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