usually, when GWT LambdaForms are profiled it means that a stand-alone nmethod
6522 // is compiled and counters aren't available since corresponding MethodHandle
6523 // isn't a compile-time constant.
6524 return false;
6525 }
6526 }
6527
6528 bool LibraryCallKit::inline_isCompileConstant() {
6529 Node* n = argument(0);
6530 set_result(n->is_Con() ? intcon(1) : intcon(0));
6531 return true;
6532 }