Thu, 20 Feb 2014 11:05:12 +0100
8035394: PPC64: Make usage of intrinsic dsqrt depend on processor recognition.
Reviewed-by: kvn
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 */
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/mathexactnode.hpp"
36 #include "opto/mulnode.hpp"
37 #include "opto/parse.hpp"
38 #include "opto/runtime.hpp"
39 #include "opto/subnode.hpp"
40 #include "prims/nativeLookup.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "trace/traceMacros.hpp"
44 class LibraryIntrinsic : public InlineCallGenerator {
45 // Extend the set of intrinsics known to the runtime:
46 public:
47 private:
48 bool _is_virtual;
49 bool _is_predicted;
50 bool _does_virtual_dispatch;
51 vmIntrinsics::ID _intrinsic_id;
53 public:
54 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, bool does_virtual_dispatch, vmIntrinsics::ID id)
55 : InlineCallGenerator(m),
56 _is_virtual(is_virtual),
57 _is_predicted(is_predicted),
58 _does_virtual_dispatch(does_virtual_dispatch),
59 _intrinsic_id(id)
60 {
61 }
62 virtual bool is_intrinsic() const { return true; }
63 virtual bool is_virtual() const { return _is_virtual; }
64 virtual bool is_predicted() const { return _is_predicted; }
65 virtual bool does_virtual_dispatch() const { return _does_virtual_dispatch; }
66 virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
67 virtual Node* generate_predicate(JVMState* jvms);
68 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
69 };
72 // Local helper class for LibraryIntrinsic:
73 class LibraryCallKit : public GraphKit {
74 private:
75 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
76 Node* _result; // the result node, if any
77 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
79 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
81 public:
82 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
83 : GraphKit(jvms),
84 _intrinsic(intrinsic),
85 _result(NULL)
86 {
87 // Check if this is a root compile. In that case we don't have a caller.
88 if (!jvms->has_method()) {
89 _reexecute_sp = sp();
90 } else {
91 // Find out how many arguments the interpreter needs when deoptimizing
92 // and save the stack pointer value so it can used by uncommon_trap.
93 // We find the argument count by looking at the declared signature.
94 bool ignored_will_link;
95 ciSignature* declared_signature = NULL;
96 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
97 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
98 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
99 }
100 }
102 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
104 ciMethod* caller() const { return jvms()->method(); }
105 int bci() const { return jvms()->bci(); }
106 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
107 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
108 ciMethod* callee() const { return _intrinsic->method(); }
110 bool try_to_inline();
111 Node* try_to_predicate();
113 void push_result() {
114 // Push the result onto the stack.
115 if (!stopped() && result() != NULL) {
116 BasicType bt = result()->bottom_type()->basic_type();
117 push_node(bt, result());
118 }
119 }
121 private:
122 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
123 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
124 }
126 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
127 void set_result(RegionNode* region, PhiNode* value);
128 Node* result() { return _result; }
130 virtual int reexecute_sp() { return _reexecute_sp; }
132 // Helper functions to inline natives
133 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
134 Node* generate_slow_guard(Node* test, RegionNode* region);
135 Node* generate_fair_guard(Node* test, RegionNode* region);
136 Node* generate_negative_guard(Node* index, RegionNode* region,
137 // resulting CastII of index:
138 Node* *pos_index = NULL);
139 Node* generate_nonpositive_guard(Node* index, bool never_negative,
140 // resulting CastII of index:
141 Node* *pos_index = NULL);
142 Node* generate_limit_guard(Node* offset, Node* subseq_length,
143 Node* array_length,
144 RegionNode* region);
145 Node* generate_current_thread(Node* &tls_output);
146 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
147 bool disjoint_bases, const char* &name, bool dest_uninitialized);
148 Node* load_mirror_from_klass(Node* klass);
149 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
150 RegionNode* region, int null_path,
151 int offset);
152 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
153 RegionNode* region, int null_path) {
154 int offset = java_lang_Class::klass_offset_in_bytes();
155 return load_klass_from_mirror_common(mirror, never_see_null,
156 region, null_path,
157 offset);
158 }
159 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
160 RegionNode* region, int null_path) {
161 int offset = java_lang_Class::array_klass_offset_in_bytes();
162 return load_klass_from_mirror_common(mirror, never_see_null,
163 region, null_path,
164 offset);
165 }
166 Node* generate_access_flags_guard(Node* kls,
167 int modifier_mask, int modifier_bits,
168 RegionNode* region);
169 Node* generate_interface_guard(Node* kls, RegionNode* region);
170 Node* generate_array_guard(Node* kls, RegionNode* region) {
171 return generate_array_guard_common(kls, region, false, false);
172 }
173 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
174 return generate_array_guard_common(kls, region, false, true);
175 }
176 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
177 return generate_array_guard_common(kls, region, true, false);
178 }
179 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
180 return generate_array_guard_common(kls, region, true, true);
181 }
182 Node* generate_array_guard_common(Node* kls, RegionNode* region,
183 bool obj_array, bool not_array);
184 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
185 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
186 bool is_virtual = false, bool is_static = false);
187 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
188 return generate_method_call(method_id, false, true);
189 }
190 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
191 return generate_method_call(method_id, true, false);
192 }
193 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
195 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
196 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
197 bool inline_string_compareTo();
198 bool inline_string_indexOf();
199 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
200 bool inline_string_equals();
201 Node* round_double_node(Node* n);
202 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
203 bool inline_math_native(vmIntrinsics::ID id);
204 bool inline_trig(vmIntrinsics::ID id);
205 bool inline_math(vmIntrinsics::ID id);
206 void inline_math_mathExact(Node* math);
207 bool inline_math_addExactI(bool is_increment);
208 bool inline_math_addExactL(bool is_increment);
209 bool inline_math_multiplyExactI();
210 bool inline_math_multiplyExactL();
211 bool inline_math_negateExactI();
212 bool inline_math_negateExactL();
213 bool inline_math_subtractExactI(bool is_decrement);
214 bool inline_math_subtractExactL(bool is_decrement);
215 bool inline_exp();
216 bool inline_pow();
217 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
218 bool inline_min_max(vmIntrinsics::ID id);
219 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
220 // This returns Type::AnyPtr, RawPtr, or OopPtr.
221 int classify_unsafe_addr(Node* &base, Node* &offset);
222 Node* make_unsafe_address(Node* base, Node* offset);
223 // Helper for inline_unsafe_access.
224 // Generates the guards that check whether the result of
225 // Unsafe.getObject should be recorded in an SATB log buffer.
226 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
227 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
228 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
229 static bool klass_needs_init_guard(Node* kls);
230 bool inline_unsafe_allocate();
231 bool inline_unsafe_copyMemory();
232 bool inline_native_currentThread();
233 #ifdef TRACE_HAVE_INTRINSICS
234 bool inline_native_classID();
235 bool inline_native_threadID();
236 #endif
237 bool inline_native_time_funcs(address method, const char* funcName);
238 bool inline_native_isInterrupted();
239 bool inline_native_Class_query(vmIntrinsics::ID id);
240 bool inline_native_subtype_check();
242 bool inline_native_newArray();
243 bool inline_native_getLength();
244 bool inline_array_copyOf(bool is_copyOfRange);
245 bool inline_array_equals();
246 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
247 bool inline_native_clone(bool is_virtual);
248 bool inline_native_Reflection_getCallerClass();
249 // Helper function for inlining native object hash method
250 bool inline_native_hashcode(bool is_virtual, bool is_static);
251 bool inline_native_getClass();
253 // Helper functions for inlining arraycopy
254 bool inline_arraycopy();
255 void generate_arraycopy(const TypePtr* adr_type,
256 BasicType basic_elem_type,
257 Node* src, Node* src_offset,
258 Node* dest, Node* dest_offset,
259 Node* copy_length,
260 bool disjoint_bases = false,
261 bool length_never_negative = false,
262 RegionNode* slow_region = NULL);
263 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
264 RegionNode* slow_region);
265 void generate_clear_array(const TypePtr* adr_type,
266 Node* dest,
267 BasicType basic_elem_type,
268 Node* slice_off,
269 Node* slice_len,
270 Node* slice_end);
271 bool generate_block_arraycopy(const TypePtr* adr_type,
272 BasicType basic_elem_type,
273 AllocateNode* alloc,
274 Node* src, Node* src_offset,
275 Node* dest, Node* dest_offset,
276 Node* dest_size, bool dest_uninitialized);
277 void generate_slow_arraycopy(const TypePtr* adr_type,
278 Node* src, Node* src_offset,
279 Node* dest, Node* dest_offset,
280 Node* copy_length, bool dest_uninitialized);
281 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
282 Node* dest_elem_klass,
283 Node* src, Node* src_offset,
284 Node* dest, Node* dest_offset,
285 Node* copy_length, bool dest_uninitialized);
286 Node* generate_generic_arraycopy(const TypePtr* adr_type,
287 Node* src, Node* src_offset,
288 Node* dest, Node* dest_offset,
289 Node* copy_length, bool dest_uninitialized);
290 void generate_unchecked_arraycopy(const TypePtr* adr_type,
291 BasicType basic_elem_type,
292 bool disjoint_bases,
293 Node* src, Node* src_offset,
294 Node* dest, Node* dest_offset,
295 Node* copy_length, bool dest_uninitialized);
296 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
297 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
298 bool inline_unsafe_ordered_store(BasicType type);
299 bool inline_unsafe_fence(vmIntrinsics::ID id);
300 bool inline_fp_conversions(vmIntrinsics::ID id);
301 bool inline_number_methods(vmIntrinsics::ID id);
302 bool inline_reference_get();
303 bool inline_aescrypt_Block(vmIntrinsics::ID id);
304 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
305 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
306 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
307 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
308 bool inline_encodeISOArray();
309 bool inline_updateCRC32();
310 bool inline_updateBytesCRC32();
311 bool inline_updateByteBufferCRC32();
312 };
315 //---------------------------make_vm_intrinsic----------------------------
316 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
317 vmIntrinsics::ID id = m->intrinsic_id();
318 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
320 if (DisableIntrinsic[0] != '\0'
321 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
322 // disabled by a user request on the command line:
323 // example: -XX:DisableIntrinsic=_hashCode,_getClass
324 return NULL;
325 }
327 if (!m->is_loaded()) {
328 // do not attempt to inline unloaded methods
329 return NULL;
330 }
332 // Only a few intrinsics implement a virtual dispatch.
333 // They are expensive calls which are also frequently overridden.
334 if (is_virtual) {
335 switch (id) {
336 case vmIntrinsics::_hashCode:
337 case vmIntrinsics::_clone:
338 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
339 break;
340 default:
341 return NULL;
342 }
343 }
345 // -XX:-InlineNatives disables nearly all intrinsics:
346 if (!InlineNatives) {
347 switch (id) {
348 case vmIntrinsics::_indexOf:
349 case vmIntrinsics::_compareTo:
350 case vmIntrinsics::_equals:
351 case vmIntrinsics::_equalsC:
352 case vmIntrinsics::_getAndAddInt:
353 case vmIntrinsics::_getAndAddLong:
354 case vmIntrinsics::_getAndSetInt:
355 case vmIntrinsics::_getAndSetLong:
356 case vmIntrinsics::_getAndSetObject:
357 case vmIntrinsics::_loadFence:
358 case vmIntrinsics::_storeFence:
359 case vmIntrinsics::_fullFence:
360 break; // InlineNatives does not control String.compareTo
361 case vmIntrinsics::_Reference_get:
362 break; // InlineNatives does not control Reference.get
363 default:
364 return NULL;
365 }
366 }
368 bool is_predicted = false;
369 bool does_virtual_dispatch = false;
371 switch (id) {
372 case vmIntrinsics::_compareTo:
373 if (!SpecialStringCompareTo) return NULL;
374 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
375 break;
376 case vmIntrinsics::_indexOf:
377 if (!SpecialStringIndexOf) return NULL;
378 break;
379 case vmIntrinsics::_equals:
380 if (!SpecialStringEquals) return NULL;
381 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
382 break;
383 case vmIntrinsics::_equalsC:
384 if (!SpecialArraysEquals) return NULL;
385 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
386 break;
387 case vmIntrinsics::_arraycopy:
388 if (!InlineArrayCopy) return NULL;
389 break;
390 case vmIntrinsics::_copyMemory:
391 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
392 if (!InlineArrayCopy) return NULL;
393 break;
394 case vmIntrinsics::_hashCode:
395 if (!InlineObjectHash) return NULL;
396 does_virtual_dispatch = true;
397 break;
398 case vmIntrinsics::_clone:
399 does_virtual_dispatch = true;
400 case vmIntrinsics::_copyOf:
401 case vmIntrinsics::_copyOfRange:
402 if (!InlineObjectCopy) return NULL;
403 // These also use the arraycopy intrinsic mechanism:
404 if (!InlineArrayCopy) return NULL;
405 break;
406 case vmIntrinsics::_encodeISOArray:
407 if (!SpecialEncodeISOArray) return NULL;
408 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
409 break;
410 case vmIntrinsics::_checkIndex:
411 // We do not intrinsify this. The optimizer does fine with it.
412 return NULL;
414 case vmIntrinsics::_getCallerClass:
415 if (!UseNewReflection) return NULL;
416 if (!InlineReflectionGetCallerClass) return NULL;
417 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
418 break;
420 case vmIntrinsics::_bitCount_i:
421 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
422 break;
424 case vmIntrinsics::_bitCount_l:
425 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
426 break;
428 case vmIntrinsics::_numberOfLeadingZeros_i:
429 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
430 break;
432 case vmIntrinsics::_numberOfLeadingZeros_l:
433 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
434 break;
436 case vmIntrinsics::_numberOfTrailingZeros_i:
437 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
438 break;
440 case vmIntrinsics::_numberOfTrailingZeros_l:
441 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
442 break;
444 case vmIntrinsics::_reverseBytes_c:
445 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
446 break;
447 case vmIntrinsics::_reverseBytes_s:
448 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
449 break;
450 case vmIntrinsics::_reverseBytes_i:
451 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
452 break;
453 case vmIntrinsics::_reverseBytes_l:
454 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
455 break;
457 case vmIntrinsics::_Reference_get:
458 // Use the intrinsic version of Reference.get() so that the value in
459 // the referent field can be registered by the G1 pre-barrier code.
460 // Also add memory barrier to prevent commoning reads from this field
461 // across safepoint since GC can change it value.
462 break;
464 case vmIntrinsics::_compareAndSwapObject:
465 #ifdef _LP64
466 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
467 #endif
468 break;
470 case vmIntrinsics::_compareAndSwapLong:
471 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
472 break;
474 case vmIntrinsics::_getAndAddInt:
475 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
476 break;
478 case vmIntrinsics::_getAndAddLong:
479 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
480 break;
482 case vmIntrinsics::_getAndSetInt:
483 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
484 break;
486 case vmIntrinsics::_getAndSetLong:
487 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
488 break;
490 case vmIntrinsics::_getAndSetObject:
491 #ifdef _LP64
492 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
493 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
494 break;
495 #else
496 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
497 break;
498 #endif
500 case vmIntrinsics::_aescrypt_encryptBlock:
501 case vmIntrinsics::_aescrypt_decryptBlock:
502 if (!UseAESIntrinsics) return NULL;
503 break;
505 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
506 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
507 if (!UseAESIntrinsics) return NULL;
508 // these two require the predicated logic
509 is_predicted = true;
510 break;
512 case vmIntrinsics::_updateCRC32:
513 case vmIntrinsics::_updateBytesCRC32:
514 case vmIntrinsics::_updateByteBufferCRC32:
515 if (!UseCRC32Intrinsics) return NULL;
516 break;
518 case vmIntrinsics::_incrementExactI:
519 case vmIntrinsics::_addExactI:
520 if (!Matcher::match_rule_supported(Op_AddExactI) || !UseMathExactIntrinsics) return NULL;
521 break;
522 case vmIntrinsics::_incrementExactL:
523 case vmIntrinsics::_addExactL:
524 if (!Matcher::match_rule_supported(Op_AddExactL) || !UseMathExactIntrinsics) return NULL;
525 break;
526 case vmIntrinsics::_decrementExactI:
527 case vmIntrinsics::_subtractExactI:
528 if (!Matcher::match_rule_supported(Op_SubExactI) || !UseMathExactIntrinsics) return NULL;
529 break;
530 case vmIntrinsics::_decrementExactL:
531 case vmIntrinsics::_subtractExactL:
532 if (!Matcher::match_rule_supported(Op_SubExactL) || !UseMathExactIntrinsics) return NULL;
533 break;
534 case vmIntrinsics::_negateExactI:
535 if (!Matcher::match_rule_supported(Op_NegExactI) || !UseMathExactIntrinsics) return NULL;
536 break;
537 case vmIntrinsics::_negateExactL:
538 if (!Matcher::match_rule_supported(Op_NegExactL) || !UseMathExactIntrinsics) return NULL;
539 break;
540 case vmIntrinsics::_multiplyExactI:
541 if (!Matcher::match_rule_supported(Op_MulExactI) || !UseMathExactIntrinsics) return NULL;
542 break;
543 case vmIntrinsics::_multiplyExactL:
544 if (!Matcher::match_rule_supported(Op_MulExactL) || !UseMathExactIntrinsics) return NULL;
545 break;
547 default:
548 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
549 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
550 break;
551 }
553 // -XX:-InlineClassNatives disables natives from the Class class.
554 // The flag applies to all reflective calls, notably Array.newArray
555 // (visible to Java programmers as Array.newInstance).
556 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
557 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
558 if (!InlineClassNatives) return NULL;
559 }
561 // -XX:-InlineThreadNatives disables natives from the Thread class.
562 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
563 if (!InlineThreadNatives) return NULL;
564 }
566 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
567 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
568 m->holder()->name() == ciSymbol::java_lang_Float() ||
569 m->holder()->name() == ciSymbol::java_lang_Double()) {
570 if (!InlineMathNatives) return NULL;
571 }
573 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
574 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
575 if (!InlineUnsafeOps) return NULL;
576 }
578 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
579 }
581 //----------------------register_library_intrinsics-----------------------
582 // Initialize this file's data structures, for each Compile instance.
583 void Compile::register_library_intrinsics() {
584 // Nothing to do here.
585 }
587 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
588 LibraryCallKit kit(jvms, this);
589 Compile* C = kit.C;
590 int nodes = C->unique();
591 #ifndef PRODUCT
592 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
593 char buf[1000];
594 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
595 tty->print_cr("Intrinsic %s", str);
596 }
597 #endif
598 ciMethod* callee = kit.callee();
599 const int bci = kit.bci();
601 // Try to inline the intrinsic.
602 if (kit.try_to_inline()) {
603 if (C->print_intrinsics() || C->print_inlining()) {
604 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
605 }
606 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
607 if (C->log()) {
608 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
609 vmIntrinsics::name_at(intrinsic_id()),
610 (is_virtual() ? " virtual='1'" : ""),
611 C->unique() - nodes);
612 }
613 // Push the result from the inlined method onto the stack.
614 kit.push_result();
615 return kit.transfer_exceptions_into_jvms();
616 }
618 // The intrinsic bailed out
619 if (C->print_intrinsics() || C->print_inlining()) {
620 if (jvms->has_method()) {
621 // Not a root compile.
622 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
623 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
624 } else {
625 // Root compile
626 tty->print("Did not generate intrinsic %s%s at bci:%d in",
627 vmIntrinsics::name_at(intrinsic_id()),
628 (is_virtual() ? " (virtual)" : ""), bci);
629 }
630 }
631 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
632 return NULL;
633 }
635 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
636 LibraryCallKit kit(jvms, this);
637 Compile* C = kit.C;
638 int nodes = C->unique();
639 #ifndef PRODUCT
640 assert(is_predicted(), "sanity");
641 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
642 char buf[1000];
643 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
644 tty->print_cr("Predicate for intrinsic %s", str);
645 }
646 #endif
647 ciMethod* callee = kit.callee();
648 const int bci = kit.bci();
650 Node* slow_ctl = kit.try_to_predicate();
651 if (!kit.failing()) {
652 if (C->print_intrinsics() || C->print_inlining()) {
653 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
654 }
655 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
656 if (C->log()) {
657 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
658 vmIntrinsics::name_at(intrinsic_id()),
659 (is_virtual() ? " virtual='1'" : ""),
660 C->unique() - nodes);
661 }
662 return slow_ctl; // Could be NULL if the check folds.
663 }
665 // The intrinsic bailed out
666 if (C->print_intrinsics() || C->print_inlining()) {
667 if (jvms->has_method()) {
668 // Not a root compile.
669 const char* msg = "failed to generate predicate for intrinsic";
670 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
671 } else {
672 // Root compile
673 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
674 vmIntrinsics::name_at(intrinsic_id()),
675 (is_virtual() ? " (virtual)" : ""), bci);
676 }
677 }
678 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
679 return NULL;
680 }
682 bool LibraryCallKit::try_to_inline() {
683 // Handle symbolic names for otherwise undistinguished boolean switches:
684 const bool is_store = true;
685 const bool is_native_ptr = true;
686 const bool is_static = true;
687 const bool is_volatile = true;
689 if (!jvms()->has_method()) {
690 // Root JVMState has a null method.
691 assert(map()->memory()->Opcode() == Op_Parm, "");
692 // Insert the memory aliasing node
693 set_all_memory(reset_memory());
694 }
695 assert(merged_memory(), "");
698 switch (intrinsic_id()) {
699 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
700 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
701 case vmIntrinsics::_getClass: return inline_native_getClass();
703 case vmIntrinsics::_dsin:
704 case vmIntrinsics::_dcos:
705 case vmIntrinsics::_dtan:
706 case vmIntrinsics::_dabs:
707 case vmIntrinsics::_datan2:
708 case vmIntrinsics::_dsqrt:
709 case vmIntrinsics::_dexp:
710 case vmIntrinsics::_dlog:
711 case vmIntrinsics::_dlog10:
712 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
714 case vmIntrinsics::_min:
715 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
717 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
718 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
719 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
720 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
721 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
722 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
723 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
724 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
725 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
726 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
727 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
728 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
730 case vmIntrinsics::_arraycopy: return inline_arraycopy();
732 case vmIntrinsics::_compareTo: return inline_string_compareTo();
733 case vmIntrinsics::_indexOf: return inline_string_indexOf();
734 case vmIntrinsics::_equals: return inline_string_equals();
736 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
737 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
738 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
739 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
740 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
741 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
742 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
743 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
744 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
746 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
747 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
748 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
749 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
750 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
751 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
752 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
753 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
754 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
756 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
757 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
758 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
759 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
760 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
761 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
762 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
763 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
765 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
766 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
767 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
768 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
769 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
770 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
771 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
772 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
774 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
775 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
776 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
777 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
778 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
779 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
780 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
781 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
782 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
784 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
785 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
786 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
787 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
788 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
789 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
790 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
791 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
792 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
794 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
795 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
796 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
797 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
799 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
800 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
801 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
803 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
804 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
805 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
807 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
808 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
809 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
810 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
811 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
813 case vmIntrinsics::_loadFence:
814 case vmIntrinsics::_storeFence:
815 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
817 case vmIntrinsics::_currentThread: return inline_native_currentThread();
818 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
820 #ifdef TRACE_HAVE_INTRINSICS
821 case vmIntrinsics::_classID: return inline_native_classID();
822 case vmIntrinsics::_threadID: return inline_native_threadID();
823 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
824 #endif
825 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
826 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
827 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
828 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
829 case vmIntrinsics::_newArray: return inline_native_newArray();
830 case vmIntrinsics::_getLength: return inline_native_getLength();
831 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
832 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
833 case vmIntrinsics::_equalsC: return inline_array_equals();
834 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
836 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
838 case vmIntrinsics::_isInstance:
839 case vmIntrinsics::_getModifiers:
840 case vmIntrinsics::_isInterface:
841 case vmIntrinsics::_isArray:
842 case vmIntrinsics::_isPrimitive:
843 case vmIntrinsics::_getSuperclass:
844 case vmIntrinsics::_getComponentType:
845 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
847 case vmIntrinsics::_floatToRawIntBits:
848 case vmIntrinsics::_floatToIntBits:
849 case vmIntrinsics::_intBitsToFloat:
850 case vmIntrinsics::_doubleToRawLongBits:
851 case vmIntrinsics::_doubleToLongBits:
852 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
854 case vmIntrinsics::_numberOfLeadingZeros_i:
855 case vmIntrinsics::_numberOfLeadingZeros_l:
856 case vmIntrinsics::_numberOfTrailingZeros_i:
857 case vmIntrinsics::_numberOfTrailingZeros_l:
858 case vmIntrinsics::_bitCount_i:
859 case vmIntrinsics::_bitCount_l:
860 case vmIntrinsics::_reverseBytes_i:
861 case vmIntrinsics::_reverseBytes_l:
862 case vmIntrinsics::_reverseBytes_s:
863 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
865 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
867 case vmIntrinsics::_Reference_get: return inline_reference_get();
869 case vmIntrinsics::_aescrypt_encryptBlock:
870 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
872 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
873 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
874 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
876 case vmIntrinsics::_encodeISOArray:
877 return inline_encodeISOArray();
879 case vmIntrinsics::_updateCRC32:
880 return inline_updateCRC32();
881 case vmIntrinsics::_updateBytesCRC32:
882 return inline_updateBytesCRC32();
883 case vmIntrinsics::_updateByteBufferCRC32:
884 return inline_updateByteBufferCRC32();
886 default:
887 // If you get here, it may be that someone has added a new intrinsic
888 // to the list in vmSymbols.hpp without implementing it here.
889 #ifndef PRODUCT
890 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
891 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
892 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
893 }
894 #endif
895 return false;
896 }
897 }
899 Node* LibraryCallKit::try_to_predicate() {
900 if (!jvms()->has_method()) {
901 // Root JVMState has a null method.
902 assert(map()->memory()->Opcode() == Op_Parm, "");
903 // Insert the memory aliasing node
904 set_all_memory(reset_memory());
905 }
906 assert(merged_memory(), "");
908 switch (intrinsic_id()) {
909 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
910 return inline_cipherBlockChaining_AESCrypt_predicate(false);
911 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
912 return inline_cipherBlockChaining_AESCrypt_predicate(true);
914 default:
915 // If you get here, it may be that someone has added a new intrinsic
916 // to the list in vmSymbols.hpp without implementing it here.
917 #ifndef PRODUCT
918 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
919 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
920 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
921 }
922 #endif
923 Node* slow_ctl = control();
924 set_control(top()); // No fast path instrinsic
925 return slow_ctl;
926 }
927 }
929 //------------------------------set_result-------------------------------
930 // Helper function for finishing intrinsics.
931 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
932 record_for_igvn(region);
933 set_control(_gvn.transform(region));
934 set_result( _gvn.transform(value));
935 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
936 }
938 //------------------------------generate_guard---------------------------
939 // Helper function for generating guarded fast-slow graph structures.
940 // The given 'test', if true, guards a slow path. If the test fails
941 // then a fast path can be taken. (We generally hope it fails.)
942 // In all cases, GraphKit::control() is updated to the fast path.
943 // The returned value represents the control for the slow path.
944 // The return value is never 'top'; it is either a valid control
945 // or NULL if it is obvious that the slow path can never be taken.
946 // Also, if region and the slow control are not NULL, the slow edge
947 // is appended to the region.
948 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
949 if (stopped()) {
950 // Already short circuited.
951 return NULL;
952 }
954 // Build an if node and its projections.
955 // If test is true we take the slow path, which we assume is uncommon.
956 if (_gvn.type(test) == TypeInt::ZERO) {
957 // The slow branch is never taken. No need to build this guard.
958 return NULL;
959 }
961 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
963 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
964 if (if_slow == top()) {
965 // The slow branch is never taken. No need to build this guard.
966 return NULL;
967 }
969 if (region != NULL)
970 region->add_req(if_slow);
972 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
973 set_control(if_fast);
975 return if_slow;
976 }
978 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
979 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
980 }
981 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
982 return generate_guard(test, region, PROB_FAIR);
983 }
985 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
986 Node* *pos_index) {
987 if (stopped())
988 return NULL; // already stopped
989 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
990 return NULL; // index is already adequately typed
991 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
992 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
993 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
994 if (is_neg != NULL && pos_index != NULL) {
995 // Emulate effect of Parse::adjust_map_after_if.
996 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
997 ccast->set_req(0, control());
998 (*pos_index) = _gvn.transform(ccast);
999 }
1000 return is_neg;
1001 }
1003 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1004 Node* *pos_index) {
1005 if (stopped())
1006 return NULL; // already stopped
1007 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1008 return NULL; // index is already adequately typed
1009 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1010 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1011 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1012 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1013 if (is_notp != NULL && pos_index != NULL) {
1014 // Emulate effect of Parse::adjust_map_after_if.
1015 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1016 ccast->set_req(0, control());
1017 (*pos_index) = _gvn.transform(ccast);
1018 }
1019 return is_notp;
1020 }
1022 // Make sure that 'position' is a valid limit index, in [0..length].
1023 // There are two equivalent plans for checking this:
1024 // A. (offset + copyLength) unsigned<= arrayLength
1025 // B. offset <= (arrayLength - copyLength)
1026 // We require that all of the values above, except for the sum and
1027 // difference, are already known to be non-negative.
1028 // Plan A is robust in the face of overflow, if offset and copyLength
1029 // are both hugely positive.
1030 //
1031 // Plan B is less direct and intuitive, but it does not overflow at
1032 // all, since the difference of two non-negatives is always
1033 // representable. Whenever Java methods must perform the equivalent
1034 // check they generally use Plan B instead of Plan A.
1035 // For the moment we use Plan A.
1036 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1037 Node* subseq_length,
1038 Node* array_length,
1039 RegionNode* region) {
1040 if (stopped())
1041 return NULL; // already stopped
1042 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1043 if (zero_offset && subseq_length->eqv_uncast(array_length))
1044 return NULL; // common case of whole-array copy
1045 Node* last = subseq_length;
1046 if (!zero_offset) // last += offset
1047 last = _gvn.transform(new (C) AddINode(last, offset));
1048 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1049 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1050 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1051 return is_over;
1052 }
1055 //--------------------------generate_current_thread--------------------
1056 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1057 ciKlass* thread_klass = env()->Thread_klass();
1058 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1059 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1060 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1061 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT, MemNode::unordered);
1062 tls_output = thread;
1063 return threadObj;
1064 }
1067 //------------------------------make_string_method_node------------------------
1068 // Helper method for String intrinsic functions. This version is called
1069 // with str1 and str2 pointing to String object nodes.
1070 //
1071 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1072 Node* no_ctrl = NULL;
1074 // Get start addr of string
1075 Node* str1_value = load_String_value(no_ctrl, str1);
1076 Node* str1_offset = load_String_offset(no_ctrl, str1);
1077 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1079 // Get length of string 1
1080 Node* str1_len = load_String_length(no_ctrl, str1);
1082 Node* str2_value = load_String_value(no_ctrl, str2);
1083 Node* str2_offset = load_String_offset(no_ctrl, str2);
1084 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1086 Node* str2_len = NULL;
1087 Node* result = NULL;
1089 switch (opcode) {
1090 case Op_StrIndexOf:
1091 // Get length of string 2
1092 str2_len = load_String_length(no_ctrl, str2);
1094 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1095 str1_start, str1_len, str2_start, str2_len);
1096 break;
1097 case Op_StrComp:
1098 // Get length of string 2
1099 str2_len = load_String_length(no_ctrl, str2);
1101 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1102 str1_start, str1_len, str2_start, str2_len);
1103 break;
1104 case Op_StrEquals:
1105 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1106 str1_start, str2_start, str1_len);
1107 break;
1108 default:
1109 ShouldNotReachHere();
1110 return NULL;
1111 }
1113 // All these intrinsics have checks.
1114 C->set_has_split_ifs(true); // Has chance for split-if optimization
1116 return _gvn.transform(result);
1117 }
1119 // Helper method for String intrinsic functions. This version is called
1120 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1121 // to Int nodes containing the lenghts of str1 and str2.
1122 //
1123 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1124 Node* result = NULL;
1125 switch (opcode) {
1126 case Op_StrIndexOf:
1127 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1128 str1_start, cnt1, str2_start, cnt2);
1129 break;
1130 case Op_StrComp:
1131 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1132 str1_start, cnt1, str2_start, cnt2);
1133 break;
1134 case Op_StrEquals:
1135 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1136 str1_start, str2_start, cnt1);
1137 break;
1138 default:
1139 ShouldNotReachHere();
1140 return NULL;
1141 }
1143 // All these intrinsics have checks.
1144 C->set_has_split_ifs(true); // Has chance for split-if optimization
1146 return _gvn.transform(result);
1147 }
1149 //------------------------------inline_string_compareTo------------------------
1150 // public int java.lang.String.compareTo(String anotherString);
1151 bool LibraryCallKit::inline_string_compareTo() {
1152 Node* receiver = null_check(argument(0));
1153 Node* arg = null_check(argument(1));
1154 if (stopped()) {
1155 return true;
1156 }
1157 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1158 return true;
1159 }
1161 //------------------------------inline_string_equals------------------------
1162 bool LibraryCallKit::inline_string_equals() {
1163 Node* receiver = null_check_receiver();
1164 // NOTE: Do not null check argument for String.equals() because spec
1165 // allows to specify NULL as argument.
1166 Node* argument = this->argument(1);
1167 if (stopped()) {
1168 return true;
1169 }
1171 // paths (plus control) merge
1172 RegionNode* region = new (C) RegionNode(5);
1173 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1175 // does source == target string?
1176 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1177 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1179 Node* if_eq = generate_slow_guard(bol, NULL);
1180 if (if_eq != NULL) {
1181 // receiver == argument
1182 phi->init_req(2, intcon(1));
1183 region->init_req(2, if_eq);
1184 }
1186 // get String klass for instanceOf
1187 ciInstanceKlass* klass = env()->String_klass();
1189 if (!stopped()) {
1190 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1191 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1192 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1194 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1195 //instanceOf == true, fallthrough
1197 if (inst_false != NULL) {
1198 phi->init_req(3, intcon(0));
1199 region->init_req(3, inst_false);
1200 }
1201 }
1203 if (!stopped()) {
1204 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1206 // Properly cast the argument to String
1207 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1208 // This path is taken only when argument's type is String:NotNull.
1209 argument = cast_not_null(argument, false);
1211 Node* no_ctrl = NULL;
1213 // Get start addr of receiver
1214 Node* receiver_val = load_String_value(no_ctrl, receiver);
1215 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1216 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1218 // Get length of receiver
1219 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1221 // Get start addr of argument
1222 Node* argument_val = load_String_value(no_ctrl, argument);
1223 Node* argument_offset = load_String_offset(no_ctrl, argument);
1224 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1226 // Get length of argument
1227 Node* argument_cnt = load_String_length(no_ctrl, argument);
1229 // Check for receiver count != argument count
1230 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1231 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1232 Node* if_ne = generate_slow_guard(bol, NULL);
1233 if (if_ne != NULL) {
1234 phi->init_req(4, intcon(0));
1235 region->init_req(4, if_ne);
1236 }
1238 // Check for count == 0 is done by assembler code for StrEquals.
1240 if (!stopped()) {
1241 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1242 phi->init_req(1, equals);
1243 region->init_req(1, control());
1244 }
1245 }
1247 // post merge
1248 set_control(_gvn.transform(region));
1249 record_for_igvn(region);
1251 set_result(_gvn.transform(phi));
1252 return true;
1253 }
1255 //------------------------------inline_array_equals----------------------------
1256 bool LibraryCallKit::inline_array_equals() {
1257 Node* arg1 = argument(0);
1258 Node* arg2 = argument(1);
1259 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1260 return true;
1261 }
1263 // Java version of String.indexOf(constant string)
1264 // class StringDecl {
1265 // StringDecl(char[] ca) {
1266 // offset = 0;
1267 // count = ca.length;
1268 // value = ca;
1269 // }
1270 // int offset;
1271 // int count;
1272 // char[] value;
1273 // }
1274 //
1275 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1276 // int targetOffset, int cache_i, int md2) {
1277 // int cache = cache_i;
1278 // int sourceOffset = string_object.offset;
1279 // int sourceCount = string_object.count;
1280 // int targetCount = target_object.length;
1281 //
1282 // int targetCountLess1 = targetCount - 1;
1283 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1284 //
1285 // char[] source = string_object.value;
1286 // char[] target = target_object;
1287 // int lastChar = target[targetCountLess1];
1288 //
1289 // outer_loop:
1290 // for (int i = sourceOffset; i < sourceEnd; ) {
1291 // int src = source[i + targetCountLess1];
1292 // if (src == lastChar) {
1293 // // With random strings and a 4-character alphabet,
1294 // // reverse matching at this point sets up 0.8% fewer
1295 // // frames, but (paradoxically) makes 0.3% more probes.
1296 // // Since those probes are nearer the lastChar probe,
1297 // // there is may be a net D$ win with reverse matching.
1298 // // But, reversing loop inhibits unroll of inner loop
1299 // // for unknown reason. So, does running outer loop from
1300 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1301 // for (int j = 0; j < targetCountLess1; j++) {
1302 // if (target[targetOffset + j] != source[i+j]) {
1303 // if ((cache & (1 << source[i+j])) == 0) {
1304 // if (md2 < j+1) {
1305 // i += j+1;
1306 // continue outer_loop;
1307 // }
1308 // }
1309 // i += md2;
1310 // continue outer_loop;
1311 // }
1312 // }
1313 // return i - sourceOffset;
1314 // }
1315 // if ((cache & (1 << src)) == 0) {
1316 // i += targetCountLess1;
1317 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1318 // i++;
1319 // }
1320 // return -1;
1321 // }
1323 //------------------------------string_indexOf------------------------
1324 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1325 jint cache_i, jint md2_i) {
1327 Node* no_ctrl = NULL;
1328 float likely = PROB_LIKELY(0.9);
1329 float unlikely = PROB_UNLIKELY(0.9);
1331 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1333 Node* source = load_String_value(no_ctrl, string_object);
1334 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1335 Node* sourceCount = load_String_length(no_ctrl, string_object);
1337 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1338 jint target_length = target_array->length();
1339 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1340 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1342 // String.value field is known to be @Stable.
1343 if (UseImplicitStableValues) {
1344 target = cast_array_to_stable(target, target_type);
1345 }
1347 IdealKit kit(this, false, true);
1348 #define __ kit.
1349 Node* zero = __ ConI(0);
1350 Node* one = __ ConI(1);
1351 Node* cache = __ ConI(cache_i);
1352 Node* md2 = __ ConI(md2_i);
1353 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1354 Node* targetCount = __ ConI(target_length);
1355 Node* targetCountLess1 = __ ConI(target_length - 1);
1356 Node* targetOffset = __ ConI(targetOffset_i);
1357 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1359 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1360 Node* outer_loop = __ make_label(2 /* goto */);
1361 Node* return_ = __ make_label(1);
1363 __ set(rtn,__ ConI(-1));
1364 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1365 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1366 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1367 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1368 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1369 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1370 Node* tpj = __ AddI(targetOffset, __ value(j));
1371 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1372 Node* ipj = __ AddI(__ value(i), __ value(j));
1373 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1374 __ if_then(targ, BoolTest::ne, src2); {
1375 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1376 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1377 __ increment(i, __ AddI(__ value(j), one));
1378 __ goto_(outer_loop);
1379 } __ end_if(); __ dead(j);
1380 }__ end_if(); __ dead(j);
1381 __ increment(i, md2);
1382 __ goto_(outer_loop);
1383 }__ end_if();
1384 __ increment(j, one);
1385 }__ end_loop(); __ dead(j);
1386 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1387 __ goto_(return_);
1388 }__ end_if();
1389 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1390 __ increment(i, targetCountLess1);
1391 }__ end_if();
1392 __ increment(i, one);
1393 __ bind(outer_loop);
1394 }__ end_loop(); __ dead(i);
1395 __ bind(return_);
1397 // Final sync IdealKit and GraphKit.
1398 final_sync(kit);
1399 Node* result = __ value(rtn);
1400 #undef __
1401 C->set_has_loops(true);
1402 return result;
1403 }
1405 //------------------------------inline_string_indexOf------------------------
1406 bool LibraryCallKit::inline_string_indexOf() {
1407 Node* receiver = argument(0);
1408 Node* arg = argument(1);
1410 Node* result;
1411 // Disable the use of pcmpestri until it can be guaranteed that
1412 // the load doesn't cross into the uncommited space.
1413 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1414 UseSSE42Intrinsics) {
1415 // Generate SSE4.2 version of indexOf
1416 // We currently only have match rules that use SSE4.2
1418 receiver = null_check(receiver);
1419 arg = null_check(arg);
1420 if (stopped()) {
1421 return true;
1422 }
1424 ciInstanceKlass* str_klass = env()->String_klass();
1425 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1427 // Make the merge point
1428 RegionNode* result_rgn = new (C) RegionNode(4);
1429 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1430 Node* no_ctrl = NULL;
1432 // Get start addr of source string
1433 Node* source = load_String_value(no_ctrl, receiver);
1434 Node* source_offset = load_String_offset(no_ctrl, receiver);
1435 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1437 // Get length of source string
1438 Node* source_cnt = load_String_length(no_ctrl, receiver);
1440 // Get start addr of substring
1441 Node* substr = load_String_value(no_ctrl, arg);
1442 Node* substr_offset = load_String_offset(no_ctrl, arg);
1443 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1445 // Get length of source string
1446 Node* substr_cnt = load_String_length(no_ctrl, arg);
1448 // Check for substr count > string count
1449 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1450 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1451 Node* if_gt = generate_slow_guard(bol, NULL);
1452 if (if_gt != NULL) {
1453 result_phi->init_req(2, intcon(-1));
1454 result_rgn->init_req(2, if_gt);
1455 }
1457 if (!stopped()) {
1458 // Check for substr count == 0
1459 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1460 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1461 Node* if_zero = generate_slow_guard(bol, NULL);
1462 if (if_zero != NULL) {
1463 result_phi->init_req(3, intcon(0));
1464 result_rgn->init_req(3, if_zero);
1465 }
1466 }
1468 if (!stopped()) {
1469 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1470 result_phi->init_req(1, result);
1471 result_rgn->init_req(1, control());
1472 }
1473 set_control(_gvn.transform(result_rgn));
1474 record_for_igvn(result_rgn);
1475 result = _gvn.transform(result_phi);
1477 } else { // Use LibraryCallKit::string_indexOf
1478 // don't intrinsify if argument isn't a constant string.
1479 if (!arg->is_Con()) {
1480 return false;
1481 }
1482 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1483 if (str_type == NULL) {
1484 return false;
1485 }
1486 ciInstanceKlass* klass = env()->String_klass();
1487 ciObject* str_const = str_type->const_oop();
1488 if (str_const == NULL || str_const->klass() != klass) {
1489 return false;
1490 }
1491 ciInstance* str = str_const->as_instance();
1492 assert(str != NULL, "must be instance");
1494 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1495 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1497 int o;
1498 int c;
1499 if (java_lang_String::has_offset_field()) {
1500 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1501 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1502 } else {
1503 o = 0;
1504 c = pat->length();
1505 }
1507 // constant strings have no offset and count == length which
1508 // simplifies the resulting code somewhat so lets optimize for that.
1509 if (o != 0 || c != pat->length()) {
1510 return false;
1511 }
1513 receiver = null_check(receiver, T_OBJECT);
1514 // NOTE: No null check on the argument is needed since it's a constant String oop.
1515 if (stopped()) {
1516 return true;
1517 }
1519 // The null string as a pattern always returns 0 (match at beginning of string)
1520 if (c == 0) {
1521 set_result(intcon(0));
1522 return true;
1523 }
1525 // Generate default indexOf
1526 jchar lastChar = pat->char_at(o + (c - 1));
1527 int cache = 0;
1528 int i;
1529 for (i = 0; i < c - 1; i++) {
1530 assert(i < pat->length(), "out of range");
1531 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1532 }
1534 int md2 = c;
1535 for (i = 0; i < c - 1; i++) {
1536 assert(i < pat->length(), "out of range");
1537 if (pat->char_at(o + i) == lastChar) {
1538 md2 = (c - 1) - i;
1539 }
1540 }
1542 result = string_indexOf(receiver, pat, o, cache, md2);
1543 }
1544 set_result(result);
1545 return true;
1546 }
1548 //--------------------------round_double_node--------------------------------
1549 // Round a double node if necessary.
1550 Node* LibraryCallKit::round_double_node(Node* n) {
1551 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1552 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1553 return n;
1554 }
1556 //------------------------------inline_math-----------------------------------
1557 // public static double Math.abs(double)
1558 // public static double Math.sqrt(double)
1559 // public static double Math.log(double)
1560 // public static double Math.log10(double)
1561 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1562 Node* arg = round_double_node(argument(0));
1563 Node* n;
1564 switch (id) {
1565 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1566 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1567 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1568 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1569 default: fatal_unexpected_iid(id); break;
1570 }
1571 set_result(_gvn.transform(n));
1572 return true;
1573 }
1575 //------------------------------inline_trig----------------------------------
1576 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1577 // argument reduction which will turn into a fast/slow diamond.
1578 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1579 Node* arg = round_double_node(argument(0));
1580 Node* n = NULL;
1582 switch (id) {
1583 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1584 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1585 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1586 default: fatal_unexpected_iid(id); break;
1587 }
1588 n = _gvn.transform(n);
1590 // Rounding required? Check for argument reduction!
1591 if (Matcher::strict_fp_requires_explicit_rounding) {
1592 static const double pi_4 = 0.7853981633974483;
1593 static const double neg_pi_4 = -0.7853981633974483;
1594 // pi/2 in 80-bit extended precision
1595 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1596 // -pi/2 in 80-bit extended precision
1597 // 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};
1598 // Cutoff value for using this argument reduction technique
1599 //static const double pi_2_minus_epsilon = 1.564660403643354;
1600 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1602 // Pseudocode for sin:
1603 // if (x <= Math.PI / 4.0) {
1604 // if (x >= -Math.PI / 4.0) return fsin(x);
1605 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1606 // } else {
1607 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1608 // }
1609 // return StrictMath.sin(x);
1611 // Pseudocode for cos:
1612 // if (x <= Math.PI / 4.0) {
1613 // if (x >= -Math.PI / 4.0) return fcos(x);
1614 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1615 // } else {
1616 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1617 // }
1618 // return StrictMath.cos(x);
1620 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1621 // requires a special machine instruction to load it. Instead we'll try
1622 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1623 // probably do the math inside the SIN encoding.
1625 // Make the merge point
1626 RegionNode* r = new (C) RegionNode(3);
1627 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1629 // Flatten arg so we need only 1 test
1630 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1631 // Node for PI/4 constant
1632 Node *pi4 = makecon(TypeD::make(pi_4));
1633 // Check PI/4 : abs(arg)
1634 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1635 // Check: If PI/4 < abs(arg) then go slow
1636 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1637 // Branch either way
1638 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1639 set_control(opt_iff(r,iff));
1641 // Set fast path result
1642 phi->init_req(2, n);
1644 // Slow path - non-blocking leaf call
1645 Node* call = NULL;
1646 switch (id) {
1647 case vmIntrinsics::_dsin:
1648 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1649 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1650 "Sin", NULL, arg, top());
1651 break;
1652 case vmIntrinsics::_dcos:
1653 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1654 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1655 "Cos", NULL, arg, top());
1656 break;
1657 case vmIntrinsics::_dtan:
1658 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1659 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1660 "Tan", NULL, arg, top());
1661 break;
1662 }
1663 assert(control()->in(0) == call, "");
1664 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1665 r->init_req(1, control());
1666 phi->init_req(1, slow_result);
1668 // Post-merge
1669 set_control(_gvn.transform(r));
1670 record_for_igvn(r);
1671 n = _gvn.transform(phi);
1673 C->set_has_split_ifs(true); // Has chance for split-if optimization
1674 }
1675 set_result(n);
1676 return true;
1677 }
1679 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1680 //-------------------
1681 //result=(result.isNaN())? funcAddr():result;
1682 // Check: If isNaN() by checking result!=result? then either trap
1683 // or go to runtime
1684 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1685 // Build the boolean node
1686 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1688 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1689 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1690 // The pow or exp intrinsic returned a NaN, which requires a call
1691 // to the runtime. Recompile with the runtime call.
1692 uncommon_trap(Deoptimization::Reason_intrinsic,
1693 Deoptimization::Action_make_not_entrant);
1694 }
1695 set_result(result);
1696 } else {
1697 // If this inlining ever returned NaN in the past, we compile a call
1698 // to the runtime to properly handle corner cases
1700 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1701 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1702 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1704 if (!if_slow->is_top()) {
1705 RegionNode* result_region = new (C) RegionNode(3);
1706 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1708 result_region->init_req(1, if_fast);
1709 result_val->init_req(1, result);
1711 set_control(if_slow);
1713 const TypePtr* no_memory_effects = NULL;
1714 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1715 no_memory_effects,
1716 x, top(), y, y ? top() : NULL);
1717 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1718 #ifdef ASSERT
1719 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1720 assert(value_top == top(), "second value must be top");
1721 #endif
1723 result_region->init_req(2, control());
1724 result_val->init_req(2, value);
1725 set_result(result_region, result_val);
1726 } else {
1727 set_result(result);
1728 }
1729 }
1730 }
1732 //------------------------------inline_exp-------------------------------------
1733 // Inline exp instructions, if possible. The Intel hardware only misses
1734 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1735 bool LibraryCallKit::inline_exp() {
1736 Node* arg = round_double_node(argument(0));
1737 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1739 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1741 C->set_has_split_ifs(true); // Has chance for split-if optimization
1742 return true;
1743 }
1745 //------------------------------inline_pow-------------------------------------
1746 // Inline power instructions, if possible.
1747 bool LibraryCallKit::inline_pow() {
1748 // Pseudocode for pow
1749 // if (x <= 0.0) {
1750 // long longy = (long)y;
1751 // if ((double)longy == y) { // if y is long
1752 // if (y + 1 == y) longy = 0; // huge number: even
1753 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1754 // } else {
1755 // result = NaN;
1756 // }
1757 // } else {
1758 // result = DPow(x,y);
1759 // }
1760 // if (result != result)? {
1761 // result = uncommon_trap() or runtime_call();
1762 // }
1763 // return result;
1765 Node* x = round_double_node(argument(0));
1766 Node* y = round_double_node(argument(2));
1768 Node* result = NULL;
1770 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1771 // Short form: skip the fancy tests and just check for NaN result.
1772 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1773 } else {
1774 // If this inlining ever returned NaN in the past, include all
1775 // checks + call to the runtime.
1777 // Set the merge point for If node with condition of (x <= 0.0)
1778 // There are four possible paths to region node and phi node
1779 RegionNode *r = new (C) RegionNode(4);
1780 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1782 // Build the first if node: if (x <= 0.0)
1783 // Node for 0 constant
1784 Node *zeronode = makecon(TypeD::ZERO);
1785 // Check x:0
1786 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1787 // Check: If (x<=0) then go complex path
1788 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1789 // Branch either way
1790 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1791 // Fast path taken; set region slot 3
1792 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1793 r->init_req(3,fast_taken); // Capture fast-control
1795 // Fast path not-taken, i.e. slow path
1796 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1798 // Set fast path result
1799 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1800 phi->init_req(3, fast_result);
1802 // Complex path
1803 // Build the second if node (if y is long)
1804 // Node for (long)y
1805 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1806 // Node for (double)((long) y)
1807 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1808 // Check (double)((long) y) : y
1809 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1810 // Check if (y isn't long) then go to slow path
1812 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1813 // Branch either way
1814 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1815 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1817 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1819 // Calculate DPow(abs(x), y)*(1 & (long)y)
1820 // Node for constant 1
1821 Node *conone = longcon(1);
1822 // 1& (long)y
1823 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1825 // A huge number is always even. Detect a huge number by checking
1826 // if y + 1 == y and set integer to be tested for parity to 0.
1827 // Required for corner case:
1828 // (long)9.223372036854776E18 = max_jlong
1829 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1830 // max_jlong is odd but 9.223372036854776E18 is even
1831 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1832 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1833 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1834 Node* correctedsign = NULL;
1835 if (ConditionalMoveLimit != 0) {
1836 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1837 } else {
1838 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1839 RegionNode *r = new (C) RegionNode(3);
1840 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1841 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1842 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1843 phi->init_req(1, signnode);
1844 phi->init_req(2, longcon(0));
1845 correctedsign = _gvn.transform(phi);
1846 ylong_path = _gvn.transform(r);
1847 record_for_igvn(r);
1848 }
1850 // zero node
1851 Node *conzero = longcon(0);
1852 // Check (1&(long)y)==0?
1853 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1854 // Check if (1&(long)y)!=0?, if so the result is negative
1855 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1856 // abs(x)
1857 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1858 // abs(x)^y
1859 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1860 // -abs(x)^y
1861 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1862 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1863 Node *signresult = NULL;
1864 if (ConditionalMoveLimit != 0) {
1865 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1866 } else {
1867 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1868 RegionNode *r = new (C) RegionNode(3);
1869 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1870 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1871 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1872 phi->init_req(1, absxpowy);
1873 phi->init_req(2, negabsxpowy);
1874 signresult = _gvn.transform(phi);
1875 ylong_path = _gvn.transform(r);
1876 record_for_igvn(r);
1877 }
1878 // Set complex path fast result
1879 r->init_req(2, ylong_path);
1880 phi->init_req(2, signresult);
1882 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1883 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1884 r->init_req(1,slow_path);
1885 phi->init_req(1,slow_result);
1887 // Post merge
1888 set_control(_gvn.transform(r));
1889 record_for_igvn(r);
1890 result = _gvn.transform(phi);
1891 }
1893 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1895 C->set_has_split_ifs(true); // Has chance for split-if optimization
1896 return true;
1897 }
1899 //------------------------------runtime_math-----------------------------
1900 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1901 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1902 "must be (DD)D or (D)D type");
1904 // Inputs
1905 Node* a = round_double_node(argument(0));
1906 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1908 const TypePtr* no_memory_effects = NULL;
1909 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1910 no_memory_effects,
1911 a, top(), b, b ? top() : NULL);
1912 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1913 #ifdef ASSERT
1914 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1915 assert(value_top == top(), "second value must be top");
1916 #endif
1918 set_result(value);
1919 return true;
1920 }
1922 //------------------------------inline_math_native-----------------------------
1923 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1924 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1925 switch (id) {
1926 // These intrinsics are not properly supported on all hardware
1927 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1928 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1929 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1930 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1931 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1932 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1934 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1935 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1936 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1937 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1939 // These intrinsics are supported on all hardware
1940 case vmIntrinsics::_dsqrt: return Matcher::match_rule_supported(Op_SqrtD) ? inline_math(id) : false;
1941 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1943 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1944 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1945 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1946 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1947 #undef FN_PTR
1949 // These intrinsics are not yet correctly implemented
1950 case vmIntrinsics::_datan2:
1951 return false;
1953 default:
1954 fatal_unexpected_iid(id);
1955 return false;
1956 }
1957 }
1959 static bool is_simple_name(Node* n) {
1960 return (n->req() == 1 // constant
1961 || (n->is_Type() && n->as_Type()->type()->singleton())
1962 || n->is_Proj() // parameter or return value
1963 || n->is_Phi() // local of some sort
1964 );
1965 }
1967 //----------------------------inline_min_max-----------------------------------
1968 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1969 set_result(generate_min_max(id, argument(0), argument(1)));
1970 return true;
1971 }
1973 void LibraryCallKit::inline_math_mathExact(Node* math) {
1974 // If we didn't get the expected opcode it means we have optimized
1975 // the node to something else and don't need the exception edge.
1976 if (!math->is_MathExact()) {
1977 set_result(math);
1978 return;
1979 }
1981 Node* result = _gvn.transform( new(C) ProjNode(math, MathExactNode::result_proj_node));
1982 Node* flags = _gvn.transform( new(C) FlagsProjNode(math, MathExactNode::flags_proj_node));
1984 Node* bol = _gvn.transform( new (C) BoolNode(flags, BoolTest::overflow) );
1985 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1986 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1987 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1989 {
1990 PreserveJVMState pjvms(this);
1991 PreserveReexecuteState preexecs(this);
1992 jvms()->set_should_reexecute(true);
1994 set_control(slow_path);
1995 set_i_o(i_o());
1997 uncommon_trap(Deoptimization::Reason_intrinsic,
1998 Deoptimization::Action_none);
1999 }
2001 set_control(fast_path);
2002 set_result(result);
2003 }
2005 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2006 Node* arg1 = argument(0);
2007 Node* arg2 = NULL;
2009 if (is_increment) {
2010 arg2 = intcon(1);
2011 } else {
2012 arg2 = argument(1);
2013 }
2015 Node* add = _gvn.transform( new(C) AddExactINode(NULL, arg1, arg2) );
2016 inline_math_mathExact(add);
2017 return true;
2018 }
2020 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2021 Node* arg1 = argument(0); // type long
2022 // argument(1) == TOP
2023 Node* arg2 = NULL;
2025 if (is_increment) {
2026 arg2 = longcon(1);
2027 } else {
2028 arg2 = argument(2); // type long
2029 // argument(3) == TOP
2030 }
2032 Node* add = _gvn.transform(new(C) AddExactLNode(NULL, arg1, arg2));
2033 inline_math_mathExact(add);
2034 return true;
2035 }
2037 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2038 Node* arg1 = argument(0);
2039 Node* arg2 = NULL;
2041 if (is_decrement) {
2042 arg2 = intcon(1);
2043 } else {
2044 arg2 = argument(1);
2045 }
2047 Node* sub = _gvn.transform(new(C) SubExactINode(NULL, arg1, arg2));
2048 inline_math_mathExact(sub);
2049 return true;
2050 }
2052 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2053 Node* arg1 = argument(0); // type long
2054 // argument(1) == TOP
2055 Node* arg2 = NULL;
2057 if (is_decrement) {
2058 arg2 = longcon(1);
2059 } else {
2060 arg2 = argument(2); // type long
2061 // argument(3) == TOP
2062 }
2064 Node* sub = _gvn.transform(new(C) SubExactLNode(NULL, arg1, arg2));
2065 inline_math_mathExact(sub);
2066 return true;
2067 }
2069 bool LibraryCallKit::inline_math_negateExactI() {
2070 Node* arg1 = argument(0);
2072 Node* neg = _gvn.transform(new(C) NegExactINode(NULL, arg1));
2073 inline_math_mathExact(neg);
2074 return true;
2075 }
2077 bool LibraryCallKit::inline_math_negateExactL() {
2078 Node* arg1 = argument(0);
2079 // argument(1) == TOP
2081 Node* neg = _gvn.transform(new(C) NegExactLNode(NULL, arg1));
2082 inline_math_mathExact(neg);
2083 return true;
2084 }
2086 bool LibraryCallKit::inline_math_multiplyExactI() {
2087 Node* arg1 = argument(0);
2088 Node* arg2 = argument(1);
2090 Node* mul = _gvn.transform(new(C) MulExactINode(NULL, arg1, arg2));
2091 inline_math_mathExact(mul);
2092 return true;
2093 }
2095 bool LibraryCallKit::inline_math_multiplyExactL() {
2096 Node* arg1 = argument(0);
2097 // argument(1) == TOP
2098 Node* arg2 = argument(2);
2099 // argument(3) == TOP
2101 Node* mul = _gvn.transform(new(C) MulExactLNode(NULL, arg1, arg2));
2102 inline_math_mathExact(mul);
2103 return true;
2104 }
2106 Node*
2107 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2108 // These are the candidate return value:
2109 Node* xvalue = x0;
2110 Node* yvalue = y0;
2112 if (xvalue == yvalue) {
2113 return xvalue;
2114 }
2116 bool want_max = (id == vmIntrinsics::_max);
2118 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2119 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2120 if (txvalue == NULL || tyvalue == NULL) return top();
2121 // This is not really necessary, but it is consistent with a
2122 // hypothetical MaxINode::Value method:
2123 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2125 // %%% This folding logic should (ideally) be in a different place.
2126 // Some should be inside IfNode, and there to be a more reliable
2127 // transformation of ?: style patterns into cmoves. We also want
2128 // more powerful optimizations around cmove and min/max.
2130 // Try to find a dominating comparison of these guys.
2131 // It can simplify the index computation for Arrays.copyOf
2132 // and similar uses of System.arraycopy.
2133 // First, compute the normalized version of CmpI(x, y).
2134 int cmp_op = Op_CmpI;
2135 Node* xkey = xvalue;
2136 Node* ykey = yvalue;
2137 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2138 if (ideal_cmpxy->is_Cmp()) {
2139 // E.g., if we have CmpI(length - offset, count),
2140 // it might idealize to CmpI(length, count + offset)
2141 cmp_op = ideal_cmpxy->Opcode();
2142 xkey = ideal_cmpxy->in(1);
2143 ykey = ideal_cmpxy->in(2);
2144 }
2146 // Start by locating any relevant comparisons.
2147 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2148 Node* cmpxy = NULL;
2149 Node* cmpyx = NULL;
2150 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2151 Node* cmp = start_from->fast_out(k);
2152 if (cmp->outcnt() > 0 && // must have prior uses
2153 cmp->in(0) == NULL && // must be context-independent
2154 cmp->Opcode() == cmp_op) { // right kind of compare
2155 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2156 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2157 }
2158 }
2160 const int NCMPS = 2;
2161 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2162 int cmpn;
2163 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2164 if (cmps[cmpn] != NULL) break; // find a result
2165 }
2166 if (cmpn < NCMPS) {
2167 // Look for a dominating test that tells us the min and max.
2168 int depth = 0; // Limit search depth for speed
2169 Node* dom = control();
2170 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2171 if (++depth >= 100) break;
2172 Node* ifproj = dom;
2173 if (!ifproj->is_Proj()) continue;
2174 Node* iff = ifproj->in(0);
2175 if (!iff->is_If()) continue;
2176 Node* bol = iff->in(1);
2177 if (!bol->is_Bool()) continue;
2178 Node* cmp = bol->in(1);
2179 if (cmp == NULL) continue;
2180 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2181 if (cmps[cmpn] == cmp) break;
2182 if (cmpn == NCMPS) continue;
2183 BoolTest::mask btest = bol->as_Bool()->_test._test;
2184 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2185 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2186 // At this point, we know that 'x btest y' is true.
2187 switch (btest) {
2188 case BoolTest::eq:
2189 // They are proven equal, so we can collapse the min/max.
2190 // Either value is the answer. Choose the simpler.
2191 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2192 return yvalue;
2193 return xvalue;
2194 case BoolTest::lt: // x < y
2195 case BoolTest::le: // x <= y
2196 return (want_max ? yvalue : xvalue);
2197 case BoolTest::gt: // x > y
2198 case BoolTest::ge: // x >= y
2199 return (want_max ? xvalue : yvalue);
2200 }
2201 }
2202 }
2204 // We failed to find a dominating test.
2205 // Let's pick a test that might GVN with prior tests.
2206 Node* best_bol = NULL;
2207 BoolTest::mask best_btest = BoolTest::illegal;
2208 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2209 Node* cmp = cmps[cmpn];
2210 if (cmp == NULL) continue;
2211 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2212 Node* bol = cmp->fast_out(j);
2213 if (!bol->is_Bool()) continue;
2214 BoolTest::mask btest = bol->as_Bool()->_test._test;
2215 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2216 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2217 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2218 best_bol = bol->as_Bool();
2219 best_btest = btest;
2220 }
2221 }
2222 }
2224 Node* answer_if_true = NULL;
2225 Node* answer_if_false = NULL;
2226 switch (best_btest) {
2227 default:
2228 if (cmpxy == NULL)
2229 cmpxy = ideal_cmpxy;
2230 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2231 // and fall through:
2232 case BoolTest::lt: // x < y
2233 case BoolTest::le: // x <= y
2234 answer_if_true = (want_max ? yvalue : xvalue);
2235 answer_if_false = (want_max ? xvalue : yvalue);
2236 break;
2237 case BoolTest::gt: // x > y
2238 case BoolTest::ge: // x >= y
2239 answer_if_true = (want_max ? xvalue : yvalue);
2240 answer_if_false = (want_max ? yvalue : xvalue);
2241 break;
2242 }
2244 jint hi, lo;
2245 if (want_max) {
2246 // We can sharpen the minimum.
2247 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2248 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2249 } else {
2250 // We can sharpen the maximum.
2251 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2252 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2253 }
2255 // Use a flow-free graph structure, to avoid creating excess control edges
2256 // which could hinder other optimizations.
2257 // Since Math.min/max is often used with arraycopy, we want
2258 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2259 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2260 answer_if_false, answer_if_true,
2261 TypeInt::make(lo, hi, widen));
2263 return _gvn.transform(cmov);
2265 /*
2266 // This is not as desirable as it may seem, since Min and Max
2267 // nodes do not have a full set of optimizations.
2268 // And they would interfere, anyway, with 'if' optimizations
2269 // and with CMoveI canonical forms.
2270 switch (id) {
2271 case vmIntrinsics::_min:
2272 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2273 case vmIntrinsics::_max:
2274 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2275 default:
2276 ShouldNotReachHere();
2277 }
2278 */
2279 }
2281 inline int
2282 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2283 const TypePtr* base_type = TypePtr::NULL_PTR;
2284 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2285 if (base_type == NULL) {
2286 // Unknown type.
2287 return Type::AnyPtr;
2288 } else if (base_type == TypePtr::NULL_PTR) {
2289 // Since this is a NULL+long form, we have to switch to a rawptr.
2290 base = _gvn.transform(new (C) CastX2PNode(offset));
2291 offset = MakeConX(0);
2292 return Type::RawPtr;
2293 } else if (base_type->base() == Type::RawPtr) {
2294 return Type::RawPtr;
2295 } else if (base_type->isa_oopptr()) {
2296 // Base is never null => always a heap address.
2297 if (base_type->ptr() == TypePtr::NotNull) {
2298 return Type::OopPtr;
2299 }
2300 // Offset is small => always a heap address.
2301 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2302 if (offset_type != NULL &&
2303 base_type->offset() == 0 && // (should always be?)
2304 offset_type->_lo >= 0 &&
2305 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2306 return Type::OopPtr;
2307 }
2308 // Otherwise, it might either be oop+off or NULL+addr.
2309 return Type::AnyPtr;
2310 } else {
2311 // No information:
2312 return Type::AnyPtr;
2313 }
2314 }
2316 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2317 int kind = classify_unsafe_addr(base, offset);
2318 if (kind == Type::RawPtr) {
2319 return basic_plus_adr(top(), base, offset);
2320 } else {
2321 return basic_plus_adr(base, offset);
2322 }
2323 }
2325 //--------------------------inline_number_methods-----------------------------
2326 // inline int Integer.numberOfLeadingZeros(int)
2327 // inline int Long.numberOfLeadingZeros(long)
2328 //
2329 // inline int Integer.numberOfTrailingZeros(int)
2330 // inline int Long.numberOfTrailingZeros(long)
2331 //
2332 // inline int Integer.bitCount(int)
2333 // inline int Long.bitCount(long)
2334 //
2335 // inline char Character.reverseBytes(char)
2336 // inline short Short.reverseBytes(short)
2337 // inline int Integer.reverseBytes(int)
2338 // inline long Long.reverseBytes(long)
2339 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2340 Node* arg = argument(0);
2341 Node* n;
2342 switch (id) {
2343 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2344 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2345 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2346 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2347 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2348 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2349 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2350 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2351 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2352 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2353 default: fatal_unexpected_iid(id); break;
2354 }
2355 set_result(_gvn.transform(n));
2356 return true;
2357 }
2359 //----------------------------inline_unsafe_access----------------------------
2361 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2363 // Helper that guards and inserts a pre-barrier.
2364 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2365 Node* pre_val, bool need_mem_bar) {
2366 // We could be accessing the referent field of a reference object. If so, when G1
2367 // is enabled, we need to log the value in the referent field in an SATB buffer.
2368 // This routine performs some compile time filters and generates suitable
2369 // runtime filters that guard the pre-barrier code.
2370 // Also add memory barrier for non volatile load from the referent field
2371 // to prevent commoning of loads across safepoint.
2372 if (!UseG1GC && !need_mem_bar)
2373 return;
2375 // Some compile time checks.
2377 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2378 const TypeX* otype = offset->find_intptr_t_type();
2379 if (otype != NULL && otype->is_con() &&
2380 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2381 // Constant offset but not the reference_offset so just return
2382 return;
2383 }
2385 // We only need to generate the runtime guards for instances.
2386 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2387 if (btype != NULL) {
2388 if (btype->isa_aryptr()) {
2389 // Array type so nothing to do
2390 return;
2391 }
2393 const TypeInstPtr* itype = btype->isa_instptr();
2394 if (itype != NULL) {
2395 // Can the klass of base_oop be statically determined to be
2396 // _not_ a sub-class of Reference and _not_ Object?
2397 ciKlass* klass = itype->klass();
2398 if ( klass->is_loaded() &&
2399 !klass->is_subtype_of(env()->Reference_klass()) &&
2400 !env()->Object_klass()->is_subtype_of(klass)) {
2401 return;
2402 }
2403 }
2404 }
2406 // The compile time filters did not reject base_oop/offset so
2407 // we need to generate the following runtime filters
2408 //
2409 // if (offset == java_lang_ref_Reference::_reference_offset) {
2410 // if (instance_of(base, java.lang.ref.Reference)) {
2411 // pre_barrier(_, pre_val, ...);
2412 // }
2413 // }
2415 float likely = PROB_LIKELY( 0.999);
2416 float unlikely = PROB_UNLIKELY(0.999);
2418 IdealKit ideal(this);
2419 #define __ ideal.
2421 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2423 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2424 // Update graphKit memory and control from IdealKit.
2425 sync_kit(ideal);
2427 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2428 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2430 // Update IdealKit memory and control from graphKit.
2431 __ sync_kit(this);
2433 Node* one = __ ConI(1);
2434 // is_instof == 0 if base_oop == NULL
2435 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2437 // Update graphKit from IdeakKit.
2438 sync_kit(ideal);
2440 // Use the pre-barrier to record the value in the referent field
2441 pre_barrier(false /* do_load */,
2442 __ ctrl(),
2443 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2444 pre_val /* pre_val */,
2445 T_OBJECT);
2446 if (need_mem_bar) {
2447 // Add memory barrier to prevent commoning reads from this field
2448 // across safepoint since GC can change its value.
2449 insert_mem_bar(Op_MemBarCPUOrder);
2450 }
2451 // Update IdealKit from graphKit.
2452 __ sync_kit(this);
2454 } __ end_if(); // _ref_type != ref_none
2455 } __ end_if(); // offset == referent_offset
2457 // Final sync IdealKit and GraphKit.
2458 final_sync(ideal);
2459 #undef __
2460 }
2463 // Interpret Unsafe.fieldOffset cookies correctly:
2464 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2466 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2467 // Attempt to infer a sharper value type from the offset and base type.
2468 ciKlass* sharpened_klass = NULL;
2470 // See if it is an instance field, with an object type.
2471 if (alias_type->field() != NULL) {
2472 assert(!is_native_ptr, "native pointer op cannot use a java address");
2473 if (alias_type->field()->type()->is_klass()) {
2474 sharpened_klass = alias_type->field()->type()->as_klass();
2475 }
2476 }
2478 // See if it is a narrow oop array.
2479 if (adr_type->isa_aryptr()) {
2480 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2481 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2482 if (elem_type != NULL) {
2483 sharpened_klass = elem_type->klass();
2484 }
2485 }
2486 }
2488 // The sharpened class might be unloaded if there is no class loader
2489 // contraint in place.
2490 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2491 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2493 #ifndef PRODUCT
2494 if (C->print_intrinsics() || C->print_inlining()) {
2495 tty->print(" from base type: "); adr_type->dump();
2496 tty->print(" sharpened value: "); tjp->dump();
2497 }
2498 #endif
2499 // Sharpen the value type.
2500 return tjp;
2501 }
2502 return NULL;
2503 }
2505 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2506 if (callee()->is_static()) return false; // caller must have the capability!
2508 #ifndef PRODUCT
2509 {
2510 ResourceMark rm;
2511 // Check the signatures.
2512 ciSignature* sig = callee()->signature();
2513 #ifdef ASSERT
2514 if (!is_store) {
2515 // Object getObject(Object base, int/long offset), etc.
2516 BasicType rtype = sig->return_type()->basic_type();
2517 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2518 rtype = T_ADDRESS; // it is really a C void*
2519 assert(rtype == type, "getter must return the expected value");
2520 if (!is_native_ptr) {
2521 assert(sig->count() == 2, "oop getter has 2 arguments");
2522 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2523 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2524 } else {
2525 assert(sig->count() == 1, "native getter has 1 argument");
2526 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2527 }
2528 } else {
2529 // void putObject(Object base, int/long offset, Object x), etc.
2530 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2531 if (!is_native_ptr) {
2532 assert(sig->count() == 3, "oop putter has 3 arguments");
2533 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2534 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2535 } else {
2536 assert(sig->count() == 2, "native putter has 2 arguments");
2537 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2538 }
2539 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2540 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2541 vtype = T_ADDRESS; // it is really a C void*
2542 assert(vtype == type, "putter must accept the expected value");
2543 }
2544 #endif // ASSERT
2545 }
2546 #endif //PRODUCT
2548 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2550 Node* receiver = argument(0); // type: oop
2552 // Build address expression. See the code in inline_unsafe_prefetch.
2553 Node* adr;
2554 Node* heap_base_oop = top();
2555 Node* offset = top();
2556 Node* val;
2558 if (!is_native_ptr) {
2559 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2560 Node* base = argument(1); // type: oop
2561 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2562 offset = argument(2); // type: long
2563 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2564 // to be plain byte offsets, which are also the same as those accepted
2565 // by oopDesc::field_base.
2566 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2567 "fieldOffset must be byte-scaled");
2568 // 32-bit machines ignore the high half!
2569 offset = ConvL2X(offset);
2570 adr = make_unsafe_address(base, offset);
2571 heap_base_oop = base;
2572 val = is_store ? argument(4) : NULL;
2573 } else {
2574 Node* ptr = argument(1); // type: long
2575 ptr = ConvL2X(ptr); // adjust Java long to machine word
2576 adr = make_unsafe_address(NULL, ptr);
2577 val = is_store ? argument(3) : NULL;
2578 }
2580 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2582 // First guess at the value type.
2583 const Type *value_type = Type::get_const_basic_type(type);
2585 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2586 // there was not enough information to nail it down.
2587 Compile::AliasType* alias_type = C->alias_type(adr_type);
2588 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2590 // We will need memory barriers unless we can determine a unique
2591 // alias category for this reference. (Note: If for some reason
2592 // the barriers get omitted and the unsafe reference begins to "pollute"
2593 // the alias analysis of the rest of the graph, either Compile::can_alias
2594 // or Compile::must_alias will throw a diagnostic assert.)
2595 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2597 // If we are reading the value of the referent field of a Reference
2598 // object (either by using Unsafe directly or through reflection)
2599 // then, if G1 is enabled, we need to record the referent in an
2600 // SATB log buffer using the pre-barrier mechanism.
2601 // Also we need to add memory barrier to prevent commoning reads
2602 // from this field across safepoint since GC can change its value.
2603 bool need_read_barrier = !is_native_ptr && !is_store &&
2604 offset != top() && heap_base_oop != top();
2606 if (!is_store && type == T_OBJECT) {
2607 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2608 if (tjp != NULL) {
2609 value_type = tjp;
2610 }
2611 }
2613 receiver = null_check(receiver);
2614 if (stopped()) {
2615 return true;
2616 }
2617 // Heap pointers get a null-check from the interpreter,
2618 // as a courtesy. However, this is not guaranteed by Unsafe,
2619 // and it is not possible to fully distinguish unintended nulls
2620 // from intended ones in this API.
2622 if (is_volatile) {
2623 // We need to emit leading and trailing CPU membars (see below) in
2624 // addition to memory membars when is_volatile. This is a little
2625 // too strong, but avoids the need to insert per-alias-type
2626 // volatile membars (for stores; compare Parse::do_put_xxx), which
2627 // we cannot do effectively here because we probably only have a
2628 // rough approximation of type.
2629 need_mem_bar = true;
2630 // For Stores, place a memory ordering barrier now.
2631 if (is_store) {
2632 insert_mem_bar(Op_MemBarRelease);
2633 } else {
2634 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
2635 insert_mem_bar(Op_MemBarVolatile);
2636 }
2637 }
2638 }
2640 // Memory barrier to prevent normal and 'unsafe' accesses from
2641 // bypassing each other. Happens after null checks, so the
2642 // exception paths do not take memory state from the memory barrier,
2643 // so there's no problems making a strong assert about mixing users
2644 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2645 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2646 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2648 if (!is_store) {
2649 Node* p = make_load(control(), adr, value_type, type, adr_type, MemNode::unordered, is_volatile);
2650 // load value
2651 switch (type) {
2652 case T_BOOLEAN:
2653 case T_CHAR:
2654 case T_BYTE:
2655 case T_SHORT:
2656 case T_INT:
2657 case T_LONG:
2658 case T_FLOAT:
2659 case T_DOUBLE:
2660 break;
2661 case T_OBJECT:
2662 if (need_read_barrier) {
2663 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2664 }
2665 break;
2666 case T_ADDRESS:
2667 // Cast to an int type.
2668 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2669 p = ConvX2L(p);
2670 break;
2671 default:
2672 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2673 break;
2674 }
2675 // The load node has the control of the preceding MemBarCPUOrder. All
2676 // following nodes will have the control of the MemBarCPUOrder inserted at
2677 // the end of this method. So, pushing the load onto the stack at a later
2678 // point is fine.
2679 set_result(p);
2680 } else {
2681 // place effect of store into memory
2682 switch (type) {
2683 case T_DOUBLE:
2684 val = dstore_rounding(val);
2685 break;
2686 case T_ADDRESS:
2687 // Repackage the long as a pointer.
2688 val = ConvL2X(val);
2689 val = _gvn.transform(new (C) CastX2PNode(val));
2690 break;
2691 }
2693 MemNode::MemOrd mo = is_volatile ? MemNode::release : MemNode::unordered;
2694 if (type != T_OBJECT ) {
2695 (void) store_to_memory(control(), adr, val, type, adr_type, mo, is_volatile);
2696 } else {
2697 // Possibly an oop being stored to Java heap or native memory
2698 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2699 // oop to Java heap.
2700 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2701 } else {
2702 // We can't tell at compile time if we are storing in the Java heap or outside
2703 // of it. So we need to emit code to conditionally do the proper type of
2704 // store.
2706 IdealKit ideal(this);
2707 #define __ ideal.
2708 // QQQ who knows what probability is here??
2709 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2710 // Sync IdealKit and graphKit.
2711 sync_kit(ideal);
2712 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type, mo);
2713 // Update IdealKit memory.
2714 __ sync_kit(this);
2715 } __ else_(); {
2716 __ store(__ ctrl(), adr, val, type, alias_type->index(), mo, is_volatile);
2717 } __ end_if();
2718 // Final sync IdealKit and GraphKit.
2719 final_sync(ideal);
2720 #undef __
2721 }
2722 }
2723 }
2725 if (is_volatile) {
2726 if (!is_store) {
2727 insert_mem_bar(Op_MemBarAcquire);
2728 } else {
2729 if (!support_IRIW_for_not_multiple_copy_atomic_cpu) {
2730 insert_mem_bar(Op_MemBarVolatile);
2731 }
2732 }
2733 }
2735 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2737 return true;
2738 }
2740 //----------------------------inline_unsafe_prefetch----------------------------
2742 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2743 #ifndef PRODUCT
2744 {
2745 ResourceMark rm;
2746 // Check the signatures.
2747 ciSignature* sig = callee()->signature();
2748 #ifdef ASSERT
2749 // Object getObject(Object base, int/long offset), etc.
2750 BasicType rtype = sig->return_type()->basic_type();
2751 if (!is_native_ptr) {
2752 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2753 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2754 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2755 } else {
2756 assert(sig->count() == 1, "native prefetch has 1 argument");
2757 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2758 }
2759 #endif // ASSERT
2760 }
2761 #endif // !PRODUCT
2763 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2765 const int idx = is_static ? 0 : 1;
2766 if (!is_static) {
2767 null_check_receiver();
2768 if (stopped()) {
2769 return true;
2770 }
2771 }
2773 // Build address expression. See the code in inline_unsafe_access.
2774 Node *adr;
2775 if (!is_native_ptr) {
2776 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2777 Node* base = argument(idx + 0); // type: oop
2778 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2779 Node* offset = argument(idx + 1); // type: long
2780 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2781 // to be plain byte offsets, which are also the same as those accepted
2782 // by oopDesc::field_base.
2783 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2784 "fieldOffset must be byte-scaled");
2785 // 32-bit machines ignore the high half!
2786 offset = ConvL2X(offset);
2787 adr = make_unsafe_address(base, offset);
2788 } else {
2789 Node* ptr = argument(idx + 0); // type: long
2790 ptr = ConvL2X(ptr); // adjust Java long to machine word
2791 adr = make_unsafe_address(NULL, ptr);
2792 }
2794 // Generate the read or write prefetch
2795 Node *prefetch;
2796 if (is_store) {
2797 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2798 } else {
2799 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2800 }
2801 prefetch->init_req(0, control());
2802 set_i_o(_gvn.transform(prefetch));
2804 return true;
2805 }
2807 //----------------------------inline_unsafe_load_store----------------------------
2808 // This method serves a couple of different customers (depending on LoadStoreKind):
2809 //
2810 // LS_cmpxchg:
2811 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2812 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2813 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2814 //
2815 // LS_xadd:
2816 // public int getAndAddInt( Object o, long offset, int delta)
2817 // public long getAndAddLong(Object o, long offset, long delta)
2818 //
2819 // LS_xchg:
2820 // int getAndSet(Object o, long offset, int newValue)
2821 // long getAndSet(Object o, long offset, long newValue)
2822 // Object getAndSet(Object o, long offset, Object newValue)
2823 //
2824 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2825 // This basic scheme here is the same as inline_unsafe_access, but
2826 // differs in enough details that combining them would make the code
2827 // overly confusing. (This is a true fact! I originally combined
2828 // them, but even I was confused by it!) As much code/comments as
2829 // possible are retained from inline_unsafe_access though to make
2830 // the correspondences clearer. - dl
2832 if (callee()->is_static()) return false; // caller must have the capability!
2834 #ifndef PRODUCT
2835 BasicType rtype;
2836 {
2837 ResourceMark rm;
2838 // Check the signatures.
2839 ciSignature* sig = callee()->signature();
2840 rtype = sig->return_type()->basic_type();
2841 if (kind == LS_xadd || kind == LS_xchg) {
2842 // Check the signatures.
2843 #ifdef ASSERT
2844 assert(rtype == type, "get and set must return the expected type");
2845 assert(sig->count() == 3, "get and set has 3 arguments");
2846 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2847 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2848 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2849 #endif // ASSERT
2850 } else if (kind == LS_cmpxchg) {
2851 // Check the signatures.
2852 #ifdef ASSERT
2853 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2854 assert(sig->count() == 4, "CAS has 4 arguments");
2855 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2856 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2857 #endif // ASSERT
2858 } else {
2859 ShouldNotReachHere();
2860 }
2861 }
2862 #endif //PRODUCT
2864 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2866 // Get arguments:
2867 Node* receiver = NULL;
2868 Node* base = NULL;
2869 Node* offset = NULL;
2870 Node* oldval = NULL;
2871 Node* newval = NULL;
2872 if (kind == LS_cmpxchg) {
2873 const bool two_slot_type = type2size[type] == 2;
2874 receiver = argument(0); // type: oop
2875 base = argument(1); // type: oop
2876 offset = argument(2); // type: long
2877 oldval = argument(4); // type: oop, int, or long
2878 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2879 } else if (kind == LS_xadd || kind == LS_xchg){
2880 receiver = argument(0); // type: oop
2881 base = argument(1); // type: oop
2882 offset = argument(2); // type: long
2883 oldval = NULL;
2884 newval = argument(4); // type: oop, int, or long
2885 }
2887 // Null check receiver.
2888 receiver = null_check(receiver);
2889 if (stopped()) {
2890 return true;
2891 }
2893 // Build field offset expression.
2894 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2895 // to be plain byte offsets, which are also the same as those accepted
2896 // by oopDesc::field_base.
2897 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2898 // 32-bit machines ignore the high half of long offsets
2899 offset = ConvL2X(offset);
2900 Node* adr = make_unsafe_address(base, offset);
2901 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2903 // For CAS, unlike inline_unsafe_access, there seems no point in
2904 // trying to refine types. Just use the coarse types here.
2905 const Type *value_type = Type::get_const_basic_type(type);
2906 Compile::AliasType* alias_type = C->alias_type(adr_type);
2907 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2909 if (kind == LS_xchg && type == T_OBJECT) {
2910 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2911 if (tjp != NULL) {
2912 value_type = tjp;
2913 }
2914 }
2916 int alias_idx = C->get_alias_index(adr_type);
2918 // Memory-model-wise, a LoadStore acts like a little synchronized
2919 // block, so needs barriers on each side. These don't translate
2920 // into actual barriers on most machines, but we still need rest of
2921 // compiler to respect ordering.
2923 insert_mem_bar(Op_MemBarRelease);
2924 insert_mem_bar(Op_MemBarCPUOrder);
2926 // 4984716: MemBars must be inserted before this
2927 // memory node in order to avoid a false
2928 // dependency which will confuse the scheduler.
2929 Node *mem = memory(alias_idx);
2931 // For now, we handle only those cases that actually exist: ints,
2932 // longs, and Object. Adding others should be straightforward.
2933 Node* load_store;
2934 switch(type) {
2935 case T_INT:
2936 if (kind == LS_xadd) {
2937 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2938 } else if (kind == LS_xchg) {
2939 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2940 } else if (kind == LS_cmpxchg) {
2941 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2942 } else {
2943 ShouldNotReachHere();
2944 }
2945 break;
2946 case T_LONG:
2947 if (kind == LS_xadd) {
2948 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2949 } else if (kind == LS_xchg) {
2950 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2951 } else if (kind == LS_cmpxchg) {
2952 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2953 } else {
2954 ShouldNotReachHere();
2955 }
2956 break;
2957 case T_OBJECT:
2958 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2959 // could be delayed during Parse (for example, in adjust_map_after_if()).
2960 // Execute transformation here to avoid barrier generation in such case.
2961 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2962 newval = _gvn.makecon(TypePtr::NULL_PTR);
2964 // Reference stores need a store barrier.
2965 if (kind == LS_xchg) {
2966 // If pre-barrier must execute before the oop store, old value will require do_load here.
2967 if (!can_move_pre_barrier()) {
2968 pre_barrier(true /* do_load*/,
2969 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2970 NULL /* pre_val*/,
2971 T_OBJECT);
2972 } // Else move pre_barrier to use load_store value, see below.
2973 } else if (kind == LS_cmpxchg) {
2974 // Same as for newval above:
2975 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2976 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2977 }
2978 // The only known value which might get overwritten is oldval.
2979 pre_barrier(false /* do_load */,
2980 control(), NULL, NULL, max_juint, NULL, NULL,
2981 oldval /* pre_val */,
2982 T_OBJECT);
2983 } else {
2984 ShouldNotReachHere();
2985 }
2987 #ifdef _LP64
2988 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2989 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2990 if (kind == LS_xchg) {
2991 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2992 newval_enc, adr_type, value_type->make_narrowoop()));
2993 } else {
2994 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2995 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2996 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2997 newval_enc, oldval_enc));
2998 }
2999 } else
3000 #endif
3001 {
3002 if (kind == LS_xchg) {
3003 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3004 } else {
3005 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3006 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3007 }
3008 }
3009 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3010 break;
3011 default:
3012 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
3013 break;
3014 }
3016 // SCMemProjNodes represent the memory state of a LoadStore. Their
3017 // main role is to prevent LoadStore nodes from being optimized away
3018 // when their results aren't used.
3019 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
3020 set_memory(proj, alias_idx);
3022 if (type == T_OBJECT && kind == LS_xchg) {
3023 #ifdef _LP64
3024 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
3025 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
3026 }
3027 #endif
3028 if (can_move_pre_barrier()) {
3029 // Don't need to load pre_val. The old value is returned by load_store.
3030 // The pre_barrier can execute after the xchg as long as no safepoint
3031 // gets inserted between them.
3032 pre_barrier(false /* do_load */,
3033 control(), NULL, NULL, max_juint, NULL, NULL,
3034 load_store /* pre_val */,
3035 T_OBJECT);
3036 }
3037 }
3039 // Add the trailing membar surrounding the access
3040 insert_mem_bar(Op_MemBarCPUOrder);
3041 insert_mem_bar(Op_MemBarAcquire);
3043 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3044 set_result(load_store);
3045 return true;
3046 }
3048 //----------------------------inline_unsafe_ordered_store----------------------
3049 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
3050 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
3051 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
3052 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3053 // This is another variant of inline_unsafe_access, differing in
3054 // that it always issues store-store ("release") barrier and ensures
3055 // store-atomicity (which only matters for "long").
3057 if (callee()->is_static()) return false; // caller must have the capability!
3059 #ifndef PRODUCT
3060 {
3061 ResourceMark rm;
3062 // Check the signatures.
3063 ciSignature* sig = callee()->signature();
3064 #ifdef ASSERT
3065 BasicType rtype = sig->return_type()->basic_type();
3066 assert(rtype == T_VOID, "must return void");
3067 assert(sig->count() == 3, "has 3 arguments");
3068 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3069 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3070 #endif // ASSERT
3071 }
3072 #endif //PRODUCT
3074 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3076 // Get arguments:
3077 Node* receiver = argument(0); // type: oop
3078 Node* base = argument(1); // type: oop
3079 Node* offset = argument(2); // type: long
3080 Node* val = argument(4); // type: oop, int, or long
3082 // Null check receiver.
3083 receiver = null_check(receiver);
3084 if (stopped()) {
3085 return true;
3086 }
3088 // Build field offset expression.
3089 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3090 // 32-bit machines ignore the high half of long offsets
3091 offset = ConvL2X(offset);
3092 Node* adr = make_unsafe_address(base, offset);
3093 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3094 const Type *value_type = Type::get_const_basic_type(type);
3095 Compile::AliasType* alias_type = C->alias_type(adr_type);
3097 insert_mem_bar(Op_MemBarRelease);
3098 insert_mem_bar(Op_MemBarCPUOrder);
3099 // Ensure that the store is atomic for longs:
3100 const bool require_atomic_access = true;
3101 Node* store;
3102 if (type == T_OBJECT) // reference stores need a store barrier.
3103 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type, MemNode::release);
3104 else {
3105 store = store_to_memory(control(), adr, val, type, adr_type, MemNode::release, require_atomic_access);
3106 }
3107 insert_mem_bar(Op_MemBarCPUOrder);
3108 return true;
3109 }
3111 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3112 // Regardless of form, don't allow previous ld/st to move down,
3113 // then issue acquire, release, or volatile mem_bar.
3114 insert_mem_bar(Op_MemBarCPUOrder);
3115 switch(id) {
3116 case vmIntrinsics::_loadFence:
3117 insert_mem_bar(Op_LoadFence);
3118 return true;
3119 case vmIntrinsics::_storeFence:
3120 insert_mem_bar(Op_StoreFence);
3121 return true;
3122 case vmIntrinsics::_fullFence:
3123 insert_mem_bar(Op_MemBarVolatile);
3124 return true;
3125 default:
3126 fatal_unexpected_iid(id);
3127 return false;
3128 }
3129 }
3131 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3132 if (!kls->is_Con()) {
3133 return true;
3134 }
3135 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3136 if (klsptr == NULL) {
3137 return true;
3138 }
3139 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3140 // don't need a guard for a klass that is already initialized
3141 return !ik->is_initialized();
3142 }
3144 //----------------------------inline_unsafe_allocate---------------------------
3145 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3146 bool LibraryCallKit::inline_unsafe_allocate() {
3147 if (callee()->is_static()) return false; // caller must have the capability!
3149 null_check_receiver(); // null-check, then ignore
3150 Node* cls = null_check(argument(1));
3151 if (stopped()) return true;
3153 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3154 kls = null_check(kls);
3155 if (stopped()) return true; // argument was like int.class
3157 Node* test = NULL;
3158 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3159 // Note: The argument might still be an illegal value like
3160 // Serializable.class or Object[].class. The runtime will handle it.
3161 // But we must make an explicit check for initialization.
3162 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3163 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3164 // can generate code to load it as unsigned byte.
3165 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN, MemNode::unordered);
3166 Node* bits = intcon(InstanceKlass::fully_initialized);
3167 test = _gvn.transform(new (C) SubINode(inst, bits));
3168 // The 'test' is non-zero if we need to take a slow path.
3169 }
3171 Node* obj = new_instance(kls, test);
3172 set_result(obj);
3173 return true;
3174 }
3176 #ifdef TRACE_HAVE_INTRINSICS
3177 /*
3178 * oop -> myklass
3179 * myklass->trace_id |= USED
3180 * return myklass->trace_id & ~0x3
3181 */
3182 bool LibraryCallKit::inline_native_classID() {
3183 null_check_receiver(); // null-check, then ignore
3184 Node* cls = null_check(argument(1), T_OBJECT);
3185 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3186 kls = null_check(kls, T_OBJECT);
3187 ByteSize offset = TRACE_ID_OFFSET;
3188 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3189 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG, MemNode::unordered);
3190 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3191 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3192 Node* clsused = longcon(0x01l); // set the class bit
3193 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3195 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3196 store_to_memory(control(), insp, orl, T_LONG, adr_type, MemNode::unordered);
3197 set_result(andl);
3198 return true;
3199 }
3201 bool LibraryCallKit::inline_native_threadID() {
3202 Node* tls_ptr = NULL;
3203 Node* cur_thr = generate_current_thread(tls_ptr);
3204 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3205 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3206 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3208 Node* threadid = NULL;
3209 size_t thread_id_size = OSThread::thread_id_size();
3210 if (thread_id_size == (size_t) BytesPerLong) {
3211 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG, MemNode::unordered));
3212 } else if (thread_id_size == (size_t) BytesPerInt) {
3213 threadid = make_load(control(), p, TypeInt::INT, T_INT, MemNode::unordered);
3214 } else {
3215 ShouldNotReachHere();
3216 }
3217 set_result(threadid);
3218 return true;
3219 }
3220 #endif
3222 //------------------------inline_native_time_funcs--------------
3223 // inline code for System.currentTimeMillis() and System.nanoTime()
3224 // these have the same type and signature
3225 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3226 const TypeFunc* tf = OptoRuntime::void_long_Type();
3227 const TypePtr* no_memory_effects = NULL;
3228 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3229 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3230 #ifdef ASSERT
3231 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3232 assert(value_top == top(), "second value must be top");
3233 #endif
3234 set_result(value);
3235 return true;
3236 }
3238 //------------------------inline_native_currentThread------------------
3239 bool LibraryCallKit::inline_native_currentThread() {
3240 Node* junk = NULL;
3241 set_result(generate_current_thread(junk));
3242 return true;
3243 }
3245 //------------------------inline_native_isInterrupted------------------
3246 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3247 bool LibraryCallKit::inline_native_isInterrupted() {
3248 // Add a fast path to t.isInterrupted(clear_int):
3249 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3250 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3251 // So, in the common case that the interrupt bit is false,
3252 // we avoid making a call into the VM. Even if the interrupt bit
3253 // is true, if the clear_int argument is false, we avoid the VM call.
3254 // However, if the receiver is not currentThread, we must call the VM,
3255 // because there must be some locking done around the operation.
3257 // We only go to the fast case code if we pass two guards.
3258 // Paths which do not pass are accumulated in the slow_region.
3260 enum {
3261 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3262 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3263 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3264 PATH_LIMIT
3265 };
3267 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3268 // out of the function.
3269 insert_mem_bar(Op_MemBarCPUOrder);
3271 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3272 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3274 RegionNode* slow_region = new (C) RegionNode(1);
3275 record_for_igvn(slow_region);
3277 // (a) Receiving thread must be the current thread.
3278 Node* rec_thr = argument(0);
3279 Node* tls_ptr = NULL;
3280 Node* cur_thr = generate_current_thread(tls_ptr);
3281 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3282 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3284 generate_slow_guard(bol_thr, slow_region);
3286 // (b) Interrupt bit on TLS must be false.
3287 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3288 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3289 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3291 // Set the control input on the field _interrupted read to prevent it floating up.
3292 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT, MemNode::unordered);
3293 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3294 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3296 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3298 // First fast path: if (!TLS._interrupted) return false;
3299 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3300 result_rgn->init_req(no_int_result_path, false_bit);
3301 result_val->init_req(no_int_result_path, intcon(0));
3303 // drop through to next case
3304 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3306 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3307 Node* clr_arg = argument(1);
3308 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3309 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3310 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3312 // Second fast path: ... else if (!clear_int) return true;
3313 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3314 result_rgn->init_req(no_clear_result_path, false_arg);
3315 result_val->init_req(no_clear_result_path, intcon(1));
3317 // drop through to next case
3318 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3320 // (d) Otherwise, go to the slow path.
3321 slow_region->add_req(control());
3322 set_control( _gvn.transform(slow_region));
3324 if (stopped()) {
3325 // There is no slow path.
3326 result_rgn->init_req(slow_result_path, top());
3327 result_val->init_req(slow_result_path, top());
3328 } else {
3329 // non-virtual because it is a private non-static
3330 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3332 Node* slow_val = set_results_for_java_call(slow_call);
3333 // this->control() comes from set_results_for_java_call
3335 Node* fast_io = slow_call->in(TypeFunc::I_O);
3336 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3338 // These two phis are pre-filled with copies of of the fast IO and Memory
3339 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3340 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3342 result_rgn->init_req(slow_result_path, control());
3343 result_io ->init_req(slow_result_path, i_o());
3344 result_mem->init_req(slow_result_path, reset_memory());
3345 result_val->init_req(slow_result_path, slow_val);
3347 set_all_memory(_gvn.transform(result_mem));
3348 set_i_o( _gvn.transform(result_io));
3349 }
3351 C->set_has_split_ifs(true); // Has chance for split-if optimization
3352 set_result(result_rgn, result_val);
3353 return true;
3354 }
3356 //---------------------------load_mirror_from_klass----------------------------
3357 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3358 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3359 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3360 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3361 }
3363 //-----------------------load_klass_from_mirror_common-------------------------
3364 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3365 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3366 // and branch to the given path on the region.
3367 // If never_see_null, take an uncommon trap on null, so we can optimistically
3368 // compile for the non-null case.
3369 // If the region is NULL, force never_see_null = true.
3370 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3371 bool never_see_null,
3372 RegionNode* region,
3373 int null_path,
3374 int offset) {
3375 if (region == NULL) never_see_null = true;
3376 Node* p = basic_plus_adr(mirror, offset);
3377 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3378 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3379 Node* null_ctl = top();
3380 kls = null_check_oop(kls, &null_ctl, never_see_null);
3381 if (region != NULL) {
3382 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3383 region->init_req(null_path, null_ctl);
3384 } else {
3385 assert(null_ctl == top(), "no loose ends");
3386 }
3387 return kls;
3388 }
3390 //--------------------(inline_native_Class_query helpers)---------------------
3391 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3392 // Fall through if (mods & mask) == bits, take the guard otherwise.
3393 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3394 // Branch around if the given klass has the given modifier bit set.
3395 // Like generate_guard, adds a new path onto the region.
3396 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3397 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT, MemNode::unordered);
3398 Node* mask = intcon(modifier_mask);
3399 Node* bits = intcon(modifier_bits);
3400 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3401 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3402 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3403 return generate_fair_guard(bol, region);
3404 }
3405 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3406 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3407 }
3409 //-------------------------inline_native_Class_query-------------------
3410 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3411 const Type* return_type = TypeInt::BOOL;
3412 Node* prim_return_value = top(); // what happens if it's a primitive class?
3413 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3414 bool expect_prim = false; // most of these guys expect to work on refs
3416 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3418 Node* mirror = argument(0);
3419 Node* obj = top();
3421 switch (id) {
3422 case vmIntrinsics::_isInstance:
3423 // nothing is an instance of a primitive type
3424 prim_return_value = intcon(0);
3425 obj = argument(1);
3426 break;
3427 case vmIntrinsics::_getModifiers:
3428 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3429 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3430 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3431 break;
3432 case vmIntrinsics::_isInterface:
3433 prim_return_value = intcon(0);
3434 break;
3435 case vmIntrinsics::_isArray:
3436 prim_return_value = intcon(0);
3437 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3438 break;
3439 case vmIntrinsics::_isPrimitive:
3440 prim_return_value = intcon(1);
3441 expect_prim = true; // obviously
3442 break;
3443 case vmIntrinsics::_getSuperclass:
3444 prim_return_value = null();
3445 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3446 break;
3447 case vmIntrinsics::_getComponentType:
3448 prim_return_value = null();
3449 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3450 break;
3451 case vmIntrinsics::_getClassAccessFlags:
3452 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3453 return_type = TypeInt::INT; // not bool! 6297094
3454 break;
3455 default:
3456 fatal_unexpected_iid(id);
3457 break;
3458 }
3460 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3461 if (mirror_con == NULL) return false; // cannot happen?
3463 #ifndef PRODUCT
3464 if (C->print_intrinsics() || C->print_inlining()) {
3465 ciType* k = mirror_con->java_mirror_type();
3466 if (k) {
3467 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3468 k->print_name();
3469 tty->cr();
3470 }
3471 }
3472 #endif
3474 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3475 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3476 record_for_igvn(region);
3477 PhiNode* phi = new (C) PhiNode(region, return_type);
3479 // The mirror will never be null of Reflection.getClassAccessFlags, however
3480 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3481 // if it is. See bug 4774291.
3483 // For Reflection.getClassAccessFlags(), the null check occurs in
3484 // the wrong place; see inline_unsafe_access(), above, for a similar
3485 // situation.
3486 mirror = null_check(mirror);
3487 // If mirror or obj is dead, only null-path is taken.
3488 if (stopped()) return true;
3490 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3492 // Now load the mirror's klass metaobject, and null-check it.
3493 // Side-effects region with the control path if the klass is null.
3494 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3495 // If kls is null, we have a primitive mirror.
3496 phi->init_req(_prim_path, prim_return_value);
3497 if (stopped()) { set_result(region, phi); return true; }
3498 bool safe_for_replace = (region->in(_prim_path) == top());
3500 Node* p; // handy temp
3501 Node* null_ctl;
3503 // Now that we have the non-null klass, we can perform the real query.
3504 // For constant classes, the query will constant-fold in LoadNode::Value.
3505 Node* query_value = top();
3506 switch (id) {
3507 case vmIntrinsics::_isInstance:
3508 // nothing is an instance of a primitive type
3509 query_value = gen_instanceof(obj, kls, safe_for_replace);
3510 break;
3512 case vmIntrinsics::_getModifiers:
3513 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3514 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3515 break;
3517 case vmIntrinsics::_isInterface:
3518 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3519 if (generate_interface_guard(kls, region) != NULL)
3520 // A guard was added. If the guard is taken, it was an interface.
3521 phi->add_req(intcon(1));
3522 // If we fall through, it's a plain class.
3523 query_value = intcon(0);
3524 break;
3526 case vmIntrinsics::_isArray:
3527 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3528 if (generate_array_guard(kls, region) != NULL)
3529 // A guard was added. If the guard is taken, it was an array.
3530 phi->add_req(intcon(1));
3531 // If we fall through, it's a plain class.
3532 query_value = intcon(0);
3533 break;
3535 case vmIntrinsics::_isPrimitive:
3536 query_value = intcon(0); // "normal" path produces false
3537 break;
3539 case vmIntrinsics::_getSuperclass:
3540 // The rules here are somewhat unfortunate, but we can still do better
3541 // with random logic than with a JNI call.
3542 // Interfaces store null or Object as _super, but must report null.
3543 // Arrays store an intermediate super as _super, but must report Object.
3544 // Other types can report the actual _super.
3545 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3546 if (generate_interface_guard(kls, region) != NULL)
3547 // A guard was added. If the guard is taken, it was an interface.
3548 phi->add_req(null());
3549 if (generate_array_guard(kls, region) != NULL)
3550 // A guard was added. If the guard is taken, it was an array.
3551 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3552 // If we fall through, it's a plain class. Get its _super.
3553 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3554 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3555 null_ctl = top();
3556 kls = null_check_oop(kls, &null_ctl);
3557 if (null_ctl != top()) {
3558 // If the guard is taken, Object.superClass is null (both klass and mirror).
3559 region->add_req(null_ctl);
3560 phi ->add_req(null());
3561 }
3562 if (!stopped()) {
3563 query_value = load_mirror_from_klass(kls);
3564 }
3565 break;
3567 case vmIntrinsics::_getComponentType:
3568 if (generate_array_guard(kls, region) != NULL) {
3569 // Be sure to pin the oop load to the guard edge just created:
3570 Node* is_array_ctrl = region->in(region->req()-1);
3571 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3572 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT, MemNode::unordered);
3573 phi->add_req(cmo);
3574 }
3575 query_value = null(); // non-array case is null
3576 break;
3578 case vmIntrinsics::_getClassAccessFlags:
3579 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3580 query_value = make_load(NULL, p, TypeInt::INT, T_INT, MemNode::unordered);
3581 break;
3583 default:
3584 fatal_unexpected_iid(id);
3585 break;
3586 }
3588 // Fall-through is the normal case of a query to a real class.
3589 phi->init_req(1, query_value);
3590 region->init_req(1, control());
3592 C->set_has_split_ifs(true); // Has chance for split-if optimization
3593 set_result(region, phi);
3594 return true;
3595 }
3597 //--------------------------inline_native_subtype_check------------------------
3598 // This intrinsic takes the JNI calls out of the heart of
3599 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3600 bool LibraryCallKit::inline_native_subtype_check() {
3601 // Pull both arguments off the stack.
3602 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3603 args[0] = argument(0);
3604 args[1] = argument(1);
3605 Node* klasses[2]; // corresponding Klasses: superk, subk
3606 klasses[0] = klasses[1] = top();
3608 enum {
3609 // A full decision tree on {superc is prim, subc is prim}:
3610 _prim_0_path = 1, // {P,N} => false
3611 // {P,P} & superc!=subc => false
3612 _prim_same_path, // {P,P} & superc==subc => true
3613 _prim_1_path, // {N,P} => false
3614 _ref_subtype_path, // {N,N} & subtype check wins => true
3615 _both_ref_path, // {N,N} & subtype check loses => false
3616 PATH_LIMIT
3617 };
3619 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3620 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3621 record_for_igvn(region);
3623 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3624 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3625 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3627 // First null-check both mirrors and load each mirror's klass metaobject.
3628 int which_arg;
3629 for (which_arg = 0; which_arg <= 1; which_arg++) {
3630 Node* arg = args[which_arg];
3631 arg = null_check(arg);
3632 if (stopped()) break;
3633 args[which_arg] = arg;
3635 Node* p = basic_plus_adr(arg, class_klass_offset);
3636 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3637 klasses[which_arg] = _gvn.transform(kls);
3638 }
3640 // Having loaded both klasses, test each for null.
3641 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3642 for (which_arg = 0; which_arg <= 1; which_arg++) {
3643 Node* kls = klasses[which_arg];
3644 Node* null_ctl = top();
3645 kls = null_check_oop(kls, &null_ctl, never_see_null);
3646 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3647 region->init_req(prim_path, null_ctl);
3648 if (stopped()) break;
3649 klasses[which_arg] = kls;
3650 }
3652 if (!stopped()) {
3653 // now we have two reference types, in klasses[0..1]
3654 Node* subk = klasses[1]; // the argument to isAssignableFrom
3655 Node* superk = klasses[0]; // the receiver
3656 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3657 // now we have a successful reference subtype check
3658 region->set_req(_ref_subtype_path, control());
3659 }
3661 // If both operands are primitive (both klasses null), then
3662 // we must return true when they are identical primitives.
3663 // It is convenient to test this after the first null klass check.
3664 set_control(region->in(_prim_0_path)); // go back to first null check
3665 if (!stopped()) {
3666 // Since superc is primitive, make a guard for the superc==subc case.
3667 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3668 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3669 generate_guard(bol_eq, region, PROB_FAIR);
3670 if (region->req() == PATH_LIMIT+1) {
3671 // A guard was added. If the added guard is taken, superc==subc.
3672 region->swap_edges(PATH_LIMIT, _prim_same_path);
3673 region->del_req(PATH_LIMIT);
3674 }
3675 region->set_req(_prim_0_path, control()); // Not equal after all.
3676 }
3678 // these are the only paths that produce 'true':
3679 phi->set_req(_prim_same_path, intcon(1));
3680 phi->set_req(_ref_subtype_path, intcon(1));
3682 // pull together the cases:
3683 assert(region->req() == PATH_LIMIT, "sane region");
3684 for (uint i = 1; i < region->req(); i++) {
3685 Node* ctl = region->in(i);
3686 if (ctl == NULL || ctl == top()) {
3687 region->set_req(i, top());
3688 phi ->set_req(i, top());
3689 } else if (phi->in(i) == NULL) {
3690 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3691 }
3692 }
3694 set_control(_gvn.transform(region));
3695 set_result(_gvn.transform(phi));
3696 return true;
3697 }
3699 //---------------------generate_array_guard_common------------------------
3700 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3701 bool obj_array, bool not_array) {
3702 // If obj_array/non_array==false/false:
3703 // Branch around if the given klass is in fact an array (either obj or prim).
3704 // If obj_array/non_array==false/true:
3705 // Branch around if the given klass is not an array klass of any kind.
3706 // If obj_array/non_array==true/true:
3707 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3708 // If obj_array/non_array==true/false:
3709 // Branch around if the kls is an oop array (Object[] or subtype)
3710 //
3711 // Like generate_guard, adds a new path onto the region.
3712 jint layout_con = 0;
3713 Node* layout_val = get_layout_helper(kls, layout_con);
3714 if (layout_val == NULL) {
3715 bool query = (obj_array
3716 ? Klass::layout_helper_is_objArray(layout_con)
3717 : Klass::layout_helper_is_array(layout_con));
3718 if (query == not_array) {
3719 return NULL; // never a branch
3720 } else { // always a branch
3721 Node* always_branch = control();
3722 if (region != NULL)
3723 region->add_req(always_branch);
3724 set_control(top());
3725 return always_branch;
3726 }
3727 }
3728 // Now test the correct condition.
3729 jint nval = (obj_array
3730 ? ((jint)Klass::_lh_array_tag_type_value
3731 << Klass::_lh_array_tag_shift)
3732 : Klass::_lh_neutral_value);
3733 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3734 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3735 // invert the test if we are looking for a non-array
3736 if (not_array) btest = BoolTest(btest).negate();
3737 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3738 return generate_fair_guard(bol, region);
3739 }
3742 //-----------------------inline_native_newArray--------------------------
3743 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3744 bool LibraryCallKit::inline_native_newArray() {
3745 Node* mirror = argument(0);
3746 Node* count_val = argument(1);
3748 mirror = null_check(mirror);
3749 // If mirror or obj is dead, only null-path is taken.
3750 if (stopped()) return true;
3752 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3753 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3754 PhiNode* result_val = new(C) PhiNode(result_reg,
3755 TypeInstPtr::NOTNULL);
3756 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3757 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3758 TypePtr::BOTTOM);
3760 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3761 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3762 result_reg, _slow_path);
3763 Node* normal_ctl = control();
3764 Node* no_array_ctl = result_reg->in(_slow_path);
3766 // Generate code for the slow case. We make a call to newArray().
3767 set_control(no_array_ctl);
3768 if (!stopped()) {
3769 // Either the input type is void.class, or else the
3770 // array klass has not yet been cached. Either the
3771 // ensuing call will throw an exception, or else it
3772 // will cache the array klass for next time.
3773 PreserveJVMState pjvms(this);
3774 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3775 Node* slow_result = set_results_for_java_call(slow_call);
3776 // this->control() comes from set_results_for_java_call
3777 result_reg->set_req(_slow_path, control());
3778 result_val->set_req(_slow_path, slow_result);
3779 result_io ->set_req(_slow_path, i_o());
3780 result_mem->set_req(_slow_path, reset_memory());
3781 }
3783 set_control(normal_ctl);
3784 if (!stopped()) {
3785 // Normal case: The array type has been cached in the java.lang.Class.
3786 // The following call works fine even if the array type is polymorphic.
3787 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3788 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3789 result_reg->init_req(_normal_path, control());
3790 result_val->init_req(_normal_path, obj);
3791 result_io ->init_req(_normal_path, i_o());
3792 result_mem->init_req(_normal_path, reset_memory());
3793 }
3795 // Return the combined state.
3796 set_i_o( _gvn.transform(result_io) );
3797 set_all_memory( _gvn.transform(result_mem));
3799 C->set_has_split_ifs(true); // Has chance for split-if optimization
3800 set_result(result_reg, result_val);
3801 return true;
3802 }
3804 //----------------------inline_native_getLength--------------------------
3805 // public static native int java.lang.reflect.Array.getLength(Object array);
3806 bool LibraryCallKit::inline_native_getLength() {
3807 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3809 Node* array = null_check(argument(0));
3810 // If array is dead, only null-path is taken.
3811 if (stopped()) return true;
3813 // Deoptimize if it is a non-array.
3814 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3816 if (non_array != NULL) {
3817 PreserveJVMState pjvms(this);
3818 set_control(non_array);
3819 uncommon_trap(Deoptimization::Reason_intrinsic,
3820 Deoptimization::Action_maybe_recompile);
3821 }
3823 // If control is dead, only non-array-path is taken.
3824 if (stopped()) return true;
3826 // The works fine even if the array type is polymorphic.
3827 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3828 Node* result = load_array_length(array);
3830 C->set_has_split_ifs(true); // Has chance for split-if optimization
3831 set_result(result);
3832 return true;
3833 }
3835 //------------------------inline_array_copyOf----------------------------
3836 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3837 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3838 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3839 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3841 // Get the arguments.
3842 Node* original = argument(0);
3843 Node* start = is_copyOfRange? argument(1): intcon(0);
3844 Node* end = is_copyOfRange? argument(2): argument(1);
3845 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3847 Node* newcopy;
3849 // Set the original stack and the reexecute bit for the interpreter to reexecute
3850 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3851 { PreserveReexecuteState preexecs(this);
3852 jvms()->set_should_reexecute(true);
3854 array_type_mirror = null_check(array_type_mirror);
3855 original = null_check(original);
3857 // Check if a null path was taken unconditionally.
3858 if (stopped()) return true;
3860 Node* orig_length = load_array_length(original);
3862 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3863 klass_node = null_check(klass_node);
3865 RegionNode* bailout = new (C) RegionNode(1);
3866 record_for_igvn(bailout);
3868 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3869 // Bail out if that is so.
3870 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3871 if (not_objArray != NULL) {
3872 // Improve the klass node's type from the new optimistic assumption:
3873 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3874 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3875 Node* cast = new (C) CastPPNode(klass_node, akls);
3876 cast->init_req(0, control());
3877 klass_node = _gvn.transform(cast);
3878 }
3880 // Bail out if either start or end is negative.
3881 generate_negative_guard(start, bailout, &start);
3882 generate_negative_guard(end, bailout, &end);
3884 Node* length = end;
3885 if (_gvn.type(start) != TypeInt::ZERO) {
3886 length = _gvn.transform(new (C) SubINode(end, start));
3887 }
3889 // Bail out if length is negative.
3890 // Without this the new_array would throw
3891 // NegativeArraySizeException but IllegalArgumentException is what
3892 // should be thrown
3893 generate_negative_guard(length, bailout, &length);
3895 if (bailout->req() > 1) {
3896 PreserveJVMState pjvms(this);
3897 set_control(_gvn.transform(bailout));
3898 uncommon_trap(Deoptimization::Reason_intrinsic,
3899 Deoptimization::Action_maybe_recompile);
3900 }
3902 if (!stopped()) {
3903 // How many elements will we copy from the original?
3904 // The answer is MinI(orig_length - start, length).
3905 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3906 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3908 newcopy = new_array(klass_node, length, 0); // no argments to push
3910 // Generate a direct call to the right arraycopy function(s).
3911 // We know the copy is disjoint but we might not know if the
3912 // oop stores need checking.
3913 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3914 // This will fail a store-check if x contains any non-nulls.
3915 bool disjoint_bases = true;
3916 // if start > orig_length then the length of the copy may be
3917 // negative.
3918 bool length_never_negative = !is_copyOfRange;
3919 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3920 original, start, newcopy, intcon(0), moved,
3921 disjoint_bases, length_never_negative);
3922 }
3923 } // original reexecute is set back here
3925 C->set_has_split_ifs(true); // Has chance for split-if optimization
3926 if (!stopped()) {
3927 set_result(newcopy);
3928 }
3929 return true;
3930 }
3933 //----------------------generate_virtual_guard---------------------------
3934 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3935 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3936 RegionNode* slow_region) {
3937 ciMethod* method = callee();
3938 int vtable_index = method->vtable_index();
3939 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3940 err_msg_res("bad index %d", vtable_index));
3941 // Get the Method* out of the appropriate vtable entry.
3942 int entry_offset = (InstanceKlass::vtable_start_offset() +
3943 vtable_index*vtableEntry::size()) * wordSize +
3944 vtableEntry::method_offset_in_bytes();
3945 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3946 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS, MemNode::unordered);
3948 // Compare the target method with the expected method (e.g., Object.hashCode).
3949 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3951 Node* native_call = makecon(native_call_addr);
3952 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3953 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3955 return generate_slow_guard(test_native, slow_region);
3956 }
3958 //-----------------------generate_method_call----------------------------
3959 // Use generate_method_call to make a slow-call to the real
3960 // method if the fast path fails. An alternative would be to
3961 // use a stub like OptoRuntime::slow_arraycopy_Java.
3962 // This only works for expanding the current library call,
3963 // not another intrinsic. (E.g., don't use this for making an
3964 // arraycopy call inside of the copyOf intrinsic.)
3965 CallJavaNode*
3966 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3967 // When compiling the intrinsic method itself, do not use this technique.
3968 guarantee(callee() != C->method(), "cannot make slow-call to self");
3970 ciMethod* method = callee();
3971 // ensure the JVMS we have will be correct for this call
3972 guarantee(method_id == method->intrinsic_id(), "must match");
3974 const TypeFunc* tf = TypeFunc::make(method);
3975 CallJavaNode* slow_call;
3976 if (is_static) {
3977 assert(!is_virtual, "");
3978 slow_call = new(C) CallStaticJavaNode(C, tf,
3979 SharedRuntime::get_resolve_static_call_stub(),
3980 method, bci());
3981 } else if (is_virtual) {
3982 null_check_receiver();
3983 int vtable_index = Method::invalid_vtable_index;
3984 if (UseInlineCaches) {
3985 // Suppress the vtable call
3986 } else {
3987 // hashCode and clone are not a miranda methods,
3988 // so the vtable index is fixed.
3989 // No need to use the linkResolver to get it.
3990 vtable_index = method->vtable_index();
3991 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3992 err_msg_res("bad index %d", vtable_index));
3993 }
3994 slow_call = new(C) CallDynamicJavaNode(tf,
3995 SharedRuntime::get_resolve_virtual_call_stub(),
3996 method, vtable_index, bci());
3997 } else { // neither virtual nor static: opt_virtual
3998 null_check_receiver();
3999 slow_call = new(C) CallStaticJavaNode(C, tf,
4000 SharedRuntime::get_resolve_opt_virtual_call_stub(),
4001 method, bci());
4002 slow_call->set_optimized_virtual(true);
4003 }
4004 set_arguments_for_java_call(slow_call);
4005 set_edges_for_java_call(slow_call);
4006 return slow_call;
4007 }
4010 //------------------------------inline_native_hashcode--------------------
4011 // Build special case code for calls to hashCode on an object.
4012 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4013 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4014 assert(!(is_virtual && is_static), "either virtual, special, or static");
4016 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4018 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4019 PhiNode* result_val = new(C) PhiNode(result_reg,
4020 TypeInt::INT);
4021 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4022 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4023 TypePtr::BOTTOM);
4024 Node* obj = NULL;
4025 if (!is_static) {
4026 // Check for hashing null object
4027 obj = null_check_receiver();
4028 if (stopped()) return true; // unconditionally null
4029 result_reg->init_req(_null_path, top());
4030 result_val->init_req(_null_path, top());
4031 } else {
4032 // Do a null check, and return zero if null.
4033 // System.identityHashCode(null) == 0
4034 obj = argument(0);
4035 Node* null_ctl = top();
4036 obj = null_check_oop(obj, &null_ctl);
4037 result_reg->init_req(_null_path, null_ctl);
4038 result_val->init_req(_null_path, _gvn.intcon(0));
4039 }
4041 // Unconditionally null? Then return right away.
4042 if (stopped()) {
4043 set_control( result_reg->in(_null_path));
4044 if (!stopped())
4045 set_result(result_val->in(_null_path));
4046 return true;
4047 }
4049 // After null check, get the object's klass.
4050 Node* obj_klass = load_object_klass(obj);
4052 // This call may be virtual (invokevirtual) or bound (invokespecial).
4053 // For each case we generate slightly different code.
4055 // We only go to the fast case code if we pass a number of guards. The
4056 // paths which do not pass are accumulated in the slow_region.
4057 RegionNode* slow_region = new (C) RegionNode(1);
4058 record_for_igvn(slow_region);
4060 // If this is a virtual call, we generate a funny guard. We pull out
4061 // the vtable entry corresponding to hashCode() from the target object.
4062 // If the target method which we are calling happens to be the native
4063 // Object hashCode() method, we pass the guard. We do not need this
4064 // guard for non-virtual calls -- the caller is known to be the native
4065 // Object hashCode().
4066 if (is_virtual) {
4067 generate_virtual_guard(obj_klass, slow_region);
4068 }
4070 // Get the header out of the object, use LoadMarkNode when available
4071 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4072 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
4074 // Test the header to see if it is unlocked.
4075 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4076 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4077 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4078 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4079 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4081 generate_slow_guard(test_unlocked, slow_region);
4083 // Get the hash value and check to see that it has been properly assigned.
4084 // We depend on hash_mask being at most 32 bits and avoid the use of
4085 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4086 // vm: see markOop.hpp.
4087 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4088 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4089 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4090 // This hack lets the hash bits live anywhere in the mark object now, as long
4091 // as the shift drops the relevant bits into the low 32 bits. Note that
4092 // Java spec says that HashCode is an int so there's no point in capturing
4093 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4094 hshifted_header = ConvX2I(hshifted_header);
4095 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4097 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4098 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4099 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4101 generate_slow_guard(test_assigned, slow_region);
4103 Node* init_mem = reset_memory();
4104 // fill in the rest of the null path:
4105 result_io ->init_req(_null_path, i_o());
4106 result_mem->init_req(_null_path, init_mem);
4108 result_val->init_req(_fast_path, hash_val);
4109 result_reg->init_req(_fast_path, control());
4110 result_io ->init_req(_fast_path, i_o());
4111 result_mem->init_req(_fast_path, init_mem);
4113 // Generate code for the slow case. We make a call to hashCode().
4114 set_control(_gvn.transform(slow_region));
4115 if (!stopped()) {
4116 // No need for PreserveJVMState, because we're using up the present state.
4117 set_all_memory(init_mem);
4118 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4119 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4120 Node* slow_result = set_results_for_java_call(slow_call);
4121 // this->control() comes from set_results_for_java_call
4122 result_reg->init_req(_slow_path, control());
4123 result_val->init_req(_slow_path, slow_result);
4124 result_io ->set_req(_slow_path, i_o());
4125 result_mem ->set_req(_slow_path, reset_memory());
4126 }
4128 // Return the combined state.
4129 set_i_o( _gvn.transform(result_io) );
4130 set_all_memory( _gvn.transform(result_mem));
4132 set_result(result_reg, result_val);
4133 return true;
4134 }
4136 //---------------------------inline_native_getClass----------------------------
4137 // public final native Class<?> java.lang.Object.getClass();
4138 //
4139 // Build special case code for calls to getClass on an object.
4140 bool LibraryCallKit::inline_native_getClass() {
4141 Node* obj = null_check_receiver();
4142 if (stopped()) return true;
4143 set_result(load_mirror_from_klass(load_object_klass(obj)));
4144 return true;
4145 }
4147 //-----------------inline_native_Reflection_getCallerClass---------------------
4148 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4149 //
4150 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4151 //
4152 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4153 // in that it must skip particular security frames and checks for
4154 // caller sensitive methods.
4155 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4156 #ifndef PRODUCT
4157 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4158 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4159 }
4160 #endif
4162 if (!jvms()->has_method()) {
4163 #ifndef PRODUCT
4164 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4165 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4166 }
4167 #endif
4168 return false;
4169 }
4171 // Walk back up the JVM state to find the caller at the required
4172 // depth.
4173 JVMState* caller_jvms = jvms();
4175 // Cf. JVM_GetCallerClass
4176 // NOTE: Start the loop at depth 1 because the current JVM state does
4177 // not include the Reflection.getCallerClass() frame.
4178 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4179 ciMethod* m = caller_jvms->method();
4180 switch (n) {
4181 case 0:
4182 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4183 break;
4184 case 1:
4185 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4186 if (!m->caller_sensitive()) {
4187 #ifndef PRODUCT
4188 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4189 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4190 }
4191 #endif
4192 return false; // bail-out; let JVM_GetCallerClass do the work
4193 }
4194 break;
4195 default:
4196 if (!m->is_ignored_by_security_stack_walk()) {
4197 // We have reached the desired frame; return the holder class.
4198 // Acquire method holder as java.lang.Class and push as constant.
4199 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4200 ciInstance* caller_mirror = caller_klass->java_mirror();
4201 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4203 #ifndef PRODUCT
4204 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4205 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());
4206 tty->print_cr(" JVM state at this point:");
4207 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4208 ciMethod* m = jvms()->of_depth(i)->method();
4209 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4210 }
4211 }
4212 #endif
4213 return true;
4214 }
4215 break;
4216 }
4217 }
4219 #ifndef PRODUCT
4220 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4221 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4222 tty->print_cr(" JVM state at this point:");
4223 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4224 ciMethod* m = jvms()->of_depth(i)->method();
4225 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4226 }
4227 }
4228 #endif
4230 return false; // bail-out; let JVM_GetCallerClass do the work
4231 }
4233 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4234 Node* arg = argument(0);
4235 Node* result;
4237 switch (id) {
4238 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4239 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4240 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4241 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4243 case vmIntrinsics::_doubleToLongBits: {
4244 // two paths (plus control) merge in a wood
4245 RegionNode *r = new (C) RegionNode(3);
4246 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4248 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4249 // Build the boolean node
4250 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4252 // Branch either way.
4253 // NaN case is less traveled, which makes all the difference.
4254 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4255 Node *opt_isnan = _gvn.transform(ifisnan);
4256 assert( opt_isnan->is_If(), "Expect an IfNode");
4257 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4258 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4260 set_control(iftrue);
4262 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4263 Node *slow_result = longcon(nan_bits); // return NaN
4264 phi->init_req(1, _gvn.transform( slow_result ));
4265 r->init_req(1, iftrue);
4267 // Else fall through
4268 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4269 set_control(iffalse);
4271 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4272 r->init_req(2, iffalse);
4274 // Post merge
4275 set_control(_gvn.transform(r));
4276 record_for_igvn(r);
4278 C->set_has_split_ifs(true); // Has chance for split-if optimization
4279 result = phi;
4280 assert(result->bottom_type()->isa_long(), "must be");
4281 break;
4282 }
4284 case vmIntrinsics::_floatToIntBits: {
4285 // two paths (plus control) merge in a wood
4286 RegionNode *r = new (C) RegionNode(3);
4287 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4289 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4290 // Build the boolean node
4291 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4293 // Branch either way.
4294 // NaN case is less traveled, which makes all the difference.
4295 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4296 Node *opt_isnan = _gvn.transform(ifisnan);
4297 assert( opt_isnan->is_If(), "Expect an IfNode");
4298 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4299 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4301 set_control(iftrue);
4303 static const jint nan_bits = 0x7fc00000;
4304 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4305 phi->init_req(1, _gvn.transform( slow_result ));
4306 r->init_req(1, iftrue);
4308 // Else fall through
4309 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4310 set_control(iffalse);
4312 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4313 r->init_req(2, iffalse);
4315 // Post merge
4316 set_control(_gvn.transform(r));
4317 record_for_igvn(r);
4319 C->set_has_split_ifs(true); // Has chance for split-if optimization
4320 result = phi;
4321 assert(result->bottom_type()->isa_int(), "must be");
4322 break;
4323 }
4325 default:
4326 fatal_unexpected_iid(id);
4327 break;
4328 }
4329 set_result(_gvn.transform(result));
4330 return true;
4331 }
4333 #ifdef _LP64
4334 #define XTOP ,top() /*additional argument*/
4335 #else //_LP64
4336 #define XTOP /*no additional argument*/
4337 #endif //_LP64
4339 //----------------------inline_unsafe_copyMemory-------------------------
4340 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4341 bool LibraryCallKit::inline_unsafe_copyMemory() {
4342 if (callee()->is_static()) return false; // caller must have the capability!
4343 null_check_receiver(); // null-check receiver
4344 if (stopped()) return true;
4346 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4348 Node* src_ptr = argument(1); // type: oop
4349 Node* src_off = ConvL2X(argument(2)); // type: long
4350 Node* dst_ptr = argument(4); // type: oop
4351 Node* dst_off = ConvL2X(argument(5)); // type: long
4352 Node* size = ConvL2X(argument(7)); // type: long
4354 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4355 "fieldOffset must be byte-scaled");
4357 Node* src = make_unsafe_address(src_ptr, src_off);
4358 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4360 // Conservatively insert a memory barrier on all memory slices.
4361 // Do not let writes of the copy source or destination float below the copy.
4362 insert_mem_bar(Op_MemBarCPUOrder);
4364 // Call it. Note that the length argument is not scaled.
4365 make_runtime_call(RC_LEAF|RC_NO_FP,
4366 OptoRuntime::fast_arraycopy_Type(),
4367 StubRoutines::unsafe_arraycopy(),
4368 "unsafe_arraycopy",
4369 TypeRawPtr::BOTTOM,
4370 src, dst, size XTOP);
4372 // Do not let reads of the copy destination float above the copy.
4373 insert_mem_bar(Op_MemBarCPUOrder);
4375 return true;
4376 }
4378 //------------------------clone_coping-----------------------------------
4379 // Helper function for inline_native_clone.
4380 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4381 assert(obj_size != NULL, "");
4382 Node* raw_obj = alloc_obj->in(1);
4383 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4385 AllocateNode* alloc = NULL;
4386 if (ReduceBulkZeroing) {
4387 // We will be completely responsible for initializing this object -
4388 // mark Initialize node as complete.
4389 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4390 // The object was just allocated - there should be no any stores!
4391 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4392 // Mark as complete_with_arraycopy so that on AllocateNode
4393 // expansion, we know this AllocateNode is initialized by an array
4394 // copy and a StoreStore barrier exists after the array copy.
4395 alloc->initialization()->set_complete_with_arraycopy();
4396 }
4398 // Copy the fastest available way.
4399 // TODO: generate fields copies for small objects instead.
4400 Node* src = obj;
4401 Node* dest = alloc_obj;
4402 Node* size = _gvn.transform(obj_size);
4404 // Exclude the header but include array length to copy by 8 bytes words.
4405 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4406 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4407 instanceOopDesc::base_offset_in_bytes();
4408 // base_off:
4409 // 8 - 32-bit VM
4410 // 12 - 64-bit VM, compressed klass
4411 // 16 - 64-bit VM, normal klass
4412 if (base_off % BytesPerLong != 0) {
4413 assert(UseCompressedClassPointers, "");
4414 if (is_array) {
4415 // Exclude length to copy by 8 bytes words.
4416 base_off += sizeof(int);
4417 } else {
4418 // Include klass to copy by 8 bytes words.
4419 base_off = instanceOopDesc::klass_offset_in_bytes();
4420 }
4421 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4422 }
4423 src = basic_plus_adr(src, base_off);
4424 dest = basic_plus_adr(dest, base_off);
4426 // Compute the length also, if needed:
4427 Node* countx = size;
4428 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4429 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4431 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4432 bool disjoint_bases = true;
4433 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4434 src, NULL, dest, NULL, countx,
4435 /*dest_uninitialized*/true);
4437 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4438 if (card_mark) {
4439 assert(!is_array, "");
4440 // Put in store barrier for any and all oops we are sticking
4441 // into this object. (We could avoid this if we could prove
4442 // that the object type contains no oop fields at all.)
4443 Node* no_particular_value = NULL;
4444 Node* no_particular_field = NULL;
4445 int raw_adr_idx = Compile::AliasIdxRaw;
4446 post_barrier(control(),
4447 memory(raw_adr_type),
4448 alloc_obj,
4449 no_particular_field,
4450 raw_adr_idx,
4451 no_particular_value,
4452 T_OBJECT,
4453 false);
4454 }
4456 // Do not let reads from the cloned object float above the arraycopy.
4457 if (alloc != NULL) {
4458 // Do not let stores that initialize this object be reordered with
4459 // a subsequent store that would make this object accessible by
4460 // other threads.
4461 // Record what AllocateNode this StoreStore protects so that
4462 // escape analysis can go from the MemBarStoreStoreNode to the
4463 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4464 // based on the escape status of the AllocateNode.
4465 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4466 } else {
4467 insert_mem_bar(Op_MemBarCPUOrder);
4468 }
4469 }
4471 //------------------------inline_native_clone----------------------------
4472 // protected native Object java.lang.Object.clone();
4473 //
4474 // Here are the simple edge cases:
4475 // null receiver => normal trap
4476 // virtual and clone was overridden => slow path to out-of-line clone
4477 // not cloneable or finalizer => slow path to out-of-line Object.clone
4478 //
4479 // The general case has two steps, allocation and copying.
4480 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4481 //
4482 // Copying also has two cases, oop arrays and everything else.
4483 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4484 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4485 //
4486 // These steps fold up nicely if and when the cloned object's klass
4487 // can be sharply typed as an object array, a type array, or an instance.
4488 //
4489 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4490 PhiNode* result_val;
4492 // Set the reexecute bit for the interpreter to reexecute
4493 // the bytecode that invokes Object.clone if deoptimization happens.
4494 { PreserveReexecuteState preexecs(this);
4495 jvms()->set_should_reexecute(true);
4497 Node* obj = null_check_receiver();
4498 if (stopped()) return true;
4500 Node* obj_klass = load_object_klass(obj);
4501 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4502 const TypeOopPtr* toop = ((tklass != NULL)
4503 ? tklass->as_instance_type()
4504 : TypeInstPtr::NOTNULL);
4506 // Conservatively insert a memory barrier on all memory slices.
4507 // Do not let writes into the original float below the clone.
4508 insert_mem_bar(Op_MemBarCPUOrder);
4510 // paths into result_reg:
4511 enum {
4512 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4513 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4514 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4515 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4516 PATH_LIMIT
4517 };
4518 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4519 result_val = new(C) PhiNode(result_reg,
4520 TypeInstPtr::NOTNULL);
4521 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4522 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4523 TypePtr::BOTTOM);
4524 record_for_igvn(result_reg);
4526 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4527 int raw_adr_idx = Compile::AliasIdxRaw;
4529 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4530 if (array_ctl != NULL) {
4531 // It's an array.
4532 PreserveJVMState pjvms(this);
4533 set_control(array_ctl);
4534 Node* obj_length = load_array_length(obj);
4535 Node* obj_size = NULL;
4536 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4538 if (!use_ReduceInitialCardMarks()) {
4539 // If it is an oop array, it requires very special treatment,
4540 // because card marking is required on each card of the array.
4541 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4542 if (is_obja != NULL) {
4543 PreserveJVMState pjvms2(this);
4544 set_control(is_obja);
4545 // Generate a direct call to the right arraycopy function(s).
4546 bool disjoint_bases = true;
4547 bool length_never_negative = true;
4548 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4549 obj, intcon(0), alloc_obj, intcon(0),
4550 obj_length,
4551 disjoint_bases, length_never_negative);
4552 result_reg->init_req(_objArray_path, control());
4553 result_val->init_req(_objArray_path, alloc_obj);
4554 result_i_o ->set_req(_objArray_path, i_o());
4555 result_mem ->set_req(_objArray_path, reset_memory());
4556 }
4557 }
4558 // Otherwise, there are no card marks to worry about.
4559 // (We can dispense with card marks if we know the allocation
4560 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4561 // causes the non-eden paths to take compensating steps to
4562 // simulate a fresh allocation, so that no further
4563 // card marks are required in compiled code to initialize
4564 // the object.)
4566 if (!stopped()) {
4567 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4569 // Present the results of the copy.
4570 result_reg->init_req(_array_path, control());
4571 result_val->init_req(_array_path, alloc_obj);
4572 result_i_o ->set_req(_array_path, i_o());
4573 result_mem ->set_req(_array_path, reset_memory());
4574 }
4575 }
4577 // We only go to the instance fast case code if we pass a number of guards.
4578 // The paths which do not pass are accumulated in the slow_region.
4579 RegionNode* slow_region = new (C) RegionNode(1);
4580 record_for_igvn(slow_region);
4581 if (!stopped()) {
4582 // It's an instance (we did array above). Make the slow-path tests.
4583 // If this is a virtual call, we generate a funny guard. We grab
4584 // the vtable entry corresponding to clone() from the target object.
4585 // If the target method which we are calling happens to be the
4586 // Object clone() method, we pass the guard. We do not need this
4587 // guard for non-virtual calls; the caller is known to be the native
4588 // Object clone().
4589 if (is_virtual) {
4590 generate_virtual_guard(obj_klass, slow_region);
4591 }
4593 // The object must be cloneable and must not have a finalizer.
4594 // Both of these conditions may be checked in a single test.
4595 // We could optimize the cloneable test further, but we don't care.
4596 generate_access_flags_guard(obj_klass,
4597 // Test both conditions:
4598 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4599 // Must be cloneable but not finalizer:
4600 JVM_ACC_IS_CLONEABLE,
4601 slow_region);
4602 }
4604 if (!stopped()) {
4605 // It's an instance, and it passed the slow-path tests.
4606 PreserveJVMState pjvms(this);
4607 Node* obj_size = NULL;
4608 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4610 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4612 // Present the results of the slow call.
4613 result_reg->init_req(_instance_path, control());
4614 result_val->init_req(_instance_path, alloc_obj);
4615 result_i_o ->set_req(_instance_path, i_o());
4616 result_mem ->set_req(_instance_path, reset_memory());
4617 }
4619 // Generate code for the slow case. We make a call to clone().
4620 set_control(_gvn.transform(slow_region));
4621 if (!stopped()) {
4622 PreserveJVMState pjvms(this);
4623 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4624 Node* slow_result = set_results_for_java_call(slow_call);
4625 // this->control() comes from set_results_for_java_call
4626 result_reg->init_req(_slow_path, control());
4627 result_val->init_req(_slow_path, slow_result);
4628 result_i_o ->set_req(_slow_path, i_o());
4629 result_mem ->set_req(_slow_path, reset_memory());
4630 }
4632 // Return the combined state.
4633 set_control( _gvn.transform(result_reg));
4634 set_i_o( _gvn.transform(result_i_o));
4635 set_all_memory( _gvn.transform(result_mem));
4636 } // original reexecute is set back here
4638 set_result(_gvn.transform(result_val));
4639 return true;
4640 }
4642 //------------------------------basictype2arraycopy----------------------------
4643 address LibraryCallKit::basictype2arraycopy(BasicType t,
4644 Node* src_offset,
4645 Node* dest_offset,
4646 bool disjoint_bases,
4647 const char* &name,
4648 bool dest_uninitialized) {
4649 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4650 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4652 bool aligned = false;
4653 bool disjoint = disjoint_bases;
4655 // if the offsets are the same, we can treat the memory regions as
4656 // disjoint, because either the memory regions are in different arrays,
4657 // or they are identical (which we can treat as disjoint.) We can also
4658 // treat a copy with a destination index less that the source index
4659 // as disjoint since a low->high copy will work correctly in this case.
4660 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4661 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4662 // both indices are constants
4663 int s_offs = src_offset_inttype->get_con();
4664 int d_offs = dest_offset_inttype->get_con();
4665 int element_size = type2aelembytes(t);
4666 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4667 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4668 if (s_offs >= d_offs) disjoint = true;
4669 } else if (src_offset == dest_offset && src_offset != NULL) {
4670 // This can occur if the offsets are identical non-constants.
4671 disjoint = true;
4672 }
4674 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4675 }
4678 //------------------------------inline_arraycopy-----------------------
4679 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4680 // Object dest, int destPos,
4681 // int length);
4682 bool LibraryCallKit::inline_arraycopy() {
4683 // Get the arguments.
4684 Node* src = argument(0); // type: oop
4685 Node* src_offset = argument(1); // type: int
4686 Node* dest = argument(2); // type: oop
4687 Node* dest_offset = argument(3); // type: int
4688 Node* length = argument(4); // type: int
4690 // Compile time checks. If any of these checks cannot be verified at compile time,
4691 // we do not make a fast path for this call. Instead, we let the call remain as it
4692 // is. The checks we choose to mandate at compile time are:
4693 //
4694 // (1) src and dest are arrays.
4695 const Type* src_type = src->Value(&_gvn);
4696 const Type* dest_type = dest->Value(&_gvn);
4697 const TypeAryPtr* top_src = src_type->isa_aryptr();
4698 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4700 // Do we have the type of src?
4701 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4702 // Do we have the type of dest?
4703 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4704 // Is the type for src from speculation?
4705 bool src_spec = false;
4706 // Is the type for dest from speculation?
4707 bool dest_spec = false;
4709 if (!has_src || !has_dest) {
4710 // We don't have sufficient type information, let's see if
4711 // speculative types can help. We need to have types for both src
4712 // and dest so that it pays off.
4714 // Do we already have or could we have type information for src
4715 bool could_have_src = has_src;
4716 // Do we already have or could we have type information for dest
4717 bool could_have_dest = has_dest;
4719 ciKlass* src_k = NULL;
4720 if (!has_src) {
4721 src_k = src_type->speculative_type();
4722 if (src_k != NULL && src_k->is_array_klass()) {
4723 could_have_src = true;
4724 }
4725 }
4727 ciKlass* dest_k = NULL;
4728 if (!has_dest) {
4729 dest_k = dest_type->speculative_type();
4730 if (dest_k != NULL && dest_k->is_array_klass()) {
4731 could_have_dest = true;
4732 }
4733 }
4735 if (could_have_src && could_have_dest) {
4736 // This is going to pay off so emit the required guards
4737 if (!has_src) {
4738 src = maybe_cast_profiled_obj(src, src_k);
4739 src_type = _gvn.type(src);
4740 top_src = src_type->isa_aryptr();
4741 has_src = (top_src != NULL && top_src->klass() != NULL);
4742 src_spec = true;
4743 }
4744 if (!has_dest) {
4745 dest = maybe_cast_profiled_obj(dest, dest_k);
4746 dest_type = _gvn.type(dest);
4747 top_dest = dest_type->isa_aryptr();
4748 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4749 dest_spec = true;
4750 }
4751 }
4752 }
4754 if (!has_src || !has_dest) {
4755 // Conservatively insert a memory barrier on all memory slices.
4756 // Do not let writes into the source float below the arraycopy.
4757 insert_mem_bar(Op_MemBarCPUOrder);
4759 // Call StubRoutines::generic_arraycopy stub.
4760 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4761 src, src_offset, dest, dest_offset, length);
4763 // Do not let reads from the destination float above the arraycopy.
4764 // Since we cannot type the arrays, we don't know which slices
4765 // might be affected. We could restrict this barrier only to those
4766 // memory slices which pertain to array elements--but don't bother.
4767 if (!InsertMemBarAfterArraycopy)
4768 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4769 insert_mem_bar(Op_MemBarCPUOrder);
4770 return true;
4771 }
4773 // (2) src and dest arrays must have elements of the same BasicType
4774 // Figure out the size and type of the elements we will be copying.
4775 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4776 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4777 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4778 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4780 if (src_elem != dest_elem || dest_elem == T_VOID) {
4781 // The component types are not the same or are not recognized. Punt.
4782 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4783 generate_slow_arraycopy(TypePtr::BOTTOM,
4784 src, src_offset, dest, dest_offset, length,
4785 /*dest_uninitialized*/false);
4786 return true;
4787 }
4789 if (src_elem == T_OBJECT) {
4790 // If both arrays are object arrays then having the exact types
4791 // for both will remove the need for a subtype check at runtime
4792 // before the call and may make it possible to pick a faster copy
4793 // routine (without a subtype check on every element)
4794 // Do we have the exact type of src?
4795 bool could_have_src = src_spec;
4796 // Do we have the exact type of dest?
4797 bool could_have_dest = dest_spec;
4798 ciKlass* src_k = top_src->klass();
4799 ciKlass* dest_k = top_dest->klass();
4800 if (!src_spec) {
4801 src_k = src_type->speculative_type();
4802 if (src_k != NULL && src_k->is_array_klass()) {
4803 could_have_src = true;
4804 }
4805 }
4806 if (!dest_spec) {
4807 dest_k = dest_type->speculative_type();
4808 if (dest_k != NULL && dest_k->is_array_klass()) {
4809 could_have_dest = true;
4810 }
4811 }
4812 if (could_have_src && could_have_dest) {
4813 // If we can have both exact types, emit the missing guards
4814 if (could_have_src && !src_spec) {
4815 src = maybe_cast_profiled_obj(src, src_k);
4816 }
4817 if (could_have_dest && !dest_spec) {
4818 dest = maybe_cast_profiled_obj(dest, dest_k);
4819 }
4820 }
4821 }
4823 //---------------------------------------------------------------------------
4824 // We will make a fast path for this call to arraycopy.
4826 // We have the following tests left to perform:
4827 //
4828 // (3) src and dest must not be null.
4829 // (4) src_offset must not be negative.
4830 // (5) dest_offset must not be negative.
4831 // (6) length must not be negative.
4832 // (7) src_offset + length must not exceed length of src.
4833 // (8) dest_offset + length must not exceed length of dest.
4834 // (9) each element of an oop array must be assignable
4836 RegionNode* slow_region = new (C) RegionNode(1);
4837 record_for_igvn(slow_region);
4839 // (3) operands must not be null
4840 // We currently perform our null checks with the null_check routine.
4841 // This means that the null exceptions will be reported in the caller
4842 // rather than (correctly) reported inside of the native arraycopy call.
4843 // This should be corrected, given time. We do our null check with the
4844 // stack pointer restored.
4845 src = null_check(src, T_ARRAY);
4846 dest = null_check(dest, T_ARRAY);
4848 // (4) src_offset must not be negative.
4849 generate_negative_guard(src_offset, slow_region);
4851 // (5) dest_offset must not be negative.
4852 generate_negative_guard(dest_offset, slow_region);
4854 // (6) length must not be negative (moved to generate_arraycopy()).
4855 // generate_negative_guard(length, slow_region);
4857 // (7) src_offset + length must not exceed length of src.
4858 generate_limit_guard(src_offset, length,
4859 load_array_length(src),
4860 slow_region);
4862 // (8) dest_offset + length must not exceed length of dest.
4863 generate_limit_guard(dest_offset, length,
4864 load_array_length(dest),
4865 slow_region);
4867 // (9) each element of an oop array must be assignable
4868 // The generate_arraycopy subroutine checks this.
4870 // This is where the memory effects are placed:
4871 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4872 generate_arraycopy(adr_type, dest_elem,
4873 src, src_offset, dest, dest_offset, length,
4874 false, false, slow_region);
4876 return true;
4877 }
4879 //-----------------------------generate_arraycopy----------------------
4880 // Generate an optimized call to arraycopy.
4881 // Caller must guard against non-arrays.
4882 // Caller must determine a common array basic-type for both arrays.
4883 // Caller must validate offsets against array bounds.
4884 // The slow_region has already collected guard failure paths
4885 // (such as out of bounds length or non-conformable array types).
4886 // The generated code has this shape, in general:
4887 //
4888 // if (length == 0) return // via zero_path
4889 // slowval = -1
4890 // if (types unknown) {
4891 // slowval = call generic copy loop
4892 // if (slowval == 0) return // via checked_path
4893 // } else if (indexes in bounds) {
4894 // if ((is object array) && !(array type check)) {
4895 // slowval = call checked copy loop
4896 // if (slowval == 0) return // via checked_path
4897 // } else {
4898 // call bulk copy loop
4899 // return // via fast_path
4900 // }
4901 // }
4902 // // adjust params for remaining work:
4903 // if (slowval != -1) {
4904 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4905 // }
4906 // slow_region:
4907 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4908 // return // via slow_call_path
4909 //
4910 // This routine is used from several intrinsics: System.arraycopy,
4911 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4912 //
4913 void
4914 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4915 BasicType basic_elem_type,
4916 Node* src, Node* src_offset,
4917 Node* dest, Node* dest_offset,
4918 Node* copy_length,
4919 bool disjoint_bases,
4920 bool length_never_negative,
4921 RegionNode* slow_region) {
4923 if (slow_region == NULL) {
4924 slow_region = new(C) RegionNode(1);
4925 record_for_igvn(slow_region);
4926 }
4928 Node* original_dest = dest;
4929 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4930 bool dest_uninitialized = false;
4932 // See if this is the initialization of a newly-allocated array.
4933 // If so, we will take responsibility here for initializing it to zero.
4934 // (Note: Because tightly_coupled_allocation performs checks on the
4935 // out-edges of the dest, we need to avoid making derived pointers
4936 // from it until we have checked its uses.)
4937 if (ReduceBulkZeroing
4938 && !ZeroTLAB // pointless if already zeroed
4939 && basic_elem_type != T_CONFLICT // avoid corner case
4940 && !src->eqv_uncast(dest)
4941 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4942 != NULL)
4943 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4944 && alloc->maybe_set_complete(&_gvn)) {
4945 // "You break it, you buy it."
4946 InitializeNode* init = alloc->initialization();
4947 assert(init->is_complete(), "we just did this");
4948 init->set_complete_with_arraycopy();
4949 assert(dest->is_CheckCastPP(), "sanity");
4950 assert(dest->in(0)->in(0) == init, "dest pinned");
4951 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4952 // From this point on, every exit path is responsible for
4953 // initializing any non-copied parts of the object to zero.
4954 // Also, if this flag is set we make sure that arraycopy interacts properly
4955 // with G1, eliding pre-barriers. See CR 6627983.
4956 dest_uninitialized = true;
4957 } else {
4958 // No zeroing elimination here.
4959 alloc = NULL;
4960 //original_dest = dest;
4961 //dest_uninitialized = false;
4962 }
4964 // Results are placed here:
4965 enum { fast_path = 1, // normal void-returning assembly stub
4966 checked_path = 2, // special assembly stub with cleanup
4967 slow_call_path = 3, // something went wrong; call the VM
4968 zero_path = 4, // bypass when length of copy is zero
4969 bcopy_path = 5, // copy primitive array by 64-bit blocks
4970 PATH_LIMIT = 6
4971 };
4972 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4973 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4974 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4975 record_for_igvn(result_region);
4976 _gvn.set_type_bottom(result_i_o);
4977 _gvn.set_type_bottom(result_memory);
4978 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4980 // The slow_control path:
4981 Node* slow_control;
4982 Node* slow_i_o = i_o();
4983 Node* slow_mem = memory(adr_type);
4984 debug_only(slow_control = (Node*) badAddress);
4986 // Checked control path:
4987 Node* checked_control = top();
4988 Node* checked_mem = NULL;
4989 Node* checked_i_o = NULL;
4990 Node* checked_value = NULL;
4992 if (basic_elem_type == T_CONFLICT) {
4993 assert(!dest_uninitialized, "");
4994 Node* cv = generate_generic_arraycopy(adr_type,
4995 src, src_offset, dest, dest_offset,
4996 copy_length, dest_uninitialized);
4997 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4998 checked_control = control();
4999 checked_i_o = i_o();
5000 checked_mem = memory(adr_type);
5001 checked_value = cv;
5002 set_control(top()); // no fast path
5003 }
5005 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
5006 if (not_pos != NULL) {
5007 PreserveJVMState pjvms(this);
5008 set_control(not_pos);
5010 // (6) length must not be negative.
5011 if (!length_never_negative) {
5012 generate_negative_guard(copy_length, slow_region);
5013 }
5015 // copy_length is 0.
5016 if (!stopped() && dest_uninitialized) {
5017 Node* dest_length = alloc->in(AllocateNode::ALength);
5018 if (copy_length->eqv_uncast(dest_length)
5019 || _gvn.find_int_con(dest_length, 1) <= 0) {
5020 // There is no zeroing to do. No need for a secondary raw memory barrier.
5021 } else {
5022 // Clear the whole thing since there are no source elements to copy.
5023 generate_clear_array(adr_type, dest, basic_elem_type,
5024 intcon(0), NULL,
5025 alloc->in(AllocateNode::AllocSize));
5026 // Use a secondary InitializeNode as raw memory barrier.
5027 // Currently it is needed only on this path since other
5028 // paths have stub or runtime calls as raw memory barriers.
5029 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5030 Compile::AliasIdxRaw,
5031 top())->as_Initialize();
5032 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5033 }
5034 }
5036 // Present the results of the fast call.
5037 result_region->init_req(zero_path, control());
5038 result_i_o ->init_req(zero_path, i_o());
5039 result_memory->init_req(zero_path, memory(adr_type));
5040 }
5042 if (!stopped() && dest_uninitialized) {
5043 // We have to initialize the *uncopied* part of the array to zero.
5044 // The copy destination is the slice dest[off..off+len]. The other slices
5045 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5046 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5047 Node* dest_length = alloc->in(AllocateNode::ALength);
5048 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
5049 copy_length));
5051 // If there is a head section that needs zeroing, do it now.
5052 if (find_int_con(dest_offset, -1) != 0) {
5053 generate_clear_array(adr_type, dest, basic_elem_type,
5054 intcon(0), dest_offset,
5055 NULL);
5056 }
5058 // Next, perform a dynamic check on the tail length.
5059 // It is often zero, and we can win big if we prove this.
5060 // There are two wins: Avoid generating the ClearArray
5061 // with its attendant messy index arithmetic, and upgrade
5062 // the copy to a more hardware-friendly word size of 64 bits.
5063 Node* tail_ctl = NULL;
5064 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5065 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
5066 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
5067 tail_ctl = generate_slow_guard(bol_lt, NULL);
5068 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5069 }
5071 // At this point, let's assume there is no tail.
5072 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5073 // There is no tail. Try an upgrade to a 64-bit copy.
5074 bool didit = false;
5075 { PreserveJVMState pjvms(this);
5076 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5077 src, src_offset, dest, dest_offset,
5078 dest_size, dest_uninitialized);
5079 if (didit) {
5080 // Present the results of the block-copying fast call.
5081 result_region->init_req(bcopy_path, control());
5082 result_i_o ->init_req(bcopy_path, i_o());
5083 result_memory->init_req(bcopy_path, memory(adr_type));
5084 }
5085 }
5086 if (didit)
5087 set_control(top()); // no regular fast path
5088 }
5090 // Clear the tail, if any.
5091 if (tail_ctl != NULL) {
5092 Node* notail_ctl = stopped() ? NULL : control();
5093 set_control(tail_ctl);
5094 if (notail_ctl == NULL) {
5095 generate_clear_array(adr_type, dest, basic_elem_type,
5096 dest_tail, NULL,
5097 dest_size);
5098 } else {
5099 // Make a local merge.
5100 Node* done_ctl = new(C) RegionNode(3);
5101 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5102 done_ctl->init_req(1, notail_ctl);
5103 done_mem->init_req(1, memory(adr_type));
5104 generate_clear_array(adr_type, dest, basic_elem_type,
5105 dest_tail, NULL,
5106 dest_size);
5107 done_ctl->init_req(2, control());
5108 done_mem->init_req(2, memory(adr_type));
5109 set_control( _gvn.transform(done_ctl));
5110 set_memory( _gvn.transform(done_mem), adr_type );
5111 }
5112 }
5113 }
5115 BasicType copy_type = basic_elem_type;
5116 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5117 if (!stopped() && copy_type == T_OBJECT) {
5118 // If src and dest have compatible element types, we can copy bits.
5119 // Types S[] and D[] are compatible if D is a supertype of S.
5120 //
5121 // If they are not, we will use checked_oop_disjoint_arraycopy,
5122 // which performs a fast optimistic per-oop check, and backs off
5123 // further to JVM_ArrayCopy on the first per-oop check that fails.
5124 // (Actually, we don't move raw bits only; the GC requires card marks.)
5126 // Get the Klass* for both src and dest
5127 Node* src_klass = load_object_klass(src);
5128 Node* dest_klass = load_object_klass(dest);
5130 // Generate the subtype check.
5131 // This might fold up statically, or then again it might not.
5132 //
5133 // Non-static example: Copying List<String>.elements to a new String[].
5134 // The backing store for a List<String> is always an Object[],
5135 // but its elements are always type String, if the generic types
5136 // are correct at the source level.
5137 //
5138 // Test S[] against D[], not S against D, because (probably)
5139 // the secondary supertype cache is less busy for S[] than S.
5140 // This usually only matters when D is an interface.
5141 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5142 // Plug failing path into checked_oop_disjoint_arraycopy
5143 if (not_subtype_ctrl != top()) {
5144 PreserveJVMState pjvms(this);
5145 set_control(not_subtype_ctrl);
5146 // (At this point we can assume disjoint_bases, since types differ.)
5147 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5148 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5149 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5150 Node* dest_elem_klass = _gvn.transform(n1);
5151 Node* cv = generate_checkcast_arraycopy(adr_type,
5152 dest_elem_klass,
5153 src, src_offset, dest, dest_offset,
5154 ConvI2X(copy_length), dest_uninitialized);
5155 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5156 checked_control = control();
5157 checked_i_o = i_o();
5158 checked_mem = memory(adr_type);
5159 checked_value = cv;
5160 }
5161 // At this point we know we do not need type checks on oop stores.
5163 // Let's see if we need card marks:
5164 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5165 // If we do not need card marks, copy using the jint or jlong stub.
5166 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5167 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5168 "sizes agree");
5169 }
5170 }
5172 if (!stopped()) {
5173 // Generate the fast path, if possible.
5174 PreserveJVMState pjvms(this);
5175 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5176 src, src_offset, dest, dest_offset,
5177 ConvI2X(copy_length), dest_uninitialized);
5179 // Present the results of the fast call.
5180 result_region->init_req(fast_path, control());
5181 result_i_o ->init_req(fast_path, i_o());
5182 result_memory->init_req(fast_path, memory(adr_type));
5183 }
5185 // Here are all the slow paths up to this point, in one bundle:
5186 slow_control = top();
5187 if (slow_region != NULL)
5188 slow_control = _gvn.transform(slow_region);
5189 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5191 set_control(checked_control);
5192 if (!stopped()) {
5193 // Clean up after the checked call.
5194 // The returned value is either 0 or -1^K,
5195 // where K = number of partially transferred array elements.
5196 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5197 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5198 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5200 // If it is 0, we are done, so transfer to the end.
5201 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5202 result_region->init_req(checked_path, checks_done);
5203 result_i_o ->init_req(checked_path, checked_i_o);
5204 result_memory->init_req(checked_path, checked_mem);
5206 // If it is not zero, merge into the slow call.
5207 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5208 RegionNode* slow_reg2 = new(C) RegionNode(3);
5209 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5210 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5211 record_for_igvn(slow_reg2);
5212 slow_reg2 ->init_req(1, slow_control);
5213 slow_i_o2 ->init_req(1, slow_i_o);
5214 slow_mem2 ->init_req(1, slow_mem);
5215 slow_reg2 ->init_req(2, control());
5216 slow_i_o2 ->init_req(2, checked_i_o);
5217 slow_mem2 ->init_req(2, checked_mem);
5219 slow_control = _gvn.transform(slow_reg2);
5220 slow_i_o = _gvn.transform(slow_i_o2);
5221 slow_mem = _gvn.transform(slow_mem2);
5223 if (alloc != NULL) {
5224 // We'll restart from the very beginning, after zeroing the whole thing.
5225 // This can cause double writes, but that's OK since dest is brand new.
5226 // So we ignore the low 31 bits of the value returned from the stub.
5227 } else {
5228 // We must continue the copy exactly where it failed, or else
5229 // another thread might see the wrong number of writes to dest.
5230 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5231 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5232 slow_offset->init_req(1, intcon(0));
5233 slow_offset->init_req(2, checked_offset);
5234 slow_offset = _gvn.transform(slow_offset);
5236 // Adjust the arguments by the conditionally incoming offset.
5237 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5238 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5239 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5241 // Tweak the node variables to adjust the code produced below:
5242 src_offset = src_off_plus;
5243 dest_offset = dest_off_plus;
5244 copy_length = length_minus;
5245 }
5246 }
5248 set_control(slow_control);
5249 if (!stopped()) {
5250 // Generate the slow path, if needed.
5251 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5253 set_memory(slow_mem, adr_type);
5254 set_i_o(slow_i_o);
5256 if (dest_uninitialized) {
5257 generate_clear_array(adr_type, dest, basic_elem_type,
5258 intcon(0), NULL,
5259 alloc->in(AllocateNode::AllocSize));
5260 }
5262 generate_slow_arraycopy(adr_type,
5263 src, src_offset, dest, dest_offset,
5264 copy_length, /*dest_uninitialized*/false);
5266 result_region->init_req(slow_call_path, control());
5267 result_i_o ->init_req(slow_call_path, i_o());
5268 result_memory->init_req(slow_call_path, memory(adr_type));
5269 }
5271 // Remove unused edges.
5272 for (uint i = 1; i < result_region->req(); i++) {
5273 if (result_region->in(i) == NULL)
5274 result_region->init_req(i, top());
5275 }
5277 // Finished; return the combined state.
5278 set_control( _gvn.transform(result_region));
5279 set_i_o( _gvn.transform(result_i_o) );
5280 set_memory( _gvn.transform(result_memory), adr_type );
5282 // The memory edges above are precise in order to model effects around
5283 // array copies accurately to allow value numbering of field loads around
5284 // arraycopy. Such field loads, both before and after, are common in Java
5285 // collections and similar classes involving header/array data structures.
5286 //
5287 // But with low number of register or when some registers are used or killed
5288 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5289 // The next memory barrier is added to avoid it. If the arraycopy can be
5290 // optimized away (which it can, sometimes) then we can manually remove
5291 // the membar also.
5292 //
5293 // Do not let reads from the cloned object float above the arraycopy.
5294 if (alloc != NULL) {
5295 // Do not let stores that initialize this object be reordered with
5296 // a subsequent store that would make this object accessible by
5297 // other threads.
5298 // Record what AllocateNode this StoreStore protects so that
5299 // escape analysis can go from the MemBarStoreStoreNode to the
5300 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5301 // based on the escape status of the AllocateNode.
5302 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5303 } else if (InsertMemBarAfterArraycopy)
5304 insert_mem_bar(Op_MemBarCPUOrder);
5305 }
5308 // Helper function which determines if an arraycopy immediately follows
5309 // an allocation, with no intervening tests or other escapes for the object.
5310 AllocateArrayNode*
5311 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5312 RegionNode* slow_region) {
5313 if (stopped()) return NULL; // no fast path
5314 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5316 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5317 if (alloc == NULL) return NULL;
5319 Node* rawmem = memory(Compile::AliasIdxRaw);
5320 // Is the allocation's memory state untouched?
5321 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5322 // Bail out if there have been raw-memory effects since the allocation.
5323 // (Example: There might have been a call or safepoint.)
5324 return NULL;
5325 }
5326 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5327 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5328 return NULL;
5329 }
5331 // There must be no unexpected observers of this allocation.
5332 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5333 Node* obs = ptr->fast_out(i);
5334 if (obs != this->map()) {
5335 return NULL;
5336 }
5337 }
5339 // This arraycopy must unconditionally follow the allocation of the ptr.
5340 Node* alloc_ctl = ptr->in(0);
5341 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5343 Node* ctl = control();
5344 while (ctl != alloc_ctl) {
5345 // There may be guards which feed into the slow_region.
5346 // Any other control flow means that we might not get a chance
5347 // to finish initializing the allocated object.
5348 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5349 IfNode* iff = ctl->in(0)->as_If();
5350 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5351 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5352 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5353 ctl = iff->in(0); // This test feeds the known slow_region.
5354 continue;
5355 }
5356 // One more try: Various low-level checks bottom out in
5357 // uncommon traps. If the debug-info of the trap omits
5358 // any reference to the allocation, as we've already
5359 // observed, then there can be no objection to the trap.
5360 bool found_trap = false;
5361 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5362 Node* obs = not_ctl->fast_out(j);
5363 if (obs->in(0) == not_ctl && obs->is_Call() &&
5364 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5365 found_trap = true; break;
5366 }
5367 }
5368 if (found_trap) {
5369 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5370 continue;
5371 }
5372 }
5373 return NULL;
5374 }
5376 // If we get this far, we have an allocation which immediately
5377 // precedes the arraycopy, and we can take over zeroing the new object.
5378 // The arraycopy will finish the initialization, and provide
5379 // a new control state to which we will anchor the destination pointer.
5381 return alloc;
5382 }
5384 // Helper for initialization of arrays, creating a ClearArray.
5385 // It writes zero bits in [start..end), within the body of an array object.
5386 // The memory effects are all chained onto the 'adr_type' alias category.
5387 //
5388 // Since the object is otherwise uninitialized, we are free
5389 // to put a little "slop" around the edges of the cleared area,
5390 // as long as it does not go back into the array's header,
5391 // or beyond the array end within the heap.
5392 //
5393 // The lower edge can be rounded down to the nearest jint and the
5394 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5395 //
5396 // Arguments:
5397 // adr_type memory slice where writes are generated
5398 // dest oop of the destination array
5399 // basic_elem_type element type of the destination
5400 // slice_idx array index of first element to store
5401 // slice_len number of elements to store (or NULL)
5402 // dest_size total size in bytes of the array object
5403 //
5404 // Exactly one of slice_len or dest_size must be non-NULL.
5405 // If dest_size is non-NULL, zeroing extends to the end of the object.
5406 // If slice_len is non-NULL, the slice_idx value must be a constant.
5407 void
5408 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5409 Node* dest,
5410 BasicType basic_elem_type,
5411 Node* slice_idx,
5412 Node* slice_len,
5413 Node* dest_size) {
5414 // one or the other but not both of slice_len and dest_size:
5415 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5416 if (slice_len == NULL) slice_len = top();
5417 if (dest_size == NULL) dest_size = top();
5419 // operate on this memory slice:
5420 Node* mem = memory(adr_type); // memory slice to operate on
5422 // scaling and rounding of indexes:
5423 int scale = exact_log2(type2aelembytes(basic_elem_type));
5424 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5425 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5426 int bump_bit = (-1 << scale) & BytesPerInt;
5428 // determine constant starts and ends
5429 const intptr_t BIG_NEG = -128;
5430 assert(BIG_NEG + 2*abase < 0, "neg enough");
5431 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5432 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5433 if (slice_len_con == 0) {
5434 return; // nothing to do here
5435 }
5436 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5437 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5438 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5439 assert(end_con < 0, "not two cons");
5440 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5441 BytesPerLong);
5442 }
5444 if (start_con >= 0 && end_con >= 0) {
5445 // Constant start and end. Simple.
5446 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5447 start_con, end_con, &_gvn);
5448 } else if (start_con >= 0 && dest_size != top()) {
5449 // Constant start, pre-rounded end after the tail of the array.
5450 Node* end = dest_size;
5451 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5452 start_con, end, &_gvn);
5453 } else if (start_con >= 0 && slice_len != top()) {
5454 // Constant start, non-constant end. End needs rounding up.
5455 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5456 intptr_t end_base = abase + (slice_idx_con << scale);
5457 int end_round = (-1 << scale) & (BytesPerLong - 1);
5458 Node* end = ConvI2X(slice_len);
5459 if (scale != 0)
5460 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5461 end_base += end_round;
5462 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5463 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5464 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5465 start_con, end, &_gvn);
5466 } else if (start_con < 0 && dest_size != top()) {
5467 // Non-constant start, pre-rounded end after the tail of the array.
5468 // This is almost certainly a "round-to-end" operation.
5469 Node* start = slice_idx;
5470 start = ConvI2X(start);
5471 if (scale != 0)
5472 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5473 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5474 if ((bump_bit | clear_low) != 0) {
5475 int to_clear = (bump_bit | clear_low);
5476 // Align up mod 8, then store a jint zero unconditionally
5477 // just before the mod-8 boundary.
5478 if (((abase + bump_bit) & ~to_clear) - bump_bit
5479 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5480 bump_bit = 0;
5481 assert((abase & to_clear) == 0, "array base must be long-aligned");
5482 } else {
5483 // Bump 'start' up to (or past) the next jint boundary:
5484 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5485 assert((abase & clear_low) == 0, "array base must be int-aligned");
5486 }
5487 // Round bumped 'start' down to jlong boundary in body of array.
5488 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5489 if (bump_bit != 0) {
5490 // Store a zero to the immediately preceding jint:
5491 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5492 Node* p1 = basic_plus_adr(dest, x1);
5493 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT, MemNode::unordered);
5494 mem = _gvn.transform(mem);
5495 }
5496 }
5497 Node* end = dest_size; // pre-rounded
5498 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5499 start, end, &_gvn);
5500 } else {
5501 // Non-constant start, unrounded non-constant end.
5502 // (Nobody zeroes a random midsection of an array using this routine.)
5503 ShouldNotReachHere(); // fix caller
5504 }
5506 // Done.
5507 set_memory(mem, adr_type);
5508 }
5511 bool
5512 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5513 BasicType basic_elem_type,
5514 AllocateNode* alloc,
5515 Node* src, Node* src_offset,
5516 Node* dest, Node* dest_offset,
5517 Node* dest_size, bool dest_uninitialized) {
5518 // See if there is an advantage from block transfer.
5519 int scale = exact_log2(type2aelembytes(basic_elem_type));
5520 if (scale >= LogBytesPerLong)
5521 return false; // it is already a block transfer
5523 // Look at the alignment of the starting offsets.
5524 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5526 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5527 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5528 if (src_off_con < 0 || dest_off_con < 0)
5529 // At present, we can only understand constants.
5530 return false;
5532 intptr_t src_off = abase + (src_off_con << scale);
5533 intptr_t dest_off = abase + (dest_off_con << scale);
5535 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5536 // Non-aligned; too bad.
5537 // One more chance: Pick off an initial 32-bit word.
5538 // This is a common case, since abase can be odd mod 8.
5539 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5540 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5541 Node* sptr = basic_plus_adr(src, src_off);
5542 Node* dptr = basic_plus_adr(dest, dest_off);
5543 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type, MemNode::unordered);
5544 store_to_memory(control(), dptr, sval, T_INT, adr_type, MemNode::unordered);
5545 src_off += BytesPerInt;
5546 dest_off += BytesPerInt;
5547 } else {
5548 return false;
5549 }
5550 }
5551 assert(src_off % BytesPerLong == 0, "");
5552 assert(dest_off % BytesPerLong == 0, "");
5554 // Do this copy by giant steps.
5555 Node* sptr = basic_plus_adr(src, src_off);
5556 Node* dptr = basic_plus_adr(dest, dest_off);
5557 Node* countx = dest_size;
5558 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5559 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5561 bool disjoint_bases = true; // since alloc != NULL
5562 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5563 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5565 return true;
5566 }
5569 // Helper function; generates code for the slow case.
5570 // We make a call to a runtime method which emulates the native method,
5571 // but without the native wrapper overhead.
5572 void
5573 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5574 Node* src, Node* src_offset,
5575 Node* dest, Node* dest_offset,
5576 Node* copy_length, bool dest_uninitialized) {
5577 assert(!dest_uninitialized, "Invariant");
5578 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5579 OptoRuntime::slow_arraycopy_Type(),
5580 OptoRuntime::slow_arraycopy_Java(),
5581 "slow_arraycopy", adr_type,
5582 src, src_offset, dest, dest_offset,
5583 copy_length);
5585 // Handle exceptions thrown by this fellow:
5586 make_slow_call_ex(call, env()->Throwable_klass(), false);
5587 }
5589 // Helper function; generates code for cases requiring runtime checks.
5590 Node*
5591 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5592 Node* dest_elem_klass,
5593 Node* src, Node* src_offset,
5594 Node* dest, Node* dest_offset,
5595 Node* copy_length, bool dest_uninitialized) {
5596 if (stopped()) return NULL;
5598 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5599 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5600 return NULL;
5601 }
5603 // Pick out the parameters required to perform a store-check
5604 // for the target array. This is an optimistic check. It will
5605 // look in each non-null element's class, at the desired klass's
5606 // super_check_offset, for the desired klass.
5607 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5608 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5609 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr(), TypeInt::INT, MemNode::unordered);
5610 Node* check_offset = ConvI2X(_gvn.transform(n3));
5611 Node* check_value = dest_elem_klass;
5613 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5614 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5616 // (We know the arrays are never conjoint, because their types differ.)
5617 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5618 OptoRuntime::checkcast_arraycopy_Type(),
5619 copyfunc_addr, "checkcast_arraycopy", adr_type,
5620 // five arguments, of which two are
5621 // intptr_t (jlong in LP64)
5622 src_start, dest_start,
5623 copy_length XTOP,
5624 check_offset XTOP,
5625 check_value);
5627 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5628 }
5631 // Helper function; generates code for cases requiring runtime checks.
5632 Node*
5633 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5634 Node* src, Node* src_offset,
5635 Node* dest, Node* dest_offset,
5636 Node* copy_length, bool dest_uninitialized) {
5637 assert(!dest_uninitialized, "Invariant");
5638 if (stopped()) return NULL;
5639 address copyfunc_addr = StubRoutines::generic_arraycopy();
5640 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5641 return NULL;
5642 }
5644 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5645 OptoRuntime::generic_arraycopy_Type(),
5646 copyfunc_addr, "generic_arraycopy", adr_type,
5647 src, src_offset, dest, dest_offset, copy_length);
5649 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5650 }
5652 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5653 void
5654 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5655 BasicType basic_elem_type,
5656 bool disjoint_bases,
5657 Node* src, Node* src_offset,
5658 Node* dest, Node* dest_offset,
5659 Node* copy_length, bool dest_uninitialized) {
5660 if (stopped()) return; // nothing to do
5662 Node* src_start = src;
5663 Node* dest_start = dest;
5664 if (src_offset != NULL || dest_offset != NULL) {
5665 assert(src_offset != NULL && dest_offset != NULL, "");
5666 src_start = array_element_address(src, src_offset, basic_elem_type);
5667 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5668 }
5670 // Figure out which arraycopy runtime method to call.
5671 const char* copyfunc_name = "arraycopy";
5672 address copyfunc_addr =
5673 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5674 disjoint_bases, copyfunc_name, dest_uninitialized);
5676 // Call it. Note that the count_ix value is not scaled to a byte-size.
5677 make_runtime_call(RC_LEAF|RC_NO_FP,
5678 OptoRuntime::fast_arraycopy_Type(),
5679 copyfunc_addr, copyfunc_name, adr_type,
5680 src_start, dest_start, copy_length XTOP);
5681 }
5683 //-------------inline_encodeISOArray-----------------------------------
5684 // encode char[] to byte[] in ISO_8859_1
5685 bool LibraryCallKit::inline_encodeISOArray() {
5686 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5687 // no receiver since it is static method
5688 Node *src = argument(0);
5689 Node *src_offset = argument(1);
5690 Node *dst = argument(2);
5691 Node *dst_offset = argument(3);
5692 Node *length = argument(4);
5694 const Type* src_type = src->Value(&_gvn);
5695 const Type* dst_type = dst->Value(&_gvn);
5696 const TypeAryPtr* top_src = src_type->isa_aryptr();
5697 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5698 if (top_src == NULL || top_src->klass() == NULL ||
5699 top_dest == NULL || top_dest->klass() == NULL) {
5700 // failed array check
5701 return false;
5702 }
5704 // Figure out the size and type of the elements we will be copying.
5705 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5706 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5707 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5708 return false;
5709 }
5710 Node* src_start = array_element_address(src, src_offset, src_elem);
5711 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5712 // 'src_start' points to src array + scaled offset
5713 // 'dst_start' points to dst array + scaled offset
5715 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5716 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5717 enc = _gvn.transform(enc);
5718 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5719 set_memory(res_mem, mtype);
5720 set_result(enc);
5721 return true;
5722 }
5724 /**
5725 * Calculate CRC32 for byte.
5726 * int java.util.zip.CRC32.update(int crc, int b)
5727 */
5728 bool LibraryCallKit::inline_updateCRC32() {
5729 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5730 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5731 // no receiver since it is static method
5732 Node* crc = argument(0); // type: int
5733 Node* b = argument(1); // type: int
5735 /*
5736 * int c = ~ crc;
5737 * b = timesXtoThe32[(b ^ c) & 0xFF];
5738 * b = b ^ (c >>> 8);
5739 * crc = ~b;
5740 */
5742 Node* M1 = intcon(-1);
5743 crc = _gvn.transform(new (C) XorINode(crc, M1));
5744 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5745 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5747 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5748 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5749 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5750 result = make_load(control(), adr, TypeInt::INT, T_INT, MemNode::unordered);
5752 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5753 result = _gvn.transform(new (C) XorINode(crc, result));
5754 result = _gvn.transform(new (C) XorINode(result, M1));
5755 set_result(result);
5756 return true;
5757 }
5759 /**
5760 * Calculate CRC32 for byte[] array.
5761 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5762 */
5763 bool LibraryCallKit::inline_updateBytesCRC32() {
5764 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5765 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5766 // no receiver since it is static method
5767 Node* crc = argument(0); // type: int
5768 Node* src = argument(1); // type: oop
5769 Node* offset = argument(2); // type: int
5770 Node* length = argument(3); // type: int
5772 const Type* src_type = src->Value(&_gvn);
5773 const TypeAryPtr* top_src = src_type->isa_aryptr();
5774 if (top_src == NULL || top_src->klass() == NULL) {
5775 // failed array check
5776 return false;
5777 }
5779 // Figure out the size and type of the elements we will be copying.
5780 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5781 if (src_elem != T_BYTE) {
5782 return false;
5783 }
5785 // 'src_start' points to src array + scaled offset
5786 Node* src_start = array_element_address(src, offset, src_elem);
5788 // We assume that range check is done by caller.
5789 // TODO: generate range check (offset+length < src.length) in debug VM.
5791 // Call the stub.
5792 address stubAddr = StubRoutines::updateBytesCRC32();
5793 const char *stubName = "updateBytesCRC32";
5795 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5796 stubAddr, stubName, TypePtr::BOTTOM,
5797 crc, src_start, length);
5798 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5799 set_result(result);
5800 return true;
5801 }
5803 /**
5804 * Calculate CRC32 for ByteBuffer.
5805 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5806 */
5807 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5808 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5809 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5810 // no receiver since it is static method
5811 Node* crc = argument(0); // type: int
5812 Node* src = argument(1); // type: long
5813 Node* offset = argument(3); // type: int
5814 Node* length = argument(4); // type: int
5816 src = ConvL2X(src); // adjust Java long to machine word
5817 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5818 offset = ConvI2X(offset);
5820 // 'src_start' points to src array + scaled offset
5821 Node* src_start = basic_plus_adr(top(), base, offset);
5823 // Call the stub.
5824 address stubAddr = StubRoutines::updateBytesCRC32();
5825 const char *stubName = "updateBytesCRC32";
5827 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5828 stubAddr, stubName, TypePtr::BOTTOM,
5829 crc, src_start, length);
5830 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5831 set_result(result);
5832 return true;
5833 }
5835 //----------------------------inline_reference_get----------------------------
5836 // public T java.lang.ref.Reference.get();
5837 bool LibraryCallKit::inline_reference_get() {
5838 const int referent_offset = java_lang_ref_Reference::referent_offset;
5839 guarantee(referent_offset > 0, "should have already been set");
5841 // Get the argument:
5842 Node* reference_obj = null_check_receiver();
5843 if (stopped()) return true;
5845 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5847 ciInstanceKlass* klass = env()->Object_klass();
5848 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5850 Node* no_ctrl = NULL;
5851 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT, MemNode::unordered);
5853 // Use the pre-barrier to record the value in the referent field
5854 pre_barrier(false /* do_load */,
5855 control(),
5856 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5857 result /* pre_val */,
5858 T_OBJECT);
5860 // Add memory barrier to prevent commoning reads from this field
5861 // across safepoint since GC can change its value.
5862 insert_mem_bar(Op_MemBarCPUOrder);
5864 set_result(result);
5865 return true;
5866 }
5869 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5870 bool is_exact=true, bool is_static=false) {
5872 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5873 assert(tinst != NULL, "obj is null");
5874 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5875 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5877 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5878 ciSymbol::make(fieldTypeString),
5879 is_static);
5880 if (field == NULL) return (Node *) NULL;
5881 assert (field != NULL, "undefined field");
5883 // Next code copied from Parse::do_get_xxx():
5885 // Compute address and memory type.
5886 int offset = field->offset_in_bytes();
5887 bool is_vol = field->is_volatile();
5888 ciType* field_klass = field->type();
5889 assert(field_klass->is_loaded(), "should be loaded");
5890 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5891 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5892 BasicType bt = field->layout_type();
5894 // Build the resultant type of the load
5895 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5897 // Build the load.
5898 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, MemNode::unordered, is_vol);
5899 return loadedField;
5900 }
5903 //------------------------------inline_aescrypt_Block-----------------------
5904 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5905 address stubAddr;
5906 const char *stubName;
5907 assert(UseAES, "need AES instruction support");
5909 switch(id) {
5910 case vmIntrinsics::_aescrypt_encryptBlock:
5911 stubAddr = StubRoutines::aescrypt_encryptBlock();
5912 stubName = "aescrypt_encryptBlock";
5913 break;
5914 case vmIntrinsics::_aescrypt_decryptBlock:
5915 stubAddr = StubRoutines::aescrypt_decryptBlock();
5916 stubName = "aescrypt_decryptBlock";
5917 break;
5918 }
5919 if (stubAddr == NULL) return false;
5921 Node* aescrypt_object = argument(0);
5922 Node* src = argument(1);
5923 Node* src_offset = argument(2);
5924 Node* dest = argument(3);
5925 Node* dest_offset = argument(4);
5927 // (1) src and dest are arrays.
5928 const Type* src_type = src->Value(&_gvn);
5929 const Type* dest_type = dest->Value(&_gvn);
5930 const TypeAryPtr* top_src = src_type->isa_aryptr();
5931 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5932 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5934 // for the quick and dirty code we will skip all the checks.
5935 // we are just trying to get the call to be generated.
5936 Node* src_start = src;
5937 Node* dest_start = dest;
5938 if (src_offset != NULL || dest_offset != NULL) {
5939 assert(src_offset != NULL && dest_offset != NULL, "");
5940 src_start = array_element_address(src, src_offset, T_BYTE);
5941 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5942 }
5944 // now need to get the start of its expanded key array
5945 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5946 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5947 if (k_start == NULL) return false;
5949 if (Matcher::pass_original_key_for_aes()) {
5950 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5951 // compatibility issues between Java key expansion and SPARC crypto instructions
5952 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5953 if (original_k_start == NULL) return false;
5955 // Call the stub.
5956 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5957 stubAddr, stubName, TypePtr::BOTTOM,
5958 src_start, dest_start, k_start, original_k_start);
5959 } else {
5960 // Call the stub.
5961 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5962 stubAddr, stubName, TypePtr::BOTTOM,
5963 src_start, dest_start, k_start);
5964 }
5966 return true;
5967 }
5969 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5970 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5971 address stubAddr;
5972 const char *stubName;
5974 assert(UseAES, "need AES instruction support");
5976 switch(id) {
5977 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5978 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5979 stubName = "cipherBlockChaining_encryptAESCrypt";
5980 break;
5981 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5982 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5983 stubName = "cipherBlockChaining_decryptAESCrypt";
5984 break;
5985 }
5986 if (stubAddr == NULL) return false;
5988 Node* cipherBlockChaining_object = argument(0);
5989 Node* src = argument(1);
5990 Node* src_offset = argument(2);
5991 Node* len = argument(3);
5992 Node* dest = argument(4);
5993 Node* dest_offset = argument(5);
5995 // (1) src and dest are arrays.
5996 const Type* src_type = src->Value(&_gvn);
5997 const Type* dest_type = dest->Value(&_gvn);
5998 const TypeAryPtr* top_src = src_type->isa_aryptr();
5999 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
6000 assert (top_src != NULL && top_src->klass() != NULL
6001 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
6003 // checks are the responsibility of the caller
6004 Node* src_start = src;
6005 Node* dest_start = dest;
6006 if (src_offset != NULL || dest_offset != NULL) {
6007 assert(src_offset != NULL && dest_offset != NULL, "");
6008 src_start = array_element_address(src, src_offset, T_BYTE);
6009 dest_start = array_element_address(dest, dest_offset, T_BYTE);
6010 }
6012 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
6013 // (because of the predicated logic executed earlier).
6014 // so we cast it here safely.
6015 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
6017 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6018 if (embeddedCipherObj == NULL) return false;
6020 // cast it to what we know it will be at runtime
6021 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
6022 assert(tinst != NULL, "CBC obj is null");
6023 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
6024 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6025 if (!klass_AESCrypt->is_loaded()) return false;
6027 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6028 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
6029 const TypeOopPtr* xtype = aklass->as_instance_type();
6030 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
6031 aescrypt_object = _gvn.transform(aescrypt_object);
6033 // we need to get the start of the aescrypt_object's expanded key array
6034 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
6035 if (k_start == NULL) return false;
6037 // similarly, get the start address of the r vector
6038 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
6039 if (objRvec == NULL) return false;
6040 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
6042 Node* cbcCrypt;
6043 if (Matcher::pass_original_key_for_aes()) {
6044 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
6045 // compatibility issues between Java key expansion and SPARC crypto instructions
6046 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
6047 if (original_k_start == NULL) return false;
6049 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
6050 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6051 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6052 stubAddr, stubName, TypePtr::BOTTOM,
6053 src_start, dest_start, k_start, r_start, len, original_k_start);
6054 } else {
6055 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
6056 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
6057 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
6058 stubAddr, stubName, TypePtr::BOTTOM,
6059 src_start, dest_start, k_start, r_start, len);
6060 }
6062 // return cipher length (int)
6063 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
6064 set_result(retvalue);
6065 return true;
6066 }
6068 //------------------------------get_key_start_from_aescrypt_object-----------------------
6069 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6070 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6071 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6072 if (objAESCryptKey == NULL) return (Node *) NULL;
6074 // now have the array, need to get the start address of the K array
6075 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6076 return k_start;
6077 }
6079 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6080 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6081 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6082 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6083 if (objAESCryptKey == NULL) return (Node *) NULL;
6085 // now have the array, need to get the start address of the lastKey array
6086 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6087 return original_k_start;
6088 }
6090 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6091 // Return node representing slow path of predicate check.
6092 // the pseudo code we want to emulate with this predicate is:
6093 // for encryption:
6094 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6095 // for decryption:
6096 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6097 // note cipher==plain is more conservative than the original java code but that's OK
6098 //
6099 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6100 // First, check receiver for NULL since it is virtual method.
6101 Node* objCBC = argument(0);
6102 objCBC = null_check(objCBC);
6104 if (stopped()) return NULL; // Always NULL
6106 // Load embeddedCipher field of CipherBlockChaining object.
6107 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6109 // get AESCrypt klass for instanceOf check
6110 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6111 // will have same classloader as CipherBlockChaining object
6112 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6113 assert(tinst != NULL, "CBCobj is null");
6114 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6116 // we want to do an instanceof comparison against the AESCrypt class
6117 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6118 if (!klass_AESCrypt->is_loaded()) {
6119 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6120 Node* ctrl = control();
6121 set_control(top()); // no regular fast path
6122 return ctrl;
6123 }
6124 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6126 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6127 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6128 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6130 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6132 // for encryption, we are done
6133 if (!decrypting)
6134 return instof_false; // even if it is NULL
6136 // for decryption, we need to add a further check to avoid
6137 // taking the intrinsic path when cipher and plain are the same
6138 // see the original java code for why.
6139 RegionNode* region = new(C) RegionNode(3);
6140 region->init_req(1, instof_false);
6141 Node* src = argument(1);
6142 Node* dest = argument(4);
6143 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6144 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6145 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6146 region->init_req(2, src_dest_conjoint);
6148 record_for_igvn(region);
6149 return _gvn.transform(region);
6150 }