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