1 /*
2 * Copyright (c) 1999, 2013, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/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 bool _does_virtual_dispatch;
50 vmIntrinsics::ID _intrinsic_id;
51
52 public:
53 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
54 : InlineCallGenerator(m),
55 _is_virtual(is_virtual),
56 _is_predicted(is_predicted),
57 _does_virtual_dispatch(does_virtual_dispatch),
58 _intrinsic_id(id)
59 {
60 }
61 virtual bool is_intrinsic() const { return true; }
62 virtual bool is_virtual() const { return _is_virtual; }
63 virtual bool is_predicted() const { return _is_predicted; }
64 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
65 virtual JVMState* generate(JVMState* jvms);
66 virtual Node* generate_predicate(JVMState* jvms);
67 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
68 };
69
70
71 // Local helper class for LibraryIntrinsic:
72 class LibraryCallKit : public GraphKit {
73 private:
74 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
75 Node* _result; // the result node, if any
76 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
77
78 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
79
80 public:
81 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
82 : GraphKit(jvms),
83 _intrinsic(intrinsic),
84 _result(NULL)
85 {
86 // Check if this is a root compile. In that case we don't have a caller.
87 if (!jvms->has_method()) {
88 _reexecute_sp = sp();
89 } else {
90 // Find out how many arguments the interpreter needs when deoptimizing
91 // and save the stack pointer value so it can used by uncommon_trap.
92 // We find the argument count by looking at the declared signature.
93 bool ignored_will_link;
94 ciSignature* declared_signature = NULL;
95 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
96 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
97 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
98 }
99 }
100
101 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
102
103 ciMethod* caller() const { return jvms()->method(); }
104 int bci() const { return jvms()->bci(); }
105 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
106 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
107 ciMethod* callee() const { return _intrinsic->method(); }
108
109 bool try_to_inline();
110 Node* try_to_predicate();
111
112 void push_result() {
113 // Push the result onto the stack.
114 if (!stopped() && result() != NULL) {
115 BasicType bt = result()->bottom_type()->basic_type();
116 push_node(bt, result());
117 }
118 }
119
120 private:
121 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
122 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
123 }
124
125 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
126 void set_result(RegionNode* region, PhiNode* value);
127 Node* result() { return _result; }
128
129 virtual int reexecute_sp() { return _reexecute_sp; }
130
131 // Helper functions to inline natives
132 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
133 Node* generate_slow_guard(Node* test, RegionNode* region);
134 Node* generate_fair_guard(Node* test, RegionNode* region);
135 Node* generate_negative_guard(Node* index, RegionNode* region,
136 // resulting CastII of index:
137 Node* *pos_index = NULL);
138 Node* generate_nonpositive_guard(Node* index, bool never_negative,
139 // resulting CastII of index:
140 Node* *pos_index = NULL);
141 Node* generate_limit_guard(Node* offset, Node* subseq_length,
142 Node* array_length,
143 RegionNode* region);
144 Node* generate_current_thread(Node* &tls_output);
145 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
146 bool disjoint_bases, const char* &name, bool dest_uninitialized);
147 Node* load_mirror_from_klass(Node* klass);
148 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
149 RegionNode* region, int null_path,
150 int offset);
151 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
152 RegionNode* region, int null_path) {
153 int offset = java_lang_Class::klass_offset_in_bytes();
154 return load_klass_from_mirror_common(mirror, never_see_null,
155 region, null_path,
156 offset);
157 }
158 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
159 RegionNode* region, int null_path) {
160 int offset = java_lang_Class::array_klass_offset_in_bytes();
161 return load_klass_from_mirror_common(mirror, never_see_null,
162 region, null_path,
163 offset);
164 }
165 Node* generate_access_flags_guard(Node* kls,
166 int modifier_mask, int modifier_bits,
167 RegionNode* region);
168 Node* generate_interface_guard(Node* kls, RegionNode* region);
169 Node* generate_array_guard(Node* kls, RegionNode* region) {
170 return generate_array_guard_common(kls, region, false, false);
171 }
172 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
173 return generate_array_guard_common(kls, region, false, true);
174 }
175 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, true, false);
177 }
178 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
179 return generate_array_guard_common(kls, region, true, true);
180 }
181 Node* generate_array_guard_common(Node* kls, RegionNode* region,
182 bool obj_array, bool not_array);
183 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
184 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
185 bool is_virtual = false, bool is_static = false);
186 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
187 return generate_method_call(method_id, false, true);
188 }
189 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
190 return generate_method_call(method_id, true, false);
191 }
192 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
193
194 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
195 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
196 bool inline_string_compareTo();
197 bool inline_string_indexOf();
198 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
199 bool inline_string_equals();
200 Node* round_double_node(Node* n);
201 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
202 bool inline_math_native(vmIntrinsics::ID id);
203 bool inline_trig(vmIntrinsics::ID id);
204 bool inline_math(vmIntrinsics::ID id);
205 bool inline_exp();
206 bool inline_pow();
207 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
208 bool inline_min_max(vmIntrinsics::ID id);
209 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
210 // This returns Type::AnyPtr, RawPtr, or OopPtr.
211 int classify_unsafe_addr(Node* &base, Node* &offset);
212 Node* make_unsafe_address(Node* base, Node* offset);
213 // Helper for inline_unsafe_access.
214 // Generates the guards that check whether the result of
215 // Unsafe.getObject should be recorded in an SATB log buffer.
216 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
217 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
218 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
219 static bool klass_needs_init_guard(Node* kls);
220 bool inline_unsafe_allocate();
221 bool inline_unsafe_copyMemory();
222 bool inline_native_currentThread();
223 #ifdef TRACE_HAVE_INTRINSICS
224 bool inline_native_classID();
225 bool inline_native_threadID();
226 #endif
227 bool inline_native_time_funcs(address method, const char* funcName);
228 bool inline_native_isInterrupted();
229 bool inline_native_Class_query(vmIntrinsics::ID id);
230 bool inline_native_subtype_check();
231
232 bool inline_native_newArray();
233 bool inline_native_getLength();
234 bool inline_array_copyOf(bool is_copyOfRange);
235 bool inline_array_equals();
236 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
237 bool inline_native_clone(bool is_virtual);
238 bool inline_native_Reflection_getCallerClass();
239 // Helper function for inlining native object hash method
240 bool inline_native_hashcode(bool is_virtual, bool is_static);
241 bool inline_native_getClass();
242
243 // Helper functions for inlining arraycopy
244 bool inline_arraycopy();
245 void generate_arraycopy(const TypePtr* adr_type,
246 BasicType basic_elem_type,
247 Node* src, Node* src_offset,
248 Node* dest, Node* dest_offset,
249 Node* copy_length,
250 bool disjoint_bases = false,
251 bool length_never_negative = false,
252 RegionNode* slow_region = NULL);
253 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
254 RegionNode* slow_region);
255 void generate_clear_array(const TypePtr* adr_type,
256 Node* dest,
257 BasicType basic_elem_type,
258 Node* slice_off,
259 Node* slice_len,
260 Node* slice_end);
261 bool generate_block_arraycopy(const TypePtr* adr_type,
262 BasicType basic_elem_type,
263 AllocateNode* alloc,
264 Node* src, Node* src_offset,
265 Node* dest, Node* dest_offset,
266 Node* dest_size, bool dest_uninitialized);
267 void generate_slow_arraycopy(const TypePtr* adr_type,
268 Node* src, Node* src_offset,
269 Node* dest, Node* dest_offset,
270 Node* copy_length, bool dest_uninitialized);
271 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
272 Node* dest_elem_klass,
273 Node* src, Node* src_offset,
274 Node* dest, Node* dest_offset,
275 Node* copy_length, bool dest_uninitialized);
276 Node* generate_generic_arraycopy(const TypePtr* adr_type,
277 Node* src, Node* src_offset,
278 Node* dest, Node* dest_offset,
279 Node* copy_length, bool dest_uninitialized);
280 void generate_unchecked_arraycopy(const TypePtr* adr_type,
281 BasicType basic_elem_type,
282 bool disjoint_bases,
283 Node* src, Node* src_offset,
284 Node* dest, Node* dest_offset,
285 Node* copy_length, bool dest_uninitialized);
286 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
287 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
288 bool inline_unsafe_ordered_store(BasicType type);
289 bool inline_unsafe_fence(vmIntrinsics::ID id);
290 bool inline_fp_conversions(vmIntrinsics::ID id);
291 bool inline_number_methods(vmIntrinsics::ID id);
292 bool inline_reference_get();
293 bool inline_aescrypt_Block(vmIntrinsics::ID id);
294 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
295 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
296 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
297 bool inline_encodeISOArray();
298 bool inline_updateCRC32();
299 bool inline_updateBytesCRC32();
300 bool inline_updateByteBufferCRC32();
301 };
302
303
304 //---------------------------make_vm_intrinsic----------------------------
305 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
306 vmIntrinsics::ID id = m->intrinsic_id();
307 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
308
309 if (DisableIntrinsic[0] != '\0'
310 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
311 // disabled by a user request on the command line:
312 // example: -XX:DisableIntrinsic=_hashCode,_getClass
313 return NULL;
314 }
315
316 if (!m->is_loaded()) {
317 // do not attempt to inline unloaded methods
318 return NULL;
319 }
320
321 // Only a few intrinsics implement a virtual dispatch.
322 // They are expensive calls which are also frequently overridden.
323 if (is_virtual) {
324 switch (id) {
325 case vmIntrinsics::_hashCode:
326 case vmIntrinsics::_clone:
327 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
328 break;
329 default:
330 return NULL;
331 }
332 }
333
334 // -XX:-InlineNatives disables nearly all intrinsics:
335 if (!InlineNatives) {
336 switch (id) {
337 case vmIntrinsics::_indexOf:
338 case vmIntrinsics::_compareTo:
339 case vmIntrinsics::_equals:
340 case vmIntrinsics::_equalsC:
341 case vmIntrinsics::_getAndAddInt:
342 case vmIntrinsics::_getAndAddLong:
343 case vmIntrinsics::_getAndSetInt:
344 case vmIntrinsics::_getAndSetLong:
345 case vmIntrinsics::_getAndSetObject:
346 case vmIntrinsics::_loadFence:
347 case vmIntrinsics::_storeFence:
348 case vmIntrinsics::_fullFence:
349 break; // InlineNatives does not control String.compareTo
350 case vmIntrinsics::_Reference_get:
351 break; // InlineNatives does not control Reference.get
352 default:
353 return NULL;
354 }
355 }
356
357 bool is_predicted = false;
358 bool does_virtual_dispatch = false;
359
360 switch (id) {
361 case vmIntrinsics::_compareTo:
362 if (!SpecialStringCompareTo) return NULL;
363 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
364 break;
365 case vmIntrinsics::_indexOf:
366 if (!SpecialStringIndexOf) return NULL;
367 break;
368 case vmIntrinsics::_equals:
369 if (!SpecialStringEquals) return NULL;
370 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
371 break;
372 case vmIntrinsics::_equalsC:
373 if (!SpecialArraysEquals) return NULL;
374 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
375 break;
376 case vmIntrinsics::_arraycopy:
377 if (!InlineArrayCopy) return NULL;
378 break;
379 case vmIntrinsics::_copyMemory:
380 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
381 if (!InlineArrayCopy) return NULL;
382 break;
383 case vmIntrinsics::_hashCode:
384 if (!InlineObjectHash) return NULL;
385 does_virtual_dispatch = true;
386 break;
387 case vmIntrinsics::_clone:
388 does_virtual_dispatch = true;
389 case vmIntrinsics::_copyOf:
390 case vmIntrinsics::_copyOfRange:
391 if (!InlineObjectCopy) return NULL;
392 // These also use the arraycopy intrinsic mechanism:
393 if (!InlineArrayCopy) return NULL;
394 break;
395 case vmIntrinsics::_encodeISOArray:
396 if (!SpecialEncodeISOArray) return NULL;
397 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
398 break;
399 case vmIntrinsics::_checkIndex:
400 // We do not intrinsify this. The optimizer does fine with it.
401 return NULL;
402
403 case vmIntrinsics::_getCallerClass:
404 if (!UseNewReflection) return NULL;
405 if (!InlineReflectionGetCallerClass) return NULL;
406 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
407 break;
408
409 case vmIntrinsics::_bitCount_i:
410 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
411 break;
412
413 case vmIntrinsics::_bitCount_l:
414 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
415 break;
416
417 case vmIntrinsics::_numberOfLeadingZeros_i:
418 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
419 break;
420
421 case vmIntrinsics::_numberOfLeadingZeros_l:
422 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
423 break;
424
425 case vmIntrinsics::_numberOfTrailingZeros_i:
426 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
427 break;
428
429 case vmIntrinsics::_numberOfTrailingZeros_l:
430 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
431 break;
432
433 case vmIntrinsics::_reverseBytes_c:
434 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
435 break;
436 case vmIntrinsics::_reverseBytes_s:
437 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
438 break;
439 case vmIntrinsics::_reverseBytes_i:
440 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
441 break;
442 case vmIntrinsics::_reverseBytes_l:
443 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
444 break;
445
446 case vmIntrinsics::_Reference_get:
447 // Use the intrinsic version of Reference.get() so that the value in
448 // the referent field can be registered by the G1 pre-barrier code.
449 // Also add memory barrier to prevent commoning reads from this field
450 // across safepoint since GC can change it value.
451 break;
452
453 case vmIntrinsics::_compareAndSwapObject:
454 #ifdef _LP64
455 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
456 #endif
457 break;
458
459 case vmIntrinsics::_compareAndSwapLong:
460 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
461 break;
462
463 case vmIntrinsics::_getAndAddInt:
464 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
465 break;
466
467 case vmIntrinsics::_getAndAddLong:
468 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
469 break;
470
471 case vmIntrinsics::_getAndSetInt:
472 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
473 break;
474
475 case vmIntrinsics::_getAndSetLong:
476 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
477 break;
478
479 case vmIntrinsics::_getAndSetObject:
480 #ifdef _LP64
481 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
482 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
483 break;
484 #else
485 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
486 break;
487 #endif
488
489 case vmIntrinsics::_aescrypt_encryptBlock:
490 case vmIntrinsics::_aescrypt_decryptBlock:
491 if (!UseAESIntrinsics) return NULL;
492 break;
493
494 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
495 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
496 if (!UseAESIntrinsics) return NULL;
497 // these two require the predicated logic
498 is_predicted = true;
499 break;
500
501 case vmIntrinsics::_updateCRC32:
502 case vmIntrinsics::_updateBytesCRC32:
503 case vmIntrinsics::_updateByteBufferCRC32:
504 if (!UseCRC32Intrinsics) return NULL;
505 break;
506
507 default:
508 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
509 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
510 break;
511 }
512
513 // -XX:-InlineClassNatives disables natives from the Class class.
514 // The flag applies to all reflective calls, notably Array.newArray
515 // (visible to Java programmers as Array.newInstance).
516 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
517 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
518 if (!InlineClassNatives) return NULL;
519 }
520
521 // -XX:-InlineThreadNatives disables natives from the Thread class.
522 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
523 if (!InlineThreadNatives) return NULL;
524 }
525
526 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
527 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
528 m->holder()->name() == ciSymbol::java_lang_Float() ||
529 m->holder()->name() == ciSymbol::java_lang_Double()) {
530 if (!InlineMathNatives) return NULL;
531 }
532
533 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
534 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
535 if (!InlineUnsafeOps) return NULL;
536 }
537
538 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
539 }
540
541 //----------------------register_library_intrinsics-----------------------
542 // Initialize this file's data structures, for each Compile instance.
543 void Compile::register_library_intrinsics() {
544 // Nothing to do here.
545 }
546
547 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
548 LibraryCallKit kit(jvms, this);
549 Compile* C = kit.C;
550 int nodes = C->unique();
551 #ifndef PRODUCT
552 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
553 char buf[1000];
554 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
555 tty->print_cr("Intrinsic %s", str);
556 }
557 #endif
558 ciMethod* callee = kit.callee();
559 const int bci = kit.bci();
560
561 // Try to inline the intrinsic.
562 if (kit.try_to_inline()) {
563 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
564 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
565 }
566 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
567 if (C->log()) {
568 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
569 vmIntrinsics::name_at(intrinsic_id()),
570 (is_virtual() ? " virtual='1'" : ""),
571 C->unique() - nodes);
572 }
573 // Push the result from the inlined method onto the stack.
574 kit.push_result();
575 return kit.transfer_exceptions_into_jvms();
576 }
577
578 // The intrinsic bailed out
579 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
580 if (jvms->has_method()) {
581 // Not a root compile.
582 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
583 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
584 } else {
585 // Root compile
586 tty->print("Did not generate intrinsic %s%s at bci:%d in",
587 vmIntrinsics::name_at(intrinsic_id()),
588 (is_virtual() ? " (virtual)" : ""), bci);
589 }
590 }
591 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
592 return NULL;
593 }
594
595 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
596 LibraryCallKit kit(jvms, this);
597 Compile* C = kit.C;
598 int nodes = C->unique();
599 #ifndef PRODUCT
600 assert(is_predicted(), "sanity");
601 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
602 char buf[1000];
603 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
604 tty->print_cr("Predicate for intrinsic %s", str);
605 }
606 #endif
607 ciMethod* callee = kit.callee();
608 const int bci = kit.bci();
609
610 Node* slow_ctl = kit.try_to_predicate();
611 if (!kit.failing()) {
612 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
613 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
614 }
615 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
616 if (C->log()) {
617 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
618 vmIntrinsics::name_at(intrinsic_id()),
619 (is_virtual() ? " virtual='1'" : ""),
620 C->unique() - nodes);
621 }
622 return slow_ctl; // Could be NULL if the check folds.
623 }
624
625 // The intrinsic bailed out
626 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
627 if (jvms->has_method()) {
628 // Not a root compile.
629 const char* msg = "failed to generate predicate for intrinsic";
630 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
631 } else {
632 // Root compile
633 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
634 vmIntrinsics::name_at(intrinsic_id()),
635 (is_virtual() ? " (virtual)" : ""), bci);
636 }
637 }
638 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
639 return NULL;
640 }
641
642 bool LibraryCallKit::try_to_inline() {
643 // Handle symbolic names for otherwise undistinguished boolean switches:
644 const bool is_store = true;
645 const bool is_native_ptr = true;
646 const bool is_static = true;
647 const bool is_volatile = true;
648
649 if (!jvms()->has_method()) {
650 // Root JVMState has a null method.
651 assert(map()->memory()->Opcode() == Op_Parm, "");
652 // Insert the memory aliasing node
653 set_all_memory(reset_memory());
654 }
655 assert(merged_memory(), "");
656
657
658 switch (intrinsic_id()) {
659 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
660 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
661 case vmIntrinsics::_getClass: return inline_native_getClass();
662
663 case vmIntrinsics::_dsin:
664 case vmIntrinsics::_dcos:
665 case vmIntrinsics::_dtan:
666 case vmIntrinsics::_dabs:
667 case vmIntrinsics::_datan2:
668 case vmIntrinsics::_dsqrt:
669 case vmIntrinsics::_dexp:
670 case vmIntrinsics::_dlog:
671 case vmIntrinsics::_dlog10:
672 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
673
674 case vmIntrinsics::_min:
675 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
676
677 case vmIntrinsics::_arraycopy: return inline_arraycopy();
678
679 case vmIntrinsics::_compareTo: return inline_string_compareTo();
680 case vmIntrinsics::_indexOf: return inline_string_indexOf();
681 case vmIntrinsics::_equals: return inline_string_equals();
682
683 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
684 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
685 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
686 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
687 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
688 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
689 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
690 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
691 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
692
693 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
694 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
695 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
696 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
697 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
698 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
699 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
700 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
701 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
702
703 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
704 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
705 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
706 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
707 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
708 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
709 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
710 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
711
712 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
713 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
714 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
715 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
716 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
717 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
718 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
719 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
720
721 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
722 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
723 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
724 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
725 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
726 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
727 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
728 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
729 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
730
731 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
732 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
733 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
734 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
735 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
736 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
737 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
738 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
739 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
740
741 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
742 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
743 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
744 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
745
746 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
747 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
748 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
749
750 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
751 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
752 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
753
754 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
755 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
756 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
757 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
758 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
759
760 case vmIntrinsics::_loadFence:
761 case vmIntrinsics::_storeFence:
762 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
763
764 case vmIntrinsics::_currentThread: return inline_native_currentThread();
765 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
766
767 #ifdef TRACE_HAVE_INTRINSICS
768 case vmIntrinsics::_classID: return inline_native_classID();
769 case vmIntrinsics::_threadID: return inline_native_threadID();
770 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
771 #endif
772 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
773 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
774 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
775 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
776 case vmIntrinsics::_newArray: return inline_native_newArray();
777 case vmIntrinsics::_getLength: return inline_native_getLength();
778 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
779 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
780 case vmIntrinsics::_equalsC: return inline_array_equals();
781 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
782
783 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
784
785 case vmIntrinsics::_isInstance:
786 case vmIntrinsics::_getModifiers:
787 case vmIntrinsics::_isInterface:
788 case vmIntrinsics::_isArray:
789 case vmIntrinsics::_isPrimitive:
790 case vmIntrinsics::_getSuperclass:
791 case vmIntrinsics::_getComponentType:
792 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
793
794 case vmIntrinsics::_floatToRawIntBits:
795 case vmIntrinsics::_floatToIntBits:
796 case vmIntrinsics::_intBitsToFloat:
797 case vmIntrinsics::_doubleToRawLongBits:
798 case vmIntrinsics::_doubleToLongBits:
799 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
800
801 case vmIntrinsics::_numberOfLeadingZeros_i:
802 case vmIntrinsics::_numberOfLeadingZeros_l:
803 case vmIntrinsics::_numberOfTrailingZeros_i:
804 case vmIntrinsics::_numberOfTrailingZeros_l:
805 case vmIntrinsics::_bitCount_i:
806 case vmIntrinsics::_bitCount_l:
807 case vmIntrinsics::_reverseBytes_i:
808 case vmIntrinsics::_reverseBytes_l:
809 case vmIntrinsics::_reverseBytes_s:
810 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
811
812 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
813
814 case vmIntrinsics::_Reference_get: return inline_reference_get();
815
816 case vmIntrinsics::_aescrypt_encryptBlock:
817 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
818
819 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
820 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
821 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
822
823 case vmIntrinsics::_encodeISOArray:
824 return inline_encodeISOArray();
825
826 case vmIntrinsics::_updateCRC32:
827 return inline_updateCRC32();
828 case vmIntrinsics::_updateBytesCRC32:
829 return inline_updateBytesCRC32();
830 case vmIntrinsics::_updateByteBufferCRC32:
831 return inline_updateByteBufferCRC32();
832
833 default:
834 // If you get here, it may be that someone has added a new intrinsic
835 // to the list in vmSymbols.hpp without implementing it here.
836 #ifndef PRODUCT
837 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
838 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
839 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
840 }
841 #endif
842 return false;
843 }
844 }
845
846 Node* LibraryCallKit::try_to_predicate() {
847 if (!jvms()->has_method()) {
848 // Root JVMState has a null method.
849 assert(map()->memory()->Opcode() == Op_Parm, "");
850 // Insert the memory aliasing node
851 set_all_memory(reset_memory());
852 }
853 assert(merged_memory(), "");
854
855 switch (intrinsic_id()) {
856 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
857 return inline_cipherBlockChaining_AESCrypt_predicate(false);
858 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
859 return inline_cipherBlockChaining_AESCrypt_predicate(true);
860
861 default:
862 // If you get here, it may be that someone has added a new intrinsic
863 // to the list in vmSymbols.hpp without implementing it here.
864 #ifndef PRODUCT
865 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
866 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
867 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
868 }
869 #endif
870 Node* slow_ctl = control();
871 set_control(top()); // No fast path instrinsic
872 return slow_ctl;
873 }
874 }
875
876 //------------------------------set_result-------------------------------
877 // Helper function for finishing intrinsics.
878 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
879 record_for_igvn(region);
880 set_control(_gvn.transform(region));
881 set_result( _gvn.transform(value));
882 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
883 }
884
885 //------------------------------generate_guard---------------------------
886 // Helper function for generating guarded fast-slow graph structures.
887 // The given 'test', if true, guards a slow path. If the test fails
888 // then a fast path can be taken. (We generally hope it fails.)
889 // In all cases, GraphKit::control() is updated to the fast path.
890 // The returned value represents the control for the slow path.
891 // The return value is never 'top'; it is either a valid control
892 // or NULL if it is obvious that the slow path can never be taken.
893 // Also, if region and the slow control are not NULL, the slow edge
894 // is appended to the region.
895 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
896 if (stopped()) {
897 // Already short circuited.
898 return NULL;
899 }
900
901 // Build an if node and its projections.
902 // If test is true we take the slow path, which we assume is uncommon.
903 if (_gvn.type(test) == TypeInt::ZERO) {
904 // The slow branch is never taken. No need to build this guard.
905 return NULL;
906 }
907
908 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
909
910 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
911 if (if_slow == top()) {
912 // The slow branch is never taken. No need to build this guard.
913 return NULL;
914 }
915
916 if (region != NULL)
917 region->add_req(if_slow);
918
919 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
920 set_control(if_fast);
921
922 return if_slow;
923 }
924
925 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
926 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
927 }
928 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
929 return generate_guard(test, region, PROB_FAIR);
930 }
931
932 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
933 Node* *pos_index) {
934 if (stopped())
935 return NULL; // already stopped
936 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
937 return NULL; // index is already adequately typed
938 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
939 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
940 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
941 if (is_neg != NULL && pos_index != NULL) {
942 // Emulate effect of Parse::adjust_map_after_if.
943 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
944 ccast->set_req(0, control());
945 (*pos_index) = _gvn.transform(ccast);
946 }
947 return is_neg;
948 }
949
950 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
951 Node* *pos_index) {
952 if (stopped())
953 return NULL; // already stopped
954 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
955 return NULL; // index is already adequately typed
956 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
957 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
958 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
959 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
960 if (is_notp != NULL && pos_index != NULL) {
961 // Emulate effect of Parse::adjust_map_after_if.
962 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
963 ccast->set_req(0, control());
964 (*pos_index) = _gvn.transform(ccast);
965 }
966 return is_notp;
967 }
968
969 // Make sure that 'position' is a valid limit index, in [0..length].
970 // There are two equivalent plans for checking this:
971 // A. (offset + copyLength) unsigned<= arrayLength
972 // B. offset <= (arrayLength - copyLength)
973 // We require that all of the values above, except for the sum and
974 // difference, are already known to be non-negative.
975 // Plan A is robust in the face of overflow, if offset and copyLength
976 // are both hugely positive.
977 //
978 // Plan B is less direct and intuitive, but it does not overflow at
979 // all, since the difference of two non-negatives is always
980 // representable. Whenever Java methods must perform the equivalent
981 // check they generally use Plan B instead of Plan A.
982 // For the moment we use Plan A.
983 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
984 Node* subseq_length,
985 Node* array_length,
986 RegionNode* region) {
987 if (stopped())
988 return NULL; // already stopped
989 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
990 if (zero_offset && subseq_length->eqv_uncast(array_length))
991 return NULL; // common case of whole-array copy
992 Node* last = subseq_length;
993 if (!zero_offset) // last += offset
994 last = _gvn.transform(new (C) AddINode(last, offset));
995 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
996 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
997 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
998 return is_over;
999 }
1000
1001
1002 //--------------------------generate_current_thread--------------------
1003 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1004 ciKlass* thread_klass = env()->Thread_klass();
1005 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1006 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1007 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1008 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1009 tls_output = thread;
1010 return threadObj;
1011 }
1012
1013
1014 //------------------------------make_string_method_node------------------------
1015 // Helper method for String intrinsic functions. This version is called
1016 // with str1 and str2 pointing to String object nodes.
1017 //
1018 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1019 Node* no_ctrl = NULL;
1020
1021 // Get start addr of string
1022 Node* str1_value = load_String_value(no_ctrl, str1);
1023 Node* str1_offset = load_String_offset(no_ctrl, str1);
1024 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1025
1026 // Get length of string 1
1027 Node* str1_len = load_String_length(no_ctrl, str1);
1028
1029 Node* str2_value = load_String_value(no_ctrl, str2);
1030 Node* str2_offset = load_String_offset(no_ctrl, str2);
1031 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1032
1033 Node* str2_len = NULL;
1034 Node* result = NULL;
1035
1036 switch (opcode) {
1037 case Op_StrIndexOf:
1038 // Get length of string 2
1039 str2_len = load_String_length(no_ctrl, str2);
1040
1041 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1042 str1_start, str1_len, str2_start, str2_len);
1043 break;
1044 case Op_StrComp:
1045 // Get length of string 2
1046 str2_len = load_String_length(no_ctrl, str2);
1047
1048 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1049 str1_start, str1_len, str2_start, str2_len);
1050 break;
1051 case Op_StrEquals:
1052 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1053 str1_start, str2_start, str1_len);
1054 break;
1055 default:
1056 ShouldNotReachHere();
1057 return NULL;
1058 }
1059
1060 // All these intrinsics have checks.
1061 C->set_has_split_ifs(true); // Has chance for split-if optimization
1062
1063 return _gvn.transform(result);
1064 }
1065
1066 // Helper method for String intrinsic functions. This version is called
1067 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1068 // to Int nodes containing the lenghts of str1 and str2.
1069 //
1070 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1071 Node* result = NULL;
1072 switch (opcode) {
1073 case Op_StrIndexOf:
1074 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1075 str1_start, cnt1, str2_start, cnt2);
1076 break;
1077 case Op_StrComp:
1078 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1079 str1_start, cnt1, str2_start, cnt2);
1080 break;
1081 case Op_StrEquals:
1082 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1083 str1_start, str2_start, cnt1);
1084 break;
1085 default:
1086 ShouldNotReachHere();
1087 return NULL;
1088 }
1089
1090 // All these intrinsics have checks.
1091 C->set_has_split_ifs(true); // Has chance for split-if optimization
1092
1093 return _gvn.transform(result);
1094 }
1095
1096 //------------------------------inline_string_compareTo------------------------
1097 // public int java.lang.String.compareTo(String anotherString);
1098 bool LibraryCallKit::inline_string_compareTo() {
1099 Node* receiver = null_check(argument(0));
1100 Node* arg = null_check(argument(1));
1101 if (stopped()) {
1102 return true;
1103 }
1104 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1105 return true;
1106 }
1107
1108 //------------------------------inline_string_equals------------------------
1109 bool LibraryCallKit::inline_string_equals() {
1110 Node* receiver = null_check_receiver();
1111 // NOTE: Do not null check argument for String.equals() because spec
1112 // allows to specify NULL as argument.
1113 Node* argument = this->argument(1);
1114 if (stopped()) {
1115 return true;
1116 }
1117
1118 // paths (plus control) merge
1119 RegionNode* region = new (C) RegionNode(5);
1120 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1121
1122 // does source == target string?
1123 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1124 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1125
1126 Node* if_eq = generate_slow_guard(bol, NULL);
1127 if (if_eq != NULL) {
1128 // receiver == argument
1129 phi->init_req(2, intcon(1));
1130 region->init_req(2, if_eq);
1131 }
1132
1133 // get String klass for instanceOf
1134 ciInstanceKlass* klass = env()->String_klass();
1135
1136 if (!stopped()) {
1137 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1138 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1139 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1140
1141 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1142 //instanceOf == true, fallthrough
1143
1144 if (inst_false != NULL) {
1145 phi->init_req(3, intcon(0));
1146 region->init_req(3, inst_false);
1147 }
1148 }
1149
1150 if (!stopped()) {
1151 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1152
1153 // Properly cast the argument to String
1154 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1155 // This path is taken only when argument's type is String:NotNull.
1156 argument = cast_not_null(argument, false);
1157
1158 Node* no_ctrl = NULL;
1159
1160 // Get start addr of receiver
1161 Node* receiver_val = load_String_value(no_ctrl, receiver);
1162 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1163 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1164
1165 // Get length of receiver
1166 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1167
1168 // Get start addr of argument
1169 Node* argument_val = load_String_value(no_ctrl, argument);
1170 Node* argument_offset = load_String_offset(no_ctrl, argument);
1171 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1172
1173 // Get length of argument
1174 Node* argument_cnt = load_String_length(no_ctrl, argument);
1175
1176 // Check for receiver count != argument count
1177 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1178 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1179 Node* if_ne = generate_slow_guard(bol, NULL);
1180 if (if_ne != NULL) {
1181 phi->init_req(4, intcon(0));
1182 region->init_req(4, if_ne);
1183 }
1184
1185 // Check for count == 0 is done by assembler code for StrEquals.
1186
1187 if (!stopped()) {
1188 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1189 phi->init_req(1, equals);
1190 region->init_req(1, control());
1191 }
1192 }
1193
1194 // post merge
1195 set_control(_gvn.transform(region));
1196 record_for_igvn(region);
1197
1198 set_result(_gvn.transform(phi));
1199 return true;
1200 }
1201
1202 //------------------------------inline_array_equals----------------------------
1203 bool LibraryCallKit::inline_array_equals() {
1204 Node* arg1 = argument(0);
1205 Node* arg2 = argument(1);
1206 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1207 return true;
1208 }
1209
1210 // Java version of String.indexOf(constant string)
1211 // class StringDecl {
1212 // StringDecl(char[] ca) {
1213 // offset = 0;
1214 // count = ca.length;
1215 // value = ca;
1216 // }
1217 // int offset;
1218 // int count;
1219 // char[] value;
1220 // }
1221 //
1222 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1223 // int targetOffset, int cache_i, int md2) {
1224 // int cache = cache_i;
1225 // int sourceOffset = string_object.offset;
1226 // int sourceCount = string_object.count;
1227 // int targetCount = target_object.length;
1228 //
1229 // int targetCountLess1 = targetCount - 1;
1230 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1231 //
1232 // char[] source = string_object.value;
1233 // char[] target = target_object;
1234 // int lastChar = target[targetCountLess1];
1235 //
1236 // outer_loop:
1237 // for (int i = sourceOffset; i < sourceEnd; ) {
1238 // int src = source[i + targetCountLess1];
1239 // if (src == lastChar) {
1240 // // With random strings and a 4-character alphabet,
1241 // // reverse matching at this point sets up 0.8% fewer
1242 // // frames, but (paradoxically) makes 0.3% more probes.
1243 // // Since those probes are nearer the lastChar probe,
1244 // // there is may be a net D$ win with reverse matching.
1245 // // But, reversing loop inhibits unroll of inner loop
1246 // // for unknown reason. So, does running outer loop from
1247 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1248 // for (int j = 0; j < targetCountLess1; j++) {
1249 // if (target[targetOffset + j] != source[i+j]) {
1250 // if ((cache & (1 << source[i+j])) == 0) {
1251 // if (md2 < j+1) {
1252 // i += j+1;
1253 // continue outer_loop;
1254 // }
1255 // }
1256 // i += md2;
1257 // continue outer_loop;
1258 // }
1259 // }
1260 // return i - sourceOffset;
1261 // }
1262 // if ((cache & (1 << src)) == 0) {
1263 // i += targetCountLess1;
1264 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1265 // i++;
1266 // }
1267 // return -1;
1268 // }
1269
1270 //------------------------------string_indexOf------------------------
1271 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1272 jint cache_i, jint md2_i) {
1273
1274 Node* no_ctrl = NULL;
1275 float likely = PROB_LIKELY(0.9);
1276 float unlikely = PROB_UNLIKELY(0.9);
1277
1278 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1279
1280 Node* source = load_String_value(no_ctrl, string_object);
1281 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1282 Node* sourceCount = load_String_length(no_ctrl, string_object);
1283
1284 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1285 jint target_length = target_array->length();
1286 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1287 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1288
1289 // String.value field is known to be @Stable.
1290 if (UseImplicitStableValues) {
1291 target = cast_array_to_stable(target, target_type);
1292 }
1293
1294 IdealKit kit(this, false, true);
1295 #define __ kit.
1296 Node* zero = __ ConI(0);
1297 Node* one = __ ConI(1);
1298 Node* cache = __ ConI(cache_i);
1299 Node* md2 = __ ConI(md2_i);
1300 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1301 Node* targetCount = __ ConI(target_length);
1302 Node* targetCountLess1 = __ ConI(target_length - 1);
1303 Node* targetOffset = __ ConI(targetOffset_i);
1304 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1305
1306 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1307 Node* outer_loop = __ make_label(2 /* goto */);
1308 Node* return_ = __ make_label(1);
1309
1310 __ set(rtn,__ ConI(-1));
1311 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1312 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1313 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1314 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1315 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1316 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1317 Node* tpj = __ AddI(targetOffset, __ value(j));
1318 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1319 Node* ipj = __ AddI(__ value(i), __ value(j));
1320 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1321 __ if_then(targ, BoolTest::ne, src2); {
1322 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1323 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1324 __ increment(i, __ AddI(__ value(j), one));
1325 __ goto_(outer_loop);
1326 } __ end_if(); __ dead(j);
1327 }__ end_if(); __ dead(j);
1328 __ increment(i, md2);
1329 __ goto_(outer_loop);
1330 }__ end_if();
1331 __ increment(j, one);
1332 }__ end_loop(); __ dead(j);
1333 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1334 __ goto_(return_);
1335 }__ end_if();
1336 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1337 __ increment(i, targetCountLess1);
1338 }__ end_if();
1339 __ increment(i, one);
1340 __ bind(outer_loop);
1341 }__ end_loop(); __ dead(i);
1342 __ bind(return_);
1343
1344 // Final sync IdealKit and GraphKit.
1345 final_sync(kit);
1346 Node* result = __ value(rtn);
1347 #undef __
1348 C->set_has_loops(true);
1349 return result;
1350 }
1351
1352 //------------------------------inline_string_indexOf------------------------
1353 bool LibraryCallKit::inline_string_indexOf() {
1354 Node* receiver = argument(0);
1355 Node* arg = argument(1);
1356
1357 Node* result;
1358 // Disable the use of pcmpestri until it can be guaranteed that
1359 // the load doesn't cross into the uncommited space.
1360 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1361 UseSSE42Intrinsics) {
1362 // Generate SSE4.2 version of indexOf
1363 // We currently only have match rules that use SSE4.2
1364
1365 receiver = null_check(receiver);
1366 arg = null_check(arg);
1367 if (stopped()) {
1368 return true;
1369 }
1370
1371 ciInstanceKlass* str_klass = env()->String_klass();
1372 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1373
1374 // Make the merge point
1375 RegionNode* result_rgn = new (C) RegionNode(4);
1376 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1377 Node* no_ctrl = NULL;
1378
1379 // Get start addr of source string
1380 Node* source = load_String_value(no_ctrl, receiver);
1381 Node* source_offset = load_String_offset(no_ctrl, receiver);
1382 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1383
1384 // Get length of source string
1385 Node* source_cnt = load_String_length(no_ctrl, receiver);
1386
1387 // Get start addr of substring
1388 Node* substr = load_String_value(no_ctrl, arg);
1389 Node* substr_offset = load_String_offset(no_ctrl, arg);
1390 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1391
1392 // Get length of source string
1393 Node* substr_cnt = load_String_length(no_ctrl, arg);
1394
1395 // Check for substr count > string count
1396 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1397 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1398 Node* if_gt = generate_slow_guard(bol, NULL);
1399 if (if_gt != NULL) {
1400 result_phi->init_req(2, intcon(-1));
1401 result_rgn->init_req(2, if_gt);
1402 }
1403
1404 if (!stopped()) {
1405 // Check for substr count == 0
1406 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1407 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1408 Node* if_zero = generate_slow_guard(bol, NULL);
1409 if (if_zero != NULL) {
1410 result_phi->init_req(3, intcon(0));
1411 result_rgn->init_req(3, if_zero);
1412 }
1413 }
1414
1415 if (!stopped()) {
1416 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1417 result_phi->init_req(1, result);
1418 result_rgn->init_req(1, control());
1419 }
1420 set_control(_gvn.transform(result_rgn));
1421 record_for_igvn(result_rgn);
1422 result = _gvn.transform(result_phi);
1423
1424 } else { // Use LibraryCallKit::string_indexOf
1425 // don't intrinsify if argument isn't a constant string.
1426 if (!arg->is_Con()) {
1427 return false;
1428 }
1429 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1430 if (str_type == NULL) {
1431 return false;
1432 }
1433 ciInstanceKlass* klass = env()->String_klass();
1434 ciObject* str_const = str_type->const_oop();
1435 if (str_const == NULL || str_const->klass() != klass) {
1436 return false;
1437 }
1438 ciInstance* str = str_const->as_instance();
1439 assert(str != NULL, "must be instance");
1440
1441 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1442 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1443
1444 int o;
1445 int c;
1446 if (java_lang_String::has_offset_field()) {
1447 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1448 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1449 } else {
1450 o = 0;
1451 c = pat->length();
1452 }
1453
1454 // constant strings have no offset and count == length which
1455 // simplifies the resulting code somewhat so lets optimize for that.
1456 if (o != 0 || c != pat->length()) {
1457 return false;
1458 }
1459
1460 receiver = null_check(receiver, T_OBJECT);
1461 // NOTE: No null check on the argument is needed since it's a constant String oop.
1462 if (stopped()) {
1463 return true;
1464 }
1465
1466 // The null string as a pattern always returns 0 (match at beginning of string)
1467 if (c == 0) {
1468 set_result(intcon(0));
1469 return true;
1470 }
1471
1472 // Generate default indexOf
1473 jchar lastChar = pat->char_at(o + (c - 1));
1474 int cache = 0;
1475 int i;
1476 for (i = 0; i < c - 1; i++) {
1477 assert(i < pat->length(), "out of range");
1478 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1479 }
1480
1481 int md2 = c;
1482 for (i = 0; i < c - 1; i++) {
1483 assert(i < pat->length(), "out of range");
1484 if (pat->char_at(o + i) == lastChar) {
1485 md2 = (c - 1) - i;
1486 }
1487 }
1488
1489 result = string_indexOf(receiver, pat, o, cache, md2);
1490 }
1491 set_result(result);
1492 return true;
1493 }
1494
1495 //--------------------------round_double_node--------------------------------
1496 // Round a double node if necessary.
1497 Node* LibraryCallKit::round_double_node(Node* n) {
1498 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1499 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1500 return n;
1501 }
1502
1503 //------------------------------inline_math-----------------------------------
1504 // public static double Math.abs(double)
1505 // public static double Math.sqrt(double)
1506 // public static double Math.log(double)
1507 // public static double Math.log10(double)
1508 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1509 Node* arg = round_double_node(argument(0));
1510 Node* n;
1511 switch (id) {
1512 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1513 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1514 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1515 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1516 default: fatal_unexpected_iid(id); break;
1517 }
1518 set_result(_gvn.transform(n));
1519 return true;
1520 }
1521
1522 //------------------------------inline_trig----------------------------------
1523 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1524 // argument reduction which will turn into a fast/slow diamond.
1525 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1526 Node* arg = round_double_node(argument(0));
1527 Node* n = NULL;
1528
1529 switch (id) {
1530 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1531 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1532 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1533 default: fatal_unexpected_iid(id); break;
1534 }
1535 n = _gvn.transform(n);
1536
1537 // Rounding required? Check for argument reduction!
1538 if (Matcher::strict_fp_requires_explicit_rounding) {
1539 static const double pi_4 = 0.7853981633974483;
1540 static const double neg_pi_4 = -0.7853981633974483;
1541 // pi/2 in 80-bit extended precision
1542 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1543 // -pi/2 in 80-bit extended precision
1544 // 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};
1545 // Cutoff value for using this argument reduction technique
1546 //static const double pi_2_minus_epsilon = 1.564660403643354;
1547 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1548
1549 // Pseudocode for sin:
1550 // if (x <= Math.PI / 4.0) {
1551 // if (x >= -Math.PI / 4.0) return fsin(x);
1552 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1553 // } else {
1554 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1555 // }
1556 // return StrictMath.sin(x);
1557
1558 // Pseudocode for cos:
1559 // if (x <= Math.PI / 4.0) {
1560 // if (x >= -Math.PI / 4.0) return fcos(x);
1561 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1562 // } else {
1563 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1564 // }
1565 // return StrictMath.cos(x);
1566
1567 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1568 // requires a special machine instruction to load it. Instead we'll try
1569 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1570 // probably do the math inside the SIN encoding.
1571
1572 // Make the merge point
1573 RegionNode* r = new (C) RegionNode(3);
1574 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1575
1576 // Flatten arg so we need only 1 test
1577 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1578 // Node for PI/4 constant
1579 Node *pi4 = makecon(TypeD::make(pi_4));
1580 // Check PI/4 : abs(arg)
1581 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1582 // Check: If PI/4 < abs(arg) then go slow
1583 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1584 // Branch either way
1585 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1586 set_control(opt_iff(r,iff));
1587
1588 // Set fast path result
1589 phi->init_req(2, n);
1590
1591 // Slow path - non-blocking leaf call
1592 Node* call = NULL;
1593 switch (id) {
1594 case vmIntrinsics::_dsin:
1595 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1596 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1597 "Sin", NULL, arg, top());
1598 break;
1599 case vmIntrinsics::_dcos:
1600 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1601 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1602 "Cos", NULL, arg, top());
1603 break;
1604 case vmIntrinsics::_dtan:
1605 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1606 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1607 "Tan", NULL, arg, top());
1608 break;
1609 }
1610 assert(control()->in(0) == call, "");
1611 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1612 r->init_req(1, control());
1613 phi->init_req(1, slow_result);
1614
1615 // Post-merge
1616 set_control(_gvn.transform(r));
1617 record_for_igvn(r);
1618 n = _gvn.transform(phi);
1619
1620 C->set_has_split_ifs(true); // Has chance for split-if optimization
1621 }
1622 set_result(n);
1623 return true;
1624 }
1625
1626 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1627 //-------------------
1628 //result=(result.isNaN())? funcAddr():result;
1629 // Check: If isNaN() by checking result!=result? then either trap
1630 // or go to runtime
1631 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1632 // Build the boolean node
1633 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1634
1635 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1636 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1637 // The pow or exp intrinsic returned a NaN, which requires a call
1638 // to the runtime. Recompile with the runtime call.
1639 uncommon_trap(Deoptimization::Reason_intrinsic,
1640 Deoptimization::Action_make_not_entrant);
1641 }
1642 set_result(result);
1643 } else {
1644 // If this inlining ever returned NaN in the past, we compile a call
1645 // to the runtime to properly handle corner cases
1646
1647 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1648 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1649 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1650
1651 if (!if_slow->is_top()) {
1652 RegionNode* result_region = new (C) RegionNode(3);
1653 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1654
1655 result_region->init_req(1, if_fast);
1656 result_val->init_req(1, result);
1657
1658 set_control(if_slow);
1659
1660 const TypePtr* no_memory_effects = NULL;
1661 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1662 no_memory_effects,
1663 x, top(), y, y ? top() : NULL);
1664 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1665 #ifdef ASSERT
1666 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1667 assert(value_top == top(), "second value must be top");
1668 #endif
1669
1670 result_region->init_req(2, control());
1671 result_val->init_req(2, value);
1672 set_result(result_region, result_val);
1673 } else {
1674 set_result(result);
1675 }
1676 }
1677 }
1678
1679 //------------------------------inline_exp-------------------------------------
1680 // Inline exp instructions, if possible. The Intel hardware only misses
1681 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1682 bool LibraryCallKit::inline_exp() {
1683 Node* arg = round_double_node(argument(0));
1684 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1685
1686 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1687
1688 C->set_has_split_ifs(true); // Has chance for split-if optimization
1689 return true;
1690 }
1691
1692 //------------------------------inline_pow-------------------------------------
1693 // Inline power instructions, if possible.
1694 bool LibraryCallKit::inline_pow() {
1695 // Pseudocode for pow
1696 // if (x <= 0.0) {
1697 // long longy = (long)y;
1698 // if ((double)longy == y) { // if y is long
1699 // if (y + 1 == y) longy = 0; // huge number: even
1700 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1701 // } else {
1702 // result = NaN;
1703 // }
1704 // } else {
1705 // result = DPow(x,y);
1706 // }
1707 // if (result != result)? {
1708 // result = uncommon_trap() or runtime_call();
1709 // }
1710 // return result;
1711
1712 Node* x = round_double_node(argument(0));
1713 Node* y = round_double_node(argument(2));
1714
1715 Node* result = NULL;
1716
1717 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1718 // Short form: skip the fancy tests and just check for NaN result.
1719 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1720 } else {
1721 // If this inlining ever returned NaN in the past, include all
1722 // checks + call to the runtime.
1723
1724 // Set the merge point for If node with condition of (x <= 0.0)
1725 // There are four possible paths to region node and phi node
1726 RegionNode *r = new (C) RegionNode(4);
1727 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1728
1729 // Build the first if node: if (x <= 0.0)
1730 // Node for 0 constant
1731 Node *zeronode = makecon(TypeD::ZERO);
1732 // Check x:0
1733 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1734 // Check: If (x<=0) then go complex path
1735 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1736 // Branch either way
1737 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1738 // Fast path taken; set region slot 3
1739 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1740 r->init_req(3,fast_taken); // Capture fast-control
1741
1742 // Fast path not-taken, i.e. slow path
1743 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1744
1745 // Set fast path result
1746 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1747 phi->init_req(3, fast_result);
1748
1749 // Complex path
1750 // Build the second if node (if y is long)
1751 // Node for (long)y
1752 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1753 // Node for (double)((long) y)
1754 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1755 // Check (double)((long) y) : y
1756 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1757 // Check if (y isn't long) then go to slow path
1758
1759 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1760 // Branch either way
1761 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1762 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1763
1764 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1765
1766 // Calculate DPow(abs(x), y)*(1 & (long)y)
1767 // Node for constant 1
1768 Node *conone = longcon(1);
1769 // 1& (long)y
1770 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1771
1772 // A huge number is always even. Detect a huge number by checking
1773 // if y + 1 == y and set integer to be tested for parity to 0.
1774 // Required for corner case:
1775 // (long)9.223372036854776E18 = max_jlong
1776 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1777 // max_jlong is odd but 9.223372036854776E18 is even
1778 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1779 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1780 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1781 Node* correctedsign = NULL;
1782 if (ConditionalMoveLimit != 0) {
1783 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1784 } else {
1785 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1786 RegionNode *r = new (C) RegionNode(3);
1787 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1788 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1789 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1790 phi->init_req(1, signnode);
1791 phi->init_req(2, longcon(0));
1792 correctedsign = _gvn.transform(phi);
1793 ylong_path = _gvn.transform(r);
1794 record_for_igvn(r);
1795 }
1796
1797 // zero node
1798 Node *conzero = longcon(0);
1799 // Check (1&(long)y)==0?
1800 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1801 // Check if (1&(long)y)!=0?, if so the result is negative
1802 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1803 // abs(x)
1804 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1805 // abs(x)^y
1806 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1807 // -abs(x)^y
1808 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1809 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1810 Node *signresult = NULL;
1811 if (ConditionalMoveLimit != 0) {
1812 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1813 } else {
1814 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1815 RegionNode *r = new (C) RegionNode(3);
1816 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1817 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1818 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1819 phi->init_req(1, absxpowy);
1820 phi->init_req(2, negabsxpowy);
1821 signresult = _gvn.transform(phi);
1822 ylong_path = _gvn.transform(r);
1823 record_for_igvn(r);
1824 }
1825 // Set complex path fast result
1826 r->init_req(2, ylong_path);
1827 phi->init_req(2, signresult);
1828
1829 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1830 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1831 r->init_req(1,slow_path);
1832 phi->init_req(1,slow_result);
1833
1834 // Post merge
1835 set_control(_gvn.transform(r));
1836 record_for_igvn(r);
1837 result = _gvn.transform(phi);
1838 }
1839
1840 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1841
1842 C->set_has_split_ifs(true); // Has chance for split-if optimization
1843 return true;
1844 }
1845
1846 //------------------------------runtime_math-----------------------------
1847 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1848 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1849 "must be (DD)D or (D)D type");
1850
1851 // Inputs
1852 Node* a = round_double_node(argument(0));
1853 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1854
1855 const TypePtr* no_memory_effects = NULL;
1856 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1857 no_memory_effects,
1858 a, top(), b, b ? top() : NULL);
1859 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1860 #ifdef ASSERT
1861 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1862 assert(value_top == top(), "second value must be top");
1863 #endif
1864
1865 set_result(value);
1866 return true;
1867 }
1868
1869 //------------------------------inline_math_native-----------------------------
1870 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1871 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1872 switch (id) {
1873 // These intrinsics are not properly supported on all hardware
1874 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1875 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1876 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1877 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1878 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1879 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1880
1881 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1882 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1883 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1884 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1885
1886 // These intrinsics are supported on all hardware
1887 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1888 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1889
1890 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1891 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1892 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1893 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1894 #undef FN_PTR
1895
1896 // These intrinsics are not yet correctly implemented
1897 case vmIntrinsics::_datan2:
1898 return false;
1899
1900 default:
1901 fatal_unexpected_iid(id);
1902 return false;
1903 }
1904 }
1905
1906 static bool is_simple_name(Node* n) {
1907 return (n->req() == 1 // constant
1908 || (n->is_Type() && n->as_Type()->type()->singleton())
1909 || n->is_Proj() // parameter or return value
1910 || n->is_Phi() // local of some sort
1911 );
1912 }
1913
1914 //----------------------------inline_min_max-----------------------------------
1915 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1916 set_result(generate_min_max(id, argument(0), argument(1)));
1917 return true;
1918 }
1919
1920 Node*
1921 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1922 // These are the candidate return value:
1923 Node* xvalue = x0;
1924 Node* yvalue = y0;
1925
1926 if (xvalue == yvalue) {
1927 return xvalue;
1928 }
1929
1930 bool want_max = (id == vmIntrinsics::_max);
1931
1932 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1933 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1934 if (txvalue == NULL || tyvalue == NULL) return top();
1935 // This is not really necessary, but it is consistent with a
1936 // hypothetical MaxINode::Value method:
1937 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1938
1939 // %%% This folding logic should (ideally) be in a different place.
1940 // Some should be inside IfNode, and there to be a more reliable
1941 // transformation of ?: style patterns into cmoves. We also want
1942 // more powerful optimizations around cmove and min/max.
1943
1944 // Try to find a dominating comparison of these guys.
1945 // It can simplify the index computation for Arrays.copyOf
1946 // and similar uses of System.arraycopy.
1947 // First, compute the normalized version of CmpI(x, y).
1948 int cmp_op = Op_CmpI;
1949 Node* xkey = xvalue;
1950 Node* ykey = yvalue;
1951 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
1952 if (ideal_cmpxy->is_Cmp()) {
1953 // E.g., if we have CmpI(length - offset, count),
1954 // it might idealize to CmpI(length, count + offset)
1955 cmp_op = ideal_cmpxy->Opcode();
1956 xkey = ideal_cmpxy->in(1);
1957 ykey = ideal_cmpxy->in(2);
1958 }
1959
1960 // Start by locating any relevant comparisons.
1961 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1962 Node* cmpxy = NULL;
1963 Node* cmpyx = NULL;
1964 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1965 Node* cmp = start_from->fast_out(k);
1966 if (cmp->outcnt() > 0 && // must have prior uses
1967 cmp->in(0) == NULL && // must be context-independent
1968 cmp->Opcode() == cmp_op) { // right kind of compare
1969 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1970 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1971 }
1972 }
1973
1974 const int NCMPS = 2;
1975 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1976 int cmpn;
1977 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1978 if (cmps[cmpn] != NULL) break; // find a result
1979 }
1980 if (cmpn < NCMPS) {
1981 // Look for a dominating test that tells us the min and max.
1982 int depth = 0; // Limit search depth for speed
1983 Node* dom = control();
1984 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1985 if (++depth >= 100) break;
1986 Node* ifproj = dom;
1987 if (!ifproj->is_Proj()) continue;
1988 Node* iff = ifproj->in(0);
1989 if (!iff->is_If()) continue;
1990 Node* bol = iff->in(1);
1991 if (!bol->is_Bool()) continue;
1992 Node* cmp = bol->in(1);
1993 if (cmp == NULL) continue;
1994 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1995 if (cmps[cmpn] == cmp) break;
1996 if (cmpn == NCMPS) continue;
1997 BoolTest::mask btest = bol->as_Bool()->_test._test;
1998 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1999 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2000 // At this point, we know that 'x btest y' is true.
2001 switch (btest) {
2002 case BoolTest::eq:
2003 // They are proven equal, so we can collapse the min/max.
2004 // Either value is the answer. Choose the simpler.
2005 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2006 return yvalue;
2007 return xvalue;
2008 case BoolTest::lt: // x < y
2009 case BoolTest::le: // x <= y
2010 return (want_max ? yvalue : xvalue);
2011 case BoolTest::gt: // x > y
2012 case BoolTest::ge: // x >= y
2013 return (want_max ? xvalue : yvalue);
2014 }
2015 }
2016 }
2017
2018 // We failed to find a dominating test.
2019 // Let's pick a test that might GVN with prior tests.
2020 Node* best_bol = NULL;
2021 BoolTest::mask best_btest = BoolTest::illegal;
2022 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2023 Node* cmp = cmps[cmpn];
2024 if (cmp == NULL) continue;
2025 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2026 Node* bol = cmp->fast_out(j);
2027 if (!bol->is_Bool()) continue;
2028 BoolTest::mask btest = bol->as_Bool()->_test._test;
2029 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2030 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2031 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2032 best_bol = bol->as_Bool();
2033 best_btest = btest;
2034 }
2035 }
2036 }
2037
2038 Node* answer_if_true = NULL;
2039 Node* answer_if_false = NULL;
2040 switch (best_btest) {
2041 default:
2042 if (cmpxy == NULL)
2043 cmpxy = ideal_cmpxy;
2044 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2045 // and fall through:
2046 case BoolTest::lt: // x < y
2047 case BoolTest::le: // x <= y
2048 answer_if_true = (want_max ? yvalue : xvalue);
2049 answer_if_false = (want_max ? xvalue : yvalue);
2050 break;
2051 case BoolTest::gt: // x > y
2052 case BoolTest::ge: // x >= y
2053 answer_if_true = (want_max ? xvalue : yvalue);
2054 answer_if_false = (want_max ? yvalue : xvalue);
2055 break;
2056 }
2057
2058 jint hi, lo;
2059 if (want_max) {
2060 // We can sharpen the minimum.
2061 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2062 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2063 } else {
2064 // We can sharpen the maximum.
2065 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2066 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2067 }
2068
2069 // Use a flow-free graph structure, to avoid creating excess control edges
2070 // which could hinder other optimizations.
2071 // Since Math.min/max is often used with arraycopy, we want
2072 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2073 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2074 answer_if_false, answer_if_true,
2075 TypeInt::make(lo, hi, widen));
2076
2077 return _gvn.transform(cmov);
2078
2079 /*
2080 // This is not as desirable as it may seem, since Min and Max
2081 // nodes do not have a full set of optimizations.
2082 // And they would interfere, anyway, with 'if' optimizations
2083 // and with CMoveI canonical forms.
2084 switch (id) {
2085 case vmIntrinsics::_min:
2086 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2087 case vmIntrinsics::_max:
2088 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2089 default:
2090 ShouldNotReachHere();
2091 }
2092 */
2093 }
2094
2095 inline int
2096 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2097 const TypePtr* base_type = TypePtr::NULL_PTR;
2098 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2099 if (base_type == NULL) {
2100 // Unknown type.
2101 return Type::AnyPtr;
2102 } else if (base_type == TypePtr::NULL_PTR) {
2103 // Since this is a NULL+long form, we have to switch to a rawptr.
2104 base = _gvn.transform(new (C) CastX2PNode(offset));
2105 offset = MakeConX(0);
2106 return Type::RawPtr;
2107 } else if (base_type->base() == Type::RawPtr) {
2108 return Type::RawPtr;
2109 } else if (base_type->isa_oopptr()) {
2110 // Base is never null => always a heap address.
2111 if (base_type->ptr() == TypePtr::NotNull) {
2112 return Type::OopPtr;
2113 }
2114 // Offset is small => always a heap address.
2115 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2116 if (offset_type != NULL &&
2117 base_type->offset() == 0 && // (should always be?)
2118 offset_type->_lo >= 0 &&
2119 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2120 return Type::OopPtr;
2121 }
2122 // Otherwise, it might either be oop+off or NULL+addr.
2123 return Type::AnyPtr;
2124 } else {
2125 // No information:
2126 return Type::AnyPtr;
2127 }
2128 }
2129
2130 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2131 int kind = classify_unsafe_addr(base, offset);
2132 if (kind == Type::RawPtr) {
2133 return basic_plus_adr(top(), base, offset);
2134 } else {
2135 return basic_plus_adr(base, offset);
2136 }
2137 }
2138
2139 //--------------------------inline_number_methods-----------------------------
2140 // inline int Integer.numberOfLeadingZeros(int)
2141 // inline int Long.numberOfLeadingZeros(long)
2142 //
2143 // inline int Integer.numberOfTrailingZeros(int)
2144 // inline int Long.numberOfTrailingZeros(long)
2145 //
2146 // inline int Integer.bitCount(int)
2147 // inline int Long.bitCount(long)
2148 //
2149 // inline char Character.reverseBytes(char)
2150 // inline short Short.reverseBytes(short)
2151 // inline int Integer.reverseBytes(int)
2152 // inline long Long.reverseBytes(long)
2153 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2154 Node* arg = argument(0);
2155 Node* n;
2156 switch (id) {
2157 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2158 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2159 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2160 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2161 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2162 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2163 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2164 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2165 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2166 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2167 default: fatal_unexpected_iid(id); break;
2168 }
2169 set_result(_gvn.transform(n));
2170 return true;
2171 }
2172
2173 //----------------------------inline_unsafe_access----------------------------
2174
2175 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2176
2177 // Helper that guards and inserts a pre-barrier.
2178 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2179 Node* pre_val, bool need_mem_bar) {
2180 // We could be accessing the referent field of a reference object. If so, when G1
2181 // is enabled, we need to log the value in the referent field in an SATB buffer.
2182 // This routine performs some compile time filters and generates suitable
2183 // runtime filters that guard the pre-barrier code.
2184 // Also add memory barrier for non volatile load from the referent field
2185 // to prevent commoning of loads across safepoint.
2186 if (!UseG1GC && !need_mem_bar)
2187 return;
2188
2189 // Some compile time checks.
2190
2191 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2192 const TypeX* otype = offset->find_intptr_t_type();
2193 if (otype != NULL && otype->is_con() &&
2194 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2195 // Constant offset but not the reference_offset so just return
2196 return;
2197 }
2198
2199 // We only need to generate the runtime guards for instances.
2200 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2201 if (btype != NULL) {
2202 if (btype->isa_aryptr()) {
2203 // Array type so nothing to do
2204 return;
2205 }
2206
2207 const TypeInstPtr* itype = btype->isa_instptr();
2208 if (itype != NULL) {
2209 // Can the klass of base_oop be statically determined to be
2210 // _not_ a sub-class of Reference and _not_ Object?
2211 ciKlass* klass = itype->klass();
2212 if ( klass->is_loaded() &&
2213 !klass->is_subtype_of(env()->Reference_klass()) &&
2214 !env()->Object_klass()->is_subtype_of(klass)) {
2215 return;
2216 }
2217 }
2218 }
2219
2220 // The compile time filters did not reject base_oop/offset so
2221 // we need to generate the following runtime filters
2222 //
2223 // if (offset == java_lang_ref_Reference::_reference_offset) {
2224 // if (instance_of(base, java.lang.ref.Reference)) {
2225 // pre_barrier(_, pre_val, ...);
2226 // }
2227 // }
2228
2229 float likely = PROB_LIKELY( 0.999);
2230 float unlikely = PROB_UNLIKELY(0.999);
2231
2232 IdealKit ideal(this);
2233 #define __ ideal.
2234
2235 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2236
2237 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2238 // Update graphKit memory and control from IdealKit.
2239 sync_kit(ideal);
2240
2241 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2242 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2243
2244 // Update IdealKit memory and control from graphKit.
2245 __ sync_kit(this);
2246
2247 Node* one = __ ConI(1);
2248 // is_instof == 0 if base_oop == NULL
2249 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2250
2251 // Update graphKit from IdeakKit.
2252 sync_kit(ideal);
2253
2254 // Use the pre-barrier to record the value in the referent field
2255 pre_barrier(false /* do_load */,
2256 __ ctrl(),
2257 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2258 pre_val /* pre_val */,
2259 T_OBJECT);
2260 if (need_mem_bar) {
2261 // Add memory barrier to prevent commoning reads from this field
2262 // across safepoint since GC can change its value.
2263 insert_mem_bar(Op_MemBarCPUOrder);
2264 }
2265 // Update IdealKit from graphKit.
2266 __ sync_kit(this);
2267
2268 } __ end_if(); // _ref_type != ref_none
2269 } __ end_if(); // offset == referent_offset
2270
2271 // Final sync IdealKit and GraphKit.
2272 final_sync(ideal);
2273 #undef __
2274 }
2275
2276
2277 // Interpret Unsafe.fieldOffset cookies correctly:
2278 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2279
2280 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2281 // Attempt to infer a sharper value type from the offset and base type.
2282 ciKlass* sharpened_klass = NULL;
2283
2284 // See if it is an instance field, with an object type.
2285 if (alias_type->field() != NULL) {
2286 assert(!is_native_ptr, "native pointer op cannot use a java address");
2287 if (alias_type->field()->type()->is_klass()) {
2288 sharpened_klass = alias_type->field()->type()->as_klass();
2289 }
2290 }
2291
2292 // See if it is a narrow oop array.
2293 if (adr_type->isa_aryptr()) {
2294 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2295 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2296 if (elem_type != NULL) {
2297 sharpened_klass = elem_type->klass();
2298 }
2299 }
2300 }
2301
2302 // The sharpened class might be unloaded if there is no class loader
2303 // contraint in place.
2304 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2305 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2306
2307 #ifndef PRODUCT
2308 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2309 tty->print(" from base type: "); adr_type->dump();
2310 tty->print(" sharpened value: "); tjp->dump();
2311 }
2312 #endif
2313 // Sharpen the value type.
2314 return tjp;
2315 }
2316 return NULL;
2317 }
2318
2319 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2320 if (callee()->is_static()) return false; // caller must have the capability!
2321
2322 #ifndef PRODUCT
2323 {
2324 ResourceMark rm;
2325 // Check the signatures.
2326 ciSignature* sig = callee()->signature();
2327 #ifdef ASSERT
2328 if (!is_store) {
2329 // Object getObject(Object base, int/long offset), etc.
2330 BasicType rtype = sig->return_type()->basic_type();
2331 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2332 rtype = T_ADDRESS; // it is really a C void*
2333 assert(rtype == type, "getter must return the expected value");
2334 if (!is_native_ptr) {
2335 assert(sig->count() == 2, "oop getter has 2 arguments");
2336 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2337 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2338 } else {
2339 assert(sig->count() == 1, "native getter has 1 argument");
2340 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2341 }
2342 } else {
2343 // void putObject(Object base, int/long offset, Object x), etc.
2344 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2345 if (!is_native_ptr) {
2346 assert(sig->count() == 3, "oop putter has 3 arguments");
2347 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2348 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2349 } else {
2350 assert(sig->count() == 2, "native putter has 2 arguments");
2351 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2352 }
2353 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2354 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2355 vtype = T_ADDRESS; // it is really a C void*
2356 assert(vtype == type, "putter must accept the expected value");
2357 }
2358 #endif // ASSERT
2359 }
2360 #endif //PRODUCT
2361
2362 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2363
2364 Node* receiver = argument(0); // type: oop
2365
2366 // Build address expression. See the code in inline_unsafe_prefetch.
2367 Node* adr;
2368 Node* heap_base_oop = top();
2369 Node* offset = top();
2370 Node* val;
2371
2372 if (!is_native_ptr) {
2373 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2374 Node* base = argument(1); // type: oop
2375 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2376 offset = argument(2); // type: long
2377 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2378 // to be plain byte offsets, which are also the same as those accepted
2379 // by oopDesc::field_base.
2380 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2381 "fieldOffset must be byte-scaled");
2382 // 32-bit machines ignore the high half!
2383 offset = ConvL2X(offset);
2384 adr = make_unsafe_address(base, offset);
2385 heap_base_oop = base;
2386 val = is_store ? argument(4) : NULL;
2387 } else {
2388 Node* ptr = argument(1); // type: long
2389 ptr = ConvL2X(ptr); // adjust Java long to machine word
2390 adr = make_unsafe_address(NULL, ptr);
2391 val = is_store ? argument(3) : NULL;
2392 }
2393
2394 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2395
2396 // First guess at the value type.
2397 const Type *value_type = Type::get_const_basic_type(type);
2398
2399 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2400 // there was not enough information to nail it down.
2401 Compile::AliasType* alias_type = C->alias_type(adr_type);
2402 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2403
2404 // We will need memory barriers unless we can determine a unique
2405 // alias category for this reference. (Note: If for some reason
2406 // the barriers get omitted and the unsafe reference begins to "pollute"
2407 // the alias analysis of the rest of the graph, either Compile::can_alias
2408 // or Compile::must_alias will throw a diagnostic assert.)
2409 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2410
2411 // If we are reading the value of the referent field of a Reference
2412 // object (either by using Unsafe directly or through reflection)
2413 // then, if G1 is enabled, we need to record the referent in an
2414 // SATB log buffer using the pre-barrier mechanism.
2415 // Also we need to add memory barrier to prevent commoning reads
2416 // from this field across safepoint since GC can change its value.
2417 bool need_read_barrier = !is_native_ptr && !is_store &&
2418 offset != top() && heap_base_oop != top();
2419
2420 if (!is_store && type == T_OBJECT) {
2421 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2422 if (tjp != NULL) {
2423 value_type = tjp;
2424 }
2425 }
2426
2427 receiver = null_check(receiver);
2428 if (stopped()) {
2429 return true;
2430 }
2431 // Heap pointers get a null-check from the interpreter,
2432 // as a courtesy. However, this is not guaranteed by Unsafe,
2433 // and it is not possible to fully distinguish unintended nulls
2434 // from intended ones in this API.
2435
2436 if (is_volatile) {
2437 // We need to emit leading and trailing CPU membars (see below) in
2438 // addition to memory membars when is_volatile. This is a little
2439 // too strong, but avoids the need to insert per-alias-type
2440 // volatile membars (for stores; compare Parse::do_put_xxx), which
2441 // we cannot do effectively here because we probably only have a
2442 // rough approximation of type.
2443 need_mem_bar = true;
2444 // For Stores, place a memory ordering barrier now.
2445 if (is_store)
2446 insert_mem_bar(Op_MemBarRelease);
2447 }
2448
2449 // Memory barrier to prevent normal and 'unsafe' accesses from
2450 // bypassing each other. Happens after null checks, so the
2451 // exception paths do not take memory state from the memory barrier,
2452 // so there's no problems making a strong assert about mixing users
2453 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2454 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2455 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2456
2457 if (!is_store) {
2458 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2459 // load value
2460 switch (type) {
2461 case T_BOOLEAN:
2462 case T_CHAR:
2463 case T_BYTE:
2464 case T_SHORT:
2465 case T_INT:
2466 case T_LONG:
2467 case T_FLOAT:
2468 case T_DOUBLE:
2469 break;
2470 case T_OBJECT:
2471 if (need_read_barrier) {
2472 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2473 }
2474 break;
2475 case T_ADDRESS:
2476 // Cast to an int type.
2477 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2478 p = ConvX2L(p);
2479 break;
2480 default:
2481 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2482 break;
2483 }
2484 // The load node has the control of the preceding MemBarCPUOrder. All
2485 // following nodes will have the control of the MemBarCPUOrder inserted at
2486 // the end of this method. So, pushing the load onto the stack at a later
2487 // point is fine.
2488 set_result(p);
2489 } else {
2490 // place effect of store into memory
2491 switch (type) {
2492 case T_DOUBLE:
2493 val = dstore_rounding(val);
2494 break;
2495 case T_ADDRESS:
2496 // Repackage the long as a pointer.
2497 val = ConvL2X(val);
2498 val = _gvn.transform(new (C) CastX2PNode(val));
2499 break;
2500 }
2501
2502 if (type != T_OBJECT ) {
2503 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2504 } else {
2505 // Possibly an oop being stored to Java heap or native memory
2506 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2507 // oop to Java heap.
2508 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2509 } else {
2510 // We can't tell at compile time if we are storing in the Java heap or outside
2511 // of it. So we need to emit code to conditionally do the proper type of
2512 // store.
2513
2514 IdealKit ideal(this);
2515 #define __ ideal.
2516 // QQQ who knows what probability is here??
2517 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2518 // Sync IdealKit and graphKit.
2519 sync_kit(ideal);
2520 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2521 // Update IdealKit memory.
2522 __ sync_kit(this);
2523 } __ else_(); {
2524 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2525 } __ end_if();
2526 // Final sync IdealKit and GraphKit.
2527 final_sync(ideal);
2528 #undef __
2529 }
2530 }
2531 }
2532
2533 if (is_volatile) {
2534 if (!is_store)
2535 insert_mem_bar(Op_MemBarAcquire);
2536 else
2537 insert_mem_bar(Op_MemBarVolatile);
2538 }
2539
2540 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2541
2542 return true;
2543 }
2544
2545 //----------------------------inline_unsafe_prefetch----------------------------
2546
2547 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2548 #ifndef PRODUCT
2549 {
2550 ResourceMark rm;
2551 // Check the signatures.
2552 ciSignature* sig = callee()->signature();
2553 #ifdef ASSERT
2554 // Object getObject(Object base, int/long offset), etc.
2555 BasicType rtype = sig->return_type()->basic_type();
2556 if (!is_native_ptr) {
2557 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2558 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2559 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2560 } else {
2561 assert(sig->count() == 1, "native prefetch has 1 argument");
2562 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2563 }
2564 #endif // ASSERT
2565 }
2566 #endif // !PRODUCT
2567
2568 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2569
2570 const int idx = is_static ? 0 : 1;
2571 if (!is_static) {
2572 null_check_receiver();
2573 if (stopped()) {
2574 return true;
2575 }
2576 }
2577
2578 // Build address expression. See the code in inline_unsafe_access.
2579 Node *adr;
2580 if (!is_native_ptr) {
2581 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2582 Node* base = argument(idx + 0); // type: oop
2583 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2584 Node* offset = argument(idx + 1); // type: long
2585 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2586 // to be plain byte offsets, which are also the same as those accepted
2587 // by oopDesc::field_base.
2588 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2589 "fieldOffset must be byte-scaled");
2590 // 32-bit machines ignore the high half!
2591 offset = ConvL2X(offset);
2592 adr = make_unsafe_address(base, offset);
2593 } else {
2594 Node* ptr = argument(idx + 0); // type: long
2595 ptr = ConvL2X(ptr); // adjust Java long to machine word
2596 adr = make_unsafe_address(NULL, ptr);
2597 }
2598
2599 // Generate the read or write prefetch
2600 Node *prefetch;
2601 if (is_store) {
2602 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2603 } else {
2604 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2605 }
2606 prefetch->init_req(0, control());
2607 set_i_o(_gvn.transform(prefetch));
2608
2609 return true;
2610 }
2611
2612 //----------------------------inline_unsafe_load_store----------------------------
2613 // This method serves a couple of different customers (depending on LoadStoreKind):
2614 //
2615 // LS_cmpxchg:
2616 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2617 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2618 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2619 //
2620 // LS_xadd:
2621 // public int getAndAddInt( Object o, long offset, int delta)
2622 // public long getAndAddLong(Object o, long offset, long delta)
2623 //
2624 // LS_xchg:
2625 // int getAndSet(Object o, long offset, int newValue)
2626 // long getAndSet(Object o, long offset, long newValue)
2627 // Object getAndSet(Object o, long offset, Object newValue)
2628 //
2629 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2630 // This basic scheme here is the same as inline_unsafe_access, but
2631 // differs in enough details that combining them would make the code
2632 // overly confusing. (This is a true fact! I originally combined
2633 // them, but even I was confused by it!) As much code/comments as
2634 // possible are retained from inline_unsafe_access though to make
2635 // the correspondences clearer. - dl
2636
2637 if (callee()->is_static()) return false; // caller must have the capability!
2638
2639 #ifndef PRODUCT
2640 BasicType rtype;
2641 {
2642 ResourceMark rm;
2643 // Check the signatures.
2644 ciSignature* sig = callee()->signature();
2645 rtype = sig->return_type()->basic_type();
2646 if (kind == LS_xadd || kind == LS_xchg) {
2647 // Check the signatures.
2648 #ifdef ASSERT
2649 assert(rtype == type, "get and set must return the expected type");
2650 assert(sig->count() == 3, "get and set has 3 arguments");
2651 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2652 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2653 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2654 #endif // ASSERT
2655 } else if (kind == LS_cmpxchg) {
2656 // Check the signatures.
2657 #ifdef ASSERT
2658 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2659 assert(sig->count() == 4, "CAS has 4 arguments");
2660 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2661 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2662 #endif // ASSERT
2663 } else {
2664 ShouldNotReachHere();
2665 }
2666 }
2667 #endif //PRODUCT
2668
2669 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2670
2671 // Get arguments:
2672 Node* receiver = NULL;
2673 Node* base = NULL;
2674 Node* offset = NULL;
2675 Node* oldval = NULL;
2676 Node* newval = NULL;
2677 if (kind == LS_cmpxchg) {
2678 const bool two_slot_type = type2size[type] == 2;
2679 receiver = argument(0); // type: oop
2680 base = argument(1); // type: oop
2681 offset = argument(2); // type: long
2682 oldval = argument(4); // type: oop, int, or long
2683 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2684 } else if (kind == LS_xadd || kind == LS_xchg){
2685 receiver = argument(0); // type: oop
2686 base = argument(1); // type: oop
2687 offset = argument(2); // type: long
2688 oldval = NULL;
2689 newval = argument(4); // type: oop, int, or long
2690 }
2691
2692 // Null check receiver.
2693 receiver = null_check(receiver);
2694 if (stopped()) {
2695 return true;
2696 }
2697
2698 // Build field offset expression.
2699 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2700 // to be plain byte offsets, which are also the same as those accepted
2701 // by oopDesc::field_base.
2702 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2703 // 32-bit machines ignore the high half of long offsets
2704 offset = ConvL2X(offset);
2705 Node* adr = make_unsafe_address(base, offset);
2706 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2707
2708 // For CAS, unlike inline_unsafe_access, there seems no point in
2709 // trying to refine types. Just use the coarse types here.
2710 const Type *value_type = Type::get_const_basic_type(type);
2711 Compile::AliasType* alias_type = C->alias_type(adr_type);
2712 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2713
2714 if (kind == LS_xchg && type == T_OBJECT) {
2715 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2716 if (tjp != NULL) {
2717 value_type = tjp;
2718 }
2719 }
2720
2721 int alias_idx = C->get_alias_index(adr_type);
2722
2723 // Memory-model-wise, a LoadStore acts like a little synchronized
2724 // block, so needs barriers on each side. These don't translate
2725 // into actual barriers on most machines, but we still need rest of
2726 // compiler to respect ordering.
2727
2728 insert_mem_bar(Op_MemBarRelease);
2729 insert_mem_bar(Op_MemBarCPUOrder);
2730
2731 // 4984716: MemBars must be inserted before this
2732 // memory node in order to avoid a false
2733 // dependency which will confuse the scheduler.
2734 Node *mem = memory(alias_idx);
2735
2736 // For now, we handle only those cases that actually exist: ints,
2737 // longs, and Object. Adding others should be straightforward.
2738 Node* load_store;
2739 switch(type) {
2740 case T_INT:
2741 if (kind == LS_xadd) {
2742 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2743 } else if (kind == LS_xchg) {
2744 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2745 } else if (kind == LS_cmpxchg) {
2746 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2747 } else {
2748 ShouldNotReachHere();
2749 }
2750 break;
2751 case T_LONG:
2752 if (kind == LS_xadd) {
2753 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2754 } else if (kind == LS_xchg) {
2755 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2756 } else if (kind == LS_cmpxchg) {
2757 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2758 } else {
2759 ShouldNotReachHere();
2760 }
2761 break;
2762 case T_OBJECT:
2763 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2764 // could be delayed during Parse (for example, in adjust_map_after_if()).
2765 // Execute transformation here to avoid barrier generation in such case.
2766 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2767 newval = _gvn.makecon(TypePtr::NULL_PTR);
2768
2769 // Reference stores need a store barrier.
2770 if (kind == LS_xchg) {
2771 // If pre-barrier must execute before the oop store, old value will require do_load here.
2772 if (!can_move_pre_barrier()) {
2773 pre_barrier(true /* do_load*/,
2774 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2775 NULL /* pre_val*/,
2776 T_OBJECT);
2777 } // Else move pre_barrier to use load_store value, see below.
2778 } else if (kind == LS_cmpxchg) {
2779 // Same as for newval above:
2780 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2781 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2782 }
2783 // The only known value which might get overwritten is oldval.
2784 pre_barrier(false /* do_load */,
2785 control(), NULL, NULL, max_juint, NULL, NULL,
2786 oldval /* pre_val */,
2787 T_OBJECT);
2788 } else {
2789 ShouldNotReachHere();
2790 }
2791
2792 #ifdef _LP64
2793 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2794 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2795 if (kind == LS_xchg) {
2796 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2797 newval_enc, adr_type, value_type->make_narrowoop()));
2798 } else {
2799 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2800 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2801 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2802 newval_enc, oldval_enc));
2803 }
2804 } else
2805 #endif
2806 {
2807 if (kind == LS_xchg) {
2808 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2809 } else {
2810 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2811 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2812 }
2813 }
2814 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2815 break;
2816 default:
2817 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2818 break;
2819 }
2820
2821 // SCMemProjNodes represent the memory state of a LoadStore. Their
2822 // main role is to prevent LoadStore nodes from being optimized away
2823 // when their results aren't used.
2824 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2825 set_memory(proj, alias_idx);
2826
2827 if (type == T_OBJECT && kind == LS_xchg) {
2828 #ifdef _LP64
2829 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2830 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2831 }
2832 #endif
2833 if (can_move_pre_barrier()) {
2834 // Don't need to load pre_val. The old value is returned by load_store.
2835 // The pre_barrier can execute after the xchg as long as no safepoint
2836 // gets inserted between them.
2837 pre_barrier(false /* do_load */,
2838 control(), NULL, NULL, max_juint, NULL, NULL,
2839 load_store /* pre_val */,
2840 T_OBJECT);
2841 }
2842 }
2843
2844 // Add the trailing membar surrounding the access
2845 insert_mem_bar(Op_MemBarCPUOrder);
2846 insert_mem_bar(Op_MemBarAcquire);
2847
2848 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2849 set_result(load_store);
2850 return true;
2851 }
2852
2853 //----------------------------inline_unsafe_ordered_store----------------------
2854 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2855 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2856 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2857 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2858 // This is another variant of inline_unsafe_access, differing in
2859 // that it always issues store-store ("release") barrier and ensures
2860 // store-atomicity (which only matters for "long").
2861
2862 if (callee()->is_static()) return false; // caller must have the capability!
2863
2864 #ifndef PRODUCT
2865 {
2866 ResourceMark rm;
2867 // Check the signatures.
2868 ciSignature* sig = callee()->signature();
2869 #ifdef ASSERT
2870 BasicType rtype = sig->return_type()->basic_type();
2871 assert(rtype == T_VOID, "must return void");
2872 assert(sig->count() == 3, "has 3 arguments");
2873 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2874 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2875 #endif // ASSERT
2876 }
2877 #endif //PRODUCT
2878
2879 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2880
2881 // Get arguments:
2882 Node* receiver = argument(0); // type: oop
2883 Node* base = argument(1); // type: oop
2884 Node* offset = argument(2); // type: long
2885 Node* val = argument(4); // type: oop, int, or long
2886
2887 // Null check receiver.
2888 receiver = null_check(receiver);
2889 if (stopped()) {
2890 return true;
2891 }
2892
2893 // Build field offset expression.
2894 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2895 // 32-bit machines ignore the high half of long offsets
2896 offset = ConvL2X(offset);
2897 Node* adr = make_unsafe_address(base, offset);
2898 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2899 const Type *value_type = Type::get_const_basic_type(type);
2900 Compile::AliasType* alias_type = C->alias_type(adr_type);
2901
2902 insert_mem_bar(Op_MemBarRelease);
2903 insert_mem_bar(Op_MemBarCPUOrder);
2904 // Ensure that the store is atomic for longs:
2905 const bool require_atomic_access = true;
2906 Node* store;
2907 if (type == T_OBJECT) // reference stores need a store barrier.
2908 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2909 else {
2910 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2911 }
2912 insert_mem_bar(Op_MemBarCPUOrder);
2913 return true;
2914 }
2915
2916 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2917 // Regardless of form, don't allow previous ld/st to move down,
2918 // then issue acquire, release, or volatile mem_bar.
2919 insert_mem_bar(Op_MemBarCPUOrder);
2920 switch(id) {
2921 case vmIntrinsics::_loadFence:
2922 insert_mem_bar(Op_MemBarAcquire);
2923 return true;
2924 case vmIntrinsics::_storeFence:
2925 insert_mem_bar(Op_MemBarRelease);
2926 return true;
2927 case vmIntrinsics::_fullFence:
2928 insert_mem_bar(Op_MemBarVolatile);
2929 return true;
2930 default:
2931 fatal_unexpected_iid(id);
2932 return false;
2933 }
2934 }
2935
2936 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
2937 if (!kls->is_Con()) {
2938 return true;
2939 }
2940 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
2941 if (klsptr == NULL) {
2942 return true;
2943 }
2944 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
2945 // don't need a guard for a klass that is already initialized
2946 return !ik->is_initialized();
2947 }
2948
2949 //----------------------------inline_unsafe_allocate---------------------------
2950 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
2951 bool LibraryCallKit::inline_unsafe_allocate() {
2952 if (callee()->is_static()) return false; // caller must have the capability!
2953
2954 null_check_receiver(); // null-check, then ignore
2955 Node* cls = null_check(argument(1));
2956 if (stopped()) return true;
2957
2958 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2959 kls = null_check(kls);
2960 if (stopped()) return true; // argument was like int.class
2961
2962 Node* test = NULL;
2963 if (LibraryCallKit::klass_needs_init_guard(kls)) {
2964 // Note: The argument might still be an illegal value like
2965 // Serializable.class or Object[].class. The runtime will handle it.
2966 // But we must make an explicit check for initialization.
2967 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2968 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2969 // can generate code to load it as unsigned byte.
2970 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2971 Node* bits = intcon(InstanceKlass::fully_initialized);
2972 test = _gvn.transform(new (C) SubINode(inst, bits));
2973 // The 'test' is non-zero if we need to take a slow path.
2974 }
2975
2976 Node* obj = new_instance(kls, test);
2977 set_result(obj);
2978 return true;
2979 }
2980
2981 #ifdef TRACE_HAVE_INTRINSICS
2982 /*
2983 * oop -> myklass
2984 * myklass->trace_id |= USED
2985 * return myklass->trace_id & ~0x3
2986 */
2987 bool LibraryCallKit::inline_native_classID() {
2988 null_check_receiver(); // null-check, then ignore
2989 Node* cls = null_check(argument(1), T_OBJECT);
2990 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2991 kls = null_check(kls, T_OBJECT);
2992 ByteSize offset = TRACE_ID_OFFSET;
2993 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2994 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2995 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2996 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2997 Node* clsused = longcon(0x01l); // set the class bit
2998 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2999
3000 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3001 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3002 set_result(andl);
3003 return true;
3004 }
3005
3006 bool LibraryCallKit::inline_native_threadID() {
3007 Node* tls_ptr = NULL;
3008 Node* cur_thr = generate_current_thread(tls_ptr);
3009 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3010 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3011 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3012
3013 Node* threadid = NULL;
3014 size_t thread_id_size = OSThread::thread_id_size();
3015 if (thread_id_size == (size_t) BytesPerLong) {
3016 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3017 } else if (thread_id_size == (size_t) BytesPerInt) {
3018 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3019 } else {
3020 ShouldNotReachHere();
3021 }
3022 set_result(threadid);
3023 return true;
3024 }
3025 #endif
3026
3027 //------------------------inline_native_time_funcs--------------
3028 // inline code for System.currentTimeMillis() and System.nanoTime()
3029 // these have the same type and signature
3030 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3031 const TypeFunc* tf = OptoRuntime::void_long_Type();
3032 const TypePtr* no_memory_effects = NULL;
3033 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3034 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3035 #ifdef ASSERT
3036 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3037 assert(value_top == top(), "second value must be top");
3038 #endif
3039 set_result(value);
3040 return true;
3041 }
3042
3043 //------------------------inline_native_currentThread------------------
3044 bool LibraryCallKit::inline_native_currentThread() {
3045 Node* junk = NULL;
3046 set_result(generate_current_thread(junk));
3047 return true;
3048 }
3049
3050 //------------------------inline_native_isInterrupted------------------
3051 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3052 bool LibraryCallKit::inline_native_isInterrupted() {
3053 // Add a fast path to t.isInterrupted(clear_int):
3054 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3055 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3056 // So, in the common case that the interrupt bit is false,
3057 // we avoid making a call into the VM. Even if the interrupt bit
3058 // is true, if the clear_int argument is false, we avoid the VM call.
3059 // However, if the receiver is not currentThread, we must call the VM,
3060 // because there must be some locking done around the operation.
3061
3062 // We only go to the fast case code if we pass two guards.
3063 // Paths which do not pass are accumulated in the slow_region.
3064
3065 enum {
3066 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3067 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3068 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3069 PATH_LIMIT
3070 };
3071
3072 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3073 // out of the function.
3074 insert_mem_bar(Op_MemBarCPUOrder);
3075
3076 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3077 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3078
3079 RegionNode* slow_region = new (C) RegionNode(1);
3080 record_for_igvn(slow_region);
3081
3082 // (a) Receiving thread must be the current thread.
3083 Node* rec_thr = argument(0);
3084 Node* tls_ptr = NULL;
3085 Node* cur_thr = generate_current_thread(tls_ptr);
3086 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3087 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3088
3089 generate_slow_guard(bol_thr, slow_region);
3090
3091 // (b) Interrupt bit on TLS must be false.
3092 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3093 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3094 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3095
3096 // Set the control input on the field _interrupted read to prevent it floating up.
3097 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3098 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3099 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3100
3101 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3102
3103 // First fast path: if (!TLS._interrupted) return false;
3104 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3105 result_rgn->init_req(no_int_result_path, false_bit);
3106 result_val->init_req(no_int_result_path, intcon(0));
3107
3108 // drop through to next case
3109 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3110
3111 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3112 Node* clr_arg = argument(1);
3113 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3114 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3115 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3116
3117 // Second fast path: ... else if (!clear_int) return true;
3118 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3119 result_rgn->init_req(no_clear_result_path, false_arg);
3120 result_val->init_req(no_clear_result_path, intcon(1));
3121
3122 // drop through to next case
3123 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3124
3125 // (d) Otherwise, go to the slow path.
3126 slow_region->add_req(control());
3127 set_control( _gvn.transform(slow_region));
3128
3129 if (stopped()) {
3130 // There is no slow path.
3131 result_rgn->init_req(slow_result_path, top());
3132 result_val->init_req(slow_result_path, top());
3133 } else {
3134 // non-virtual because it is a private non-static
3135 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3136
3137 Node* slow_val = set_results_for_java_call(slow_call);
3138 // this->control() comes from set_results_for_java_call
3139
3140 Node* fast_io = slow_call->in(TypeFunc::I_O);
3141 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3142
3143 // These two phis are pre-filled with copies of of the fast IO and Memory
3144 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3145 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3146
3147 result_rgn->init_req(slow_result_path, control());
3148 result_io ->init_req(slow_result_path, i_o());
3149 result_mem->init_req(slow_result_path, reset_memory());
3150 result_val->init_req(slow_result_path, slow_val);
3151
3152 set_all_memory(_gvn.transform(result_mem));
3153 set_i_o( _gvn.transform(result_io));
3154 }
3155
3156 C->set_has_split_ifs(true); // Has chance for split-if optimization
3157 set_result(result_rgn, result_val);
3158 return true;
3159 }
3160
3161 //---------------------------load_mirror_from_klass----------------------------
3162 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3163 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3164 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3165 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3166 }
3167
3168 //-----------------------load_klass_from_mirror_common-------------------------
3169 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3170 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3171 // and branch to the given path on the region.
3172 // If never_see_null, take an uncommon trap on null, so we can optimistically
3173 // compile for the non-null case.
3174 // If the region is NULL, force never_see_null = true.
3175 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3176 bool never_see_null,
3177 RegionNode* region,
3178 int null_path,
3179 int offset) {
3180 if (region == NULL) never_see_null = true;
3181 Node* p = basic_plus_adr(mirror, offset);
3182 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3183 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3184 Node* null_ctl = top();
3185 kls = null_check_oop(kls, &null_ctl, never_see_null);
3186 if (region != NULL) {
3187 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3188 region->init_req(null_path, null_ctl);
3189 } else {
3190 assert(null_ctl == top(), "no loose ends");
3191 }
3192 return kls;
3193 }
3194
3195 //--------------------(inline_native_Class_query helpers)---------------------
3196 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3197 // Fall through if (mods & mask) == bits, take the guard otherwise.
3198 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3199 // Branch around if the given klass has the given modifier bit set.
3200 // Like generate_guard, adds a new path onto the region.
3201 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3202 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3203 Node* mask = intcon(modifier_mask);
3204 Node* bits = intcon(modifier_bits);
3205 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3206 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3207 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3208 return generate_fair_guard(bol, region);
3209 }
3210 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3211 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3212 }
3213
3214 //-------------------------inline_native_Class_query-------------------
3215 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3216 const Type* return_type = TypeInt::BOOL;
3217 Node* prim_return_value = top(); // what happens if it's a primitive class?
3218 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3219 bool expect_prim = false; // most of these guys expect to work on refs
3220
3221 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3222
3223 Node* mirror = argument(0);
3224 Node* obj = top();
3225
3226 switch (id) {
3227 case vmIntrinsics::_isInstance:
3228 // nothing is an instance of a primitive type
3229 prim_return_value = intcon(0);
3230 obj = argument(1);
3231 break;
3232 case vmIntrinsics::_getModifiers:
3233 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3234 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3235 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3236 break;
3237 case vmIntrinsics::_isInterface:
3238 prim_return_value = intcon(0);
3239 break;
3240 case vmIntrinsics::_isArray:
3241 prim_return_value = intcon(0);
3242 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3243 break;
3244 case vmIntrinsics::_isPrimitive:
3245 prim_return_value = intcon(1);
3246 expect_prim = true; // obviously
3247 break;
3248 case vmIntrinsics::_getSuperclass:
3249 prim_return_value = null();
3250 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3251 break;
3252 case vmIntrinsics::_getComponentType:
3253 prim_return_value = null();
3254 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3255 break;
3256 case vmIntrinsics::_getClassAccessFlags:
3257 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3258 return_type = TypeInt::INT; // not bool! 6297094
3259 break;
3260 default:
3261 fatal_unexpected_iid(id);
3262 break;
3263 }
3264
3265 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3266 if (mirror_con == NULL) return false; // cannot happen?
3267
3268 #ifndef PRODUCT
3269 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3270 ciType* k = mirror_con->java_mirror_type();
3271 if (k) {
3272 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3273 k->print_name();
3274 tty->cr();
3275 }
3276 }
3277 #endif
3278
3279 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3280 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3281 record_for_igvn(region);
3282 PhiNode* phi = new (C) PhiNode(region, return_type);
3283
3284 // The mirror will never be null of Reflection.getClassAccessFlags, however
3285 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3286 // if it is. See bug 4774291.
3287
3288 // For Reflection.getClassAccessFlags(), the null check occurs in
3289 // the wrong place; see inline_unsafe_access(), above, for a similar
3290 // situation.
3291 mirror = null_check(mirror);
3292 // If mirror or obj is dead, only null-path is taken.
3293 if (stopped()) return true;
3294
3295 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3296
3297 // Now load the mirror's klass metaobject, and null-check it.
3298 // Side-effects region with the control path if the klass is null.
3299 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3300 // If kls is null, we have a primitive mirror.
3301 phi->init_req(_prim_path, prim_return_value);
3302 if (stopped()) { set_result(region, phi); return true; }
3303
3304 Node* p; // handy temp
3305 Node* null_ctl;
3306
3307 // Now that we have the non-null klass, we can perform the real query.
3308 // For constant classes, the query will constant-fold in LoadNode::Value.
3309 Node* query_value = top();
3310 switch (id) {
3311 case vmIntrinsics::_isInstance:
3312 // nothing is an instance of a primitive type
3313 query_value = gen_instanceof(obj, kls);
3314 break;
3315
3316 case vmIntrinsics::_getModifiers:
3317 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3318 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3319 break;
3320
3321 case vmIntrinsics::_isInterface:
3322 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3323 if (generate_interface_guard(kls, region) != NULL)
3324 // A guard was added. If the guard is taken, it was an interface.
3325 phi->add_req(intcon(1));
3326 // If we fall through, it's a plain class.
3327 query_value = intcon(0);
3328 break;
3329
3330 case vmIntrinsics::_isArray:
3331 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3332 if (generate_array_guard(kls, region) != NULL)
3333 // A guard was added. If the guard is taken, it was an array.
3334 phi->add_req(intcon(1));
3335 // If we fall through, it's a plain class.
3336 query_value = intcon(0);
3337 break;
3338
3339 case vmIntrinsics::_isPrimitive:
3340 query_value = intcon(0); // "normal" path produces false
3341 break;
3342
3343 case vmIntrinsics::_getSuperclass:
3344 // The rules here are somewhat unfortunate, but we can still do better
3345 // with random logic than with a JNI call.
3346 // Interfaces store null or Object as _super, but must report null.
3347 // Arrays store an intermediate super as _super, but must report Object.
3348 // Other types can report the actual _super.
3349 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3350 if (generate_interface_guard(kls, region) != NULL)
3351 // A guard was added. If the guard is taken, it was an interface.
3352 phi->add_req(null());
3353 if (generate_array_guard(kls, region) != NULL)
3354 // A guard was added. If the guard is taken, it was an array.
3355 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3356 // If we fall through, it's a plain class. Get its _super.
3357 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3358 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3359 null_ctl = top();
3360 kls = null_check_oop(kls, &null_ctl);
3361 if (null_ctl != top()) {
3362 // If the guard is taken, Object.superClass is null (both klass and mirror).
3363 region->add_req(null_ctl);
3364 phi ->add_req(null());
3365 }
3366 if (!stopped()) {
3367 query_value = load_mirror_from_klass(kls);
3368 }
3369 break;
3370
3371 case vmIntrinsics::_getComponentType:
3372 if (generate_array_guard(kls, region) != NULL) {
3373 // Be sure to pin the oop load to the guard edge just created:
3374 Node* is_array_ctrl = region->in(region->req()-1);
3375 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3376 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3377 phi->add_req(cmo);
3378 }
3379 query_value = null(); // non-array case is null
3380 break;
3381
3382 case vmIntrinsics::_getClassAccessFlags:
3383 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3384 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3385 break;
3386
3387 default:
3388 fatal_unexpected_iid(id);
3389 break;
3390 }
3391
3392 // Fall-through is the normal case of a query to a real class.
3393 phi->init_req(1, query_value);
3394 region->init_req(1, control());
3395
3396 C->set_has_split_ifs(true); // Has chance for split-if optimization
3397 set_result(region, phi);
3398 return true;
3399 }
3400
3401 //--------------------------inline_native_subtype_check------------------------
3402 // This intrinsic takes the JNI calls out of the heart of
3403 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3404 bool LibraryCallKit::inline_native_subtype_check() {
3405 // Pull both arguments off the stack.
3406 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3407 args[0] = argument(0);
3408 args[1] = argument(1);
3409 Node* klasses[2]; // corresponding Klasses: superk, subk
3410 klasses[0] = klasses[1] = top();
3411
3412 enum {
3413 // A full decision tree on {superc is prim, subc is prim}:
3414 _prim_0_path = 1, // {P,N} => false
3415 // {P,P} & superc!=subc => false
3416 _prim_same_path, // {P,P} & superc==subc => true
3417 _prim_1_path, // {N,P} => false
3418 _ref_subtype_path, // {N,N} & subtype check wins => true
3419 _both_ref_path, // {N,N} & subtype check loses => false
3420 PATH_LIMIT
3421 };
3422
3423 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3424 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3425 record_for_igvn(region);
3426
3427 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3428 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3429 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3430
3431 // First null-check both mirrors and load each mirror's klass metaobject.
3432 int which_arg;
3433 for (which_arg = 0; which_arg <= 1; which_arg++) {
3434 Node* arg = args[which_arg];
3435 arg = null_check(arg);
3436 if (stopped()) break;
3437 args[which_arg] = arg;
3438
3439 Node* p = basic_plus_adr(arg, class_klass_offset);
3440 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3441 klasses[which_arg] = _gvn.transform(kls);
3442 }
3443
3444 // Having loaded both klasses, test each for null.
3445 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3446 for (which_arg = 0; which_arg <= 1; which_arg++) {
3447 Node* kls = klasses[which_arg];
3448 Node* null_ctl = top();
3449 kls = null_check_oop(kls, &null_ctl, never_see_null);
3450 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3451 region->init_req(prim_path, null_ctl);
3452 if (stopped()) break;
3453 klasses[which_arg] = kls;
3454 }
3455
3456 if (!stopped()) {
3457 // now we have two reference types, in klasses[0..1]
3458 Node* subk = klasses[1]; // the argument to isAssignableFrom
3459 Node* superk = klasses[0]; // the receiver
3460 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3461 // now we have a successful reference subtype check
3462 region->set_req(_ref_subtype_path, control());
3463 }
3464
3465 // If both operands are primitive (both klasses null), then
3466 // we must return true when they are identical primitives.
3467 // It is convenient to test this after the first null klass check.
3468 set_control(region->in(_prim_0_path)); // go back to first null check
3469 if (!stopped()) {
3470 // Since superc is primitive, make a guard for the superc==subc case.
3471 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3472 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3473 generate_guard(bol_eq, region, PROB_FAIR);
3474 if (region->req() == PATH_LIMIT+1) {
3475 // A guard was added. If the added guard is taken, superc==subc.
3476 region->swap_edges(PATH_LIMIT, _prim_same_path);
3477 region->del_req(PATH_LIMIT);
3478 }
3479 region->set_req(_prim_0_path, control()); // Not equal after all.
3480 }
3481
3482 // these are the only paths that produce 'true':
3483 phi->set_req(_prim_same_path, intcon(1));
3484 phi->set_req(_ref_subtype_path, intcon(1));
3485
3486 // pull together the cases:
3487 assert(region->req() == PATH_LIMIT, "sane region");
3488 for (uint i = 1; i < region->req(); i++) {
3489 Node* ctl = region->in(i);
3490 if (ctl == NULL || ctl == top()) {
3491 region->set_req(i, top());
3492 phi ->set_req(i, top());
3493 } else if (phi->in(i) == NULL) {
3494 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3495 }
3496 }
3497
3498 set_control(_gvn.transform(region));
3499 set_result(_gvn.transform(phi));
3500 return true;
3501 }
3502
3503 //---------------------generate_array_guard_common------------------------
3504 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3505 bool obj_array, bool not_array) {
3506 // If obj_array/non_array==false/false:
3507 // Branch around if the given klass is in fact an array (either obj or prim).
3508 // If obj_array/non_array==false/true:
3509 // Branch around if the given klass is not an array klass of any kind.
3510 // If obj_array/non_array==true/true:
3511 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3512 // If obj_array/non_array==true/false:
3513 // Branch around if the kls is an oop array (Object[] or subtype)
3514 //
3515 // Like generate_guard, adds a new path onto the region.
3516 jint layout_con = 0;
3517 Node* layout_val = get_layout_helper(kls, layout_con);
3518 if (layout_val == NULL) {
3519 bool query = (obj_array
3520 ? Klass::layout_helper_is_objArray(layout_con)
3521 : Klass::layout_helper_is_array(layout_con));
3522 if (query == not_array) {
3523 return NULL; // never a branch
3524 } else { // always a branch
3525 Node* always_branch = control();
3526 if (region != NULL)
3527 region->add_req(always_branch);
3528 set_control(top());
3529 return always_branch;
3530 }
3531 }
3532 // Now test the correct condition.
3533 jint nval = (obj_array
3534 ? ((jint)Klass::_lh_array_tag_type_value
3535 << Klass::_lh_array_tag_shift)
3536 : Klass::_lh_neutral_value);
3537 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3538 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3539 // invert the test if we are looking for a non-array
3540 if (not_array) btest = BoolTest(btest).negate();
3541 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3542 return generate_fair_guard(bol, region);
3543 }
3544
3545
3546 //-----------------------inline_native_newArray--------------------------
3547 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3548 bool LibraryCallKit::inline_native_newArray() {
3549 Node* mirror = argument(0);
3550 Node* count_val = argument(1);
3551
3552 mirror = null_check(mirror);
3553 // If mirror or obj is dead, only null-path is taken.
3554 if (stopped()) return true;
3555
3556 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3557 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3558 PhiNode* result_val = new(C) PhiNode(result_reg,
3559 TypeInstPtr::NOTNULL);
3560 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3561 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3562 TypePtr::BOTTOM);
3563
3564 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3565 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3566 result_reg, _slow_path);
3567 Node* normal_ctl = control();
3568 Node* no_array_ctl = result_reg->in(_slow_path);
3569
3570 // Generate code for the slow case. We make a call to newArray().
3571 set_control(no_array_ctl);
3572 if (!stopped()) {
3573 // Either the input type is void.class, or else the
3574 // array klass has not yet been cached. Either the
3575 // ensuing call will throw an exception, or else it
3576 // will cache the array klass for next time.
3577 PreserveJVMState pjvms(this);
3578 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3579 Node* slow_result = set_results_for_java_call(slow_call);
3580 // this->control() comes from set_results_for_java_call
3581 result_reg->set_req(_slow_path, control());
3582 result_val->set_req(_slow_path, slow_result);
3583 result_io ->set_req(_slow_path, i_o());
3584 result_mem->set_req(_slow_path, reset_memory());
3585 }
3586
3587 set_control(normal_ctl);
3588 if (!stopped()) {
3589 // Normal case: The array type has been cached in the java.lang.Class.
3590 // The following call works fine even if the array type is polymorphic.
3591 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3592 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3593 result_reg->init_req(_normal_path, control());
3594 result_val->init_req(_normal_path, obj);
3595 result_io ->init_req(_normal_path, i_o());
3596 result_mem->init_req(_normal_path, reset_memory());
3597 }
3598
3599 // Return the combined state.
3600 set_i_o( _gvn.transform(result_io) );
3601 set_all_memory( _gvn.transform(result_mem));
3602
3603 C->set_has_split_ifs(true); // Has chance for split-if optimization
3604 set_result(result_reg, result_val);
3605 return true;
3606 }
3607
3608 //----------------------inline_native_getLength--------------------------
3609 // public static native int java.lang.reflect.Array.getLength(Object array);
3610 bool LibraryCallKit::inline_native_getLength() {
3611 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3612
3613 Node* array = null_check(argument(0));
3614 // If array is dead, only null-path is taken.
3615 if (stopped()) return true;
3616
3617 // Deoptimize if it is a non-array.
3618 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3619
3620 if (non_array != NULL) {
3621 PreserveJVMState pjvms(this);
3622 set_control(non_array);
3623 uncommon_trap(Deoptimization::Reason_intrinsic,
3624 Deoptimization::Action_maybe_recompile);
3625 }
3626
3627 // If control is dead, only non-array-path is taken.
3628 if (stopped()) return true;
3629
3630 // The works fine even if the array type is polymorphic.
3631 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3632 Node* result = load_array_length(array);
3633
3634 C->set_has_split_ifs(true); // Has chance for split-if optimization
3635 set_result(result);
3636 return true;
3637 }
3638
3639 //------------------------inline_array_copyOf----------------------------
3640 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3641 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3642 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3643 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3644
3645 // Get the arguments.
3646 Node* original = argument(0);
3647 Node* start = is_copyOfRange? argument(1): intcon(0);
3648 Node* end = is_copyOfRange? argument(2): argument(1);
3649 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3650
3651 Node* newcopy;
3652
3653 // Set the original stack and the reexecute bit for the interpreter to reexecute
3654 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3655 { PreserveReexecuteState preexecs(this);
3656 jvms()->set_should_reexecute(true);
3657
3658 array_type_mirror = null_check(array_type_mirror);
3659 original = null_check(original);
3660
3661 // Check if a null path was taken unconditionally.
3662 if (stopped()) return true;
3663
3664 Node* orig_length = load_array_length(original);
3665
3666 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3667 klass_node = null_check(klass_node);
3668
3669 RegionNode* bailout = new (C) RegionNode(1);
3670 record_for_igvn(bailout);
3671
3672 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3673 // Bail out if that is so.
3674 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3675 if (not_objArray != NULL) {
3676 // Improve the klass node's type from the new optimistic assumption:
3677 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3678 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3679 Node* cast = new (C) CastPPNode(klass_node, akls);
3680 cast->init_req(0, control());
3681 klass_node = _gvn.transform(cast);
3682 }
3683
3684 // Bail out if either start or end is negative.
3685 generate_negative_guard(start, bailout, &start);
3686 generate_negative_guard(end, bailout, &end);
3687
3688 Node* length = end;
3689 if (_gvn.type(start) != TypeInt::ZERO) {
3690 length = _gvn.transform(new (C) SubINode(end, start));
3691 }
3692
3693 // Bail out if length is negative.
3694 // Without this the new_array would throw
3695 // NegativeArraySizeException but IllegalArgumentException is what
3696 // should be thrown
3697 generate_negative_guard(length, bailout, &length);
3698
3699 if (bailout->req() > 1) {
3700 PreserveJVMState pjvms(this);
3701 set_control(_gvn.transform(bailout));
3702 uncommon_trap(Deoptimization::Reason_intrinsic,
3703 Deoptimization::Action_maybe_recompile);
3704 }
3705
3706 if (!stopped()) {
3707 // How many elements will we copy from the original?
3708 // The answer is MinI(orig_length - start, length).
3709 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3710 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3711
3712 newcopy = new_array(klass_node, length, 0); // no argments to push
3713
3714 // Generate a direct call to the right arraycopy function(s).
3715 // We know the copy is disjoint but we might not know if the
3716 // oop stores need checking.
3717 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3718 // This will fail a store-check if x contains any non-nulls.
3719 bool disjoint_bases = true;
3720 // if start > orig_length then the length of the copy may be
3721 // negative.
3722 bool length_never_negative = !is_copyOfRange;
3723 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3724 original, start, newcopy, intcon(0), moved,
3725 disjoint_bases, length_never_negative);
3726 }
3727 } // original reexecute is set back here
3728
3729 C->set_has_split_ifs(true); // Has chance for split-if optimization
3730 if (!stopped()) {
3731 set_result(newcopy);
3732 }
3733 return true;
3734 }
3735
3736
3737 //----------------------generate_virtual_guard---------------------------
3738 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3739 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3740 RegionNode* slow_region) {
3741 ciMethod* method = callee();
3742 int vtable_index = method->vtable_index();
3743 // Get the Method* out of the appropriate vtable entry.
3744 int entry_offset = (InstanceKlass::vtable_start_offset() +
3745 vtable_index*vtableEntry::size()) * wordSize +
3746 vtableEntry::method_offset_in_bytes();
3747 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3748 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3749
3750 // Compare the target method with the expected method (e.g., Object.hashCode).
3751 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3752
3753 Node* native_call = makecon(native_call_addr);
3754 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3755 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3756
3757 return generate_slow_guard(test_native, slow_region);
3758 }
3759
3760 //-----------------------generate_method_call----------------------------
3761 // Use generate_method_call to make a slow-call to the real
3762 // method if the fast path fails. An alternative would be to
3763 // use a stub like OptoRuntime::slow_arraycopy_Java.
3764 // This only works for expanding the current library call,
3765 // not another intrinsic. (E.g., don't use this for making an
3766 // arraycopy call inside of the copyOf intrinsic.)
3767 CallJavaNode*
3768 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3769 // When compiling the intrinsic method itself, do not use this technique.
3770 guarantee(callee() != C->method(), "cannot make slow-call to self");
3771
3772 ciMethod* method = callee();
3773 // ensure the JVMS we have will be correct for this call
3774 guarantee(method_id == method->intrinsic_id(), "must match");
3775
3776 const TypeFunc* tf = TypeFunc::make(method);
3777 CallJavaNode* slow_call;
3778 if (is_static) {
3779 assert(!is_virtual, "");
3780 slow_call = new(C) CallStaticJavaNode(C, tf,
3781 SharedRuntime::get_resolve_static_call_stub(),
3782 method, bci());
3783 } else if (is_virtual) {
3784 null_check_receiver();
3785 int vtable_index = Method::invalid_vtable_index;
3786 if (UseInlineCaches) {
3787 // Suppress the vtable call
3788 } else {
3789 // hashCode and clone are not a miranda methods,
3790 // so the vtable index is fixed.
3791 // No need to use the linkResolver to get it.
3792 vtable_index = method->vtable_index();
3793 }
3794 slow_call = new(C) CallDynamicJavaNode(tf,
3795 SharedRuntime::get_resolve_virtual_call_stub(),
3796 method, vtable_index, bci());
3797 } else { // neither virtual nor static: opt_virtual
3798 null_check_receiver();
3799 slow_call = new(C) CallStaticJavaNode(C, tf,
3800 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3801 method, bci());
3802 slow_call->set_optimized_virtual(true);
3803 }
3804 set_arguments_for_java_call(slow_call);
3805 set_edges_for_java_call(slow_call);
3806 return slow_call;
3807 }
3808
3809
3810 //------------------------------inline_native_hashcode--------------------
3811 // Build special case code for calls to hashCode on an object.
3812 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3813 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3814 assert(!(is_virtual && is_static), "either virtual, special, or static");
3815
3816 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3817
3818 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3819 PhiNode* result_val = new(C) PhiNode(result_reg,
3820 TypeInt::INT);
3821 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3822 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3823 TypePtr::BOTTOM);
3824 Node* obj = NULL;
3825 if (!is_static) {
3826 // Check for hashing null object
3827 obj = null_check_receiver();
3828 if (stopped()) return true; // unconditionally null
3829 result_reg->init_req(_null_path, top());
3830 result_val->init_req(_null_path, top());
3831 } else {
3832 // Do a null check, and return zero if null.
3833 // System.identityHashCode(null) == 0
3834 obj = argument(0);
3835 Node* null_ctl = top();
3836 obj = null_check_oop(obj, &null_ctl);
3837 result_reg->init_req(_null_path, null_ctl);
3838 result_val->init_req(_null_path, _gvn.intcon(0));
3839 }
3840
3841 // Unconditionally null? Then return right away.
3842 if (stopped()) {
3843 set_control( result_reg->in(_null_path));
3844 if (!stopped())
3845 set_result(result_val->in(_null_path));
3846 return true;
3847 }
3848
3849 // After null check, get the object's klass.
3850 Node* obj_klass = load_object_klass(obj);
3851
3852 // This call may be virtual (invokevirtual) or bound (invokespecial).
3853 // For each case we generate slightly different code.
3854
3855 // We only go to the fast case code if we pass a number of guards. The
3856 // paths which do not pass are accumulated in the slow_region.
3857 RegionNode* slow_region = new (C) RegionNode(1);
3858 record_for_igvn(slow_region);
3859
3860 // If this is a virtual call, we generate a funny guard. We pull out
3861 // the vtable entry corresponding to hashCode() from the target object.
3862 // If the target method which we are calling happens to be the native
3863 // Object hashCode() method, we pass the guard. We do not need this
3864 // guard for non-virtual calls -- the caller is known to be the native
3865 // Object hashCode().
3866 if (is_virtual) {
3867 generate_virtual_guard(obj_klass, slow_region);
3868 }
3869
3870 // Get the header out of the object, use LoadMarkNode when available
3871 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3872 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3873
3874 // Test the header to see if it is unlocked.
3875 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3876 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
3877 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3878 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
3879 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
3880
3881 generate_slow_guard(test_unlocked, slow_region);
3882
3883 // Get the hash value and check to see that it has been properly assigned.
3884 // We depend on hash_mask being at most 32 bits and avoid the use of
3885 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3886 // vm: see markOop.hpp.
3887 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3888 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3889 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
3890 // This hack lets the hash bits live anywhere in the mark object now, as long
3891 // as the shift drops the relevant bits into the low 32 bits. Note that
3892 // Java spec says that HashCode is an int so there's no point in capturing
3893 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3894 hshifted_header = ConvX2I(hshifted_header);
3895 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
3896
3897 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3898 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
3899 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
3900
3901 generate_slow_guard(test_assigned, slow_region);
3902
3903 Node* init_mem = reset_memory();
3904 // fill in the rest of the null path:
3905 result_io ->init_req(_null_path, i_o());
3906 result_mem->init_req(_null_path, init_mem);
3907
3908 result_val->init_req(_fast_path, hash_val);
3909 result_reg->init_req(_fast_path, control());
3910 result_io ->init_req(_fast_path, i_o());
3911 result_mem->init_req(_fast_path, init_mem);
3912
3913 // Generate code for the slow case. We make a call to hashCode().
3914 set_control(_gvn.transform(slow_region));
3915 if (!stopped()) {
3916 // No need for PreserveJVMState, because we're using up the present state.
3917 set_all_memory(init_mem);
3918 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3919 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3920 Node* slow_result = set_results_for_java_call(slow_call);
3921 // this->control() comes from set_results_for_java_call
3922 result_reg->init_req(_slow_path, control());
3923 result_val->init_req(_slow_path, slow_result);
3924 result_io ->set_req(_slow_path, i_o());
3925 result_mem ->set_req(_slow_path, reset_memory());
3926 }
3927
3928 // Return the combined state.
3929 set_i_o( _gvn.transform(result_io) );
3930 set_all_memory( _gvn.transform(result_mem));
3931
3932 set_result(result_reg, result_val);
3933 return true;
3934 }
3935
3936 //---------------------------inline_native_getClass----------------------------
3937 // public final native Class<?> java.lang.Object.getClass();
3938 //
3939 // Build special case code for calls to getClass on an object.
3940 bool LibraryCallKit::inline_native_getClass() {
3941 Node* obj = null_check_receiver();
3942 if (stopped()) return true;
3943 set_result(load_mirror_from_klass(load_object_klass(obj)));
3944 return true;
3945 }
3946
3947 //-----------------inline_native_Reflection_getCallerClass---------------------
3948 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
3949 //
3950 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3951 //
3952 // NOTE: This code must perform the same logic as JVM_GetCallerClass
3953 // in that it must skip particular security frames and checks for
3954 // caller sensitive methods.
3955 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3956 #ifndef PRODUCT
3957 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3958 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3959 }
3960 #endif
3961
3962 if (!jvms()->has_method()) {
3963 #ifndef PRODUCT
3964 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3965 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3966 }
3967 #endif
3968 return false;
3969 }
3970
3971 // Walk back up the JVM state to find the caller at the required
3972 // depth.
3973 JVMState* caller_jvms = jvms();
3974
3975 // Cf. JVM_GetCallerClass
3976 // NOTE: Start the loop at depth 1 because the current JVM state does
3977 // not include the Reflection.getCallerClass() frame.
3978 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
3979 ciMethod* m = caller_jvms->method();
3980 switch (n) {
3981 case 0:
3982 fatal("current JVM state does not include the Reflection.getCallerClass frame");
3983 break;
3984 case 1:
3985 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
3986 if (!m->caller_sensitive()) {
3987 #ifndef PRODUCT
3988 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3989 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
3990 }
3991 #endif
3992 return false; // bail-out; let JVM_GetCallerClass do the work
3993 }
3994 break;
3995 default:
3996 if (!m->is_ignored_by_security_stack_walk()) {
3997 // We have reached the desired frame; return the holder class.
3998 // Acquire method holder as java.lang.Class and push as constant.
3999 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4000 ciInstance* caller_mirror = caller_klass->java_mirror();
4001 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4002
4003 #ifndef PRODUCT
4004 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4005 tty->print_cr(" Succeeded: caller = %d) %s.%s, JVMS depth = %d", n, caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), jvms()->depth());
4006 tty->print_cr(" JVM state at this point:");
4007 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4008 ciMethod* m = jvms()->of_depth(i)->method();
4009 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4010 }
4011 }
4012 #endif
4013 return true;
4014 }
4015 break;
4016 }
4017 }
4018
4019 #ifndef PRODUCT
4020 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4021 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4022 tty->print_cr(" JVM state at this point:");
4023 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4024 ciMethod* m = jvms()->of_depth(i)->method();
4025 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4026 }
4027 }
4028 #endif
4029
4030 return false; // bail-out; let JVM_GetCallerClass do the work
4031 }
4032
4033 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4034 Node* arg = argument(0);
4035 Node* result;
4036
4037 switch (id) {
4038 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4039 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4040 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4041 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4042
4043 case vmIntrinsics::_doubleToLongBits: {
4044 // two paths (plus control) merge in a wood
4045 RegionNode *r = new (C) RegionNode(3);
4046 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4047
4048 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4049 // Build the boolean node
4050 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4051
4052 // Branch either way.
4053 // NaN case is less traveled, which makes all the difference.
4054 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4055 Node *opt_isnan = _gvn.transform(ifisnan);
4056 assert( opt_isnan->is_If(), "Expect an IfNode");
4057 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4058 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4059
4060 set_control(iftrue);
4061
4062 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4063 Node *slow_result = longcon(nan_bits); // return NaN
4064 phi->init_req(1, _gvn.transform( slow_result ));
4065 r->init_req(1, iftrue);
4066
4067 // Else fall through
4068 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4069 set_control(iffalse);
4070
4071 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4072 r->init_req(2, iffalse);
4073
4074 // Post merge
4075 set_control(_gvn.transform(r));
4076 record_for_igvn(r);
4077
4078 C->set_has_split_ifs(true); // Has chance for split-if optimization
4079 result = phi;
4080 assert(result->bottom_type()->isa_long(), "must be");
4081 break;
4082 }
4083
4084 case vmIntrinsics::_floatToIntBits: {
4085 // two paths (plus control) merge in a wood
4086 RegionNode *r = new (C) RegionNode(3);
4087 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4088
4089 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4090 // Build the boolean node
4091 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4092
4093 // Branch either way.
4094 // NaN case is less traveled, which makes all the difference.
4095 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4096 Node *opt_isnan = _gvn.transform(ifisnan);
4097 assert( opt_isnan->is_If(), "Expect an IfNode");
4098 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4099 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4100
4101 set_control(iftrue);
4102
4103 static const jint nan_bits = 0x7fc00000;
4104 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4105 phi->init_req(1, _gvn.transform( slow_result ));
4106 r->init_req(1, iftrue);
4107
4108 // Else fall through
4109 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4110 set_control(iffalse);
4111
4112 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4113 r->init_req(2, iffalse);
4114
4115 // Post merge
4116 set_control(_gvn.transform(r));
4117 record_for_igvn(r);
4118
4119 C->set_has_split_ifs(true); // Has chance for split-if optimization
4120 result = phi;
4121 assert(result->bottom_type()->isa_int(), "must be");
4122 break;
4123 }
4124
4125 default:
4126 fatal_unexpected_iid(id);
4127 break;
4128 }
4129 set_result(_gvn.transform(result));
4130 return true;
4131 }
4132
4133 #ifdef _LP64
4134 #define XTOP ,top() /*additional argument*/
4135 #else //_LP64
4136 #define XTOP /*no additional argument*/
4137 #endif //_LP64
4138
4139 //----------------------inline_unsafe_copyMemory-------------------------
4140 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4141 bool LibraryCallKit::inline_unsafe_copyMemory() {
4142 if (callee()->is_static()) return false; // caller must have the capability!
4143 null_check_receiver(); // null-check receiver
4144 if (stopped()) return true;
4145
4146 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4147
4148 Node* src_ptr = argument(1); // type: oop
4149 Node* src_off = ConvL2X(argument(2)); // type: long
4150 Node* dst_ptr = argument(4); // type: oop
4151 Node* dst_off = ConvL2X(argument(5)); // type: long
4152 Node* size = ConvL2X(argument(7)); // type: long
4153
4154 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4155 "fieldOffset must be byte-scaled");
4156
4157 Node* src = make_unsafe_address(src_ptr, src_off);
4158 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4159
4160 // Conservatively insert a memory barrier on all memory slices.
4161 // Do not let writes of the copy source or destination float below the copy.
4162 insert_mem_bar(Op_MemBarCPUOrder);
4163
4164 // Call it. Note that the length argument is not scaled.
4165 make_runtime_call(RC_LEAF|RC_NO_FP,
4166 OptoRuntime::fast_arraycopy_Type(),
4167 StubRoutines::unsafe_arraycopy(),
4168 "unsafe_arraycopy",
4169 TypeRawPtr::BOTTOM,
4170 src, dst, size XTOP);
4171
4172 // Do not let reads of the copy destination float above the copy.
4173 insert_mem_bar(Op_MemBarCPUOrder);
4174
4175 return true;
4176 }
4177
4178 //------------------------clone_coping-----------------------------------
4179 // Helper function for inline_native_clone.
4180 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4181 assert(obj_size != NULL, "");
4182 Node* raw_obj = alloc_obj->in(1);
4183 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4184
4185 AllocateNode* alloc = NULL;
4186 if (ReduceBulkZeroing) {
4187 // We will be completely responsible for initializing this object -
4188 // mark Initialize node as complete.
4189 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4190 // The object was just allocated - there should be no any stores!
4191 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4192 // Mark as complete_with_arraycopy so that on AllocateNode
4193 // expansion, we know this AllocateNode is initialized by an array
4194 // copy and a StoreStore barrier exists after the array copy.
4195 alloc->initialization()->set_complete_with_arraycopy();
4196 }
4197
4198 // Copy the fastest available way.
4199 // TODO: generate fields copies for small objects instead.
4200 Node* src = obj;
4201 Node* dest = alloc_obj;
4202 Node* size = _gvn.transform(obj_size);
4203
4204 // Exclude the header but include array length to copy by 8 bytes words.
4205 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4206 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4207 instanceOopDesc::base_offset_in_bytes();
4208 // base_off:
4209 // 8 - 32-bit VM
4210 // 12 - 64-bit VM, compressed klass
4211 // 16 - 64-bit VM, normal klass
4212 if (base_off % BytesPerLong != 0) {
4213 assert(UseCompressedKlassPointers, "");
4214 if (is_array) {
4215 // Exclude length to copy by 8 bytes words.
4216 base_off += sizeof(int);
4217 } else {
4218 // Include klass to copy by 8 bytes words.
4219 base_off = instanceOopDesc::klass_offset_in_bytes();
4220 }
4221 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4222 }
4223 src = basic_plus_adr(src, base_off);
4224 dest = basic_plus_adr(dest, base_off);
4225
4226 // Compute the length also, if needed:
4227 Node* countx = size;
4228 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4229 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4230
4231 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4232 bool disjoint_bases = true;
4233 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4234 src, NULL, dest, NULL, countx,
4235 /*dest_uninitialized*/true);
4236
4237 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4238 if (card_mark) {
4239 assert(!is_array, "");
4240 // Put in store barrier for any and all oops we are sticking
4241 // into this object. (We could avoid this if we could prove
4242 // that the object type contains no oop fields at all.)
4243 Node* no_particular_value = NULL;
4244 Node* no_particular_field = NULL;
4245 int raw_adr_idx = Compile::AliasIdxRaw;
4246 post_barrier(control(),
4247 memory(raw_adr_type),
4248 alloc_obj,
4249 no_particular_field,
4250 raw_adr_idx,
4251 no_particular_value,
4252 T_OBJECT,
4253 false);
4254 }
4255
4256 // Do not let reads from the cloned object float above the arraycopy.
4257 if (alloc != NULL) {
4258 // Do not let stores that initialize this object be reordered with
4259 // a subsequent store that would make this object accessible by
4260 // other threads.
4261 // Record what AllocateNode this StoreStore protects so that
4262 // escape analysis can go from the MemBarStoreStoreNode to the
4263 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4264 // based on the escape status of the AllocateNode.
4265 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4266 } else {
4267 insert_mem_bar(Op_MemBarCPUOrder);
4268 }
4269 }
4270
4271 //------------------------inline_native_clone----------------------------
4272 // protected native Object java.lang.Object.clone();
4273 //
4274 // Here are the simple edge cases:
4275 // null receiver => normal trap
4276 // virtual and clone was overridden => slow path to out-of-line clone
4277 // not cloneable or finalizer => slow path to out-of-line Object.clone
4278 //
4279 // The general case has two steps, allocation and copying.
4280 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4281 //
4282 // Copying also has two cases, oop arrays and everything else.
4283 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4284 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4285 //
4286 // These steps fold up nicely if and when the cloned object's klass
4287 // can be sharply typed as an object array, a type array, or an instance.
4288 //
4289 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4290 PhiNode* result_val;
4291
4292 // Set the reexecute bit for the interpreter to reexecute
4293 // the bytecode that invokes Object.clone if deoptimization happens.
4294 { PreserveReexecuteState preexecs(this);
4295 jvms()->set_should_reexecute(true);
4296
4297 Node* obj = null_check_receiver();
4298 if (stopped()) return true;
4299
4300 Node* obj_klass = load_object_klass(obj);
4301 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4302 const TypeOopPtr* toop = ((tklass != NULL)
4303 ? tklass->as_instance_type()
4304 : TypeInstPtr::NOTNULL);
4305
4306 // Conservatively insert a memory barrier on all memory slices.
4307 // Do not let writes into the original float below the clone.
4308 insert_mem_bar(Op_MemBarCPUOrder);
4309
4310 // paths into result_reg:
4311 enum {
4312 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4313 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4314 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4315 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4316 PATH_LIMIT
4317 };
4318 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4319 result_val = new(C) PhiNode(result_reg,
4320 TypeInstPtr::NOTNULL);
4321 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4322 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4323 TypePtr::BOTTOM);
4324 record_for_igvn(result_reg);
4325
4326 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4327 int raw_adr_idx = Compile::AliasIdxRaw;
4328
4329 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4330 if (array_ctl != NULL) {
4331 // It's an array.
4332 PreserveJVMState pjvms(this);
4333 set_control(array_ctl);
4334 Node* obj_length = load_array_length(obj);
4335 Node* obj_size = NULL;
4336 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4337
4338 if (!use_ReduceInitialCardMarks()) {
4339 // If it is an oop array, it requires very special treatment,
4340 // because card marking is required on each card of the array.
4341 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4342 if (is_obja != NULL) {
4343 PreserveJVMState pjvms2(this);
4344 set_control(is_obja);
4345 // Generate a direct call to the right arraycopy function(s).
4346 bool disjoint_bases = true;
4347 bool length_never_negative = true;
4348 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4349 obj, intcon(0), alloc_obj, intcon(0),
4350 obj_length,
4351 disjoint_bases, length_never_negative);
4352 result_reg->init_req(_objArray_path, control());
4353 result_val->init_req(_objArray_path, alloc_obj);
4354 result_i_o ->set_req(_objArray_path, i_o());
4355 result_mem ->set_req(_objArray_path, reset_memory());
4356 }
4357 }
4358 // Otherwise, there are no card marks to worry about.
4359 // (We can dispense with card marks if we know the allocation
4360 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4361 // causes the non-eden paths to take compensating steps to
4362 // simulate a fresh allocation, so that no further
4363 // card marks are required in compiled code to initialize
4364 // the object.)
4365
4366 if (!stopped()) {
4367 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4368
4369 // Present the results of the copy.
4370 result_reg->init_req(_array_path, control());
4371 result_val->init_req(_array_path, alloc_obj);
4372 result_i_o ->set_req(_array_path, i_o());
4373 result_mem ->set_req(_array_path, reset_memory());
4374 }
4375 }
4376
4377 // We only go to the instance fast case code if we pass a number of guards.
4378 // The paths which do not pass are accumulated in the slow_region.
4379 RegionNode* slow_region = new (C) RegionNode(1);
4380 record_for_igvn(slow_region);
4381 if (!stopped()) {
4382 // It's an instance (we did array above). Make the slow-path tests.
4383 // If this is a virtual call, we generate a funny guard. We grab
4384 // the vtable entry corresponding to clone() from the target object.
4385 // If the target method which we are calling happens to be the
4386 // Object clone() method, we pass the guard. We do not need this
4387 // guard for non-virtual calls; the caller is known to be the native
4388 // Object clone().
4389 if (is_virtual) {
4390 generate_virtual_guard(obj_klass, slow_region);
4391 }
4392
4393 // The object must be cloneable and must not have a finalizer.
4394 // Both of these conditions may be checked in a single test.
4395 // We could optimize the cloneable test further, but we don't care.
4396 generate_access_flags_guard(obj_klass,
4397 // Test both conditions:
4398 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4399 // Must be cloneable but not finalizer:
4400 JVM_ACC_IS_CLONEABLE,
4401 slow_region);
4402 }
4403
4404 if (!stopped()) {
4405 // It's an instance, and it passed the slow-path tests.
4406 PreserveJVMState pjvms(this);
4407 Node* obj_size = NULL;
4408 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4409
4410 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4411
4412 // Present the results of the slow call.
4413 result_reg->init_req(_instance_path, control());
4414 result_val->init_req(_instance_path, alloc_obj);
4415 result_i_o ->set_req(_instance_path, i_o());
4416 result_mem ->set_req(_instance_path, reset_memory());
4417 }
4418
4419 // Generate code for the slow case. We make a call to clone().
4420 set_control(_gvn.transform(slow_region));
4421 if (!stopped()) {
4422 PreserveJVMState pjvms(this);
4423 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4424 Node* slow_result = set_results_for_java_call(slow_call);
4425 // this->control() comes from set_results_for_java_call
4426 result_reg->init_req(_slow_path, control());
4427 result_val->init_req(_slow_path, slow_result);
4428 result_i_o ->set_req(_slow_path, i_o());
4429 result_mem ->set_req(_slow_path, reset_memory());
4430 }
4431
4432 // Return the combined state.
4433 set_control( _gvn.transform(result_reg));
4434 set_i_o( _gvn.transform(result_i_o));
4435 set_all_memory( _gvn.transform(result_mem));
4436 } // original reexecute is set back here
4437
4438 set_result(_gvn.transform(result_val));
4439 return true;
4440 }
4441
4442 //------------------------------basictype2arraycopy----------------------------
4443 address LibraryCallKit::basictype2arraycopy(BasicType t,
4444 Node* src_offset,
4445 Node* dest_offset,
4446 bool disjoint_bases,
4447 const char* &name,
4448 bool dest_uninitialized) {
4449 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4450 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4451
4452 bool aligned = false;
4453 bool disjoint = disjoint_bases;
4454
4455 // if the offsets are the same, we can treat the memory regions as
4456 // disjoint, because either the memory regions are in different arrays,
4457 // or they are identical (which we can treat as disjoint.) We can also
4458 // treat a copy with a destination index less that the source index
4459 // as disjoint since a low->high copy will work correctly in this case.
4460 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4461 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4462 // both indices are constants
4463 int s_offs = src_offset_inttype->get_con();
4464 int d_offs = dest_offset_inttype->get_con();
4465 int element_size = type2aelembytes(t);
4466 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4467 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4468 if (s_offs >= d_offs) disjoint = true;
4469 } else if (src_offset == dest_offset && src_offset != NULL) {
4470 // This can occur if the offsets are identical non-constants.
4471 disjoint = true;
4472 }
4473
4474 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4475 }
4476
4477
4478 //------------------------------inline_arraycopy-----------------------
4479 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4480 // Object dest, int destPos,
4481 // int length);
4482 bool LibraryCallKit::inline_arraycopy() {
4483 // Get the arguments.
4484 Node* src = argument(0); // type: oop
4485 Node* src_offset = argument(1); // type: int
4486 Node* dest = argument(2); // type: oop
4487 Node* dest_offset = argument(3); // type: int
4488 Node* length = argument(4); // type: int
4489
4490 // Compile time checks. If any of these checks cannot be verified at compile time,
4491 // we do not make a fast path for this call. Instead, we let the call remain as it
4492 // is. The checks we choose to mandate at compile time are:
4493 //
4494 // (1) src and dest are arrays.
4495 const Type* src_type = src->Value(&_gvn);
4496 const Type* dest_type = dest->Value(&_gvn);
4497 const TypeAryPtr* top_src = src_type->isa_aryptr();
4498 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4499 if (top_src == NULL || top_src->klass() == NULL ||
4500 top_dest == NULL || top_dest->klass() == NULL) {
4501 // Conservatively insert a memory barrier on all memory slices.
4502 // Do not let writes into the source float below the arraycopy.
4503 insert_mem_bar(Op_MemBarCPUOrder);
4504
4505 // Call StubRoutines::generic_arraycopy stub.
4506 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4507 src, src_offset, dest, dest_offset, length);
4508
4509 // Do not let reads from the destination float above the arraycopy.
4510 // Since we cannot type the arrays, we don't know which slices
4511 // might be affected. We could restrict this barrier only to those
4512 // memory slices which pertain to array elements--but don't bother.
4513 if (!InsertMemBarAfterArraycopy)
4514 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4515 insert_mem_bar(Op_MemBarCPUOrder);
4516 return true;
4517 }
4518
4519 // (2) src and dest arrays must have elements of the same BasicType
4520 // Figure out the size and type of the elements we will be copying.
4521 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4522 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4523 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4524 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4525
4526 if (src_elem != dest_elem || dest_elem == T_VOID) {
4527 // The component types are not the same or are not recognized. Punt.
4528 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4529 generate_slow_arraycopy(TypePtr::BOTTOM,
4530 src, src_offset, dest, dest_offset, length,
4531 /*dest_uninitialized*/false);
4532 return true;
4533 }
4534
4535 //---------------------------------------------------------------------------
4536 // We will make a fast path for this call to arraycopy.
4537
4538 // We have the following tests left to perform:
4539 //
4540 // (3) src and dest must not be null.
4541 // (4) src_offset must not be negative.
4542 // (5) dest_offset must not be negative.
4543 // (6) length must not be negative.
4544 // (7) src_offset + length must not exceed length of src.
4545 // (8) dest_offset + length must not exceed length of dest.
4546 // (9) each element of an oop array must be assignable
4547
4548 RegionNode* slow_region = new (C) RegionNode(1);
4549 record_for_igvn(slow_region);
4550
4551 // (3) operands must not be null
4552 // We currently perform our null checks with the null_check routine.
4553 // This means that the null exceptions will be reported in the caller
4554 // rather than (correctly) reported inside of the native arraycopy call.
4555 // This should be corrected, given time. We do our null check with the
4556 // stack pointer restored.
4557 src = null_check(src, T_ARRAY);
4558 dest = null_check(dest, T_ARRAY);
4559
4560 // (4) src_offset must not be negative.
4561 generate_negative_guard(src_offset, slow_region);
4562
4563 // (5) dest_offset must not be negative.
4564 generate_negative_guard(dest_offset, slow_region);
4565
4566 // (6) length must not be negative (moved to generate_arraycopy()).
4567 // generate_negative_guard(length, slow_region);
4568
4569 // (7) src_offset + length must not exceed length of src.
4570 generate_limit_guard(src_offset, length,
4571 load_array_length(src),
4572 slow_region);
4573
4574 // (8) dest_offset + length must not exceed length of dest.
4575 generate_limit_guard(dest_offset, length,
4576 load_array_length(dest),
4577 slow_region);
4578
4579 // (9) each element of an oop array must be assignable
4580 // The generate_arraycopy subroutine checks this.
4581
4582 // This is where the memory effects are placed:
4583 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4584 generate_arraycopy(adr_type, dest_elem,
4585 src, src_offset, dest, dest_offset, length,
4586 false, false, slow_region);
4587
4588 return true;
4589 }
4590
4591 //-----------------------------generate_arraycopy----------------------
4592 // Generate an optimized call to arraycopy.
4593 // Caller must guard against non-arrays.
4594 // Caller must determine a common array basic-type for both arrays.
4595 // Caller must validate offsets against array bounds.
4596 // The slow_region has already collected guard failure paths
4597 // (such as out of bounds length or non-conformable array types).
4598 // The generated code has this shape, in general:
4599 //
4600 // if (length == 0) return // via zero_path
4601 // slowval = -1
4602 // if (types unknown) {
4603 // slowval = call generic copy loop
4604 // if (slowval == 0) return // via checked_path
4605 // } else if (indexes in bounds) {
4606 // if ((is object array) && !(array type check)) {
4607 // slowval = call checked copy loop
4608 // if (slowval == 0) return // via checked_path
4609 // } else {
4610 // call bulk copy loop
4611 // return // via fast_path
4612 // }
4613 // }
4614 // // adjust params for remaining work:
4615 // if (slowval != -1) {
4616 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4617 // }
4618 // slow_region:
4619 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4620 // return // via slow_call_path
4621 //
4622 // This routine is used from several intrinsics: System.arraycopy,
4623 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4624 //
4625 void
4626 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4627 BasicType basic_elem_type,
4628 Node* src, Node* src_offset,
4629 Node* dest, Node* dest_offset,
4630 Node* copy_length,
4631 bool disjoint_bases,
4632 bool length_never_negative,
4633 RegionNode* slow_region) {
4634
4635 if (slow_region == NULL) {
4636 slow_region = new(C) RegionNode(1);
4637 record_for_igvn(slow_region);
4638 }
4639
4640 Node* original_dest = dest;
4641 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4642 bool dest_uninitialized = false;
4643
4644 // See if this is the initialization of a newly-allocated array.
4645 // If so, we will take responsibility here for initializing it to zero.
4646 // (Note: Because tightly_coupled_allocation performs checks on the
4647 // out-edges of the dest, we need to avoid making derived pointers
4648 // from it until we have checked its uses.)
4649 if (ReduceBulkZeroing
4650 && !ZeroTLAB // pointless if already zeroed
4651 && basic_elem_type != T_CONFLICT // avoid corner case
4652 && !src->eqv_uncast(dest)
4653 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4654 != NULL)
4655 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4656 && alloc->maybe_set_complete(&_gvn)) {
4657 // "You break it, you buy it."
4658 InitializeNode* init = alloc->initialization();
4659 assert(init->is_complete(), "we just did this");
4660 init->set_complete_with_arraycopy();
4661 assert(dest->is_CheckCastPP(), "sanity");
4662 assert(dest->in(0)->in(0) == init, "dest pinned");
4663 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4664 // From this point on, every exit path is responsible for
4665 // initializing any non-copied parts of the object to zero.
4666 // Also, if this flag is set we make sure that arraycopy interacts properly
4667 // with G1, eliding pre-barriers. See CR 6627983.
4668 dest_uninitialized = true;
4669 } else {
4670 // No zeroing elimination here.
4671 alloc = NULL;
4672 //original_dest = dest;
4673 //dest_uninitialized = false;
4674 }
4675
4676 // Results are placed here:
4677 enum { fast_path = 1, // normal void-returning assembly stub
4678 checked_path = 2, // special assembly stub with cleanup
4679 slow_call_path = 3, // something went wrong; call the VM
4680 zero_path = 4, // bypass when length of copy is zero
4681 bcopy_path = 5, // copy primitive array by 64-bit blocks
4682 PATH_LIMIT = 6
4683 };
4684 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4685 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4686 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4687 record_for_igvn(result_region);
4688 _gvn.set_type_bottom(result_i_o);
4689 _gvn.set_type_bottom(result_memory);
4690 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4691
4692 // The slow_control path:
4693 Node* slow_control;
4694 Node* slow_i_o = i_o();
4695 Node* slow_mem = memory(adr_type);
4696 debug_only(slow_control = (Node*) badAddress);
4697
4698 // Checked control path:
4699 Node* checked_control = top();
4700 Node* checked_mem = NULL;
4701 Node* checked_i_o = NULL;
4702 Node* checked_value = NULL;
4703
4704 if (basic_elem_type == T_CONFLICT) {
4705 assert(!dest_uninitialized, "");
4706 Node* cv = generate_generic_arraycopy(adr_type,
4707 src, src_offset, dest, dest_offset,
4708 copy_length, dest_uninitialized);
4709 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4710 checked_control = control();
4711 checked_i_o = i_o();
4712 checked_mem = memory(adr_type);
4713 checked_value = cv;
4714 set_control(top()); // no fast path
4715 }
4716
4717 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4718 if (not_pos != NULL) {
4719 PreserveJVMState pjvms(this);
4720 set_control(not_pos);
4721
4722 // (6) length must not be negative.
4723 if (!length_never_negative) {
4724 generate_negative_guard(copy_length, slow_region);
4725 }
4726
4727 // copy_length is 0.
4728 if (!stopped() && dest_uninitialized) {
4729 Node* dest_length = alloc->in(AllocateNode::ALength);
4730 if (copy_length->eqv_uncast(dest_length)
4731 || _gvn.find_int_con(dest_length, 1) <= 0) {
4732 // There is no zeroing to do. No need for a secondary raw memory barrier.
4733 } else {
4734 // Clear the whole thing since there are no source elements to copy.
4735 generate_clear_array(adr_type, dest, basic_elem_type,
4736 intcon(0), NULL,
4737 alloc->in(AllocateNode::AllocSize));
4738 // Use a secondary InitializeNode as raw memory barrier.
4739 // Currently it is needed only on this path since other
4740 // paths have stub or runtime calls as raw memory barriers.
4741 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4742 Compile::AliasIdxRaw,
4743 top())->as_Initialize();
4744 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4745 }
4746 }
4747
4748 // Present the results of the fast call.
4749 result_region->init_req(zero_path, control());
4750 result_i_o ->init_req(zero_path, i_o());
4751 result_memory->init_req(zero_path, memory(adr_type));
4752 }
4753
4754 if (!stopped() && dest_uninitialized) {
4755 // We have to initialize the *uncopied* part of the array to zero.
4756 // The copy destination is the slice dest[off..off+len]. The other slices
4757 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4758 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4759 Node* dest_length = alloc->in(AllocateNode::ALength);
4760 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4761 copy_length));
4762
4763 // If there is a head section that needs zeroing, do it now.
4764 if (find_int_con(dest_offset, -1) != 0) {
4765 generate_clear_array(adr_type, dest, basic_elem_type,
4766 intcon(0), dest_offset,
4767 NULL);
4768 }
4769
4770 // Next, perform a dynamic check on the tail length.
4771 // It is often zero, and we can win big if we prove this.
4772 // There are two wins: Avoid generating the ClearArray
4773 // with its attendant messy index arithmetic, and upgrade
4774 // the copy to a more hardware-friendly word size of 64 bits.
4775 Node* tail_ctl = NULL;
4776 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4777 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
4778 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
4779 tail_ctl = generate_slow_guard(bol_lt, NULL);
4780 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4781 }
4782
4783 // At this point, let's assume there is no tail.
4784 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4785 // There is no tail. Try an upgrade to a 64-bit copy.
4786 bool didit = false;
4787 { PreserveJVMState pjvms(this);
4788 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4789 src, src_offset, dest, dest_offset,
4790 dest_size, dest_uninitialized);
4791 if (didit) {
4792 // Present the results of the block-copying fast call.
4793 result_region->init_req(bcopy_path, control());
4794 result_i_o ->init_req(bcopy_path, i_o());
4795 result_memory->init_req(bcopy_path, memory(adr_type));
4796 }
4797 }
4798 if (didit)
4799 set_control(top()); // no regular fast path
4800 }
4801
4802 // Clear the tail, if any.
4803 if (tail_ctl != NULL) {
4804 Node* notail_ctl = stopped() ? NULL : control();
4805 set_control(tail_ctl);
4806 if (notail_ctl == NULL) {
4807 generate_clear_array(adr_type, dest, basic_elem_type,
4808 dest_tail, NULL,
4809 dest_size);
4810 } else {
4811 // Make a local merge.
4812 Node* done_ctl = new(C) RegionNode(3);
4813 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4814 done_ctl->init_req(1, notail_ctl);
4815 done_mem->init_req(1, memory(adr_type));
4816 generate_clear_array(adr_type, dest, basic_elem_type,
4817 dest_tail, NULL,
4818 dest_size);
4819 done_ctl->init_req(2, control());
4820 done_mem->init_req(2, memory(adr_type));
4821 set_control( _gvn.transform(done_ctl));
4822 set_memory( _gvn.transform(done_mem), adr_type );
4823 }
4824 }
4825 }
4826
4827 BasicType copy_type = basic_elem_type;
4828 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4829 if (!stopped() && copy_type == T_OBJECT) {
4830 // If src and dest have compatible element types, we can copy bits.
4831 // Types S[] and D[] are compatible if D is a supertype of S.
4832 //
4833 // If they are not, we will use checked_oop_disjoint_arraycopy,
4834 // which performs a fast optimistic per-oop check, and backs off
4835 // further to JVM_ArrayCopy on the first per-oop check that fails.
4836 // (Actually, we don't move raw bits only; the GC requires card marks.)
4837
4838 // Get the Klass* for both src and dest
4839 Node* src_klass = load_object_klass(src);
4840 Node* dest_klass = load_object_klass(dest);
4841
4842 // Generate the subtype check.
4843 // This might fold up statically, or then again it might not.
4844 //
4845 // Non-static example: Copying List<String>.elements to a new String[].
4846 // The backing store for a List<String> is always an Object[],
4847 // but its elements are always type String, if the generic types
4848 // are correct at the source level.
4849 //
4850 // Test S[] against D[], not S against D, because (probably)
4851 // the secondary supertype cache is less busy for S[] than S.
4852 // This usually only matters when D is an interface.
4853 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4854 // Plug failing path into checked_oop_disjoint_arraycopy
4855 if (not_subtype_ctrl != top()) {
4856 PreserveJVMState pjvms(this);
4857 set_control(not_subtype_ctrl);
4858 // (At this point we can assume disjoint_bases, since types differ.)
4859 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
4860 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4861 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4862 Node* dest_elem_klass = _gvn.transform(n1);
4863 Node* cv = generate_checkcast_arraycopy(adr_type,
4864 dest_elem_klass,
4865 src, src_offset, dest, dest_offset,
4866 ConvI2X(copy_length), dest_uninitialized);
4867 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4868 checked_control = control();
4869 checked_i_o = i_o();
4870 checked_mem = memory(adr_type);
4871 checked_value = cv;
4872 }
4873 // At this point we know we do not need type checks on oop stores.
4874
4875 // Let's see if we need card marks:
4876 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4877 // If we do not need card marks, copy using the jint or jlong stub.
4878 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4879 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4880 "sizes agree");
4881 }
4882 }
4883
4884 if (!stopped()) {
4885 // Generate the fast path, if possible.
4886 PreserveJVMState pjvms(this);
4887 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4888 src, src_offset, dest, dest_offset,
4889 ConvI2X(copy_length), dest_uninitialized);
4890
4891 // Present the results of the fast call.
4892 result_region->init_req(fast_path, control());
4893 result_i_o ->init_req(fast_path, i_o());
4894 result_memory->init_req(fast_path, memory(adr_type));
4895 }
4896
4897 // Here are all the slow paths up to this point, in one bundle:
4898 slow_control = top();
4899 if (slow_region != NULL)
4900 slow_control = _gvn.transform(slow_region);
4901 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4902
4903 set_control(checked_control);
4904 if (!stopped()) {
4905 // Clean up after the checked call.
4906 // The returned value is either 0 or -1^K,
4907 // where K = number of partially transferred array elements.
4908 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
4909 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
4910 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4911
4912 // If it is 0, we are done, so transfer to the end.
4913 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
4914 result_region->init_req(checked_path, checks_done);
4915 result_i_o ->init_req(checked_path, checked_i_o);
4916 result_memory->init_req(checked_path, checked_mem);
4917
4918 // If it is not zero, merge into the slow call.
4919 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
4920 RegionNode* slow_reg2 = new(C) RegionNode(3);
4921 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4922 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4923 record_for_igvn(slow_reg2);
4924 slow_reg2 ->init_req(1, slow_control);
4925 slow_i_o2 ->init_req(1, slow_i_o);
4926 slow_mem2 ->init_req(1, slow_mem);
4927 slow_reg2 ->init_req(2, control());
4928 slow_i_o2 ->init_req(2, checked_i_o);
4929 slow_mem2 ->init_req(2, checked_mem);
4930
4931 slow_control = _gvn.transform(slow_reg2);
4932 slow_i_o = _gvn.transform(slow_i_o2);
4933 slow_mem = _gvn.transform(slow_mem2);
4934
4935 if (alloc != NULL) {
4936 // We'll restart from the very beginning, after zeroing the whole thing.
4937 // This can cause double writes, but that's OK since dest is brand new.
4938 // So we ignore the low 31 bits of the value returned from the stub.
4939 } else {
4940 // We must continue the copy exactly where it failed, or else
4941 // another thread might see the wrong number of writes to dest.
4942 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
4943 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4944 slow_offset->init_req(1, intcon(0));
4945 slow_offset->init_req(2, checked_offset);
4946 slow_offset = _gvn.transform(slow_offset);
4947
4948 // Adjust the arguments by the conditionally incoming offset.
4949 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
4950 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
4951 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
4952
4953 // Tweak the node variables to adjust the code produced below:
4954 src_offset = src_off_plus;
4955 dest_offset = dest_off_plus;
4956 copy_length = length_minus;
4957 }
4958 }
4959
4960 set_control(slow_control);
4961 if (!stopped()) {
4962 // Generate the slow path, if needed.
4963 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4964
4965 set_memory(slow_mem, adr_type);
4966 set_i_o(slow_i_o);
4967
4968 if (dest_uninitialized) {
4969 generate_clear_array(adr_type, dest, basic_elem_type,
4970 intcon(0), NULL,
4971 alloc->in(AllocateNode::AllocSize));
4972 }
4973
4974 generate_slow_arraycopy(adr_type,
4975 src, src_offset, dest, dest_offset,
4976 copy_length, /*dest_uninitialized*/false);
4977
4978 result_region->init_req(slow_call_path, control());
4979 result_i_o ->init_req(slow_call_path, i_o());
4980 result_memory->init_req(slow_call_path, memory(adr_type));
4981 }
4982
4983 // Remove unused edges.
4984 for (uint i = 1; i < result_region->req(); i++) {
4985 if (result_region->in(i) == NULL)
4986 result_region->init_req(i, top());
4987 }
4988
4989 // Finished; return the combined state.
4990 set_control( _gvn.transform(result_region));
4991 set_i_o( _gvn.transform(result_i_o) );
4992 set_memory( _gvn.transform(result_memory), adr_type );
4993
4994 // The memory edges above are precise in order to model effects around
4995 // array copies accurately to allow value numbering of field loads around
4996 // arraycopy. Such field loads, both before and after, are common in Java
4997 // collections and similar classes involving header/array data structures.
4998 //
4999 // But with low number of register or when some registers are used or killed
5000 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5001 // The next memory barrier is added to avoid it. If the arraycopy can be
5002 // optimized away (which it can, sometimes) then we can manually remove
5003 // the membar also.
5004 //
5005 // Do not let reads from the cloned object float above the arraycopy.
5006 if (alloc != NULL) {
5007 // Do not let stores that initialize this object be reordered with
5008 // a subsequent store that would make this object accessible by
5009 // other threads.
5010 // Record what AllocateNode this StoreStore protects so that
5011 // escape analysis can go from the MemBarStoreStoreNode to the
5012 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5013 // based on the escape status of the AllocateNode.
5014 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5015 } else if (InsertMemBarAfterArraycopy)
5016 insert_mem_bar(Op_MemBarCPUOrder);
5017 }
5018
5019
5020 // Helper function which determines if an arraycopy immediately follows
5021 // an allocation, with no intervening tests or other escapes for the object.
5022 AllocateArrayNode*
5023 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5024 RegionNode* slow_region) {
5025 if (stopped()) return NULL; // no fast path
5026 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5027
5028 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5029 if (alloc == NULL) return NULL;
5030
5031 Node* rawmem = memory(Compile::AliasIdxRaw);
5032 // Is the allocation's memory state untouched?
5033 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5034 // Bail out if there have been raw-memory effects since the allocation.
5035 // (Example: There might have been a call or safepoint.)
5036 return NULL;
5037 }
5038 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5039 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5040 return NULL;
5041 }
5042
5043 // There must be no unexpected observers of this allocation.
5044 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5045 Node* obs = ptr->fast_out(i);
5046 if (obs != this->map()) {
5047 return NULL;
5048 }
5049 }
5050
5051 // This arraycopy must unconditionally follow the allocation of the ptr.
5052 Node* alloc_ctl = ptr->in(0);
5053 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5054
5055 Node* ctl = control();
5056 while (ctl != alloc_ctl) {
5057 // There may be guards which feed into the slow_region.
5058 // Any other control flow means that we might not get a chance
5059 // to finish initializing the allocated object.
5060 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5061 IfNode* iff = ctl->in(0)->as_If();
5062 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5063 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5064 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5065 ctl = iff->in(0); // This test feeds the known slow_region.
5066 continue;
5067 }
5068 // One more try: Various low-level checks bottom out in
5069 // uncommon traps. If the debug-info of the trap omits
5070 // any reference to the allocation, as we've already
5071 // observed, then there can be no objection to the trap.
5072 bool found_trap = false;
5073 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5074 Node* obs = not_ctl->fast_out(j);
5075 if (obs->in(0) == not_ctl && obs->is_Call() &&
5076 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5077 found_trap = true; break;
5078 }
5079 }
5080 if (found_trap) {
5081 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5082 continue;
5083 }
5084 }
5085 return NULL;
5086 }
5087
5088 // If we get this far, we have an allocation which immediately
5089 // precedes the arraycopy, and we can take over zeroing the new object.
5090 // The arraycopy will finish the initialization, and provide
5091 // a new control state to which we will anchor the destination pointer.
5092
5093 return alloc;
5094 }
5095
5096 // Helper for initialization of arrays, creating a ClearArray.
5097 // It writes zero bits in [start..end), within the body of an array object.
5098 // The memory effects are all chained onto the 'adr_type' alias category.
5099 //
5100 // Since the object is otherwise uninitialized, we are free
5101 // to put a little "slop" around the edges of the cleared area,
5102 // as long as it does not go back into the array's header,
5103 // or beyond the array end within the heap.
5104 //
5105 // The lower edge can be rounded down to the nearest jint and the
5106 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5107 //
5108 // Arguments:
5109 // adr_type memory slice where writes are generated
5110 // dest oop of the destination array
5111 // basic_elem_type element type of the destination
5112 // slice_idx array index of first element to store
5113 // slice_len number of elements to store (or NULL)
5114 // dest_size total size in bytes of the array object
5115 //
5116 // Exactly one of slice_len or dest_size must be non-NULL.
5117 // If dest_size is non-NULL, zeroing extends to the end of the object.
5118 // If slice_len is non-NULL, the slice_idx value must be a constant.
5119 void
5120 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5121 Node* dest,
5122 BasicType basic_elem_type,
5123 Node* slice_idx,
5124 Node* slice_len,
5125 Node* dest_size) {
5126 // one or the other but not both of slice_len and dest_size:
5127 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5128 if (slice_len == NULL) slice_len = top();
5129 if (dest_size == NULL) dest_size = top();
5130
5131 // operate on this memory slice:
5132 Node* mem = memory(adr_type); // memory slice to operate on
5133
5134 // scaling and rounding of indexes:
5135 int scale = exact_log2(type2aelembytes(basic_elem_type));
5136 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5137 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5138 int bump_bit = (-1 << scale) & BytesPerInt;
5139
5140 // determine constant starts and ends
5141 const intptr_t BIG_NEG = -128;
5142 assert(BIG_NEG + 2*abase < 0, "neg enough");
5143 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5144 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5145 if (slice_len_con == 0) {
5146 return; // nothing to do here
5147 }
5148 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5149 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5150 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5151 assert(end_con < 0, "not two cons");
5152 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5153 BytesPerLong);
5154 }
5155
5156 if (start_con >= 0 && end_con >= 0) {
5157 // Constant start and end. Simple.
5158 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5159 start_con, end_con, &_gvn);
5160 } else if (start_con >= 0 && dest_size != top()) {
5161 // Constant start, pre-rounded end after the tail of the array.
5162 Node* end = dest_size;
5163 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5164 start_con, end, &_gvn);
5165 } else if (start_con >= 0 && slice_len != top()) {
5166 // Constant start, non-constant end. End needs rounding up.
5167 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5168 intptr_t end_base = abase + (slice_idx_con << scale);
5169 int end_round = (-1 << scale) & (BytesPerLong - 1);
5170 Node* end = ConvI2X(slice_len);
5171 if (scale != 0)
5172 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5173 end_base += end_round;
5174 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5175 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5176 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5177 start_con, end, &_gvn);
5178 } else if (start_con < 0 && dest_size != top()) {
5179 // Non-constant start, pre-rounded end after the tail of the array.
5180 // This is almost certainly a "round-to-end" operation.
5181 Node* start = slice_idx;
5182 start = ConvI2X(start);
5183 if (scale != 0)
5184 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5185 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5186 if ((bump_bit | clear_low) != 0) {
5187 int to_clear = (bump_bit | clear_low);
5188 // Align up mod 8, then store a jint zero unconditionally
5189 // just before the mod-8 boundary.
5190 if (((abase + bump_bit) & ~to_clear) - bump_bit
5191 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5192 bump_bit = 0;
5193 assert((abase & to_clear) == 0, "array base must be long-aligned");
5194 } else {
5195 // Bump 'start' up to (or past) the next jint boundary:
5196 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5197 assert((abase & clear_low) == 0, "array base must be int-aligned");
5198 }
5199 // Round bumped 'start' down to jlong boundary in body of array.
5200 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5201 if (bump_bit != 0) {
5202 // Store a zero to the immediately preceding jint:
5203 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5204 Node* p1 = basic_plus_adr(dest, x1);
5205 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5206 mem = _gvn.transform(mem);
5207 }
5208 }
5209 Node* end = dest_size; // pre-rounded
5210 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5211 start, end, &_gvn);
5212 } else {
5213 // Non-constant start, unrounded non-constant end.
5214 // (Nobody zeroes a random midsection of an array using this routine.)
5215 ShouldNotReachHere(); // fix caller
5216 }
5217
5218 // Done.
5219 set_memory(mem, adr_type);
5220 }
5221
5222
5223 bool
5224 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5225 BasicType basic_elem_type,
5226 AllocateNode* alloc,
5227 Node* src, Node* src_offset,
5228 Node* dest, Node* dest_offset,
5229 Node* dest_size, bool dest_uninitialized) {
5230 // See if there is an advantage from block transfer.
5231 int scale = exact_log2(type2aelembytes(basic_elem_type));
5232 if (scale >= LogBytesPerLong)
5233 return false; // it is already a block transfer
5234
5235 // Look at the alignment of the starting offsets.
5236 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5237
5238 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5239 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5240 if (src_off_con < 0 || dest_off_con < 0)
5241 // At present, we can only understand constants.
5242 return false;
5243
5244 intptr_t src_off = abase + (src_off_con << scale);
5245 intptr_t dest_off = abase + (dest_off_con << scale);
5246
5247 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5248 // Non-aligned; too bad.
5249 // One more chance: Pick off an initial 32-bit word.
5250 // This is a common case, since abase can be odd mod 8.
5251 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5252 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5253 Node* sptr = basic_plus_adr(src, src_off);
5254 Node* dptr = basic_plus_adr(dest, dest_off);
5255 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5256 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5257 src_off += BytesPerInt;
5258 dest_off += BytesPerInt;
5259 } else {
5260 return false;
5261 }
5262 }
5263 assert(src_off % BytesPerLong == 0, "");
5264 assert(dest_off % BytesPerLong == 0, "");
5265
5266 // Do this copy by giant steps.
5267 Node* sptr = basic_plus_adr(src, src_off);
5268 Node* dptr = basic_plus_adr(dest, dest_off);
5269 Node* countx = dest_size;
5270 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5271 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5272
5273 bool disjoint_bases = true; // since alloc != NULL
5274 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5275 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5276
5277 return true;
5278 }
5279
5280
5281 // Helper function; generates code for the slow case.
5282 // We make a call to a runtime method which emulates the native method,
5283 // but without the native wrapper overhead.
5284 void
5285 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5286 Node* src, Node* src_offset,
5287 Node* dest, Node* dest_offset,
5288 Node* copy_length, bool dest_uninitialized) {
5289 assert(!dest_uninitialized, "Invariant");
5290 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5291 OptoRuntime::slow_arraycopy_Type(),
5292 OptoRuntime::slow_arraycopy_Java(),
5293 "slow_arraycopy", adr_type,
5294 src, src_offset, dest, dest_offset,
5295 copy_length);
5296
5297 // Handle exceptions thrown by this fellow:
5298 make_slow_call_ex(call, env()->Throwable_klass(), false);
5299 }
5300
5301 // Helper function; generates code for cases requiring runtime checks.
5302 Node*
5303 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5304 Node* dest_elem_klass,
5305 Node* src, Node* src_offset,
5306 Node* dest, Node* dest_offset,
5307 Node* copy_length, bool dest_uninitialized) {
5308 if (stopped()) return NULL;
5309
5310 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5311 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5312 return NULL;
5313 }
5314
5315 // Pick out the parameters required to perform a store-check
5316 // for the target array. This is an optimistic check. It will
5317 // look in each non-null element's class, at the desired klass's
5318 // super_check_offset, for the desired klass.
5319 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5320 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5321 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5322 Node* check_offset = ConvI2X(_gvn.transform(n3));
5323 Node* check_value = dest_elem_klass;
5324
5325 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5326 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5327
5328 // (We know the arrays are never conjoint, because their types differ.)
5329 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5330 OptoRuntime::checkcast_arraycopy_Type(),
5331 copyfunc_addr, "checkcast_arraycopy", adr_type,
5332 // five arguments, of which two are
5333 // intptr_t (jlong in LP64)
5334 src_start, dest_start,
5335 copy_length XTOP,
5336 check_offset XTOP,
5337 check_value);
5338
5339 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5340 }
5341
5342
5343 // Helper function; generates code for cases requiring runtime checks.
5344 Node*
5345 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5346 Node* src, Node* src_offset,
5347 Node* dest, Node* dest_offset,
5348 Node* copy_length, bool dest_uninitialized) {
5349 assert(!dest_uninitialized, "Invariant");
5350 if (stopped()) return NULL;
5351 address copyfunc_addr = StubRoutines::generic_arraycopy();
5352 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5353 return NULL;
5354 }
5355
5356 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5357 OptoRuntime::generic_arraycopy_Type(),
5358 copyfunc_addr, "generic_arraycopy", adr_type,
5359 src, src_offset, dest, dest_offset, copy_length);
5360
5361 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5362 }
5363
5364 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5365 void
5366 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5367 BasicType basic_elem_type,
5368 bool disjoint_bases,
5369 Node* src, Node* src_offset,
5370 Node* dest, Node* dest_offset,
5371 Node* copy_length, bool dest_uninitialized) {
5372 if (stopped()) return; // nothing to do
5373
5374 Node* src_start = src;
5375 Node* dest_start = dest;
5376 if (src_offset != NULL || dest_offset != NULL) {
5377 assert(src_offset != NULL && dest_offset != NULL, "");
5378 src_start = array_element_address(src, src_offset, basic_elem_type);
5379 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5380 }
5381
5382 // Figure out which arraycopy runtime method to call.
5383 const char* copyfunc_name = "arraycopy";
5384 address copyfunc_addr =
5385 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5386 disjoint_bases, copyfunc_name, dest_uninitialized);
5387
5388 // Call it. Note that the count_ix value is not scaled to a byte-size.
5389 make_runtime_call(RC_LEAF|RC_NO_FP,
5390 OptoRuntime::fast_arraycopy_Type(),
5391 copyfunc_addr, copyfunc_name, adr_type,
5392 src_start, dest_start, copy_length XTOP);
5393 }
5394
5395 //-------------inline_encodeISOArray-----------------------------------
5396 // encode char[] to byte[] in ISO_8859_1
5397 bool LibraryCallKit::inline_encodeISOArray() {
5398 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5399 // no receiver since it is static method
5400 Node *src = argument(0);
5401 Node *src_offset = argument(1);
5402 Node *dst = argument(2);
5403 Node *dst_offset = argument(3);
5404 Node *length = argument(4);
5405
5406 const Type* src_type = src->Value(&_gvn);
5407 const Type* dst_type = dst->Value(&_gvn);
5408 const TypeAryPtr* top_src = src_type->isa_aryptr();
5409 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5410 if (top_src == NULL || top_src->klass() == NULL ||
5411 top_dest == NULL || top_dest->klass() == NULL) {
5412 // failed array check
5413 return false;
5414 }
5415
5416 // Figure out the size and type of the elements we will be copying.
5417 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5418 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5419 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5420 return false;
5421 }
5422 Node* src_start = array_element_address(src, src_offset, src_elem);
5423 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5424 // 'src_start' points to src array + scaled offset
5425 // 'dst_start' points to dst array + scaled offset
5426
5427 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5428 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5429 enc = _gvn.transform(enc);
5430 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5431 set_memory(res_mem, mtype);
5432 set_result(enc);
5433 return true;
5434 }
5435
5436 /**
5437 * Calculate CRC32 for byte.
5438 * int java.util.zip.CRC32.update(int crc, int b)
5439 */
5440 bool LibraryCallKit::inline_updateCRC32() {
5441 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5442 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5443 // no receiver since it is static method
5444 Node* crc = argument(0); // type: int
5445 Node* b = argument(1); // type: int
5446
5447 /*
5448 * int c = ~ crc;
5449 * b = timesXtoThe32[(b ^ c) & 0xFF];
5450 * b = b ^ (c >>> 8);
5451 * crc = ~b;
5452 */
5453
5454 Node* M1 = intcon(-1);
5455 crc = _gvn.transform(new (C) XorINode(crc, M1));
5456 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5457 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5458
5459 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5460 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5461 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5462 result = make_load(control(), adr, TypeInt::INT, T_INT);
5463
5464 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5465 result = _gvn.transform(new (C) XorINode(crc, result));
5466 result = _gvn.transform(new (C) XorINode(result, M1));
5467 set_result(result);
5468 return true;
5469 }
5470
5471 /**
5472 * Calculate CRC32 for byte[] array.
5473 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5474 */
5475 bool LibraryCallKit::inline_updateBytesCRC32() {
5476 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5477 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5478 // no receiver since it is static method
5479 Node* crc = argument(0); // type: int
5480 Node* src = argument(1); // type: oop
5481 Node* offset = argument(2); // type: int
5482 Node* length = argument(3); // type: int
5483
5484 const Type* src_type = src->Value(&_gvn);
5485 const TypeAryPtr* top_src = src_type->isa_aryptr();
5486 if (top_src == NULL || top_src->klass() == NULL) {
5487 // failed array check
5488 return false;
5489 }
5490
5491 // Figure out the size and type of the elements we will be copying.
5492 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5493 if (src_elem != T_BYTE) {
5494 return false;
5495 }
5496
5497 // 'src_start' points to src array + scaled offset
5498 Node* src_start = array_element_address(src, offset, src_elem);
5499
5500 // We assume that range check is done by caller.
5501 // TODO: generate range check (offset+length < src.length) in debug VM.
5502
5503 // Call the stub.
5504 address stubAddr = StubRoutines::updateBytesCRC32();
5505 const char *stubName = "updateBytesCRC32";
5506
5507 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5508 stubAddr, stubName, TypePtr::BOTTOM,
5509 crc, src_start, length);
5510 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5511 set_result(result);
5512 return true;
5513 }
5514
5515 /**
5516 * Calculate CRC32 for ByteBuffer.
5517 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5518 */
5519 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5520 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5521 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5522 // no receiver since it is static method
5523 Node* crc = argument(0); // type: int
5524 Node* src = argument(1); // type: long
5525 Node* offset = argument(3); // type: int
5526 Node* length = argument(4); // type: int
5527
5528 src = ConvL2X(src); // adjust Java long to machine word
5529 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5530 offset = ConvI2X(offset);
5531
5532 // 'src_start' points to src array + scaled offset
5533 Node* src_start = basic_plus_adr(top(), base, offset);
5534
5535 // Call the stub.
5536 address stubAddr = StubRoutines::updateBytesCRC32();
5537 const char *stubName = "updateBytesCRC32";
5538
5539 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5540 stubAddr, stubName, TypePtr::BOTTOM,
5541 crc, src_start, length);
5542 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5543 set_result(result);
5544 return true;
5545 }
5546
5547 //----------------------------inline_reference_get----------------------------
5548 // public T java.lang.ref.Reference.get();
5549 bool LibraryCallKit::inline_reference_get() {
5550 const int referent_offset = java_lang_ref_Reference::referent_offset;
5551 guarantee(referent_offset > 0, "should have already been set");
5552
5553 // Get the argument:
5554 Node* reference_obj = null_check_receiver();
5555 if (stopped()) return true;
5556
5557 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5558
5559 ciInstanceKlass* klass = env()->Object_klass();
5560 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5561
5562 Node* no_ctrl = NULL;
5563 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5564
5565 // Use the pre-barrier to record the value in the referent field
5566 pre_barrier(false /* do_load */,
5567 control(),
5568 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5569 result /* pre_val */,
5570 T_OBJECT);
5571
5572 // Add memory barrier to prevent commoning reads from this field
5573 // across safepoint since GC can change its value.
5574 insert_mem_bar(Op_MemBarCPUOrder);
5575
5576 set_result(result);
5577 return true;
5578 }
5579
5580
5581 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5582 bool is_exact=true, bool is_static=false) {
5583
5584 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5585 assert(tinst != NULL, "obj is null");
5586 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5587 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5588
5589 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5590 ciSymbol::make(fieldTypeString),
5591 is_static);
5592 if (field == NULL) return (Node *) NULL;
5593 assert (field != NULL, "undefined field");
5594
5595 // Next code copied from Parse::do_get_xxx():
5596
5597 // Compute address and memory type.
5598 int offset = field->offset_in_bytes();
5599 bool is_vol = field->is_volatile();
5600 ciType* field_klass = field->type();
5601 assert(field_klass->is_loaded(), "should be loaded");
5602 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5603 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5604 BasicType bt = field->layout_type();
5605
5606 // Build the resultant type of the load
5607 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5608
5609 // Build the load.
5610 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5611 return loadedField;
5612 }
5613
5614
5615 //------------------------------inline_aescrypt_Block-----------------------
5616 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5617 address stubAddr;
5618 const char *stubName;
5619 assert(UseAES, "need AES instruction support");
5620
5621 switch(id) {
5622 case vmIntrinsics::_aescrypt_encryptBlock:
5623 stubAddr = StubRoutines::aescrypt_encryptBlock();
5624 stubName = "aescrypt_encryptBlock";
5625 break;
5626 case vmIntrinsics::_aescrypt_decryptBlock:
5627 stubAddr = StubRoutines::aescrypt_decryptBlock();
5628 stubName = "aescrypt_decryptBlock";
5629 break;
5630 }
5631 if (stubAddr == NULL) return false;
5632
5633 Node* aescrypt_object = argument(0);
5634 Node* src = argument(1);
5635 Node* src_offset = argument(2);
5636 Node* dest = argument(3);
5637 Node* dest_offset = argument(4);
5638
5639 // (1) src and dest are arrays.
5640 const Type* src_type = src->Value(&_gvn);
5641 const Type* dest_type = dest->Value(&_gvn);
5642 const TypeAryPtr* top_src = src_type->isa_aryptr();
5643 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5644 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5645
5646 // for the quick and dirty code we will skip all the checks.
5647 // we are just trying to get the call to be generated.
5648 Node* src_start = src;
5649 Node* dest_start = dest;
5650 if (src_offset != NULL || dest_offset != NULL) {
5651 assert(src_offset != NULL && dest_offset != NULL, "");
5652 src_start = array_element_address(src, src_offset, T_BYTE);
5653 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5654 }
5655
5656 // now need to get the start of its expanded key array
5657 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5658 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5659 if (k_start == NULL) return false;
5660
5661 // Call the stub.
5662 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5663 stubAddr, stubName, TypePtr::BOTTOM,
5664 src_start, dest_start, k_start);
5665
5666 return true;
5667 }
5668
5669 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5670 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5671 address stubAddr;
5672 const char *stubName;
5673
5674 assert(UseAES, "need AES instruction support");
5675
5676 switch(id) {
5677 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5678 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5679 stubName = "cipherBlockChaining_encryptAESCrypt";
5680 break;
5681 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5682 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5683 stubName = "cipherBlockChaining_decryptAESCrypt";
5684 break;
5685 }
5686 if (stubAddr == NULL) return false;
5687
5688 Node* cipherBlockChaining_object = argument(0);
5689 Node* src = argument(1);
5690 Node* src_offset = argument(2);
5691 Node* len = argument(3);
5692 Node* dest = argument(4);
5693 Node* dest_offset = argument(5);
5694
5695 // (1) src and dest are arrays.
5696 const Type* src_type = src->Value(&_gvn);
5697 const Type* dest_type = dest->Value(&_gvn);
5698 const TypeAryPtr* top_src = src_type->isa_aryptr();
5699 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5700 assert (top_src != NULL && top_src->klass() != NULL
5701 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5702
5703 // checks are the responsibility of the caller
5704 Node* src_start = src;
5705 Node* dest_start = dest;
5706 if (src_offset != NULL || dest_offset != NULL) {
5707 assert(src_offset != NULL && dest_offset != NULL, "");
5708 src_start = array_element_address(src, src_offset, T_BYTE);
5709 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5710 }
5711
5712 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5713 // (because of the predicated logic executed earlier).
5714 // so we cast it here safely.
5715 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5716
5717 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5718 if (embeddedCipherObj == NULL) return false;
5719
5720 // cast it to what we know it will be at runtime
5721 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5722 assert(tinst != NULL, "CBC obj is null");
5723 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5724 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5725 if (!klass_AESCrypt->is_loaded()) return false;
5726
5727 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5728 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5729 const TypeOopPtr* xtype = aklass->as_instance_type();
5730 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5731 aescrypt_object = _gvn.transform(aescrypt_object);
5732
5733 // we need to get the start of the aescrypt_object's expanded key array
5734 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5735 if (k_start == NULL) return false;
5736
5737 // similarly, get the start address of the r vector
5738 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5739 if (objRvec == NULL) return false;
5740 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5741
5742 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5743 make_runtime_call(RC_LEAF|RC_NO_FP,
5744 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5745 stubAddr, stubName, TypePtr::BOTTOM,
5746 src_start, dest_start, k_start, r_start, len);
5747
5748 // return is void so no result needs to be pushed
5749
5750 return true;
5751 }
5752
5753 //------------------------------get_key_start_from_aescrypt_object-----------------------
5754 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5755 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5756 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5757 if (objAESCryptKey == NULL) return (Node *) NULL;
5758
5759 // now have the array, need to get the start address of the K array
5760 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5761 return k_start;
5762 }
5763
5764 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5765 // Return node representing slow path of predicate check.
5766 // the pseudo code we want to emulate with this predicate is:
5767 // for encryption:
5768 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5769 // for decryption:
5770 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5771 // note cipher==plain is more conservative than the original java code but that's OK
5772 //
5773 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5774 // First, check receiver for NULL since it is virtual method.
5775 Node* objCBC = argument(0);
5776 objCBC = null_check(objCBC);
5777
5778 if (stopped()) return NULL; // Always NULL
5779
5780 // Load embeddedCipher field of CipherBlockChaining object.
5781 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5782
5783 // get AESCrypt klass for instanceOf check
5784 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5785 // will have same classloader as CipherBlockChaining object
5786 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5787 assert(tinst != NULL, "CBCobj is null");
5788 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5789
5790 // we want to do an instanceof comparison against the AESCrypt class
5791 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5792 if (!klass_AESCrypt->is_loaded()) {
5793 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5794 Node* ctrl = control();
5795 set_control(top()); // no regular fast path
5796 return ctrl;
5797 }
5798 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5799
5800 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5801 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5802 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5803
5804 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5805
5806 // for encryption, we are done
5807 if (!decrypting)
5808 return instof_false; // even if it is NULL
5809
5810 // for decryption, we need to add a further check to avoid
5811 // taking the intrinsic path when cipher and plain are the same
5812 // see the original java code for why.
5813 RegionNode* region = new(C) RegionNode(3);
5814 region->init_req(1, instof_false);
5815 Node* src = argument(1);
5816 Node* dest = argument(4);
5817 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5818 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5819 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5820 region->init_req(2, src_dest_conjoint);
5821
5822 record_for_igvn(region);
5823 return _gvn.transform(region);
5824 }