1 /*
2 * Copyright (c) 1999, 2013, 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 "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mathexactnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "prims/nativeLookup.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "trace/traceMacros.hpp"
43
44 class LibraryIntrinsic : public InlineCallGenerator {
45 // Extend the set of intrinsics known to the runtime:
46 public:
47 private:
48 bool _is_virtual;
49 bool _is_predicted;
50 bool _does_virtual_dispatch;
51 vmIntrinsics::ID _intrinsic_id;
52
53 public:
54 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
55 : InlineCallGenerator(m),
56 _is_virtual(is_virtual),
57 _is_predicted(is_predicted),
58 _does_virtual_dispatch(does_virtual_dispatch),
59 _intrinsic_id(id)
60 {
61 }
62 virtual bool is_intrinsic() const { return true; }
63 virtual bool is_virtual() const { return _is_virtual; }
64 virtual bool is_predicted() const { return _is_predicted; }
65 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
66 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
67 virtual Node* generate_predicate(JVMState* jvms);
68 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
69 };
70
71
72 // Local helper class for LibraryIntrinsic:
73 class LibraryCallKit : public GraphKit {
74 private:
75 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
76 Node* _result; // the result node, if any
77 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
78
79 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
80
81 public:
82 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
83 : GraphKit(jvms),
84 _intrinsic(intrinsic),
85 _result(NULL)
86 {
87 // Check if this is a root compile. In that case we don't have a caller.
88 if (!jvms->has_method()) {
89 _reexecute_sp = sp();
90 } else {
91 // Find out how many arguments the interpreter needs when deoptimizing
92 // and save the stack pointer value so it can used by uncommon_trap.
93 // We find the argument count by looking at the declared signature.
94 bool ignored_will_link;
95 ciSignature* declared_signature = NULL;
96 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
97 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
98 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
99 }
100 }
101
102 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
103
104 ciMethod* caller() const { return jvms()->method(); }
105 int bci() const { return jvms()->bci(); }
106 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
107 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
108 ciMethod* callee() const { return _intrinsic->method(); }
109
110 bool try_to_inline();
111 Node* try_to_predicate();
112
113 void push_result() {
114 // Push the result onto the stack.
115 if (!stopped() && result() != NULL) {
116 BasicType bt = result()->bottom_type()->basic_type();
117 push_node(bt, result());
118 }
119 }
120
121 private:
122 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
123 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
124 }
125
126 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
127 void set_result(RegionNode* region, PhiNode* value);
128 Node* result() { return _result; }
129
130 virtual int reexecute_sp() { return _reexecute_sp; }
131
132 // Helper functions to inline natives
133 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
134 Node* generate_slow_guard(Node* test, RegionNode* region);
135 Node* generate_fair_guard(Node* test, RegionNode* region);
136 Node* generate_negative_guard(Node* index, RegionNode* region,
137 // resulting CastII of index:
138 Node* *pos_index = NULL);
139 Node* generate_nonpositive_guard(Node* index, bool never_negative,
140 // resulting CastII of index:
141 Node* *pos_index = NULL);
142 Node* generate_limit_guard(Node* offset, Node* subseq_length,
143 Node* array_length,
144 RegionNode* region);
145 Node* generate_current_thread(Node* &tls_output);
146 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
147 bool disjoint_bases, const char* &name, bool dest_uninitialized);
148 Node* load_mirror_from_klass(Node* klass);
149 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
150 RegionNode* region, int null_path,
151 int offset);
152 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
153 RegionNode* region, int null_path) {
154 int offset = java_lang_Class::klass_offset_in_bytes();
155 return load_klass_from_mirror_common(mirror, never_see_null,
156 region, null_path,
157 offset);
158 }
159 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::array_klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* generate_access_flags_guard(Node* kls,
167 int modifier_mask, int modifier_bits,
168 RegionNode* region);
169 Node* generate_interface_guard(Node* kls, RegionNode* region);
170 Node* generate_array_guard(Node* kls, RegionNode* region) {
171 return generate_array_guard_common(kls, region, false, false);
172 }
173 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
174 return generate_array_guard_common(kls, region, false, true);
175 }
176 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, true, false);
178 }
179 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, true, true);
181 }
182 Node* generate_array_guard_common(Node* kls, RegionNode* region,
183 bool obj_array, bool not_array);
184 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
185 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
186 bool is_virtual = false, bool is_static = false);
187 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
188 return generate_method_call(method_id, false, true);
189 }
190 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
191 return generate_method_call(method_id, true, false);
192 }
193 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
194
195 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
196 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
197 bool inline_string_compareTo();
198 bool inline_string_indexOf();
199 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
200 bool inline_string_equals();
201 Node* round_double_node(Node* n);
202 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
203 bool inline_math_native(vmIntrinsics::ID id);
204 bool inline_trig(vmIntrinsics::ID id);
205 bool inline_math(vmIntrinsics::ID id);
206 bool inline_math_mathExact(Node* math);
207 bool inline_math_addExact();
208 bool inline_exp();
209 bool inline_pow();
210 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
211 bool inline_min_max(vmIntrinsics::ID id);
212 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
213 // This returns Type::AnyPtr, RawPtr, or OopPtr.
214 int classify_unsafe_addr(Node* &base, Node* &offset);
215 Node* make_unsafe_address(Node* base, Node* offset);
216 // Helper for inline_unsafe_access.
217 // Generates the guards that check whether the result of
218 // Unsafe.getObject should be recorded in an SATB log buffer.
219 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
220 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
221 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
222 static bool klass_needs_init_guard(Node* kls);
223 bool inline_unsafe_allocate();
224 bool inline_unsafe_copyMemory();
225 bool inline_native_currentThread();
226 #ifdef TRACE_HAVE_INTRINSICS
227 bool inline_native_classID();
228 bool inline_native_threadID();
229 #endif
230 bool inline_native_time_funcs(address method, const char* funcName);
231 bool inline_native_isInterrupted();
232 bool inline_native_Class_query(vmIntrinsics::ID id);
233 bool inline_native_subtype_check();
234
235 bool inline_native_newArray();
236 bool inline_native_getLength();
237 bool inline_array_copyOf(bool is_copyOfRange);
238 bool inline_array_equals();
239 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
240 bool inline_native_clone(bool is_virtual);
241 bool inline_native_Reflection_getCallerClass();
242 // Helper function for inlining native object hash method
243 bool inline_native_hashcode(bool is_virtual, bool is_static);
244 bool inline_native_getClass();
245
246 // Helper functions for inlining arraycopy
247 bool inline_arraycopy();
248 void generate_arraycopy(const TypePtr* adr_type,
249 BasicType basic_elem_type,
250 Node* src, Node* src_offset,
251 Node* dest, Node* dest_offset,
252 Node* copy_length,
253 bool disjoint_bases = false,
254 bool length_never_negative = false,
255 RegionNode* slow_region = NULL);
256 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
257 RegionNode* slow_region);
258 void generate_clear_array(const TypePtr* adr_type,
259 Node* dest,
260 BasicType basic_elem_type,
261 Node* slice_off,
262 Node* slice_len,
263 Node* slice_end);
264 bool generate_block_arraycopy(const TypePtr* adr_type,
265 BasicType basic_elem_type,
266 AllocateNode* alloc,
267 Node* src, Node* src_offset,
268 Node* dest, Node* dest_offset,
269 Node* dest_size, bool dest_uninitialized);
270 void generate_slow_arraycopy(const TypePtr* adr_type,
271 Node* src, Node* src_offset,
272 Node* dest, Node* dest_offset,
273 Node* copy_length, bool dest_uninitialized);
274 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
275 Node* dest_elem_klass,
276 Node* src, Node* src_offset,
277 Node* dest, Node* dest_offset,
278 Node* copy_length, bool dest_uninitialized);
279 Node* generate_generic_arraycopy(const TypePtr* adr_type,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* copy_length, bool dest_uninitialized);
283 void generate_unchecked_arraycopy(const TypePtr* adr_type,
284 BasicType basic_elem_type,
285 bool disjoint_bases,
286 Node* src, Node* src_offset,
287 Node* dest, Node* dest_offset,
288 Node* copy_length, bool dest_uninitialized);
289 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
290 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
291 bool inline_unsafe_ordered_store(BasicType type);
292 bool inline_unsafe_fence(vmIntrinsics::ID id);
293 bool inline_fp_conversions(vmIntrinsics::ID id);
294 bool inline_number_methods(vmIntrinsics::ID id);
295 bool inline_reference_get();
296 bool inline_aescrypt_Block(vmIntrinsics::ID id);
297 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
298 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
299 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
300 bool inline_encodeISOArray();
301 bool inline_updateCRC32();
302 bool inline_updateBytesCRC32();
303 bool inline_updateByteBufferCRC32();
304 };
305
306
307 //---------------------------make_vm_intrinsic----------------------------
308 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
309 vmIntrinsics::ID id = m->intrinsic_id();
310 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
311
312 if (DisableIntrinsic[0] != '\0'
313 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
314 // disabled by a user request on the command line:
315 // example: -XX:DisableIntrinsic=_hashCode,_getClass
316 return NULL;
317 }
318
319 if (!m->is_loaded()) {
320 // do not attempt to inline unloaded methods
321 return NULL;
322 }
323
324 // Only a few intrinsics implement a virtual dispatch.
325 // They are expensive calls which are also frequently overridden.
326 if (is_virtual) {
327 switch (id) {
328 case vmIntrinsics::_hashCode:
329 case vmIntrinsics::_clone:
330 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
331 break;
332 default:
333 return NULL;
334 }
335 }
336
337 // -XX:-InlineNatives disables nearly all intrinsics:
338 if (!InlineNatives) {
339 switch (id) {
340 case vmIntrinsics::_indexOf:
341 case vmIntrinsics::_compareTo:
342 case vmIntrinsics::_equals:
343 case vmIntrinsics::_equalsC:
344 case vmIntrinsics::_getAndAddInt:
345 case vmIntrinsics::_getAndAddLong:
346 case vmIntrinsics::_getAndSetInt:
347 case vmIntrinsics::_getAndSetLong:
348 case vmIntrinsics::_getAndSetObject:
349 case vmIntrinsics::_loadFence:
350 case vmIntrinsics::_storeFence:
351 case vmIntrinsics::_fullFence:
352 break; // InlineNatives does not control String.compareTo
353 case vmIntrinsics::_Reference_get:
354 break; // InlineNatives does not control Reference.get
355 default:
356 return NULL;
357 }
358 }
359
360 bool is_predicted = false;
361 bool does_virtual_dispatch = false;
362
363 switch (id) {
364 case vmIntrinsics::_compareTo:
365 if (!SpecialStringCompareTo) return NULL;
366 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
367 break;
368 case vmIntrinsics::_indexOf:
369 if (!SpecialStringIndexOf) return NULL;
370 break;
371 case vmIntrinsics::_equals:
372 if (!SpecialStringEquals) return NULL;
373 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
374 break;
375 case vmIntrinsics::_equalsC:
376 if (!SpecialArraysEquals) return NULL;
377 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
378 break;
379 case vmIntrinsics::_arraycopy:
380 if (!InlineArrayCopy) return NULL;
381 break;
382 case vmIntrinsics::_copyMemory:
383 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
384 if (!InlineArrayCopy) return NULL;
385 break;
386 case vmIntrinsics::_hashCode:
387 if (!InlineObjectHash) return NULL;
388 does_virtual_dispatch = true;
389 break;
390 case vmIntrinsics::_clone:
391 does_virtual_dispatch = true;
392 case vmIntrinsics::_copyOf:
393 case vmIntrinsics::_copyOfRange:
394 if (!InlineObjectCopy) return NULL;
395 // These also use the arraycopy intrinsic mechanism:
396 if (!InlineArrayCopy) return NULL;
397 break;
398 case vmIntrinsics::_encodeISOArray:
399 if (!SpecialEncodeISOArray) return NULL;
400 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
401 break;
402 case vmIntrinsics::_checkIndex:
403 // We do not intrinsify this. The optimizer does fine with it.
404 return NULL;
405
406 case vmIntrinsics::_getCallerClass:
407 if (!UseNewReflection) return NULL;
408 if (!InlineReflectionGetCallerClass) return NULL;
409 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
410 break;
411
412 case vmIntrinsics::_bitCount_i:
413 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
414 break;
415
416 case vmIntrinsics::_bitCount_l:
417 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
418 break;
419
420 case vmIntrinsics::_numberOfLeadingZeros_i:
421 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
422 break;
423
424 case vmIntrinsics::_numberOfLeadingZeros_l:
425 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
426 break;
427
428 case vmIntrinsics::_numberOfTrailingZeros_i:
429 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
430 break;
431
432 case vmIntrinsics::_numberOfTrailingZeros_l:
433 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
434 break;
435
436 case vmIntrinsics::_reverseBytes_c:
437 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
438 break;
439 case vmIntrinsics::_reverseBytes_s:
440 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
441 break;
442 case vmIntrinsics::_reverseBytes_i:
443 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
444 break;
445 case vmIntrinsics::_reverseBytes_l:
446 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
447 break;
448
449 case vmIntrinsics::_Reference_get:
450 // Use the intrinsic version of Reference.get() so that the value in
451 // the referent field can be registered by the G1 pre-barrier code.
452 // Also add memory barrier to prevent commoning reads from this field
453 // across safepoint since GC can change it value.
454 break;
455
456 case vmIntrinsics::_compareAndSwapObject:
457 #ifdef _LP64
458 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
459 #endif
460 break;
461
462 case vmIntrinsics::_compareAndSwapLong:
463 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
464 break;
465
466 case vmIntrinsics::_getAndAddInt:
467 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
468 break;
469
470 case vmIntrinsics::_getAndAddLong:
471 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
472 break;
473
474 case vmIntrinsics::_getAndSetInt:
475 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
476 break;
477
478 case vmIntrinsics::_getAndSetLong:
479 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
480 break;
481
482 case vmIntrinsics::_getAndSetObject:
483 #ifdef _LP64
484 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
485 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
486 break;
487 #else
488 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
489 break;
490 #endif
491
492 case vmIntrinsics::_aescrypt_encryptBlock:
493 case vmIntrinsics::_aescrypt_decryptBlock:
494 if (!UseAESIntrinsics) return NULL;
495 break;
496
497 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
498 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
499 if (!UseAESIntrinsics) return NULL;
500 // these two require the predicated logic
501 is_predicted = true;
502 break;
503
504 case vmIntrinsics::_updateCRC32:
505 case vmIntrinsics::_updateBytesCRC32:
506 case vmIntrinsics::_updateByteBufferCRC32:
507 if (!UseCRC32Intrinsics) return NULL;
508 break;
509
510 case vmIntrinsics::_addExact:
511 if (!Matcher::match_rule_supported(Op_AddExactI)) {
512 return NULL;
513 }
514 if (!UseMathExactIntrinsics) {
515 return NULL;
516 }
517 break;
518
519 default:
520 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
521 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
522 break;
523 }
524
525 // -XX:-InlineClassNatives disables natives from the Class class.
526 // The flag applies to all reflective calls, notably Array.newArray
527 // (visible to Java programmers as Array.newInstance).
528 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
529 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
530 if (!InlineClassNatives) return NULL;
531 }
532
533 // -XX:-InlineThreadNatives disables natives from the Thread class.
534 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
535 if (!InlineThreadNatives) return NULL;
536 }
537
538 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
539 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
540 m->holder()->name() == ciSymbol::java_lang_Float() ||
541 m->holder()->name() == ciSymbol::java_lang_Double()) {
542 if (!InlineMathNatives) return NULL;
543 }
544
545 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
546 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
547 if (!InlineUnsafeOps) return NULL;
548 }
549
550 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
551 }
552
553 //----------------------register_library_intrinsics-----------------------
554 // Initialize this file's data structures, for each Compile instance.
555 void Compile::register_library_intrinsics() {
556 // Nothing to do here.
557 }
558
559 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
560 LibraryCallKit kit(jvms, this);
561 Compile* C = kit.C;
562 int nodes = C->unique();
563 #ifndef PRODUCT
564 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
565 char buf[1000];
566 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
567 tty->print_cr("Intrinsic %s", str);
568 }
569 #endif
570 ciMethod* callee = kit.callee();
571 const int bci = kit.bci();
572
573 // Try to inline the intrinsic.
574 if (kit.try_to_inline()) {
575 if (C->print_intrinsics() || C->print_inlining()) {
576 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
577 }
578 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
579 if (C->log()) {
580 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
581 vmIntrinsics::name_at(intrinsic_id()),
582 (is_virtual() ? " virtual='1'" : ""),
583 C->unique() - nodes);
584 }
585 // Push the result from the inlined method onto the stack.
586 kit.push_result();
587 return kit.transfer_exceptions_into_jvms();
588 }
589
590 // The intrinsic bailed out
591 if (C->print_intrinsics() || C->print_inlining()) {
592 if (jvms->has_method()) {
593 // Not a root compile.
594 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
595 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
596 } else {
597 // Root compile
598 tty->print("Did not generate intrinsic %s%s at bci:%d in",
599 vmIntrinsics::name_at(intrinsic_id()),
600 (is_virtual() ? " (virtual)" : ""), bci);
601 }
602 }
603 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
604 return NULL;
605 }
606
607 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
608 LibraryCallKit kit(jvms, this);
609 Compile* C = kit.C;
610 int nodes = C->unique();
611 #ifndef PRODUCT
612 assert(is_predicted(), "sanity");
613 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
614 char buf[1000];
615 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
616 tty->print_cr("Predicate for intrinsic %s", str);
617 }
618 #endif
619 ciMethod* callee = kit.callee();
620 const int bci = kit.bci();
621
622 Node* slow_ctl = kit.try_to_predicate();
623 if (!kit.failing()) {
624 if (C->print_intrinsics() || C->print_inlining()) {
625 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
626 }
627 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
628 if (C->log()) {
629 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
630 vmIntrinsics::name_at(intrinsic_id()),
631 (is_virtual() ? " virtual='1'" : ""),
632 C->unique() - nodes);
633 }
634 return slow_ctl; // Could be NULL if the check folds.
635 }
636
637 // The intrinsic bailed out
638 if (C->print_intrinsics() || C->print_inlining()) {
639 if (jvms->has_method()) {
640 // Not a root compile.
641 const char* msg = "failed to generate predicate for intrinsic";
642 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
643 } else {
644 // Root compile
645 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
646 vmIntrinsics::name_at(intrinsic_id()),
647 (is_virtual() ? " (virtual)" : ""), bci);
648 }
649 }
650 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
651 return NULL;
652 }
653
654 bool LibraryCallKit::try_to_inline() {
655 // Handle symbolic names for otherwise undistinguished boolean switches:
656 const bool is_store = true;
657 const bool is_native_ptr = true;
658 const bool is_static = true;
659 const bool is_volatile = true;
660
661 if (!jvms()->has_method()) {
662 // Root JVMState has a null method.
663 assert(map()->memory()->Opcode() == Op_Parm, "");
664 // Insert the memory aliasing node
665 set_all_memory(reset_memory());
666 }
667 assert(merged_memory(), "");
668
669
670 switch (intrinsic_id()) {
671 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
672 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
673 case vmIntrinsics::_getClass: return inline_native_getClass();
674
675 case vmIntrinsics::_dsin:
676 case vmIntrinsics::_dcos:
677 case vmIntrinsics::_dtan:
678 case vmIntrinsics::_dabs:
679 case vmIntrinsics::_datan2:
680 case vmIntrinsics::_dsqrt:
681 case vmIntrinsics::_dexp:
682 case vmIntrinsics::_dlog:
683 case vmIntrinsics::_dlog10:
684 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
685
686 case vmIntrinsics::_min:
687 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
688
689 case vmIntrinsics::_addExact: return inline_math_addExact();
690
691 case vmIntrinsics::_arraycopy: return inline_arraycopy();
692
693 case vmIntrinsics::_compareTo: return inline_string_compareTo();
694 case vmIntrinsics::_indexOf: return inline_string_indexOf();
695 case vmIntrinsics::_equals: return inline_string_equals();
696
697 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
698 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
699 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
700 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
701 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
702 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
703 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
704 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
705 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
706
707 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
708 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
709 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
710 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
711 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
712 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
713 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
714 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
715 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
716
717 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
718 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
719 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
720 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
721 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
722 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
723 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
724 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
725
726 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
727 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
728 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
729 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
730 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
731 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
732 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
733 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
734
735 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
736 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
737 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
738 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
739 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
740 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
741 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
742 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
743 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
744
745 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
746 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
747 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
748 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
749 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
750 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
751 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
752 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
753 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
754
755 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
756 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
757 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
758 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
759
760 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
761 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
762 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
763
764 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
765 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
766 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
767
768 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
769 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
770 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
771 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
772 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
773
774 case vmIntrinsics::_loadFence:
775 case vmIntrinsics::_storeFence:
776 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
777
778 case vmIntrinsics::_currentThread: return inline_native_currentThread();
779 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
780
781 #ifdef TRACE_HAVE_INTRINSICS
782 case vmIntrinsics::_classID: return inline_native_classID();
783 case vmIntrinsics::_threadID: return inline_native_threadID();
784 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
785 #endif
786 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
787 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
788 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
789 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
790 case vmIntrinsics::_newArray: return inline_native_newArray();
791 case vmIntrinsics::_getLength: return inline_native_getLength();
792 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
793 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
794 case vmIntrinsics::_equalsC: return inline_array_equals();
795 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
796
797 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
798
799 case vmIntrinsics::_isInstance:
800 case vmIntrinsics::_getModifiers:
801 case vmIntrinsics::_isInterface:
802 case vmIntrinsics::_isArray:
803 case vmIntrinsics::_isPrimitive:
804 case vmIntrinsics::_getSuperclass:
805 case vmIntrinsics::_getComponentType:
806 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
807
808 case vmIntrinsics::_floatToRawIntBits:
809 case vmIntrinsics::_floatToIntBits:
810 case vmIntrinsics::_intBitsToFloat:
811 case vmIntrinsics::_doubleToRawLongBits:
812 case vmIntrinsics::_doubleToLongBits:
813 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
814
815 case vmIntrinsics::_numberOfLeadingZeros_i:
816 case vmIntrinsics::_numberOfLeadingZeros_l:
817 case vmIntrinsics::_numberOfTrailingZeros_i:
818 case vmIntrinsics::_numberOfTrailingZeros_l:
819 case vmIntrinsics::_bitCount_i:
820 case vmIntrinsics::_bitCount_l:
821 case vmIntrinsics::_reverseBytes_i:
822 case vmIntrinsics::_reverseBytes_l:
823 case vmIntrinsics::_reverseBytes_s:
824 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
825
826 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
827
828 case vmIntrinsics::_Reference_get: return inline_reference_get();
829
830 case vmIntrinsics::_aescrypt_encryptBlock:
831 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
832
833 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
834 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
835 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
836
837 case vmIntrinsics::_encodeISOArray:
838 return inline_encodeISOArray();
839
840 case vmIntrinsics::_updateCRC32:
841 return inline_updateCRC32();
842 case vmIntrinsics::_updateBytesCRC32:
843 return inline_updateBytesCRC32();
844 case vmIntrinsics::_updateByteBufferCRC32:
845 return inline_updateByteBufferCRC32();
846
847 default:
848 // If you get here, it may be that someone has added a new intrinsic
849 // to the list in vmSymbols.hpp without implementing it here.
850 #ifndef PRODUCT
851 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
852 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
853 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
854 }
855 #endif
856 return false;
857 }
858 }
859
860 Node* LibraryCallKit::try_to_predicate() {
861 if (!jvms()->has_method()) {
862 // Root JVMState has a null method.
863 assert(map()->memory()->Opcode() == Op_Parm, "");
864 // Insert the memory aliasing node
865 set_all_memory(reset_memory());
866 }
867 assert(merged_memory(), "");
868
869 switch (intrinsic_id()) {
870 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
871 return inline_cipherBlockChaining_AESCrypt_predicate(false);
872 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
873 return inline_cipherBlockChaining_AESCrypt_predicate(true);
874
875 default:
876 // If you get here, it may be that someone has added a new intrinsic
877 // to the list in vmSymbols.hpp without implementing it here.
878 #ifndef PRODUCT
879 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
880 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
881 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
882 }
883 #endif
884 Node* slow_ctl = control();
885 set_control(top()); // No fast path instrinsic
886 return slow_ctl;
887 }
888 }
889
890 //------------------------------set_result-------------------------------
891 // Helper function for finishing intrinsics.
892 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
893 record_for_igvn(region);
894 set_control(_gvn.transform(region));
895 set_result( _gvn.transform(value));
896 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
897 }
898
899 //------------------------------generate_guard---------------------------
900 // Helper function for generating guarded fast-slow graph structures.
901 // The given 'test', if true, guards a slow path. If the test fails
902 // then a fast path can be taken. (We generally hope it fails.)
903 // In all cases, GraphKit::control() is updated to the fast path.
904 // The returned value represents the control for the slow path.
905 // The return value is never 'top'; it is either a valid control
906 // or NULL if it is obvious that the slow path can never be taken.
907 // Also, if region and the slow control are not NULL, the slow edge
908 // is appended to the region.
909 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
910 if (stopped()) {
911 // Already short circuited.
912 return NULL;
913 }
914
915 // Build an if node and its projections.
916 // If test is true we take the slow path, which we assume is uncommon.
917 if (_gvn.type(test) == TypeInt::ZERO) {
918 // The slow branch is never taken. No need to build this guard.
919 return NULL;
920 }
921
922 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
923
924 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
925 if (if_slow == top()) {
926 // The slow branch is never taken. No need to build this guard.
927 return NULL;
928 }
929
930 if (region != NULL)
931 region->add_req(if_slow);
932
933 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
934 set_control(if_fast);
935
936 return if_slow;
937 }
938
939 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
940 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
941 }
942 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
943 return generate_guard(test, region, PROB_FAIR);
944 }
945
946 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
947 Node* *pos_index) {
948 if (stopped())
949 return NULL; // already stopped
950 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
951 return NULL; // index is already adequately typed
952 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
953 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
954 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
955 if (is_neg != NULL && pos_index != NULL) {
956 // Emulate effect of Parse::adjust_map_after_if.
957 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
958 ccast->set_req(0, control());
959 (*pos_index) = _gvn.transform(ccast);
960 }
961 return is_neg;
962 }
963
964 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
965 Node* *pos_index) {
966 if (stopped())
967 return NULL; // already stopped
968 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
969 return NULL; // index is already adequately typed
970 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
971 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
972 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
973 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
974 if (is_notp != NULL && pos_index != NULL) {
975 // Emulate effect of Parse::adjust_map_after_if.
976 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
977 ccast->set_req(0, control());
978 (*pos_index) = _gvn.transform(ccast);
979 }
980 return is_notp;
981 }
982
983 // Make sure that 'position' is a valid limit index, in [0..length].
984 // There are two equivalent plans for checking this:
985 // A. (offset + copyLength) unsigned<= arrayLength
986 // B. offset <= (arrayLength - copyLength)
987 // We require that all of the values above, except for the sum and
988 // difference, are already known to be non-negative.
989 // Plan A is robust in the face of overflow, if offset and copyLength
990 // are both hugely positive.
991 //
992 // Plan B is less direct and intuitive, but it does not overflow at
993 // all, since the difference of two non-negatives is always
994 // representable. Whenever Java methods must perform the equivalent
995 // check they generally use Plan B instead of Plan A.
996 // For the moment we use Plan A.
997 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
998 Node* subseq_length,
999 Node* array_length,
1000 RegionNode* region) {
1001 if (stopped())
1002 return NULL; // already stopped
1003 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1004 if (zero_offset && subseq_length->eqv_uncast(array_length))
1005 return NULL; // common case of whole-array copy
1006 Node* last = subseq_length;
1007 if (!zero_offset) // last += offset
1008 last = _gvn.transform(new (C) AddINode(last, offset));
1009 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1010 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1011 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1012 return is_over;
1013 }
1014
1015
1016 //--------------------------generate_current_thread--------------------
1017 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1018 ciKlass* thread_klass = env()->Thread_klass();
1019 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1020 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1021 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1022 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1023 tls_output = thread;
1024 return threadObj;
1025 }
1026
1027
1028 //------------------------------make_string_method_node------------------------
1029 // Helper method for String intrinsic functions. This version is called
1030 // with str1 and str2 pointing to String object nodes.
1031 //
1032 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1033 Node* no_ctrl = NULL;
1034
1035 // Get start addr of string
1036 Node* str1_value = load_String_value(no_ctrl, str1);
1037 Node* str1_offset = load_String_offset(no_ctrl, str1);
1038 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1039
1040 // Get length of string 1
1041 Node* str1_len = load_String_length(no_ctrl, str1);
1042
1043 Node* str2_value = load_String_value(no_ctrl, str2);
1044 Node* str2_offset = load_String_offset(no_ctrl, str2);
1045 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1046
1047 Node* str2_len = NULL;
1048 Node* result = NULL;
1049
1050 switch (opcode) {
1051 case Op_StrIndexOf:
1052 // Get length of string 2
1053 str2_len = load_String_length(no_ctrl, str2);
1054
1055 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1056 str1_start, str1_len, str2_start, str2_len);
1057 break;
1058 case Op_StrComp:
1059 // Get length of string 2
1060 str2_len = load_String_length(no_ctrl, str2);
1061
1062 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1063 str1_start, str1_len, str2_start, str2_len);
1064 break;
1065 case Op_StrEquals:
1066 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1067 str1_start, str2_start, str1_len);
1068 break;
1069 default:
1070 ShouldNotReachHere();
1071 return NULL;
1072 }
1073
1074 // All these intrinsics have checks.
1075 C->set_has_split_ifs(true); // Has chance for split-if optimization
1076
1077 return _gvn.transform(result);
1078 }
1079
1080 // Helper method for String intrinsic functions. This version is called
1081 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1082 // to Int nodes containing the lenghts of str1 and str2.
1083 //
1084 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1085 Node* result = NULL;
1086 switch (opcode) {
1087 case Op_StrIndexOf:
1088 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1089 str1_start, cnt1, str2_start, cnt2);
1090 break;
1091 case Op_StrComp:
1092 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1093 str1_start, cnt1, str2_start, cnt2);
1094 break;
1095 case Op_StrEquals:
1096 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1097 str1_start, str2_start, cnt1);
1098 break;
1099 default:
1100 ShouldNotReachHere();
1101 return NULL;
1102 }
1103
1104 // All these intrinsics have checks.
1105 C->set_has_split_ifs(true); // Has chance for split-if optimization
1106
1107 return _gvn.transform(result);
1108 }
1109
1110 //------------------------------inline_string_compareTo------------------------
1111 // public int java.lang.String.compareTo(String anotherString);
1112 bool LibraryCallKit::inline_string_compareTo() {
1113 Node* receiver = null_check(argument(0));
1114 Node* arg = null_check(argument(1));
1115 if (stopped()) {
1116 return true;
1117 }
1118 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1119 return true;
1120 }
1121
1122 //------------------------------inline_string_equals------------------------
1123 bool LibraryCallKit::inline_string_equals() {
1124 Node* receiver = null_check_receiver();
1125 // NOTE: Do not null check argument for String.equals() because spec
1126 // allows to specify NULL as argument.
1127 Node* argument = this->argument(1);
1128 if (stopped()) {
1129 return true;
1130 }
1131
1132 // paths (plus control) merge
1133 RegionNode* region = new (C) RegionNode(5);
1134 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1135
1136 // does source == target string?
1137 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1138 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1139
1140 Node* if_eq = generate_slow_guard(bol, NULL);
1141 if (if_eq != NULL) {
1142 // receiver == argument
1143 phi->init_req(2, intcon(1));
1144 region->init_req(2, if_eq);
1145 }
1146
1147 // get String klass for instanceOf
1148 ciInstanceKlass* klass = env()->String_klass();
1149
1150 if (!stopped()) {
1151 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1152 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1153 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1154
1155 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1156 //instanceOf == true, fallthrough
1157
1158 if (inst_false != NULL) {
1159 phi->init_req(3, intcon(0));
1160 region->init_req(3, inst_false);
1161 }
1162 }
1163
1164 if (!stopped()) {
1165 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1166
1167 // Properly cast the argument to String
1168 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1169 // This path is taken only when argument's type is String:NotNull.
1170 argument = cast_not_null(argument, false);
1171
1172 Node* no_ctrl = NULL;
1173
1174 // Get start addr of receiver
1175 Node* receiver_val = load_String_value(no_ctrl, receiver);
1176 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1177 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1178
1179 // Get length of receiver
1180 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1181
1182 // Get start addr of argument
1183 Node* argument_val = load_String_value(no_ctrl, argument);
1184 Node* argument_offset = load_String_offset(no_ctrl, argument);
1185 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1186
1187 // Get length of argument
1188 Node* argument_cnt = load_String_length(no_ctrl, argument);
1189
1190 // Check for receiver count != argument count
1191 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1192 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1193 Node* if_ne = generate_slow_guard(bol, NULL);
1194 if (if_ne != NULL) {
1195 phi->init_req(4, intcon(0));
1196 region->init_req(4, if_ne);
1197 }
1198
1199 // Check for count == 0 is done by assembler code for StrEquals.
1200
1201 if (!stopped()) {
1202 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1203 phi->init_req(1, equals);
1204 region->init_req(1, control());
1205 }
1206 }
1207
1208 // post merge
1209 set_control(_gvn.transform(region));
1210 record_for_igvn(region);
1211
1212 set_result(_gvn.transform(phi));
1213 return true;
1214 }
1215
1216 //------------------------------inline_array_equals----------------------------
1217 bool LibraryCallKit::inline_array_equals() {
1218 Node* arg1 = argument(0);
1219 Node* arg2 = argument(1);
1220 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1221 return true;
1222 }
1223
1224 // Java version of String.indexOf(constant string)
1225 // class StringDecl {
1226 // StringDecl(char[] ca) {
1227 // offset = 0;
1228 // count = ca.length;
1229 // value = ca;
1230 // }
1231 // int offset;
1232 // int count;
1233 // char[] value;
1234 // }
1235 //
1236 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1237 // int targetOffset, int cache_i, int md2) {
1238 // int cache = cache_i;
1239 // int sourceOffset = string_object.offset;
1240 // int sourceCount = string_object.count;
1241 // int targetCount = target_object.length;
1242 //
1243 // int targetCountLess1 = targetCount - 1;
1244 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1245 //
1246 // char[] source = string_object.value;
1247 // char[] target = target_object;
1248 // int lastChar = target[targetCountLess1];
1249 //
1250 // outer_loop:
1251 // for (int i = sourceOffset; i < sourceEnd; ) {
1252 // int src = source[i + targetCountLess1];
1253 // if (src == lastChar) {
1254 // // With random strings and a 4-character alphabet,
1255 // // reverse matching at this point sets up 0.8% fewer
1256 // // frames, but (paradoxically) makes 0.3% more probes.
1257 // // Since those probes are nearer the lastChar probe,
1258 // // there is may be a net D$ win with reverse matching.
1259 // // But, reversing loop inhibits unroll of inner loop
1260 // // for unknown reason. So, does running outer loop from
1261 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1262 // for (int j = 0; j < targetCountLess1; j++) {
1263 // if (target[targetOffset + j] != source[i+j]) {
1264 // if ((cache & (1 << source[i+j])) == 0) {
1265 // if (md2 < j+1) {
1266 // i += j+1;
1267 // continue outer_loop;
1268 // }
1269 // }
1270 // i += md2;
1271 // continue outer_loop;
1272 // }
1273 // }
1274 // return i - sourceOffset;
1275 // }
1276 // if ((cache & (1 << src)) == 0) {
1277 // i += targetCountLess1;
1278 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1279 // i++;
1280 // }
1281 // return -1;
1282 // }
1283
1284 //------------------------------string_indexOf------------------------
1285 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1286 jint cache_i, jint md2_i) {
1287
1288 Node* no_ctrl = NULL;
1289 float likely = PROB_LIKELY(0.9);
1290 float unlikely = PROB_UNLIKELY(0.9);
1291
1292 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1293
1294 Node* source = load_String_value(no_ctrl, string_object);
1295 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1296 Node* sourceCount = load_String_length(no_ctrl, string_object);
1297
1298 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1299 jint target_length = target_array->length();
1300 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1301 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1302
1303 // String.value field is known to be @Stable.
1304 if (UseImplicitStableValues) {
1305 target = cast_array_to_stable(target, target_type);
1306 }
1307
1308 IdealKit kit(this, false, true);
1309 #define __ kit.
1310 Node* zero = __ ConI(0);
1311 Node* one = __ ConI(1);
1312 Node* cache = __ ConI(cache_i);
1313 Node* md2 = __ ConI(md2_i);
1314 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1315 Node* targetCount = __ ConI(target_length);
1316 Node* targetCountLess1 = __ ConI(target_length - 1);
1317 Node* targetOffset = __ ConI(targetOffset_i);
1318 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1319
1320 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1321 Node* outer_loop = __ make_label(2 /* goto */);
1322 Node* return_ = __ make_label(1);
1323
1324 __ set(rtn,__ ConI(-1));
1325 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1326 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1327 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1328 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1329 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1330 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1331 Node* tpj = __ AddI(targetOffset, __ value(j));
1332 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1333 Node* ipj = __ AddI(__ value(i), __ value(j));
1334 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1335 __ if_then(targ, BoolTest::ne, src2); {
1336 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1337 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1338 __ increment(i, __ AddI(__ value(j), one));
1339 __ goto_(outer_loop);
1340 } __ end_if(); __ dead(j);
1341 }__ end_if(); __ dead(j);
1342 __ increment(i, md2);
1343 __ goto_(outer_loop);
1344 }__ end_if();
1345 __ increment(j, one);
1346 }__ end_loop(); __ dead(j);
1347 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1348 __ goto_(return_);
1349 }__ end_if();
1350 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1351 __ increment(i, targetCountLess1);
1352 }__ end_if();
1353 __ increment(i, one);
1354 __ bind(outer_loop);
1355 }__ end_loop(); __ dead(i);
1356 __ bind(return_);
1357
1358 // Final sync IdealKit and GraphKit.
1359 final_sync(kit);
1360 Node* result = __ value(rtn);
1361 #undef __
1362 C->set_has_loops(true);
1363 return result;
1364 }
1365
1366 //------------------------------inline_string_indexOf------------------------
1367 bool LibraryCallKit::inline_string_indexOf() {
1368 Node* receiver = argument(0);
1369 Node* arg = argument(1);
1370
1371 Node* result;
1372 // Disable the use of pcmpestri until it can be guaranteed that
1373 // the load doesn't cross into the uncommited space.
1374 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1375 UseSSE42Intrinsics) {
1376 // Generate SSE4.2 version of indexOf
1377 // We currently only have match rules that use SSE4.2
1378
1379 receiver = null_check(receiver);
1380 arg = null_check(arg);
1381 if (stopped()) {
1382 return true;
1383 }
1384
1385 ciInstanceKlass* str_klass = env()->String_klass();
1386 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1387
1388 // Make the merge point
1389 RegionNode* result_rgn = new (C) RegionNode(4);
1390 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1391 Node* no_ctrl = NULL;
1392
1393 // Get start addr of source string
1394 Node* source = load_String_value(no_ctrl, receiver);
1395 Node* source_offset = load_String_offset(no_ctrl, receiver);
1396 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1397
1398 // Get length of source string
1399 Node* source_cnt = load_String_length(no_ctrl, receiver);
1400
1401 // Get start addr of substring
1402 Node* substr = load_String_value(no_ctrl, arg);
1403 Node* substr_offset = load_String_offset(no_ctrl, arg);
1404 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1405
1406 // Get length of source string
1407 Node* substr_cnt = load_String_length(no_ctrl, arg);
1408
1409 // Check for substr count > string count
1410 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1411 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1412 Node* if_gt = generate_slow_guard(bol, NULL);
1413 if (if_gt != NULL) {
1414 result_phi->init_req(2, intcon(-1));
1415 result_rgn->init_req(2, if_gt);
1416 }
1417
1418 if (!stopped()) {
1419 // Check for substr count == 0
1420 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1421 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1422 Node* if_zero = generate_slow_guard(bol, NULL);
1423 if (if_zero != NULL) {
1424 result_phi->init_req(3, intcon(0));
1425 result_rgn->init_req(3, if_zero);
1426 }
1427 }
1428
1429 if (!stopped()) {
1430 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1431 result_phi->init_req(1, result);
1432 result_rgn->init_req(1, control());
1433 }
1434 set_control(_gvn.transform(result_rgn));
1435 record_for_igvn(result_rgn);
1436 result = _gvn.transform(result_phi);
1437
1438 } else { // Use LibraryCallKit::string_indexOf
1439 // don't intrinsify if argument isn't a constant string.
1440 if (!arg->is_Con()) {
1441 return false;
1442 }
1443 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1444 if (str_type == NULL) {
1445 return false;
1446 }
1447 ciInstanceKlass* klass = env()->String_klass();
1448 ciObject* str_const = str_type->const_oop();
1449 if (str_const == NULL || str_const->klass() != klass) {
1450 return false;
1451 }
1452 ciInstance* str = str_const->as_instance();
1453 assert(str != NULL, "must be instance");
1454
1455 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1456 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1457
1458 int o;
1459 int c;
1460 if (java_lang_String::has_offset_field()) {
1461 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1462 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1463 } else {
1464 o = 0;
1465 c = pat->length();
1466 }
1467
1468 // constant strings have no offset and count == length which
1469 // simplifies the resulting code somewhat so lets optimize for that.
1470 if (o != 0 || c != pat->length()) {
1471 return false;
1472 }
1473
1474 receiver = null_check(receiver, T_OBJECT);
1475 // NOTE: No null check on the argument is needed since it's a constant String oop.
1476 if (stopped()) {
1477 return true;
1478 }
1479
1480 // The null string as a pattern always returns 0 (match at beginning of string)
1481 if (c == 0) {
1482 set_result(intcon(0));
1483 return true;
1484 }
1485
1486 // Generate default indexOf
1487 jchar lastChar = pat->char_at(o + (c - 1));
1488 int cache = 0;
1489 int i;
1490 for (i = 0; i < c - 1; i++) {
1491 assert(i < pat->length(), "out of range");
1492 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1493 }
1494
1495 int md2 = c;
1496 for (i = 0; i < c - 1; i++) {
1497 assert(i < pat->length(), "out of range");
1498 if (pat->char_at(o + i) == lastChar) {
1499 md2 = (c - 1) - i;
1500 }
1501 }
1502
1503 result = string_indexOf(receiver, pat, o, cache, md2);
1504 }
1505 set_result(result);
1506 return true;
1507 }
1508
1509 //--------------------------round_double_node--------------------------------
1510 // Round a double node if necessary.
1511 Node* LibraryCallKit::round_double_node(Node* n) {
1512 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1513 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1514 return n;
1515 }
1516
1517 //------------------------------inline_math-----------------------------------
1518 // public static double Math.abs(double)
1519 // public static double Math.sqrt(double)
1520 // public static double Math.log(double)
1521 // public static double Math.log10(double)
1522 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1523 Node* arg = round_double_node(argument(0));
1524 Node* n;
1525 switch (id) {
1526 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1527 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1528 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1529 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1530 default: fatal_unexpected_iid(id); break;
1531 }
1532 set_result(_gvn.transform(n));
1533 return true;
1534 }
1535
1536 //------------------------------inline_trig----------------------------------
1537 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1538 // argument reduction which will turn into a fast/slow diamond.
1539 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1540 Node* arg = round_double_node(argument(0));
1541 Node* n = NULL;
1542
1543 switch (id) {
1544 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1545 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1546 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1547 default: fatal_unexpected_iid(id); break;
1548 }
1549 n = _gvn.transform(n);
1550
1551 // Rounding required? Check for argument reduction!
1552 if (Matcher::strict_fp_requires_explicit_rounding) {
1553 static const double pi_4 = 0.7853981633974483;
1554 static const double neg_pi_4 = -0.7853981633974483;
1555 // pi/2 in 80-bit extended precision
1556 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1557 // -pi/2 in 80-bit extended precision
1558 // 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};
1559 // Cutoff value for using this argument reduction technique
1560 //static const double pi_2_minus_epsilon = 1.564660403643354;
1561 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1562
1563 // Pseudocode for sin:
1564 // if (x <= Math.PI / 4.0) {
1565 // if (x >= -Math.PI / 4.0) return fsin(x);
1566 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1567 // } else {
1568 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1569 // }
1570 // return StrictMath.sin(x);
1571
1572 // Pseudocode for cos:
1573 // if (x <= Math.PI / 4.0) {
1574 // if (x >= -Math.PI / 4.0) return fcos(x);
1575 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1576 // } else {
1577 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1578 // }
1579 // return StrictMath.cos(x);
1580
1581 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1582 // requires a special machine instruction to load it. Instead we'll try
1583 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1584 // probably do the math inside the SIN encoding.
1585
1586 // Make the merge point
1587 RegionNode* r = new (C) RegionNode(3);
1588 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1589
1590 // Flatten arg so we need only 1 test
1591 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1592 // Node for PI/4 constant
1593 Node *pi4 = makecon(TypeD::make(pi_4));
1594 // Check PI/4 : abs(arg)
1595 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1596 // Check: If PI/4 < abs(arg) then go slow
1597 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1598 // Branch either way
1599 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1600 set_control(opt_iff(r,iff));
1601
1602 // Set fast path result
1603 phi->init_req(2, n);
1604
1605 // Slow path - non-blocking leaf call
1606 Node* call = NULL;
1607 switch (id) {
1608 case vmIntrinsics::_dsin:
1609 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1610 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1611 "Sin", NULL, arg, top());
1612 break;
1613 case vmIntrinsics::_dcos:
1614 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1615 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1616 "Cos", NULL, arg, top());
1617 break;
1618 case vmIntrinsics::_dtan:
1619 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1620 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1621 "Tan", NULL, arg, top());
1622 break;
1623 }
1624 assert(control()->in(0) == call, "");
1625 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1626 r->init_req(1, control());
1627 phi->init_req(1, slow_result);
1628
1629 // Post-merge
1630 set_control(_gvn.transform(r));
1631 record_for_igvn(r);
1632 n = _gvn.transform(phi);
1633
1634 C->set_has_split_ifs(true); // Has chance for split-if optimization
1635 }
1636 set_result(n);
1637 return true;
1638 }
1639
1640 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1641 //-------------------
1642 //result=(result.isNaN())? funcAddr():result;
1643 // Check: If isNaN() by checking result!=result? then either trap
1644 // or go to runtime
1645 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1646 // Build the boolean node
1647 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1648
1649 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1650 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1651 // The pow or exp intrinsic returned a NaN, which requires a call
1652 // to the runtime. Recompile with the runtime call.
1653 uncommon_trap(Deoptimization::Reason_intrinsic,
1654 Deoptimization::Action_make_not_entrant);
1655 }
1656 set_result(result);
1657 } else {
1658 // If this inlining ever returned NaN in the past, we compile a call
1659 // to the runtime to properly handle corner cases
1660
1661 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1662 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1663 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1664
1665 if (!if_slow->is_top()) {
1666 RegionNode* result_region = new (C) RegionNode(3);
1667 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1668
1669 result_region->init_req(1, if_fast);
1670 result_val->init_req(1, result);
1671
1672 set_control(if_slow);
1673
1674 const TypePtr* no_memory_effects = NULL;
1675 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1676 no_memory_effects,
1677 x, top(), y, y ? top() : NULL);
1678 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1679 #ifdef ASSERT
1680 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1681 assert(value_top == top(), "second value must be top");
1682 #endif
1683
1684 result_region->init_req(2, control());
1685 result_val->init_req(2, value);
1686 set_result(result_region, result_val);
1687 } else {
1688 set_result(result);
1689 }
1690 }
1691 }
1692
1693 //------------------------------inline_exp-------------------------------------
1694 // Inline exp instructions, if possible. The Intel hardware only misses
1695 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1696 bool LibraryCallKit::inline_exp() {
1697 Node* arg = round_double_node(argument(0));
1698 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1699
1700 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1701
1702 C->set_has_split_ifs(true); // Has chance for split-if optimization
1703 return true;
1704 }
1705
1706 //------------------------------inline_pow-------------------------------------
1707 // Inline power instructions, if possible.
1708 bool LibraryCallKit::inline_pow() {
1709 // Pseudocode for pow
1710 // if (x <= 0.0) {
1711 // long longy = (long)y;
1712 // if ((double)longy == y) { // if y is long
1713 // if (y + 1 == y) longy = 0; // huge number: even
1714 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1715 // } else {
1716 // result = NaN;
1717 // }
1718 // } else {
1719 // result = DPow(x,y);
1720 // }
1721 // if (result != result)? {
1722 // result = uncommon_trap() or runtime_call();
1723 // }
1724 // return result;
1725
1726 Node* x = round_double_node(argument(0));
1727 Node* y = round_double_node(argument(2));
1728
1729 Node* result = NULL;
1730
1731 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1732 // Short form: skip the fancy tests and just check for NaN result.
1733 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1734 } else {
1735 // If this inlining ever returned NaN in the past, include all
1736 // checks + call to the runtime.
1737
1738 // Set the merge point for If node with condition of (x <= 0.0)
1739 // There are four possible paths to region node and phi node
1740 RegionNode *r = new (C) RegionNode(4);
1741 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1742
1743 // Build the first if node: if (x <= 0.0)
1744 // Node for 0 constant
1745 Node *zeronode = makecon(TypeD::ZERO);
1746 // Check x:0
1747 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1748 // Check: If (x<=0) then go complex path
1749 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1750 // Branch either way
1751 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1752 // Fast path taken; set region slot 3
1753 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1754 r->init_req(3,fast_taken); // Capture fast-control
1755
1756 // Fast path not-taken, i.e. slow path
1757 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1758
1759 // Set fast path result
1760 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1761 phi->init_req(3, fast_result);
1762
1763 // Complex path
1764 // Build the second if node (if y is long)
1765 // Node for (long)y
1766 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1767 // Node for (double)((long) y)
1768 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1769 // Check (double)((long) y) : y
1770 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1771 // Check if (y isn't long) then go to slow path
1772
1773 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1774 // Branch either way
1775 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1776 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1777
1778 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1779
1780 // Calculate DPow(abs(x), y)*(1 & (long)y)
1781 // Node for constant 1
1782 Node *conone = longcon(1);
1783 // 1& (long)y
1784 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1785
1786 // A huge number is always even. Detect a huge number by checking
1787 // if y + 1 == y and set integer to be tested for parity to 0.
1788 // Required for corner case:
1789 // (long)9.223372036854776E18 = max_jlong
1790 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1791 // max_jlong is odd but 9.223372036854776E18 is even
1792 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1793 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1794 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1795 Node* correctedsign = NULL;
1796 if (ConditionalMoveLimit != 0) {
1797 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1798 } else {
1799 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1800 RegionNode *r = new (C) RegionNode(3);
1801 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1802 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1803 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1804 phi->init_req(1, signnode);
1805 phi->init_req(2, longcon(0));
1806 correctedsign = _gvn.transform(phi);
1807 ylong_path = _gvn.transform(r);
1808 record_for_igvn(r);
1809 }
1810
1811 // zero node
1812 Node *conzero = longcon(0);
1813 // Check (1&(long)y)==0?
1814 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1815 // Check if (1&(long)y)!=0?, if so the result is negative
1816 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1817 // abs(x)
1818 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1819 // abs(x)^y
1820 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1821 // -abs(x)^y
1822 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1823 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1824 Node *signresult = NULL;
1825 if (ConditionalMoveLimit != 0) {
1826 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1827 } else {
1828 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1829 RegionNode *r = new (C) RegionNode(3);
1830 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1831 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1832 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1833 phi->init_req(1, absxpowy);
1834 phi->init_req(2, negabsxpowy);
1835 signresult = _gvn.transform(phi);
1836 ylong_path = _gvn.transform(r);
1837 record_for_igvn(r);
1838 }
1839 // Set complex path fast result
1840 r->init_req(2, ylong_path);
1841 phi->init_req(2, signresult);
1842
1843 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1844 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1845 r->init_req(1,slow_path);
1846 phi->init_req(1,slow_result);
1847
1848 // Post merge
1849 set_control(_gvn.transform(r));
1850 record_for_igvn(r);
1851 result = _gvn.transform(phi);
1852 }
1853
1854 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1855
1856 C->set_has_split_ifs(true); // Has chance for split-if optimization
1857 return true;
1858 }
1859
1860 //------------------------------runtime_math-----------------------------
1861 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1862 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1863 "must be (DD)D or (D)D type");
1864
1865 // Inputs
1866 Node* a = round_double_node(argument(0));
1867 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1868
1869 const TypePtr* no_memory_effects = NULL;
1870 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1871 no_memory_effects,
1872 a, top(), b, b ? top() : NULL);
1873 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1874 #ifdef ASSERT
1875 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1876 assert(value_top == top(), "second value must be top");
1877 #endif
1878
1879 set_result(value);
1880 return true;
1881 }
1882
1883 //------------------------------inline_math_native-----------------------------
1884 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1885 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1886 switch (id) {
1887 // These intrinsics are not properly supported on all hardware
1888 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1889 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1890 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1891 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1892 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1893 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1894
1895 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1896 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1897 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1898 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1899
1900 // These intrinsics are supported on all hardware
1901 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1902 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1903
1904 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1905 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1906 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1907 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1908 #undef FN_PTR
1909
1910 // These intrinsics are not yet correctly implemented
1911 case vmIntrinsics::_datan2:
1912 return false;
1913
1914 default:
1915 fatal_unexpected_iid(id);
1916 return false;
1917 }
1918 }
1919
1920 static bool is_simple_name(Node* n) {
1921 return (n->req() == 1 // constant
1922 || (n->is_Type() && n->as_Type()->type()->singleton())
1923 || n->is_Proj() // parameter or return value
1924 || n->is_Phi() // local of some sort
1925 );
1926 }
1927
1928 //----------------------------inline_min_max-----------------------------------
1929 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1930 set_result(generate_min_max(id, argument(0), argument(1)));
1931 return true;
1932 }
1933
1934 bool LibraryCallKit::inline_math_mathExact(Node* math) {
1935 Node* result = _gvn.transform( new(C) ProjNode(math, MathExactNode::result_proj_node));
1936 Node* flags = _gvn.transform( new(C) FlagsProjNode(math, MathExactNode::flags_proj_node));
1937
1938 Node* bol = _gvn.transform( new (C) BoolNode(flags, BoolTest::overflow) );
1939 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1940 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1941 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1942
1943 {
1944 PreserveJVMState pjvms(this);
1945 PreserveReexecuteState preexecs(this);
1946 jvms()->set_should_reexecute(true);
1947
1948 set_control(slow_path);
1949 set_i_o(i_o());
1950
1951 uncommon_trap(Deoptimization::Reason_intrinsic,
1952 Deoptimization::Action_none);
1953 }
1954
1955 set_control(fast_path);
1956 set_result(result);
1957 return true;
1958 }
1959
1960 bool LibraryCallKit::inline_math_addExact() {
1961 Node* arg1 = argument(0);
1962 Node* arg2 = argument(1);
1963
1964 Node* add = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2) );
1965 if (add->Opcode() == Op_AddExactI) {
1966 return inline_math_mathExact(add);
1967 } else {
1968 set_result(add);
1969 }
1970 return true;
1971 }
1972
1973 Node*
1974 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1975 // These are the candidate return value:
1976 Node* xvalue = x0;
1977 Node* yvalue = y0;
1978
1979 if (xvalue == yvalue) {
1980 return xvalue;
1981 }
1982
1983 bool want_max = (id == vmIntrinsics::_max);
1984
1985 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1986 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1987 if (txvalue == NULL || tyvalue == NULL) return top();
1988 // This is not really necessary, but it is consistent with a
1989 // hypothetical MaxINode::Value method:
1990 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1991
1992 // %%% This folding logic should (ideally) be in a different place.
1993 // Some should be inside IfNode, and there to be a more reliable
1994 // transformation of ?: style patterns into cmoves. We also want
1995 // more powerful optimizations around cmove and min/max.
1996
1997 // Try to find a dominating comparison of these guys.
1998 // It can simplify the index computation for Arrays.copyOf
1999 // and similar uses of System.arraycopy.
2000 // First, compute the normalized version of CmpI(x, y).
2001 int cmp_op = Op_CmpI;
2002 Node* xkey = xvalue;
2003 Node* ykey = yvalue;
2004 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2005 if (ideal_cmpxy->is_Cmp()) {
2006 // E.g., if we have CmpI(length - offset, count),
2007 // it might idealize to CmpI(length, count + offset)
2008 cmp_op = ideal_cmpxy->Opcode();
2009 xkey = ideal_cmpxy->in(1);
2010 ykey = ideal_cmpxy->in(2);
2011 }
2012
2013 // Start by locating any relevant comparisons.
2014 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2015 Node* cmpxy = NULL;
2016 Node* cmpyx = NULL;
2017 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2018 Node* cmp = start_from->fast_out(k);
2019 if (cmp->outcnt() > 0 && // must have prior uses
2020 cmp->in(0) == NULL && // must be context-independent
2021 cmp->Opcode() == cmp_op) { // right kind of compare
2022 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2023 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2024 }
2025 }
2026
2027 const int NCMPS = 2;
2028 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2029 int cmpn;
2030 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2031 if (cmps[cmpn] != NULL) break; // find a result
2032 }
2033 if (cmpn < NCMPS) {
2034 // Look for a dominating test that tells us the min and max.
2035 int depth = 0; // Limit search depth for speed
2036 Node* dom = control();
2037 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2038 if (++depth >= 100) break;
2039 Node* ifproj = dom;
2040 if (!ifproj->is_Proj()) continue;
2041 Node* iff = ifproj->in(0);
2042 if (!iff->is_If()) continue;
2043 Node* bol = iff->in(1);
2044 if (!bol->is_Bool()) continue;
2045 Node* cmp = bol->in(1);
2046 if (cmp == NULL) continue;
2047 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2048 if (cmps[cmpn] == cmp) break;
2049 if (cmpn == NCMPS) continue;
2050 BoolTest::mask btest = bol->as_Bool()->_test._test;
2051 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2052 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2053 // At this point, we know that 'x btest y' is true.
2054 switch (btest) {
2055 case BoolTest::eq:
2056 // They are proven equal, so we can collapse the min/max.
2057 // Either value is the answer. Choose the simpler.
2058 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2059 return yvalue;
2060 return xvalue;
2061 case BoolTest::lt: // x < y
2062 case BoolTest::le: // x <= y
2063 return (want_max ? yvalue : xvalue);
2064 case BoolTest::gt: // x > y
2065 case BoolTest::ge: // x >= y
2066 return (want_max ? xvalue : yvalue);
2067 }
2068 }
2069 }
2070
2071 // We failed to find a dominating test.
2072 // Let's pick a test that might GVN with prior tests.
2073 Node* best_bol = NULL;
2074 BoolTest::mask best_btest = BoolTest::illegal;
2075 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2076 Node* cmp = cmps[cmpn];
2077 if (cmp == NULL) continue;
2078 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2079 Node* bol = cmp->fast_out(j);
2080 if (!bol->is_Bool()) continue;
2081 BoolTest::mask btest = bol->as_Bool()->_test._test;
2082 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2083 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2084 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2085 best_bol = bol->as_Bool();
2086 best_btest = btest;
2087 }
2088 }
2089 }
2090
2091 Node* answer_if_true = NULL;
2092 Node* answer_if_false = NULL;
2093 switch (best_btest) {
2094 default:
2095 if (cmpxy == NULL)
2096 cmpxy = ideal_cmpxy;
2097 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2098 // and fall through:
2099 case BoolTest::lt: // x < y
2100 case BoolTest::le: // x <= y
2101 answer_if_true = (want_max ? yvalue : xvalue);
2102 answer_if_false = (want_max ? xvalue : yvalue);
2103 break;
2104 case BoolTest::gt: // x > y
2105 case BoolTest::ge: // x >= y
2106 answer_if_true = (want_max ? xvalue : yvalue);
2107 answer_if_false = (want_max ? yvalue : xvalue);
2108 break;
2109 }
2110
2111 jint hi, lo;
2112 if (want_max) {
2113 // We can sharpen the minimum.
2114 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2115 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2116 } else {
2117 // We can sharpen the maximum.
2118 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2119 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2120 }
2121
2122 // Use a flow-free graph structure, to avoid creating excess control edges
2123 // which could hinder other optimizations.
2124 // Since Math.min/max is often used with arraycopy, we want
2125 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2126 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2127 answer_if_false, answer_if_true,
2128 TypeInt::make(lo, hi, widen));
2129
2130 return _gvn.transform(cmov);
2131
2132 /*
2133 // This is not as desirable as it may seem, since Min and Max
2134 // nodes do not have a full set of optimizations.
2135 // And they would interfere, anyway, with 'if' optimizations
2136 // and with CMoveI canonical forms.
2137 switch (id) {
2138 case vmIntrinsics::_min:
2139 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2140 case vmIntrinsics::_max:
2141 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2142 default:
2143 ShouldNotReachHere();
2144 }
2145 */
2146 }
2147
2148 inline int
2149 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2150 const TypePtr* base_type = TypePtr::NULL_PTR;
2151 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2152 if (base_type == NULL) {
2153 // Unknown type.
2154 return Type::AnyPtr;
2155 } else if (base_type == TypePtr::NULL_PTR) {
2156 // Since this is a NULL+long form, we have to switch to a rawptr.
2157 base = _gvn.transform(new (C) CastX2PNode(offset));
2158 offset = MakeConX(0);
2159 return Type::RawPtr;
2160 } else if (base_type->base() == Type::RawPtr) {
2161 return Type::RawPtr;
2162 } else if (base_type->isa_oopptr()) {
2163 // Base is never null => always a heap address.
2164 if (base_type->ptr() == TypePtr::NotNull) {
2165 return Type::OopPtr;
2166 }
2167 // Offset is small => always a heap address.
2168 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2169 if (offset_type != NULL &&
2170 base_type->offset() == 0 && // (should always be?)
2171 offset_type->_lo >= 0 &&
2172 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2173 return Type::OopPtr;
2174 }
2175 // Otherwise, it might either be oop+off or NULL+addr.
2176 return Type::AnyPtr;
2177 } else {
2178 // No information:
2179 return Type::AnyPtr;
2180 }
2181 }
2182
2183 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2184 int kind = classify_unsafe_addr(base, offset);
2185 if (kind == Type::RawPtr) {
2186 return basic_plus_adr(top(), base, offset);
2187 } else {
2188 return basic_plus_adr(base, offset);
2189 }
2190 }
2191
2192 //--------------------------inline_number_methods-----------------------------
2193 // inline int Integer.numberOfLeadingZeros(int)
2194 // inline int Long.numberOfLeadingZeros(long)
2195 //
2196 // inline int Integer.numberOfTrailingZeros(int)
2197 // inline int Long.numberOfTrailingZeros(long)
2198 //
2199 // inline int Integer.bitCount(int)
2200 // inline int Long.bitCount(long)
2201 //
2202 // inline char Character.reverseBytes(char)
2203 // inline short Short.reverseBytes(short)
2204 // inline int Integer.reverseBytes(int)
2205 // inline long Long.reverseBytes(long)
2206 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2207 Node* arg = argument(0);
2208 Node* n;
2209 switch (id) {
2210 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2211 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2212 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2213 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2214 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2215 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2216 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2217 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2218 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2219 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2220 default: fatal_unexpected_iid(id); break;
2221 }
2222 set_result(_gvn.transform(n));
2223 return true;
2224 }
2225
2226 //----------------------------inline_unsafe_access----------------------------
2227
2228 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2229
2230 // Helper that guards and inserts a pre-barrier.
2231 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2232 Node* pre_val, bool need_mem_bar) {
2233 // We could be accessing the referent field of a reference object. If so, when G1
2234 // is enabled, we need to log the value in the referent field in an SATB buffer.
2235 // This routine performs some compile time filters and generates suitable
2236 // runtime filters that guard the pre-barrier code.
2237 // Also add memory barrier for non volatile load from the referent field
2238 // to prevent commoning of loads across safepoint.
2239 if (!UseG1GC && !need_mem_bar)
2240 return;
2241
2242 // Some compile time checks.
2243
2244 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2245 const TypeX* otype = offset->find_intptr_t_type();
2246 if (otype != NULL && otype->is_con() &&
2247 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2248 // Constant offset but not the reference_offset so just return
2249 return;
2250 }
2251
2252 // We only need to generate the runtime guards for instances.
2253 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2254 if (btype != NULL) {
2255 if (btype->isa_aryptr()) {
2256 // Array type so nothing to do
2257 return;
2258 }
2259
2260 const TypeInstPtr* itype = btype->isa_instptr();
2261 if (itype != NULL) {
2262 // Can the klass of base_oop be statically determined to be
2263 // _not_ a sub-class of Reference and _not_ Object?
2264 ciKlass* klass = itype->klass();
2265 if ( klass->is_loaded() &&
2266 !klass->is_subtype_of(env()->Reference_klass()) &&
2267 !env()->Object_klass()->is_subtype_of(klass)) {
2268 return;
2269 }
2270 }
2271 }
2272
2273 // The compile time filters did not reject base_oop/offset so
2274 // we need to generate the following runtime filters
2275 //
2276 // if (offset == java_lang_ref_Reference::_reference_offset) {
2277 // if (instance_of(base, java.lang.ref.Reference)) {
2278 // pre_barrier(_, pre_val, ...);
2279 // }
2280 // }
2281
2282 float likely = PROB_LIKELY( 0.999);
2283 float unlikely = PROB_UNLIKELY(0.999);
2284
2285 IdealKit ideal(this);
2286 #define __ ideal.
2287
2288 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2289
2290 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2291 // Update graphKit memory and control from IdealKit.
2292 sync_kit(ideal);
2293
2294 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2295 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2296
2297 // Update IdealKit memory and control from graphKit.
2298 __ sync_kit(this);
2299
2300 Node* one = __ ConI(1);
2301 // is_instof == 0 if base_oop == NULL
2302 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2303
2304 // Update graphKit from IdeakKit.
2305 sync_kit(ideal);
2306
2307 // Use the pre-barrier to record the value in the referent field
2308 pre_barrier(false /* do_load */,
2309 __ ctrl(),
2310 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2311 pre_val /* pre_val */,
2312 T_OBJECT);
2313 if (need_mem_bar) {
2314 // Add memory barrier to prevent commoning reads from this field
2315 // across safepoint since GC can change its value.
2316 insert_mem_bar(Op_MemBarCPUOrder);
2317 }
2318 // Update IdealKit from graphKit.
2319 __ sync_kit(this);
2320
2321 } __ end_if(); // _ref_type != ref_none
2322 } __ end_if(); // offset == referent_offset
2323
2324 // Final sync IdealKit and GraphKit.
2325 final_sync(ideal);
2326 #undef __
2327 }
2328
2329
2330 // Interpret Unsafe.fieldOffset cookies correctly:
2331 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2332
2333 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2334 // Attempt to infer a sharper value type from the offset and base type.
2335 ciKlass* sharpened_klass = NULL;
2336
2337 // See if it is an instance field, with an object type.
2338 if (alias_type->field() != NULL) {
2339 assert(!is_native_ptr, "native pointer op cannot use a java address");
2340 if (alias_type->field()->type()->is_klass()) {
2341 sharpened_klass = alias_type->field()->type()->as_klass();
2342 }
2343 }
2344
2345 // See if it is a narrow oop array.
2346 if (adr_type->isa_aryptr()) {
2347 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2348 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2349 if (elem_type != NULL) {
2350 sharpened_klass = elem_type->klass();
2351 }
2352 }
2353 }
2354
2355 // The sharpened class might be unloaded if there is no class loader
2356 // contraint in place.
2357 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2358 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2359
2360 #ifndef PRODUCT
2361 if (C->print_intrinsics() || C->print_inlining()) {
2362 tty->print(" from base type: "); adr_type->dump();
2363 tty->print(" sharpened value: "); tjp->dump();
2364 }
2365 #endif
2366 // Sharpen the value type.
2367 return tjp;
2368 }
2369 return NULL;
2370 }
2371
2372 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2373 if (callee()->is_static()) return false; // caller must have the capability!
2374
2375 #ifndef PRODUCT
2376 {
2377 ResourceMark rm;
2378 // Check the signatures.
2379 ciSignature* sig = callee()->signature();
2380 #ifdef ASSERT
2381 if (!is_store) {
2382 // Object getObject(Object base, int/long offset), etc.
2383 BasicType rtype = sig->return_type()->basic_type();
2384 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2385 rtype = T_ADDRESS; // it is really a C void*
2386 assert(rtype == type, "getter must return the expected value");
2387 if (!is_native_ptr) {
2388 assert(sig->count() == 2, "oop getter has 2 arguments");
2389 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2390 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2391 } else {
2392 assert(sig->count() == 1, "native getter has 1 argument");
2393 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2394 }
2395 } else {
2396 // void putObject(Object base, int/long offset, Object x), etc.
2397 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2398 if (!is_native_ptr) {
2399 assert(sig->count() == 3, "oop putter has 3 arguments");
2400 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2401 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2402 } else {
2403 assert(sig->count() == 2, "native putter has 2 arguments");
2404 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2405 }
2406 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2407 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2408 vtype = T_ADDRESS; // it is really a C void*
2409 assert(vtype == type, "putter must accept the expected value");
2410 }
2411 #endif // ASSERT
2412 }
2413 #endif //PRODUCT
2414
2415 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2416
2417 Node* receiver = argument(0); // type: oop
2418
2419 // Build address expression. See the code in inline_unsafe_prefetch.
2420 Node* adr;
2421 Node* heap_base_oop = top();
2422 Node* offset = top();
2423 Node* val;
2424
2425 if (!is_native_ptr) {
2426 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2427 Node* base = argument(1); // type: oop
2428 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2429 offset = argument(2); // type: long
2430 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2431 // to be plain byte offsets, which are also the same as those accepted
2432 // by oopDesc::field_base.
2433 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2434 "fieldOffset must be byte-scaled");
2435 // 32-bit machines ignore the high half!
2436 offset = ConvL2X(offset);
2437 adr = make_unsafe_address(base, offset);
2438 heap_base_oop = base;
2439 val = is_store ? argument(4) : NULL;
2440 } else {
2441 Node* ptr = argument(1); // type: long
2442 ptr = ConvL2X(ptr); // adjust Java long to machine word
2443 adr = make_unsafe_address(NULL, ptr);
2444 val = is_store ? argument(3) : NULL;
2445 }
2446
2447 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2448
2449 // First guess at the value type.
2450 const Type *value_type = Type::get_const_basic_type(type);
2451
2452 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2453 // there was not enough information to nail it down.
2454 Compile::AliasType* alias_type = C->alias_type(adr_type);
2455 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2456
2457 // We will need memory barriers unless we can determine a unique
2458 // alias category for this reference. (Note: If for some reason
2459 // the barriers get omitted and the unsafe reference begins to "pollute"
2460 // the alias analysis of the rest of the graph, either Compile::can_alias
2461 // or Compile::must_alias will throw a diagnostic assert.)
2462 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2463
2464 // If we are reading the value of the referent field of a Reference
2465 // object (either by using Unsafe directly or through reflection)
2466 // then, if G1 is enabled, we need to record the referent in an
2467 // SATB log buffer using the pre-barrier mechanism.
2468 // Also we need to add memory barrier to prevent commoning reads
2469 // from this field across safepoint since GC can change its value.
2470 bool need_read_barrier = !is_native_ptr && !is_store &&
2471 offset != top() && heap_base_oop != top();
2472
2473 if (!is_store && type == T_OBJECT) {
2474 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2475 if (tjp != NULL) {
2476 value_type = tjp;
2477 }
2478 }
2479
2480 receiver = null_check(receiver);
2481 if (stopped()) {
2482 return true;
2483 }
2484 // Heap pointers get a null-check from the interpreter,
2485 // as a courtesy. However, this is not guaranteed by Unsafe,
2486 // and it is not possible to fully distinguish unintended nulls
2487 // from intended ones in this API.
2488
2489 if (is_volatile) {
2490 // We need to emit leading and trailing CPU membars (see below) in
2491 // addition to memory membars when is_volatile. This is a little
2492 // too strong, but avoids the need to insert per-alias-type
2493 // volatile membars (for stores; compare Parse::do_put_xxx), which
2494 // we cannot do effectively here because we probably only have a
2495 // rough approximation of type.
2496 need_mem_bar = true;
2497 // For Stores, place a memory ordering barrier now.
2498 if (is_store)
2499 insert_mem_bar(Op_MemBarRelease);
2500 }
2501
2502 // Memory barrier to prevent normal and 'unsafe' accesses from
2503 // bypassing each other. Happens after null checks, so the
2504 // exception paths do not take memory state from the memory barrier,
2505 // so there's no problems making a strong assert about mixing users
2506 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2507 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2508 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2509
2510 if (!is_store) {
2511 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2512 // load value
2513 switch (type) {
2514 case T_BOOLEAN:
2515 case T_CHAR:
2516 case T_BYTE:
2517 case T_SHORT:
2518 case T_INT:
2519 case T_LONG:
2520 case T_FLOAT:
2521 case T_DOUBLE:
2522 break;
2523 case T_OBJECT:
2524 if (need_read_barrier) {
2525 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2526 }
2527 break;
2528 case T_ADDRESS:
2529 // Cast to an int type.
2530 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2531 p = ConvX2L(p);
2532 break;
2533 default:
2534 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2535 break;
2536 }
2537 // The load node has the control of the preceding MemBarCPUOrder. All
2538 // following nodes will have the control of the MemBarCPUOrder inserted at
2539 // the end of this method. So, pushing the load onto the stack at a later
2540 // point is fine.
2541 set_result(p);
2542 } else {
2543 // place effect of store into memory
2544 switch (type) {
2545 case T_DOUBLE:
2546 val = dstore_rounding(val);
2547 break;
2548 case T_ADDRESS:
2549 // Repackage the long as a pointer.
2550 val = ConvL2X(val);
2551 val = _gvn.transform(new (C) CastX2PNode(val));
2552 break;
2553 }
2554
2555 if (type != T_OBJECT ) {
2556 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2557 } else {
2558 // Possibly an oop being stored to Java heap or native memory
2559 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2560 // oop to Java heap.
2561 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2562 } else {
2563 // We can't tell at compile time if we are storing in the Java heap or outside
2564 // of it. So we need to emit code to conditionally do the proper type of
2565 // store.
2566
2567 IdealKit ideal(this);
2568 #define __ ideal.
2569 // QQQ who knows what probability is here??
2570 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2571 // Sync IdealKit and graphKit.
2572 sync_kit(ideal);
2573 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2574 // Update IdealKit memory.
2575 __ sync_kit(this);
2576 } __ else_(); {
2577 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2578 } __ end_if();
2579 // Final sync IdealKit and GraphKit.
2580 final_sync(ideal);
2581 #undef __
2582 }
2583 }
2584 }
2585
2586 if (is_volatile) {
2587 if (!is_store)
2588 insert_mem_bar(Op_MemBarAcquire);
2589 else
2590 insert_mem_bar(Op_MemBarVolatile);
2591 }
2592
2593 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2594
2595 return true;
2596 }
2597
2598 //----------------------------inline_unsafe_prefetch----------------------------
2599
2600 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2601 #ifndef PRODUCT
2602 {
2603 ResourceMark rm;
2604 // Check the signatures.
2605 ciSignature* sig = callee()->signature();
2606 #ifdef ASSERT
2607 // Object getObject(Object base, int/long offset), etc.
2608 BasicType rtype = sig->return_type()->basic_type();
2609 if (!is_native_ptr) {
2610 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2611 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2612 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2613 } else {
2614 assert(sig->count() == 1, "native prefetch has 1 argument");
2615 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2616 }
2617 #endif // ASSERT
2618 }
2619 #endif // !PRODUCT
2620
2621 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2622
2623 const int idx = is_static ? 0 : 1;
2624 if (!is_static) {
2625 null_check_receiver();
2626 if (stopped()) {
2627 return true;
2628 }
2629 }
2630
2631 // Build address expression. See the code in inline_unsafe_access.
2632 Node *adr;
2633 if (!is_native_ptr) {
2634 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2635 Node* base = argument(idx + 0); // type: oop
2636 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2637 Node* offset = argument(idx + 1); // type: long
2638 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2639 // to be plain byte offsets, which are also the same as those accepted
2640 // by oopDesc::field_base.
2641 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2642 "fieldOffset must be byte-scaled");
2643 // 32-bit machines ignore the high half!
2644 offset = ConvL2X(offset);
2645 adr = make_unsafe_address(base, offset);
2646 } else {
2647 Node* ptr = argument(idx + 0); // type: long
2648 ptr = ConvL2X(ptr); // adjust Java long to machine word
2649 adr = make_unsafe_address(NULL, ptr);
2650 }
2651
2652 // Generate the read or write prefetch
2653 Node *prefetch;
2654 if (is_store) {
2655 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2656 } else {
2657 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2658 }
2659 prefetch->init_req(0, control());
2660 set_i_o(_gvn.transform(prefetch));
2661
2662 return true;
2663 }
2664
2665 //----------------------------inline_unsafe_load_store----------------------------
2666 // This method serves a couple of different customers (depending on LoadStoreKind):
2667 //
2668 // LS_cmpxchg:
2669 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2670 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2671 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2672 //
2673 // LS_xadd:
2674 // public int getAndAddInt( Object o, long offset, int delta)
2675 // public long getAndAddLong(Object o, long offset, long delta)
2676 //
2677 // LS_xchg:
2678 // int getAndSet(Object o, long offset, int newValue)
2679 // long getAndSet(Object o, long offset, long newValue)
2680 // Object getAndSet(Object o, long offset, Object newValue)
2681 //
2682 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2683 // This basic scheme here is the same as inline_unsafe_access, but
2684 // differs in enough details that combining them would make the code
2685 // overly confusing. (This is a true fact! I originally combined
2686 // them, but even I was confused by it!) As much code/comments as
2687 // possible are retained from inline_unsafe_access though to make
2688 // the correspondences clearer. - dl
2689
2690 if (callee()->is_static()) return false; // caller must have the capability!
2691
2692 #ifndef PRODUCT
2693 BasicType rtype;
2694 {
2695 ResourceMark rm;
2696 // Check the signatures.
2697 ciSignature* sig = callee()->signature();
2698 rtype = sig->return_type()->basic_type();
2699 if (kind == LS_xadd || kind == LS_xchg) {
2700 // Check the signatures.
2701 #ifdef ASSERT
2702 assert(rtype == type, "get and set must return the expected type");
2703 assert(sig->count() == 3, "get and set has 3 arguments");
2704 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2705 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2706 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2707 #endif // ASSERT
2708 } else if (kind == LS_cmpxchg) {
2709 // Check the signatures.
2710 #ifdef ASSERT
2711 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2712 assert(sig->count() == 4, "CAS has 4 arguments");
2713 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2714 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2715 #endif // ASSERT
2716 } else {
2717 ShouldNotReachHere();
2718 }
2719 }
2720 #endif //PRODUCT
2721
2722 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2723
2724 // Get arguments:
2725 Node* receiver = NULL;
2726 Node* base = NULL;
2727 Node* offset = NULL;
2728 Node* oldval = NULL;
2729 Node* newval = NULL;
2730 if (kind == LS_cmpxchg) {
2731 const bool two_slot_type = type2size[type] == 2;
2732 receiver = argument(0); // type: oop
2733 base = argument(1); // type: oop
2734 offset = argument(2); // type: long
2735 oldval = argument(4); // type: oop, int, or long
2736 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2737 } else if (kind == LS_xadd || kind == LS_xchg){
2738 receiver = argument(0); // type: oop
2739 base = argument(1); // type: oop
2740 offset = argument(2); // type: long
2741 oldval = NULL;
2742 newval = argument(4); // type: oop, int, or long
2743 }
2744
2745 // Null check receiver.
2746 receiver = null_check(receiver);
2747 if (stopped()) {
2748 return true;
2749 }
2750
2751 // Build field offset expression.
2752 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2753 // to be plain byte offsets, which are also the same as those accepted
2754 // by oopDesc::field_base.
2755 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2756 // 32-bit machines ignore the high half of long offsets
2757 offset = ConvL2X(offset);
2758 Node* adr = make_unsafe_address(base, offset);
2759 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2760
2761 // For CAS, unlike inline_unsafe_access, there seems no point in
2762 // trying to refine types. Just use the coarse types here.
2763 const Type *value_type = Type::get_const_basic_type(type);
2764 Compile::AliasType* alias_type = C->alias_type(adr_type);
2765 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2766
2767 if (kind == LS_xchg && type == T_OBJECT) {
2768 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2769 if (tjp != NULL) {
2770 value_type = tjp;
2771 }
2772 }
2773
2774 int alias_idx = C->get_alias_index(adr_type);
2775
2776 // Memory-model-wise, a LoadStore acts like a little synchronized
2777 // block, so needs barriers on each side. These don't translate
2778 // into actual barriers on most machines, but we still need rest of
2779 // compiler to respect ordering.
2780
2781 insert_mem_bar(Op_MemBarRelease);
2782 insert_mem_bar(Op_MemBarCPUOrder);
2783
2784 // 4984716: MemBars must be inserted before this
2785 // memory node in order to avoid a false
2786 // dependency which will confuse the scheduler.
2787 Node *mem = memory(alias_idx);
2788
2789 // For now, we handle only those cases that actually exist: ints,
2790 // longs, and Object. Adding others should be straightforward.
2791 Node* load_store;
2792 switch(type) {
2793 case T_INT:
2794 if (kind == LS_xadd) {
2795 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2796 } else if (kind == LS_xchg) {
2797 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2798 } else if (kind == LS_cmpxchg) {
2799 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2800 } else {
2801 ShouldNotReachHere();
2802 }
2803 break;
2804 case T_LONG:
2805 if (kind == LS_xadd) {
2806 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2807 } else if (kind == LS_xchg) {
2808 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2809 } else if (kind == LS_cmpxchg) {
2810 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2811 } else {
2812 ShouldNotReachHere();
2813 }
2814 break;
2815 case T_OBJECT:
2816 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2817 // could be delayed during Parse (for example, in adjust_map_after_if()).
2818 // Execute transformation here to avoid barrier generation in such case.
2819 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2820 newval = _gvn.makecon(TypePtr::NULL_PTR);
2821
2822 // Reference stores need a store barrier.
2823 if (kind == LS_xchg) {
2824 // If pre-barrier must execute before the oop store, old value will require do_load here.
2825 if (!can_move_pre_barrier()) {
2826 pre_barrier(true /* do_load*/,
2827 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2828 NULL /* pre_val*/,
2829 T_OBJECT);
2830 } // Else move pre_barrier to use load_store value, see below.
2831 } else if (kind == LS_cmpxchg) {
2832 // Same as for newval above:
2833 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2834 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2835 }
2836 // The only known value which might get overwritten is oldval.
2837 pre_barrier(false /* do_load */,
2838 control(), NULL, NULL, max_juint, NULL, NULL,
2839 oldval /* pre_val */,
2840 T_OBJECT);
2841 } else {
2842 ShouldNotReachHere();
2843 }
2844
2845 #ifdef _LP64
2846 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2847 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2848 if (kind == LS_xchg) {
2849 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2850 newval_enc, adr_type, value_type->make_narrowoop()));
2851 } else {
2852 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2853 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2854 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2855 newval_enc, oldval_enc));
2856 }
2857 } else
2858 #endif
2859 {
2860 if (kind == LS_xchg) {
2861 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2862 } else {
2863 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2864 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2865 }
2866 }
2867 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2868 break;
2869 default:
2870 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2871 break;
2872 }
2873
2874 // SCMemProjNodes represent the memory state of a LoadStore. Their
2875 // main role is to prevent LoadStore nodes from being optimized away
2876 // when their results aren't used.
2877 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2878 set_memory(proj, alias_idx);
2879
2880 if (type == T_OBJECT && kind == LS_xchg) {
2881 #ifdef _LP64
2882 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2883 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2884 }
2885 #endif
2886 if (can_move_pre_barrier()) {
2887 // Don't need to load pre_val. The old value is returned by load_store.
2888 // The pre_barrier can execute after the xchg as long as no safepoint
2889 // gets inserted between them.
2890 pre_barrier(false /* do_load */,
2891 control(), NULL, NULL, max_juint, NULL, NULL,
2892 load_store /* pre_val */,
2893 T_OBJECT);
2894 }
2895 }
2896
2897 // Add the trailing membar surrounding the access
2898 insert_mem_bar(Op_MemBarCPUOrder);
2899 insert_mem_bar(Op_MemBarAcquire);
2900
2901 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2902 set_result(load_store);
2903 return true;
2904 }
2905
2906 //----------------------------inline_unsafe_ordered_store----------------------
2907 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2908 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2909 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2910 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2911 // This is another variant of inline_unsafe_access, differing in
2912 // that it always issues store-store ("release") barrier and ensures
2913 // store-atomicity (which only matters for "long").
2914
2915 if (callee()->is_static()) return false; // caller must have the capability!
2916
2917 #ifndef PRODUCT
2918 {
2919 ResourceMark rm;
2920 // Check the signatures.
2921 ciSignature* sig = callee()->signature();
2922 #ifdef ASSERT
2923 BasicType rtype = sig->return_type()->basic_type();
2924 assert(rtype == T_VOID, "must return void");
2925 assert(sig->count() == 3, "has 3 arguments");
2926 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2927 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2928 #endif // ASSERT
2929 }
2930 #endif //PRODUCT
2931
2932 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2933
2934 // Get arguments:
2935 Node* receiver = argument(0); // type: oop
2936 Node* base = argument(1); // type: oop
2937 Node* offset = argument(2); // type: long
2938 Node* val = argument(4); // type: oop, int, or long
2939
2940 // Null check receiver.
2941 receiver = null_check(receiver);
2942 if (stopped()) {
2943 return true;
2944 }
2945
2946 // Build field offset expression.
2947 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2948 // 32-bit machines ignore the high half of long offsets
2949 offset = ConvL2X(offset);
2950 Node* adr = make_unsafe_address(base, offset);
2951 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2952 const Type *value_type = Type::get_const_basic_type(type);
2953 Compile::AliasType* alias_type = C->alias_type(adr_type);
2954
2955 insert_mem_bar(Op_MemBarRelease);
2956 insert_mem_bar(Op_MemBarCPUOrder);
2957 // Ensure that the store is atomic for longs:
2958 const bool require_atomic_access = true;
2959 Node* store;
2960 if (type == T_OBJECT) // reference stores need a store barrier.
2961 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2962 else {
2963 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2964 }
2965 insert_mem_bar(Op_MemBarCPUOrder);
2966 return true;
2967 }
2968
2969 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2970 // Regardless of form, don't allow previous ld/st to move down,
2971 // then issue acquire, release, or volatile mem_bar.
2972 insert_mem_bar(Op_MemBarCPUOrder);
2973 switch(id) {
2974 case vmIntrinsics::_loadFence:
2975 insert_mem_bar(Op_MemBarAcquire);
2976 return true;
2977 case vmIntrinsics::_storeFence:
2978 insert_mem_bar(Op_MemBarRelease);
2979 return true;
2980 case vmIntrinsics::_fullFence:
2981 insert_mem_bar(Op_MemBarVolatile);
2982 return true;
2983 default:
2984 fatal_unexpected_iid(id);
2985 return false;
2986 }
2987 }
2988
2989 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2990 if (!kls->is_Con()) {
2991 return true;
2992 }
2993 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2994 if (klsptr == NULL) {
2995 return true;
2996 }
2997 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2998 // don't need a guard for a klass that is already initialized
2999 return !ik->is_initialized();
3000 }
3001
3002 //----------------------------inline_unsafe_allocate---------------------------
3003 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3004 bool LibraryCallKit::inline_unsafe_allocate() {
3005 if (callee()->is_static()) return false; // caller must have the capability!
3006
3007 null_check_receiver(); // null-check, then ignore
3008 Node* cls = null_check(argument(1));
3009 if (stopped()) return true;
3010
3011 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3012 kls = null_check(kls);
3013 if (stopped()) return true; // argument was like int.class
3014
3015 Node* test = NULL;
3016 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3017 // Note: The argument might still be an illegal value like
3018 // Serializable.class or Object[].class. The runtime will handle it.
3019 // But we must make an explicit check for initialization.
3020 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3021 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3022 // can generate code to load it as unsigned byte.
3023 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3024 Node* bits = intcon(InstanceKlass::fully_initialized);
3025 test = _gvn.transform(new (C) SubINode(inst, bits));
3026 // The 'test' is non-zero if we need to take a slow path.
3027 }
3028
3029 Node* obj = new_instance(kls, test);
3030 set_result(obj);
3031 return true;
3032 }
3033
3034 #ifdef TRACE_HAVE_INTRINSICS
3035 /*
3036 * oop -> myklass
3037 * myklass->trace_id |= USED
3038 * return myklass->trace_id & ~0x3
3039 */
3040 bool LibraryCallKit::inline_native_classID() {
3041 null_check_receiver(); // null-check, then ignore
3042 Node* cls = null_check(argument(1), T_OBJECT);
3043 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3044 kls = null_check(kls, T_OBJECT);
3045 ByteSize offset = TRACE_ID_OFFSET;
3046 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3047 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3048 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3049 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3050 Node* clsused = longcon(0x01l); // set the class bit
3051 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3052
3053 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3054 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3055 set_result(andl);
3056 return true;
3057 }
3058
3059 bool LibraryCallKit::inline_native_threadID() {
3060 Node* tls_ptr = NULL;
3061 Node* cur_thr = generate_current_thread(tls_ptr);
3062 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3063 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3064 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3065
3066 Node* threadid = NULL;
3067 size_t thread_id_size = OSThread::thread_id_size();
3068 if (thread_id_size == (size_t) BytesPerLong) {
3069 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3070 } else if (thread_id_size == (size_t) BytesPerInt) {
3071 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3072 } else {
3073 ShouldNotReachHere();
3074 }
3075 set_result(threadid);
3076 return true;
3077 }
3078 #endif
3079
3080 //------------------------inline_native_time_funcs--------------
3081 // inline code for System.currentTimeMillis() and System.nanoTime()
3082 // these have the same type and signature
3083 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3084 const TypeFunc* tf = OptoRuntime::void_long_Type();
3085 const TypePtr* no_memory_effects = NULL;
3086 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3087 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3088 #ifdef ASSERT
3089 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3090 assert(value_top == top(), "second value must be top");
3091 #endif
3092 set_result(value);
3093 return true;
3094 }
3095
3096 //------------------------inline_native_currentThread------------------
3097 bool LibraryCallKit::inline_native_currentThread() {
3098 Node* junk = NULL;
3099 set_result(generate_current_thread(junk));
3100 return true;
3101 }
3102
3103 //------------------------inline_native_isInterrupted------------------
3104 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3105 bool LibraryCallKit::inline_native_isInterrupted() {
3106 // Add a fast path to t.isInterrupted(clear_int):
3107 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3108 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3109 // So, in the common case that the interrupt bit is false,
3110 // we avoid making a call into the VM. Even if the interrupt bit
3111 // is true, if the clear_int argument is false, we avoid the VM call.
3112 // However, if the receiver is not currentThread, we must call the VM,
3113 // because there must be some locking done around the operation.
3114
3115 // We only go to the fast case code if we pass two guards.
3116 // Paths which do not pass are accumulated in the slow_region.
3117
3118 enum {
3119 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3120 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3121 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3122 PATH_LIMIT
3123 };
3124
3125 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3126 // out of the function.
3127 insert_mem_bar(Op_MemBarCPUOrder);
3128
3129 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3130 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3131
3132 RegionNode* slow_region = new (C) RegionNode(1);
3133 record_for_igvn(slow_region);
3134
3135 // (a) Receiving thread must be the current thread.
3136 Node* rec_thr = argument(0);
3137 Node* tls_ptr = NULL;
3138 Node* cur_thr = generate_current_thread(tls_ptr);
3139 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3140 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3141
3142 generate_slow_guard(bol_thr, slow_region);
3143
3144 // (b) Interrupt bit on TLS must be false.
3145 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3146 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3147 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3148
3149 // Set the control input on the field _interrupted read to prevent it floating up.
3150 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3151 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3152 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3153
3154 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3155
3156 // First fast path: if (!TLS._interrupted) return false;
3157 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3158 result_rgn->init_req(no_int_result_path, false_bit);
3159 result_val->init_req(no_int_result_path, intcon(0));
3160
3161 // drop through to next case
3162 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3163
3164 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3165 Node* clr_arg = argument(1);
3166 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3167 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3168 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3169
3170 // Second fast path: ... else if (!clear_int) return true;
3171 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3172 result_rgn->init_req(no_clear_result_path, false_arg);
3173 result_val->init_req(no_clear_result_path, intcon(1));
3174
3175 // drop through to next case
3176 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
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);
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, 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);
3256 Node* mask = intcon(modifier_mask);
3257 Node* bits = intcon(modifier_bits);
3258 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3259 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3260 Node* bol = _gvn.transform(new (C) 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 Node* obj = top();
3278
3279 switch (id) {
3280 case vmIntrinsics::_isInstance:
3281 // nothing is an instance of a primitive type
3282 prim_return_value = intcon(0);
3283 obj = argument(1);
3284 break;
3285 case vmIntrinsics::_getModifiers:
3286 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3287 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3288 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3289 break;
3290 case vmIntrinsics::_isInterface:
3291 prim_return_value = intcon(0);
3292 break;
3293 case vmIntrinsics::_isArray:
3294 prim_return_value = intcon(0);
3295 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3296 break;
3297 case vmIntrinsics::_isPrimitive:
3298 prim_return_value = intcon(1);
3299 expect_prim = true; // obviously
3300 break;
3301 case vmIntrinsics::_getSuperclass:
3302 prim_return_value = null();
3303 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3304 break;
3305 case vmIntrinsics::_getComponentType:
3306 prim_return_value = null();
3307 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3308 break;
3309 case vmIntrinsics::_getClassAccessFlags:
3310 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3311 return_type = TypeInt::INT; // not bool! 6297094
3312 break;
3313 default:
3314 fatal_unexpected_iid(id);
3315 break;
3316 }
3317
3318 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3319 if (mirror_con == NULL) return false; // cannot happen?
3320
3321 #ifndef PRODUCT
3322 if (C->print_intrinsics() || C->print_inlining()) {
3323 ciType* k = mirror_con->java_mirror_type();
3324 if (k) {
3325 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3326 k->print_name();
3327 tty->cr();
3328 }
3329 }
3330 #endif
3331
3332 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3333 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3334 record_for_igvn(region);
3335 PhiNode* phi = new (C) PhiNode(region, return_type);
3336
3337 // The mirror will never be null of Reflection.getClassAccessFlags, however
3338 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3339 // if it is. See bug 4774291.
3340
3341 // For Reflection.getClassAccessFlags(), the null check occurs in
3342 // the wrong place; see inline_unsafe_access(), above, for a similar
3343 // situation.
3344 mirror = null_check(mirror);
3345 // If mirror or obj is dead, only null-path is taken.
3346 if (stopped()) return true;
3347
3348 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3349
3350 // Now load the mirror's klass metaobject, and null-check it.
3351 // Side-effects region with the control path if the klass is null.
3352 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3353 // If kls is null, we have a primitive mirror.
3354 phi->init_req(_prim_path, prim_return_value);
3355 if (stopped()) { set_result(region, phi); return true; }
3356 bool safe_for_replace = (region->in(_prim_path) == top());
3357
3358 Node* p; // handy temp
3359 Node* null_ctl;
3360
3361 // Now that we have the non-null klass, we can perform the real query.
3362 // For constant classes, the query will constant-fold in LoadNode::Value.
3363 Node* query_value = top();
3364 switch (id) {
3365 case vmIntrinsics::_isInstance:
3366 // nothing is an instance of a primitive type
3367 query_value = gen_instanceof(obj, kls, safe_for_replace);
3368 break;
3369
3370 case vmIntrinsics::_getModifiers:
3371 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3372 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3373 break;
3374
3375 case vmIntrinsics::_isInterface:
3376 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3377 if (generate_interface_guard(kls, region) != NULL)
3378 // A guard was added. If the guard is taken, it was an interface.
3379 phi->add_req(intcon(1));
3380 // If we fall through, it's a plain class.
3381 query_value = intcon(0);
3382 break;
3383
3384 case vmIntrinsics::_isArray:
3385 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3386 if (generate_array_guard(kls, region) != NULL)
3387 // A guard was added. If the guard is taken, it was an array.
3388 phi->add_req(intcon(1));
3389 // If we fall through, it's a plain class.
3390 query_value = intcon(0);
3391 break;
3392
3393 case vmIntrinsics::_isPrimitive:
3394 query_value = intcon(0); // "normal" path produces false
3395 break;
3396
3397 case vmIntrinsics::_getSuperclass:
3398 // The rules here are somewhat unfortunate, but we can still do better
3399 // with random logic than with a JNI call.
3400 // Interfaces store null or Object as _super, but must report null.
3401 // Arrays store an intermediate super as _super, but must report Object.
3402 // Other types can report the actual _super.
3403 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3404 if (generate_interface_guard(kls, region) != NULL)
3405 // A guard was added. If the guard is taken, it was an interface.
3406 phi->add_req(null());
3407 if (generate_array_guard(kls, region) != NULL)
3408 // A guard was added. If the guard is taken, it was an array.
3409 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3410 // If we fall through, it's a plain class. Get its _super.
3411 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3412 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3413 null_ctl = top();
3414 kls = null_check_oop(kls, &null_ctl);
3415 if (null_ctl != top()) {
3416 // If the guard is taken, Object.superClass is null (both klass and mirror).
3417 region->add_req(null_ctl);
3418 phi ->add_req(null());
3419 }
3420 if (!stopped()) {
3421 query_value = load_mirror_from_klass(kls);
3422 }
3423 break;
3424
3425 case vmIntrinsics::_getComponentType:
3426 if (generate_array_guard(kls, region) != NULL) {
3427 // Be sure to pin the oop load to the guard edge just created:
3428 Node* is_array_ctrl = region->in(region->req()-1);
3429 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3430 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3431 phi->add_req(cmo);
3432 }
3433 query_value = null(); // non-array case is null
3434 break;
3435
3436 case vmIntrinsics::_getClassAccessFlags:
3437 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3438 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3439 break;
3440
3441 default:
3442 fatal_unexpected_iid(id);
3443 break;
3444 }
3445
3446 // Fall-through is the normal case of a query to a real class.
3447 phi->init_req(1, query_value);
3448 region->init_req(1, control());
3449
3450 C->set_has_split_ifs(true); // Has chance for split-if optimization
3451 set_result(region, phi);
3452 return true;
3453 }
3454
3455 //--------------------------inline_native_subtype_check------------------------
3456 // This intrinsic takes the JNI calls out of the heart of
3457 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3458 bool LibraryCallKit::inline_native_subtype_check() {
3459 // Pull both arguments off the stack.
3460 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3461 args[0] = argument(0);
3462 args[1] = argument(1);
3463 Node* klasses[2]; // corresponding Klasses: superk, subk
3464 klasses[0] = klasses[1] = top();
3465
3466 enum {
3467 // A full decision tree on {superc is prim, subc is prim}:
3468 _prim_0_path = 1, // {P,N} => false
3469 // {P,P} & superc!=subc => false
3470 _prim_same_path, // {P,P} & superc==subc => true
3471 _prim_1_path, // {N,P} => false
3472 _ref_subtype_path, // {N,N} & subtype check wins => true
3473 _both_ref_path, // {N,N} & subtype check loses => false
3474 PATH_LIMIT
3475 };
3476
3477 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3478 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3479 record_for_igvn(region);
3480
3481 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3482 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3483 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3484
3485 // First null-check both mirrors and load each mirror's klass metaobject.
3486 int which_arg;
3487 for (which_arg = 0; which_arg <= 1; which_arg++) {
3488 Node* arg = args[which_arg];
3489 arg = null_check(arg);
3490 if (stopped()) break;
3491 args[which_arg] = arg;
3492
3493 Node* p = basic_plus_adr(arg, class_klass_offset);
3494 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3495 klasses[which_arg] = _gvn.transform(kls);
3496 }
3497
3498 // Having loaded both klasses, test each for null.
3499 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3500 for (which_arg = 0; which_arg <= 1; which_arg++) {
3501 Node* kls = klasses[which_arg];
3502 Node* null_ctl = top();
3503 kls = null_check_oop(kls, &null_ctl, never_see_null);
3504 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3505 region->init_req(prim_path, null_ctl);
3506 if (stopped()) break;
3507 klasses[which_arg] = kls;
3508 }
3509
3510 if (!stopped()) {
3511 // now we have two reference types, in klasses[0..1]
3512 Node* subk = klasses[1]; // the argument to isAssignableFrom
3513 Node* superk = klasses[0]; // the receiver
3514 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3515 // now we have a successful reference subtype check
3516 region->set_req(_ref_subtype_path, control());
3517 }
3518
3519 // If both operands are primitive (both klasses null), then
3520 // we must return true when they are identical primitives.
3521 // It is convenient to test this after the first null klass check.
3522 set_control(region->in(_prim_0_path)); // go back to first null check
3523 if (!stopped()) {
3524 // Since superc is primitive, make a guard for the superc==subc case.
3525 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3526 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3527 generate_guard(bol_eq, region, PROB_FAIR);
3528 if (region->req() == PATH_LIMIT+1) {
3529 // A guard was added. If the added guard is taken, superc==subc.
3530 region->swap_edges(PATH_LIMIT, _prim_same_path);
3531 region->del_req(PATH_LIMIT);
3532 }
3533 region->set_req(_prim_0_path, control()); // Not equal after all.
3534 }
3535
3536 // these are the only paths that produce 'true':
3537 phi->set_req(_prim_same_path, intcon(1));
3538 phi->set_req(_ref_subtype_path, intcon(1));
3539
3540 // pull together the cases:
3541 assert(region->req() == PATH_LIMIT, "sane region");
3542 for (uint i = 1; i < region->req(); i++) {
3543 Node* ctl = region->in(i);
3544 if (ctl == NULL || ctl == top()) {
3545 region->set_req(i, top());
3546 phi ->set_req(i, top());
3547 } else if (phi->in(i) == NULL) {
3548 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3549 }
3550 }
3551
3552 set_control(_gvn.transform(region));
3553 set_result(_gvn.transform(phi));
3554 return true;
3555 }
3556
3557 //---------------------generate_array_guard_common------------------------
3558 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3559 bool obj_array, bool not_array) {
3560 // If obj_array/non_array==false/false:
3561 // Branch around if the given klass is in fact an array (either obj or prim).
3562 // If obj_array/non_array==false/true:
3563 // Branch around if the given klass is not an array klass of any kind.
3564 // If obj_array/non_array==true/true:
3565 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3566 // If obj_array/non_array==true/false:
3567 // Branch around if the kls is an oop array (Object[] or subtype)
3568 //
3569 // Like generate_guard, adds a new path onto the region.
3570 jint layout_con = 0;
3571 Node* layout_val = get_layout_helper(kls, layout_con);
3572 if (layout_val == NULL) {
3573 bool query = (obj_array
3574 ? Klass::layout_helper_is_objArray(layout_con)
3575 : Klass::layout_helper_is_array(layout_con));
3576 if (query == not_array) {
3577 return NULL; // never a branch
3578 } else { // always a branch
3579 Node* always_branch = control();
3580 if (region != NULL)
3581 region->add_req(always_branch);
3582 set_control(top());
3583 return always_branch;
3584 }
3585 }
3586 // Now test the correct condition.
3587 jint nval = (obj_array
3588 ? ((jint)Klass::_lh_array_tag_type_value
3589 << Klass::_lh_array_tag_shift)
3590 : Klass::_lh_neutral_value);
3591 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3592 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3593 // invert the test if we are looking for a non-array
3594 if (not_array) btest = BoolTest(btest).negate();
3595 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3596 return generate_fair_guard(bol, region);
3597 }
3598
3599
3600 //-----------------------inline_native_newArray--------------------------
3601 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3602 bool LibraryCallKit::inline_native_newArray() {
3603 Node* mirror = argument(0);
3604 Node* count_val = argument(1);
3605
3606 mirror = null_check(mirror);
3607 // If mirror or obj is dead, only null-path is taken.
3608 if (stopped()) return true;
3609
3610 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3611 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3612 PhiNode* result_val = new(C) PhiNode(result_reg,
3613 TypeInstPtr::NOTNULL);
3614 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3615 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3616 TypePtr::BOTTOM);
3617
3618 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3619 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3620 result_reg, _slow_path);
3621 Node* normal_ctl = control();
3622 Node* no_array_ctl = result_reg->in(_slow_path);
3623
3624 // Generate code for the slow case. We make a call to newArray().
3625 set_control(no_array_ctl);
3626 if (!stopped()) {
3627 // Either the input type is void.class, or else the
3628 // array klass has not yet been cached. Either the
3629 // ensuing call will throw an exception, or else it
3630 // will cache the array klass for next time.
3631 PreserveJVMState pjvms(this);
3632 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3633 Node* slow_result = set_results_for_java_call(slow_call);
3634 // this->control() comes from set_results_for_java_call
3635 result_reg->set_req(_slow_path, control());
3636 result_val->set_req(_slow_path, slow_result);
3637 result_io ->set_req(_slow_path, i_o());
3638 result_mem->set_req(_slow_path, reset_memory());
3639 }
3640
3641 set_control(normal_ctl);
3642 if (!stopped()) {
3643 // Normal case: The array type has been cached in the java.lang.Class.
3644 // The following call works fine even if the array type is polymorphic.
3645 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3646 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3647 result_reg->init_req(_normal_path, control());
3648 result_val->init_req(_normal_path, obj);
3649 result_io ->init_req(_normal_path, i_o());
3650 result_mem->init_req(_normal_path, reset_memory());
3651 }
3652
3653 // Return the combined state.
3654 set_i_o( _gvn.transform(result_io) );
3655 set_all_memory( _gvn.transform(result_mem));
3656
3657 C->set_has_split_ifs(true); // Has chance for split-if optimization
3658 set_result(result_reg, result_val);
3659 return true;
3660 }
3661
3662 //----------------------inline_native_getLength--------------------------
3663 // public static native int java.lang.reflect.Array.getLength(Object array);
3664 bool LibraryCallKit::inline_native_getLength() {
3665 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3666
3667 Node* array = null_check(argument(0));
3668 // If array is dead, only null-path is taken.
3669 if (stopped()) return true;
3670
3671 // Deoptimize if it is a non-array.
3672 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3673
3674 if (non_array != NULL) {
3675 PreserveJVMState pjvms(this);
3676 set_control(non_array);
3677 uncommon_trap(Deoptimization::Reason_intrinsic,
3678 Deoptimization::Action_maybe_recompile);
3679 }
3680
3681 // If control is dead, only non-array-path is taken.
3682 if (stopped()) return true;
3683
3684 // The works fine even if the array type is polymorphic.
3685 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3686 Node* result = load_array_length(array);
3687
3688 C->set_has_split_ifs(true); // Has chance for split-if optimization
3689 set_result(result);
3690 return true;
3691 }
3692
3693 //------------------------inline_array_copyOf----------------------------
3694 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3695 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3696 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3697 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3698
3699 // Get the arguments.
3700 Node* original = argument(0);
3701 Node* start = is_copyOfRange? argument(1): intcon(0);
3702 Node* end = is_copyOfRange? argument(2): argument(1);
3703 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3704
3705 Node* newcopy;
3706
3707 // Set the original stack and the reexecute bit for the interpreter to reexecute
3708 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3709 { PreserveReexecuteState preexecs(this);
3710 jvms()->set_should_reexecute(true);
3711
3712 array_type_mirror = null_check(array_type_mirror);
3713 original = null_check(original);
3714
3715 // Check if a null path was taken unconditionally.
3716 if (stopped()) return true;
3717
3718 Node* orig_length = load_array_length(original);
3719
3720 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3721 klass_node = null_check(klass_node);
3722
3723 RegionNode* bailout = new (C) RegionNode(1);
3724 record_for_igvn(bailout);
3725
3726 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3727 // Bail out if that is so.
3728 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3729 if (not_objArray != NULL) {
3730 // Improve the klass node's type from the new optimistic assumption:
3731 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3732 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3733 Node* cast = new (C) CastPPNode(klass_node, akls);
3734 cast->init_req(0, control());
3735 klass_node = _gvn.transform(cast);
3736 }
3737
3738 // Bail out if either start or end is negative.
3739 generate_negative_guard(start, bailout, &start);
3740 generate_negative_guard(end, bailout, &end);
3741
3742 Node* length = end;
3743 if (_gvn.type(start) != TypeInt::ZERO) {
3744 length = _gvn.transform(new (C) SubINode(end, start));
3745 }
3746
3747 // Bail out if length is negative.
3748 // Without this the new_array would throw
3749 // NegativeArraySizeException but IllegalArgumentException is what
3750 // should be thrown
3751 generate_negative_guard(length, bailout, &length);
3752
3753 if (bailout->req() > 1) {
3754 PreserveJVMState pjvms(this);
3755 set_control(_gvn.transform(bailout));
3756 uncommon_trap(Deoptimization::Reason_intrinsic,
3757 Deoptimization::Action_maybe_recompile);
3758 }
3759
3760 if (!stopped()) {
3761 // How many elements will we copy from the original?
3762 // The answer is MinI(orig_length - start, length).
3763 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3764 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3765
3766 newcopy = new_array(klass_node, length, 0); // no argments to push
3767
3768 // Generate a direct call to the right arraycopy function(s).
3769 // We know the copy is disjoint but we might not know if the
3770 // oop stores need checking.
3771 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3772 // This will fail a store-check if x contains any non-nulls.
3773 bool disjoint_bases = true;
3774 // if start > orig_length then the length of the copy may be
3775 // negative.
3776 bool length_never_negative = !is_copyOfRange;
3777 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3778 original, start, newcopy, intcon(0), moved,
3779 disjoint_bases, length_never_negative);
3780 }
3781 } // original reexecute is set back here
3782
3783 C->set_has_split_ifs(true); // Has chance for split-if optimization
3784 if (!stopped()) {
3785 set_result(newcopy);
3786 }
3787 return true;
3788 }
3789
3790
3791 //----------------------generate_virtual_guard---------------------------
3792 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3793 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3794 RegionNode* slow_region) {
3795 ciMethod* method = callee();
3796 int vtable_index = method->vtable_index();
3797 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3798 err_msg_res("bad index %d", vtable_index));
3799 // Get the Method* out of the appropriate vtable entry.
3800 int entry_offset = (InstanceKlass::vtable_start_offset() +
3801 vtable_index*vtableEntry::size()) * wordSize +
3802 vtableEntry::method_offset_in_bytes();
3803 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3804 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3805
3806 // Compare the target method with the expected method (e.g., Object.hashCode).
3807 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3808
3809 Node* native_call = makecon(native_call_addr);
3810 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3811 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3812
3813 return generate_slow_guard(test_native, slow_region);
3814 }
3815
3816 //-----------------------generate_method_call----------------------------
3817 // Use generate_method_call to make a slow-call to the real
3818 // method if the fast path fails. An alternative would be to
3819 // use a stub like OptoRuntime::slow_arraycopy_Java.
3820 // This only works for expanding the current library call,
3821 // not another intrinsic. (E.g., don't use this for making an
3822 // arraycopy call inside of the copyOf intrinsic.)
3823 CallJavaNode*
3824 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3825 // When compiling the intrinsic method itself, do not use this technique.
3826 guarantee(callee() != C->method(), "cannot make slow-call to self");
3827
3828 ciMethod* method = callee();
3829 // ensure the JVMS we have will be correct for this call
3830 guarantee(method_id == method->intrinsic_id(), "must match");
3831
3832 const TypeFunc* tf = TypeFunc::make(method);
3833 CallJavaNode* slow_call;
3834 if (is_static) {
3835 assert(!is_virtual, "");
3836 slow_call = new(C) CallStaticJavaNode(C, tf,
3837 SharedRuntime::get_resolve_static_call_stub(),
3838 method, bci());
3839 } else if (is_virtual) {
3840 null_check_receiver();
3841 int vtable_index = Method::invalid_vtable_index;
3842 if (UseInlineCaches) {
3843 // Suppress the vtable call
3844 } else {
3845 // hashCode and clone are not a miranda methods,
3846 // so the vtable index is fixed.
3847 // No need to use the linkResolver to get it.
3848 vtable_index = method->vtable_index();
3849 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3850 err_msg_res("bad index %d", vtable_index));
3851 }
3852 slow_call = new(C) CallDynamicJavaNode(tf,
3853 SharedRuntime::get_resolve_virtual_call_stub(),
3854 method, vtable_index, bci());
3855 } else { // neither virtual nor static: opt_virtual
3856 null_check_receiver();
3857 slow_call = new(C) CallStaticJavaNode(C, tf,
3858 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3859 method, bci());
3860 slow_call->set_optimized_virtual(true);
3861 }
3862 set_arguments_for_java_call(slow_call);
3863 set_edges_for_java_call(slow_call);
3864 return slow_call;
3865 }
3866
3867
3868 //------------------------------inline_native_hashcode--------------------
3869 // Build special case code for calls to hashCode on an object.
3870 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3871 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3872 assert(!(is_virtual && is_static), "either virtual, special, or static");
3873
3874 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3875
3876 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3877 PhiNode* result_val = new(C) PhiNode(result_reg,
3878 TypeInt::INT);
3879 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3880 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3881 TypePtr::BOTTOM);
3882 Node* obj = NULL;
3883 if (!is_static) {
3884 // Check for hashing null object
3885 obj = null_check_receiver();
3886 if (stopped()) return true; // unconditionally null
3887 result_reg->init_req(_null_path, top());
3888 result_val->init_req(_null_path, top());
3889 } else {
3890 // Do a null check, and return zero if null.
3891 // System.identityHashCode(null) == 0
3892 obj = argument(0);
3893 Node* null_ctl = top();
3894 obj = null_check_oop(obj, &null_ctl);
3895 result_reg->init_req(_null_path, null_ctl);
3896 result_val->init_req(_null_path, _gvn.intcon(0));
3897 }
3898
3899 // Unconditionally null? Then return right away.
3900 if (stopped()) {
3901 set_control( result_reg->in(_null_path));
3902 if (!stopped())
3903 set_result(result_val->in(_null_path));
3904 return true;
3905 }
3906
3907 // After null check, get the object's klass.
3908 Node* obj_klass = load_object_klass(obj);
3909
3910 // This call may be virtual (invokevirtual) or bound (invokespecial).
3911 // For each case we generate slightly different code.
3912
3913 // We only go to the fast case code if we pass a number of guards. The
3914 // paths which do not pass are accumulated in the slow_region.
3915 RegionNode* slow_region = new (C) RegionNode(1);
3916 record_for_igvn(slow_region);
3917
3918 // If this is a virtual call, we generate a funny guard. We pull out
3919 // the vtable entry corresponding to hashCode() from the target object.
3920 // If the target method which we are calling happens to be the native
3921 // Object hashCode() method, we pass the guard. We do not need this
3922 // guard for non-virtual calls -- the caller is known to be the native
3923 // Object hashCode().
3924 if (is_virtual) {
3925 generate_virtual_guard(obj_klass, slow_region);
3926 }
3927
3928 // Get the header out of the object, use LoadMarkNode when available
3929 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3930 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3931
3932 // Test the header to see if it is unlocked.
3933 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3934 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
3935 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3936 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
3937 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
3938
3939 generate_slow_guard(test_unlocked, slow_region);
3940
3941 // Get the hash value and check to see that it has been properly assigned.
3942 // We depend on hash_mask being at most 32 bits and avoid the use of
3943 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3944 // vm: see markOop.hpp.
3945 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3946 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3947 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
3948 // This hack lets the hash bits live anywhere in the mark object now, as long
3949 // as the shift drops the relevant bits into the low 32 bits. Note that
3950 // Java spec says that HashCode is an int so there's no point in capturing
3951 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3952 hshifted_header = ConvX2I(hshifted_header);
3953 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
3954
3955 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3956 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
3957 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
3958
3959 generate_slow_guard(test_assigned, slow_region);
3960
3961 Node* init_mem = reset_memory();
3962 // fill in the rest of the null path:
3963 result_io ->init_req(_null_path, i_o());
3964 result_mem->init_req(_null_path, init_mem);
3965
3966 result_val->init_req(_fast_path, hash_val);
3967 result_reg->init_req(_fast_path, control());
3968 result_io ->init_req(_fast_path, i_o());
3969 result_mem->init_req(_fast_path, init_mem);
3970
3971 // Generate code for the slow case. We make a call to hashCode().
3972 set_control(_gvn.transform(slow_region));
3973 if (!stopped()) {
3974 // No need for PreserveJVMState, because we're using up the present state.
3975 set_all_memory(init_mem);
3976 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3977 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3978 Node* slow_result = set_results_for_java_call(slow_call);
3979 // this->control() comes from set_results_for_java_call
3980 result_reg->init_req(_slow_path, control());
3981 result_val->init_req(_slow_path, slow_result);
3982 result_io ->set_req(_slow_path, i_o());
3983 result_mem ->set_req(_slow_path, reset_memory());
3984 }
3985
3986 // Return the combined state.
3987 set_i_o( _gvn.transform(result_io) );
3988 set_all_memory( _gvn.transform(result_mem));
3989
3990 set_result(result_reg, result_val);
3991 return true;
3992 }
3993
3994 //---------------------------inline_native_getClass----------------------------
3995 // public final native Class<?> java.lang.Object.getClass();
3996 //
3997 // Build special case code for calls to getClass on an object.
3998 bool LibraryCallKit::inline_native_getClass() {
3999 Node* obj = null_check_receiver();
4000 if (stopped()) return true;
4001 set_result(load_mirror_from_klass(load_object_klass(obj)));
4002 return true;
4003 }
4004
4005 //-----------------inline_native_Reflection_getCallerClass---------------------
4006 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4007 //
4008 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4009 //
4010 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4011 // in that it must skip particular security frames and checks for
4012 // caller sensitive methods.
4013 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4014 #ifndef PRODUCT
4015 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4016 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4017 }
4018 #endif
4019
4020 if (!jvms()->has_method()) {
4021 #ifndef PRODUCT
4022 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4023 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4024 }
4025 #endif
4026 return false;
4027 }
4028
4029 // Walk back up the JVM state to find the caller at the required
4030 // depth.
4031 JVMState* caller_jvms = jvms();
4032
4033 // Cf. JVM_GetCallerClass
4034 // NOTE: Start the loop at depth 1 because the current JVM state does
4035 // not include the Reflection.getCallerClass() frame.
4036 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4037 ciMethod* m = caller_jvms->method();
4038 switch (n) {
4039 case 0:
4040 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4041 break;
4042 case 1:
4043 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4044 if (!m->caller_sensitive()) {
4045 #ifndef PRODUCT
4046 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4047 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4048 }
4049 #endif
4050 return false; // bail-out; let JVM_GetCallerClass do the work
4051 }
4052 break;
4053 default:
4054 if (!m->is_ignored_by_security_stack_walk()) {
4055 // We have reached the desired frame; return the holder class.
4056 // Acquire method holder as java.lang.Class and push as constant.
4057 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4058 ciInstance* caller_mirror = caller_klass->java_mirror();
4059 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4060
4061 #ifndef PRODUCT
4062 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4063 tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
4064 tty->print_cr(" JVM state at this point:");
4065 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4066 ciMethod* m = jvms()->of_depth(i)->method();
4067 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4068 }
4069 }
4070 #endif
4071 return true;
4072 }
4073 break;
4074 }
4075 }
4076
4077 #ifndef PRODUCT
4078 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4079 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4080 tty->print_cr(" JVM state at this point:");
4081 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4082 ciMethod* m = jvms()->of_depth(i)->method();
4083 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4084 }
4085 }
4086 #endif
4087
4088 return false; // bail-out; let JVM_GetCallerClass do the work
4089 }
4090
4091 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4092 Node* arg = argument(0);
4093 Node* result;
4094
4095 switch (id) {
4096 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4097 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4098 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4099 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4100
4101 case vmIntrinsics::_doubleToLongBits: {
4102 // two paths (plus control) merge in a wood
4103 RegionNode *r = new (C) RegionNode(3);
4104 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4105
4106 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4107 // Build the boolean node
4108 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4109
4110 // Branch either way.
4111 // NaN case is less traveled, which makes all the difference.
4112 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4113 Node *opt_isnan = _gvn.transform(ifisnan);
4114 assert( opt_isnan->is_If(), "Expect an IfNode");
4115 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4116 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4117
4118 set_control(iftrue);
4119
4120 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4121 Node *slow_result = longcon(nan_bits); // return NaN
4122 phi->init_req(1, _gvn.transform( slow_result ));
4123 r->init_req(1, iftrue);
4124
4125 // Else fall through
4126 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4127 set_control(iffalse);
4128
4129 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4130 r->init_req(2, iffalse);
4131
4132 // Post merge
4133 set_control(_gvn.transform(r));
4134 record_for_igvn(r);
4135
4136 C->set_has_split_ifs(true); // Has chance for split-if optimization
4137 result = phi;
4138 assert(result->bottom_type()->isa_long(), "must be");
4139 break;
4140 }
4141
4142 case vmIntrinsics::_floatToIntBits: {
4143 // two paths (plus control) merge in a wood
4144 RegionNode *r = new (C) RegionNode(3);
4145 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4146
4147 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4148 // Build the boolean node
4149 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4150
4151 // Branch either way.
4152 // NaN case is less traveled, which makes all the difference.
4153 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4154 Node *opt_isnan = _gvn.transform(ifisnan);
4155 assert( opt_isnan->is_If(), "Expect an IfNode");
4156 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4157 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4158
4159 set_control(iftrue);
4160
4161 static const jint nan_bits = 0x7fc00000;
4162 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4163 phi->init_req(1, _gvn.transform( slow_result ));
4164 r->init_req(1, iftrue);
4165
4166 // Else fall through
4167 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4168 set_control(iffalse);
4169
4170 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4171 r->init_req(2, iffalse);
4172
4173 // Post merge
4174 set_control(_gvn.transform(r));
4175 record_for_igvn(r);
4176
4177 C->set_has_split_ifs(true); // Has chance for split-if optimization
4178 result = phi;
4179 assert(result->bottom_type()->isa_int(), "must be");
4180 break;
4181 }
4182
4183 default:
4184 fatal_unexpected_iid(id);
4185 break;
4186 }
4187 set_result(_gvn.transform(result));
4188 return true;
4189 }
4190
4191 #ifdef _LP64
4192 #define XTOP ,top() /*additional argument*/
4193 #else //_LP64
4194 #define XTOP /*no additional argument*/
4195 #endif //_LP64
4196
4197 //----------------------inline_unsafe_copyMemory-------------------------
4198 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4199 bool LibraryCallKit::inline_unsafe_copyMemory() {
4200 if (callee()->is_static()) return false; // caller must have the capability!
4201 null_check_receiver(); // null-check receiver
4202 if (stopped()) return true;
4203
4204 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4205
4206 Node* src_ptr = argument(1); // type: oop
4207 Node* src_off = ConvL2X(argument(2)); // type: long
4208 Node* dst_ptr = argument(4); // type: oop
4209 Node* dst_off = ConvL2X(argument(5)); // type: long
4210 Node* size = ConvL2X(argument(7)); // type: long
4211
4212 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4213 "fieldOffset must be byte-scaled");
4214
4215 Node* src = make_unsafe_address(src_ptr, src_off);
4216 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4217
4218 // Conservatively insert a memory barrier on all memory slices.
4219 // Do not let writes of the copy source or destination float below the copy.
4220 insert_mem_bar(Op_MemBarCPUOrder);
4221
4222 // Call it. Note that the length argument is not scaled.
4223 make_runtime_call(RC_LEAF|RC_NO_FP,
4224 OptoRuntime::fast_arraycopy_Type(),
4225 StubRoutines::unsafe_arraycopy(),
4226 "unsafe_arraycopy",
4227 TypeRawPtr::BOTTOM,
4228 src, dst, size XTOP);
4229
4230 // Do not let reads of the copy destination float above the copy.
4231 insert_mem_bar(Op_MemBarCPUOrder);
4232
4233 return true;
4234 }
4235
4236 //------------------------clone_coping-----------------------------------
4237 // Helper function for inline_native_clone.
4238 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4239 assert(obj_size != NULL, "");
4240 Node* raw_obj = alloc_obj->in(1);
4241 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4242
4243 AllocateNode* alloc = NULL;
4244 if (ReduceBulkZeroing) {
4245 // We will be completely responsible for initializing this object -
4246 // mark Initialize node as complete.
4247 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4248 // The object was just allocated - there should be no any stores!
4249 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4250 // Mark as complete_with_arraycopy so that on AllocateNode
4251 // expansion, we know this AllocateNode is initialized by an array
4252 // copy and a StoreStore barrier exists after the array copy.
4253 alloc->initialization()->set_complete_with_arraycopy();
4254 }
4255
4256 // Copy the fastest available way.
4257 // TODO: generate fields copies for small objects instead.
4258 Node* src = obj;
4259 Node* dest = alloc_obj;
4260 Node* size = _gvn.transform(obj_size);
4261
4262 // Exclude the header but include array length to copy by 8 bytes words.
4263 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4264 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4265 instanceOopDesc::base_offset_in_bytes();
4266 // base_off:
4267 // 8 - 32-bit VM
4268 // 12 - 64-bit VM, compressed klass
4269 // 16 - 64-bit VM, normal klass
4270 if (base_off % BytesPerLong != 0) {
4271 assert(UseCompressedClassPointers, "");
4272 if (is_array) {
4273 // Exclude length to copy by 8 bytes words.
4274 base_off += sizeof(int);
4275 } else {
4276 // Include klass to copy by 8 bytes words.
4277 base_off = instanceOopDesc::klass_offset_in_bytes();
4278 }
4279 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4280 }
4281 src = basic_plus_adr(src, base_off);
4282 dest = basic_plus_adr(dest, base_off);
4283
4284 // Compute the length also, if needed:
4285 Node* countx = size;
4286 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4287 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4288
4289 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4290 bool disjoint_bases = true;
4291 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4292 src, NULL, dest, NULL, countx,
4293 /*dest_uninitialized*/true);
4294
4295 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4296 if (card_mark) {
4297 assert(!is_array, "");
4298 // Put in store barrier for any and all oops we are sticking
4299 // into this object. (We could avoid this if we could prove
4300 // that the object type contains no oop fields at all.)
4301 Node* no_particular_value = NULL;
4302 Node* no_particular_field = NULL;
4303 int raw_adr_idx = Compile::AliasIdxRaw;
4304 post_barrier(control(),
4305 memory(raw_adr_type),
4306 alloc_obj,
4307 no_particular_field,
4308 raw_adr_idx,
4309 no_particular_value,
4310 T_OBJECT,
4311 false);
4312 }
4313
4314 // Do not let reads from the cloned object float above the arraycopy.
4315 if (alloc != NULL) {
4316 // Do not let stores that initialize this object be reordered with
4317 // a subsequent store that would make this object accessible by
4318 // other threads.
4319 // Record what AllocateNode this StoreStore protects so that
4320 // escape analysis can go from the MemBarStoreStoreNode to the
4321 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4322 // based on the escape status of the AllocateNode.
4323 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4324 } else {
4325 insert_mem_bar(Op_MemBarCPUOrder);
4326 }
4327 }
4328
4329 //------------------------inline_native_clone----------------------------
4330 // protected native Object java.lang.Object.clone();
4331 //
4332 // Here are the simple edge cases:
4333 // null receiver => normal trap
4334 // virtual and clone was overridden => slow path to out-of-line clone
4335 // not cloneable or finalizer => slow path to out-of-line Object.clone
4336 //
4337 // The general case has two steps, allocation and copying.
4338 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4339 //
4340 // Copying also has two cases, oop arrays and everything else.
4341 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4342 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4343 //
4344 // These steps fold up nicely if and when the cloned object's klass
4345 // can be sharply typed as an object array, a type array, or an instance.
4346 //
4347 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4348 PhiNode* result_val;
4349
4350 // Set the reexecute bit for the interpreter to reexecute
4351 // the bytecode that invokes Object.clone if deoptimization happens.
4352 { PreserveReexecuteState preexecs(this);
4353 jvms()->set_should_reexecute(true);
4354
4355 Node* obj = null_check_receiver();
4356 if (stopped()) return true;
4357
4358 Node* obj_klass = load_object_klass(obj);
4359 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4360 const TypeOopPtr* toop = ((tklass != NULL)
4361 ? tklass->as_instance_type()
4362 : TypeInstPtr::NOTNULL);
4363
4364 // Conservatively insert a memory barrier on all memory slices.
4365 // Do not let writes into the original float below the clone.
4366 insert_mem_bar(Op_MemBarCPUOrder);
4367
4368 // paths into result_reg:
4369 enum {
4370 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4371 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4372 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4373 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4374 PATH_LIMIT
4375 };
4376 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4377 result_val = new(C) PhiNode(result_reg,
4378 TypeInstPtr::NOTNULL);
4379 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4380 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4381 TypePtr::BOTTOM);
4382 record_for_igvn(result_reg);
4383
4384 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4385 int raw_adr_idx = Compile::AliasIdxRaw;
4386
4387 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4388 if (array_ctl != NULL) {
4389 // It's an array.
4390 PreserveJVMState pjvms(this);
4391 set_control(array_ctl);
4392 Node* obj_length = load_array_length(obj);
4393 Node* obj_size = NULL;
4394 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4395
4396 if (!use_ReduceInitialCardMarks()) {
4397 // If it is an oop array, it requires very special treatment,
4398 // because card marking is required on each card of the array.
4399 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4400 if (is_obja != NULL) {
4401 PreserveJVMState pjvms2(this);
4402 set_control(is_obja);
4403 // Generate a direct call to the right arraycopy function(s).
4404 bool disjoint_bases = true;
4405 bool length_never_negative = true;
4406 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4407 obj, intcon(0), alloc_obj, intcon(0),
4408 obj_length,
4409 disjoint_bases, length_never_negative);
4410 result_reg->init_req(_objArray_path, control());
4411 result_val->init_req(_objArray_path, alloc_obj);
4412 result_i_o ->set_req(_objArray_path, i_o());
4413 result_mem ->set_req(_objArray_path, reset_memory());
4414 }
4415 }
4416 // Otherwise, there are no card marks to worry about.
4417 // (We can dispense with card marks if we know the allocation
4418 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4419 // causes the non-eden paths to take compensating steps to
4420 // simulate a fresh allocation, so that no further
4421 // card marks are required in compiled code to initialize
4422 // the object.)
4423
4424 if (!stopped()) {
4425 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4426
4427 // Present the results of the copy.
4428 result_reg->init_req(_array_path, control());
4429 result_val->init_req(_array_path, alloc_obj);
4430 result_i_o ->set_req(_array_path, i_o());
4431 result_mem ->set_req(_array_path, reset_memory());
4432 }
4433 }
4434
4435 // We only go to the instance fast case code if we pass a number of guards.
4436 // The paths which do not pass are accumulated in the slow_region.
4437 RegionNode* slow_region = new (C) RegionNode(1);
4438 record_for_igvn(slow_region);
4439 if (!stopped()) {
4440 // It's an instance (we did array above). Make the slow-path tests.
4441 // If this is a virtual call, we generate a funny guard. We grab
4442 // the vtable entry corresponding to clone() from the target object.
4443 // If the target method which we are calling happens to be the
4444 // Object clone() method, we pass the guard. We do not need this
4445 // guard for non-virtual calls; the caller is known to be the native
4446 // Object clone().
4447 if (is_virtual) {
4448 generate_virtual_guard(obj_klass, slow_region);
4449 }
4450
4451 // The object must be cloneable and must not have a finalizer.
4452 // Both of these conditions may be checked in a single test.
4453 // We could optimize the cloneable test further, but we don't care.
4454 generate_access_flags_guard(obj_klass,
4455 // Test both conditions:
4456 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4457 // Must be cloneable but not finalizer:
4458 JVM_ACC_IS_CLONEABLE,
4459 slow_region);
4460 }
4461
4462 if (!stopped()) {
4463 // It's an instance, and it passed the slow-path tests.
4464 PreserveJVMState pjvms(this);
4465 Node* obj_size = NULL;
4466 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4467
4468 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4469
4470 // Present the results of the slow call.
4471 result_reg->init_req(_instance_path, control());
4472 result_val->init_req(_instance_path, alloc_obj);
4473 result_i_o ->set_req(_instance_path, i_o());
4474 result_mem ->set_req(_instance_path, reset_memory());
4475 }
4476
4477 // Generate code for the slow case. We make a call to clone().
4478 set_control(_gvn.transform(slow_region));
4479 if (!stopped()) {
4480 PreserveJVMState pjvms(this);
4481 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4482 Node* slow_result = set_results_for_java_call(slow_call);
4483 // this->control() comes from set_results_for_java_call
4484 result_reg->init_req(_slow_path, control());
4485 result_val->init_req(_slow_path, slow_result);
4486 result_i_o ->set_req(_slow_path, i_o());
4487 result_mem ->set_req(_slow_path, reset_memory());
4488 }
4489
4490 // Return the combined state.
4491 set_control( _gvn.transform(result_reg));
4492 set_i_o( _gvn.transform(result_i_o));
4493 set_all_memory( _gvn.transform(result_mem));
4494 } // original reexecute is set back here
4495
4496 set_result(_gvn.transform(result_val));
4497 return true;
4498 }
4499
4500 //------------------------------basictype2arraycopy----------------------------
4501 address LibraryCallKit::basictype2arraycopy(BasicType t,
4502 Node* src_offset,
4503 Node* dest_offset,
4504 bool disjoint_bases,
4505 const char* &name,
4506 bool dest_uninitialized) {
4507 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4508 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4509
4510 bool aligned = false;
4511 bool disjoint = disjoint_bases;
4512
4513 // if the offsets are the same, we can treat the memory regions as
4514 // disjoint, because either the memory regions are in different arrays,
4515 // or they are identical (which we can treat as disjoint.) We can also
4516 // treat a copy with a destination index less that the source index
4517 // as disjoint since a low->high copy will work correctly in this case.
4518 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4519 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4520 // both indices are constants
4521 int s_offs = src_offset_inttype->get_con();
4522 int d_offs = dest_offset_inttype->get_con();
4523 int element_size = type2aelembytes(t);
4524 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4525 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4526 if (s_offs >= d_offs) disjoint = true;
4527 } else if (src_offset == dest_offset && src_offset != NULL) {
4528 // This can occur if the offsets are identical non-constants.
4529 disjoint = true;
4530 }
4531
4532 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4533 }
4534
4535
4536 //------------------------------inline_arraycopy-----------------------
4537 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4538 // Object dest, int destPos,
4539 // int length);
4540 bool LibraryCallKit::inline_arraycopy() {
4541 // Get the arguments.
4542 Node* src = argument(0); // type: oop
4543 Node* src_offset = argument(1); // type: int
4544 Node* dest = argument(2); // type: oop
4545 Node* dest_offset = argument(3); // type: int
4546 Node* length = argument(4); // type: int
4547
4548 // Compile time checks. If any of these checks cannot be verified at compile time,
4549 // we do not make a fast path for this call. Instead, we let the call remain as it
4550 // is. The checks we choose to mandate at compile time are:
4551 //
4552 // (1) src and dest are arrays.
4553 const Type* src_type = src->Value(&_gvn);
4554 const Type* dest_type = dest->Value(&_gvn);
4555 const TypeAryPtr* top_src = src_type->isa_aryptr();
4556 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4557
4558 // Do we have the type of src?
4559 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4560 // Do we have the type of dest?
4561 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4562 // Is the type for src from speculation?
4563 bool src_spec = false;
4564 // Is the type for dest from speculation?
4565 bool dest_spec = false;
4566
4567 if (!has_src || !has_dest) {
4568 // We don't have sufficient type information, let's see if
4569 // speculative types can help. We need to have types for both src
4570 // and dest so that it pays off.
4571
4572 // Do we already have or could we have type information for src
4573 bool could_have_src = has_src;
4574 // Do we already have or could we have type information for dest
4575 bool could_have_dest = has_dest;
4576
4577 ciKlass* src_k = NULL;
4578 if (!has_src) {
4579 src_k = src_type->is_oopptr()->speculative_type();
4580 if (src_k != NULL && src_k->is_array_klass()) {
4581 could_have_src = true;
4582 }
4583 }
4584
4585 ciKlass* dest_k = NULL;
4586 if (!has_dest) {
4587 dest_k = dest_type->speculative_type();
4588 if (dest_k != NULL && dest_k->is_array_klass()) {
4589 could_have_dest = true;
4590 }
4591 }
4592
4593 if (could_have_src && could_have_dest) {
4594 // This is going to pay off so emit the required guards
4595 if (!has_src) {
4596 src = maybe_cast_profiled_obj(src, src_k);
4597 src_type = _gvn.type(src);
4598 top_src = src_type->isa_aryptr();
4599 has_src = (top_src != NULL && top_src->klass() != NULL);
4600 src_spec = true;
4601 }
4602 if (!has_dest) {
4603 dest = maybe_cast_profiled_obj(dest, dest_k);
4604 dest_type = _gvn.type(dest);
4605 top_dest = dest_type->isa_aryptr();
4606 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4607 dest_spec = true;
4608 }
4609 }
4610 }
4611
4612 if (!has_src || !has_dest) {
4613 // Conservatively insert a memory barrier on all memory slices.
4614 // Do not let writes into the source float below the arraycopy.
4615 insert_mem_bar(Op_MemBarCPUOrder);
4616
4617 // Call StubRoutines::generic_arraycopy stub.
4618 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4619 src, src_offset, dest, dest_offset, length);
4620
4621 // Do not let reads from the destination float above the arraycopy.
4622 // Since we cannot type the arrays, we don't know which slices
4623 // might be affected. We could restrict this barrier only to those
4624 // memory slices which pertain to array elements--but don't bother.
4625 if (!InsertMemBarAfterArraycopy)
4626 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4627 insert_mem_bar(Op_MemBarCPUOrder);
4628 return true;
4629 }
4630
4631 // (2) src and dest arrays must have elements of the same BasicType
4632 // Figure out the size and type of the elements we will be copying.
4633 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4634 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4635 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4636 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4637
4638 if (src_elem != dest_elem || dest_elem == T_VOID) {
4639 // The component types are not the same or are not recognized. Punt.
4640 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4641 generate_slow_arraycopy(TypePtr::BOTTOM,
4642 src, src_offset, dest, dest_offset, length,
4643 /*dest_uninitialized*/false);
4644 return true;
4645 }
4646
4647 if (src_elem == T_OBJECT) {
4648 // If both arrays are object arrays then having the exact types
4649 // for both will remove the need for a subtype check at runtime
4650 // before the call and may make it possible to pick a faster copy
4651 // routine (without a subtype check on every element)
4652 // Do we have the exact type of src?
4653 bool could_have_src = src_spec;
4654 // Do we have the exact type of dest?
4655 bool could_have_dest = dest_spec;
4656 ciKlass* src_k = top_src->klass(), *dest_k = top_dest->klass();
4657 if (!src_spec) {
4658 src_k = src_type->speculative_type();
4659 if (src_k != NULL && src_k->is_array_klass()) {
4660 could_have_src = true;
4661 }
4662 }
4663 if (!dest_spec) {
4664 dest_k = dest_type->speculative_type();
4665 if (dest_k != NULL && dest_k->is_array_klass()) {
4666 could_have_dest = true;
4667 }
4668 }
4669 if (could_have_src && could_have_dest) {
4670 // If we can have both exact types, emit the missing guards
4671 if (could_have_src && !src_spec) {
4672 src = maybe_cast_profiled_obj(src, src_k);
4673 }
4674 if (could_have_dest && !dest_spec) {
4675 dest = maybe_cast_profiled_obj(dest, dest_k);
4676 }
4677 }
4678 }
4679
4680 //---------------------------------------------------------------------------
4681 // We will make a fast path for this call to arraycopy.
4682
4683 // We have the following tests left to perform:
4684 //
4685 // (3) src and dest must not be null.
4686 // (4) src_offset must not be negative.
4687 // (5) dest_offset must not be negative.
4688 // (6) length must not be negative.
4689 // (7) src_offset + length must not exceed length of src.
4690 // (8) dest_offset + length must not exceed length of dest.
4691 // (9) each element of an oop array must be assignable
4692
4693 RegionNode* slow_region = new (C) RegionNode(1);
4694 record_for_igvn(slow_region);
4695
4696 // (3) operands must not be null
4697 // We currently perform our null checks with the null_check routine.
4698 // This means that the null exceptions will be reported in the caller
4699 // rather than (correctly) reported inside of the native arraycopy call.
4700 // This should be corrected, given time. We do our null check with the
4701 // stack pointer restored.
4702 src = null_check(src, T_ARRAY);
4703 dest = null_check(dest, T_ARRAY);
4704
4705 // (4) src_offset must not be negative.
4706 generate_negative_guard(src_offset, slow_region);
4707
4708 // (5) dest_offset must not be negative.
4709 generate_negative_guard(dest_offset, slow_region);
4710
4711 // (6) length must not be negative (moved to generate_arraycopy()).
4712 // generate_negative_guard(length, slow_region);
4713
4714 // (7) src_offset + length must not exceed length of src.
4715 generate_limit_guard(src_offset, length,
4716 load_array_length(src),
4717 slow_region);
4718
4719 // (8) dest_offset + length must not exceed length of dest.
4720 generate_limit_guard(dest_offset, length,
4721 load_array_length(dest),
4722 slow_region);
4723
4724 // (9) each element of an oop array must be assignable
4725 // The generate_arraycopy subroutine checks this.
4726
4727 // This is where the memory effects are placed:
4728 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4729 generate_arraycopy(adr_type, dest_elem,
4730 src, src_offset, dest, dest_offset, length,
4731 false, false, slow_region);
4732
4733 return true;
4734 }
4735
4736 //-----------------------------generate_arraycopy----------------------
4737 // Generate an optimized call to arraycopy.
4738 // Caller must guard against non-arrays.
4739 // Caller must determine a common array basic-type for both arrays.
4740 // Caller must validate offsets against array bounds.
4741 // The slow_region has already collected guard failure paths
4742 // (such as out of bounds length or non-conformable array types).
4743 // The generated code has this shape, in general:
4744 //
4745 // if (length == 0) return // via zero_path
4746 // slowval = -1
4747 // if (types unknown) {
4748 // slowval = call generic copy loop
4749 // if (slowval == 0) return // via checked_path
4750 // } else if (indexes in bounds) {
4751 // if ((is object array) && !(array type check)) {
4752 // slowval = call checked copy loop
4753 // if (slowval == 0) return // via checked_path
4754 // } else {
4755 // call bulk copy loop
4756 // return // via fast_path
4757 // }
4758 // }
4759 // // adjust params for remaining work:
4760 // if (slowval != -1) {
4761 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4762 // }
4763 // slow_region:
4764 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4765 // return // via slow_call_path
4766 //
4767 // This routine is used from several intrinsics: System.arraycopy,
4768 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4769 //
4770 void
4771 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4772 BasicType basic_elem_type,
4773 Node* src, Node* src_offset,
4774 Node* dest, Node* dest_offset,
4775 Node* copy_length,
4776 bool disjoint_bases,
4777 bool length_never_negative,
4778 RegionNode* slow_region) {
4779
4780 if (slow_region == NULL) {
4781 slow_region = new(C) RegionNode(1);
4782 record_for_igvn(slow_region);
4783 }
4784
4785 Node* original_dest = dest;
4786 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4787 bool dest_uninitialized = false;
4788
4789 // See if this is the initialization of a newly-allocated array.
4790 // If so, we will take responsibility here for initializing it to zero.
4791 // (Note: Because tightly_coupled_allocation performs checks on the
4792 // out-edges of the dest, we need to avoid making derived pointers
4793 // from it until we have checked its uses.)
4794 if (ReduceBulkZeroing
4795 && !ZeroTLAB // pointless if already zeroed
4796 && basic_elem_type != T_CONFLICT // avoid corner case
4797 && !src->eqv_uncast(dest)
4798 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4799 != NULL)
4800 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4801 && alloc->maybe_set_complete(&_gvn)) {
4802 // "You break it, you buy it."
4803 InitializeNode* init = alloc->initialization();
4804 assert(init->is_complete(), "we just did this");
4805 init->set_complete_with_arraycopy();
4806 assert(dest->is_CheckCastPP(), "sanity");
4807 assert(dest->in(0)->in(0) == init, "dest pinned");
4808 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4809 // From this point on, every exit path is responsible for
4810 // initializing any non-copied parts of the object to zero.
4811 // Also, if this flag is set we make sure that arraycopy interacts properly
4812 // with G1, eliding pre-barriers. See CR 6627983.
4813 dest_uninitialized = true;
4814 } else {
4815 // No zeroing elimination here.
4816 alloc = NULL;
4817 //original_dest = dest;
4818 //dest_uninitialized = false;
4819 }
4820
4821 // Results are placed here:
4822 enum { fast_path = 1, // normal void-returning assembly stub
4823 checked_path = 2, // special assembly stub with cleanup
4824 slow_call_path = 3, // something went wrong; call the VM
4825 zero_path = 4, // bypass when length of copy is zero
4826 bcopy_path = 5, // copy primitive array by 64-bit blocks
4827 PATH_LIMIT = 6
4828 };
4829 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4830 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4831 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4832 record_for_igvn(result_region);
4833 _gvn.set_type_bottom(result_i_o);
4834 _gvn.set_type_bottom(result_memory);
4835 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4836
4837 // The slow_control path:
4838 Node* slow_control;
4839 Node* slow_i_o = i_o();
4840 Node* slow_mem = memory(adr_type);
4841 debug_only(slow_control = (Node*) badAddress);
4842
4843 // Checked control path:
4844 Node* checked_control = top();
4845 Node* checked_mem = NULL;
4846 Node* checked_i_o = NULL;
4847 Node* checked_value = NULL;
4848
4849 if (basic_elem_type == T_CONFLICT) {
4850 assert(!dest_uninitialized, "");
4851 Node* cv = generate_generic_arraycopy(adr_type,
4852 src, src_offset, dest, dest_offset,
4853 copy_length, dest_uninitialized);
4854 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4855 checked_control = control();
4856 checked_i_o = i_o();
4857 checked_mem = memory(adr_type);
4858 checked_value = cv;
4859 set_control(top()); // no fast path
4860 }
4861
4862 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4863 if (not_pos != NULL) {
4864 PreserveJVMState pjvms(this);
4865 set_control(not_pos);
4866
4867 // (6) length must not be negative.
4868 if (!length_never_negative) {
4869 generate_negative_guard(copy_length, slow_region);
4870 }
4871
4872 // copy_length is 0.
4873 if (!stopped() && dest_uninitialized) {
4874 Node* dest_length = alloc->in(AllocateNode::ALength);
4875 if (copy_length->eqv_uncast(dest_length)
4876 || _gvn.find_int_con(dest_length, 1) <= 0) {
4877 // There is no zeroing to do. No need for a secondary raw memory barrier.
4878 } else {
4879 // Clear the whole thing since there are no source elements to copy.
4880 generate_clear_array(adr_type, dest, basic_elem_type,
4881 intcon(0), NULL,
4882 alloc->in(AllocateNode::AllocSize));
4883 // Use a secondary InitializeNode as raw memory barrier.
4884 // Currently it is needed only on this path since other
4885 // paths have stub or runtime calls as raw memory barriers.
4886 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4887 Compile::AliasIdxRaw,
4888 top())->as_Initialize();
4889 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4890 }
4891 }
4892
4893 // Present the results of the fast call.
4894 result_region->init_req(zero_path, control());
4895 result_i_o ->init_req(zero_path, i_o());
4896 result_memory->init_req(zero_path, memory(adr_type));
4897 }
4898
4899 if (!stopped() && dest_uninitialized) {
4900 // We have to initialize the *uncopied* part of the array to zero.
4901 // The copy destination is the slice dest[off..off+len]. The other slices
4902 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4903 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4904 Node* dest_length = alloc->in(AllocateNode::ALength);
4905 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4906 copy_length));
4907
4908 // If there is a head section that needs zeroing, do it now.
4909 if (find_int_con(dest_offset, -1) != 0) {
4910 generate_clear_array(adr_type, dest, basic_elem_type,
4911 intcon(0), dest_offset,
4912 NULL);
4913 }
4914
4915 // Next, perform a dynamic check on the tail length.
4916 // It is often zero, and we can win big if we prove this.
4917 // There are two wins: Avoid generating the ClearArray
4918 // with its attendant messy index arithmetic, and upgrade
4919 // the copy to a more hardware-friendly word size of 64 bits.
4920 Node* tail_ctl = NULL;
4921 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4922 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
4923 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
4924 tail_ctl = generate_slow_guard(bol_lt, NULL);
4925 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4926 }
4927
4928 // At this point, let's assume there is no tail.
4929 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4930 // There is no tail. Try an upgrade to a 64-bit copy.
4931 bool didit = false;
4932 { PreserveJVMState pjvms(this);
4933 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4934 src, src_offset, dest, dest_offset,
4935 dest_size, dest_uninitialized);
4936 if (didit) {
4937 // Present the results of the block-copying fast call.
4938 result_region->init_req(bcopy_path, control());
4939 result_i_o ->init_req(bcopy_path, i_o());
4940 result_memory->init_req(bcopy_path, memory(adr_type));
4941 }
4942 }
4943 if (didit)
4944 set_control(top()); // no regular fast path
4945 }
4946
4947 // Clear the tail, if any.
4948 if (tail_ctl != NULL) {
4949 Node* notail_ctl = stopped() ? NULL : control();
4950 set_control(tail_ctl);
4951 if (notail_ctl == NULL) {
4952 generate_clear_array(adr_type, dest, basic_elem_type,
4953 dest_tail, NULL,
4954 dest_size);
4955 } else {
4956 // Make a local merge.
4957 Node* done_ctl = new(C) RegionNode(3);
4958 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4959 done_ctl->init_req(1, notail_ctl);
4960 done_mem->init_req(1, memory(adr_type));
4961 generate_clear_array(adr_type, dest, basic_elem_type,
4962 dest_tail, NULL,
4963 dest_size);
4964 done_ctl->init_req(2, control());
4965 done_mem->init_req(2, memory(adr_type));
4966 set_control( _gvn.transform(done_ctl));
4967 set_memory( _gvn.transform(done_mem), adr_type );
4968 }
4969 }
4970 }
4971
4972 BasicType copy_type = basic_elem_type;
4973 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4974 if (!stopped() && copy_type == T_OBJECT) {
4975 // If src and dest have compatible element types, we can copy bits.
4976 // Types S[] and D[] are compatible if D is a supertype of S.
4977 //
4978 // If they are not, we will use checked_oop_disjoint_arraycopy,
4979 // which performs a fast optimistic per-oop check, and backs off
4980 // further to JVM_ArrayCopy on the first per-oop check that fails.
4981 // (Actually, we don't move raw bits only; the GC requires card marks.)
4982
4983 // Get the Klass* for both src and dest
4984 Node* src_klass = load_object_klass(src);
4985 Node* dest_klass = load_object_klass(dest);
4986
4987 // Generate the subtype check.
4988 // This might fold up statically, or then again it might not.
4989 //
4990 // Non-static example: Copying List<String>.elements to a new String[].
4991 // The backing store for a List<String> is always an Object[],
4992 // but its elements are always type String, if the generic types
4993 // are correct at the source level.
4994 //
4995 // Test S[] against D[], not S against D, because (probably)
4996 // the secondary supertype cache is less busy for S[] than S.
4997 // This usually only matters when D is an interface.
4998 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4999 // Plug failing path into checked_oop_disjoint_arraycopy
5000 if (not_subtype_ctrl != top()) {
5001 PreserveJVMState pjvms(this);
5002 set_control(not_subtype_ctrl);
5003 // (At this point we can assume disjoint_bases, since types differ.)
5004 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5005 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5006 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5007 Node* dest_elem_klass = _gvn.transform(n1);
5008 Node* cv = generate_checkcast_arraycopy(adr_type,
5009 dest_elem_klass,
5010 src, src_offset, dest, dest_offset,
5011 ConvI2X(copy_length), dest_uninitialized);
5012 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5013 checked_control = control();
5014 checked_i_o = i_o();
5015 checked_mem = memory(adr_type);
5016 checked_value = cv;
5017 }
5018 // At this point we know we do not need type checks on oop stores.
5019
5020 // Let's see if we need card marks:
5021 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5022 // If we do not need card marks, copy using the jint or jlong stub.
5023 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5024 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5025 "sizes agree");
5026 }
5027 }
5028
5029 if (!stopped()) {
5030 // Generate the fast path, if possible.
5031 PreserveJVMState pjvms(this);
5032 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5033 src, src_offset, dest, dest_offset,
5034 ConvI2X(copy_length), dest_uninitialized);
5035
5036 // Present the results of the fast call.
5037 result_region->init_req(fast_path, control());
5038 result_i_o ->init_req(fast_path, i_o());
5039 result_memory->init_req(fast_path, memory(adr_type));
5040 }
5041
5042 // Here are all the slow paths up to this point, in one bundle:
5043 slow_control = top();
5044 if (slow_region != NULL)
5045 slow_control = _gvn.transform(slow_region);
5046 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5047
5048 set_control(checked_control);
5049 if (!stopped()) {
5050 // Clean up after the checked call.
5051 // The returned value is either 0 or -1^K,
5052 // where K = number of partially transferred array elements.
5053 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5054 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5055 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5056
5057 // If it is 0, we are done, so transfer to the end.
5058 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5059 result_region->init_req(checked_path, checks_done);
5060 result_i_o ->init_req(checked_path, checked_i_o);
5061 result_memory->init_req(checked_path, checked_mem);
5062
5063 // If it is not zero, merge into the slow call.
5064 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5065 RegionNode* slow_reg2 = new(C) RegionNode(3);
5066 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5067 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5068 record_for_igvn(slow_reg2);
5069 slow_reg2 ->init_req(1, slow_control);
5070 slow_i_o2 ->init_req(1, slow_i_o);
5071 slow_mem2 ->init_req(1, slow_mem);
5072 slow_reg2 ->init_req(2, control());
5073 slow_i_o2 ->init_req(2, checked_i_o);
5074 slow_mem2 ->init_req(2, checked_mem);
5075
5076 slow_control = _gvn.transform(slow_reg2);
5077 slow_i_o = _gvn.transform(slow_i_o2);
5078 slow_mem = _gvn.transform(slow_mem2);
5079
5080 if (alloc != NULL) {
5081 // We'll restart from the very beginning, after zeroing the whole thing.
5082 // This can cause double writes, but that's OK since dest is brand new.
5083 // So we ignore the low 31 bits of the value returned from the stub.
5084 } else {
5085 // We must continue the copy exactly where it failed, or else
5086 // another thread might see the wrong number of writes to dest.
5087 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5088 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5089 slow_offset->init_req(1, intcon(0));
5090 slow_offset->init_req(2, checked_offset);
5091 slow_offset = _gvn.transform(slow_offset);
5092
5093 // Adjust the arguments by the conditionally incoming offset.
5094 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5095 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5096 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5097
5098 // Tweak the node variables to adjust the code produced below:
5099 src_offset = src_off_plus;
5100 dest_offset = dest_off_plus;
5101 copy_length = length_minus;
5102 }
5103 }
5104
5105 set_control(slow_control);
5106 if (!stopped()) {
5107 // Generate the slow path, if needed.
5108 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5109
5110 set_memory(slow_mem, adr_type);
5111 set_i_o(slow_i_o);
5112
5113 if (dest_uninitialized) {
5114 generate_clear_array(adr_type, dest, basic_elem_type,
5115 intcon(0), NULL,
5116 alloc->in(AllocateNode::AllocSize));
5117 }
5118
5119 generate_slow_arraycopy(adr_type,
5120 src, src_offset, dest, dest_offset,
5121 copy_length, /*dest_uninitialized*/false);
5122
5123 result_region->init_req(slow_call_path, control());
5124 result_i_o ->init_req(slow_call_path, i_o());
5125 result_memory->init_req(slow_call_path, memory(adr_type));
5126 }
5127
5128 // Remove unused edges.
5129 for (uint i = 1; i < result_region->req(); i++) {
5130 if (result_region->in(i) == NULL)
5131 result_region->init_req(i, top());
5132 }
5133
5134 // Finished; return the combined state.
5135 set_control( _gvn.transform(result_region));
5136 set_i_o( _gvn.transform(result_i_o) );
5137 set_memory( _gvn.transform(result_memory), adr_type );
5138
5139 // The memory edges above are precise in order to model effects around
5140 // array copies accurately to allow value numbering of field loads around
5141 // arraycopy. Such field loads, both before and after, are common in Java
5142 // collections and similar classes involving header/array data structures.
5143 //
5144 // But with low number of register or when some registers are used or killed
5145 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5146 // The next memory barrier is added to avoid it. If the arraycopy can be
5147 // optimized away (which it can, sometimes) then we can manually remove
5148 // the membar also.
5149 //
5150 // Do not let reads from the cloned object float above the arraycopy.
5151 if (alloc != NULL) {
5152 // Do not let stores that initialize this object be reordered with
5153 // a subsequent store that would make this object accessible by
5154 // other threads.
5155 // Record what AllocateNode this StoreStore protects so that
5156 // escape analysis can go from the MemBarStoreStoreNode to the
5157 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5158 // based on the escape status of the AllocateNode.
5159 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5160 } else if (InsertMemBarAfterArraycopy)
5161 insert_mem_bar(Op_MemBarCPUOrder);
5162 }
5163
5164
5165 // Helper function which determines if an arraycopy immediately follows
5166 // an allocation, with no intervening tests or other escapes for the object.
5167 AllocateArrayNode*
5168 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5169 RegionNode* slow_region) {
5170 if (stopped()) return NULL; // no fast path
5171 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5172
5173 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5174 if (alloc == NULL) return NULL;
5175
5176 Node* rawmem = memory(Compile::AliasIdxRaw);
5177 // Is the allocation's memory state untouched?
5178 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5179 // Bail out if there have been raw-memory effects since the allocation.
5180 // (Example: There might have been a call or safepoint.)
5181 return NULL;
5182 }
5183 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5184 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5185 return NULL;
5186 }
5187
5188 // There must be no unexpected observers of this allocation.
5189 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5190 Node* obs = ptr->fast_out(i);
5191 if (obs != this->map()) {
5192 return NULL;
5193 }
5194 }
5195
5196 // This arraycopy must unconditionally follow the allocation of the ptr.
5197 Node* alloc_ctl = ptr->in(0);
5198 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5199
5200 Node* ctl = control();
5201 while (ctl != alloc_ctl) {
5202 // There may be guards which feed into the slow_region.
5203 // Any other control flow means that we might not get a chance
5204 // to finish initializing the allocated object.
5205 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5206 IfNode* iff = ctl->in(0)->as_If();
5207 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5208 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5209 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5210 ctl = iff->in(0); // This test feeds the known slow_region.
5211 continue;
5212 }
5213 // One more try: Various low-level checks bottom out in
5214 // uncommon traps. If the debug-info of the trap omits
5215 // any reference to the allocation, as we've already
5216 // observed, then there can be no objection to the trap.
5217 bool found_trap = false;
5218 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5219 Node* obs = not_ctl->fast_out(j);
5220 if (obs->in(0) == not_ctl && obs->is_Call() &&
5221 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5222 found_trap = true; break;
5223 }
5224 }
5225 if (found_trap) {
5226 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5227 continue;
5228 }
5229 }
5230 return NULL;
5231 }
5232
5233 // If we get this far, we have an allocation which immediately
5234 // precedes the arraycopy, and we can take over zeroing the new object.
5235 // The arraycopy will finish the initialization, and provide
5236 // a new control state to which we will anchor the destination pointer.
5237
5238 return alloc;
5239 }
5240
5241 // Helper for initialization of arrays, creating a ClearArray.
5242 // It writes zero bits in [start..end), within the body of an array object.
5243 // The memory effects are all chained onto the 'adr_type' alias category.
5244 //
5245 // Since the object is otherwise uninitialized, we are free
5246 // to put a little "slop" around the edges of the cleared area,
5247 // as long as it does not go back into the array's header,
5248 // or beyond the array end within the heap.
5249 //
5250 // The lower edge can be rounded down to the nearest jint and the
5251 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5252 //
5253 // Arguments:
5254 // adr_type memory slice where writes are generated
5255 // dest oop of the destination array
5256 // basic_elem_type element type of the destination
5257 // slice_idx array index of first element to store
5258 // slice_len number of elements to store (or NULL)
5259 // dest_size total size in bytes of the array object
5260 //
5261 // Exactly one of slice_len or dest_size must be non-NULL.
5262 // If dest_size is non-NULL, zeroing extends to the end of the object.
5263 // If slice_len is non-NULL, the slice_idx value must be a constant.
5264 void
5265 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5266 Node* dest,
5267 BasicType basic_elem_type,
5268 Node* slice_idx,
5269 Node* slice_len,
5270 Node* dest_size) {
5271 // one or the other but not both of slice_len and dest_size:
5272 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5273 if (slice_len == NULL) slice_len = top();
5274 if (dest_size == NULL) dest_size = top();
5275
5276 // operate on this memory slice:
5277 Node* mem = memory(adr_type); // memory slice to operate on
5278
5279 // scaling and rounding of indexes:
5280 int scale = exact_log2(type2aelembytes(basic_elem_type));
5281 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5282 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5283 int bump_bit = (-1 << scale) & BytesPerInt;
5284
5285 // determine constant starts and ends
5286 const intptr_t BIG_NEG = -128;
5287 assert(BIG_NEG + 2*abase < 0, "neg enough");
5288 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5289 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5290 if (slice_len_con == 0) {
5291 return; // nothing to do here
5292 }
5293 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5294 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5295 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5296 assert(end_con < 0, "not two cons");
5297 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5298 BytesPerLong);
5299 }
5300
5301 if (start_con >= 0 && end_con >= 0) {
5302 // Constant start and end. Simple.
5303 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5304 start_con, end_con, &_gvn);
5305 } else if (start_con >= 0 && dest_size != top()) {
5306 // Constant start, pre-rounded end after the tail of the array.
5307 Node* end = dest_size;
5308 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5309 start_con, end, &_gvn);
5310 } else if (start_con >= 0 && slice_len != top()) {
5311 // Constant start, non-constant end. End needs rounding up.
5312 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5313 intptr_t end_base = abase + (slice_idx_con << scale);
5314 int end_round = (-1 << scale) & (BytesPerLong - 1);
5315 Node* end = ConvI2X(slice_len);
5316 if (scale != 0)
5317 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5318 end_base += end_round;
5319 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5320 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5321 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5322 start_con, end, &_gvn);
5323 } else if (start_con < 0 && dest_size != top()) {
5324 // Non-constant start, pre-rounded end after the tail of the array.
5325 // This is almost certainly a "round-to-end" operation.
5326 Node* start = slice_idx;
5327 start = ConvI2X(start);
5328 if (scale != 0)
5329 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5330 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5331 if ((bump_bit | clear_low) != 0) {
5332 int to_clear = (bump_bit | clear_low);
5333 // Align up mod 8, then store a jint zero unconditionally
5334 // just before the mod-8 boundary.
5335 if (((abase + bump_bit) & ~to_clear) - bump_bit
5336 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5337 bump_bit = 0;
5338 assert((abase & to_clear) == 0, "array base must be long-aligned");
5339 } else {
5340 // Bump 'start' up to (or past) the next jint boundary:
5341 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5342 assert((abase & clear_low) == 0, "array base must be int-aligned");
5343 }
5344 // Round bumped 'start' down to jlong boundary in body of array.
5345 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5346 if (bump_bit != 0) {
5347 // Store a zero to the immediately preceding jint:
5348 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5349 Node* p1 = basic_plus_adr(dest, x1);
5350 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5351 mem = _gvn.transform(mem);
5352 }
5353 }
5354 Node* end = dest_size; // pre-rounded
5355 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5356 start, end, &_gvn);
5357 } else {
5358 // Non-constant start, unrounded non-constant end.
5359 // (Nobody zeroes a random midsection of an array using this routine.)
5360 ShouldNotReachHere(); // fix caller
5361 }
5362
5363 // Done.
5364 set_memory(mem, adr_type);
5365 }
5366
5367
5368 bool
5369 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5370 BasicType basic_elem_type,
5371 AllocateNode* alloc,
5372 Node* src, Node* src_offset,
5373 Node* dest, Node* dest_offset,
5374 Node* dest_size, bool dest_uninitialized) {
5375 // See if there is an advantage from block transfer.
5376 int scale = exact_log2(type2aelembytes(basic_elem_type));
5377 if (scale >= LogBytesPerLong)
5378 return false; // it is already a block transfer
5379
5380 // Look at the alignment of the starting offsets.
5381 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5382
5383 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5384 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5385 if (src_off_con < 0 || dest_off_con < 0)
5386 // At present, we can only understand constants.
5387 return false;
5388
5389 intptr_t src_off = abase + (src_off_con << scale);
5390 intptr_t dest_off = abase + (dest_off_con << scale);
5391
5392 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5393 // Non-aligned; too bad.
5394 // One more chance: Pick off an initial 32-bit word.
5395 // This is a common case, since abase can be odd mod 8.
5396 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5397 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5398 Node* sptr = basic_plus_adr(src, src_off);
5399 Node* dptr = basic_plus_adr(dest, dest_off);
5400 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5401 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5402 src_off += BytesPerInt;
5403 dest_off += BytesPerInt;
5404 } else {
5405 return false;
5406 }
5407 }
5408 assert(src_off % BytesPerLong == 0, "");
5409 assert(dest_off % BytesPerLong == 0, "");
5410
5411 // Do this copy by giant steps.
5412 Node* sptr = basic_plus_adr(src, src_off);
5413 Node* dptr = basic_plus_adr(dest, dest_off);
5414 Node* countx = dest_size;
5415 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5416 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5417
5418 bool disjoint_bases = true; // since alloc != NULL
5419 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5420 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5421
5422 return true;
5423 }
5424
5425
5426 // Helper function; generates code for the slow case.
5427 // We make a call to a runtime method which emulates the native method,
5428 // but without the native wrapper overhead.
5429 void
5430 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5431 Node* src, Node* src_offset,
5432 Node* dest, Node* dest_offset,
5433 Node* copy_length, bool dest_uninitialized) {
5434 assert(!dest_uninitialized, "Invariant");
5435 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5436 OptoRuntime::slow_arraycopy_Type(),
5437 OptoRuntime::slow_arraycopy_Java(),
5438 "slow_arraycopy", adr_type,
5439 src, src_offset, dest, dest_offset,
5440 copy_length);
5441
5442 // Handle exceptions thrown by this fellow:
5443 make_slow_call_ex(call, env()->Throwable_klass(), false);
5444 }
5445
5446 // Helper function; generates code for cases requiring runtime checks.
5447 Node*
5448 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5449 Node* dest_elem_klass,
5450 Node* src, Node* src_offset,
5451 Node* dest, Node* dest_offset,
5452 Node* copy_length, bool dest_uninitialized) {
5453 if (stopped()) return NULL;
5454
5455 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5456 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5457 return NULL;
5458 }
5459
5460 // Pick out the parameters required to perform a store-check
5461 // for the target array. This is an optimistic check. It will
5462 // look in each non-null element's class, at the desired klass's
5463 // super_check_offset, for the desired klass.
5464 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5465 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5466 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5467 Node* check_offset = ConvI2X(_gvn.transform(n3));
5468 Node* check_value = dest_elem_klass;
5469
5470 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5471 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5472
5473 // (We know the arrays are never conjoint, because their types differ.)
5474 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5475 OptoRuntime::checkcast_arraycopy_Type(),
5476 copyfunc_addr, "checkcast_arraycopy", adr_type,
5477 // five arguments, of which two are
5478 // intptr_t (jlong in LP64)
5479 src_start, dest_start,
5480 copy_length XTOP,
5481 check_offset XTOP,
5482 check_value);
5483
5484 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5485 }
5486
5487
5488 // Helper function; generates code for cases requiring runtime checks.
5489 Node*
5490 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5491 Node* src, Node* src_offset,
5492 Node* dest, Node* dest_offset,
5493 Node* copy_length, bool dest_uninitialized) {
5494 assert(!dest_uninitialized, "Invariant");
5495 if (stopped()) return NULL;
5496 address copyfunc_addr = StubRoutines::generic_arraycopy();
5497 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5498 return NULL;
5499 }
5500
5501 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5502 OptoRuntime::generic_arraycopy_Type(),
5503 copyfunc_addr, "generic_arraycopy", adr_type,
5504 src, src_offset, dest, dest_offset, copy_length);
5505
5506 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5507 }
5508
5509 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5510 void
5511 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5512 BasicType basic_elem_type,
5513 bool disjoint_bases,
5514 Node* src, Node* src_offset,
5515 Node* dest, Node* dest_offset,
5516 Node* copy_length, bool dest_uninitialized) {
5517 if (stopped()) return; // nothing to do
5518
5519 Node* src_start = src;
5520 Node* dest_start = dest;
5521 if (src_offset != NULL || dest_offset != NULL) {
5522 assert(src_offset != NULL && dest_offset != NULL, "");
5523 src_start = array_element_address(src, src_offset, basic_elem_type);
5524 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5525 }
5526
5527 // Figure out which arraycopy runtime method to call.
5528 const char* copyfunc_name = "arraycopy";
5529 address copyfunc_addr =
5530 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5531 disjoint_bases, copyfunc_name, dest_uninitialized);
5532
5533 // Call it. Note that the count_ix value is not scaled to a byte-size.
5534 make_runtime_call(RC_LEAF|RC_NO_FP,
5535 OptoRuntime::fast_arraycopy_Type(),
5536 copyfunc_addr, copyfunc_name, adr_type,
5537 src_start, dest_start, copy_length XTOP);
5538 }
5539
5540 //-------------inline_encodeISOArray-----------------------------------
5541 // encode char[] to byte[] in ISO_8859_1
5542 bool LibraryCallKit::inline_encodeISOArray() {
5543 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5544 // no receiver since it is static method
5545 Node *src = argument(0);
5546 Node *src_offset = argument(1);
5547 Node *dst = argument(2);
5548 Node *dst_offset = argument(3);
5549 Node *length = argument(4);
5550
5551 const Type* src_type = src->Value(&_gvn);
5552 const Type* dst_type = dst->Value(&_gvn);
5553 const TypeAryPtr* top_src = src_type->isa_aryptr();
5554 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5555 if (top_src == NULL || top_src->klass() == NULL ||
5556 top_dest == NULL || top_dest->klass() == NULL) {
5557 // failed array check
5558 return false;
5559 }
5560
5561 // Figure out the size and type of the elements we will be copying.
5562 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5563 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5564 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5565 return false;
5566 }
5567 Node* src_start = array_element_address(src, src_offset, src_elem);
5568 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5569 // 'src_start' points to src array + scaled offset
5570 // 'dst_start' points to dst array + scaled offset
5571
5572 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5573 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5574 enc = _gvn.transform(enc);
5575 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5576 set_memory(res_mem, mtype);
5577 set_result(enc);
5578 return true;
5579 }
5580
5581 /**
5582 * Calculate CRC32 for byte.
5583 * int java.util.zip.CRC32.update(int crc, int b)
5584 */
5585 bool LibraryCallKit::inline_updateCRC32() {
5586 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5587 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5588 // no receiver since it is static method
5589 Node* crc = argument(0); // type: int
5590 Node* b = argument(1); // type: int
5591
5592 /*
5593 * int c = ~ crc;
5594 * b = timesXtoThe32[(b ^ c) & 0xFF];
5595 * b = b ^ (c >>> 8);
5596 * crc = ~b;
5597 */
5598
5599 Node* M1 = intcon(-1);
5600 crc = _gvn.transform(new (C) XorINode(crc, M1));
5601 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5602 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5603
5604 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5605 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5606 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5607 result = make_load(control(), adr, TypeInt::INT, T_INT);
5608
5609 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5610 result = _gvn.transform(new (C) XorINode(crc, result));
5611 result = _gvn.transform(new (C) XorINode(result, M1));
5612 set_result(result);
5613 return true;
5614 }
5615
5616 /**
5617 * Calculate CRC32 for byte[] array.
5618 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5619 */
5620 bool LibraryCallKit::inline_updateBytesCRC32() {
5621 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5622 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5623 // no receiver since it is static method
5624 Node* crc = argument(0); // type: int
5625 Node* src = argument(1); // type: oop
5626 Node* offset = argument(2); // type: int
5627 Node* length = argument(3); // type: int
5628
5629 const Type* src_type = src->Value(&_gvn);
5630 const TypeAryPtr* top_src = src_type->isa_aryptr();
5631 if (top_src == NULL || top_src->klass() == NULL) {
5632 // failed array check
5633 return false;
5634 }
5635
5636 // Figure out the size and type of the elements we will be copying.
5637 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5638 if (src_elem != T_BYTE) {
5639 return false;
5640 }
5641
5642 // 'src_start' points to src array + scaled offset
5643 Node* src_start = array_element_address(src, offset, src_elem);
5644
5645 // We assume that range check is done by caller.
5646 // TODO: generate range check (offset+length < src.length) in debug VM.
5647
5648 // Call the stub.
5649 address stubAddr = StubRoutines::updateBytesCRC32();
5650 const char *stubName = "updateBytesCRC32";
5651
5652 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5653 stubAddr, stubName, TypePtr::BOTTOM,
5654 crc, src_start, length);
5655 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5656 set_result(result);
5657 return true;
5658 }
5659
5660 /**
5661 * Calculate CRC32 for ByteBuffer.
5662 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5663 */
5664 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5665 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5666 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5667 // no receiver since it is static method
5668 Node* crc = argument(0); // type: int
5669 Node* src = argument(1); // type: long
5670 Node* offset = argument(3); // type: int
5671 Node* length = argument(4); // type: int
5672
5673 src = ConvL2X(src); // adjust Java long to machine word
5674 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5675 offset = ConvI2X(offset);
5676
5677 // 'src_start' points to src array + scaled offset
5678 Node* src_start = basic_plus_adr(top(), base, offset);
5679
5680 // Call the stub.
5681 address stubAddr = StubRoutines::updateBytesCRC32();
5682 const char *stubName = "updateBytesCRC32";
5683
5684 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5685 stubAddr, stubName, TypePtr::BOTTOM,
5686 crc, src_start, length);
5687 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5688 set_result(result);
5689 return true;
5690 }
5691
5692 //----------------------------inline_reference_get----------------------------
5693 // public T java.lang.ref.Reference.get();
5694 bool LibraryCallKit::inline_reference_get() {
5695 const int referent_offset = java_lang_ref_Reference::referent_offset;
5696 guarantee(referent_offset > 0, "should have already been set");
5697
5698 // Get the argument:
5699 Node* reference_obj = null_check_receiver();
5700 if (stopped()) return true;
5701
5702 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5703
5704 ciInstanceKlass* klass = env()->Object_klass();
5705 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5706
5707 Node* no_ctrl = NULL;
5708 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5709
5710 // Use the pre-barrier to record the value in the referent field
5711 pre_barrier(false /* do_load */,
5712 control(),
5713 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5714 result /* pre_val */,
5715 T_OBJECT);
5716
5717 // Add memory barrier to prevent commoning reads from this field
5718 // across safepoint since GC can change its value.
5719 insert_mem_bar(Op_MemBarCPUOrder);
5720
5721 set_result(result);
5722 return true;
5723 }
5724
5725
5726 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5727 bool is_exact=true, bool is_static=false) {
5728
5729 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5730 assert(tinst != NULL, "obj is null");
5731 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5732 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5733
5734 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5735 ciSymbol::make(fieldTypeString),
5736 is_static);
5737 if (field == NULL) return (Node *) NULL;
5738 assert (field != NULL, "undefined field");
5739
5740 // Next code copied from Parse::do_get_xxx():
5741
5742 // Compute address and memory type.
5743 int offset = field->offset_in_bytes();
5744 bool is_vol = field->is_volatile();
5745 ciType* field_klass = field->type();
5746 assert(field_klass->is_loaded(), "should be loaded");
5747 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5748 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5749 BasicType bt = field->layout_type();
5750
5751 // Build the resultant type of the load
5752 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5753
5754 // Build the load.
5755 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5756 return loadedField;
5757 }
5758
5759
5760 //------------------------------inline_aescrypt_Block-----------------------
5761 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5762 address stubAddr;
5763 const char *stubName;
5764 assert(UseAES, "need AES instruction support");
5765
5766 switch(id) {
5767 case vmIntrinsics::_aescrypt_encryptBlock:
5768 stubAddr = StubRoutines::aescrypt_encryptBlock();
5769 stubName = "aescrypt_encryptBlock";
5770 break;
5771 case vmIntrinsics::_aescrypt_decryptBlock:
5772 stubAddr = StubRoutines::aescrypt_decryptBlock();
5773 stubName = "aescrypt_decryptBlock";
5774 break;
5775 }
5776 if (stubAddr == NULL) return false;
5777
5778 Node* aescrypt_object = argument(0);
5779 Node* src = argument(1);
5780 Node* src_offset = argument(2);
5781 Node* dest = argument(3);
5782 Node* dest_offset = argument(4);
5783
5784 // (1) src and dest are arrays.
5785 const Type* src_type = src->Value(&_gvn);
5786 const Type* dest_type = dest->Value(&_gvn);
5787 const TypeAryPtr* top_src = src_type->isa_aryptr();
5788 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5789 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5790
5791 // for the quick and dirty code we will skip all the checks.
5792 // we are just trying to get the call to be generated.
5793 Node* src_start = src;
5794 Node* dest_start = dest;
5795 if (src_offset != NULL || dest_offset != NULL) {
5796 assert(src_offset != NULL && dest_offset != NULL, "");
5797 src_start = array_element_address(src, src_offset, T_BYTE);
5798 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5799 }
5800
5801 // now need to get the start of its expanded key array
5802 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5803 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5804 if (k_start == NULL) return false;
5805
5806 // Call the stub.
5807 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5808 stubAddr, stubName, TypePtr::BOTTOM,
5809 src_start, dest_start, k_start);
5810
5811 return true;
5812 }
5813
5814 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5815 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5816 address stubAddr;
5817 const char *stubName;
5818
5819 assert(UseAES, "need AES instruction support");
5820
5821 switch(id) {
5822 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5823 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5824 stubName = "cipherBlockChaining_encryptAESCrypt";
5825 break;
5826 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5827 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5828 stubName = "cipherBlockChaining_decryptAESCrypt";
5829 break;
5830 }
5831 if (stubAddr == NULL) return false;
5832
5833 Node* cipherBlockChaining_object = argument(0);
5834 Node* src = argument(1);
5835 Node* src_offset = argument(2);
5836 Node* len = argument(3);
5837 Node* dest = argument(4);
5838 Node* dest_offset = argument(5);
5839
5840 // (1) src and dest are arrays.
5841 const Type* src_type = src->Value(&_gvn);
5842 const Type* dest_type = dest->Value(&_gvn);
5843 const TypeAryPtr* top_src = src_type->isa_aryptr();
5844 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5845 assert (top_src != NULL && top_src->klass() != NULL
5846 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5847
5848 // checks are the responsibility of the caller
5849 Node* src_start = src;
5850 Node* dest_start = dest;
5851 if (src_offset != NULL || dest_offset != NULL) {
5852 assert(src_offset != NULL && dest_offset != NULL, "");
5853 src_start = array_element_address(src, src_offset, T_BYTE);
5854 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5855 }
5856
5857 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5858 // (because of the predicated logic executed earlier).
5859 // so we cast it here safely.
5860 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5861
5862 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5863 if (embeddedCipherObj == NULL) return false;
5864
5865 // cast it to what we know it will be at runtime
5866 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5867 assert(tinst != NULL, "CBC obj is null");
5868 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5869 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5870 if (!klass_AESCrypt->is_loaded()) return false;
5871
5872 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5873 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5874 const TypeOopPtr* xtype = aklass->as_instance_type();
5875 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5876 aescrypt_object = _gvn.transform(aescrypt_object);
5877
5878 // we need to get the start of the aescrypt_object's expanded key array
5879 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5880 if (k_start == NULL) return false;
5881
5882 // similarly, get the start address of the r vector
5883 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5884 if (objRvec == NULL) return false;
5885 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5886
5887 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5888 make_runtime_call(RC_LEAF|RC_NO_FP,
5889 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5890 stubAddr, stubName, TypePtr::BOTTOM,
5891 src_start, dest_start, k_start, r_start, len);
5892
5893 // return is void so no result needs to be pushed
5894
5895 return true;
5896 }
5897
5898 //------------------------------get_key_start_from_aescrypt_object-----------------------
5899 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5900 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5901 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5902 if (objAESCryptKey == NULL) return (Node *) NULL;
5903
5904 // now have the array, need to get the start address of the K array
5905 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5906 return k_start;
5907 }
5908
5909 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5910 // Return node representing slow path of predicate check.
5911 // the pseudo code we want to emulate with this predicate is:
5912 // for encryption:
5913 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5914 // for decryption:
5915 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5916 // note cipher==plain is more conservative than the original java code but that's OK
5917 //
5918 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5919 // First, check receiver for NULL since it is virtual method.
5920 Node* objCBC = argument(0);
5921 objCBC = null_check(objCBC);
5922
5923 if (stopped()) return NULL; // Always NULL
5924
5925 // Load embeddedCipher field of CipherBlockChaining object.
5926 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5927
5928 // get AESCrypt klass for instanceOf check
5929 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5930 // will have same classloader as CipherBlockChaining object
5931 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5932 assert(tinst != NULL, "CBCobj is null");
5933 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5934
5935 // we want to do an instanceof comparison against the AESCrypt class
5936 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5937 if (!klass_AESCrypt->is_loaded()) {
5938 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5939 Node* ctrl = control();
5940 set_control(top()); // no regular fast path
5941 return ctrl;
5942 }
5943 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5944
5945 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5946 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5947 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5948
5949 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5950
5951 // for encryption, we are done
5952 if (!decrypting)
5953 return instof_false; // even if it is NULL
5954
5955 // for decryption, we need to add a further check to avoid
5956 // taking the intrinsic path when cipher and plain are the same
5957 // see the original java code for why.
5958 RegionNode* region = new(C) RegionNode(3);
5959 region->init_req(1, instof_false);
5960 Node* src = argument(1);
5961 Node* dest = argument(4);
5962 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5963 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5964 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5965 region->init_req(2, src_dest_conjoint);
5966
5967 record_for_igvn(region);
5968 return _gvn.transform(region);
5969 }