1 /*
2 * Copyright (c) 1999, 2014, 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/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "trace/traceMacros.hpp"
42
43 class LibraryIntrinsic : public InlineCallGenerator {
44 // Extend the set of intrinsics known to the runtime:
45 public:
46 private:
47 bool _is_virtual;
48 bool _is_predicted;
49 vmIntrinsics::ID _intrinsic_id;
50
51 public:
52 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id)
53 : InlineCallGenerator(m),
54 _is_virtual(is_virtual),
55 _is_predicted(is_predicted),
56 _intrinsic_id(id)
57 {
58 }
59 virtual bool is_intrinsic() const { return true; }
60 virtual bool is_virtual() const { return _is_virtual; }
61 virtual bool is_predicted() const { return _is_predicted; }
62 virtual JVMState* generate(JVMState* jvms);
63 virtual Node* generate_predicate(JVMState* jvms);
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
65 };
66
67
68 // Local helper class for LibraryIntrinsic:
69 class LibraryCallKit : public GraphKit {
70 private:
71 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
72 Node* _result; // the result node, if any
73 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
74
75 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
76
77 public:
78 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
79 : GraphKit(jvms),
80 _intrinsic(intrinsic),
81 _result(NULL)
82 {
83 // Check if this is a root compile. In that case we don't have a caller.
84 if (!jvms->has_method()) {
85 _reexecute_sp = sp();
86 } else {
87 // Find out how many arguments the interpreter needs when deoptimizing
88 // and save the stack pointer value so it can used by uncommon_trap.
89 // We find the argument count by looking at the declared signature.
90 bool ignored_will_link;
91 ciSignature* declared_signature = NULL;
92 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
93 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
94 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
95 }
96 }
97
98 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
99
100 ciMethod* caller() const { return jvms()->method(); }
101 int bci() const { return jvms()->bci(); }
102 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
103 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
104 ciMethod* callee() const { return _intrinsic->method(); }
105
106 bool try_to_inline();
107 Node* try_to_predicate();
108
109 void push_result() {
110 // Push the result onto the stack.
111 if (!stopped() && result() != NULL) {
112 BasicType bt = result()->bottom_type()->basic_type();
113 push_node(bt, result());
114 }
115 }
116
117 private:
118 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
119 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
120 }
121
122 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
123 void set_result(RegionNode* region, PhiNode* value);
124 Node* result() { return _result; }
125
126 virtual int reexecute_sp() { return _reexecute_sp; }
127
128 // Helper functions to inline natives
129 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
130 Node* generate_slow_guard(Node* test, RegionNode* region);
131 Node* generate_fair_guard(Node* test, RegionNode* region);
132 Node* generate_negative_guard(Node* index, RegionNode* region,
133 // resulting CastII of index:
134 Node* *pos_index = NULL);
135 Node* generate_nonpositive_guard(Node* index, bool never_negative,
136 // resulting CastII of index:
137 Node* *pos_index = NULL);
138 Node* generate_limit_guard(Node* offset, Node* subseq_length,
139 Node* array_length,
140 RegionNode* region);
141 Node* generate_current_thread(Node* &tls_output);
142 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
143 bool disjoint_bases, const char* &name, bool dest_uninitialized);
144 Node* load_mirror_from_klass(Node* klass);
145 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
146 RegionNode* region, int null_path,
147 int offset);
148 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
149 RegionNode* region, int null_path) {
150 int offset = java_lang_Class::klass_offset_in_bytes();
151 return load_klass_from_mirror_common(mirror, never_see_null,
152 region, null_path,
153 offset);
154 }
155 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
156 RegionNode* region, int null_path) {
157 int offset = java_lang_Class::array_klass_offset_in_bytes();
158 return load_klass_from_mirror_common(mirror, never_see_null,
159 region, null_path,
160 offset);
161 }
162 Node* generate_access_flags_guard(Node* kls,
163 int modifier_mask, int modifier_bits,
164 RegionNode* region);
165 Node* generate_interface_guard(Node* kls, RegionNode* region);
166 Node* generate_array_guard(Node* kls, RegionNode* region) {
167 return generate_array_guard_common(kls, region, false, false);
168 }
169 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
170 return generate_array_guard_common(kls, region, false, true);
171 }
172 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
173 return generate_array_guard_common(kls, region, true, false);
174 }
175 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, true, true);
177 }
178 Node* generate_array_guard_common(Node* kls, RegionNode* region,
179 bool obj_array, bool not_array);
180 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
181 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
182 bool is_virtual = false, bool is_static = false);
183 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
184 return generate_method_call(method_id, false, true);
185 }
186 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
187 return generate_method_call(method_id, true, false);
188 }
189 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
190
191 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
192 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
193 bool inline_string_compareTo();
194 bool inline_string_indexOf();
195 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
196 bool inline_string_equals();
197 Node* round_double_node(Node* n);
198 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
199 bool inline_math_native(vmIntrinsics::ID id);
200 bool inline_trig(vmIntrinsics::ID id);
201 bool inline_math(vmIntrinsics::ID id);
202 bool inline_exp();
203 bool inline_pow();
204 Node* finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
205 bool inline_min_max(vmIntrinsics::ID id);
206 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
207 // This returns Type::AnyPtr, RawPtr, or OopPtr.
208 int classify_unsafe_addr(Node* &base, Node* &offset);
209 Node* make_unsafe_address(Node* base, Node* offset);
210 // Helper for inline_unsafe_access.
211 // Generates the guards that check whether the result of
212 // Unsafe.getObject should be recorded in an SATB log buffer.
213 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
214 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
215 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
216 bool inline_unsafe_allocate();
217 bool inline_unsafe_copyMemory();
218 bool inline_native_currentThread();
219 #ifdef TRACE_HAVE_INTRINSICS
220 bool inline_native_classID();
221 bool inline_native_threadID();
222 #endif
223 bool inline_native_time_funcs(address method, const char* funcName);
224 bool inline_native_isInterrupted();
225 bool inline_native_Class_query(vmIntrinsics::ID id);
226 bool inline_native_subtype_check();
227
228 bool inline_native_newArray();
229 bool inline_native_getLength();
230 bool inline_array_copyOf(bool is_copyOfRange);
231 bool inline_array_equals();
232 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
233 bool inline_native_clone(bool is_virtual);
234 bool inline_native_Reflection_getCallerClass();
235 bool is_method_invoke_or_aux_frame(JVMState* jvms);
236 // Helper function for inlining native object hash method
237 bool inline_native_hashcode(bool is_virtual, bool is_static);
238 bool inline_native_getClass();
239
240 // Helper functions for inlining arraycopy
241 bool inline_arraycopy();
242 void generate_arraycopy(const TypePtr* adr_type,
243 BasicType basic_elem_type,
244 Node* src, Node* src_offset,
245 Node* dest, Node* dest_offset,
246 Node* copy_length,
247 bool disjoint_bases = false,
248 bool length_never_negative = false,
249 RegionNode* slow_region = NULL);
250 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
251 RegionNode* slow_region);
252 void generate_clear_array(const TypePtr* adr_type,
253 Node* dest,
254 BasicType basic_elem_type,
255 Node* slice_off,
256 Node* slice_len,
257 Node* slice_end);
258 bool generate_block_arraycopy(const TypePtr* adr_type,
259 BasicType basic_elem_type,
260 AllocateNode* alloc,
261 Node* src, Node* src_offset,
262 Node* dest, Node* dest_offset,
263 Node* dest_size, bool dest_uninitialized);
264 void generate_slow_arraycopy(const TypePtr* adr_type,
265 Node* src, Node* src_offset,
266 Node* dest, Node* dest_offset,
267 Node* copy_length, bool dest_uninitialized);
268 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
269 Node* dest_elem_klass,
270 Node* src, Node* src_offset,
271 Node* dest, Node* dest_offset,
272 Node* copy_length, bool dest_uninitialized);
273 Node* generate_generic_arraycopy(const TypePtr* adr_type,
274 Node* src, Node* src_offset,
275 Node* dest, Node* dest_offset,
276 Node* copy_length, bool dest_uninitialized);
277 void generate_unchecked_arraycopy(const TypePtr* adr_type,
278 BasicType basic_elem_type,
279 bool disjoint_bases,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* copy_length, bool dest_uninitialized);
283 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
284 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
285 bool inline_unsafe_ordered_store(BasicType type);
286 bool inline_fp_conversions(vmIntrinsics::ID id);
287 bool inline_number_methods(vmIntrinsics::ID id);
288 bool inline_reference_get();
289 bool inline_aescrypt_Block(vmIntrinsics::ID id);
290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
292 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
293 };
294
295
296 //---------------------------make_vm_intrinsic----------------------------
297 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
298 vmIntrinsics::ID id = m->intrinsic_id();
299 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
300
301 if (DisableIntrinsic[0] != '\0'
302 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
303 // disabled by a user request on the command line:
304 // example: -XX:DisableIntrinsic=_hashCode,_getClass
305 return NULL;
306 }
307
308 if (!m->is_loaded()) {
309 // do not attempt to inline unloaded methods
310 return NULL;
311 }
312
313 // Only a few intrinsics implement a virtual dispatch.
314 // They are expensive calls which are also frequently overridden.
315 if (is_virtual) {
316 switch (id) {
317 case vmIntrinsics::_hashCode:
318 case vmIntrinsics::_clone:
319 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
320 break;
321 default:
322 return NULL;
323 }
324 }
325
326 // -XX:-InlineNatives disables nearly all intrinsics:
327 if (!InlineNatives) {
328 switch (id) {
329 case vmIntrinsics::_indexOf:
330 case vmIntrinsics::_compareTo:
331 case vmIntrinsics::_equals:
332 case vmIntrinsics::_equalsC:
333 case vmIntrinsics::_getAndAddInt:
334 case vmIntrinsics::_getAndAddLong:
335 case vmIntrinsics::_getAndSetInt:
336 case vmIntrinsics::_getAndSetLong:
337 case vmIntrinsics::_getAndSetObject:
338 break; // InlineNatives does not control String.compareTo
339 case vmIntrinsics::_Reference_get:
340 break; // InlineNatives does not control Reference.get
341 default:
342 return NULL;
343 }
344 }
345
346 bool is_predicted = false;
347
348 switch (id) {
349 case vmIntrinsics::_compareTo:
350 if (!SpecialStringCompareTo) return NULL;
351 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
352 break;
353 case vmIntrinsics::_indexOf:
354 if (!SpecialStringIndexOf) return NULL;
355 break;
356 case vmIntrinsics::_equals:
357 if (!SpecialStringEquals) return NULL;
358 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
359 break;
360 case vmIntrinsics::_equalsC:
361 if (!SpecialArraysEquals) return NULL;
362 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
363 break;
364 case vmIntrinsics::_arraycopy:
365 if (!InlineArrayCopy) return NULL;
366 break;
367 case vmIntrinsics::_copyMemory:
368 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
369 if (!InlineArrayCopy) return NULL;
370 break;
371 case vmIntrinsics::_hashCode:
372 if (!InlineObjectHash) return NULL;
373 break;
374 case vmIntrinsics::_clone:
375 case vmIntrinsics::_copyOf:
376 case vmIntrinsics::_copyOfRange:
377 if (!InlineObjectCopy) return NULL;
378 // These also use the arraycopy intrinsic mechanism:
379 if (!InlineArrayCopy) return NULL;
380 break;
381 case vmIntrinsics::_checkIndex:
382 // We do not intrinsify this. The optimizer does fine with it.
383 return NULL;
384
385 case vmIntrinsics::_getCallerClass:
386 if (!UseNewReflection) return NULL;
387 if (!InlineReflectionGetCallerClass) return NULL;
388 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
389 break;
390
391 case vmIntrinsics::_bitCount_i:
392 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
393 break;
394
395 case vmIntrinsics::_bitCount_l:
396 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
397 break;
398
399 case vmIntrinsics::_numberOfLeadingZeros_i:
400 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
401 break;
402
403 case vmIntrinsics::_numberOfLeadingZeros_l:
404 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
405 break;
406
407 case vmIntrinsics::_numberOfTrailingZeros_i:
408 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
409 break;
410
411 case vmIntrinsics::_numberOfTrailingZeros_l:
412 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
413 break;
414
415 case vmIntrinsics::_reverseBytes_c:
416 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
417 break;
418 case vmIntrinsics::_reverseBytes_s:
419 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
420 break;
421 case vmIntrinsics::_reverseBytes_i:
422 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
423 break;
424 case vmIntrinsics::_reverseBytes_l:
425 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
426 break;
427
428 case vmIntrinsics::_Reference_get:
429 // Use the intrinsic version of Reference.get() so that the value in
430 // the referent field can be registered by the G1 pre-barrier code.
431 // Also add memory barrier to prevent commoning reads from this field
432 // across safepoint since GC can change it value.
433 break;
434
435 case vmIntrinsics::_compareAndSwapObject:
436 #ifdef _LP64
437 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
438 #endif
439 break;
440
441 case vmIntrinsics::_compareAndSwapLong:
442 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
443 break;
444
445 case vmIntrinsics::_getAndAddInt:
446 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
447 break;
448
449 case vmIntrinsics::_getAndAddLong:
450 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
451 break;
452
453 case vmIntrinsics::_getAndSetInt:
454 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
455 break;
456
457 case vmIntrinsics::_getAndSetLong:
458 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
459 break;
460
461 case vmIntrinsics::_getAndSetObject:
462 #ifdef _LP64
463 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
464 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
465 break;
466 #else
467 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
468 break;
469 #endif
470
471 case vmIntrinsics::_aescrypt_encryptBlock:
472 case vmIntrinsics::_aescrypt_decryptBlock:
473 if (!UseAESIntrinsics) return NULL;
474 break;
475
476 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
477 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
478 if (!UseAESIntrinsics) return NULL;
479 // these two require the predicated logic
480 is_predicted = true;
481 break;
482
483 default:
484 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
485 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
486 break;
487 }
488
489 // -XX:-InlineClassNatives disables natives from the Class class.
490 // The flag applies to all reflective calls, notably Array.newArray
491 // (visible to Java programmers as Array.newInstance).
492 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
493 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
494 if (!InlineClassNatives) return NULL;
495 }
496
497 // -XX:-InlineThreadNatives disables natives from the Thread class.
498 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
499 if (!InlineThreadNatives) return NULL;
500 }
501
502 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
503 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
504 m->holder()->name() == ciSymbol::java_lang_Float() ||
505 m->holder()->name() == ciSymbol::java_lang_Double()) {
506 if (!InlineMathNatives) return NULL;
507 }
508
509 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
510 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
511 if (!InlineUnsafeOps) return NULL;
512 }
513
514 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
515 }
516
517 //----------------------register_library_intrinsics-----------------------
518 // Initialize this file's data structures, for each Compile instance.
519 void Compile::register_library_intrinsics() {
520 // Nothing to do here.
521 }
522
523 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
524 LibraryCallKit kit(jvms, this);
525 Compile* C = kit.C;
526 int nodes = C->unique();
527 #ifndef PRODUCT
528 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
529 char buf[1000];
530 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
531 tty->print_cr("Intrinsic %s", str);
532 }
533 #endif
534 ciMethod* callee = kit.callee();
535 const int bci = kit.bci();
536
537 // Try to inline the intrinsic.
538 if (kit.try_to_inline()) {
539 if (C->print_intrinsics() || C->print_inlining()) {
540 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
541 }
542 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
543 if (C->log()) {
544 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
545 vmIntrinsics::name_at(intrinsic_id()),
546 (is_virtual() ? " virtual='1'" : ""),
547 C->unique() - nodes);
548 }
549 // Push the result from the inlined method onto the stack.
550 kit.push_result();
551 return kit.transfer_exceptions_into_jvms();
552 }
553
554 // The intrinsic bailed out
555 if (C->print_intrinsics() || C->print_inlining()) {
556 if (jvms->has_method()) {
557 // Not a root compile.
558 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
559 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
560 } else {
561 // Root compile
562 tty->print("Did not generate intrinsic %s%s at bci:%d in",
563 vmIntrinsics::name_at(intrinsic_id()),
564 (is_virtual() ? " (virtual)" : ""), bci);
565 }
566 }
567 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
568 return NULL;
569 }
570
571 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
572 LibraryCallKit kit(jvms, this);
573 Compile* C = kit.C;
574 int nodes = C->unique();
575 #ifndef PRODUCT
576 assert(is_predicted(), "sanity");
577 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
578 char buf[1000];
579 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
580 tty->print_cr("Predicate for intrinsic %s", str);
581 }
582 #endif
583 ciMethod* callee = kit.callee();
584 const int bci = kit.bci();
585
586 Node* slow_ctl = kit.try_to_predicate();
587 if (!kit.failing()) {
588 if (C->print_intrinsics() || C->print_inlining()) {
589 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
590 }
591 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
592 if (C->log()) {
593 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
594 vmIntrinsics::name_at(intrinsic_id()),
595 (is_virtual() ? " virtual='1'" : ""),
596 C->unique() - nodes);
597 }
598 return slow_ctl; // Could be NULL if the check folds.
599 }
600
601 // The intrinsic bailed out
602 if (C->print_intrinsics() || C->print_inlining()) {
603 if (jvms->has_method()) {
604 // Not a root compile.
605 const char* msg = "failed to generate predicate for intrinsic";
606 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
607 } else {
608 // Root compile
609 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
610 vmIntrinsics::name_at(intrinsic_id()),
611 (is_virtual() ? " (virtual)" : ""), bci);
612 }
613 }
614 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
615 return NULL;
616 }
617
618 bool LibraryCallKit::try_to_inline() {
619 // Handle symbolic names for otherwise undistinguished boolean switches:
620 const bool is_store = true;
621 const bool is_native_ptr = true;
622 const bool is_static = true;
623 const bool is_volatile = true;
624
625 if (!jvms()->has_method()) {
626 // Root JVMState has a null method.
627 assert(map()->memory()->Opcode() == Op_Parm, "");
628 // Insert the memory aliasing node
629 set_all_memory(reset_memory());
630 }
631 assert(merged_memory(), "");
632
633
634 switch (intrinsic_id()) {
635 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
636 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
637 case vmIntrinsics::_getClass: return inline_native_getClass();
638
639 case vmIntrinsics::_dsin:
640 case vmIntrinsics::_dcos:
641 case vmIntrinsics::_dtan:
642 case vmIntrinsics::_dabs:
643 case vmIntrinsics::_datan2:
644 case vmIntrinsics::_dsqrt:
645 case vmIntrinsics::_dexp:
646 case vmIntrinsics::_dlog:
647 case vmIntrinsics::_dlog10:
648 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
649
650 case vmIntrinsics::_min:
651 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
652
653 case vmIntrinsics::_arraycopy: return inline_arraycopy();
654
655 case vmIntrinsics::_compareTo: return inline_string_compareTo();
656 case vmIntrinsics::_indexOf: return inline_string_indexOf();
657 case vmIntrinsics::_equals: return inline_string_equals();
658
659 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
660 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
661 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
662 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
663 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
664 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
665 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
666 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
667 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
668
669 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
670 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
671 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
672 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
673 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
674 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
675 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
676 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
677 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
678
679 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
680 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
681 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
682 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
683 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
684 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
685 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
686 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
687
688 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
689 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
690 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
691 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
692 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
693 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
694 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
695 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
696
697 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
698 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
699 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
700 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
701 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
702 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
703 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
704 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
705 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
706
707 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
708 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
709 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
710 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
711 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
712 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
713 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
714 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
715 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
716
717 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
718 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
719 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
720 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
721
722 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
723 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
724 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
725
726 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
727 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
728 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
729
730 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
731 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
732 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
733 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
734 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
735
736 case vmIntrinsics::_currentThread: return inline_native_currentThread();
737 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
738
739 #ifdef TRACE_HAVE_INTRINSICS
740 case vmIntrinsics::_classID: return inline_native_classID();
741 case vmIntrinsics::_threadID: return inline_native_threadID();
742 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
743 #endif
744 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
745 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
746 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
747 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
748 case vmIntrinsics::_newArray: return inline_native_newArray();
749 case vmIntrinsics::_getLength: return inline_native_getLength();
750 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
751 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
752 case vmIntrinsics::_equalsC: return inline_array_equals();
753 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
754
755 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
756
757 case vmIntrinsics::_isInstance:
758 case vmIntrinsics::_getModifiers:
759 case vmIntrinsics::_isInterface:
760 case vmIntrinsics::_isArray:
761 case vmIntrinsics::_isPrimitive:
762 case vmIntrinsics::_getSuperclass:
763 case vmIntrinsics::_getComponentType:
764 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
765
766 case vmIntrinsics::_floatToRawIntBits:
767 case vmIntrinsics::_floatToIntBits:
768 case vmIntrinsics::_intBitsToFloat:
769 case vmIntrinsics::_doubleToRawLongBits:
770 case vmIntrinsics::_doubleToLongBits:
771 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
772
773 case vmIntrinsics::_numberOfLeadingZeros_i:
774 case vmIntrinsics::_numberOfLeadingZeros_l:
775 case vmIntrinsics::_numberOfTrailingZeros_i:
776 case vmIntrinsics::_numberOfTrailingZeros_l:
777 case vmIntrinsics::_bitCount_i:
778 case vmIntrinsics::_bitCount_l:
779 case vmIntrinsics::_reverseBytes_i:
780 case vmIntrinsics::_reverseBytes_l:
781 case vmIntrinsics::_reverseBytes_s:
782 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
783
784 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
785
786 case vmIntrinsics::_Reference_get: return inline_reference_get();
787
788 case vmIntrinsics::_aescrypt_encryptBlock:
789 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
790
791 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
792 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
793 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
794
795 default:
796 // If you get here, it may be that someone has added a new intrinsic
797 // to the list in vmSymbols.hpp without implementing it here.
798 #ifndef PRODUCT
799 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
800 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
801 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
802 }
803 #endif
804 return false;
805 }
806 }
807
808 Node* LibraryCallKit::try_to_predicate() {
809 if (!jvms()->has_method()) {
810 // Root JVMState has a null method.
811 assert(map()->memory()->Opcode() == Op_Parm, "");
812 // Insert the memory aliasing node
813 set_all_memory(reset_memory());
814 }
815 assert(merged_memory(), "");
816
817 switch (intrinsic_id()) {
818 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
819 return inline_cipherBlockChaining_AESCrypt_predicate(false);
820 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
821 return inline_cipherBlockChaining_AESCrypt_predicate(true);
822
823 default:
824 // If you get here, it may be that someone has added a new intrinsic
825 // to the list in vmSymbols.hpp without implementing it here.
826 #ifndef PRODUCT
827 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
828 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
829 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
830 }
831 #endif
832 Node* slow_ctl = control();
833 set_control(top()); // No fast path instrinsic
834 return slow_ctl;
835 }
836 }
837
838 //------------------------------set_result-------------------------------
839 // Helper function for finishing intrinsics.
840 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
841 record_for_igvn(region);
842 set_control(_gvn.transform(region));
843 set_result( _gvn.transform(value));
844 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
845 }
846
847 //------------------------------generate_guard---------------------------
848 // Helper function for generating guarded fast-slow graph structures.
849 // The given 'test', if true, guards a slow path. If the test fails
850 // then a fast path can be taken. (We generally hope it fails.)
851 // In all cases, GraphKit::control() is updated to the fast path.
852 // The returned value represents the control for the slow path.
853 // The return value is never 'top'; it is either a valid control
854 // or NULL if it is obvious that the slow path can never be taken.
855 // Also, if region and the slow control are not NULL, the slow edge
856 // is appended to the region.
857 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
858 if (stopped()) {
859 // Already short circuited.
860 return NULL;
861 }
862
863 // Build an if node and its projections.
864 // If test is true we take the slow path, which we assume is uncommon.
865 if (_gvn.type(test) == TypeInt::ZERO) {
866 // The slow branch is never taken. No need to build this guard.
867 return NULL;
868 }
869
870 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
871
872 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
873 if (if_slow == top()) {
874 // The slow branch is never taken. No need to build this guard.
875 return NULL;
876 }
877
878 if (region != NULL)
879 region->add_req(if_slow);
880
881 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
882 set_control(if_fast);
883
884 return if_slow;
885 }
886
887 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
888 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
889 }
890 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
891 return generate_guard(test, region, PROB_FAIR);
892 }
893
894 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
895 Node* *pos_index) {
896 if (stopped())
897 return NULL; // already stopped
898 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
899 return NULL; // index is already adequately typed
900 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
901 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
902 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
903 if (is_neg != NULL && pos_index != NULL) {
904 // Emulate effect of Parse::adjust_map_after_if.
905 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
906 ccast->set_req(0, control());
907 (*pos_index) = _gvn.transform(ccast);
908 }
909 return is_neg;
910 }
911
912 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
913 Node* *pos_index) {
914 if (stopped())
915 return NULL; // already stopped
916 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
917 return NULL; // index is already adequately typed
918 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
919 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
920 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
921 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
922 if (is_notp != NULL && pos_index != NULL) {
923 // Emulate effect of Parse::adjust_map_after_if.
924 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
925 ccast->set_req(0, control());
926 (*pos_index) = _gvn.transform(ccast);
927 }
928 return is_notp;
929 }
930
931 // Make sure that 'position' is a valid limit index, in [0..length].
932 // There are two equivalent plans for checking this:
933 // A. (offset + copyLength) unsigned<= arrayLength
934 // B. offset <= (arrayLength - copyLength)
935 // We require that all of the values above, except for the sum and
936 // difference, are already known to be non-negative.
937 // Plan A is robust in the face of overflow, if offset and copyLength
938 // are both hugely positive.
939 //
940 // Plan B is less direct and intuitive, but it does not overflow at
941 // all, since the difference of two non-negatives is always
942 // representable. Whenever Java methods must perform the equivalent
943 // check they generally use Plan B instead of Plan A.
944 // For the moment we use Plan A.
945 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
946 Node* subseq_length,
947 Node* array_length,
948 RegionNode* region) {
949 if (stopped())
950 return NULL; // already stopped
951 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
952 if (zero_offset && subseq_length->eqv_uncast(array_length))
953 return NULL; // common case of whole-array copy
954 Node* last = subseq_length;
955 if (!zero_offset) // last += offset
956 last = _gvn.transform( new (C) AddINode(last, offset));
957 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
958 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
959 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
960 return is_over;
961 }
962
963
964 //--------------------------generate_current_thread--------------------
965 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
966 ciKlass* thread_klass = env()->Thread_klass();
967 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
968 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
969 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
970 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
971 tls_output = thread;
972 return threadObj;
973 }
974
975
976 //------------------------------make_string_method_node------------------------
977 // Helper method for String intrinsic functions. This version is called
978 // with str1 and str2 pointing to String object nodes.
979 //
980 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
981 Node* no_ctrl = NULL;
982
983 // Get start addr of string
984 Node* str1_value = load_String_value(no_ctrl, str1);
985 Node* str1_offset = load_String_offset(no_ctrl, str1);
986 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
987
988 // Get length of string 1
989 Node* str1_len = load_String_length(no_ctrl, str1);
990
991 Node* str2_value = load_String_value(no_ctrl, str2);
992 Node* str2_offset = load_String_offset(no_ctrl, str2);
993 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
994
995 Node* str2_len = NULL;
996 Node* result = NULL;
997
998 switch (opcode) {
999 case Op_StrIndexOf:
1000 // Get length of string 2
1001 str2_len = load_String_length(no_ctrl, str2);
1002
1003 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1004 str1_start, str1_len, str2_start, str2_len);
1005 break;
1006 case Op_StrComp:
1007 // Get length of string 2
1008 str2_len = load_String_length(no_ctrl, str2);
1009
1010 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1011 str1_start, str1_len, str2_start, str2_len);
1012 break;
1013 case Op_StrEquals:
1014 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1015 str1_start, str2_start, str1_len);
1016 break;
1017 default:
1018 ShouldNotReachHere();
1019 return NULL;
1020 }
1021
1022 // All these intrinsics have checks.
1023 C->set_has_split_ifs(true); // Has chance for split-if optimization
1024
1025 return _gvn.transform(result);
1026 }
1027
1028 // Helper method for String intrinsic functions. This version is called
1029 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1030 // to Int nodes containing the lenghts of str1 and str2.
1031 //
1032 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1033 Node* result = NULL;
1034 switch (opcode) {
1035 case Op_StrIndexOf:
1036 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1037 str1_start, cnt1, str2_start, cnt2);
1038 break;
1039 case Op_StrComp:
1040 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1041 str1_start, cnt1, str2_start, cnt2);
1042 break;
1043 case Op_StrEquals:
1044 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1045 str1_start, str2_start, cnt1);
1046 break;
1047 default:
1048 ShouldNotReachHere();
1049 return NULL;
1050 }
1051
1052 // All these intrinsics have checks.
1053 C->set_has_split_ifs(true); // Has chance for split-if optimization
1054
1055 return _gvn.transform(result);
1056 }
1057
1058 //------------------------------inline_string_compareTo------------------------
1059 // public int java.lang.String.compareTo(String anotherString);
1060 bool LibraryCallKit::inline_string_compareTo() {
1061 Node* receiver = null_check(argument(0));
1062 Node* arg = null_check(argument(1));
1063 if (stopped()) {
1064 return true;
1065 }
1066 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1067 return true;
1068 }
1069
1070 //------------------------------inline_string_equals------------------------
1071 bool LibraryCallKit::inline_string_equals() {
1072 Node* receiver = null_check_receiver();
1073 // NOTE: Do not null check argument for String.equals() because spec
1074 // allows to specify NULL as argument.
1075 Node* argument = this->argument(1);
1076 if (stopped()) {
1077 return true;
1078 }
1079
1080 // paths (plus control) merge
1081 RegionNode* region = new (C) RegionNode(5);
1082 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1083
1084 // does source == target string?
1085 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1086 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1087
1088 Node* if_eq = generate_slow_guard(bol, NULL);
1089 if (if_eq != NULL) {
1090 // receiver == argument
1091 phi->init_req(2, intcon(1));
1092 region->init_req(2, if_eq);
1093 }
1094
1095 // get String klass for instanceOf
1096 ciInstanceKlass* klass = env()->String_klass();
1097
1098 if (!stopped()) {
1099 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1100 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1101 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1102
1103 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1104 //instanceOf == true, fallthrough
1105
1106 if (inst_false != NULL) {
1107 phi->init_req(3, intcon(0));
1108 region->init_req(3, inst_false);
1109 }
1110 }
1111
1112 if (!stopped()) {
1113 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1114
1115 // Properly cast the argument to String
1116 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1117 // This path is taken only when argument's type is String:NotNull.
1118 argument = cast_not_null(argument, false);
1119
1120 Node* no_ctrl = NULL;
1121
1122 // Get start addr of receiver
1123 Node* receiver_val = load_String_value(no_ctrl, receiver);
1124 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1125 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1126
1127 // Get length of receiver
1128 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1129
1130 // Get start addr of argument
1131 Node* argument_val = load_String_value(no_ctrl, argument);
1132 Node* argument_offset = load_String_offset(no_ctrl, argument);
1133 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1134
1135 // Get length of argument
1136 Node* argument_cnt = load_String_length(no_ctrl, argument);
1137
1138 // Check for receiver count != argument count
1139 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1140 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1141 Node* if_ne = generate_slow_guard(bol, NULL);
1142 if (if_ne != NULL) {
1143 phi->init_req(4, intcon(0));
1144 region->init_req(4, if_ne);
1145 }
1146
1147 // Check for count == 0 is done by assembler code for StrEquals.
1148
1149 if (!stopped()) {
1150 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1151 phi->init_req(1, equals);
1152 region->init_req(1, control());
1153 }
1154 }
1155
1156 // post merge
1157 set_control(_gvn.transform(region));
1158 record_for_igvn(region);
1159
1160 set_result(_gvn.transform(phi));
1161 return true;
1162 }
1163
1164 //------------------------------inline_array_equals----------------------------
1165 bool LibraryCallKit::inline_array_equals() {
1166 Node* arg1 = argument(0);
1167 Node* arg2 = argument(1);
1168 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1169 return true;
1170 }
1171
1172 // Java version of String.indexOf(constant string)
1173 // class StringDecl {
1174 // StringDecl(char[] ca) {
1175 // offset = 0;
1176 // count = ca.length;
1177 // value = ca;
1178 // }
1179 // int offset;
1180 // int count;
1181 // char[] value;
1182 // }
1183 //
1184 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1185 // int targetOffset, int cache_i, int md2) {
1186 // int cache = cache_i;
1187 // int sourceOffset = string_object.offset;
1188 // int sourceCount = string_object.count;
1189 // int targetCount = target_object.length;
1190 //
1191 // int targetCountLess1 = targetCount - 1;
1192 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1193 //
1194 // char[] source = string_object.value;
1195 // char[] target = target_object;
1196 // int lastChar = target[targetCountLess1];
1197 //
1198 // outer_loop:
1199 // for (int i = sourceOffset; i < sourceEnd; ) {
1200 // int src = source[i + targetCountLess1];
1201 // if (src == lastChar) {
1202 // // With random strings and a 4-character alphabet,
1203 // // reverse matching at this point sets up 0.8% fewer
1204 // // frames, but (paradoxically) makes 0.3% more probes.
1205 // // Since those probes are nearer the lastChar probe,
1206 // // there is may be a net D$ win with reverse matching.
1207 // // But, reversing loop inhibits unroll of inner loop
1208 // // for unknown reason. So, does running outer loop from
1209 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1210 // for (int j = 0; j < targetCountLess1; j++) {
1211 // if (target[targetOffset + j] != source[i+j]) {
1212 // if ((cache & (1 << source[i+j])) == 0) {
1213 // if (md2 < j+1) {
1214 // i += j+1;
1215 // continue outer_loop;
1216 // }
1217 // }
1218 // i += md2;
1219 // continue outer_loop;
1220 // }
1221 // }
1222 // return i - sourceOffset;
1223 // }
1224 // if ((cache & (1 << src)) == 0) {
1225 // i += targetCountLess1;
1226 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1227 // i++;
1228 // }
1229 // return -1;
1230 // }
1231
1232 //------------------------------string_indexOf------------------------
1233 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1234 jint cache_i, jint md2_i) {
1235
1236 Node* no_ctrl = NULL;
1237 float likely = PROB_LIKELY(0.9);
1238 float unlikely = PROB_UNLIKELY(0.9);
1239
1240 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1241
1242 Node* source = load_String_value(no_ctrl, string_object);
1243 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1244 Node* sourceCount = load_String_length(no_ctrl, string_object);
1245
1246 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1247 jint target_length = target_array->length();
1248 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1249 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1250
1251 IdealKit kit(this, false, true);
1252 #define __ kit.
1253 Node* zero = __ ConI(0);
1254 Node* one = __ ConI(1);
1255 Node* cache = __ ConI(cache_i);
1256 Node* md2 = __ ConI(md2_i);
1257 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1258 Node* targetCount = __ ConI(target_length);
1259 Node* targetCountLess1 = __ ConI(target_length - 1);
1260 Node* targetOffset = __ ConI(targetOffset_i);
1261 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1262
1263 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1264 Node* outer_loop = __ make_label(2 /* goto */);
1265 Node* return_ = __ make_label(1);
1266
1267 __ set(rtn,__ ConI(-1));
1268 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1269 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1270 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1271 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1272 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1273 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1274 Node* tpj = __ AddI(targetOffset, __ value(j));
1275 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1276 Node* ipj = __ AddI(__ value(i), __ value(j));
1277 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1278 __ if_then(targ, BoolTest::ne, src2); {
1279 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1280 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1281 __ increment(i, __ AddI(__ value(j), one));
1282 __ goto_(outer_loop);
1283 } __ end_if(); __ dead(j);
1284 }__ end_if(); __ dead(j);
1285 __ increment(i, md2);
1286 __ goto_(outer_loop);
1287 }__ end_if();
1288 __ increment(j, one);
1289 }__ end_loop(); __ dead(j);
1290 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1291 __ goto_(return_);
1292 }__ end_if();
1293 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1294 __ increment(i, targetCountLess1);
1295 }__ end_if();
1296 __ increment(i, one);
1297 __ bind(outer_loop);
1298 }__ end_loop(); __ dead(i);
1299 __ bind(return_);
1300
1301 // Final sync IdealKit and GraphKit.
1302 final_sync(kit);
1303 Node* result = __ value(rtn);
1304 #undef __
1305 C->set_has_loops(true);
1306 return result;
1307 }
1308
1309 //------------------------------inline_string_indexOf------------------------
1310 bool LibraryCallKit::inline_string_indexOf() {
1311 Node* receiver = argument(0);
1312 Node* arg = argument(1);
1313
1314 Node* result;
1315 // Disable the use of pcmpestri until it can be guaranteed that
1316 // the load doesn't cross into the uncommited space.
1317 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1318 UseSSE42Intrinsics) {
1319 // Generate SSE4.2 version of indexOf
1320 // We currently only have match rules that use SSE4.2
1321
1322 receiver = null_check(receiver);
1323 arg = null_check(arg);
1324 if (stopped()) {
1325 return true;
1326 }
1327
1328 ciInstanceKlass* str_klass = env()->String_klass();
1329 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1330
1331 // Make the merge point
1332 RegionNode* result_rgn = new (C) RegionNode(4);
1333 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1334 Node* no_ctrl = NULL;
1335
1336 // Get start addr of source string
1337 Node* source = load_String_value(no_ctrl, receiver);
1338 Node* source_offset = load_String_offset(no_ctrl, receiver);
1339 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1340
1341 // Get length of source string
1342 Node* source_cnt = load_String_length(no_ctrl, receiver);
1343
1344 // Get start addr of substring
1345 Node* substr = load_String_value(no_ctrl, arg);
1346 Node* substr_offset = load_String_offset(no_ctrl, arg);
1347 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1348
1349 // Get length of source string
1350 Node* substr_cnt = load_String_length(no_ctrl, arg);
1351
1352 // Check for substr count > string count
1353 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1354 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1355 Node* if_gt = generate_slow_guard(bol, NULL);
1356 if (if_gt != NULL) {
1357 result_phi->init_req(2, intcon(-1));
1358 result_rgn->init_req(2, if_gt);
1359 }
1360
1361 if (!stopped()) {
1362 // Check for substr count == 0
1363 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1364 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1365 Node* if_zero = generate_slow_guard(bol, NULL);
1366 if (if_zero != NULL) {
1367 result_phi->init_req(3, intcon(0));
1368 result_rgn->init_req(3, if_zero);
1369 }
1370 }
1371
1372 if (!stopped()) {
1373 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1374 result_phi->init_req(1, result);
1375 result_rgn->init_req(1, control());
1376 }
1377 set_control(_gvn.transform(result_rgn));
1378 record_for_igvn(result_rgn);
1379 result = _gvn.transform(result_phi);
1380
1381 } else { // Use LibraryCallKit::string_indexOf
1382 // don't intrinsify if argument isn't a constant string.
1383 if (!arg->is_Con()) {
1384 return false;
1385 }
1386 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1387 if (str_type == NULL) {
1388 return false;
1389 }
1390 ciInstanceKlass* klass = env()->String_klass();
1391 ciObject* str_const = str_type->const_oop();
1392 if (str_const == NULL || str_const->klass() != klass) {
1393 return false;
1394 }
1395 ciInstance* str = str_const->as_instance();
1396 assert(str != NULL, "must be instance");
1397
1398 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1399 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1400
1401 int o;
1402 int c;
1403 if (java_lang_String::has_offset_field()) {
1404 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1405 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1406 } else {
1407 o = 0;
1408 c = pat->length();
1409 }
1410
1411 // constant strings have no offset and count == length which
1412 // simplifies the resulting code somewhat so lets optimize for that.
1413 if (o != 0 || c != pat->length()) {
1414 return false;
1415 }
1416
1417 receiver = null_check(receiver, T_OBJECT);
1418 // NOTE: No null check on the argument is needed since it's a constant String oop.
1419 if (stopped()) {
1420 return true;
1421 }
1422
1423 // The null string as a pattern always returns 0 (match at beginning of string)
1424 if (c == 0) {
1425 set_result(intcon(0));
1426 return true;
1427 }
1428
1429 // Generate default indexOf
1430 jchar lastChar = pat->char_at(o + (c - 1));
1431 int cache = 0;
1432 int i;
1433 for (i = 0; i < c - 1; i++) {
1434 assert(i < pat->length(), "out of range");
1435 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1436 }
1437
1438 int md2 = c;
1439 for (i = 0; i < c - 1; i++) {
1440 assert(i < pat->length(), "out of range");
1441 if (pat->char_at(o + i) == lastChar) {
1442 md2 = (c - 1) - i;
1443 }
1444 }
1445
1446 result = string_indexOf(receiver, pat, o, cache, md2);
1447 }
1448 set_result(result);
1449 return true;
1450 }
1451
1452 //--------------------------round_double_node--------------------------------
1453 // Round a double node if necessary.
1454 Node* LibraryCallKit::round_double_node(Node* n) {
1455 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1456 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1457 return n;
1458 }
1459
1460 //------------------------------inline_math-----------------------------------
1461 // public static double Math.abs(double)
1462 // public static double Math.sqrt(double)
1463 // public static double Math.log(double)
1464 // public static double Math.log10(double)
1465 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1466 Node* arg = round_double_node(argument(0));
1467 Node* n;
1468 switch (id) {
1469 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1470 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1471 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1472 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1473 default: fatal_unexpected_iid(id); break;
1474 }
1475 set_result(_gvn.transform(n));
1476 return true;
1477 }
1478
1479 //------------------------------inline_trig----------------------------------
1480 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1481 // argument reduction which will turn into a fast/slow diamond.
1482 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1483 Node* arg = round_double_node(argument(0));
1484 Node* n = NULL;
1485
1486 switch (id) {
1487 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1488 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1489 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1490 default: fatal_unexpected_iid(id); break;
1491 }
1492 n = _gvn.transform(n);
1493
1494 // Rounding required? Check for argument reduction!
1495 if (Matcher::strict_fp_requires_explicit_rounding) {
1496 static const double pi_4 = 0.7853981633974483;
1497 static const double neg_pi_4 = -0.7853981633974483;
1498 // pi/2 in 80-bit extended precision
1499 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1500 // -pi/2 in 80-bit extended precision
1501 // 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};
1502 // Cutoff value for using this argument reduction technique
1503 //static const double pi_2_minus_epsilon = 1.564660403643354;
1504 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1505
1506 // Pseudocode for sin:
1507 // if (x <= Math.PI / 4.0) {
1508 // if (x >= -Math.PI / 4.0) return fsin(x);
1509 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1510 // } else {
1511 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1512 // }
1513 // return StrictMath.sin(x);
1514
1515 // Pseudocode for cos:
1516 // if (x <= Math.PI / 4.0) {
1517 // if (x >= -Math.PI / 4.0) return fcos(x);
1518 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1519 // } else {
1520 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1521 // }
1522 // return StrictMath.cos(x);
1523
1524 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1525 // requires a special machine instruction to load it. Instead we'll try
1526 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1527 // probably do the math inside the SIN encoding.
1528
1529 // Make the merge point
1530 RegionNode* r = new (C) RegionNode(3);
1531 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1532
1533 // Flatten arg so we need only 1 test
1534 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1535 // Node for PI/4 constant
1536 Node *pi4 = makecon(TypeD::make(pi_4));
1537 // Check PI/4 : abs(arg)
1538 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1539 // Check: If PI/4 < abs(arg) then go slow
1540 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1541 // Branch either way
1542 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1543 set_control(opt_iff(r,iff));
1544
1545 // Set fast path result
1546 phi->init_req(2, n);
1547
1548 // Slow path - non-blocking leaf call
1549 Node* call = NULL;
1550 switch (id) {
1551 case vmIntrinsics::_dsin:
1552 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1553 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1554 "Sin", NULL, arg, top());
1555 break;
1556 case vmIntrinsics::_dcos:
1557 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1558 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1559 "Cos", NULL, arg, top());
1560 break;
1561 case vmIntrinsics::_dtan:
1562 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1563 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1564 "Tan", NULL, arg, top());
1565 break;
1566 }
1567 assert(control()->in(0) == call, "");
1568 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1569 r->init_req(1, control());
1570 phi->init_req(1, slow_result);
1571
1572 // Post-merge
1573 set_control(_gvn.transform(r));
1574 record_for_igvn(r);
1575 n = _gvn.transform(phi);
1576
1577 C->set_has_split_ifs(true); // Has chance for split-if optimization
1578 }
1579 set_result(n);
1580 return true;
1581 }
1582
1583 Node* LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1584 //-------------------
1585 //result=(result.isNaN())? funcAddr():result;
1586 // Check: If isNaN() by checking result!=result? then either trap
1587 // or go to runtime
1588 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1589 // Build the boolean node
1590 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1591
1592 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1593 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1594 // The pow or exp intrinsic returned a NaN, which requires a call
1595 // to the runtime. Recompile with the runtime call.
1596 uncommon_trap(Deoptimization::Reason_intrinsic,
1597 Deoptimization::Action_make_not_entrant);
1598 }
1599 return result;
1600 } else {
1601 // If this inlining ever returned NaN in the past, we compile a call
1602 // to the runtime to properly handle corner cases
1603
1604 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1605 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1606 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1607
1608 if (!if_slow->is_top()) {
1609 RegionNode* result_region = new (C) RegionNode(3);
1610 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1611
1612 result_region->init_req(1, if_fast);
1613 result_val->init_req(1, result);
1614
1615 set_control(if_slow);
1616
1617 const TypePtr* no_memory_effects = NULL;
1618 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1619 no_memory_effects,
1620 x, top(), y, y ? top() : NULL);
1621 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1622 #ifdef ASSERT
1623 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1624 assert(value_top == top(), "second value must be top");
1625 #endif
1626
1627 result_region->init_req(2, control());
1628 result_val->init_req(2, value);
1629 set_control(_gvn.transform(result_region));
1630 return _gvn.transform(result_val);
1631 } else {
1632 return result;
1633 }
1634 }
1635 }
1636
1637 //------------------------------inline_exp-------------------------------------
1638 // Inline exp instructions, if possible. The Intel hardware only misses
1639 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1640 bool LibraryCallKit::inline_exp() {
1641 Node* arg = round_double_node(argument(0));
1642 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1643
1644 n = finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1645 set_result(n);
1646
1647 C->set_has_split_ifs(true); // Has chance for split-if optimization
1648 return true;
1649 }
1650
1651 //------------------------------inline_pow-------------------------------------
1652 // Inline power instructions, if possible.
1653 bool LibraryCallKit::inline_pow() {
1654 // Pseudocode for pow
1655 // if (y == 2) {
1656 // return x * x;
1657 // } else {
1658 // if (x <= 0.0) {
1659 // long longy = (long)y;
1660 // if ((double)longy == y) { // if y is long
1661 // if (y + 1 == y) longy = 0; // huge number: even
1662 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1663 // } else {
1664 // result = NaN;
1665 // }
1666 // } else {
1667 // result = DPow(x,y);
1668 // }
1669 // if (result != result)? {
1670 // result = uncommon_trap() or runtime_call();
1671 // }
1672 // return result;
1673 // }
1674
1675 Node* x = round_double_node(argument(0));
1676 Node* y = round_double_node(argument(2));
1677
1678 Node* result = NULL;
1679
1680 Node* const_two_node = makecon(TypeD::make(2.0));
1681 Node* cmp_node = _gvn.transform(new (C) CmpDNode(y, const_two_node));
1682 Node* bool_node = _gvn.transform(new (C) BoolNode(cmp_node, BoolTest::eq));
1683 IfNode* if_node = create_and_xform_if(control(), bool_node, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1684 Node* if_true = _gvn.transform(new (C) IfTrueNode(if_node));
1685 Node* if_false = _gvn.transform(new (C) IfFalseNode(if_node));
1686
1687 RegionNode* region_node = new (C) RegionNode(3);
1688 region_node->init_req(1, if_true);
1689
1690 Node* phi_node = new (C) PhiNode(region_node, Type::DOUBLE);
1691 // special case for x^y where y == 2, we can convert it to x * x
1692 phi_node->init_req(1, _gvn.transform(new (C) MulDNode(x, x)));
1693
1694 // set control to if_false since we will now process the false branch
1695 set_control(if_false);
1696
1697 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1698 // Short form: skip the fancy tests and just check for NaN result.
1699 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1700 } else {
1701 // If this inlining ever returned NaN in the past, include all
1702 // checks + call to the runtime.
1703
1704 // Set the merge point for If node with condition of (x <= 0.0)
1705 // There are four possible paths to region node and phi node
1706 RegionNode *r = new (C) RegionNode(4);
1707 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1708
1709 // Build the first if node: if (x <= 0.0)
1710 // Node for 0 constant
1711 Node *zeronode = makecon(TypeD::ZERO);
1712 // Check x:0
1713 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1714 // Check: If (x<=0) then go complex path
1715 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1716 // Branch either way
1717 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1718 // Fast path taken; set region slot 3
1719 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1720 r->init_req(3,fast_taken); // Capture fast-control
1721
1722 // Fast path not-taken, i.e. slow path
1723 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1724
1725 // Set fast path result
1726 Node *fast_result = _gvn.transform( new (C) PowDNode(C, control(), x, y) );
1727 phi->init_req(3, fast_result);
1728
1729 // Complex path
1730 // Build the second if node (if y is long)
1731 // Node for (long)y
1732 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1733 // Node for (double)((long) y)
1734 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1735 // Check (double)((long) y) : y
1736 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1737 // Check if (y isn't long) then go to slow path
1738
1739 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1740 // Branch either way
1741 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1742 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1743
1744 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1745
1746 // Calculate DPow(abs(x), y)*(1 & (long)y)
1747 // Node for constant 1
1748 Node *conone = longcon(1);
1749 // 1& (long)y
1750 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1751
1752 // A huge number is always even. Detect a huge number by checking
1753 // if y + 1 == y and set integer to be tested for parity to 0.
1754 // Required for corner case:
1755 // (long)9.223372036854776E18 = max_jlong
1756 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1757 // max_jlong is odd but 9.223372036854776E18 is even
1758 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1759 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1760 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1761 Node* correctedsign = NULL;
1762 if (ConditionalMoveLimit != 0) {
1763 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1764 } else {
1765 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1766 RegionNode *r = new (C) RegionNode(3);
1767 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1768 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1769 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1770 phi->init_req(1, signnode);
1771 phi->init_req(2, longcon(0));
1772 correctedsign = _gvn.transform(phi);
1773 ylong_path = _gvn.transform(r);
1774 record_for_igvn(r);
1775 }
1776
1777 // zero node
1778 Node *conzero = longcon(0);
1779 // Check (1&(long)y)==0?
1780 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1781 // Check if (1&(long)y)!=0?, if so the result is negative
1782 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1783 // abs(x)
1784 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1785 // abs(x)^y
1786 Node *absxpowy = _gvn.transform( new (C) PowDNode(C, control(), absx, y) );
1787 // -abs(x)^y
1788 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1789 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1790 Node *signresult = NULL;
1791 if (ConditionalMoveLimit != 0) {
1792 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1793 } else {
1794 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1795 RegionNode *r = new (C) RegionNode(3);
1796 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1797 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1798 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1799 phi->init_req(1, absxpowy);
1800 phi->init_req(2, negabsxpowy);
1801 signresult = _gvn.transform(phi);
1802 ylong_path = _gvn.transform(r);
1803 record_for_igvn(r);
1804 }
1805 // Set complex path fast result
1806 r->init_req(2, ylong_path);
1807 phi->init_req(2, signresult);
1808
1809 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1810 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1811 r->init_req(1,slow_path);
1812 phi->init_req(1,slow_result);
1813
1814 // Post merge
1815 set_control(_gvn.transform(r));
1816 record_for_igvn(r);
1817 result = _gvn.transform(phi);
1818 }
1819
1820 result = finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1821
1822 // control from finish_pow_exp is now input to the region node
1823 region_node->set_req(2, control());
1824 // the result from finish_pow_exp is now input to the phi node
1825 phi_node->init_req(2, result);
1826 set_control(_gvn.transform(region_node));
1827 record_for_igvn(region_node);
1828 set_result(_gvn.transform(phi_node));
1829
1830 C->set_has_split_ifs(true); // Has chance for split-if optimization
1831 return true;
1832 }
1833
1834 //------------------------------runtime_math-----------------------------
1835 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1836 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1837 "must be (DD)D or (D)D type");
1838
1839 // Inputs
1840 Node* a = round_double_node(argument(0));
1841 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1842
1843 const TypePtr* no_memory_effects = NULL;
1844 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1845 no_memory_effects,
1846 a, top(), b, b ? top() : NULL);
1847 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1848 #ifdef ASSERT
1849 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1850 assert(value_top == top(), "second value must be top");
1851 #endif
1852
1853 set_result(value);
1854 return true;
1855 }
1856
1857 //------------------------------inline_math_native-----------------------------
1858 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1859 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1860 switch (id) {
1861 // These intrinsics are not properly supported on all hardware
1862 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1863 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1864 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1865 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1866 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1867 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1868
1869 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1870 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1871 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1872 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1873
1874 // These intrinsics are supported on all hardware
1875 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1876 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1877
1878 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1879 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1880 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1881 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1882 #undef FN_PTR
1883
1884 // These intrinsics are not yet correctly implemented
1885 case vmIntrinsics::_datan2:
1886 return false;
1887
1888 default:
1889 fatal_unexpected_iid(id);
1890 return false;
1891 }
1892 }
1893
1894 static bool is_simple_name(Node* n) {
1895 return (n->req() == 1 // constant
1896 || (n->is_Type() && n->as_Type()->type()->singleton())
1897 || n->is_Proj() // parameter or return value
1898 || n->is_Phi() // local of some sort
1899 );
1900 }
1901
1902 //----------------------------inline_min_max-----------------------------------
1903 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1904 set_result(generate_min_max(id, argument(0), argument(1)));
1905 return true;
1906 }
1907
1908 Node*
1909 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1910 // These are the candidate return value:
1911 Node* xvalue = x0;
1912 Node* yvalue = y0;
1913
1914 if (xvalue == yvalue) {
1915 return xvalue;
1916 }
1917
1918 bool want_max = (id == vmIntrinsics::_max);
1919
1920 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1921 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1922 if (txvalue == NULL || tyvalue == NULL) return top();
1923 // This is not really necessary, but it is consistent with a
1924 // hypothetical MaxINode::Value method:
1925 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1926
1927 // %%% This folding logic should (ideally) be in a different place.
1928 // Some should be inside IfNode, and there to be a more reliable
1929 // transformation of ?: style patterns into cmoves. We also want
1930 // more powerful optimizations around cmove and min/max.
1931
1932 // Try to find a dominating comparison of these guys.
1933 // It can simplify the index computation for Arrays.copyOf
1934 // and similar uses of System.arraycopy.
1935 // First, compute the normalized version of CmpI(x, y).
1936 int cmp_op = Op_CmpI;
1937 Node* xkey = xvalue;
1938 Node* ykey = yvalue;
1939 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1940 if (ideal_cmpxy->is_Cmp()) {
1941 // E.g., if we have CmpI(length - offset, count),
1942 // it might idealize to CmpI(length, count + offset)
1943 cmp_op = ideal_cmpxy->Opcode();
1944 xkey = ideal_cmpxy->in(1);
1945 ykey = ideal_cmpxy->in(2);
1946 }
1947
1948 // Start by locating any relevant comparisons.
1949 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1950 Node* cmpxy = NULL;
1951 Node* cmpyx = NULL;
1952 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1953 Node* cmp = start_from->fast_out(k);
1954 if (cmp->outcnt() > 0 && // must have prior uses
1955 cmp->in(0) == NULL && // must be context-independent
1956 cmp->Opcode() == cmp_op) { // right kind of compare
1957 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1958 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1959 }
1960 }
1961
1962 const int NCMPS = 2;
1963 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1964 int cmpn;
1965 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1966 if (cmps[cmpn] != NULL) break; // find a result
1967 }
1968 if (cmpn < NCMPS) {
1969 // Look for a dominating test that tells us the min and max.
1970 int depth = 0; // Limit search depth for speed
1971 Node* dom = control();
1972 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1973 if (++depth >= 100) break;
1974 Node* ifproj = dom;
1975 if (!ifproj->is_Proj()) continue;
1976 Node* iff = ifproj->in(0);
1977 if (!iff->is_If()) continue;
1978 Node* bol = iff->in(1);
1979 if (!bol->is_Bool()) continue;
1980 Node* cmp = bol->in(1);
1981 if (cmp == NULL) continue;
1982 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1983 if (cmps[cmpn] == cmp) break;
1984 if (cmpn == NCMPS) continue;
1985 BoolTest::mask btest = bol->as_Bool()->_test._test;
1986 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1987 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1988 // At this point, we know that 'x btest y' is true.
1989 switch (btest) {
1990 case BoolTest::eq:
1991 // They are proven equal, so we can collapse the min/max.
1992 // Either value is the answer. Choose the simpler.
1993 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1994 return yvalue;
1995 return xvalue;
1996 case BoolTest::lt: // x < y
1997 case BoolTest::le: // x <= y
1998 return (want_max ? yvalue : xvalue);
1999 case BoolTest::gt: // x > y
2000 case BoolTest::ge: // x >= y
2001 return (want_max ? xvalue : yvalue);
2002 }
2003 }
2004 }
2005
2006 // We failed to find a dominating test.
2007 // Let's pick a test that might GVN with prior tests.
2008 Node* best_bol = NULL;
2009 BoolTest::mask best_btest = BoolTest::illegal;
2010 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2011 Node* cmp = cmps[cmpn];
2012 if (cmp == NULL) continue;
2013 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2014 Node* bol = cmp->fast_out(j);
2015 if (!bol->is_Bool()) continue;
2016 BoolTest::mask btest = bol->as_Bool()->_test._test;
2017 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2018 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2019 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2020 best_bol = bol->as_Bool();
2021 best_btest = btest;
2022 }
2023 }
2024 }
2025
2026 Node* answer_if_true = NULL;
2027 Node* answer_if_false = NULL;
2028 switch (best_btest) {
2029 default:
2030 if (cmpxy == NULL)
2031 cmpxy = ideal_cmpxy;
2032 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2033 // and fall through:
2034 case BoolTest::lt: // x < y
2035 case BoolTest::le: // x <= y
2036 answer_if_true = (want_max ? yvalue : xvalue);
2037 answer_if_false = (want_max ? xvalue : yvalue);
2038 break;
2039 case BoolTest::gt: // x > y
2040 case BoolTest::ge: // x >= y
2041 answer_if_true = (want_max ? xvalue : yvalue);
2042 answer_if_false = (want_max ? yvalue : xvalue);
2043 break;
2044 }
2045
2046 jint hi, lo;
2047 if (want_max) {
2048 // We can sharpen the minimum.
2049 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2050 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2051 } else {
2052 // We can sharpen the maximum.
2053 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2054 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2055 }
2056
2057 // Use a flow-free graph structure, to avoid creating excess control edges
2058 // which could hinder other optimizations.
2059 // Since Math.min/max is often used with arraycopy, we want
2060 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2061 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2062 answer_if_false, answer_if_true,
2063 TypeInt::make(lo, hi, widen));
2064
2065 return _gvn.transform(cmov);
2066
2067 /*
2068 // This is not as desirable as it may seem, since Min and Max
2069 // nodes do not have a full set of optimizations.
2070 // And they would interfere, anyway, with 'if' optimizations
2071 // and with CMoveI canonical forms.
2072 switch (id) {
2073 case vmIntrinsics::_min:
2074 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2075 case vmIntrinsics::_max:
2076 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2077 default:
2078 ShouldNotReachHere();
2079 }
2080 */
2081 }
2082
2083 inline int
2084 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2085 const TypePtr* base_type = TypePtr::NULL_PTR;
2086 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2087 if (base_type == NULL) {
2088 // Unknown type.
2089 return Type::AnyPtr;
2090 } else if (base_type == TypePtr::NULL_PTR) {
2091 // Since this is a NULL+long form, we have to switch to a rawptr.
2092 base = _gvn.transform( new (C) CastX2PNode(offset) );
2093 offset = MakeConX(0);
2094 return Type::RawPtr;
2095 } else if (base_type->base() == Type::RawPtr) {
2096 return Type::RawPtr;
2097 } else if (base_type->isa_oopptr()) {
2098 // Base is never null => always a heap address.
2099 if (base_type->ptr() == TypePtr::NotNull) {
2100 return Type::OopPtr;
2101 }
2102 // Offset is small => always a heap address.
2103 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2104 if (offset_type != NULL &&
2105 base_type->offset() == 0 && // (should always be?)
2106 offset_type->_lo >= 0 &&
2107 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2108 return Type::OopPtr;
2109 }
2110 // Otherwise, it might either be oop+off or NULL+addr.
2111 return Type::AnyPtr;
2112 } else {
2113 // No information:
2114 return Type::AnyPtr;
2115 }
2116 }
2117
2118 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2119 int kind = classify_unsafe_addr(base, offset);
2120 if (kind == Type::RawPtr) {
2121 return basic_plus_adr(top(), base, offset);
2122 } else {
2123 return basic_plus_adr(base, offset);
2124 }
2125 }
2126
2127 //--------------------------inline_number_methods-----------------------------
2128 // inline int Integer.numberOfLeadingZeros(int)
2129 // inline int Long.numberOfLeadingZeros(long)
2130 //
2131 // inline int Integer.numberOfTrailingZeros(int)
2132 // inline int Long.numberOfTrailingZeros(long)
2133 //
2134 // inline int Integer.bitCount(int)
2135 // inline int Long.bitCount(long)
2136 //
2137 // inline char Character.reverseBytes(char)
2138 // inline short Short.reverseBytes(short)
2139 // inline int Integer.reverseBytes(int)
2140 // inline long Long.reverseBytes(long)
2141 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2142 Node* arg = argument(0);
2143 Node* n;
2144 switch (id) {
2145 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2146 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2147 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2148 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2149 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2150 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2151 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2152 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2153 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2154 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2155 default: fatal_unexpected_iid(id); break;
2156 }
2157 set_result(_gvn.transform(n));
2158 return true;
2159 }
2160
2161 //----------------------------inline_unsafe_access----------------------------
2162
2163 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2164
2165 // Helper that guards and inserts a pre-barrier.
2166 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2167 Node* pre_val, bool need_mem_bar) {
2168 // We could be accessing the referent field of a reference object. If so, when G1
2169 // is enabled, we need to log the value in the referent field in an SATB buffer.
2170 // This routine performs some compile time filters and generates suitable
2171 // runtime filters that guard the pre-barrier code.
2172 // Also add memory barrier for non volatile load from the referent field
2173 // to prevent commoning of loads across safepoint.
2174 if (!UseG1GC && !need_mem_bar)
2175 return;
2176
2177 // Some compile time checks.
2178
2179 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2180 const TypeX* otype = offset->find_intptr_t_type();
2181 if (otype != NULL && otype->is_con() &&
2182 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2183 // Constant offset but not the reference_offset so just return
2184 return;
2185 }
2186
2187 // We only need to generate the runtime guards for instances.
2188 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2189 if (btype != NULL) {
2190 if (btype->isa_aryptr()) {
2191 // Array type so nothing to do
2192 return;
2193 }
2194
2195 const TypeInstPtr* itype = btype->isa_instptr();
2196 if (itype != NULL) {
2197 // Can the klass of base_oop be statically determined to be
2198 // _not_ a sub-class of Reference and _not_ Object?
2199 ciKlass* klass = itype->klass();
2200 if ( klass->is_loaded() &&
2201 !klass->is_subtype_of(env()->Reference_klass()) &&
2202 !env()->Object_klass()->is_subtype_of(klass)) {
2203 return;
2204 }
2205 }
2206 }
2207
2208 // The compile time filters did not reject base_oop/offset so
2209 // we need to generate the following runtime filters
2210 //
2211 // if (offset == java_lang_ref_Reference::_reference_offset) {
2212 // if (instance_of(base, java.lang.ref.Reference)) {
2213 // pre_barrier(_, pre_val, ...);
2214 // }
2215 // }
2216
2217 float likely = PROB_LIKELY( 0.999);
2218 float unlikely = PROB_UNLIKELY(0.999);
2219
2220 IdealKit ideal(this);
2221 #define __ ideal.
2222
2223 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2224
2225 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2226 // Update graphKit memory and control from IdealKit.
2227 sync_kit(ideal);
2228
2229 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2230 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2231
2232 // Update IdealKit memory and control from graphKit.
2233 __ sync_kit(this);
2234
2235 Node* one = __ ConI(1);
2236 // is_instof == 0 if base_oop == NULL
2237 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2238
2239 // Update graphKit from IdeakKit.
2240 sync_kit(ideal);
2241
2242 // Use the pre-barrier to record the value in the referent field
2243 pre_barrier(false /* do_load */,
2244 __ ctrl(),
2245 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2246 pre_val /* pre_val */,
2247 T_OBJECT);
2248 if (need_mem_bar) {
2249 // Add memory barrier to prevent commoning reads from this field
2250 // across safepoint since GC can change its value.
2251 insert_mem_bar(Op_MemBarCPUOrder);
2252 }
2253 // Update IdealKit from graphKit.
2254 __ sync_kit(this);
2255
2256 } __ end_if(); // _ref_type != ref_none
2257 } __ end_if(); // offset == referent_offset
2258
2259 // Final sync IdealKit and GraphKit.
2260 final_sync(ideal);
2261 #undef __
2262 }
2263
2264
2265 // Interpret Unsafe.fieldOffset cookies correctly:
2266 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2267
2268 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2269 // Attempt to infer a sharper value type from the offset and base type.
2270 ciKlass* sharpened_klass = NULL;
2271
2272 // See if it is an instance field, with an object type.
2273 if (alias_type->field() != NULL) {
2274 assert(!is_native_ptr, "native pointer op cannot use a java address");
2275 if (alias_type->field()->type()->is_klass()) {
2276 sharpened_klass = alias_type->field()->type()->as_klass();
2277 }
2278 }
2279
2280 // See if it is a narrow oop array.
2281 if (adr_type->isa_aryptr()) {
2282 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2283 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2284 if (elem_type != NULL) {
2285 sharpened_klass = elem_type->klass();
2286 }
2287 }
2288 }
2289
2290 // The sharpened class might be unloaded if there is no class loader
2291 // contraint in place.
2292 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2293 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2294
2295 #ifndef PRODUCT
2296 if (C->print_intrinsics() || C->print_inlining()) {
2297 tty->print(" from base type: "); adr_type->dump();
2298 tty->print(" sharpened value: "); tjp->dump();
2299 }
2300 #endif
2301 // Sharpen the value type.
2302 return tjp;
2303 }
2304 return NULL;
2305 }
2306
2307 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2308 if (callee()->is_static()) return false; // caller must have the capability!
2309
2310 #ifndef PRODUCT
2311 {
2312 ResourceMark rm;
2313 // Check the signatures.
2314 ciSignature* sig = callee()->signature();
2315 #ifdef ASSERT
2316 if (!is_store) {
2317 // Object getObject(Object base, int/long offset), etc.
2318 BasicType rtype = sig->return_type()->basic_type();
2319 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2320 rtype = T_ADDRESS; // it is really a C void*
2321 assert(rtype == type, "getter must return the expected value");
2322 if (!is_native_ptr) {
2323 assert(sig->count() == 2, "oop getter has 2 arguments");
2324 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2325 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2326 } else {
2327 assert(sig->count() == 1, "native getter has 1 argument");
2328 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2329 }
2330 } else {
2331 // void putObject(Object base, int/long offset, Object x), etc.
2332 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2333 if (!is_native_ptr) {
2334 assert(sig->count() == 3, "oop putter has 3 arguments");
2335 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2336 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2337 } else {
2338 assert(sig->count() == 2, "native putter has 2 arguments");
2339 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2340 }
2341 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2342 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2343 vtype = T_ADDRESS; // it is really a C void*
2344 assert(vtype == type, "putter must accept the expected value");
2345 }
2346 #endif // ASSERT
2347 }
2348 #endif //PRODUCT
2349
2350 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2351
2352 Node* receiver = argument(0); // type: oop
2353
2354 // Build address expression. See the code in inline_unsafe_prefetch.
2355 Node* adr;
2356 Node* heap_base_oop = top();
2357 Node* offset = top();
2358 Node* val;
2359
2360 if (!is_native_ptr) {
2361 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2362 Node* base = argument(1); // type: oop
2363 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2364 offset = argument(2); // type: long
2365 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2366 // to be plain byte offsets, which are also the same as those accepted
2367 // by oopDesc::field_base.
2368 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2369 "fieldOffset must be byte-scaled");
2370 // 32-bit machines ignore the high half!
2371 offset = ConvL2X(offset);
2372 adr = make_unsafe_address(base, offset);
2373 heap_base_oop = base;
2374 val = is_store ? argument(4) : NULL;
2375 } else {
2376 Node* ptr = argument(1); // type: long
2377 ptr = ConvL2X(ptr); // adjust Java long to machine word
2378 adr = make_unsafe_address(NULL, ptr);
2379 val = is_store ? argument(3) : NULL;
2380 }
2381
2382 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2383
2384 // First guess at the value type.
2385 const Type *value_type = Type::get_const_basic_type(type);
2386
2387 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2388 // there was not enough information to nail it down.
2389 Compile::AliasType* alias_type = C->alias_type(adr_type);
2390 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2391
2392 // We will need memory barriers unless we can determine a unique
2393 // alias category for this reference. (Note: If for some reason
2394 // the barriers get omitted and the unsafe reference begins to "pollute"
2395 // the alias analysis of the rest of the graph, either Compile::can_alias
2396 // or Compile::must_alias will throw a diagnostic assert.)
2397 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2398
2399 // If we are reading the value of the referent field of a Reference
2400 // object (either by using Unsafe directly or through reflection)
2401 // then, if G1 is enabled, we need to record the referent in an
2402 // SATB log buffer using the pre-barrier mechanism.
2403 // Also we need to add memory barrier to prevent commoning reads
2404 // from this field across safepoint since GC can change its value.
2405 bool need_read_barrier = !is_native_ptr && !is_store &&
2406 offset != top() && heap_base_oop != top();
2407
2408 if (!is_store && type == T_OBJECT) {
2409 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2410 if (tjp != NULL) {
2411 value_type = tjp;
2412 }
2413 }
2414
2415 receiver = null_check(receiver);
2416 if (stopped()) {
2417 return true;
2418 }
2419 // Heap pointers get a null-check from the interpreter,
2420 // as a courtesy. However, this is not guaranteed by Unsafe,
2421 // and it is not possible to fully distinguish unintended nulls
2422 // from intended ones in this API.
2423
2424 if (is_volatile) {
2425 // We need to emit leading and trailing CPU membars (see below) in
2426 // addition to memory membars when is_volatile. This is a little
2427 // too strong, but avoids the need to insert per-alias-type
2428 // volatile membars (for stores; compare Parse::do_put_xxx), which
2429 // we cannot do effectively here because we probably only have a
2430 // rough approximation of type.
2431 need_mem_bar = true;
2432 // For Stores, place a memory ordering barrier now.
2433 if (is_store)
2434 insert_mem_bar(Op_MemBarRelease);
2435 #ifdef PPC64
2436 // Support ordering of "Independent Reads of Independent Writes" (see Parse::do_get_xxx).
2437 // Solution: implement volatile read as fence-load-acquire
2438 else
2439 insert_mem_bar(Op_MemBarVolatile);
2440 #endif
2441 }
2442
2443 // Memory barrier to prevent normal and 'unsafe' accesses from
2444 // bypassing each other. Happens after null checks, so the
2445 // exception paths do not take memory state from the memory barrier,
2446 // so there's no problems making a strong assert about mixing users
2447 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2448 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2449 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2450
2451 if (!is_store) {
2452 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2453 // load value
2454 switch (type) {
2455 case T_BOOLEAN:
2456 case T_CHAR:
2457 case T_BYTE:
2458 case T_SHORT:
2459 case T_INT:
2460 case T_LONG:
2461 case T_FLOAT:
2462 case T_DOUBLE:
2463 break;
2464 case T_OBJECT:
2465 if (need_read_barrier) {
2466 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2467 }
2468 break;
2469 case T_ADDRESS:
2470 // Cast to an int type.
2471 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2472 p = ConvX2UL(p);
2473 break;
2474 default:
2475 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2476 break;
2477 }
2478 // The load node has the control of the preceding MemBarCPUOrder. All
2479 // following nodes will have the control of the MemBarCPUOrder inserted at
2480 // the end of this method. So, pushing the load onto the stack at a later
2481 // point is fine.
2482 set_result(p);
2483 } else {
2484 // place effect of store into memory
2485 switch (type) {
2486 case T_DOUBLE:
2487 val = dstore_rounding(val);
2488 break;
2489 case T_ADDRESS:
2490 // Repackage the long as a pointer.
2491 val = ConvL2X(val);
2492 val = _gvn.transform( new (C) CastX2PNode(val) );
2493 break;
2494 }
2495
2496 StoreNode::Sem sem = is_volatile ? StoreNode::release : StoreNode::unordered;
2497 if (type != T_OBJECT ) {
2498 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile, sem);
2499 } else {
2500 // Possibly an oop being stored to Java heap or native memory
2501 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2502 // oop to Java heap.
2503 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, sem);
2504 } else {
2505 // We can't tell at compile time if we are storing in the Java heap or outside
2506 // of it. So we need to emit code to conditionally do the proper type of
2507 // store.
2508
2509 IdealKit ideal(this);
2510 #define __ ideal.
2511 // QQQ who knows what probability is here??
2512 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2513 // Sync IdealKit and graphKit.
2514 sync_kit(ideal);
2515 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, sem);
2516 // Update IdealKit memory.
2517 __ sync_kit(this);
2518 } __ else_(); {
2519 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile, sem);
2520 } __ end_if();
2521 // Final sync IdealKit and GraphKit.
2522 final_sync(ideal);
2523 #undef __
2524 }
2525 }
2526 }
2527
2528 if (is_volatile) {
2529 if (!is_store)
2530 insert_mem_bar(Op_MemBarAcquire);
2531 #ifndef PPC64
2532 // Changed volatiles/Unsafe: lwsync-store, fence-load-acquire.
2533 else
2534 insert_mem_bar(Op_MemBarVolatile);
2535 #endif
2536 }
2537
2538 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2539
2540 return true;
2541 }
2542
2543 //----------------------------inline_unsafe_prefetch----------------------------
2544
2545 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2546 #ifndef PRODUCT
2547 {
2548 ResourceMark rm;
2549 // Check the signatures.
2550 ciSignature* sig = callee()->signature();
2551 #ifdef ASSERT
2552 // Object getObject(Object base, int/long offset), etc.
2553 BasicType rtype = sig->return_type()->basic_type();
2554 if (!is_native_ptr) {
2555 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2556 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2557 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2558 } else {
2559 assert(sig->count() == 1, "native prefetch has 1 argument");
2560 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2561 }
2562 #endif // ASSERT
2563 }
2564 #endif // !PRODUCT
2565
2566 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2567
2568 const int idx = is_static ? 0 : 1;
2569 if (!is_static) {
2570 null_check_receiver();
2571 if (stopped()) {
2572 return true;
2573 }
2574 }
2575
2576 // Build address expression. See the code in inline_unsafe_access.
2577 Node *adr;
2578 if (!is_native_ptr) {
2579 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2580 Node* base = argument(idx + 0); // type: oop
2581 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2582 Node* offset = argument(idx + 1); // type: long
2583 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2584 // to be plain byte offsets, which are also the same as those accepted
2585 // by oopDesc::field_base.
2586 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2587 "fieldOffset must be byte-scaled");
2588 // 32-bit machines ignore the high half!
2589 offset = ConvL2X(offset);
2590 adr = make_unsafe_address(base, offset);
2591 } else {
2592 Node* ptr = argument(idx + 0); // type: long
2593 ptr = ConvL2X(ptr); // adjust Java long to machine word
2594 adr = make_unsafe_address(NULL, ptr);
2595 }
2596
2597 // Generate the read or write prefetch
2598 Node *prefetch;
2599 if (is_store) {
2600 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2601 } else {
2602 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2603 }
2604 prefetch->init_req(0, control());
2605 set_i_o(_gvn.transform(prefetch));
2606
2607 return true;
2608 }
2609
2610 //----------------------------inline_unsafe_load_store----------------------------
2611 // This method serves a couple of different customers (depending on LoadStoreKind):
2612 //
2613 // LS_cmpxchg:
2614 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2615 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2616 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2617 //
2618 // LS_xadd:
2619 // public int getAndAddInt( Object o, long offset, int delta)
2620 // public long getAndAddLong(Object o, long offset, long delta)
2621 //
2622 // LS_xchg:
2623 // int getAndSet(Object o, long offset, int newValue)
2624 // long getAndSet(Object o, long offset, long newValue)
2625 // Object getAndSet(Object o, long offset, Object newValue)
2626 //
2627 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2628 // This basic scheme here is the same as inline_unsafe_access, but
2629 // differs in enough details that combining them would make the code
2630 // overly confusing. (This is a true fact! I originally combined
2631 // them, but even I was confused by it!) As much code/comments as
2632 // possible are retained from inline_unsafe_access though to make
2633 // the correspondences clearer. - dl
2634
2635 if (callee()->is_static()) return false; // caller must have the capability!
2636
2637 #ifndef PRODUCT
2638 BasicType rtype;
2639 {
2640 ResourceMark rm;
2641 // Check the signatures.
2642 ciSignature* sig = callee()->signature();
2643 rtype = sig->return_type()->basic_type();
2644 if (kind == LS_xadd || kind == LS_xchg) {
2645 // Check the signatures.
2646 #ifdef ASSERT
2647 assert(rtype == type, "get and set must return the expected type");
2648 assert(sig->count() == 3, "get and set has 3 arguments");
2649 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2650 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2651 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2652 #endif // ASSERT
2653 } else if (kind == LS_cmpxchg) {
2654 // Check the signatures.
2655 #ifdef ASSERT
2656 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2657 assert(sig->count() == 4, "CAS has 4 arguments");
2658 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2659 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2660 #endif // ASSERT
2661 } else {
2662 ShouldNotReachHere();
2663 }
2664 }
2665 #endif //PRODUCT
2666
2667 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2668
2669 // Get arguments:
2670 Node* receiver = NULL;
2671 Node* base = NULL;
2672 Node* offset = NULL;
2673 Node* oldval = NULL;
2674 Node* newval = NULL;
2675 if (kind == LS_cmpxchg) {
2676 const bool two_slot_type = type2size[type] == 2;
2677 receiver = argument(0); // type: oop
2678 base = argument(1); // type: oop
2679 offset = argument(2); // type: long
2680 oldval = argument(4); // type: oop, int, or long
2681 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2682 } else if (kind == LS_xadd || kind == LS_xchg){
2683 receiver = argument(0); // type: oop
2684 base = argument(1); // type: oop
2685 offset = argument(2); // type: long
2686 oldval = NULL;
2687 newval = argument(4); // type: oop, int, or long
2688 }
2689
2690 // Null check receiver.
2691 receiver = null_check(receiver);
2692 if (stopped()) {
2693 return true;
2694 }
2695
2696 // Build field offset expression.
2697 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2698 // to be plain byte offsets, which are also the same as those accepted
2699 // by oopDesc::field_base.
2700 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2701 // 32-bit machines ignore the high half of long offsets
2702 offset = ConvL2X(offset);
2703 Node* adr = make_unsafe_address(base, offset);
2704 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2705
2706 // For CAS, unlike inline_unsafe_access, there seems no point in
2707 // trying to refine types. Just use the coarse types here.
2708 const Type *value_type = Type::get_const_basic_type(type);
2709 Compile::AliasType* alias_type = C->alias_type(adr_type);
2710 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2711
2712 if (kind == LS_xchg && type == T_OBJECT) {
2713 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2714 if (tjp != NULL) {
2715 value_type = tjp;
2716 }
2717 }
2718
2719 int alias_idx = C->get_alias_index(adr_type);
2720
2721 // Memory-model-wise, a LoadStore acts like a little synchronized
2722 // block, so needs barriers on each side. These don't translate
2723 // into actual barriers on most machines, but we still need rest of
2724 // compiler to respect ordering.
2725
2726 insert_mem_bar(Op_MemBarRelease);
2727 insert_mem_bar(Op_MemBarCPUOrder);
2728
2729 // 4984716: MemBars must be inserted before this
2730 // memory node in order to avoid a false
2731 // dependency which will confuse the scheduler.
2732 Node *mem = memory(alias_idx);
2733
2734 // For now, we handle only those cases that actually exist: ints,
2735 // longs, and Object. Adding others should be straightforward.
2736 Node* load_store;
2737 switch(type) {
2738 case T_INT:
2739 if (kind == LS_xadd) {
2740 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2741 } else if (kind == LS_xchg) {
2742 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2743 } else if (kind == LS_cmpxchg) {
2744 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2745 } else {
2746 ShouldNotReachHere();
2747 }
2748 break;
2749 case T_LONG:
2750 if (kind == LS_xadd) {
2751 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2752 } else if (kind == LS_xchg) {
2753 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2754 } else if (kind == LS_cmpxchg) {
2755 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2756 } else {
2757 ShouldNotReachHere();
2758 }
2759 break;
2760 case T_OBJECT:
2761 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2762 // could be delayed during Parse (for example, in adjust_map_after_if()).
2763 // Execute transformation here to avoid barrier generation in such case.
2764 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2765 newval = _gvn.makecon(TypePtr::NULL_PTR);
2766
2767 // Reference stores need a store barrier.
2768 pre_barrier(true /* do_load*/,
2769 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2770 NULL /* pre_val*/,
2771 T_OBJECT);
2772 #ifdef _LP64
2773 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2774 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2775 if (kind == LS_xchg) {
2776 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2777 newval_enc, adr_type, value_type->make_narrowoop()));
2778 } else {
2779 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2780 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2781 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2782 newval_enc, oldval_enc));
2783 }
2784 } else
2785 #endif
2786 {
2787 if (kind == LS_xchg) {
2788 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2789 } else {
2790 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2791 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2792 }
2793 }
2794 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2795 break;
2796 default:
2797 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2798 break;
2799 }
2800
2801 // SCMemProjNodes represent the memory state of a LoadStore. Their
2802 // main role is to prevent LoadStore nodes from being optimized away
2803 // when their results aren't used.
2804 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2805 set_memory(proj, alias_idx);
2806
2807 // Add the trailing membar surrounding the access
2808 insert_mem_bar(Op_MemBarCPUOrder);
2809 // On power we need a fence to prevent succeeding loads from floating
2810 // above the store of the compare-exchange.
2811 #ifdef PPC64
2812 insert_mem_bar(Op_MemBarVolatile);
2813 #else
2814 insert_mem_bar(Op_MemBarAcquire);
2815 #endif
2816
2817 #ifdef _LP64
2818 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2819 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2820 }
2821 #endif
2822
2823 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2824 set_result(load_store);
2825 return true;
2826 }
2827
2828 //----------------------------inline_unsafe_ordered_store----------------------
2829 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2830 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2831 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2832 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2833 // This is another variant of inline_unsafe_access, differing in
2834 // that it always issues store-store ("release") barrier and ensures
2835 // store-atomicity (which only matters for "long").
2836
2837 if (callee()->is_static()) return false; // caller must have the capability!
2838
2839 #ifndef PRODUCT
2840 {
2841 ResourceMark rm;
2842 // Check the signatures.
2843 ciSignature* sig = callee()->signature();
2844 #ifdef ASSERT
2845 BasicType rtype = sig->return_type()->basic_type();
2846 assert(rtype == T_VOID, "must return void");
2847 assert(sig->count() == 3, "has 3 arguments");
2848 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2849 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2850 #endif // ASSERT
2851 }
2852 #endif //PRODUCT
2853
2854 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2855
2856 // Get arguments:
2857 Node* receiver = argument(0); // type: oop
2858 Node* base = argument(1); // type: oop
2859 Node* offset = argument(2); // type: long
2860 Node* val = argument(4); // type: oop, int, or long
2861
2862 // Null check receiver.
2863 receiver = null_check(receiver);
2864 if (stopped()) {
2865 return true;
2866 }
2867
2868 // Build field offset expression.
2869 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2870 // 32-bit machines ignore the high half of long offsets
2871 offset = ConvL2X(offset);
2872 Node* adr = make_unsafe_address(base, offset);
2873 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2874 const Type *value_type = Type::get_const_basic_type(type);
2875 Compile::AliasType* alias_type = C->alias_type(adr_type);
2876
2877 insert_mem_bar(Op_MemBarRelease);
2878 insert_mem_bar(Op_MemBarCPUOrder);
2879 // Ensure that the store is atomic for longs:
2880 const bool require_atomic_access = true;
2881 Node* store;
2882 if (type == T_OBJECT) // reference stores need a store barrier.
2883 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2884 else {
2885 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2886 }
2887 insert_mem_bar(Op_MemBarCPUOrder);
2888 return true;
2889 }
2890
2891 //----------------------------inline_unsafe_allocate---------------------------
2892 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2893 bool LibraryCallKit::inline_unsafe_allocate() {
2894 if (callee()->is_static()) return false; // caller must have the capability!
2895
2896 null_check_receiver(); // null-check, then ignore
2897 Node* cls = null_check(argument(1));
2898 if (stopped()) return true;
2899
2900 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2901 kls = null_check(kls);
2902 if (stopped()) return true; // argument was like int.class
2903
2904 // Note: The argument might still be an illegal value like
2905 // Serializable.class or Object[].class. The runtime will handle it.
2906 // But we must make an explicit check for initialization.
2907 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2908 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2909 // can generate code to load it as unsigned byte.
2910 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2911 Node* bits = intcon(instanceKlass::fully_initialized);
2912 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2913 // The 'test' is non-zero if we need to take a slow path.
2914
2915 Node* obj = new_instance(kls, test);
2916 set_result(obj);
2917 return true;
2918 }
2919
2920 #ifdef TRACE_HAVE_INTRINSICS
2921 /*
2922 * oop -> myklass
2923 * myklass->trace_id |= USED
2924 * return myklass->trace_id & ~0x3
2925 */
2926 bool LibraryCallKit::inline_native_classID() {
2927 null_check_receiver(); // null-check, then ignore
2928 Node* cls = null_check(argument(1), T_OBJECT);
2929 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2930 kls = null_check(kls, T_OBJECT);
2931 ByteSize offset = TRACE_ID_OFFSET;
2932 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2933 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2934 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2935 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2936 Node* clsused = longcon(0x01l); // set the class bit
2937 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2938
2939 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2940 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2941 set_result(andl);
2942 return true;
2943 }
2944
2945 bool LibraryCallKit::inline_native_threadID() {
2946 Node* tls_ptr = NULL;
2947 Node* cur_thr = generate_current_thread(tls_ptr);
2948 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2949 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2950 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2951
2952 Node* threadid = NULL;
2953 size_t thread_id_size = OSThread::thread_id_size();
2954 if (thread_id_size == (size_t) BytesPerLong) {
2955 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2956 } else if (thread_id_size == (size_t) BytesPerInt) {
2957 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2958 } else {
2959 ShouldNotReachHere();
2960 }
2961 set_result(threadid);
2962 return true;
2963 }
2964 #endif
2965
2966 //------------------------inline_native_time_funcs--------------
2967 // inline code for System.currentTimeMillis() and System.nanoTime()
2968 // these have the same type and signature
2969 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2970 const TypeFunc* tf = OptoRuntime::void_long_Type();
2971 const TypePtr* no_memory_effects = NULL;
2972 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2973 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2974 #ifdef ASSERT
2975 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2976 assert(value_top == top(), "second value must be top");
2977 #endif
2978 set_result(value);
2979 return true;
2980 }
2981
2982 //------------------------inline_native_currentThread------------------
2983 bool LibraryCallKit::inline_native_currentThread() {
2984 Node* junk = NULL;
2985 set_result(generate_current_thread(junk));
2986 return true;
2987 }
2988
2989 //------------------------inline_native_isInterrupted------------------
2990 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2991 bool LibraryCallKit::inline_native_isInterrupted() {
2992 // Add a fast path to t.isInterrupted(clear_int):
2993 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2994 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2995 // So, in the common case that the interrupt bit is false,
2996 // we avoid making a call into the VM. Even if the interrupt bit
2997 // is true, if the clear_int argument is false, we avoid the VM call.
2998 // However, if the receiver is not currentThread, we must call the VM,
2999 // because there must be some locking done around the operation.
3000
3001 // We only go to the fast case code if we pass two guards.
3002 // Paths which do not pass are accumulated in the slow_region.
3003
3004 enum {
3005 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3006 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3007 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3008 PATH_LIMIT
3009 };
3010
3011 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3012 // out of the function.
3013 insert_mem_bar(Op_MemBarCPUOrder);
3014
3015 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3016 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3017
3018 RegionNode* slow_region = new (C) RegionNode(1);
3019 record_for_igvn(slow_region);
3020
3021 // (a) Receiving thread must be the current thread.
3022 Node* rec_thr = argument(0);
3023 Node* tls_ptr = NULL;
3024 Node* cur_thr = generate_current_thread(tls_ptr);
3025 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3026 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3027
3028 generate_slow_guard(bol_thr, slow_region);
3029
3030 // (b) Interrupt bit on TLS must be false.
3031 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3032 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3033 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3034
3035 // Set the control input on the field _interrupted read to prevent it floating up.
3036 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3037 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3038 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3039
3040 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3041
3042 // First fast path: if (!TLS._interrupted) return false;
3043 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3044 result_rgn->init_req(no_int_result_path, false_bit);
3045 result_val->init_req(no_int_result_path, intcon(0));
3046
3047 // drop through to next case
3048 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3049
3050 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3051 Node* clr_arg = argument(1);
3052 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3053 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3054 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3055
3056 // Second fast path: ... else if (!clear_int) return true;
3057 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3058 result_rgn->init_req(no_clear_result_path, false_arg);
3059 result_val->init_req(no_clear_result_path, intcon(1));
3060
3061 // drop through to next case
3062 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3063
3064 // (d) Otherwise, go to the slow path.
3065 slow_region->add_req(control());
3066 set_control( _gvn.transform(slow_region) );
3067
3068 if (stopped()) {
3069 // There is no slow path.
3070 result_rgn->init_req(slow_result_path, top());
3071 result_val->init_req(slow_result_path, top());
3072 } else {
3073 // non-virtual because it is a private non-static
3074 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3075
3076 Node* slow_val = set_results_for_java_call(slow_call);
3077 // this->control() comes from set_results_for_java_call
3078
3079 Node* fast_io = slow_call->in(TypeFunc::I_O);
3080 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3081
3082 // These two phis are pre-filled with copies of of the fast IO and Memory
3083 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3084 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3085
3086 result_rgn->init_req(slow_result_path, control());
3087 result_io ->init_req(slow_result_path, i_o());
3088 result_mem->init_req(slow_result_path, reset_memory());
3089 result_val->init_req(slow_result_path, slow_val);
3090
3091 set_all_memory(_gvn.transform(result_mem));
3092 set_i_o( _gvn.transform(result_io));
3093 }
3094
3095 C->set_has_split_ifs(true); // Has chance for split-if optimization
3096 set_result(result_rgn, result_val);
3097 return true;
3098 }
3099
3100 //---------------------------load_mirror_from_klass----------------------------
3101 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3102 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3103 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3104 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3105 }
3106
3107 //-----------------------load_klass_from_mirror_common-------------------------
3108 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3109 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3110 // and branch to the given path on the region.
3111 // If never_see_null, take an uncommon trap on null, so we can optimistically
3112 // compile for the non-null case.
3113 // If the region is NULL, force never_see_null = true.
3114 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3115 bool never_see_null,
3116 RegionNode* region,
3117 int null_path,
3118 int offset) {
3119 if (region == NULL) never_see_null = true;
3120 Node* p = basic_plus_adr(mirror, offset);
3121 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3122 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3123 Node* null_ctl = top();
3124 kls = null_check_oop(kls, &null_ctl, never_see_null);
3125 if (region != NULL) {
3126 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3127 region->init_req(null_path, null_ctl);
3128 } else {
3129 assert(null_ctl == top(), "no loose ends");
3130 }
3131 return kls;
3132 }
3133
3134 //--------------------(inline_native_Class_query helpers)---------------------
3135 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3136 // Fall through if (mods & mask) == bits, take the guard otherwise.
3137 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3138 // Branch around if the given klass has the given modifier bit set.
3139 // Like generate_guard, adds a new path onto the region.
3140 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3141 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3142 Node* mask = intcon(modifier_mask);
3143 Node* bits = intcon(modifier_bits);
3144 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3145 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3146 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3147 return generate_fair_guard(bol, region);
3148 }
3149 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3150 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3151 }
3152
3153 //-------------------------inline_native_Class_query-------------------
3154 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3155 const Type* return_type = TypeInt::BOOL;
3156 Node* prim_return_value = top(); // what happens if it's a primitive class?
3157 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3158 bool expect_prim = false; // most of these guys expect to work on refs
3159
3160 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3161
3162 Node* mirror = argument(0);
3163 Node* obj = top();
3164
3165 switch (id) {
3166 case vmIntrinsics::_isInstance:
3167 // nothing is an instance of a primitive type
3168 prim_return_value = intcon(0);
3169 obj = argument(1);
3170 break;
3171 case vmIntrinsics::_getModifiers:
3172 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3173 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3174 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3175 break;
3176 case vmIntrinsics::_isInterface:
3177 prim_return_value = intcon(0);
3178 break;
3179 case vmIntrinsics::_isArray:
3180 prim_return_value = intcon(0);
3181 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3182 break;
3183 case vmIntrinsics::_isPrimitive:
3184 prim_return_value = intcon(1);
3185 expect_prim = true; // obviously
3186 break;
3187 case vmIntrinsics::_getSuperclass:
3188 prim_return_value = null();
3189 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3190 break;
3191 case vmIntrinsics::_getComponentType:
3192 prim_return_value = null();
3193 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3194 break;
3195 case vmIntrinsics::_getClassAccessFlags:
3196 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3197 return_type = TypeInt::INT; // not bool! 6297094
3198 break;
3199 default:
3200 fatal_unexpected_iid(id);
3201 break;
3202 }
3203
3204 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3205 if (mirror_con == NULL) return false; // cannot happen?
3206
3207 #ifndef PRODUCT
3208 if (C->print_intrinsics() || C->print_inlining()) {
3209 ciType* k = mirror_con->java_mirror_type();
3210 if (k) {
3211 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3212 k->print_name();
3213 tty->cr();
3214 }
3215 }
3216 #endif
3217
3218 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3219 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3220 record_for_igvn(region);
3221 PhiNode* phi = new (C) PhiNode(region, return_type);
3222
3223 // The mirror will never be null of Reflection.getClassAccessFlags, however
3224 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3225 // if it is. See bug 4774291.
3226
3227 // For Reflection.getClassAccessFlags(), the null check occurs in
3228 // the wrong place; see inline_unsafe_access(), above, for a similar
3229 // situation.
3230 mirror = null_check(mirror);
3231 // If mirror or obj is dead, only null-path is taken.
3232 if (stopped()) return true;
3233
3234 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3235
3236 // Now load the mirror's klass metaobject, and null-check it.
3237 // Side-effects region with the control path if the klass is null.
3238 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3239 // If kls is null, we have a primitive mirror.
3240 phi->init_req(_prim_path, prim_return_value);
3241 if (stopped()) { set_result(region, phi); return true; }
3242
3243 Node* p; // handy temp
3244 Node* null_ctl;
3245
3246 // Now that we have the non-null klass, we can perform the real query.
3247 // For constant classes, the query will constant-fold in LoadNode::Value.
3248 Node* query_value = top();
3249 switch (id) {
3250 case vmIntrinsics::_isInstance:
3251 // nothing is an instance of a primitive type
3252 query_value = gen_instanceof(obj, kls);
3253 break;
3254
3255 case vmIntrinsics::_getModifiers:
3256 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3257 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3258 break;
3259
3260 case vmIntrinsics::_isInterface:
3261 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3262 if (generate_interface_guard(kls, region) != NULL)
3263 // A guard was added. If the guard is taken, it was an interface.
3264 phi->add_req(intcon(1));
3265 // If we fall through, it's a plain class.
3266 query_value = intcon(0);
3267 break;
3268
3269 case vmIntrinsics::_isArray:
3270 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3271 if (generate_array_guard(kls, region) != NULL)
3272 // A guard was added. If the guard is taken, it was an array.
3273 phi->add_req(intcon(1));
3274 // If we fall through, it's a plain class.
3275 query_value = intcon(0);
3276 break;
3277
3278 case vmIntrinsics::_isPrimitive:
3279 query_value = intcon(0); // "normal" path produces false
3280 break;
3281
3282 case vmIntrinsics::_getSuperclass:
3283 // The rules here are somewhat unfortunate, but we can still do better
3284 // with random logic than with a JNI call.
3285 // Interfaces store null or Object as _super, but must report null.
3286 // Arrays store an intermediate super as _super, but must report Object.
3287 // Other types can report the actual _super.
3288 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3289 if (generate_interface_guard(kls, region) != NULL)
3290 // A guard was added. If the guard is taken, it was an interface.
3291 phi->add_req(null());
3292 if (generate_array_guard(kls, region) != NULL)
3293 // A guard was added. If the guard is taken, it was an array.
3294 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3295 // If we fall through, it's a plain class. Get its _super.
3296 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3297 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3298 null_ctl = top();
3299 kls = null_check_oop(kls, &null_ctl);
3300 if (null_ctl != top()) {
3301 // If the guard is taken, Object.superClass is null (both klass and mirror).
3302 region->add_req(null_ctl);
3303 phi ->add_req(null());
3304 }
3305 if (!stopped()) {
3306 query_value = load_mirror_from_klass(kls);
3307 }
3308 break;
3309
3310 case vmIntrinsics::_getComponentType:
3311 if (generate_array_guard(kls, region) != NULL) {
3312 // Be sure to pin the oop load to the guard edge just created:
3313 Node* is_array_ctrl = region->in(region->req()-1);
3314 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3315 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3316 phi->add_req(cmo);
3317 }
3318 query_value = null(); // non-array case is null
3319 break;
3320
3321 case vmIntrinsics::_getClassAccessFlags:
3322 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3323 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3324 break;
3325
3326 default:
3327 fatal_unexpected_iid(id);
3328 break;
3329 }
3330
3331 // Fall-through is the normal case of a query to a real class.
3332 phi->init_req(1, query_value);
3333 region->init_req(1, control());
3334
3335 C->set_has_split_ifs(true); // Has chance for split-if optimization
3336 set_result(region, phi);
3337 return true;
3338 }
3339
3340 //--------------------------inline_native_subtype_check------------------------
3341 // This intrinsic takes the JNI calls out of the heart of
3342 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3343 bool LibraryCallKit::inline_native_subtype_check() {
3344 // Pull both arguments off the stack.
3345 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3346 args[0] = argument(0);
3347 args[1] = argument(1);
3348 Node* klasses[2]; // corresponding Klasses: superk, subk
3349 klasses[0] = klasses[1] = top();
3350
3351 enum {
3352 // A full decision tree on {superc is prim, subc is prim}:
3353 _prim_0_path = 1, // {P,N} => false
3354 // {P,P} & superc!=subc => false
3355 _prim_same_path, // {P,P} & superc==subc => true
3356 _prim_1_path, // {N,P} => false
3357 _ref_subtype_path, // {N,N} & subtype check wins => true
3358 _both_ref_path, // {N,N} & subtype check loses => false
3359 PATH_LIMIT
3360 };
3361
3362 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3363 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3364 record_for_igvn(region);
3365
3366 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3367 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3368 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3369
3370 // First null-check both mirrors and load each mirror's klass metaobject.
3371 int which_arg;
3372 for (which_arg = 0; which_arg <= 1; which_arg++) {
3373 Node* arg = args[which_arg];
3374 arg = null_check(arg);
3375 if (stopped()) break;
3376 args[which_arg] = arg;
3377
3378 Node* p = basic_plus_adr(arg, class_klass_offset);
3379 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3380 klasses[which_arg] = _gvn.transform(kls);
3381 }
3382
3383 // Having loaded both klasses, test each for null.
3384 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3385 for (which_arg = 0; which_arg <= 1; which_arg++) {
3386 Node* kls = klasses[which_arg];
3387 Node* null_ctl = top();
3388 kls = null_check_oop(kls, &null_ctl, never_see_null);
3389 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3390 region->init_req(prim_path, null_ctl);
3391 if (stopped()) break;
3392 klasses[which_arg] = kls;
3393 }
3394
3395 if (!stopped()) {
3396 // now we have two reference types, in klasses[0..1]
3397 Node* subk = klasses[1]; // the argument to isAssignableFrom
3398 Node* superk = klasses[0]; // the receiver
3399 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3400 // now we have a successful reference subtype check
3401 region->set_req(_ref_subtype_path, control());
3402 }
3403
3404 // If both operands are primitive (both klasses null), then
3405 // we must return true when they are identical primitives.
3406 // It is convenient to test this after the first null klass check.
3407 set_control(region->in(_prim_0_path)); // go back to first null check
3408 if (!stopped()) {
3409 // Since superc is primitive, make a guard for the superc==subc case.
3410 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3411 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3412 generate_guard(bol_eq, region, PROB_FAIR);
3413 if (region->req() == PATH_LIMIT+1) {
3414 // A guard was added. If the added guard is taken, superc==subc.
3415 region->swap_edges(PATH_LIMIT, _prim_same_path);
3416 region->del_req(PATH_LIMIT);
3417 }
3418 region->set_req(_prim_0_path, control()); // Not equal after all.
3419 }
3420
3421 // these are the only paths that produce 'true':
3422 phi->set_req(_prim_same_path, intcon(1));
3423 phi->set_req(_ref_subtype_path, intcon(1));
3424
3425 // pull together the cases:
3426 assert(region->req() == PATH_LIMIT, "sane region");
3427 for (uint i = 1; i < region->req(); i++) {
3428 Node* ctl = region->in(i);
3429 if (ctl == NULL || ctl == top()) {
3430 region->set_req(i, top());
3431 phi ->set_req(i, top());
3432 } else if (phi->in(i) == NULL) {
3433 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3434 }
3435 }
3436
3437 set_control(_gvn.transform(region));
3438 set_result(_gvn.transform(phi));
3439 return true;
3440 }
3441
3442 //---------------------generate_array_guard_common------------------------
3443 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3444 bool obj_array, bool not_array) {
3445 // If obj_array/non_array==false/false:
3446 // Branch around if the given klass is in fact an array (either obj or prim).
3447 // If obj_array/non_array==false/true:
3448 // Branch around if the given klass is not an array klass of any kind.
3449 // If obj_array/non_array==true/true:
3450 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3451 // If obj_array/non_array==true/false:
3452 // Branch around if the kls is an oop array (Object[] or subtype)
3453 //
3454 // Like generate_guard, adds a new path onto the region.
3455 jint layout_con = 0;
3456 Node* layout_val = get_layout_helper(kls, layout_con);
3457 if (layout_val == NULL) {
3458 bool query = (obj_array
3459 ? Klass::layout_helper_is_objArray(layout_con)
3460 : Klass::layout_helper_is_javaArray(layout_con));
3461 if (query == not_array) {
3462 return NULL; // never a branch
3463 } else { // always a branch
3464 Node* always_branch = control();
3465 if (region != NULL)
3466 region->add_req(always_branch);
3467 set_control(top());
3468 return always_branch;
3469 }
3470 }
3471 // Now test the correct condition.
3472 jint nval = (obj_array
3473 ? ((jint)Klass::_lh_array_tag_type_value
3474 << Klass::_lh_array_tag_shift)
3475 : Klass::_lh_neutral_value);
3476 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3477 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3478 // invert the test if we are looking for a non-array
3479 if (not_array) btest = BoolTest(btest).negate();
3480 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3481 return generate_fair_guard(bol, region);
3482 }
3483
3484
3485 //-----------------------inline_native_newArray--------------------------
3486 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3487 bool LibraryCallKit::inline_native_newArray() {
3488 Node* mirror = argument(0);
3489 Node* count_val = argument(1);
3490
3491 mirror = null_check(mirror);
3492 // If mirror or obj is dead, only null-path is taken.
3493 if (stopped()) return true;
3494
3495 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3496 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3497 PhiNode* result_val = new(C) PhiNode(result_reg,
3498 TypeInstPtr::NOTNULL);
3499 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3500 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3501 TypePtr::BOTTOM);
3502
3503 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3504 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3505 result_reg, _slow_path);
3506 Node* normal_ctl = control();
3507 Node* no_array_ctl = result_reg->in(_slow_path);
3508
3509 // Generate code for the slow case. We make a call to newArray().
3510 set_control(no_array_ctl);
3511 if (!stopped()) {
3512 // Either the input type is void.class, or else the
3513 // array klass has not yet been cached. Either the
3514 // ensuing call will throw an exception, or else it
3515 // will cache the array klass for next time.
3516 PreserveJVMState pjvms(this);
3517 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3518 Node* slow_result = set_results_for_java_call(slow_call);
3519 // this->control() comes from set_results_for_java_call
3520 result_reg->set_req(_slow_path, control());
3521 result_val->set_req(_slow_path, slow_result);
3522 result_io ->set_req(_slow_path, i_o());
3523 result_mem->set_req(_slow_path, reset_memory());
3524 }
3525
3526 set_control(normal_ctl);
3527 if (!stopped()) {
3528 // Normal case: The array type has been cached in the java.lang.Class.
3529 // The following call works fine even if the array type is polymorphic.
3530 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3531 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3532 result_reg->init_req(_normal_path, control());
3533 result_val->init_req(_normal_path, obj);
3534 result_io ->init_req(_normal_path, i_o());
3535 result_mem->init_req(_normal_path, reset_memory());
3536 }
3537
3538 // Return the combined state.
3539 set_i_o( _gvn.transform(result_io) );
3540 set_all_memory( _gvn.transform(result_mem) );
3541
3542 C->set_has_split_ifs(true); // Has chance for split-if optimization
3543 set_result(result_reg, result_val);
3544 return true;
3545 }
3546
3547 //----------------------inline_native_getLength--------------------------
3548 // public static native int java.lang.reflect.Array.getLength(Object array);
3549 bool LibraryCallKit::inline_native_getLength() {
3550 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3551
3552 Node* array = null_check(argument(0));
3553 // If array is dead, only null-path is taken.
3554 if (stopped()) return true;
3555
3556 // Deoptimize if it is a non-array.
3557 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3558
3559 if (non_array != NULL) {
3560 PreserveJVMState pjvms(this);
3561 set_control(non_array);
3562 uncommon_trap(Deoptimization::Reason_intrinsic,
3563 Deoptimization::Action_maybe_recompile);
3564 }
3565
3566 // If control is dead, only non-array-path is taken.
3567 if (stopped()) return true;
3568
3569 // The works fine even if the array type is polymorphic.
3570 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3571 Node* result = load_array_length(array);
3572
3573 C->set_has_split_ifs(true); // Has chance for split-if optimization
3574 set_result(result);
3575 return true;
3576 }
3577
3578 //------------------------inline_array_copyOf----------------------------
3579 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3580 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3581 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3582 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3583
3584 // Get the arguments.
3585 Node* original = argument(0);
3586 Node* start = is_copyOfRange? argument(1): intcon(0);
3587 Node* end = is_copyOfRange? argument(2): argument(1);
3588 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3589
3590 Node* newcopy;
3591
3592 // Set the original stack and the reexecute bit for the interpreter to reexecute
3593 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3594 { PreserveReexecuteState preexecs(this);
3595 jvms()->set_should_reexecute(true);
3596
3597 array_type_mirror = null_check(array_type_mirror);
3598 original = null_check(original);
3599
3600 // Check if a null path was taken unconditionally.
3601 if (stopped()) return true;
3602
3603 Node* orig_length = load_array_length(original);
3604
3605 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3606 klass_node = null_check(klass_node);
3607
3608 RegionNode* bailout = new (C) RegionNode(1);
3609 record_for_igvn(bailout);
3610
3611 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3612 // Bail out if that is so.
3613 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3614 if (not_objArray != NULL) {
3615 // Improve the klass node's type from the new optimistic assumption:
3616 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3617 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3618 Node* cast = new (C) CastPPNode(klass_node, akls);
3619 cast->init_req(0, control());
3620 klass_node = _gvn.transform(cast);
3621 }
3622
3623 // Bail out if either start or end is negative.
3624 generate_negative_guard(start, bailout, &start);
3625 generate_negative_guard(end, bailout, &end);
3626
3627 Node* length = end;
3628 if (_gvn.type(start) != TypeInt::ZERO) {
3629 length = _gvn.transform(new (C) SubINode(end, start));
3630 }
3631
3632 // Bail out if length is negative.
3633 // Without this the new_array would throw
3634 // NegativeArraySizeException but IllegalArgumentException is what
3635 // should be thrown
3636 generate_negative_guard(length, bailout, &length);
3637
3638 if (bailout->req() > 1) {
3639 PreserveJVMState pjvms(this);
3640 set_control(_gvn.transform(bailout));
3641 uncommon_trap(Deoptimization::Reason_intrinsic,
3642 Deoptimization::Action_maybe_recompile);
3643 }
3644
3645 if (!stopped()) {
3646 // How many elements will we copy from the original?
3647 // The answer is MinI(orig_length - start, length).
3648 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3649 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3650
3651 newcopy = new_array(klass_node, length, 0); // no argments to push
3652
3653 // Generate a direct call to the right arraycopy function(s).
3654 // We know the copy is disjoint but we might not know if the
3655 // oop stores need checking.
3656 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3657 // This will fail a store-check if x contains any non-nulls.
3658 bool disjoint_bases = true;
3659 // if start > orig_length then the length of the copy may be
3660 // negative.
3661 bool length_never_negative = !is_copyOfRange;
3662 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3663 original, start, newcopy, intcon(0), moved,
3664 disjoint_bases, length_never_negative);
3665 }
3666 } // original reexecute is set back here
3667
3668 C->set_has_split_ifs(true); // Has chance for split-if optimization
3669 if (!stopped()) {
3670 set_result(newcopy);
3671 }
3672 return true;
3673 }
3674
3675
3676 //----------------------generate_virtual_guard---------------------------
3677 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3678 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3679 RegionNode* slow_region) {
3680 ciMethod* method = callee();
3681 int vtable_index = method->vtable_index();
3682 // Get the methodOop out of the appropriate vtable entry.
3683 int entry_offset = (instanceKlass::vtable_start_offset() +
3684 vtable_index*vtableEntry::size()) * wordSize +
3685 vtableEntry::method_offset_in_bytes();
3686 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3687 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3688
3689 // Compare the target method with the expected method (e.g., Object.hashCode).
3690 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3691
3692 Node* native_call = makecon(native_call_addr);
3693 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3694 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3695
3696 return generate_slow_guard(test_native, slow_region);
3697 }
3698
3699 //-----------------------generate_method_call----------------------------
3700 // Use generate_method_call to make a slow-call to the real
3701 // method if the fast path fails. An alternative would be to
3702 // use a stub like OptoRuntime::slow_arraycopy_Java.
3703 // This only works for expanding the current library call,
3704 // not another intrinsic. (E.g., don't use this for making an
3705 // arraycopy call inside of the copyOf intrinsic.)
3706 CallJavaNode*
3707 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3708 // When compiling the intrinsic method itself, do not use this technique.
3709 guarantee(callee() != C->method(), "cannot make slow-call to self");
3710
3711 ciMethod* method = callee();
3712 // ensure the JVMS we have will be correct for this call
3713 guarantee(method_id == method->intrinsic_id(), "must match");
3714
3715 const TypeFunc* tf = TypeFunc::make(method);
3716 CallJavaNode* slow_call;
3717 if (is_static) {
3718 assert(!is_virtual, "");
3719 slow_call = new(C) CallStaticJavaNode(tf,
3720 SharedRuntime::get_resolve_static_call_stub(),
3721 method, bci());
3722 } else if (is_virtual) {
3723 null_check_receiver();
3724 int vtable_index = methodOopDesc::invalid_vtable_index;
3725 if (UseInlineCaches) {
3726 // Suppress the vtable call
3727 } else {
3728 // hashCode and clone are not a miranda methods,
3729 // so the vtable index is fixed.
3730 // No need to use the linkResolver to get it.
3731 vtable_index = method->vtable_index();
3732 }
3733 slow_call = new(C) CallDynamicJavaNode(tf,
3734 SharedRuntime::get_resolve_virtual_call_stub(),
3735 method, vtable_index, bci());
3736 } else { // neither virtual nor static: opt_virtual
3737 null_check_receiver();
3738 slow_call = new(C) CallStaticJavaNode(tf,
3739 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3740 method, bci());
3741 slow_call->set_optimized_virtual(true);
3742 }
3743 set_arguments_for_java_call(slow_call);
3744 set_edges_for_java_call(slow_call);
3745 return slow_call;
3746 }
3747
3748
3749 /**
3750 * Build special case code for calls to hashCode on an object. This call may
3751 * be virtual (invokevirtual) or bound (invokespecial). For each case we generate
3752 * slightly different code.
3753 */
3754 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3755 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3756 assert(!(is_virtual && is_static), "either virtual, special, or static");
3757
3758 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3759
3760 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3761 PhiNode* result_val = new(C) PhiNode(result_reg, TypeInt::INT);
3762 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3763 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY, TypePtr::BOTTOM);
3764 Node* obj = NULL;
3765 if (!is_static) {
3766 // Check for hashing null object
3767 obj = null_check_receiver();
3768 if (stopped()) return true; // unconditionally null
3769 result_reg->init_req(_null_path, top());
3770 result_val->init_req(_null_path, top());
3771 } else {
3772 // Do a null check, and return zero if null.
3773 // System.identityHashCode(null) == 0
3774 obj = argument(0);
3775 Node* null_ctl = top();
3776 obj = null_check_oop(obj, &null_ctl);
3777 result_reg->init_req(_null_path, null_ctl);
3778 result_val->init_req(_null_path, _gvn.intcon(0));
3779 }
3780
3781 // Unconditionally null? Then return right away.
3782 if (stopped()) {
3783 set_control( result_reg->in(_null_path));
3784 if (!stopped())
3785 set_result(result_val->in(_null_path));
3786 return true;
3787 }
3788
3789 // We only go to the fast case code if we pass a number of guards. The
3790 // paths which do not pass are accumulated in the slow_region.
3791 RegionNode* slow_region = new (C) RegionNode(1);
3792 record_for_igvn(slow_region);
3793
3794 // If this is a virtual call, we generate a funny guard. We pull out
3795 // the vtable entry corresponding to hashCode() from the target object.
3796 // If the target method which we are calling happens to be the native
3797 // Object hashCode() method, we pass the guard. We do not need this
3798 // guard for non-virtual calls -- the caller is known to be the native
3799 // Object hashCode().
3800 if (is_virtual) {
3801 // After null check, get the object's klass.
3802 Node* obj_klass = load_object_klass(obj);
3803 generate_virtual_guard(obj_klass, slow_region);
3804 }
3805
3806 // Get the header out of the object, use LoadMarkNode when available
3807 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3808 // The control of the load must be NULL. Otherwise, the load can move before
3809 // the null check after castPP removal.
3810 Node* no_ctrl = NULL;
3811 Node* header = make_load(no_ctrl, header_addr, TypeX_X, TypeX_X->basic_type());
3812
3813 // Test the header to see if it is unlocked.
3814 Node* lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3815 Node* lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
3816 Node* unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3817 Node* chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
3818 Node* test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
3819
3820 generate_slow_guard(test_unlocked, slow_region);
3821
3822 // Get the hash value and check to see that it has been properly assigned.
3823 // We depend on hash_mask being at most 32 bits and avoid the use of
3824 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3825 // vm: see markOop.hpp.
3826 Node* hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3827 Node* hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3828 Node* hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
3829 // This hack lets the hash bits live anywhere in the mark object now, as long
3830 // as the shift drops the relevant bits into the low 32 bits. Note that
3831 // Java spec says that HashCode is an int so there's no point in capturing
3832 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3833 hshifted_header = ConvX2I(hshifted_header);
3834 Node* hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
3835
3836 Node* no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3837 Node* chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
3838 Node* test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
3839
3840 generate_slow_guard(test_assigned, slow_region);
3841
3842 Node* init_mem = reset_memory();
3843 // fill in the rest of the null path:
3844 result_io ->init_req(_null_path, i_o());
3845 result_mem->init_req(_null_path, init_mem);
3846
3847 result_val->init_req(_fast_path, hash_val);
3848 result_reg->init_req(_fast_path, control());
3849 result_io ->init_req(_fast_path, i_o());
3850 result_mem->init_req(_fast_path, init_mem);
3851
3852 // Generate code for the slow case. We make a call to hashCode().
3853 set_control(_gvn.transform(slow_region));
3854 if (!stopped()) {
3855 // No need for PreserveJVMState, because we're using up the present state.
3856 set_all_memory(init_mem);
3857 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3858 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3859 Node* slow_result = set_results_for_java_call(slow_call);
3860 // this->control() comes from set_results_for_java_call
3861 result_reg->init_req(_slow_path, control());
3862 result_val->init_req(_slow_path, slow_result);
3863 result_io ->set_req(_slow_path, i_o());
3864 result_mem ->set_req(_slow_path, reset_memory());
3865 }
3866
3867 // Return the combined state.
3868 set_i_o( _gvn.transform(result_io) );
3869 set_all_memory( _gvn.transform(result_mem) );
3870
3871 set_result(result_reg, result_val);
3872 return true;
3873 }
3874
3875 //---------------------------inline_native_getClass----------------------------
3876 // public final native Class<?> java.lang.Object.getClass();
3877 //
3878 // Build special case code for calls to getClass on an object.
3879 bool LibraryCallKit::inline_native_getClass() {
3880 Node* obj = null_check_receiver();
3881 if (stopped()) return true;
3882 set_result(load_mirror_from_klass(load_object_klass(obj)));
3883 return true;
3884 }
3885
3886 //-----------------inline_native_Reflection_getCallerClass---------------------
3887 // public static native Class<?> sun.reflect.Reflection.getCallerClass(int realFramesToSkip);
3888 //
3889 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3890 //
3891 // NOTE that this code must perform the same logic as
3892 // vframeStream::security_get_caller_frame in that it must skip
3893 // Method.invoke() and auxiliary frames.
3894 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3895 #ifndef PRODUCT
3896 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3897 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3898 }
3899 #endif
3900
3901 Node* caller_depth_node = argument(0);
3902
3903 // The depth value must be a constant in order for the runtime call
3904 // to be eliminated.
3905 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3906 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3907 #ifndef PRODUCT
3908 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3909 tty->print_cr(" Bailing out because caller depth was not a constant");
3910 }
3911 #endif
3912 return false;
3913 }
3914 // Note that the JVM state at this point does not include the
3915 // getCallerClass() frame which we are trying to inline. The
3916 // semantics of getCallerClass(), however, are that the "first"
3917 // frame is the getCallerClass() frame, so we subtract one from the
3918 // requested depth before continuing. We don't inline requests of
3919 // getCallerClass(0).
3920 int caller_depth = caller_depth_type->get_con() - 1;
3921 if (caller_depth < 0) {
3922 #ifndef PRODUCT
3923 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3924 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3925 }
3926 #endif
3927 return false;
3928 }
3929
3930 if (!jvms()->has_method()) {
3931 #ifndef PRODUCT
3932 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3933 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3934 }
3935 #endif
3936 return false;
3937 }
3938 int _depth = jvms()->depth(); // cache call chain depth
3939
3940 // Walk back up the JVM state to find the caller at the required
3941 // depth. NOTE that this code must perform the same logic as
3942 // vframeStream::security_get_caller_frame in that it must skip
3943 // Method.invoke() and auxiliary frames. Note also that depth is
3944 // 1-based (1 is the bottom of the inlining).
3945 int inlining_depth = _depth;
3946 JVMState* caller_jvms = NULL;
3947
3948 if (inlining_depth > 0) {
3949 caller_jvms = jvms();
3950 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3951 do {
3952 // The following if-tests should be performed in this order
3953 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3954 // Skip a Method.invoke() or auxiliary frame
3955 } else if (caller_depth > 0) {
3956 // Skip real frame
3957 --caller_depth;
3958 } else {
3959 // We're done: reached desired caller after skipping.
3960 break;
3961 }
3962 caller_jvms = caller_jvms->caller();
3963 --inlining_depth;
3964 } while (inlining_depth > 0);
3965 }
3966
3967 if (inlining_depth == 0) {
3968 #ifndef PRODUCT
3969 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3970 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3971 tty->print_cr(" JVM state at this point:");
3972 for (int i = _depth; i >= 1; i--) {
3973 ciMethod* m = jvms()->of_depth(i)->method();
3974 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3975 }
3976 }
3977 #endif
3978 return false; // Reached end of inlining
3979 }
3980
3981 // Acquire method holder as java.lang.Class
3982 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3983 ciInstance* caller_mirror = caller_klass->java_mirror();
3984
3985 // Push this as a constant
3986 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3987
3988 #ifndef PRODUCT
3989 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
3990 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
3991 tty->print_cr(" JVM state at this point:");
3992 for (int i = _depth; i >= 1; i--) {
3993 ciMethod* m = jvms()->of_depth(i)->method();
3994 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3995 }
3996 }
3997 #endif
3998 return true;
3999 }
4000
4001 // Helper routine for above
4002 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
4003 ciMethod* method = jvms->method();
4004
4005 // Is this the Method.invoke method itself?
4006 if (method->intrinsic_id() == vmIntrinsics::_invoke)
4007 return true;
4008
4009 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
4010 ciKlass* k = method->holder();
4011 if (k->is_instance_klass()) {
4012 ciInstanceKlass* ik = k->as_instance_klass();
4013 for (; ik != NULL; ik = ik->super()) {
4014 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
4015 ik == env()->find_system_klass(ik->name())) {
4016 return true;
4017 }
4018 }
4019 }
4020
4021 if (method->is_method_handle_intrinsic() ||
4022 method->is_compiled_lambda_form()) {
4023 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4024 return true;
4025 }
4026
4027 return false;
4028 }
4029
4030 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4031 Node* arg = argument(0);
4032 Node* result;
4033
4034 switch (id) {
4035 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4036 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4037 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4038 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4039
4040 case vmIntrinsics::_doubleToLongBits: {
4041 // two paths (plus control) merge in a wood
4042 RegionNode *r = new (C) RegionNode(3);
4043 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4044
4045 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4046 // Build the boolean node
4047 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4048
4049 // Branch either way.
4050 // NaN case is less traveled, which makes all the difference.
4051 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4052 Node *opt_isnan = _gvn.transform(ifisnan);
4053 assert( opt_isnan->is_If(), "Expect an IfNode");
4054 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4055 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4056
4057 set_control(iftrue);
4058
4059 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4060 Node *slow_result = longcon(nan_bits); // return NaN
4061 phi->init_req(1, _gvn.transform( slow_result ));
4062 r->init_req(1, iftrue);
4063
4064 // Else fall through
4065 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4066 set_control(iffalse);
4067
4068 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4069 r->init_req(2, iffalse);
4070
4071 // Post merge
4072 set_control(_gvn.transform(r));
4073 record_for_igvn(r);
4074
4075 C->set_has_split_ifs(true); // Has chance for split-if optimization
4076 result = phi;
4077 assert(result->bottom_type()->isa_long(), "must be");
4078 break;
4079 }
4080
4081 case vmIntrinsics::_floatToIntBits: {
4082 // two paths (plus control) merge in a wood
4083 RegionNode *r = new (C) RegionNode(3);
4084 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4085
4086 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4087 // Build the boolean node
4088 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4089
4090 // Branch either way.
4091 // NaN case is less traveled, which makes all the difference.
4092 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4093 Node *opt_isnan = _gvn.transform(ifisnan);
4094 assert( opt_isnan->is_If(), "Expect an IfNode");
4095 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4096 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4097
4098 set_control(iftrue);
4099
4100 static const jint nan_bits = 0x7fc00000;
4101 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4102 phi->init_req(1, _gvn.transform( slow_result ));
4103 r->init_req(1, iftrue);
4104
4105 // Else fall through
4106 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4107 set_control(iffalse);
4108
4109 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4110 r->init_req(2, iffalse);
4111
4112 // Post merge
4113 set_control(_gvn.transform(r));
4114 record_for_igvn(r);
4115
4116 C->set_has_split_ifs(true); // Has chance for split-if optimization
4117 result = phi;
4118 assert(result->bottom_type()->isa_int(), "must be");
4119 break;
4120 }
4121
4122 default:
4123 fatal_unexpected_iid(id);
4124 break;
4125 }
4126 set_result(_gvn.transform(result));
4127 return true;
4128 }
4129
4130 #ifdef _LP64
4131 #define XTOP ,top() /*additional argument*/
4132 #else //_LP64
4133 #define XTOP /*no additional argument*/
4134 #endif //_LP64
4135
4136 //----------------------inline_unsafe_copyMemory-------------------------
4137 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4138 bool LibraryCallKit::inline_unsafe_copyMemory() {
4139 if (callee()->is_static()) return false; // caller must have the capability!
4140 null_check_receiver(); // null-check receiver
4141 if (stopped()) return true;
4142
4143 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4144
4145 Node* src_ptr = argument(1); // type: oop
4146 Node* src_off = ConvL2X(argument(2)); // type: long
4147 Node* dst_ptr = argument(4); // type: oop
4148 Node* dst_off = ConvL2X(argument(5)); // type: long
4149 Node* size = ConvL2X(argument(7)); // type: long
4150
4151 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4152 "fieldOffset must be byte-scaled");
4153
4154 Node* src = make_unsafe_address(src_ptr, src_off);
4155 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4156
4157 // Conservatively insert a memory barrier on all memory slices.
4158 // Do not let writes of the copy source or destination float below the copy.
4159 insert_mem_bar(Op_MemBarCPUOrder);
4160
4161 // Call it. Note that the length argument is not scaled.
4162 make_runtime_call(RC_LEAF|RC_NO_FP,
4163 OptoRuntime::fast_arraycopy_Type(),
4164 StubRoutines::unsafe_arraycopy(),
4165 "unsafe_arraycopy",
4166 TypeRawPtr::BOTTOM,
4167 src, dst, size XTOP);
4168
4169 // Do not let reads of the copy destination float above the copy.
4170 insert_mem_bar(Op_MemBarCPUOrder);
4171
4172 return true;
4173 }
4174
4175 //------------------------clone_coping-----------------------------------
4176 // Helper function for inline_native_clone.
4177 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4178 assert(obj_size != NULL, "");
4179 Node* raw_obj = alloc_obj->in(1);
4180 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4181
4182 AllocateNode* alloc = NULL;
4183 if (ReduceBulkZeroing) {
4184 // We will be completely responsible for initializing this object -
4185 // mark Initialize node as complete.
4186 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4187 // The object was just allocated - there should be no any stores!
4188 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4189 // Mark as complete_with_arraycopy so that on AllocateNode
4190 // expansion, we know this AllocateNode is initialized by an array
4191 // copy and a StoreStore barrier exists after the array copy.
4192 alloc->initialization()->set_complete_with_arraycopy();
4193 }
4194
4195 // Copy the fastest available way.
4196 // TODO: generate fields copies for small objects instead.
4197 Node* src = obj;
4198 Node* dest = alloc_obj;
4199 Node* size = _gvn.transform(obj_size);
4200
4201 // Exclude the header but include array length to copy by 8 bytes words.
4202 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4203 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4204 instanceOopDesc::base_offset_in_bytes();
4205 // base_off:
4206 // 8 - 32-bit VM
4207 // 12 - 64-bit VM, compressed oops
4208 // 16 - 64-bit VM, normal oops
4209 if (base_off % BytesPerLong != 0) {
4210 assert(UseCompressedOops, "");
4211 if (is_array) {
4212 // Exclude length to copy by 8 bytes words.
4213 base_off += sizeof(int);
4214 } else {
4215 // Include klass to copy by 8 bytes words.
4216 base_off = instanceOopDesc::klass_offset_in_bytes();
4217 }
4218 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4219 }
4220 src = basic_plus_adr(src, base_off);
4221 dest = basic_plus_adr(dest, base_off);
4222
4223 // Compute the length also, if needed:
4224 Node* countx = size;
4225 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4226 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4227
4228 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4229 bool disjoint_bases = true;
4230 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4231 src, NULL, dest, NULL, countx,
4232 /*dest_uninitialized*/true);
4233
4234 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4235 if (card_mark) {
4236 assert(!is_array, "");
4237 // Put in store barrier for any and all oops we are sticking
4238 // into this object. (We could avoid this if we could prove
4239 // that the object type contains no oop fields at all.)
4240 Node* no_particular_value = NULL;
4241 Node* no_particular_field = NULL;
4242 int raw_adr_idx = Compile::AliasIdxRaw;
4243 post_barrier(control(),
4244 memory(raw_adr_type),
4245 alloc_obj,
4246 no_particular_field,
4247 raw_adr_idx,
4248 no_particular_value,
4249 T_OBJECT,
4250 false);
4251 }
4252
4253 // Do not let reads from the cloned object float above the arraycopy.
4254 if (alloc != NULL) {
4255 // Do not let stores that initialize this object be reordered with
4256 // a subsequent store that would make this object accessible by
4257 // other threads.
4258 // Record what AllocateNode this StoreStore protects so that
4259 // escape analysis can go from the MemBarStoreStoreNode to the
4260 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4261 // based on the escape status of the AllocateNode.
4262 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4263 } else {
4264 insert_mem_bar(Op_MemBarCPUOrder);
4265 }
4266 }
4267
4268 //------------------------inline_native_clone----------------------------
4269 // protected native Object java.lang.Object.clone();
4270 //
4271 // Here are the simple edge cases:
4272 // null receiver => normal trap
4273 // virtual and clone was overridden => slow path to out-of-line clone
4274 // not cloneable or finalizer => slow path to out-of-line Object.clone
4275 //
4276 // The general case has two steps, allocation and copying.
4277 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4278 //
4279 // Copying also has two cases, oop arrays and everything else.
4280 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4281 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4282 //
4283 // These steps fold up nicely if and when the cloned object's klass
4284 // can be sharply typed as an object array, a type array, or an instance.
4285 //
4286 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4287 PhiNode* result_val;
4288
4289 // Set the reexecute bit for the interpreter to reexecute
4290 // the bytecode that invokes Object.clone if deoptimization happens.
4291 { PreserveReexecuteState preexecs(this);
4292 jvms()->set_should_reexecute(true);
4293
4294 Node* obj = null_check_receiver();
4295 if (stopped()) return true;
4296
4297 Node* obj_klass = load_object_klass(obj);
4298 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4299 const TypeOopPtr* toop = ((tklass != NULL)
4300 ? tklass->as_instance_type()
4301 : TypeInstPtr::NOTNULL);
4302
4303 // Conservatively insert a memory barrier on all memory slices.
4304 // Do not let writes into the original float below the clone.
4305 insert_mem_bar(Op_MemBarCPUOrder);
4306
4307 // paths into result_reg:
4308 enum {
4309 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4310 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4311 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4312 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4313 PATH_LIMIT
4314 };
4315 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4316 result_val = new(C) PhiNode(result_reg,
4317 TypeInstPtr::NOTNULL);
4318 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4319 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4320 TypePtr::BOTTOM);
4321 record_for_igvn(result_reg);
4322
4323 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4324 int raw_adr_idx = Compile::AliasIdxRaw;
4325
4326 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4327 if (array_ctl != NULL) {
4328 // It's an array.
4329 PreserveJVMState pjvms(this);
4330 set_control(array_ctl);
4331 Node* obj_length = load_array_length(obj);
4332 Node* obj_size = NULL;
4333 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4334
4335 if (!use_ReduceInitialCardMarks()) {
4336 // If it is an oop array, it requires very special treatment,
4337 // because card marking is required on each card of the array.
4338 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4339 if (is_obja != NULL) {
4340 PreserveJVMState pjvms2(this);
4341 set_control(is_obja);
4342 // Generate a direct call to the right arraycopy function(s).
4343 bool disjoint_bases = true;
4344 bool length_never_negative = true;
4345 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4346 obj, intcon(0), alloc_obj, intcon(0),
4347 obj_length,
4348 disjoint_bases, length_never_negative);
4349 result_reg->init_req(_objArray_path, control());
4350 result_val->init_req(_objArray_path, alloc_obj);
4351 result_i_o ->set_req(_objArray_path, i_o());
4352 result_mem ->set_req(_objArray_path, reset_memory());
4353 }
4354 }
4355 // Otherwise, there are no card marks to worry about.
4356 // (We can dispense with card marks if we know the allocation
4357 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4358 // causes the non-eden paths to take compensating steps to
4359 // simulate a fresh allocation, so that no further
4360 // card marks are required in compiled code to initialize
4361 // the object.)
4362
4363 if (!stopped()) {
4364 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4365
4366 // Present the results of the copy.
4367 result_reg->init_req(_array_path, control());
4368 result_val->init_req(_array_path, alloc_obj);
4369 result_i_o ->set_req(_array_path, i_o());
4370 result_mem ->set_req(_array_path, reset_memory());
4371 }
4372 }
4373
4374 // We only go to the instance fast case code if we pass a number of guards.
4375 // The paths which do not pass are accumulated in the slow_region.
4376 RegionNode* slow_region = new (C) RegionNode(1);
4377 record_for_igvn(slow_region);
4378 if (!stopped()) {
4379 // It's an instance (we did array above). Make the slow-path tests.
4380 // If this is a virtual call, we generate a funny guard. We grab
4381 // the vtable entry corresponding to clone() from the target object.
4382 // If the target method which we are calling happens to be the
4383 // Object clone() method, we pass the guard. We do not need this
4384 // guard for non-virtual calls; the caller is known to be the native
4385 // Object clone().
4386 if (is_virtual) {
4387 generate_virtual_guard(obj_klass, slow_region);
4388 }
4389
4390 // The object must be cloneable and must not have a finalizer.
4391 // Both of these conditions may be checked in a single test.
4392 // We could optimize the cloneable test further, but we don't care.
4393 generate_access_flags_guard(obj_klass,
4394 // Test both conditions:
4395 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4396 // Must be cloneable but not finalizer:
4397 JVM_ACC_IS_CLONEABLE,
4398 slow_region);
4399 }
4400
4401 if (!stopped()) {
4402 // It's an instance, and it passed the slow-path tests.
4403 PreserveJVMState pjvms(this);
4404 Node* obj_size = NULL;
4405 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4406
4407 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4408
4409 // Present the results of the slow call.
4410 result_reg->init_req(_instance_path, control());
4411 result_val->init_req(_instance_path, alloc_obj);
4412 result_i_o ->set_req(_instance_path, i_o());
4413 result_mem ->set_req(_instance_path, reset_memory());
4414 }
4415
4416 // Generate code for the slow case. We make a call to clone().
4417 set_control(_gvn.transform(slow_region));
4418 if (!stopped()) {
4419 PreserveJVMState pjvms(this);
4420 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4421 Node* slow_result = set_results_for_java_call(slow_call);
4422 // this->control() comes from set_results_for_java_call
4423 result_reg->init_req(_slow_path, control());
4424 result_val->init_req(_slow_path, slow_result);
4425 result_i_o ->set_req(_slow_path, i_o());
4426 result_mem ->set_req(_slow_path, reset_memory());
4427 }
4428
4429 // Return the combined state.
4430 set_control( _gvn.transform(result_reg) );
4431 set_i_o( _gvn.transform(result_i_o) );
4432 set_all_memory( _gvn.transform(result_mem) );
4433 } // original reexecute is set back here
4434
4435 set_result(_gvn.transform(result_val));
4436 return true;
4437 }
4438
4439 //------------------------------basictype2arraycopy----------------------------
4440 address LibraryCallKit::basictype2arraycopy(BasicType t,
4441 Node* src_offset,
4442 Node* dest_offset,
4443 bool disjoint_bases,
4444 const char* &name,
4445 bool dest_uninitialized) {
4446 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4447 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4448
4449 bool aligned = false;
4450 bool disjoint = disjoint_bases;
4451
4452 // if the offsets are the same, we can treat the memory regions as
4453 // disjoint, because either the memory regions are in different arrays,
4454 // or they are identical (which we can treat as disjoint.) We can also
4455 // treat a copy with a destination index less that the source index
4456 // as disjoint since a low->high copy will work correctly in this case.
4457 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4458 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4459 // both indices are constants
4460 int s_offs = src_offset_inttype->get_con();
4461 int d_offs = dest_offset_inttype->get_con();
4462 int element_size = type2aelembytes(t);
4463 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4464 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4465 if (s_offs >= d_offs) disjoint = true;
4466 } else if (src_offset == dest_offset && src_offset != NULL) {
4467 // This can occur if the offsets are identical non-constants.
4468 disjoint = true;
4469 }
4470
4471 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4472 }
4473
4474
4475 //------------------------------inline_arraycopy-----------------------
4476 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4477 // Object dest, int destPos,
4478 // int length);
4479 bool LibraryCallKit::inline_arraycopy() {
4480 // Get the arguments.
4481 Node* src = argument(0); // type: oop
4482 Node* src_offset = argument(1); // type: int
4483 Node* dest = argument(2); // type: oop
4484 Node* dest_offset = argument(3); // type: int
4485 Node* length = argument(4); // type: int
4486
4487 // Compile time checks. If any of these checks cannot be verified at compile time,
4488 // we do not make a fast path for this call. Instead, we let the call remain as it
4489 // is. The checks we choose to mandate at compile time are:
4490 //
4491 // (1) src and dest are arrays.
4492 const Type* src_type = src->Value(&_gvn);
4493 const Type* dest_type = dest->Value(&_gvn);
4494 const TypeAryPtr* top_src = src_type->isa_aryptr();
4495 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4496 if (top_src == NULL || top_src->klass() == NULL ||
4497 top_dest == NULL || top_dest->klass() == NULL) {
4498 // Conservatively insert a memory barrier on all memory slices.
4499 // Do not let writes into the source float below the arraycopy.
4500 insert_mem_bar(Op_MemBarCPUOrder);
4501
4502 // Call StubRoutines::generic_arraycopy stub.
4503 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4504 src, src_offset, dest, dest_offset, length);
4505
4506 // Do not let reads from the destination float above the arraycopy.
4507 // Since we cannot type the arrays, we don't know which slices
4508 // might be affected. We could restrict this barrier only to those
4509 // memory slices which pertain to array elements--but don't bother.
4510 if (!InsertMemBarAfterArraycopy)
4511 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4512 insert_mem_bar(Op_MemBarCPUOrder);
4513 return true;
4514 }
4515
4516 // (2) src and dest arrays must have elements of the same BasicType
4517 // Figure out the size and type of the elements we will be copying.
4518 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4519 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4520 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4521 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4522
4523 if (src_elem != dest_elem || dest_elem == T_VOID) {
4524 // The component types are not the same or are not recognized. Punt.
4525 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4526 generate_slow_arraycopy(TypePtr::BOTTOM,
4527 src, src_offset, dest, dest_offset, length,
4528 /*dest_uninitialized*/false);
4529 return true;
4530 }
4531
4532 //---------------------------------------------------------------------------
4533 // We will make a fast path for this call to arraycopy.
4534
4535 // We have the following tests left to perform:
4536 //
4537 // (3) src and dest must not be null.
4538 // (4) src_offset must not be negative.
4539 // (5) dest_offset must not be negative.
4540 // (6) length must not be negative.
4541 // (7) src_offset + length must not exceed length of src.
4542 // (8) dest_offset + length must not exceed length of dest.
4543 // (9) each element of an oop array must be assignable
4544
4545 RegionNode* slow_region = new (C) RegionNode(1);
4546 record_for_igvn(slow_region);
4547
4548 // (3) operands must not be null
4549 // We currently perform our null checks with the null_check routine.
4550 // This means that the null exceptions will be reported in the caller
4551 // rather than (correctly) reported inside of the native arraycopy call.
4552 // This should be corrected, given time. We do our null check with the
4553 // stack pointer restored.
4554 src = null_check(src, T_ARRAY);
4555 dest = null_check(dest, T_ARRAY);
4556
4557 // (4) src_offset must not be negative.
4558 generate_negative_guard(src_offset, slow_region);
4559
4560 // (5) dest_offset must not be negative.
4561 generate_negative_guard(dest_offset, slow_region);
4562
4563 // (6) length must not be negative (moved to generate_arraycopy()).
4564 // generate_negative_guard(length, slow_region);
4565
4566 // (7) src_offset + length must not exceed length of src.
4567 generate_limit_guard(src_offset, length,
4568 load_array_length(src),
4569 slow_region);
4570
4571 // (8) dest_offset + length must not exceed length of dest.
4572 generate_limit_guard(dest_offset, length,
4573 load_array_length(dest),
4574 slow_region);
4575
4576 // (9) each element of an oop array must be assignable
4577 // The generate_arraycopy subroutine checks this.
4578
4579 // This is where the memory effects are placed:
4580 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4581 generate_arraycopy(adr_type, dest_elem,
4582 src, src_offset, dest, dest_offset, length,
4583 false, false, slow_region);
4584
4585 return true;
4586 }
4587
4588 //-----------------------------generate_arraycopy----------------------
4589 // Generate an optimized call to arraycopy.
4590 // Caller must guard against non-arrays.
4591 // Caller must determine a common array basic-type for both arrays.
4592 // Caller must validate offsets against array bounds.
4593 // The slow_region has already collected guard failure paths
4594 // (such as out of bounds length or non-conformable array types).
4595 // The generated code has this shape, in general:
4596 //
4597 // if (length == 0) return // via zero_path
4598 // slowval = -1
4599 // if (types unknown) {
4600 // slowval = call generic copy loop
4601 // if (slowval == 0) return // via checked_path
4602 // } else if (indexes in bounds) {
4603 // if ((is object array) && !(array type check)) {
4604 // slowval = call checked copy loop
4605 // if (slowval == 0) return // via checked_path
4606 // } else {
4607 // call bulk copy loop
4608 // return // via fast_path
4609 // }
4610 // }
4611 // // adjust params for remaining work:
4612 // if (slowval != -1) {
4613 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4614 // }
4615 // slow_region:
4616 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4617 // return // via slow_call_path
4618 //
4619 // This routine is used from several intrinsics: System.arraycopy,
4620 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4621 //
4622 void
4623 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4624 BasicType basic_elem_type,
4625 Node* src, Node* src_offset,
4626 Node* dest, Node* dest_offset,
4627 Node* copy_length,
4628 bool disjoint_bases,
4629 bool length_never_negative,
4630 RegionNode* slow_region) {
4631
4632 if (slow_region == NULL) {
4633 slow_region = new(C) RegionNode(1);
4634 record_for_igvn(slow_region);
4635 }
4636
4637 Node* original_dest = dest;
4638 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4639 bool dest_uninitialized = false;
4640
4641 // See if this is the initialization of a newly-allocated array.
4642 // If so, we will take responsibility here for initializing it to zero.
4643 // (Note: Because tightly_coupled_allocation performs checks on the
4644 // out-edges of the dest, we need to avoid making derived pointers
4645 // from it until we have checked its uses.)
4646 if (ReduceBulkZeroing
4647 && !ZeroTLAB // pointless if already zeroed
4648 && basic_elem_type != T_CONFLICT // avoid corner case
4649 && !src->eqv_uncast(dest)
4650 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4651 != NULL)
4652 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4653 && alloc->maybe_set_complete(&_gvn)) {
4654 // "You break it, you buy it."
4655 InitializeNode* init = alloc->initialization();
4656 assert(init->is_complete(), "we just did this");
4657 init->set_complete_with_arraycopy();
4658 assert(dest->is_CheckCastPP(), "sanity");
4659 assert(dest->in(0)->in(0) == init, "dest pinned");
4660 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4661 // From this point on, every exit path is responsible for
4662 // initializing any non-copied parts of the object to zero.
4663 // Also, if this flag is set we make sure that arraycopy interacts properly
4664 // with G1, eliding pre-barriers. See CR 6627983.
4665 dest_uninitialized = true;
4666 } else {
4667 // No zeroing elimination here.
4668 alloc = NULL;
4669 //original_dest = dest;
4670 //dest_uninitialized = false;
4671 }
4672
4673 // Results are placed here:
4674 enum { fast_path = 1, // normal void-returning assembly stub
4675 checked_path = 2, // special assembly stub with cleanup
4676 slow_call_path = 3, // something went wrong; call the VM
4677 zero_path = 4, // bypass when length of copy is zero
4678 bcopy_path = 5, // copy primitive array by 64-bit blocks
4679 PATH_LIMIT = 6
4680 };
4681 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4682 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4683 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4684 record_for_igvn(result_region);
4685 _gvn.set_type_bottom(result_i_o);
4686 _gvn.set_type_bottom(result_memory);
4687 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4688
4689 // The slow_control path:
4690 Node* slow_control;
4691 Node* slow_i_o = i_o();
4692 Node* slow_mem = memory(adr_type);
4693 debug_only(slow_control = (Node*) badAddress);
4694
4695 // Checked control path:
4696 Node* checked_control = top();
4697 Node* checked_mem = NULL;
4698 Node* checked_i_o = NULL;
4699 Node* checked_value = NULL;
4700
4701 if (basic_elem_type == T_CONFLICT) {
4702 assert(!dest_uninitialized, "");
4703 Node* cv = generate_generic_arraycopy(adr_type,
4704 src, src_offset, dest, dest_offset,
4705 copy_length, dest_uninitialized);
4706 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4707 checked_control = control();
4708 checked_i_o = i_o();
4709 checked_mem = memory(adr_type);
4710 checked_value = cv;
4711 set_control(top()); // no fast path
4712 }
4713
4714 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4715 if (not_pos != NULL) {
4716 PreserveJVMState pjvms(this);
4717 set_control(not_pos);
4718
4719 // (6) length must not be negative.
4720 if (!length_never_negative) {
4721 generate_negative_guard(copy_length, slow_region);
4722 }
4723
4724 // copy_length is 0.
4725 if (!stopped() && dest_uninitialized) {
4726 Node* dest_length = alloc->in(AllocateNode::ALength);
4727 if (copy_length->eqv_uncast(dest_length)
4728 || _gvn.find_int_con(dest_length, 1) <= 0) {
4729 // There is no zeroing to do. No need for a secondary raw memory barrier.
4730 } else {
4731 // Clear the whole thing since there are no source elements to copy.
4732 generate_clear_array(adr_type, dest, basic_elem_type,
4733 intcon(0), NULL,
4734 alloc->in(AllocateNode::AllocSize));
4735 // Use a secondary InitializeNode as raw memory barrier.
4736 // Currently it is needed only on this path since other
4737 // paths have stub or runtime calls as raw memory barriers.
4738 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4739 Compile::AliasIdxRaw,
4740 top())->as_Initialize();
4741 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4742 }
4743 }
4744
4745 // Present the results of the fast call.
4746 result_region->init_req(zero_path, control());
4747 result_i_o ->init_req(zero_path, i_o());
4748 result_memory->init_req(zero_path, memory(adr_type));
4749 }
4750
4751 if (!stopped() && dest_uninitialized) {
4752 // We have to initialize the *uncopied* part of the array to zero.
4753 // The copy destination is the slice dest[off..off+len]. The other slices
4754 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4755 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4756 Node* dest_length = alloc->in(AllocateNode::ALength);
4757 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4758 copy_length) );
4759
4760 // If there is a head section that needs zeroing, do it now.
4761 if (find_int_con(dest_offset, -1) != 0) {
4762 generate_clear_array(adr_type, dest, basic_elem_type,
4763 intcon(0), dest_offset,
4764 NULL);
4765 }
4766
4767 // Next, perform a dynamic check on the tail length.
4768 // It is often zero, and we can win big if we prove this.
4769 // There are two wins: Avoid generating the ClearArray
4770 // with its attendant messy index arithmetic, and upgrade
4771 // the copy to a more hardware-friendly word size of 64 bits.
4772 Node* tail_ctl = NULL;
4773 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4774 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4775 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4776 tail_ctl = generate_slow_guard(bol_lt, NULL);
4777 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4778 }
4779
4780 // At this point, let's assume there is no tail.
4781 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4782 // There is no tail. Try an upgrade to a 64-bit copy.
4783 bool didit = false;
4784 { PreserveJVMState pjvms(this);
4785 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4786 src, src_offset, dest, dest_offset,
4787 dest_size, dest_uninitialized);
4788 if (didit) {
4789 // Present the results of the block-copying fast call.
4790 result_region->init_req(bcopy_path, control());
4791 result_i_o ->init_req(bcopy_path, i_o());
4792 result_memory->init_req(bcopy_path, memory(adr_type));
4793 }
4794 }
4795 if (didit)
4796 set_control(top()); // no regular fast path
4797 }
4798
4799 // Clear the tail, if any.
4800 if (tail_ctl != NULL) {
4801 Node* notail_ctl = stopped() ? NULL : control();
4802 set_control(tail_ctl);
4803 if (notail_ctl == NULL) {
4804 generate_clear_array(adr_type, dest, basic_elem_type,
4805 dest_tail, NULL,
4806 dest_size);
4807 } else {
4808 // Make a local merge.
4809 Node* done_ctl = new(C) RegionNode(3);
4810 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4811 done_ctl->init_req(1, notail_ctl);
4812 done_mem->init_req(1, memory(adr_type));
4813 generate_clear_array(adr_type, dest, basic_elem_type,
4814 dest_tail, NULL,
4815 dest_size);
4816 done_ctl->init_req(2, control());
4817 done_mem->init_req(2, memory(adr_type));
4818 set_control( _gvn.transform(done_ctl) );
4819 set_memory( _gvn.transform(done_mem), adr_type );
4820 }
4821 }
4822 }
4823
4824 BasicType copy_type = basic_elem_type;
4825 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4826 if (!stopped() && copy_type == T_OBJECT) {
4827 // If src and dest have compatible element types, we can copy bits.
4828 // Types S[] and D[] are compatible if D is a supertype of S.
4829 //
4830 // If they are not, we will use checked_oop_disjoint_arraycopy,
4831 // which performs a fast optimistic per-oop check, and backs off
4832 // further to JVM_ArrayCopy on the first per-oop check that fails.
4833 // (Actually, we don't move raw bits only; the GC requires card marks.)
4834
4835 // Get the klassOop for both src and dest
4836 Node* src_klass = load_object_klass(src);
4837 Node* dest_klass = load_object_klass(dest);
4838
4839 // Generate the subtype check.
4840 // This might fold up statically, or then again it might not.
4841 //
4842 // Non-static example: Copying List<String>.elements to a new String[].
4843 // The backing store for a List<String> is always an Object[],
4844 // but its elements are always type String, if the generic types
4845 // are correct at the source level.
4846 //
4847 // Test S[] against D[], not S against D, because (probably)
4848 // the secondary supertype cache is less busy for S[] than S.
4849 // This usually only matters when D is an interface.
4850 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4851 // Plug failing path into checked_oop_disjoint_arraycopy
4852 if (not_subtype_ctrl != top()) {
4853 PreserveJVMState pjvms(this);
4854 set_control(not_subtype_ctrl);
4855 // (At this point we can assume disjoint_bases, since types differ.)
4856 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4857 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4858 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4859 Node* dest_elem_klass = _gvn.transform(n1);
4860 Node* cv = generate_checkcast_arraycopy(adr_type,
4861 dest_elem_klass,
4862 src, src_offset, dest, dest_offset,
4863 ConvI2X(copy_length), dest_uninitialized);
4864 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4865 checked_control = control();
4866 checked_i_o = i_o();
4867 checked_mem = memory(adr_type);
4868 checked_value = cv;
4869 }
4870 // At this point we know we do not need type checks on oop stores.
4871
4872 // Let's see if we need card marks:
4873 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4874 // If we do not need card marks, copy using the jint or jlong stub.
4875 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4876 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4877 "sizes agree");
4878 }
4879 }
4880
4881 if (!stopped()) {
4882 // Generate the fast path, if possible.
4883 PreserveJVMState pjvms(this);
4884 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4885 src, src_offset, dest, dest_offset,
4886 ConvI2X(copy_length), dest_uninitialized);
4887
4888 // Present the results of the fast call.
4889 result_region->init_req(fast_path, control());
4890 result_i_o ->init_req(fast_path, i_o());
4891 result_memory->init_req(fast_path, memory(adr_type));
4892 }
4893
4894 // Here are all the slow paths up to this point, in one bundle:
4895 slow_control = top();
4896 if (slow_region != NULL)
4897 slow_control = _gvn.transform(slow_region);
4898 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4899
4900 set_control(checked_control);
4901 if (!stopped()) {
4902 // Clean up after the checked call.
4903 // The returned value is either 0 or -1^K,
4904 // where K = number of partially transferred array elements.
4905 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4906 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4907 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4908
4909 // If it is 0, we are done, so transfer to the end.
4910 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4911 result_region->init_req(checked_path, checks_done);
4912 result_i_o ->init_req(checked_path, checked_i_o);
4913 result_memory->init_req(checked_path, checked_mem);
4914
4915 // If it is not zero, merge into the slow call.
4916 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4917 RegionNode* slow_reg2 = new(C) RegionNode(3);
4918 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4919 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4920 record_for_igvn(slow_reg2);
4921 slow_reg2 ->init_req(1, slow_control);
4922 slow_i_o2 ->init_req(1, slow_i_o);
4923 slow_mem2 ->init_req(1, slow_mem);
4924 slow_reg2 ->init_req(2, control());
4925 slow_i_o2 ->init_req(2, checked_i_o);
4926 slow_mem2 ->init_req(2, checked_mem);
4927
4928 slow_control = _gvn.transform(slow_reg2);
4929 slow_i_o = _gvn.transform(slow_i_o2);
4930 slow_mem = _gvn.transform(slow_mem2);
4931
4932 if (alloc != NULL) {
4933 // We'll restart from the very beginning, after zeroing the whole thing.
4934 // This can cause double writes, but that's OK since dest is brand new.
4935 // So we ignore the low 31 bits of the value returned from the stub.
4936 } else {
4937 // We must continue the copy exactly where it failed, or else
4938 // another thread might see the wrong number of writes to dest.
4939 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4940 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4941 slow_offset->init_req(1, intcon(0));
4942 slow_offset->init_req(2, checked_offset);
4943 slow_offset = _gvn.transform(slow_offset);
4944
4945 // Adjust the arguments by the conditionally incoming offset.
4946 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4947 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4948 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4949
4950 // Tweak the node variables to adjust the code produced below:
4951 src_offset = src_off_plus;
4952 dest_offset = dest_off_plus;
4953 copy_length = length_minus;
4954 }
4955 }
4956
4957 set_control(slow_control);
4958 if (!stopped()) {
4959 // Generate the slow path, if needed.
4960 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4961
4962 set_memory(slow_mem, adr_type);
4963 set_i_o(slow_i_o);
4964
4965 if (dest_uninitialized) {
4966 generate_clear_array(adr_type, dest, basic_elem_type,
4967 intcon(0), NULL,
4968 alloc->in(AllocateNode::AllocSize));
4969 }
4970
4971 generate_slow_arraycopy(adr_type,
4972 src, src_offset, dest, dest_offset,
4973 copy_length, /*dest_uninitialized*/false);
4974
4975 result_region->init_req(slow_call_path, control());
4976 result_i_o ->init_req(slow_call_path, i_o());
4977 result_memory->init_req(slow_call_path, memory(adr_type));
4978 }
4979
4980 // Remove unused edges.
4981 for (uint i = 1; i < result_region->req(); i++) {
4982 if (result_region->in(i) == NULL)
4983 result_region->init_req(i, top());
4984 }
4985
4986 // Finished; return the combined state.
4987 set_control( _gvn.transform(result_region) );
4988 set_i_o( _gvn.transform(result_i_o) );
4989 set_memory( _gvn.transform(result_memory), adr_type );
4990
4991 // The memory edges above are precise in order to model effects around
4992 // array copies accurately to allow value numbering of field loads around
4993 // arraycopy. Such field loads, both before and after, are common in Java
4994 // collections and similar classes involving header/array data structures.
4995 //
4996 // But with low number of register or when some registers are used or killed
4997 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4998 // The next memory barrier is added to avoid it. If the arraycopy can be
4999 // optimized away (which it can, sometimes) then we can manually remove
5000 // the membar also.
5001 //
5002 // Do not let reads from the cloned object float above the arraycopy.
5003 if (alloc != NULL) {
5004 // Do not let stores that initialize this object be reordered with
5005 // a subsequent store that would make this object accessible by
5006 // other threads.
5007 // Record what AllocateNode this StoreStore protects so that
5008 // escape analysis can go from the MemBarStoreStoreNode to the
5009 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5010 // based on the escape status of the AllocateNode.
5011 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5012 } else if (InsertMemBarAfterArraycopy)
5013 insert_mem_bar(Op_MemBarCPUOrder);
5014 }
5015
5016
5017 // Helper function which determines if an arraycopy immediately follows
5018 // an allocation, with no intervening tests or other escapes for the object.
5019 AllocateArrayNode*
5020 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5021 RegionNode* slow_region) {
5022 if (stopped()) return NULL; // no fast path
5023 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5024
5025 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5026 if (alloc == NULL) return NULL;
5027
5028 Node* rawmem = memory(Compile::AliasIdxRaw);
5029 // Is the allocation's memory state untouched?
5030 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5031 // Bail out if there have been raw-memory effects since the allocation.
5032 // (Example: There might have been a call or safepoint.)
5033 return NULL;
5034 }
5035 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5036 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5037 return NULL;
5038 }
5039
5040 // There must be no unexpected observers of this allocation.
5041 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5042 Node* obs = ptr->fast_out(i);
5043 if (obs != this->map()) {
5044 return NULL;
5045 }
5046 }
5047
5048 // This arraycopy must unconditionally follow the allocation of the ptr.
5049 Node* alloc_ctl = ptr->in(0);
5050 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5051
5052 Node* ctl = control();
5053 while (ctl != alloc_ctl) {
5054 // There may be guards which feed into the slow_region.
5055 // Any other control flow means that we might not get a chance
5056 // to finish initializing the allocated object.
5057 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5058 IfNode* iff = ctl->in(0)->as_If();
5059 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5060 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5061 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5062 ctl = iff->in(0); // This test feeds the known slow_region.
5063 continue;
5064 }
5065 // One more try: Various low-level checks bottom out in
5066 // uncommon traps. If the debug-info of the trap omits
5067 // any reference to the allocation, as we've already
5068 // observed, then there can be no objection to the trap.
5069 bool found_trap = false;
5070 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5071 Node* obs = not_ctl->fast_out(j);
5072 if (obs->in(0) == not_ctl && obs->is_Call() &&
5073 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5074 found_trap = true; break;
5075 }
5076 }
5077 if (found_trap) {
5078 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5079 continue;
5080 }
5081 }
5082 return NULL;
5083 }
5084
5085 // If we get this far, we have an allocation which immediately
5086 // precedes the arraycopy, and we can take over zeroing the new object.
5087 // The arraycopy will finish the initialization, and provide
5088 // a new control state to which we will anchor the destination pointer.
5089
5090 return alloc;
5091 }
5092
5093 // Helper for initialization of arrays, creating a ClearArray.
5094 // It writes zero bits in [start..end), within the body of an array object.
5095 // The memory effects are all chained onto the 'adr_type' alias category.
5096 //
5097 // Since the object is otherwise uninitialized, we are free
5098 // to put a little "slop" around the edges of the cleared area,
5099 // as long as it does not go back into the array's header,
5100 // or beyond the array end within the heap.
5101 //
5102 // The lower edge can be rounded down to the nearest jint and the
5103 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5104 //
5105 // Arguments:
5106 // adr_type memory slice where writes are generated
5107 // dest oop of the destination array
5108 // basic_elem_type element type of the destination
5109 // slice_idx array index of first element to store
5110 // slice_len number of elements to store (or NULL)
5111 // dest_size total size in bytes of the array object
5112 //
5113 // Exactly one of slice_len or dest_size must be non-NULL.
5114 // If dest_size is non-NULL, zeroing extends to the end of the object.
5115 // If slice_len is non-NULL, the slice_idx value must be a constant.
5116 void
5117 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5118 Node* dest,
5119 BasicType basic_elem_type,
5120 Node* slice_idx,
5121 Node* slice_len,
5122 Node* dest_size) {
5123 // one or the other but not both of slice_len and dest_size:
5124 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5125 if (slice_len == NULL) slice_len = top();
5126 if (dest_size == NULL) dest_size = top();
5127
5128 // operate on this memory slice:
5129 Node* mem = memory(adr_type); // memory slice to operate on
5130
5131 // scaling and rounding of indexes:
5132 int scale = exact_log2(type2aelembytes(basic_elem_type));
5133 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5134 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5135 int bump_bit = (-1 << scale) & BytesPerInt;
5136
5137 // determine constant starts and ends
5138 const intptr_t BIG_NEG = -128;
5139 assert(BIG_NEG + 2*abase < 0, "neg enough");
5140 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5141 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5142 if (slice_len_con == 0) {
5143 return; // nothing to do here
5144 }
5145 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5146 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5147 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5148 assert(end_con < 0, "not two cons");
5149 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5150 BytesPerLong);
5151 }
5152
5153 if (start_con >= 0 && end_con >= 0) {
5154 // Constant start and end. Simple.
5155 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5156 start_con, end_con, &_gvn);
5157 } else if (start_con >= 0 && dest_size != top()) {
5158 // Constant start, pre-rounded end after the tail of the array.
5159 Node* end = dest_size;
5160 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5161 start_con, end, &_gvn);
5162 } else if (start_con >= 0 && slice_len != top()) {
5163 // Constant start, non-constant end. End needs rounding up.
5164 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5165 intptr_t end_base = abase + (slice_idx_con << scale);
5166 int end_round = (-1 << scale) & (BytesPerLong - 1);
5167 Node* end = ConvI2X(slice_len);
5168 if (scale != 0)
5169 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5170 end_base += end_round;
5171 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5172 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5173 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5174 start_con, end, &_gvn);
5175 } else if (start_con < 0 && dest_size != top()) {
5176 // Non-constant start, pre-rounded end after the tail of the array.
5177 // This is almost certainly a "round-to-end" operation.
5178 Node* start = slice_idx;
5179 start = ConvI2X(start);
5180 if (scale != 0)
5181 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5182 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5183 if ((bump_bit | clear_low) != 0) {
5184 int to_clear = (bump_bit | clear_low);
5185 // Align up mod 8, then store a jint zero unconditionally
5186 // just before the mod-8 boundary.
5187 if (((abase + bump_bit) & ~to_clear) - bump_bit
5188 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5189 bump_bit = 0;
5190 assert((abase & to_clear) == 0, "array base must be long-aligned");
5191 } else {
5192 // Bump 'start' up to (or past) the next jint boundary:
5193 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5194 assert((abase & clear_low) == 0, "array base must be int-aligned");
5195 }
5196 // Round bumped 'start' down to jlong boundary in body of array.
5197 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5198 if (bump_bit != 0) {
5199 // Store a zero to the immediately preceding jint:
5200 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5201 Node* p1 = basic_plus_adr(dest, x1);
5202 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, StoreNode::unordered);
5203 mem = _gvn.transform(mem);
5204 }
5205 }
5206 Node* end = dest_size; // pre-rounded
5207 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5208 start, end, &_gvn);
5209 } else {
5210 // Non-constant start, unrounded non-constant end.
5211 // (Nobody zeroes a random midsection of an array using this routine.)
5212 ShouldNotReachHere(); // fix caller
5213 }
5214
5215 // Done.
5216 set_memory(mem, adr_type);
5217 }
5218
5219
5220 bool
5221 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5222 BasicType basic_elem_type,
5223 AllocateNode* alloc,
5224 Node* src, Node* src_offset,
5225 Node* dest, Node* dest_offset,
5226 Node* dest_size, bool dest_uninitialized) {
5227 // See if there is an advantage from block transfer.
5228 int scale = exact_log2(type2aelembytes(basic_elem_type));
5229 if (scale >= LogBytesPerLong)
5230 return false; // it is already a block transfer
5231
5232 // Look at the alignment of the starting offsets.
5233 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5234
5235 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5236 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5237 if (src_off_con < 0 || dest_off_con < 0)
5238 // At present, we can only understand constants.
5239 return false;
5240
5241 intptr_t src_off = abase + (src_off_con << scale);
5242 intptr_t dest_off = abase + (dest_off_con << scale);
5243
5244 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5245 // Non-aligned; too bad.
5246 // One more chance: Pick off an initial 32-bit word.
5247 // This is a common case, since abase can be odd mod 8.
5248 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5249 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5250 Node* sptr = basic_plus_adr(src, src_off);
5251 Node* dptr = basic_plus_adr(dest, dest_off);
5252 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5253 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5254 src_off += BytesPerInt;
5255 dest_off += BytesPerInt;
5256 } else {
5257 return false;
5258 }
5259 }
5260 assert(src_off % BytesPerLong == 0, "");
5261 assert(dest_off % BytesPerLong == 0, "");
5262
5263 // Do this copy by giant steps.
5264 Node* sptr = basic_plus_adr(src, src_off);
5265 Node* dptr = basic_plus_adr(dest, dest_off);
5266 Node* countx = dest_size;
5267 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5268 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5269
5270 bool disjoint_bases = true; // since alloc != NULL
5271 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5272 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5273
5274 return true;
5275 }
5276
5277
5278 // Helper function; generates code for the slow case.
5279 // We make a call to a runtime method which emulates the native method,
5280 // but without the native wrapper overhead.
5281 void
5282 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5283 Node* src, Node* src_offset,
5284 Node* dest, Node* dest_offset,
5285 Node* copy_length, bool dest_uninitialized) {
5286 assert(!dest_uninitialized, "Invariant");
5287 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5288 OptoRuntime::slow_arraycopy_Type(),
5289 OptoRuntime::slow_arraycopy_Java(),
5290 "slow_arraycopy", adr_type,
5291 src, src_offset, dest, dest_offset,
5292 copy_length);
5293
5294 // Handle exceptions thrown by this fellow:
5295 make_slow_call_ex(call, env()->Throwable_klass(), false);
5296 }
5297
5298 // Helper function; generates code for cases requiring runtime checks.
5299 Node*
5300 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5301 Node* dest_elem_klass,
5302 Node* src, Node* src_offset,
5303 Node* dest, Node* dest_offset,
5304 Node* copy_length, bool dest_uninitialized) {
5305 if (stopped()) return NULL;
5306
5307 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5308 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5309 return NULL;
5310 }
5311
5312 // Pick out the parameters required to perform a store-check
5313 // for the target array. This is an optimistic check. It will
5314 // look in each non-null element's class, at the desired klass's
5315 // super_check_offset, for the desired klass.
5316 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5317 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5318 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5319 Node* check_offset = ConvI2X(_gvn.transform(n3));
5320 Node* check_value = dest_elem_klass;
5321
5322 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5323 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5324
5325 // (We know the arrays are never conjoint, because their types differ.)
5326 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5327 OptoRuntime::checkcast_arraycopy_Type(),
5328 copyfunc_addr, "checkcast_arraycopy", adr_type,
5329 // five arguments, of which two are
5330 // intptr_t (jlong in LP64)
5331 src_start, dest_start,
5332 copy_length XTOP,
5333 check_offset XTOP,
5334 check_value);
5335
5336 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5337 }
5338
5339
5340 // Helper function; generates code for cases requiring runtime checks.
5341 Node*
5342 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5343 Node* src, Node* src_offset,
5344 Node* dest, Node* dest_offset,
5345 Node* copy_length, bool dest_uninitialized) {
5346 assert(!dest_uninitialized, "Invariant");
5347 if (stopped()) return NULL;
5348 address copyfunc_addr = StubRoutines::generic_arraycopy();
5349 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5350 return NULL;
5351 }
5352
5353 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5354 OptoRuntime::generic_arraycopy_Type(),
5355 copyfunc_addr, "generic_arraycopy", adr_type,
5356 src, src_offset, dest, dest_offset, copy_length);
5357
5358 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5359 }
5360
5361 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5362 void
5363 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5364 BasicType basic_elem_type,
5365 bool disjoint_bases,
5366 Node* src, Node* src_offset,
5367 Node* dest, Node* dest_offset,
5368 Node* copy_length, bool dest_uninitialized) {
5369 if (stopped()) return; // nothing to do
5370
5371 Node* src_start = src;
5372 Node* dest_start = dest;
5373 if (src_offset != NULL || dest_offset != NULL) {
5374 assert(src_offset != NULL && dest_offset != NULL, "");
5375 src_start = array_element_address(src, src_offset, basic_elem_type);
5376 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5377 }
5378
5379 // Figure out which arraycopy runtime method to call.
5380 const char* copyfunc_name = "arraycopy";
5381 address copyfunc_addr =
5382 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5383 disjoint_bases, copyfunc_name, dest_uninitialized);
5384
5385 // Call it. Note that the count_ix value is not scaled to a byte-size.
5386 make_runtime_call(RC_LEAF|RC_NO_FP,
5387 OptoRuntime::fast_arraycopy_Type(),
5388 copyfunc_addr, copyfunc_name, adr_type,
5389 src_start, dest_start, copy_length XTOP);
5390 }
5391
5392 //----------------------------inline_reference_get----------------------------
5393 // public T java.lang.ref.Reference.get();
5394 bool LibraryCallKit::inline_reference_get() {
5395 const int referent_offset = java_lang_ref_Reference::referent_offset;
5396 guarantee(referent_offset > 0, "should have already been set");
5397
5398 // Get the argument:
5399 Node* reference_obj = null_check_receiver();
5400 if (stopped()) return true;
5401
5402 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5403
5404 ciInstanceKlass* klass = env()->Object_klass();
5405 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5406
5407 Node* no_ctrl = NULL;
5408 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5409
5410 // Use the pre-barrier to record the value in the referent field
5411 pre_barrier(false /* do_load */,
5412 control(),
5413 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5414 result /* pre_val */,
5415 T_OBJECT);
5416
5417 // Add memory barrier to prevent commoning reads from this field
5418 // across safepoint since GC can change its value.
5419 insert_mem_bar(Op_MemBarCPUOrder);
5420
5421 set_result(result);
5422 return true;
5423 }
5424
5425
5426 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5427 bool is_exact=true, bool is_static=false) {
5428
5429 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5430 assert(tinst != NULL, "obj is null");
5431 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5432 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5433
5434 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5435 ciSymbol::make(fieldTypeString),
5436 is_static);
5437 if (field == NULL) return (Node *) NULL;
5438 assert (field != NULL, "undefined field");
5439
5440 // Next code copied from Parse::do_get_xxx():
5441
5442 // Compute address and memory type.
5443 int offset = field->offset_in_bytes();
5444 bool is_vol = field->is_volatile();
5445 ciType* field_klass = field->type();
5446 assert(field_klass->is_loaded(), "should be loaded");
5447 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5448 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5449 BasicType bt = field->layout_type();
5450
5451 // Build the resultant type of the load
5452 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5453
5454 // Build the load.
5455 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5456 return loadedField;
5457 }
5458
5459
5460 //------------------------------inline_aescrypt_Block-----------------------
5461 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5462 address stubAddr;
5463 const char *stubName;
5464 assert(UseAES, "need AES instruction support");
5465
5466 switch(id) {
5467 case vmIntrinsics::_aescrypt_encryptBlock:
5468 stubAddr = StubRoutines::aescrypt_encryptBlock();
5469 stubName = "aescrypt_encryptBlock";
5470 break;
5471 case vmIntrinsics::_aescrypt_decryptBlock:
5472 stubAddr = StubRoutines::aescrypt_decryptBlock();
5473 stubName = "aescrypt_decryptBlock";
5474 break;
5475 }
5476 if (stubAddr == NULL) return false;
5477
5478 Node* aescrypt_object = argument(0);
5479 Node* src = argument(1);
5480 Node* src_offset = argument(2);
5481 Node* dest = argument(3);
5482 Node* dest_offset = argument(4);
5483
5484 // (1) src and dest are arrays.
5485 const Type* src_type = src->Value(&_gvn);
5486 const Type* dest_type = dest->Value(&_gvn);
5487 const TypeAryPtr* top_src = src_type->isa_aryptr();
5488 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5489 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5490
5491 // for the quick and dirty code we will skip all the checks.
5492 // we are just trying to get the call to be generated.
5493 Node* src_start = src;
5494 Node* dest_start = dest;
5495 if (src_offset != NULL || dest_offset != NULL) {
5496 assert(src_offset != NULL && dest_offset != NULL, "");
5497 src_start = array_element_address(src, src_offset, T_BYTE);
5498 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5499 }
5500
5501 // now need to get the start of its expanded key array
5502 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5503 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5504 if (k_start == NULL) return false;
5505
5506 // Call the stub.
5507 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5508 stubAddr, stubName, TypePtr::BOTTOM,
5509 src_start, dest_start, k_start);
5510
5511 return true;
5512 }
5513
5514 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5515 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5516 address stubAddr;
5517 const char *stubName;
5518
5519 assert(UseAES, "need AES instruction support");
5520
5521 switch(id) {
5522 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5523 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5524 stubName = "cipherBlockChaining_encryptAESCrypt";
5525 break;
5526 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5527 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5528 stubName = "cipherBlockChaining_decryptAESCrypt";
5529 break;
5530 }
5531 if (stubAddr == NULL) return false;
5532
5533 Node* cipherBlockChaining_object = argument(0);
5534 Node* src = argument(1);
5535 Node* src_offset = argument(2);
5536 Node* len = argument(3);
5537 Node* dest = argument(4);
5538 Node* dest_offset = argument(5);
5539
5540 // (1) src and dest are arrays.
5541 const Type* src_type = src->Value(&_gvn);
5542 const Type* dest_type = dest->Value(&_gvn);
5543 const TypeAryPtr* top_src = src_type->isa_aryptr();
5544 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5545 assert (top_src != NULL && top_src->klass() != NULL
5546 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5547
5548 // checks are the responsibility of the caller
5549 Node* src_start = src;
5550 Node* dest_start = dest;
5551 if (src_offset != NULL || dest_offset != NULL) {
5552 assert(src_offset != NULL && dest_offset != NULL, "");
5553 src_start = array_element_address(src, src_offset, T_BYTE);
5554 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5555 }
5556
5557 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5558 // (because of the predicated logic executed earlier).
5559 // so we cast it here safely.
5560 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5561
5562 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5563 if (embeddedCipherObj == NULL) return false;
5564
5565 // cast it to what we know it will be at runtime
5566 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5567 assert(tinst != NULL, "CBC obj is null");
5568 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5569 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5570 if (!klass_AESCrypt->is_loaded()) return false;
5571
5572 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5573 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5574 const TypeOopPtr* xtype = aklass->as_instance_type();
5575 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5576 aescrypt_object = _gvn.transform(aescrypt_object);
5577
5578 // we need to get the start of the aescrypt_object's expanded key array
5579 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5580 if (k_start == NULL) return false;
5581
5582 // similarly, get the start address of the r vector
5583 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5584 if (objRvec == NULL) return false;
5585 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5586
5587 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5588 make_runtime_call(RC_LEAF|RC_NO_FP,
5589 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5590 stubAddr, stubName, TypePtr::BOTTOM,
5591 src_start, dest_start, k_start, r_start, len);
5592
5593 // return is void so no result needs to be pushed
5594
5595 return true;
5596 }
5597
5598 //------------------------------get_key_start_from_aescrypt_object-----------------------
5599 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5600 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5601 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5602 if (objAESCryptKey == NULL) return (Node *) NULL;
5603
5604 // now have the array, need to get the start address of the K array
5605 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5606 return k_start;
5607 }
5608
5609 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5610 // Return node representing slow path of predicate check.
5611 // the pseudo code we want to emulate with this predicate is:
5612 // for encryption:
5613 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5614 // for decryption:
5615 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5616 // note cipher==plain is more conservative than the original java code but that's OK
5617 //
5618 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5619 // First, check receiver for NULL since it is virtual method.
5620 Node* objCBC = argument(0);
5621 objCBC = null_check(objCBC);
5622
5623 if (stopped()) return NULL; // Always NULL
5624
5625 // Load embeddedCipher field of CipherBlockChaining object.
5626 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5627
5628 // get AESCrypt klass for instanceOf check
5629 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5630 // will have same classloader as CipherBlockChaining object
5631 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5632 assert(tinst != NULL, "CBCobj is null");
5633 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5634
5635 // we want to do an instanceof comparison against the AESCrypt class
5636 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5637 if (!klass_AESCrypt->is_loaded()) {
5638 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5639 Node* ctrl = control();
5640 set_control(top()); // no regular fast path
5641 return ctrl;
5642 }
5643 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5644
5645 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5646 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5647 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5648
5649 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5650
5651 // for encryption, we are done
5652 if (!decrypting)
5653 return instof_false; // even if it is NULL
5654
5655 // for decryption, we need to add a further check to avoid
5656 // taking the intrinsic path when cipher and plain are the same
5657 // see the original java code for why.
5658 RegionNode* region = new(C) RegionNode(3);
5659 region->init_req(1, instof_false);
5660 Node* src = argument(1);
5661 Node* dest = argument(4);
5662 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5663 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5664 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5665 region->init_req(2, src_dest_conjoint);
5666
5667 record_for_igvn(region);
5668 return _gvn.transform(region);
5669 }