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