Thu, 23 Jan 2014 12:08:28 +0100
8027754: Enable loop optimizations for loops with MathExact inside
Reviewed-by: kvn, iveresov
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 template <typename OverflowOp>
207 bool inline_math_overflow(Node* arg1, Node* arg2);
208 void inline_math_mathExact(Node* math, Node* test);
209 bool inline_math_addExactI(bool is_increment);
210 bool inline_math_addExactL(bool is_increment);
211 bool inline_math_multiplyExactI();
212 bool inline_math_multiplyExactL();
213 bool inline_math_negateExactI();
214 bool inline_math_negateExactL();
215 bool inline_math_subtractExactI(bool is_decrement);
216 bool inline_math_subtractExactL(bool is_decrement);
217 bool inline_exp();
218 bool inline_pow();
219 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
220 bool inline_min_max(vmIntrinsics::ID id);
221 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
222 // This returns Type::AnyPtr, RawPtr, or OopPtr.
223 int classify_unsafe_addr(Node* &base, Node* &offset);
224 Node* make_unsafe_address(Node* base, Node* offset);
225 // Helper for inline_unsafe_access.
226 // Generates the guards that check whether the result of
227 // Unsafe.getObject should be recorded in an SATB log buffer.
228 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
229 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
230 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
231 static bool klass_needs_init_guard(Node* kls);
232 bool inline_unsafe_allocate();
233 bool inline_unsafe_copyMemory();
234 bool inline_native_currentThread();
235 #ifdef TRACE_HAVE_INTRINSICS
236 bool inline_native_classID();
237 bool inline_native_threadID();
238 #endif
239 bool inline_native_time_funcs(address method, const char* funcName);
240 bool inline_native_isInterrupted();
241 bool inline_native_Class_query(vmIntrinsics::ID id);
242 bool inline_native_subtype_check();
244 bool inline_native_newArray();
245 bool inline_native_getLength();
246 bool inline_array_copyOf(bool is_copyOfRange);
247 bool inline_array_equals();
248 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
249 bool inline_native_clone(bool is_virtual);
250 bool inline_native_Reflection_getCallerClass();
251 // Helper function for inlining native object hash method
252 bool inline_native_hashcode(bool is_virtual, bool is_static);
253 bool inline_native_getClass();
255 // Helper functions for inlining arraycopy
256 bool inline_arraycopy();
257 void generate_arraycopy(const TypePtr* adr_type,
258 BasicType basic_elem_type,
259 Node* src, Node* src_offset,
260 Node* dest, Node* dest_offset,
261 Node* copy_length,
262 bool disjoint_bases = false,
263 bool length_never_negative = false,
264 RegionNode* slow_region = NULL);
265 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
266 RegionNode* slow_region);
267 void generate_clear_array(const TypePtr* adr_type,
268 Node* dest,
269 BasicType basic_elem_type,
270 Node* slice_off,
271 Node* slice_len,
272 Node* slice_end);
273 bool generate_block_arraycopy(const TypePtr* adr_type,
274 BasicType basic_elem_type,
275 AllocateNode* alloc,
276 Node* src, Node* src_offset,
277 Node* dest, Node* dest_offset,
278 Node* dest_size, bool dest_uninitialized);
279 void generate_slow_arraycopy(const TypePtr* adr_type,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* copy_length, bool dest_uninitialized);
283 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
284 Node* dest_elem_klass,
285 Node* src, Node* src_offset,
286 Node* dest, Node* dest_offset,
287 Node* copy_length, bool dest_uninitialized);
288 Node* generate_generic_arraycopy(const TypePtr* adr_type,
289 Node* src, Node* src_offset,
290 Node* dest, Node* dest_offset,
291 Node* copy_length, bool dest_uninitialized);
292 void generate_unchecked_arraycopy(const TypePtr* adr_type,
293 BasicType basic_elem_type,
294 bool disjoint_bases,
295 Node* src, Node* src_offset,
296 Node* dest, Node* dest_offset,
297 Node* copy_length, bool dest_uninitialized);
298 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
299 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
300 bool inline_unsafe_ordered_store(BasicType type);
301 bool inline_unsafe_fence(vmIntrinsics::ID id);
302 bool inline_fp_conversions(vmIntrinsics::ID id);
303 bool inline_number_methods(vmIntrinsics::ID id);
304 bool inline_reference_get();
305 bool inline_aescrypt_Block(vmIntrinsics::ID id);
306 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
307 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
308 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
309 Node* get_original_key_start_from_aescrypt_object(Node* aescrypt_object);
310 bool inline_encodeISOArray();
311 bool inline_updateCRC32();
312 bool inline_updateBytesCRC32();
313 bool inline_updateByteBufferCRC32();
314 };
317 //---------------------------make_vm_intrinsic----------------------------
318 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
319 vmIntrinsics::ID id = m->intrinsic_id();
320 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
322 if (DisableIntrinsic[0] != '\0'
323 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
324 // disabled by a user request on the command line:
325 // example: -XX:DisableIntrinsic=_hashCode,_getClass
326 return NULL;
327 }
329 if (!m->is_loaded()) {
330 // do not attempt to inline unloaded methods
331 return NULL;
332 }
334 // Only a few intrinsics implement a virtual dispatch.
335 // They are expensive calls which are also frequently overridden.
336 if (is_virtual) {
337 switch (id) {
338 case vmIntrinsics::_hashCode:
339 case vmIntrinsics::_clone:
340 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
341 break;
342 default:
343 return NULL;
344 }
345 }
347 // -XX:-InlineNatives disables nearly all intrinsics:
348 if (!InlineNatives) {
349 switch (id) {
350 case vmIntrinsics::_indexOf:
351 case vmIntrinsics::_compareTo:
352 case vmIntrinsics::_equals:
353 case vmIntrinsics::_equalsC:
354 case vmIntrinsics::_getAndAddInt:
355 case vmIntrinsics::_getAndAddLong:
356 case vmIntrinsics::_getAndSetInt:
357 case vmIntrinsics::_getAndSetLong:
358 case vmIntrinsics::_getAndSetObject:
359 case vmIntrinsics::_loadFence:
360 case vmIntrinsics::_storeFence:
361 case vmIntrinsics::_fullFence:
362 break; // InlineNatives does not control String.compareTo
363 case vmIntrinsics::_Reference_get:
364 break; // InlineNatives does not control Reference.get
365 default:
366 return NULL;
367 }
368 }
370 bool is_predicted = false;
371 bool does_virtual_dispatch = false;
373 switch (id) {
374 case vmIntrinsics::_compareTo:
375 if (!SpecialStringCompareTo) return NULL;
376 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
377 break;
378 case vmIntrinsics::_indexOf:
379 if (!SpecialStringIndexOf) return NULL;
380 break;
381 case vmIntrinsics::_equals:
382 if (!SpecialStringEquals) return NULL;
383 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
384 break;
385 case vmIntrinsics::_equalsC:
386 if (!SpecialArraysEquals) return NULL;
387 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
388 break;
389 case vmIntrinsics::_arraycopy:
390 if (!InlineArrayCopy) return NULL;
391 break;
392 case vmIntrinsics::_copyMemory:
393 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
394 if (!InlineArrayCopy) return NULL;
395 break;
396 case vmIntrinsics::_hashCode:
397 if (!InlineObjectHash) return NULL;
398 does_virtual_dispatch = true;
399 break;
400 case vmIntrinsics::_clone:
401 does_virtual_dispatch = true;
402 case vmIntrinsics::_copyOf:
403 case vmIntrinsics::_copyOfRange:
404 if (!InlineObjectCopy) return NULL;
405 // These also use the arraycopy intrinsic mechanism:
406 if (!InlineArrayCopy) return NULL;
407 break;
408 case vmIntrinsics::_encodeISOArray:
409 if (!SpecialEncodeISOArray) return NULL;
410 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
411 break;
412 case vmIntrinsics::_checkIndex:
413 // We do not intrinsify this. The optimizer does fine with it.
414 return NULL;
416 case vmIntrinsics::_getCallerClass:
417 if (!UseNewReflection) return NULL;
418 if (!InlineReflectionGetCallerClass) return NULL;
419 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
420 break;
422 case vmIntrinsics::_bitCount_i:
423 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
424 break;
426 case vmIntrinsics::_bitCount_l:
427 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
428 break;
430 case vmIntrinsics::_numberOfLeadingZeros_i:
431 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
432 break;
434 case vmIntrinsics::_numberOfLeadingZeros_l:
435 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
436 break;
438 case vmIntrinsics::_numberOfTrailingZeros_i:
439 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
440 break;
442 case vmIntrinsics::_numberOfTrailingZeros_l:
443 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
444 break;
446 case vmIntrinsics::_reverseBytes_c:
447 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
448 break;
449 case vmIntrinsics::_reverseBytes_s:
450 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
451 break;
452 case vmIntrinsics::_reverseBytes_i:
453 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
454 break;
455 case vmIntrinsics::_reverseBytes_l:
456 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
457 break;
459 case vmIntrinsics::_Reference_get:
460 // Use the intrinsic version of Reference.get() so that the value in
461 // the referent field can be registered by the G1 pre-barrier code.
462 // Also add memory barrier to prevent commoning reads from this field
463 // across safepoint since GC can change it value.
464 break;
466 case vmIntrinsics::_compareAndSwapObject:
467 #ifdef _LP64
468 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
469 #endif
470 break;
472 case vmIntrinsics::_compareAndSwapLong:
473 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
474 break;
476 case vmIntrinsics::_getAndAddInt:
477 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
478 break;
480 case vmIntrinsics::_getAndAddLong:
481 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
482 break;
484 case vmIntrinsics::_getAndSetInt:
485 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
486 break;
488 case vmIntrinsics::_getAndSetLong:
489 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
490 break;
492 case vmIntrinsics::_getAndSetObject:
493 #ifdef _LP64
494 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
495 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
496 break;
497 #else
498 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
499 break;
500 #endif
502 case vmIntrinsics::_aescrypt_encryptBlock:
503 case vmIntrinsics::_aescrypt_decryptBlock:
504 if (!UseAESIntrinsics) return NULL;
505 break;
507 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
508 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
509 if (!UseAESIntrinsics) return NULL;
510 // these two require the predicated logic
511 is_predicted = true;
512 break;
514 case vmIntrinsics::_updateCRC32:
515 case vmIntrinsics::_updateBytesCRC32:
516 case vmIntrinsics::_updateByteBufferCRC32:
517 if (!UseCRC32Intrinsics) return NULL;
518 break;
520 case vmIntrinsics::_incrementExactI:
521 case vmIntrinsics::_addExactI:
522 if (!Matcher::match_rule_supported(Op_OverflowAddI) || !UseMathExactIntrinsics) return NULL;
523 break;
524 case vmIntrinsics::_incrementExactL:
525 case vmIntrinsics::_addExactL:
526 if (!Matcher::match_rule_supported(Op_OverflowAddL) || !UseMathExactIntrinsics) return NULL;
527 break;
528 case vmIntrinsics::_decrementExactI:
529 case vmIntrinsics::_subtractExactI:
530 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
531 break;
532 case vmIntrinsics::_decrementExactL:
533 case vmIntrinsics::_subtractExactL:
534 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
535 break;
536 case vmIntrinsics::_negateExactI:
537 if (!Matcher::match_rule_supported(Op_OverflowSubI) || !UseMathExactIntrinsics) return NULL;
538 break;
539 case vmIntrinsics::_negateExactL:
540 if (!Matcher::match_rule_supported(Op_OverflowSubL) || !UseMathExactIntrinsics) return NULL;
541 break;
542 case vmIntrinsics::_multiplyExactI:
543 if (!Matcher::match_rule_supported(Op_OverflowMulI) || !UseMathExactIntrinsics) return NULL;
544 break;
545 case vmIntrinsics::_multiplyExactL:
546 if (!Matcher::match_rule_supported(Op_OverflowMulL) || !UseMathExactIntrinsics) return NULL;
547 break;
549 default:
550 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
551 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
552 break;
553 }
555 // -XX:-InlineClassNatives disables natives from the Class class.
556 // The flag applies to all reflective calls, notably Array.newArray
557 // (visible to Java programmers as Array.newInstance).
558 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
559 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
560 if (!InlineClassNatives) return NULL;
561 }
563 // -XX:-InlineThreadNatives disables natives from the Thread class.
564 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
565 if (!InlineThreadNatives) return NULL;
566 }
568 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
569 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
570 m->holder()->name() == ciSymbol::java_lang_Float() ||
571 m->holder()->name() == ciSymbol::java_lang_Double()) {
572 if (!InlineMathNatives) return NULL;
573 }
575 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
576 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
577 if (!InlineUnsafeOps) return NULL;
578 }
580 return new LibraryIntrinsic(m, is_virtual, is_predicted, does_virtual_dispatch, (vmIntrinsics::ID) id);
581 }
583 //----------------------register_library_intrinsics-----------------------
584 // Initialize this file's data structures, for each Compile instance.
585 void Compile::register_library_intrinsics() {
586 // Nothing to do here.
587 }
589 JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
590 LibraryCallKit kit(jvms, this);
591 Compile* C = kit.C;
592 int nodes = C->unique();
593 #ifndef PRODUCT
594 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
595 char buf[1000];
596 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
597 tty->print_cr("Intrinsic %s", str);
598 }
599 #endif
600 ciMethod* callee = kit.callee();
601 const int bci = kit.bci();
603 // Try to inline the intrinsic.
604 if (kit.try_to_inline()) {
605 if (C->print_intrinsics() || C->print_inlining()) {
606 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
607 }
608 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
609 if (C->log()) {
610 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
611 vmIntrinsics::name_at(intrinsic_id()),
612 (is_virtual() ? " virtual='1'" : ""),
613 C->unique() - nodes);
614 }
615 // Push the result from the inlined method onto the stack.
616 kit.push_result();
617 return kit.transfer_exceptions_into_jvms();
618 }
620 // The intrinsic bailed out
621 if (C->print_intrinsics() || C->print_inlining()) {
622 if (jvms->has_method()) {
623 // Not a root compile.
624 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
625 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
626 } else {
627 // Root compile
628 tty->print("Did not generate intrinsic %s%s at bci:%d in",
629 vmIntrinsics::name_at(intrinsic_id()),
630 (is_virtual() ? " (virtual)" : ""), bci);
631 }
632 }
633 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
634 return NULL;
635 }
637 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
638 LibraryCallKit kit(jvms, this);
639 Compile* C = kit.C;
640 int nodes = C->unique();
641 #ifndef PRODUCT
642 assert(is_predicted(), "sanity");
643 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
644 char buf[1000];
645 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
646 tty->print_cr("Predicate for intrinsic %s", str);
647 }
648 #endif
649 ciMethod* callee = kit.callee();
650 const int bci = kit.bci();
652 Node* slow_ctl = kit.try_to_predicate();
653 if (!kit.failing()) {
654 if (C->print_intrinsics() || C->print_inlining()) {
655 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
656 }
657 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
658 if (C->log()) {
659 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
660 vmIntrinsics::name_at(intrinsic_id()),
661 (is_virtual() ? " virtual='1'" : ""),
662 C->unique() - nodes);
663 }
664 return slow_ctl; // Could be NULL if the check folds.
665 }
667 // The intrinsic bailed out
668 if (C->print_intrinsics() || C->print_inlining()) {
669 if (jvms->has_method()) {
670 // Not a root compile.
671 const char* msg = "failed to generate predicate for intrinsic";
672 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
673 } else {
674 // Root compile
675 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
676 vmIntrinsics::name_at(intrinsic_id()),
677 (is_virtual() ? " (virtual)" : ""), bci);
678 }
679 }
680 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
681 return NULL;
682 }
684 bool LibraryCallKit::try_to_inline() {
685 // Handle symbolic names for otherwise undistinguished boolean switches:
686 const bool is_store = true;
687 const bool is_native_ptr = true;
688 const bool is_static = true;
689 const bool is_volatile = true;
691 if (!jvms()->has_method()) {
692 // Root JVMState has a null method.
693 assert(map()->memory()->Opcode() == Op_Parm, "");
694 // Insert the memory aliasing node
695 set_all_memory(reset_memory());
696 }
697 assert(merged_memory(), "");
700 switch (intrinsic_id()) {
701 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
702 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
703 case vmIntrinsics::_getClass: return inline_native_getClass();
705 case vmIntrinsics::_dsin:
706 case vmIntrinsics::_dcos:
707 case vmIntrinsics::_dtan:
708 case vmIntrinsics::_dabs:
709 case vmIntrinsics::_datan2:
710 case vmIntrinsics::_dsqrt:
711 case vmIntrinsics::_dexp:
712 case vmIntrinsics::_dlog:
713 case vmIntrinsics::_dlog10:
714 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
716 case vmIntrinsics::_min:
717 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
719 case vmIntrinsics::_addExactI: return inline_math_addExactI(false /* add */);
720 case vmIntrinsics::_addExactL: return inline_math_addExactL(false /* add */);
721 case vmIntrinsics::_decrementExactI: return inline_math_subtractExactI(true /* decrement */);
722 case vmIntrinsics::_decrementExactL: return inline_math_subtractExactL(true /* decrement */);
723 case vmIntrinsics::_incrementExactI: return inline_math_addExactI(true /* increment */);
724 case vmIntrinsics::_incrementExactL: return inline_math_addExactL(true /* increment */);
725 case vmIntrinsics::_multiplyExactI: return inline_math_multiplyExactI();
726 case vmIntrinsics::_multiplyExactL: return inline_math_multiplyExactL();
727 case vmIntrinsics::_negateExactI: return inline_math_negateExactI();
728 case vmIntrinsics::_negateExactL: return inline_math_negateExactL();
729 case vmIntrinsics::_subtractExactI: return inline_math_subtractExactI(false /* subtract */);
730 case vmIntrinsics::_subtractExactL: return inline_math_subtractExactL(false /* subtract */);
732 case vmIntrinsics::_arraycopy: return inline_arraycopy();
734 case vmIntrinsics::_compareTo: return inline_string_compareTo();
735 case vmIntrinsics::_indexOf: return inline_string_indexOf();
736 case vmIntrinsics::_equals: return inline_string_equals();
738 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
739 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
740 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
741 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
742 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
743 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
744 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
745 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
746 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
748 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
749 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
750 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
751 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
752 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
753 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
754 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
755 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
756 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
758 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
759 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
760 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
761 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
762 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
763 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
764 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
765 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
767 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
768 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
769 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
770 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
771 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
772 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
773 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
774 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
776 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
777 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
778 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
779 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
780 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
781 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
782 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
783 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
784 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
786 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
787 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
788 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
789 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
790 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
791 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
792 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
793 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
794 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
796 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
797 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
798 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
799 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
801 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
802 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
803 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
805 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
806 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
807 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
809 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
810 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
811 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
812 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
813 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
815 case vmIntrinsics::_loadFence:
816 case vmIntrinsics::_storeFence:
817 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
819 case vmIntrinsics::_currentThread: return inline_native_currentThread();
820 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
822 #ifdef TRACE_HAVE_INTRINSICS
823 case vmIntrinsics::_classID: return inline_native_classID();
824 case vmIntrinsics::_threadID: return inline_native_threadID();
825 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
826 #endif
827 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
828 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
829 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
830 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
831 case vmIntrinsics::_newArray: return inline_native_newArray();
832 case vmIntrinsics::_getLength: return inline_native_getLength();
833 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
834 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
835 case vmIntrinsics::_equalsC: return inline_array_equals();
836 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
838 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
840 case vmIntrinsics::_isInstance:
841 case vmIntrinsics::_getModifiers:
842 case vmIntrinsics::_isInterface:
843 case vmIntrinsics::_isArray:
844 case vmIntrinsics::_isPrimitive:
845 case vmIntrinsics::_getSuperclass:
846 case vmIntrinsics::_getComponentType:
847 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
849 case vmIntrinsics::_floatToRawIntBits:
850 case vmIntrinsics::_floatToIntBits:
851 case vmIntrinsics::_intBitsToFloat:
852 case vmIntrinsics::_doubleToRawLongBits:
853 case vmIntrinsics::_doubleToLongBits:
854 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
856 case vmIntrinsics::_numberOfLeadingZeros_i:
857 case vmIntrinsics::_numberOfLeadingZeros_l:
858 case vmIntrinsics::_numberOfTrailingZeros_i:
859 case vmIntrinsics::_numberOfTrailingZeros_l:
860 case vmIntrinsics::_bitCount_i:
861 case vmIntrinsics::_bitCount_l:
862 case vmIntrinsics::_reverseBytes_i:
863 case vmIntrinsics::_reverseBytes_l:
864 case vmIntrinsics::_reverseBytes_s:
865 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
867 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
869 case vmIntrinsics::_Reference_get: return inline_reference_get();
871 case vmIntrinsics::_aescrypt_encryptBlock:
872 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
874 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
875 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
876 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
878 case vmIntrinsics::_encodeISOArray:
879 return inline_encodeISOArray();
881 case vmIntrinsics::_updateCRC32:
882 return inline_updateCRC32();
883 case vmIntrinsics::_updateBytesCRC32:
884 return inline_updateBytesCRC32();
885 case vmIntrinsics::_updateByteBufferCRC32:
886 return inline_updateByteBufferCRC32();
888 default:
889 // If you get here, it may be that someone has added a new intrinsic
890 // to the list in vmSymbols.hpp without implementing it here.
891 #ifndef PRODUCT
892 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
893 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
894 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
895 }
896 #endif
897 return false;
898 }
899 }
901 Node* LibraryCallKit::try_to_predicate() {
902 if (!jvms()->has_method()) {
903 // Root JVMState has a null method.
904 assert(map()->memory()->Opcode() == Op_Parm, "");
905 // Insert the memory aliasing node
906 set_all_memory(reset_memory());
907 }
908 assert(merged_memory(), "");
910 switch (intrinsic_id()) {
911 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
912 return inline_cipherBlockChaining_AESCrypt_predicate(false);
913 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
914 return inline_cipherBlockChaining_AESCrypt_predicate(true);
916 default:
917 // If you get here, it may be that someone has added a new intrinsic
918 // to the list in vmSymbols.hpp without implementing it here.
919 #ifndef PRODUCT
920 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
921 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
922 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
923 }
924 #endif
925 Node* slow_ctl = control();
926 set_control(top()); // No fast path instrinsic
927 return slow_ctl;
928 }
929 }
931 //------------------------------set_result-------------------------------
932 // Helper function for finishing intrinsics.
933 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
934 record_for_igvn(region);
935 set_control(_gvn.transform(region));
936 set_result( _gvn.transform(value));
937 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
938 }
940 //------------------------------generate_guard---------------------------
941 // Helper function for generating guarded fast-slow graph structures.
942 // The given 'test', if true, guards a slow path. If the test fails
943 // then a fast path can be taken. (We generally hope it fails.)
944 // In all cases, GraphKit::control() is updated to the fast path.
945 // The returned value represents the control for the slow path.
946 // The return value is never 'top'; it is either a valid control
947 // or NULL if it is obvious that the slow path can never be taken.
948 // Also, if region and the slow control are not NULL, the slow edge
949 // is appended to the region.
950 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
951 if (stopped()) {
952 // Already short circuited.
953 return NULL;
954 }
956 // Build an if node and its projections.
957 // If test is true we take the slow path, which we assume is uncommon.
958 if (_gvn.type(test) == TypeInt::ZERO) {
959 // The slow branch is never taken. No need to build this guard.
960 return NULL;
961 }
963 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
965 Node* if_slow = _gvn.transform(new (C) IfTrueNode(iff));
966 if (if_slow == top()) {
967 // The slow branch is never taken. No need to build this guard.
968 return NULL;
969 }
971 if (region != NULL)
972 region->add_req(if_slow);
974 Node* if_fast = _gvn.transform(new (C) IfFalseNode(iff));
975 set_control(if_fast);
977 return if_slow;
978 }
980 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
981 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
982 }
983 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
984 return generate_guard(test, region, PROB_FAIR);
985 }
987 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
988 Node* *pos_index) {
989 if (stopped())
990 return NULL; // already stopped
991 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
992 return NULL; // index is already adequately typed
993 Node* cmp_lt = _gvn.transform(new (C) CmpINode(index, intcon(0)));
994 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
995 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
996 if (is_neg != NULL && pos_index != NULL) {
997 // Emulate effect of Parse::adjust_map_after_if.
998 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
999 ccast->set_req(0, control());
1000 (*pos_index) = _gvn.transform(ccast);
1001 }
1002 return is_neg;
1003 }
1005 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
1006 Node* *pos_index) {
1007 if (stopped())
1008 return NULL; // already stopped
1009 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
1010 return NULL; // index is already adequately typed
1011 Node* cmp_le = _gvn.transform(new (C) CmpINode(index, intcon(0)));
1012 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
1013 Node* bol_le = _gvn.transform(new (C) BoolNode(cmp_le, le_or_eq));
1014 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
1015 if (is_notp != NULL && pos_index != NULL) {
1016 // Emulate effect of Parse::adjust_map_after_if.
1017 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
1018 ccast->set_req(0, control());
1019 (*pos_index) = _gvn.transform(ccast);
1020 }
1021 return is_notp;
1022 }
1024 // Make sure that 'position' is a valid limit index, in [0..length].
1025 // There are two equivalent plans for checking this:
1026 // A. (offset + copyLength) unsigned<= arrayLength
1027 // B. offset <= (arrayLength - copyLength)
1028 // We require that all of the values above, except for the sum and
1029 // difference, are already known to be non-negative.
1030 // Plan A is robust in the face of overflow, if offset and copyLength
1031 // are both hugely positive.
1032 //
1033 // Plan B is less direct and intuitive, but it does not overflow at
1034 // all, since the difference of two non-negatives is always
1035 // representable. Whenever Java methods must perform the equivalent
1036 // check they generally use Plan B instead of Plan A.
1037 // For the moment we use Plan A.
1038 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
1039 Node* subseq_length,
1040 Node* array_length,
1041 RegionNode* region) {
1042 if (stopped())
1043 return NULL; // already stopped
1044 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1045 if (zero_offset && subseq_length->eqv_uncast(array_length))
1046 return NULL; // common case of whole-array copy
1047 Node* last = subseq_length;
1048 if (!zero_offset) // last += offset
1049 last = _gvn.transform(new (C) AddINode(last, offset));
1050 Node* cmp_lt = _gvn.transform(new (C) CmpUNode(array_length, last));
1051 Node* bol_lt = _gvn.transform(new (C) BoolNode(cmp_lt, BoolTest::lt));
1052 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1053 return is_over;
1054 }
1057 //--------------------------generate_current_thread--------------------
1058 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1059 ciKlass* thread_klass = env()->Thread_klass();
1060 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1061 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1062 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1063 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1064 tls_output = thread;
1065 return threadObj;
1066 }
1069 //------------------------------make_string_method_node------------------------
1070 // Helper method for String intrinsic functions. This version is called
1071 // with str1 and str2 pointing to String object nodes.
1072 //
1073 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1074 Node* no_ctrl = NULL;
1076 // Get start addr of string
1077 Node* str1_value = load_String_value(no_ctrl, str1);
1078 Node* str1_offset = load_String_offset(no_ctrl, str1);
1079 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1081 // Get length of string 1
1082 Node* str1_len = load_String_length(no_ctrl, str1);
1084 Node* str2_value = load_String_value(no_ctrl, str2);
1085 Node* str2_offset = load_String_offset(no_ctrl, str2);
1086 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1088 Node* str2_len = NULL;
1089 Node* result = NULL;
1091 switch (opcode) {
1092 case Op_StrIndexOf:
1093 // Get length of string 2
1094 str2_len = load_String_length(no_ctrl, str2);
1096 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1097 str1_start, str1_len, str2_start, str2_len);
1098 break;
1099 case Op_StrComp:
1100 // Get length of string 2
1101 str2_len = load_String_length(no_ctrl, str2);
1103 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1104 str1_start, str1_len, str2_start, str2_len);
1105 break;
1106 case Op_StrEquals:
1107 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1108 str1_start, str2_start, str1_len);
1109 break;
1110 default:
1111 ShouldNotReachHere();
1112 return NULL;
1113 }
1115 // All these intrinsics have checks.
1116 C->set_has_split_ifs(true); // Has chance for split-if optimization
1118 return _gvn.transform(result);
1119 }
1121 // Helper method for String intrinsic functions. This version is called
1122 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1123 // to Int nodes containing the lenghts of str1 and str2.
1124 //
1125 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1126 Node* result = NULL;
1127 switch (opcode) {
1128 case Op_StrIndexOf:
1129 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1130 str1_start, cnt1, str2_start, cnt2);
1131 break;
1132 case Op_StrComp:
1133 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1134 str1_start, cnt1, str2_start, cnt2);
1135 break;
1136 case Op_StrEquals:
1137 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1138 str1_start, str2_start, cnt1);
1139 break;
1140 default:
1141 ShouldNotReachHere();
1142 return NULL;
1143 }
1145 // All these intrinsics have checks.
1146 C->set_has_split_ifs(true); // Has chance for split-if optimization
1148 return _gvn.transform(result);
1149 }
1151 //------------------------------inline_string_compareTo------------------------
1152 // public int java.lang.String.compareTo(String anotherString);
1153 bool LibraryCallKit::inline_string_compareTo() {
1154 Node* receiver = null_check(argument(0));
1155 Node* arg = null_check(argument(1));
1156 if (stopped()) {
1157 return true;
1158 }
1159 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1160 return true;
1161 }
1163 //------------------------------inline_string_equals------------------------
1164 bool LibraryCallKit::inline_string_equals() {
1165 Node* receiver = null_check_receiver();
1166 // NOTE: Do not null check argument for String.equals() because spec
1167 // allows to specify NULL as argument.
1168 Node* argument = this->argument(1);
1169 if (stopped()) {
1170 return true;
1171 }
1173 // paths (plus control) merge
1174 RegionNode* region = new (C) RegionNode(5);
1175 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1177 // does source == target string?
1178 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1179 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1181 Node* if_eq = generate_slow_guard(bol, NULL);
1182 if (if_eq != NULL) {
1183 // receiver == argument
1184 phi->init_req(2, intcon(1));
1185 region->init_req(2, if_eq);
1186 }
1188 // get String klass for instanceOf
1189 ciInstanceKlass* klass = env()->String_klass();
1191 if (!stopped()) {
1192 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1193 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1194 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1196 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1197 //instanceOf == true, fallthrough
1199 if (inst_false != NULL) {
1200 phi->init_req(3, intcon(0));
1201 region->init_req(3, inst_false);
1202 }
1203 }
1205 if (!stopped()) {
1206 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1208 // Properly cast the argument to String
1209 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1210 // This path is taken only when argument's type is String:NotNull.
1211 argument = cast_not_null(argument, false);
1213 Node* no_ctrl = NULL;
1215 // Get start addr of receiver
1216 Node* receiver_val = load_String_value(no_ctrl, receiver);
1217 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1218 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1220 // Get length of receiver
1221 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1223 // Get start addr of argument
1224 Node* argument_val = load_String_value(no_ctrl, argument);
1225 Node* argument_offset = load_String_offset(no_ctrl, argument);
1226 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1228 // Get length of argument
1229 Node* argument_cnt = load_String_length(no_ctrl, argument);
1231 // Check for receiver count != argument count
1232 Node* cmp = _gvn.transform(new(C) CmpINode(receiver_cnt, argument_cnt));
1233 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::ne));
1234 Node* if_ne = generate_slow_guard(bol, NULL);
1235 if (if_ne != NULL) {
1236 phi->init_req(4, intcon(0));
1237 region->init_req(4, if_ne);
1238 }
1240 // Check for count == 0 is done by assembler code for StrEquals.
1242 if (!stopped()) {
1243 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1244 phi->init_req(1, equals);
1245 region->init_req(1, control());
1246 }
1247 }
1249 // post merge
1250 set_control(_gvn.transform(region));
1251 record_for_igvn(region);
1253 set_result(_gvn.transform(phi));
1254 return true;
1255 }
1257 //------------------------------inline_array_equals----------------------------
1258 bool LibraryCallKit::inline_array_equals() {
1259 Node* arg1 = argument(0);
1260 Node* arg2 = argument(1);
1261 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1262 return true;
1263 }
1265 // Java version of String.indexOf(constant string)
1266 // class StringDecl {
1267 // StringDecl(char[] ca) {
1268 // offset = 0;
1269 // count = ca.length;
1270 // value = ca;
1271 // }
1272 // int offset;
1273 // int count;
1274 // char[] value;
1275 // }
1276 //
1277 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1278 // int targetOffset, int cache_i, int md2) {
1279 // int cache = cache_i;
1280 // int sourceOffset = string_object.offset;
1281 // int sourceCount = string_object.count;
1282 // int targetCount = target_object.length;
1283 //
1284 // int targetCountLess1 = targetCount - 1;
1285 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1286 //
1287 // char[] source = string_object.value;
1288 // char[] target = target_object;
1289 // int lastChar = target[targetCountLess1];
1290 //
1291 // outer_loop:
1292 // for (int i = sourceOffset; i < sourceEnd; ) {
1293 // int src = source[i + targetCountLess1];
1294 // if (src == lastChar) {
1295 // // With random strings and a 4-character alphabet,
1296 // // reverse matching at this point sets up 0.8% fewer
1297 // // frames, but (paradoxically) makes 0.3% more probes.
1298 // // Since those probes are nearer the lastChar probe,
1299 // // there is may be a net D$ win with reverse matching.
1300 // // But, reversing loop inhibits unroll of inner loop
1301 // // for unknown reason. So, does running outer loop from
1302 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1303 // for (int j = 0; j < targetCountLess1; j++) {
1304 // if (target[targetOffset + j] != source[i+j]) {
1305 // if ((cache & (1 << source[i+j])) == 0) {
1306 // if (md2 < j+1) {
1307 // i += j+1;
1308 // continue outer_loop;
1309 // }
1310 // }
1311 // i += md2;
1312 // continue outer_loop;
1313 // }
1314 // }
1315 // return i - sourceOffset;
1316 // }
1317 // if ((cache & (1 << src)) == 0) {
1318 // i += targetCountLess1;
1319 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1320 // i++;
1321 // }
1322 // return -1;
1323 // }
1325 //------------------------------string_indexOf------------------------
1326 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1327 jint cache_i, jint md2_i) {
1329 Node* no_ctrl = NULL;
1330 float likely = PROB_LIKELY(0.9);
1331 float unlikely = PROB_UNLIKELY(0.9);
1333 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1335 Node* source = load_String_value(no_ctrl, string_object);
1336 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1337 Node* sourceCount = load_String_length(no_ctrl, string_object);
1339 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)));
1340 jint target_length = target_array->length();
1341 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1342 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1344 // String.value field is known to be @Stable.
1345 if (UseImplicitStableValues) {
1346 target = cast_array_to_stable(target, target_type);
1347 }
1349 IdealKit kit(this, false, true);
1350 #define __ kit.
1351 Node* zero = __ ConI(0);
1352 Node* one = __ ConI(1);
1353 Node* cache = __ ConI(cache_i);
1354 Node* md2 = __ ConI(md2_i);
1355 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1356 Node* targetCount = __ ConI(target_length);
1357 Node* targetCountLess1 = __ ConI(target_length - 1);
1358 Node* targetOffset = __ ConI(targetOffset_i);
1359 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1361 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1362 Node* outer_loop = __ make_label(2 /* goto */);
1363 Node* return_ = __ make_label(1);
1365 __ set(rtn,__ ConI(-1));
1366 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1367 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1368 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1369 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1370 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1371 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1372 Node* tpj = __ AddI(targetOffset, __ value(j));
1373 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1374 Node* ipj = __ AddI(__ value(i), __ value(j));
1375 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1376 __ if_then(targ, BoolTest::ne, src2); {
1377 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1378 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1379 __ increment(i, __ AddI(__ value(j), one));
1380 __ goto_(outer_loop);
1381 } __ end_if(); __ dead(j);
1382 }__ end_if(); __ dead(j);
1383 __ increment(i, md2);
1384 __ goto_(outer_loop);
1385 }__ end_if();
1386 __ increment(j, one);
1387 }__ end_loop(); __ dead(j);
1388 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1389 __ goto_(return_);
1390 }__ end_if();
1391 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1392 __ increment(i, targetCountLess1);
1393 }__ end_if();
1394 __ increment(i, one);
1395 __ bind(outer_loop);
1396 }__ end_loop(); __ dead(i);
1397 __ bind(return_);
1399 // Final sync IdealKit and GraphKit.
1400 final_sync(kit);
1401 Node* result = __ value(rtn);
1402 #undef __
1403 C->set_has_loops(true);
1404 return result;
1405 }
1407 //------------------------------inline_string_indexOf------------------------
1408 bool LibraryCallKit::inline_string_indexOf() {
1409 Node* receiver = argument(0);
1410 Node* arg = argument(1);
1412 Node* result;
1413 // Disable the use of pcmpestri until it can be guaranteed that
1414 // the load doesn't cross into the uncommited space.
1415 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1416 UseSSE42Intrinsics) {
1417 // Generate SSE4.2 version of indexOf
1418 // We currently only have match rules that use SSE4.2
1420 receiver = null_check(receiver);
1421 arg = null_check(arg);
1422 if (stopped()) {
1423 return true;
1424 }
1426 ciInstanceKlass* str_klass = env()->String_klass();
1427 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1429 // Make the merge point
1430 RegionNode* result_rgn = new (C) RegionNode(4);
1431 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1432 Node* no_ctrl = NULL;
1434 // Get start addr of source string
1435 Node* source = load_String_value(no_ctrl, receiver);
1436 Node* source_offset = load_String_offset(no_ctrl, receiver);
1437 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1439 // Get length of source string
1440 Node* source_cnt = load_String_length(no_ctrl, receiver);
1442 // Get start addr of substring
1443 Node* substr = load_String_value(no_ctrl, arg);
1444 Node* substr_offset = load_String_offset(no_ctrl, arg);
1445 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1447 // Get length of source string
1448 Node* substr_cnt = load_String_length(no_ctrl, arg);
1450 // Check for substr count > string count
1451 Node* cmp = _gvn.transform(new(C) CmpINode(substr_cnt, source_cnt));
1452 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::gt));
1453 Node* if_gt = generate_slow_guard(bol, NULL);
1454 if (if_gt != NULL) {
1455 result_phi->init_req(2, intcon(-1));
1456 result_rgn->init_req(2, if_gt);
1457 }
1459 if (!stopped()) {
1460 // Check for substr count == 0
1461 cmp = _gvn.transform(new(C) CmpINode(substr_cnt, intcon(0)));
1462 bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
1463 Node* if_zero = generate_slow_guard(bol, NULL);
1464 if (if_zero != NULL) {
1465 result_phi->init_req(3, intcon(0));
1466 result_rgn->init_req(3, if_zero);
1467 }
1468 }
1470 if (!stopped()) {
1471 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1472 result_phi->init_req(1, result);
1473 result_rgn->init_req(1, control());
1474 }
1475 set_control(_gvn.transform(result_rgn));
1476 record_for_igvn(result_rgn);
1477 result = _gvn.transform(result_phi);
1479 } else { // Use LibraryCallKit::string_indexOf
1480 // don't intrinsify if argument isn't a constant string.
1481 if (!arg->is_Con()) {
1482 return false;
1483 }
1484 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1485 if (str_type == NULL) {
1486 return false;
1487 }
1488 ciInstanceKlass* klass = env()->String_klass();
1489 ciObject* str_const = str_type->const_oop();
1490 if (str_const == NULL || str_const->klass() != klass) {
1491 return false;
1492 }
1493 ciInstance* str = str_const->as_instance();
1494 assert(str != NULL, "must be instance");
1496 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1497 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1499 int o;
1500 int c;
1501 if (java_lang_String::has_offset_field()) {
1502 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1503 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1504 } else {
1505 o = 0;
1506 c = pat->length();
1507 }
1509 // constant strings have no offset and count == length which
1510 // simplifies the resulting code somewhat so lets optimize for that.
1511 if (o != 0 || c != pat->length()) {
1512 return false;
1513 }
1515 receiver = null_check(receiver, T_OBJECT);
1516 // NOTE: No null check on the argument is needed since it's a constant String oop.
1517 if (stopped()) {
1518 return true;
1519 }
1521 // The null string as a pattern always returns 0 (match at beginning of string)
1522 if (c == 0) {
1523 set_result(intcon(0));
1524 return true;
1525 }
1527 // Generate default indexOf
1528 jchar lastChar = pat->char_at(o + (c - 1));
1529 int cache = 0;
1530 int i;
1531 for (i = 0; i < c - 1; i++) {
1532 assert(i < pat->length(), "out of range");
1533 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1534 }
1536 int md2 = c;
1537 for (i = 0; i < c - 1; i++) {
1538 assert(i < pat->length(), "out of range");
1539 if (pat->char_at(o + i) == lastChar) {
1540 md2 = (c - 1) - i;
1541 }
1542 }
1544 result = string_indexOf(receiver, pat, o, cache, md2);
1545 }
1546 set_result(result);
1547 return true;
1548 }
1550 //--------------------------round_double_node--------------------------------
1551 // Round a double node if necessary.
1552 Node* LibraryCallKit::round_double_node(Node* n) {
1553 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1554 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1555 return n;
1556 }
1558 //------------------------------inline_math-----------------------------------
1559 // public static double Math.abs(double)
1560 // public static double Math.sqrt(double)
1561 // public static double Math.log(double)
1562 // public static double Math.log10(double)
1563 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1564 Node* arg = round_double_node(argument(0));
1565 Node* n;
1566 switch (id) {
1567 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1568 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1569 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1570 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1571 default: fatal_unexpected_iid(id); break;
1572 }
1573 set_result(_gvn.transform(n));
1574 return true;
1575 }
1577 //------------------------------inline_trig----------------------------------
1578 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1579 // argument reduction which will turn into a fast/slow diamond.
1580 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1581 Node* arg = round_double_node(argument(0));
1582 Node* n = NULL;
1584 switch (id) {
1585 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1586 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1587 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1588 default: fatal_unexpected_iid(id); break;
1589 }
1590 n = _gvn.transform(n);
1592 // Rounding required? Check for argument reduction!
1593 if (Matcher::strict_fp_requires_explicit_rounding) {
1594 static const double pi_4 = 0.7853981633974483;
1595 static const double neg_pi_4 = -0.7853981633974483;
1596 // pi/2 in 80-bit extended precision
1597 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1598 // -pi/2 in 80-bit extended precision
1599 // 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};
1600 // Cutoff value for using this argument reduction technique
1601 //static const double pi_2_minus_epsilon = 1.564660403643354;
1602 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1604 // Pseudocode for sin:
1605 // if (x <= Math.PI / 4.0) {
1606 // if (x >= -Math.PI / 4.0) return fsin(x);
1607 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1608 // } else {
1609 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1610 // }
1611 // return StrictMath.sin(x);
1613 // Pseudocode for cos:
1614 // if (x <= Math.PI / 4.0) {
1615 // if (x >= -Math.PI / 4.0) return fcos(x);
1616 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1617 // } else {
1618 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1619 // }
1620 // return StrictMath.cos(x);
1622 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1623 // requires a special machine instruction to load it. Instead we'll try
1624 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1625 // probably do the math inside the SIN encoding.
1627 // Make the merge point
1628 RegionNode* r = new (C) RegionNode(3);
1629 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1631 // Flatten arg so we need only 1 test
1632 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1633 // Node for PI/4 constant
1634 Node *pi4 = makecon(TypeD::make(pi_4));
1635 // Check PI/4 : abs(arg)
1636 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1637 // Check: If PI/4 < abs(arg) then go slow
1638 Node *bol = _gvn.transform(new (C) BoolNode( cmp, BoolTest::lt ));
1639 // Branch either way
1640 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1641 set_control(opt_iff(r,iff));
1643 // Set fast path result
1644 phi->init_req(2, n);
1646 // Slow path - non-blocking leaf call
1647 Node* call = NULL;
1648 switch (id) {
1649 case vmIntrinsics::_dsin:
1650 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1651 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1652 "Sin", NULL, arg, top());
1653 break;
1654 case vmIntrinsics::_dcos:
1655 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1656 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1657 "Cos", NULL, arg, top());
1658 break;
1659 case vmIntrinsics::_dtan:
1660 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1661 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1662 "Tan", NULL, arg, top());
1663 break;
1664 }
1665 assert(control()->in(0) == call, "");
1666 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1667 r->init_req(1, control());
1668 phi->init_req(1, slow_result);
1670 // Post-merge
1671 set_control(_gvn.transform(r));
1672 record_for_igvn(r);
1673 n = _gvn.transform(phi);
1675 C->set_has_split_ifs(true); // Has chance for split-if optimization
1676 }
1677 set_result(n);
1678 return true;
1679 }
1681 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1682 //-------------------
1683 //result=(result.isNaN())? funcAddr():result;
1684 // Check: If isNaN() by checking result!=result? then either trap
1685 // or go to runtime
1686 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1687 // Build the boolean node
1688 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1690 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1691 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1692 // The pow or exp intrinsic returned a NaN, which requires a call
1693 // to the runtime. Recompile with the runtime call.
1694 uncommon_trap(Deoptimization::Reason_intrinsic,
1695 Deoptimization::Action_make_not_entrant);
1696 }
1697 set_result(result);
1698 } else {
1699 // If this inlining ever returned NaN in the past, we compile a call
1700 // to the runtime to properly handle corner cases
1702 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1703 Node* if_slow = _gvn.transform(new (C) IfFalseNode(iff));
1704 Node* if_fast = _gvn.transform(new (C) IfTrueNode(iff));
1706 if (!if_slow->is_top()) {
1707 RegionNode* result_region = new (C) RegionNode(3);
1708 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1710 result_region->init_req(1, if_fast);
1711 result_val->init_req(1, result);
1713 set_control(if_slow);
1715 const TypePtr* no_memory_effects = NULL;
1716 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1717 no_memory_effects,
1718 x, top(), y, y ? top() : NULL);
1719 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1720 #ifdef ASSERT
1721 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1722 assert(value_top == top(), "second value must be top");
1723 #endif
1725 result_region->init_req(2, control());
1726 result_val->init_req(2, value);
1727 set_result(result_region, result_val);
1728 } else {
1729 set_result(result);
1730 }
1731 }
1732 }
1734 //------------------------------inline_exp-------------------------------------
1735 // Inline exp instructions, if possible. The Intel hardware only misses
1736 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1737 bool LibraryCallKit::inline_exp() {
1738 Node* arg = round_double_node(argument(0));
1739 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1741 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1743 C->set_has_split_ifs(true); // Has chance for split-if optimization
1744 return true;
1745 }
1747 //------------------------------inline_pow-------------------------------------
1748 // Inline power instructions, if possible.
1749 bool LibraryCallKit::inline_pow() {
1750 // Pseudocode for pow
1751 // if (x <= 0.0) {
1752 // long longy = (long)y;
1753 // if ((double)longy == y) { // if y is long
1754 // if (y + 1 == y) longy = 0; // huge number: even
1755 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1756 // } else {
1757 // result = NaN;
1758 // }
1759 // } else {
1760 // result = DPow(x,y);
1761 // }
1762 // if (result != result)? {
1763 // result = uncommon_trap() or runtime_call();
1764 // }
1765 // return result;
1767 Node* x = round_double_node(argument(0));
1768 Node* y = round_double_node(argument(2));
1770 Node* result = NULL;
1772 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1773 // Short form: skip the fancy tests and just check for NaN result.
1774 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1775 } else {
1776 // If this inlining ever returned NaN in the past, include all
1777 // checks + call to the runtime.
1779 // Set the merge point for If node with condition of (x <= 0.0)
1780 // There are four possible paths to region node and phi node
1781 RegionNode *r = new (C) RegionNode(4);
1782 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1784 // Build the first if node: if (x <= 0.0)
1785 // Node for 0 constant
1786 Node *zeronode = makecon(TypeD::ZERO);
1787 // Check x:0
1788 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1789 // Check: If (x<=0) then go complex path
1790 Node *bol1 = _gvn.transform(new (C) BoolNode( cmp, BoolTest::le ));
1791 // Branch either way
1792 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1793 // Fast path taken; set region slot 3
1794 Node *fast_taken = _gvn.transform(new (C) IfFalseNode(if1));
1795 r->init_req(3,fast_taken); // Capture fast-control
1797 // Fast path not-taken, i.e. slow path
1798 Node *complex_path = _gvn.transform(new (C) IfTrueNode(if1));
1800 // Set fast path result
1801 Node *fast_result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1802 phi->init_req(3, fast_result);
1804 // Complex path
1805 // Build the second if node (if y is long)
1806 // Node for (long)y
1807 Node *longy = _gvn.transform(new (C) ConvD2LNode(y));
1808 // Node for (double)((long) y)
1809 Node *doublelongy= _gvn.transform(new (C) ConvL2DNode(longy));
1810 // Check (double)((long) y) : y
1811 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1812 // Check if (y isn't long) then go to slow path
1814 Node *bol2 = _gvn.transform(new (C) BoolNode( cmplongy, BoolTest::ne ));
1815 // Branch either way
1816 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1817 Node* ylong_path = _gvn.transform(new (C) IfFalseNode(if2));
1819 Node *slow_path = _gvn.transform(new (C) IfTrueNode(if2));
1821 // Calculate DPow(abs(x), y)*(1 & (long)y)
1822 // Node for constant 1
1823 Node *conone = longcon(1);
1824 // 1& (long)y
1825 Node *signnode= _gvn.transform(new (C) AndLNode(conone, longy));
1827 // A huge number is always even. Detect a huge number by checking
1828 // if y + 1 == y and set integer to be tested for parity to 0.
1829 // Required for corner case:
1830 // (long)9.223372036854776E18 = max_jlong
1831 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1832 // max_jlong is odd but 9.223372036854776E18 is even
1833 Node* yplus1 = _gvn.transform(new (C) AddDNode(y, makecon(TypeD::make(1))));
1834 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1835 Node *bolyplus1 = _gvn.transform(new (C) BoolNode( cmpyplus1, BoolTest::eq ));
1836 Node* correctedsign = NULL;
1837 if (ConditionalMoveLimit != 0) {
1838 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1839 } else {
1840 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1841 RegionNode *r = new (C) RegionNode(3);
1842 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1843 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyplus1)));
1844 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyplus1)));
1845 phi->init_req(1, signnode);
1846 phi->init_req(2, longcon(0));
1847 correctedsign = _gvn.transform(phi);
1848 ylong_path = _gvn.transform(r);
1849 record_for_igvn(r);
1850 }
1852 // zero node
1853 Node *conzero = longcon(0);
1854 // Check (1&(long)y)==0?
1855 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1856 // Check if (1&(long)y)!=0?, if so the result is negative
1857 Node *bol3 = _gvn.transform(new (C) BoolNode( cmpeq1, BoolTest::ne ));
1858 // abs(x)
1859 Node *absx=_gvn.transform(new (C) AbsDNode(x));
1860 // abs(x)^y
1861 Node *absxpowy = _gvn.transform(new (C) PowDNode(C, control(), absx, y));
1862 // -abs(x)^y
1863 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1864 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1865 Node *signresult = NULL;
1866 if (ConditionalMoveLimit != 0) {
1867 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1868 } else {
1869 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1870 RegionNode *r = new (C) RegionNode(3);
1871 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1872 r->init_req(1, _gvn.transform(new (C) IfFalseNode(ifyeven)));
1873 r->init_req(2, _gvn.transform(new (C) IfTrueNode(ifyeven)));
1874 phi->init_req(1, absxpowy);
1875 phi->init_req(2, negabsxpowy);
1876 signresult = _gvn.transform(phi);
1877 ylong_path = _gvn.transform(r);
1878 record_for_igvn(r);
1879 }
1880 // Set complex path fast result
1881 r->init_req(2, ylong_path);
1882 phi->init_req(2, signresult);
1884 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1885 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1886 r->init_req(1,slow_path);
1887 phi->init_req(1,slow_result);
1889 // Post merge
1890 set_control(_gvn.transform(r));
1891 record_for_igvn(r);
1892 result = _gvn.transform(phi);
1893 }
1895 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1897 C->set_has_split_ifs(true); // Has chance for split-if optimization
1898 return true;
1899 }
1901 //------------------------------runtime_math-----------------------------
1902 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1903 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1904 "must be (DD)D or (D)D type");
1906 // Inputs
1907 Node* a = round_double_node(argument(0));
1908 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1910 const TypePtr* no_memory_effects = NULL;
1911 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1912 no_memory_effects,
1913 a, top(), b, b ? top() : NULL);
1914 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1915 #ifdef ASSERT
1916 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1917 assert(value_top == top(), "second value must be top");
1918 #endif
1920 set_result(value);
1921 return true;
1922 }
1924 //------------------------------inline_math_native-----------------------------
1925 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1926 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1927 switch (id) {
1928 // These intrinsics are not properly supported on all hardware
1929 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1930 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1931 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1932 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1933 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1934 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1936 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1937 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1938 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1939 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1941 // These intrinsics are supported on all hardware
1942 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1943 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1945 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1946 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1947 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1948 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1949 #undef FN_PTR
1951 // These intrinsics are not yet correctly implemented
1952 case vmIntrinsics::_datan2:
1953 return false;
1955 default:
1956 fatal_unexpected_iid(id);
1957 return false;
1958 }
1959 }
1961 static bool is_simple_name(Node* n) {
1962 return (n->req() == 1 // constant
1963 || (n->is_Type() && n->as_Type()->type()->singleton())
1964 || n->is_Proj() // parameter or return value
1965 || n->is_Phi() // local of some sort
1966 );
1967 }
1969 //----------------------------inline_min_max-----------------------------------
1970 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1971 set_result(generate_min_max(id, argument(0), argument(1)));
1972 return true;
1973 }
1975 void LibraryCallKit::inline_math_mathExact(Node* math, Node *test) {
1976 Node* bol = _gvn.transform( new (C) BoolNode(test, BoolTest::overflow) );
1977 IfNode* check = create_and_map_if(control(), bol, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
1978 Node* fast_path = _gvn.transform( new (C) IfFalseNode(check));
1979 Node* slow_path = _gvn.transform( new (C) IfTrueNode(check) );
1981 {
1982 PreserveJVMState pjvms(this);
1983 PreserveReexecuteState preexecs(this);
1984 jvms()->set_should_reexecute(true);
1986 set_control(slow_path);
1987 set_i_o(i_o());
1989 uncommon_trap(Deoptimization::Reason_intrinsic,
1990 Deoptimization::Action_none);
1991 }
1993 set_control(fast_path);
1994 set_result(math);
1995 }
1997 template <typename OverflowOp>
1998 bool LibraryCallKit::inline_math_overflow(Node* arg1, Node* arg2) {
1999 typedef typename OverflowOp::MathOp MathOp;
2001 MathOp* mathOp = new(C) MathOp(arg1, arg2);
2002 Node* operation = _gvn.transform( mathOp );
2003 Node* ofcheck = _gvn.transform( new(C) OverflowOp(arg1, arg2) );
2004 inline_math_mathExact(operation, ofcheck);
2005 return true;
2006 }
2008 bool LibraryCallKit::inline_math_addExactI(bool is_increment) {
2009 return inline_math_overflow<OverflowAddINode>(argument(0), is_increment ? intcon(1) : argument(1));
2010 }
2012 bool LibraryCallKit::inline_math_addExactL(bool is_increment) {
2013 return inline_math_overflow<OverflowAddLNode>(argument(0), is_increment ? longcon(1) : argument(2));
2014 }
2016 bool LibraryCallKit::inline_math_subtractExactI(bool is_decrement) {
2017 return inline_math_overflow<OverflowSubINode>(argument(0), is_decrement ? intcon(1) : argument(1));
2018 }
2020 bool LibraryCallKit::inline_math_subtractExactL(bool is_decrement) {
2021 return inline_math_overflow<OverflowSubLNode>(argument(0), is_decrement ? longcon(1) : argument(2));
2022 }
2024 bool LibraryCallKit::inline_math_negateExactI() {
2025 return inline_math_overflow<OverflowSubINode>(intcon(0), argument(0));
2026 }
2028 bool LibraryCallKit::inline_math_negateExactL() {
2029 return inline_math_overflow<OverflowSubLNode>(longcon(0), argument(0));
2030 }
2032 bool LibraryCallKit::inline_math_multiplyExactI() {
2033 return inline_math_overflow<OverflowMulINode>(argument(0), argument(1));
2034 }
2036 bool LibraryCallKit::inline_math_multiplyExactL() {
2037 return inline_math_overflow<OverflowMulLNode>(argument(0), argument(2));
2038 }
2040 Node*
2041 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2042 // These are the candidate return value:
2043 Node* xvalue = x0;
2044 Node* yvalue = y0;
2046 if (xvalue == yvalue) {
2047 return xvalue;
2048 }
2050 bool want_max = (id == vmIntrinsics::_max);
2052 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2053 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2054 if (txvalue == NULL || tyvalue == NULL) return top();
2055 // This is not really necessary, but it is consistent with a
2056 // hypothetical MaxINode::Value method:
2057 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2059 // %%% This folding logic should (ideally) be in a different place.
2060 // Some should be inside IfNode, and there to be a more reliable
2061 // transformation of ?: style patterns into cmoves. We also want
2062 // more powerful optimizations around cmove and min/max.
2064 // Try to find a dominating comparison of these guys.
2065 // It can simplify the index computation for Arrays.copyOf
2066 // and similar uses of System.arraycopy.
2067 // First, compute the normalized version of CmpI(x, y).
2068 int cmp_op = Op_CmpI;
2069 Node* xkey = xvalue;
2070 Node* ykey = yvalue;
2071 Node* ideal_cmpxy = _gvn.transform(new(C) CmpINode(xkey, ykey));
2072 if (ideal_cmpxy->is_Cmp()) {
2073 // E.g., if we have CmpI(length - offset, count),
2074 // it might idealize to CmpI(length, count + offset)
2075 cmp_op = ideal_cmpxy->Opcode();
2076 xkey = ideal_cmpxy->in(1);
2077 ykey = ideal_cmpxy->in(2);
2078 }
2080 // Start by locating any relevant comparisons.
2081 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2082 Node* cmpxy = NULL;
2083 Node* cmpyx = NULL;
2084 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2085 Node* cmp = start_from->fast_out(k);
2086 if (cmp->outcnt() > 0 && // must have prior uses
2087 cmp->in(0) == NULL && // must be context-independent
2088 cmp->Opcode() == cmp_op) { // right kind of compare
2089 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2090 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2091 }
2092 }
2094 const int NCMPS = 2;
2095 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2096 int cmpn;
2097 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2098 if (cmps[cmpn] != NULL) break; // find a result
2099 }
2100 if (cmpn < NCMPS) {
2101 // Look for a dominating test that tells us the min and max.
2102 int depth = 0; // Limit search depth for speed
2103 Node* dom = control();
2104 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2105 if (++depth >= 100) break;
2106 Node* ifproj = dom;
2107 if (!ifproj->is_Proj()) continue;
2108 Node* iff = ifproj->in(0);
2109 if (!iff->is_If()) continue;
2110 Node* bol = iff->in(1);
2111 if (!bol->is_Bool()) continue;
2112 Node* cmp = bol->in(1);
2113 if (cmp == NULL) continue;
2114 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2115 if (cmps[cmpn] == cmp) break;
2116 if (cmpn == NCMPS) continue;
2117 BoolTest::mask btest = bol->as_Bool()->_test._test;
2118 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2119 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2120 // At this point, we know that 'x btest y' is true.
2121 switch (btest) {
2122 case BoolTest::eq:
2123 // They are proven equal, so we can collapse the min/max.
2124 // Either value is the answer. Choose the simpler.
2125 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2126 return yvalue;
2127 return xvalue;
2128 case BoolTest::lt: // x < y
2129 case BoolTest::le: // x <= y
2130 return (want_max ? yvalue : xvalue);
2131 case BoolTest::gt: // x > y
2132 case BoolTest::ge: // x >= y
2133 return (want_max ? xvalue : yvalue);
2134 }
2135 }
2136 }
2138 // We failed to find a dominating test.
2139 // Let's pick a test that might GVN with prior tests.
2140 Node* best_bol = NULL;
2141 BoolTest::mask best_btest = BoolTest::illegal;
2142 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2143 Node* cmp = cmps[cmpn];
2144 if (cmp == NULL) continue;
2145 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2146 Node* bol = cmp->fast_out(j);
2147 if (!bol->is_Bool()) continue;
2148 BoolTest::mask btest = bol->as_Bool()->_test._test;
2149 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2150 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2151 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2152 best_bol = bol->as_Bool();
2153 best_btest = btest;
2154 }
2155 }
2156 }
2158 Node* answer_if_true = NULL;
2159 Node* answer_if_false = NULL;
2160 switch (best_btest) {
2161 default:
2162 if (cmpxy == NULL)
2163 cmpxy = ideal_cmpxy;
2164 best_bol = _gvn.transform(new(C) BoolNode(cmpxy, BoolTest::lt));
2165 // and fall through:
2166 case BoolTest::lt: // x < y
2167 case BoolTest::le: // x <= y
2168 answer_if_true = (want_max ? yvalue : xvalue);
2169 answer_if_false = (want_max ? xvalue : yvalue);
2170 break;
2171 case BoolTest::gt: // x > y
2172 case BoolTest::ge: // x >= y
2173 answer_if_true = (want_max ? xvalue : yvalue);
2174 answer_if_false = (want_max ? yvalue : xvalue);
2175 break;
2176 }
2178 jint hi, lo;
2179 if (want_max) {
2180 // We can sharpen the minimum.
2181 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2182 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2183 } else {
2184 // We can sharpen the maximum.
2185 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2186 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2187 }
2189 // Use a flow-free graph structure, to avoid creating excess control edges
2190 // which could hinder other optimizations.
2191 // Since Math.min/max is often used with arraycopy, we want
2192 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2193 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2194 answer_if_false, answer_if_true,
2195 TypeInt::make(lo, hi, widen));
2197 return _gvn.transform(cmov);
2199 /*
2200 // This is not as desirable as it may seem, since Min and Max
2201 // nodes do not have a full set of optimizations.
2202 // And they would interfere, anyway, with 'if' optimizations
2203 // and with CMoveI canonical forms.
2204 switch (id) {
2205 case vmIntrinsics::_min:
2206 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2207 case vmIntrinsics::_max:
2208 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2209 default:
2210 ShouldNotReachHere();
2211 }
2212 */
2213 }
2215 inline int
2216 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2217 const TypePtr* base_type = TypePtr::NULL_PTR;
2218 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2219 if (base_type == NULL) {
2220 // Unknown type.
2221 return Type::AnyPtr;
2222 } else if (base_type == TypePtr::NULL_PTR) {
2223 // Since this is a NULL+long form, we have to switch to a rawptr.
2224 base = _gvn.transform(new (C) CastX2PNode(offset));
2225 offset = MakeConX(0);
2226 return Type::RawPtr;
2227 } else if (base_type->base() == Type::RawPtr) {
2228 return Type::RawPtr;
2229 } else if (base_type->isa_oopptr()) {
2230 // Base is never null => always a heap address.
2231 if (base_type->ptr() == TypePtr::NotNull) {
2232 return Type::OopPtr;
2233 }
2234 // Offset is small => always a heap address.
2235 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2236 if (offset_type != NULL &&
2237 base_type->offset() == 0 && // (should always be?)
2238 offset_type->_lo >= 0 &&
2239 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2240 return Type::OopPtr;
2241 }
2242 // Otherwise, it might either be oop+off or NULL+addr.
2243 return Type::AnyPtr;
2244 } else {
2245 // No information:
2246 return Type::AnyPtr;
2247 }
2248 }
2250 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2251 int kind = classify_unsafe_addr(base, offset);
2252 if (kind == Type::RawPtr) {
2253 return basic_plus_adr(top(), base, offset);
2254 } else {
2255 return basic_plus_adr(base, offset);
2256 }
2257 }
2259 //--------------------------inline_number_methods-----------------------------
2260 // inline int Integer.numberOfLeadingZeros(int)
2261 // inline int Long.numberOfLeadingZeros(long)
2262 //
2263 // inline int Integer.numberOfTrailingZeros(int)
2264 // inline int Long.numberOfTrailingZeros(long)
2265 //
2266 // inline int Integer.bitCount(int)
2267 // inline int Long.bitCount(long)
2268 //
2269 // inline char Character.reverseBytes(char)
2270 // inline short Short.reverseBytes(short)
2271 // inline int Integer.reverseBytes(int)
2272 // inline long Long.reverseBytes(long)
2273 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2274 Node* arg = argument(0);
2275 Node* n;
2276 switch (id) {
2277 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2278 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2279 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2280 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2281 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2282 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2283 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2284 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2285 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2286 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2287 default: fatal_unexpected_iid(id); break;
2288 }
2289 set_result(_gvn.transform(n));
2290 return true;
2291 }
2293 //----------------------------inline_unsafe_access----------------------------
2295 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2297 // Helper that guards and inserts a pre-barrier.
2298 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2299 Node* pre_val, bool need_mem_bar) {
2300 // We could be accessing the referent field of a reference object. If so, when G1
2301 // is enabled, we need to log the value in the referent field in an SATB buffer.
2302 // This routine performs some compile time filters and generates suitable
2303 // runtime filters that guard the pre-barrier code.
2304 // Also add memory barrier for non volatile load from the referent field
2305 // to prevent commoning of loads across safepoint.
2306 if (!UseG1GC && !need_mem_bar)
2307 return;
2309 // Some compile time checks.
2311 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2312 const TypeX* otype = offset->find_intptr_t_type();
2313 if (otype != NULL && otype->is_con() &&
2314 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2315 // Constant offset but not the reference_offset so just return
2316 return;
2317 }
2319 // We only need to generate the runtime guards for instances.
2320 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2321 if (btype != NULL) {
2322 if (btype->isa_aryptr()) {
2323 // Array type so nothing to do
2324 return;
2325 }
2327 const TypeInstPtr* itype = btype->isa_instptr();
2328 if (itype != NULL) {
2329 // Can the klass of base_oop be statically determined to be
2330 // _not_ a sub-class of Reference and _not_ Object?
2331 ciKlass* klass = itype->klass();
2332 if ( klass->is_loaded() &&
2333 !klass->is_subtype_of(env()->Reference_klass()) &&
2334 !env()->Object_klass()->is_subtype_of(klass)) {
2335 return;
2336 }
2337 }
2338 }
2340 // The compile time filters did not reject base_oop/offset so
2341 // we need to generate the following runtime filters
2342 //
2343 // if (offset == java_lang_ref_Reference::_reference_offset) {
2344 // if (instance_of(base, java.lang.ref.Reference)) {
2345 // pre_barrier(_, pre_val, ...);
2346 // }
2347 // }
2349 float likely = PROB_LIKELY( 0.999);
2350 float unlikely = PROB_UNLIKELY(0.999);
2352 IdealKit ideal(this);
2353 #define __ ideal.
2355 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2357 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2358 // Update graphKit memory and control from IdealKit.
2359 sync_kit(ideal);
2361 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2362 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2364 // Update IdealKit memory and control from graphKit.
2365 __ sync_kit(this);
2367 Node* one = __ ConI(1);
2368 // is_instof == 0 if base_oop == NULL
2369 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2371 // Update graphKit from IdeakKit.
2372 sync_kit(ideal);
2374 // Use the pre-barrier to record the value in the referent field
2375 pre_barrier(false /* do_load */,
2376 __ ctrl(),
2377 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2378 pre_val /* pre_val */,
2379 T_OBJECT);
2380 if (need_mem_bar) {
2381 // Add memory barrier to prevent commoning reads from this field
2382 // across safepoint since GC can change its value.
2383 insert_mem_bar(Op_MemBarCPUOrder);
2384 }
2385 // Update IdealKit from graphKit.
2386 __ sync_kit(this);
2388 } __ end_if(); // _ref_type != ref_none
2389 } __ end_if(); // offset == referent_offset
2391 // Final sync IdealKit and GraphKit.
2392 final_sync(ideal);
2393 #undef __
2394 }
2397 // Interpret Unsafe.fieldOffset cookies correctly:
2398 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2400 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2401 // Attempt to infer a sharper value type from the offset and base type.
2402 ciKlass* sharpened_klass = NULL;
2404 // See if it is an instance field, with an object type.
2405 if (alias_type->field() != NULL) {
2406 assert(!is_native_ptr, "native pointer op cannot use a java address");
2407 if (alias_type->field()->type()->is_klass()) {
2408 sharpened_klass = alias_type->field()->type()->as_klass();
2409 }
2410 }
2412 // See if it is a narrow oop array.
2413 if (adr_type->isa_aryptr()) {
2414 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2415 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2416 if (elem_type != NULL) {
2417 sharpened_klass = elem_type->klass();
2418 }
2419 }
2420 }
2422 // The sharpened class might be unloaded if there is no class loader
2423 // contraint in place.
2424 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2425 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2427 #ifndef PRODUCT
2428 if (C->print_intrinsics() || C->print_inlining()) {
2429 tty->print(" from base type: "); adr_type->dump();
2430 tty->print(" sharpened value: "); tjp->dump();
2431 }
2432 #endif
2433 // Sharpen the value type.
2434 return tjp;
2435 }
2436 return NULL;
2437 }
2439 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2440 if (callee()->is_static()) return false; // caller must have the capability!
2442 #ifndef PRODUCT
2443 {
2444 ResourceMark rm;
2445 // Check the signatures.
2446 ciSignature* sig = callee()->signature();
2447 #ifdef ASSERT
2448 if (!is_store) {
2449 // Object getObject(Object base, int/long offset), etc.
2450 BasicType rtype = sig->return_type()->basic_type();
2451 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2452 rtype = T_ADDRESS; // it is really a C void*
2453 assert(rtype == type, "getter must return the expected value");
2454 if (!is_native_ptr) {
2455 assert(sig->count() == 2, "oop getter has 2 arguments");
2456 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2457 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2458 } else {
2459 assert(sig->count() == 1, "native getter has 1 argument");
2460 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2461 }
2462 } else {
2463 // void putObject(Object base, int/long offset, Object x), etc.
2464 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2465 if (!is_native_ptr) {
2466 assert(sig->count() == 3, "oop putter has 3 arguments");
2467 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2468 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2469 } else {
2470 assert(sig->count() == 2, "native putter has 2 arguments");
2471 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2472 }
2473 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2474 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2475 vtype = T_ADDRESS; // it is really a C void*
2476 assert(vtype == type, "putter must accept the expected value");
2477 }
2478 #endif // ASSERT
2479 }
2480 #endif //PRODUCT
2482 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2484 Node* receiver = argument(0); // type: oop
2486 // Build address expression. See the code in inline_unsafe_prefetch.
2487 Node* adr;
2488 Node* heap_base_oop = top();
2489 Node* offset = top();
2490 Node* val;
2492 if (!is_native_ptr) {
2493 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2494 Node* base = argument(1); // type: oop
2495 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2496 offset = argument(2); // type: long
2497 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2498 // to be plain byte offsets, which are also the same as those accepted
2499 // by oopDesc::field_base.
2500 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2501 "fieldOffset must be byte-scaled");
2502 // 32-bit machines ignore the high half!
2503 offset = ConvL2X(offset);
2504 adr = make_unsafe_address(base, offset);
2505 heap_base_oop = base;
2506 val = is_store ? argument(4) : NULL;
2507 } else {
2508 Node* ptr = argument(1); // type: long
2509 ptr = ConvL2X(ptr); // adjust Java long to machine word
2510 adr = make_unsafe_address(NULL, ptr);
2511 val = is_store ? argument(3) : NULL;
2512 }
2514 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2516 // First guess at the value type.
2517 const Type *value_type = Type::get_const_basic_type(type);
2519 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2520 // there was not enough information to nail it down.
2521 Compile::AliasType* alias_type = C->alias_type(adr_type);
2522 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2524 // We will need memory barriers unless we can determine a unique
2525 // alias category for this reference. (Note: If for some reason
2526 // the barriers get omitted and the unsafe reference begins to "pollute"
2527 // the alias analysis of the rest of the graph, either Compile::can_alias
2528 // or Compile::must_alias will throw a diagnostic assert.)
2529 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2531 // If we are reading the value of the referent field of a Reference
2532 // object (either by using Unsafe directly or through reflection)
2533 // then, if G1 is enabled, we need to record the referent in an
2534 // SATB log buffer using the pre-barrier mechanism.
2535 // Also we need to add memory barrier to prevent commoning reads
2536 // from this field across safepoint since GC can change its value.
2537 bool need_read_barrier = !is_native_ptr && !is_store &&
2538 offset != top() && heap_base_oop != top();
2540 if (!is_store && type == T_OBJECT) {
2541 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2542 if (tjp != NULL) {
2543 value_type = tjp;
2544 }
2545 }
2547 receiver = null_check(receiver);
2548 if (stopped()) {
2549 return true;
2550 }
2551 // Heap pointers get a null-check from the interpreter,
2552 // as a courtesy. However, this is not guaranteed by Unsafe,
2553 // and it is not possible to fully distinguish unintended nulls
2554 // from intended ones in this API.
2556 if (is_volatile) {
2557 // We need to emit leading and trailing CPU membars (see below) in
2558 // addition to memory membars when is_volatile. This is a little
2559 // too strong, but avoids the need to insert per-alias-type
2560 // volatile membars (for stores; compare Parse::do_put_xxx), which
2561 // we cannot do effectively here because we probably only have a
2562 // rough approximation of type.
2563 need_mem_bar = true;
2564 // For Stores, place a memory ordering barrier now.
2565 if (is_store)
2566 insert_mem_bar(Op_MemBarRelease);
2567 }
2569 // Memory barrier to prevent normal and 'unsafe' accesses from
2570 // bypassing each other. Happens after null checks, so the
2571 // exception paths do not take memory state from the memory barrier,
2572 // so there's no problems making a strong assert about mixing users
2573 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2574 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2575 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2577 if (!is_store) {
2578 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2579 // load value
2580 switch (type) {
2581 case T_BOOLEAN:
2582 case T_CHAR:
2583 case T_BYTE:
2584 case T_SHORT:
2585 case T_INT:
2586 case T_LONG:
2587 case T_FLOAT:
2588 case T_DOUBLE:
2589 break;
2590 case T_OBJECT:
2591 if (need_read_barrier) {
2592 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2593 }
2594 break;
2595 case T_ADDRESS:
2596 // Cast to an int type.
2597 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2598 p = ConvX2L(p);
2599 break;
2600 default:
2601 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2602 break;
2603 }
2604 // The load node has the control of the preceding MemBarCPUOrder. All
2605 // following nodes will have the control of the MemBarCPUOrder inserted at
2606 // the end of this method. So, pushing the load onto the stack at a later
2607 // point is fine.
2608 set_result(p);
2609 } else {
2610 // place effect of store into memory
2611 switch (type) {
2612 case T_DOUBLE:
2613 val = dstore_rounding(val);
2614 break;
2615 case T_ADDRESS:
2616 // Repackage the long as a pointer.
2617 val = ConvL2X(val);
2618 val = _gvn.transform(new (C) CastX2PNode(val));
2619 break;
2620 }
2622 if (type != T_OBJECT ) {
2623 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2624 } else {
2625 // Possibly an oop being stored to Java heap or native memory
2626 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2627 // oop to Java heap.
2628 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2629 } else {
2630 // We can't tell at compile time if we are storing in the Java heap or outside
2631 // of it. So we need to emit code to conditionally do the proper type of
2632 // store.
2634 IdealKit ideal(this);
2635 #define __ ideal.
2636 // QQQ who knows what probability is here??
2637 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2638 // Sync IdealKit and graphKit.
2639 sync_kit(ideal);
2640 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2641 // Update IdealKit memory.
2642 __ sync_kit(this);
2643 } __ else_(); {
2644 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2645 } __ end_if();
2646 // Final sync IdealKit and GraphKit.
2647 final_sync(ideal);
2648 #undef __
2649 }
2650 }
2651 }
2653 if (is_volatile) {
2654 if (!is_store)
2655 insert_mem_bar(Op_MemBarAcquire);
2656 else
2657 insert_mem_bar(Op_MemBarVolatile);
2658 }
2660 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2662 return true;
2663 }
2665 //----------------------------inline_unsafe_prefetch----------------------------
2667 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2668 #ifndef PRODUCT
2669 {
2670 ResourceMark rm;
2671 // Check the signatures.
2672 ciSignature* sig = callee()->signature();
2673 #ifdef ASSERT
2674 // Object getObject(Object base, int/long offset), etc.
2675 BasicType rtype = sig->return_type()->basic_type();
2676 if (!is_native_ptr) {
2677 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2678 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2679 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2680 } else {
2681 assert(sig->count() == 1, "native prefetch has 1 argument");
2682 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2683 }
2684 #endif // ASSERT
2685 }
2686 #endif // !PRODUCT
2688 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2690 const int idx = is_static ? 0 : 1;
2691 if (!is_static) {
2692 null_check_receiver();
2693 if (stopped()) {
2694 return true;
2695 }
2696 }
2698 // Build address expression. See the code in inline_unsafe_access.
2699 Node *adr;
2700 if (!is_native_ptr) {
2701 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2702 Node* base = argument(idx + 0); // type: oop
2703 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2704 Node* offset = argument(idx + 1); // type: long
2705 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2706 // to be plain byte offsets, which are also the same as those accepted
2707 // by oopDesc::field_base.
2708 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2709 "fieldOffset must be byte-scaled");
2710 // 32-bit machines ignore the high half!
2711 offset = ConvL2X(offset);
2712 adr = make_unsafe_address(base, offset);
2713 } else {
2714 Node* ptr = argument(idx + 0); // type: long
2715 ptr = ConvL2X(ptr); // adjust Java long to machine word
2716 adr = make_unsafe_address(NULL, ptr);
2717 }
2719 // Generate the read or write prefetch
2720 Node *prefetch;
2721 if (is_store) {
2722 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2723 } else {
2724 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2725 }
2726 prefetch->init_req(0, control());
2727 set_i_o(_gvn.transform(prefetch));
2729 return true;
2730 }
2732 //----------------------------inline_unsafe_load_store----------------------------
2733 // This method serves a couple of different customers (depending on LoadStoreKind):
2734 //
2735 // LS_cmpxchg:
2736 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2737 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2738 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2739 //
2740 // LS_xadd:
2741 // public int getAndAddInt( Object o, long offset, int delta)
2742 // public long getAndAddLong(Object o, long offset, long delta)
2743 //
2744 // LS_xchg:
2745 // int getAndSet(Object o, long offset, int newValue)
2746 // long getAndSet(Object o, long offset, long newValue)
2747 // Object getAndSet(Object o, long offset, Object newValue)
2748 //
2749 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2750 // This basic scheme here is the same as inline_unsafe_access, but
2751 // differs in enough details that combining them would make the code
2752 // overly confusing. (This is a true fact! I originally combined
2753 // them, but even I was confused by it!) As much code/comments as
2754 // possible are retained from inline_unsafe_access though to make
2755 // the correspondences clearer. - dl
2757 if (callee()->is_static()) return false; // caller must have the capability!
2759 #ifndef PRODUCT
2760 BasicType rtype;
2761 {
2762 ResourceMark rm;
2763 // Check the signatures.
2764 ciSignature* sig = callee()->signature();
2765 rtype = sig->return_type()->basic_type();
2766 if (kind == LS_xadd || kind == LS_xchg) {
2767 // Check the signatures.
2768 #ifdef ASSERT
2769 assert(rtype == type, "get and set must return the expected type");
2770 assert(sig->count() == 3, "get and set has 3 arguments");
2771 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2772 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2773 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2774 #endif // ASSERT
2775 } else if (kind == LS_cmpxchg) {
2776 // Check the signatures.
2777 #ifdef ASSERT
2778 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2779 assert(sig->count() == 4, "CAS has 4 arguments");
2780 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2781 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2782 #endif // ASSERT
2783 } else {
2784 ShouldNotReachHere();
2785 }
2786 }
2787 #endif //PRODUCT
2789 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2791 // Get arguments:
2792 Node* receiver = NULL;
2793 Node* base = NULL;
2794 Node* offset = NULL;
2795 Node* oldval = NULL;
2796 Node* newval = NULL;
2797 if (kind == LS_cmpxchg) {
2798 const bool two_slot_type = type2size[type] == 2;
2799 receiver = argument(0); // type: oop
2800 base = argument(1); // type: oop
2801 offset = argument(2); // type: long
2802 oldval = argument(4); // type: oop, int, or long
2803 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2804 } else if (kind == LS_xadd || kind == LS_xchg){
2805 receiver = argument(0); // type: oop
2806 base = argument(1); // type: oop
2807 offset = argument(2); // type: long
2808 oldval = NULL;
2809 newval = argument(4); // type: oop, int, or long
2810 }
2812 // Null check receiver.
2813 receiver = null_check(receiver);
2814 if (stopped()) {
2815 return true;
2816 }
2818 // Build field offset expression.
2819 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2820 // to be plain byte offsets, which are also the same as those accepted
2821 // by oopDesc::field_base.
2822 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2823 // 32-bit machines ignore the high half of long offsets
2824 offset = ConvL2X(offset);
2825 Node* adr = make_unsafe_address(base, offset);
2826 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2828 // For CAS, unlike inline_unsafe_access, there seems no point in
2829 // trying to refine types. Just use the coarse types here.
2830 const Type *value_type = Type::get_const_basic_type(type);
2831 Compile::AliasType* alias_type = C->alias_type(adr_type);
2832 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2834 if (kind == LS_xchg && type == T_OBJECT) {
2835 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2836 if (tjp != NULL) {
2837 value_type = tjp;
2838 }
2839 }
2841 int alias_idx = C->get_alias_index(adr_type);
2843 // Memory-model-wise, a LoadStore acts like a little synchronized
2844 // block, so needs barriers on each side. These don't translate
2845 // into actual barriers on most machines, but we still need rest of
2846 // compiler to respect ordering.
2848 insert_mem_bar(Op_MemBarRelease);
2849 insert_mem_bar(Op_MemBarCPUOrder);
2851 // 4984716: MemBars must be inserted before this
2852 // memory node in order to avoid a false
2853 // dependency which will confuse the scheduler.
2854 Node *mem = memory(alias_idx);
2856 // For now, we handle only those cases that actually exist: ints,
2857 // longs, and Object. Adding others should be straightforward.
2858 Node* load_store;
2859 switch(type) {
2860 case T_INT:
2861 if (kind == LS_xadd) {
2862 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2863 } else if (kind == LS_xchg) {
2864 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2865 } else if (kind == LS_cmpxchg) {
2866 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2867 } else {
2868 ShouldNotReachHere();
2869 }
2870 break;
2871 case T_LONG:
2872 if (kind == LS_xadd) {
2873 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2874 } else if (kind == LS_xchg) {
2875 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2876 } else if (kind == LS_cmpxchg) {
2877 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2878 } else {
2879 ShouldNotReachHere();
2880 }
2881 break;
2882 case T_OBJECT:
2883 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2884 // could be delayed during Parse (for example, in adjust_map_after_if()).
2885 // Execute transformation here to avoid barrier generation in such case.
2886 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2887 newval = _gvn.makecon(TypePtr::NULL_PTR);
2889 // Reference stores need a store barrier.
2890 if (kind == LS_xchg) {
2891 // If pre-barrier must execute before the oop store, old value will require do_load here.
2892 if (!can_move_pre_barrier()) {
2893 pre_barrier(true /* do_load*/,
2894 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2895 NULL /* pre_val*/,
2896 T_OBJECT);
2897 } // Else move pre_barrier to use load_store value, see below.
2898 } else if (kind == LS_cmpxchg) {
2899 // Same as for newval above:
2900 if (_gvn.type(oldval) == TypePtr::NULL_PTR) {
2901 oldval = _gvn.makecon(TypePtr::NULL_PTR);
2902 }
2903 // The only known value which might get overwritten is oldval.
2904 pre_barrier(false /* do_load */,
2905 control(), NULL, NULL, max_juint, NULL, NULL,
2906 oldval /* pre_val */,
2907 T_OBJECT);
2908 } else {
2909 ShouldNotReachHere();
2910 }
2912 #ifdef _LP64
2913 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2914 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2915 if (kind == LS_xchg) {
2916 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2917 newval_enc, adr_type, value_type->make_narrowoop()));
2918 } else {
2919 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2920 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2921 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2922 newval_enc, oldval_enc));
2923 }
2924 } else
2925 #endif
2926 {
2927 if (kind == LS_xchg) {
2928 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2929 } else {
2930 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2931 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2932 }
2933 }
2934 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2935 break;
2936 default:
2937 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2938 break;
2939 }
2941 // SCMemProjNodes represent the memory state of a LoadStore. Their
2942 // main role is to prevent LoadStore nodes from being optimized away
2943 // when their results aren't used.
2944 Node* proj = _gvn.transform(new (C) SCMemProjNode(load_store));
2945 set_memory(proj, alias_idx);
2947 if (type == T_OBJECT && kind == LS_xchg) {
2948 #ifdef _LP64
2949 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2950 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2951 }
2952 #endif
2953 if (can_move_pre_barrier()) {
2954 // Don't need to load pre_val. The old value is returned by load_store.
2955 // The pre_barrier can execute after the xchg as long as no safepoint
2956 // gets inserted between them.
2957 pre_barrier(false /* do_load */,
2958 control(), NULL, NULL, max_juint, NULL, NULL,
2959 load_store /* pre_val */,
2960 T_OBJECT);
2961 }
2962 }
2964 // Add the trailing membar surrounding the access
2965 insert_mem_bar(Op_MemBarCPUOrder);
2966 insert_mem_bar(Op_MemBarAcquire);
2968 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2969 set_result(load_store);
2970 return true;
2971 }
2973 //----------------------------inline_unsafe_ordered_store----------------------
2974 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2975 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2976 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2977 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2978 // This is another variant of inline_unsafe_access, differing in
2979 // that it always issues store-store ("release") barrier and ensures
2980 // store-atomicity (which only matters for "long").
2982 if (callee()->is_static()) return false; // caller must have the capability!
2984 #ifndef PRODUCT
2985 {
2986 ResourceMark rm;
2987 // Check the signatures.
2988 ciSignature* sig = callee()->signature();
2989 #ifdef ASSERT
2990 BasicType rtype = sig->return_type()->basic_type();
2991 assert(rtype == T_VOID, "must return void");
2992 assert(sig->count() == 3, "has 3 arguments");
2993 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2994 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2995 #endif // ASSERT
2996 }
2997 #endif //PRODUCT
2999 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3001 // Get arguments:
3002 Node* receiver = argument(0); // type: oop
3003 Node* base = argument(1); // type: oop
3004 Node* offset = argument(2); // type: long
3005 Node* val = argument(4); // type: oop, int, or long
3007 // Null check receiver.
3008 receiver = null_check(receiver);
3009 if (stopped()) {
3010 return true;
3011 }
3013 // Build field offset expression.
3014 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3015 // 32-bit machines ignore the high half of long offsets
3016 offset = ConvL2X(offset);
3017 Node* adr = make_unsafe_address(base, offset);
3018 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3019 const Type *value_type = Type::get_const_basic_type(type);
3020 Compile::AliasType* alias_type = C->alias_type(adr_type);
3022 insert_mem_bar(Op_MemBarRelease);
3023 insert_mem_bar(Op_MemBarCPUOrder);
3024 // Ensure that the store is atomic for longs:
3025 const bool require_atomic_access = true;
3026 Node* store;
3027 if (type == T_OBJECT) // reference stores need a store barrier.
3028 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
3029 else {
3030 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
3031 }
3032 insert_mem_bar(Op_MemBarCPUOrder);
3033 return true;
3034 }
3036 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
3037 // Regardless of form, don't allow previous ld/st to move down,
3038 // then issue acquire, release, or volatile mem_bar.
3039 insert_mem_bar(Op_MemBarCPUOrder);
3040 switch(id) {
3041 case vmIntrinsics::_loadFence:
3042 insert_mem_bar(Op_MemBarAcquire);
3043 return true;
3044 case vmIntrinsics::_storeFence:
3045 insert_mem_bar(Op_MemBarRelease);
3046 return true;
3047 case vmIntrinsics::_fullFence:
3048 insert_mem_bar(Op_MemBarVolatile);
3049 return true;
3050 default:
3051 fatal_unexpected_iid(id);
3052 return false;
3053 }
3054 }
3056 bool LibraryCallKit::klass_needs_init_guard(Node* kls) {
3057 if (!kls->is_Con()) {
3058 return true;
3059 }
3060 const TypeKlassPtr* klsptr = kls->bottom_type()->isa_klassptr();
3061 if (klsptr == NULL) {
3062 return true;
3063 }
3064 ciInstanceKlass* ik = klsptr->klass()->as_instance_klass();
3065 // don't need a guard for a klass that is already initialized
3066 return !ik->is_initialized();
3067 }
3069 //----------------------------inline_unsafe_allocate---------------------------
3070 // public native Object sun.misc.Unsafe.allocateInstance(Class<?> cls);
3071 bool LibraryCallKit::inline_unsafe_allocate() {
3072 if (callee()->is_static()) return false; // caller must have the capability!
3074 null_check_receiver(); // null-check, then ignore
3075 Node* cls = null_check(argument(1));
3076 if (stopped()) return true;
3078 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3079 kls = null_check(kls);
3080 if (stopped()) return true; // argument was like int.class
3082 Node* test = NULL;
3083 if (LibraryCallKit::klass_needs_init_guard(kls)) {
3084 // Note: The argument might still be an illegal value like
3085 // Serializable.class or Object[].class. The runtime will handle it.
3086 // But we must make an explicit check for initialization.
3087 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3088 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3089 // can generate code to load it as unsigned byte.
3090 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3091 Node* bits = intcon(InstanceKlass::fully_initialized);
3092 test = _gvn.transform(new (C) SubINode(inst, bits));
3093 // The 'test' is non-zero if we need to take a slow path.
3094 }
3096 Node* obj = new_instance(kls, test);
3097 set_result(obj);
3098 return true;
3099 }
3101 #ifdef TRACE_HAVE_INTRINSICS
3102 /*
3103 * oop -> myklass
3104 * myklass->trace_id |= USED
3105 * return myklass->trace_id & ~0x3
3106 */
3107 bool LibraryCallKit::inline_native_classID() {
3108 null_check_receiver(); // null-check, then ignore
3109 Node* cls = null_check(argument(1), T_OBJECT);
3110 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
3111 kls = null_check(kls, T_OBJECT);
3112 ByteSize offset = TRACE_ID_OFFSET;
3113 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3114 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3115 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3116 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3117 Node* clsused = longcon(0x01l); // set the class bit
3118 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3120 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3121 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3122 set_result(andl);
3123 return true;
3124 }
3126 bool LibraryCallKit::inline_native_threadID() {
3127 Node* tls_ptr = NULL;
3128 Node* cur_thr = generate_current_thread(tls_ptr);
3129 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3130 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3131 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3133 Node* threadid = NULL;
3134 size_t thread_id_size = OSThread::thread_id_size();
3135 if (thread_id_size == (size_t) BytesPerLong) {
3136 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3137 } else if (thread_id_size == (size_t) BytesPerInt) {
3138 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3139 } else {
3140 ShouldNotReachHere();
3141 }
3142 set_result(threadid);
3143 return true;
3144 }
3145 #endif
3147 //------------------------inline_native_time_funcs--------------
3148 // inline code for System.currentTimeMillis() and System.nanoTime()
3149 // these have the same type and signature
3150 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3151 const TypeFunc* tf = OptoRuntime::void_long_Type();
3152 const TypePtr* no_memory_effects = NULL;
3153 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3154 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3155 #ifdef ASSERT
3156 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
3157 assert(value_top == top(), "second value must be top");
3158 #endif
3159 set_result(value);
3160 return true;
3161 }
3163 //------------------------inline_native_currentThread------------------
3164 bool LibraryCallKit::inline_native_currentThread() {
3165 Node* junk = NULL;
3166 set_result(generate_current_thread(junk));
3167 return true;
3168 }
3170 //------------------------inline_native_isInterrupted------------------
3171 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
3172 bool LibraryCallKit::inline_native_isInterrupted() {
3173 // Add a fast path to t.isInterrupted(clear_int):
3174 // (t == Thread.current() &&
3175 // (!TLS._osthread._interrupted || WINDOWS_ONLY(false) NOT_WINDOWS(!clear_int)))
3176 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3177 // So, in the common case that the interrupt bit is false,
3178 // we avoid making a call into the VM. Even if the interrupt bit
3179 // is true, if the clear_int argument is false, we avoid the VM call.
3180 // However, if the receiver is not currentThread, we must call the VM,
3181 // because there must be some locking done around the operation.
3183 // We only go to the fast case code if we pass two guards.
3184 // Paths which do not pass are accumulated in the slow_region.
3186 enum {
3187 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
3188 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
3189 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
3190 PATH_LIMIT
3191 };
3193 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
3194 // out of the function.
3195 insert_mem_bar(Op_MemBarCPUOrder);
3197 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3198 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3200 RegionNode* slow_region = new (C) RegionNode(1);
3201 record_for_igvn(slow_region);
3203 // (a) Receiving thread must be the current thread.
3204 Node* rec_thr = argument(0);
3205 Node* tls_ptr = NULL;
3206 Node* cur_thr = generate_current_thread(tls_ptr);
3207 Node* cmp_thr = _gvn.transform(new (C) CmpPNode(cur_thr, rec_thr));
3208 Node* bol_thr = _gvn.transform(new (C) BoolNode(cmp_thr, BoolTest::ne));
3210 generate_slow_guard(bol_thr, slow_region);
3212 // (b) Interrupt bit on TLS must be false.
3213 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3214 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3215 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3217 // Set the control input on the field _interrupted read to prevent it floating up.
3218 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3219 Node* cmp_bit = _gvn.transform(new (C) CmpINode(int_bit, intcon(0)));
3220 Node* bol_bit = _gvn.transform(new (C) BoolNode(cmp_bit, BoolTest::ne));
3222 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3224 // First fast path: if (!TLS._interrupted) return false;
3225 Node* false_bit = _gvn.transform(new (C) IfFalseNode(iff_bit));
3226 result_rgn->init_req(no_int_result_path, false_bit);
3227 result_val->init_req(no_int_result_path, intcon(0));
3229 // drop through to next case
3230 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)));
3232 #ifndef TARGET_OS_FAMILY_windows
3233 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3234 Node* clr_arg = argument(1);
3235 Node* cmp_arg = _gvn.transform(new (C) CmpINode(clr_arg, intcon(0)));
3236 Node* bol_arg = _gvn.transform(new (C) BoolNode(cmp_arg, BoolTest::ne));
3237 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3239 // Second fast path: ... else if (!clear_int) return true;
3240 Node* false_arg = _gvn.transform(new (C) IfFalseNode(iff_arg));
3241 result_rgn->init_req(no_clear_result_path, false_arg);
3242 result_val->init_req(no_clear_result_path, intcon(1));
3244 // drop through to next case
3245 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)));
3246 #else
3247 // To return true on Windows you must read the _interrupted field
3248 // and check the the event state i.e. take the slow path.
3249 #endif // TARGET_OS_FAMILY_windows
3251 // (d) Otherwise, go to the slow path.
3252 slow_region->add_req(control());
3253 set_control( _gvn.transform(slow_region));
3255 if (stopped()) {
3256 // There is no slow path.
3257 result_rgn->init_req(slow_result_path, top());
3258 result_val->init_req(slow_result_path, top());
3259 } else {
3260 // non-virtual because it is a private non-static
3261 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3263 Node* slow_val = set_results_for_java_call(slow_call);
3264 // this->control() comes from set_results_for_java_call
3266 Node* fast_io = slow_call->in(TypeFunc::I_O);
3267 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3269 // These two phis are pre-filled with copies of of the fast IO and Memory
3270 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3271 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3273 result_rgn->init_req(slow_result_path, control());
3274 result_io ->init_req(slow_result_path, i_o());
3275 result_mem->init_req(slow_result_path, reset_memory());
3276 result_val->init_req(slow_result_path, slow_val);
3278 set_all_memory(_gvn.transform(result_mem));
3279 set_i_o( _gvn.transform(result_io));
3280 }
3282 C->set_has_split_ifs(true); // Has chance for split-if optimization
3283 set_result(result_rgn, result_val);
3284 return true;
3285 }
3287 //---------------------------load_mirror_from_klass----------------------------
3288 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3289 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3290 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3291 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3292 }
3294 //-----------------------load_klass_from_mirror_common-------------------------
3295 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3296 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3297 // and branch to the given path on the region.
3298 // If never_see_null, take an uncommon trap on null, so we can optimistically
3299 // compile for the non-null case.
3300 // If the region is NULL, force never_see_null = true.
3301 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3302 bool never_see_null,
3303 RegionNode* region,
3304 int null_path,
3305 int offset) {
3306 if (region == NULL) never_see_null = true;
3307 Node* p = basic_plus_adr(mirror, offset);
3308 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3309 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type));
3310 Node* null_ctl = top();
3311 kls = null_check_oop(kls, &null_ctl, never_see_null);
3312 if (region != NULL) {
3313 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3314 region->init_req(null_path, null_ctl);
3315 } else {
3316 assert(null_ctl == top(), "no loose ends");
3317 }
3318 return kls;
3319 }
3321 //--------------------(inline_native_Class_query helpers)---------------------
3322 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3323 // Fall through if (mods & mask) == bits, take the guard otherwise.
3324 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3325 // Branch around if the given klass has the given modifier bit set.
3326 // Like generate_guard, adds a new path onto the region.
3327 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3328 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3329 Node* mask = intcon(modifier_mask);
3330 Node* bits = intcon(modifier_bits);
3331 Node* mbit = _gvn.transform(new (C) AndINode(mods, mask));
3332 Node* cmp = _gvn.transform(new (C) CmpINode(mbit, bits));
3333 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
3334 return generate_fair_guard(bol, region);
3335 }
3336 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3337 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3338 }
3340 //-------------------------inline_native_Class_query-------------------
3341 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3342 const Type* return_type = TypeInt::BOOL;
3343 Node* prim_return_value = top(); // what happens if it's a primitive class?
3344 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3345 bool expect_prim = false; // most of these guys expect to work on refs
3347 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3349 Node* mirror = argument(0);
3350 Node* obj = top();
3352 switch (id) {
3353 case vmIntrinsics::_isInstance:
3354 // nothing is an instance of a primitive type
3355 prim_return_value = intcon(0);
3356 obj = argument(1);
3357 break;
3358 case vmIntrinsics::_getModifiers:
3359 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3360 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3361 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3362 break;
3363 case vmIntrinsics::_isInterface:
3364 prim_return_value = intcon(0);
3365 break;
3366 case vmIntrinsics::_isArray:
3367 prim_return_value = intcon(0);
3368 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3369 break;
3370 case vmIntrinsics::_isPrimitive:
3371 prim_return_value = intcon(1);
3372 expect_prim = true; // obviously
3373 break;
3374 case vmIntrinsics::_getSuperclass:
3375 prim_return_value = null();
3376 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3377 break;
3378 case vmIntrinsics::_getComponentType:
3379 prim_return_value = null();
3380 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3381 break;
3382 case vmIntrinsics::_getClassAccessFlags:
3383 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3384 return_type = TypeInt::INT; // not bool! 6297094
3385 break;
3386 default:
3387 fatal_unexpected_iid(id);
3388 break;
3389 }
3391 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3392 if (mirror_con == NULL) return false; // cannot happen?
3394 #ifndef PRODUCT
3395 if (C->print_intrinsics() || C->print_inlining()) {
3396 ciType* k = mirror_con->java_mirror_type();
3397 if (k) {
3398 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3399 k->print_name();
3400 tty->cr();
3401 }
3402 }
3403 #endif
3405 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3406 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3407 record_for_igvn(region);
3408 PhiNode* phi = new (C) PhiNode(region, return_type);
3410 // The mirror will never be null of Reflection.getClassAccessFlags, however
3411 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3412 // if it is. See bug 4774291.
3414 // For Reflection.getClassAccessFlags(), the null check occurs in
3415 // the wrong place; see inline_unsafe_access(), above, for a similar
3416 // situation.
3417 mirror = null_check(mirror);
3418 // If mirror or obj is dead, only null-path is taken.
3419 if (stopped()) return true;
3421 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3423 // Now load the mirror's klass metaobject, and null-check it.
3424 // Side-effects region with the control path if the klass is null.
3425 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3426 // If kls is null, we have a primitive mirror.
3427 phi->init_req(_prim_path, prim_return_value);
3428 if (stopped()) { set_result(region, phi); return true; }
3429 bool safe_for_replace = (region->in(_prim_path) == top());
3431 Node* p; // handy temp
3432 Node* null_ctl;
3434 // Now that we have the non-null klass, we can perform the real query.
3435 // For constant classes, the query will constant-fold in LoadNode::Value.
3436 Node* query_value = top();
3437 switch (id) {
3438 case vmIntrinsics::_isInstance:
3439 // nothing is an instance of a primitive type
3440 query_value = gen_instanceof(obj, kls, safe_for_replace);
3441 break;
3443 case vmIntrinsics::_getModifiers:
3444 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3445 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3446 break;
3448 case vmIntrinsics::_isInterface:
3449 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3450 if (generate_interface_guard(kls, region) != NULL)
3451 // A guard was added. If the guard is taken, it was an interface.
3452 phi->add_req(intcon(1));
3453 // If we fall through, it's a plain class.
3454 query_value = intcon(0);
3455 break;
3457 case vmIntrinsics::_isArray:
3458 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3459 if (generate_array_guard(kls, region) != NULL)
3460 // A guard was added. If the guard is taken, it was an array.
3461 phi->add_req(intcon(1));
3462 // If we fall through, it's a plain class.
3463 query_value = intcon(0);
3464 break;
3466 case vmIntrinsics::_isPrimitive:
3467 query_value = intcon(0); // "normal" path produces false
3468 break;
3470 case vmIntrinsics::_getSuperclass:
3471 // The rules here are somewhat unfortunate, but we can still do better
3472 // with random logic than with a JNI call.
3473 // Interfaces store null or Object as _super, but must report null.
3474 // Arrays store an intermediate super as _super, but must report Object.
3475 // Other types can report the actual _super.
3476 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3477 if (generate_interface_guard(kls, region) != NULL)
3478 // A guard was added. If the guard is taken, it was an interface.
3479 phi->add_req(null());
3480 if (generate_array_guard(kls, region) != NULL)
3481 // A guard was added. If the guard is taken, it was an array.
3482 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3483 // If we fall through, it's a plain class. Get its _super.
3484 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3485 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL));
3486 null_ctl = top();
3487 kls = null_check_oop(kls, &null_ctl);
3488 if (null_ctl != top()) {
3489 // If the guard is taken, Object.superClass is null (both klass and mirror).
3490 region->add_req(null_ctl);
3491 phi ->add_req(null());
3492 }
3493 if (!stopped()) {
3494 query_value = load_mirror_from_klass(kls);
3495 }
3496 break;
3498 case vmIntrinsics::_getComponentType:
3499 if (generate_array_guard(kls, region) != NULL) {
3500 // Be sure to pin the oop load to the guard edge just created:
3501 Node* is_array_ctrl = region->in(region->req()-1);
3502 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3503 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3504 phi->add_req(cmo);
3505 }
3506 query_value = null(); // non-array case is null
3507 break;
3509 case vmIntrinsics::_getClassAccessFlags:
3510 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3511 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3512 break;
3514 default:
3515 fatal_unexpected_iid(id);
3516 break;
3517 }
3519 // Fall-through is the normal case of a query to a real class.
3520 phi->init_req(1, query_value);
3521 region->init_req(1, control());
3523 C->set_has_split_ifs(true); // Has chance for split-if optimization
3524 set_result(region, phi);
3525 return true;
3526 }
3528 //--------------------------inline_native_subtype_check------------------------
3529 // This intrinsic takes the JNI calls out of the heart of
3530 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3531 bool LibraryCallKit::inline_native_subtype_check() {
3532 // Pull both arguments off the stack.
3533 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3534 args[0] = argument(0);
3535 args[1] = argument(1);
3536 Node* klasses[2]; // corresponding Klasses: superk, subk
3537 klasses[0] = klasses[1] = top();
3539 enum {
3540 // A full decision tree on {superc is prim, subc is prim}:
3541 _prim_0_path = 1, // {P,N} => false
3542 // {P,P} & superc!=subc => false
3543 _prim_same_path, // {P,P} & superc==subc => true
3544 _prim_1_path, // {N,P} => false
3545 _ref_subtype_path, // {N,N} & subtype check wins => true
3546 _both_ref_path, // {N,N} & subtype check loses => false
3547 PATH_LIMIT
3548 };
3550 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3551 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3552 record_for_igvn(region);
3554 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3555 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3556 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3558 // First null-check both mirrors and load each mirror's klass metaobject.
3559 int which_arg;
3560 for (which_arg = 0; which_arg <= 1; which_arg++) {
3561 Node* arg = args[which_arg];
3562 arg = null_check(arg);
3563 if (stopped()) break;
3564 args[which_arg] = arg;
3566 Node* p = basic_plus_adr(arg, class_klass_offset);
3567 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3568 klasses[which_arg] = _gvn.transform(kls);
3569 }
3571 // Having loaded both klasses, test each for null.
3572 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3573 for (which_arg = 0; which_arg <= 1; which_arg++) {
3574 Node* kls = klasses[which_arg];
3575 Node* null_ctl = top();
3576 kls = null_check_oop(kls, &null_ctl, never_see_null);
3577 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3578 region->init_req(prim_path, null_ctl);
3579 if (stopped()) break;
3580 klasses[which_arg] = kls;
3581 }
3583 if (!stopped()) {
3584 // now we have two reference types, in klasses[0..1]
3585 Node* subk = klasses[1]; // the argument to isAssignableFrom
3586 Node* superk = klasses[0]; // the receiver
3587 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3588 // now we have a successful reference subtype check
3589 region->set_req(_ref_subtype_path, control());
3590 }
3592 // If both operands are primitive (both klasses null), then
3593 // we must return true when they are identical primitives.
3594 // It is convenient to test this after the first null klass check.
3595 set_control(region->in(_prim_0_path)); // go back to first null check
3596 if (!stopped()) {
3597 // Since superc is primitive, make a guard for the superc==subc case.
3598 Node* cmp_eq = _gvn.transform(new (C) CmpPNode(args[0], args[1]));
3599 Node* bol_eq = _gvn.transform(new (C) BoolNode(cmp_eq, BoolTest::eq));
3600 generate_guard(bol_eq, region, PROB_FAIR);
3601 if (region->req() == PATH_LIMIT+1) {
3602 // A guard was added. If the added guard is taken, superc==subc.
3603 region->swap_edges(PATH_LIMIT, _prim_same_path);
3604 region->del_req(PATH_LIMIT);
3605 }
3606 region->set_req(_prim_0_path, control()); // Not equal after all.
3607 }
3609 // these are the only paths that produce 'true':
3610 phi->set_req(_prim_same_path, intcon(1));
3611 phi->set_req(_ref_subtype_path, intcon(1));
3613 // pull together the cases:
3614 assert(region->req() == PATH_LIMIT, "sane region");
3615 for (uint i = 1; i < region->req(); i++) {
3616 Node* ctl = region->in(i);
3617 if (ctl == NULL || ctl == top()) {
3618 region->set_req(i, top());
3619 phi ->set_req(i, top());
3620 } else if (phi->in(i) == NULL) {
3621 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3622 }
3623 }
3625 set_control(_gvn.transform(region));
3626 set_result(_gvn.transform(phi));
3627 return true;
3628 }
3630 //---------------------generate_array_guard_common------------------------
3631 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3632 bool obj_array, bool not_array) {
3633 // If obj_array/non_array==false/false:
3634 // Branch around if the given klass is in fact an array (either obj or prim).
3635 // If obj_array/non_array==false/true:
3636 // Branch around if the given klass is not an array klass of any kind.
3637 // If obj_array/non_array==true/true:
3638 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3639 // If obj_array/non_array==true/false:
3640 // Branch around if the kls is an oop array (Object[] or subtype)
3641 //
3642 // Like generate_guard, adds a new path onto the region.
3643 jint layout_con = 0;
3644 Node* layout_val = get_layout_helper(kls, layout_con);
3645 if (layout_val == NULL) {
3646 bool query = (obj_array
3647 ? Klass::layout_helper_is_objArray(layout_con)
3648 : Klass::layout_helper_is_array(layout_con));
3649 if (query == not_array) {
3650 return NULL; // never a branch
3651 } else { // always a branch
3652 Node* always_branch = control();
3653 if (region != NULL)
3654 region->add_req(always_branch);
3655 set_control(top());
3656 return always_branch;
3657 }
3658 }
3659 // Now test the correct condition.
3660 jint nval = (obj_array
3661 ? ((jint)Klass::_lh_array_tag_type_value
3662 << Klass::_lh_array_tag_shift)
3663 : Klass::_lh_neutral_value);
3664 Node* cmp = _gvn.transform(new(C) CmpINode(layout_val, intcon(nval)));
3665 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3666 // invert the test if we are looking for a non-array
3667 if (not_array) btest = BoolTest(btest).negate();
3668 Node* bol = _gvn.transform(new(C) BoolNode(cmp, btest));
3669 return generate_fair_guard(bol, region);
3670 }
3673 //-----------------------inline_native_newArray--------------------------
3674 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3675 bool LibraryCallKit::inline_native_newArray() {
3676 Node* mirror = argument(0);
3677 Node* count_val = argument(1);
3679 mirror = null_check(mirror);
3680 // If mirror or obj is dead, only null-path is taken.
3681 if (stopped()) return true;
3683 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3684 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3685 PhiNode* result_val = new(C) PhiNode(result_reg,
3686 TypeInstPtr::NOTNULL);
3687 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3688 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3689 TypePtr::BOTTOM);
3691 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3692 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3693 result_reg, _slow_path);
3694 Node* normal_ctl = control();
3695 Node* no_array_ctl = result_reg->in(_slow_path);
3697 // Generate code for the slow case. We make a call to newArray().
3698 set_control(no_array_ctl);
3699 if (!stopped()) {
3700 // Either the input type is void.class, or else the
3701 // array klass has not yet been cached. Either the
3702 // ensuing call will throw an exception, or else it
3703 // will cache the array klass for next time.
3704 PreserveJVMState pjvms(this);
3705 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3706 Node* slow_result = set_results_for_java_call(slow_call);
3707 // this->control() comes from set_results_for_java_call
3708 result_reg->set_req(_slow_path, control());
3709 result_val->set_req(_slow_path, slow_result);
3710 result_io ->set_req(_slow_path, i_o());
3711 result_mem->set_req(_slow_path, reset_memory());
3712 }
3714 set_control(normal_ctl);
3715 if (!stopped()) {
3716 // Normal case: The array type has been cached in the java.lang.Class.
3717 // The following call works fine even if the array type is polymorphic.
3718 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3719 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3720 result_reg->init_req(_normal_path, control());
3721 result_val->init_req(_normal_path, obj);
3722 result_io ->init_req(_normal_path, i_o());
3723 result_mem->init_req(_normal_path, reset_memory());
3724 }
3726 // Return the combined state.
3727 set_i_o( _gvn.transform(result_io) );
3728 set_all_memory( _gvn.transform(result_mem));
3730 C->set_has_split_ifs(true); // Has chance for split-if optimization
3731 set_result(result_reg, result_val);
3732 return true;
3733 }
3735 //----------------------inline_native_getLength--------------------------
3736 // public static native int java.lang.reflect.Array.getLength(Object array);
3737 bool LibraryCallKit::inline_native_getLength() {
3738 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3740 Node* array = null_check(argument(0));
3741 // If array is dead, only null-path is taken.
3742 if (stopped()) return true;
3744 // Deoptimize if it is a non-array.
3745 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3747 if (non_array != NULL) {
3748 PreserveJVMState pjvms(this);
3749 set_control(non_array);
3750 uncommon_trap(Deoptimization::Reason_intrinsic,
3751 Deoptimization::Action_maybe_recompile);
3752 }
3754 // If control is dead, only non-array-path is taken.
3755 if (stopped()) return true;
3757 // The works fine even if the array type is polymorphic.
3758 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3759 Node* result = load_array_length(array);
3761 C->set_has_split_ifs(true); // Has chance for split-if optimization
3762 set_result(result);
3763 return true;
3764 }
3766 //------------------------inline_array_copyOf----------------------------
3767 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3768 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3769 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3770 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3772 // Get the arguments.
3773 Node* original = argument(0);
3774 Node* start = is_copyOfRange? argument(1): intcon(0);
3775 Node* end = is_copyOfRange? argument(2): argument(1);
3776 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3778 Node* newcopy;
3780 // Set the original stack and the reexecute bit for the interpreter to reexecute
3781 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3782 { PreserveReexecuteState preexecs(this);
3783 jvms()->set_should_reexecute(true);
3785 array_type_mirror = null_check(array_type_mirror);
3786 original = null_check(original);
3788 // Check if a null path was taken unconditionally.
3789 if (stopped()) return true;
3791 Node* orig_length = load_array_length(original);
3793 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3794 klass_node = null_check(klass_node);
3796 RegionNode* bailout = new (C) RegionNode(1);
3797 record_for_igvn(bailout);
3799 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3800 // Bail out if that is so.
3801 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3802 if (not_objArray != NULL) {
3803 // Improve the klass node's type from the new optimistic assumption:
3804 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3805 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3806 Node* cast = new (C) CastPPNode(klass_node, akls);
3807 cast->init_req(0, control());
3808 klass_node = _gvn.transform(cast);
3809 }
3811 // Bail out if either start or end is negative.
3812 generate_negative_guard(start, bailout, &start);
3813 generate_negative_guard(end, bailout, &end);
3815 Node* length = end;
3816 if (_gvn.type(start) != TypeInt::ZERO) {
3817 length = _gvn.transform(new (C) SubINode(end, start));
3818 }
3820 // Bail out if length is negative.
3821 // Without this the new_array would throw
3822 // NegativeArraySizeException but IllegalArgumentException is what
3823 // should be thrown
3824 generate_negative_guard(length, bailout, &length);
3826 if (bailout->req() > 1) {
3827 PreserveJVMState pjvms(this);
3828 set_control(_gvn.transform(bailout));
3829 uncommon_trap(Deoptimization::Reason_intrinsic,
3830 Deoptimization::Action_maybe_recompile);
3831 }
3833 if (!stopped()) {
3834 // How many elements will we copy from the original?
3835 // The answer is MinI(orig_length - start, length).
3836 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3837 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3839 newcopy = new_array(klass_node, length, 0); // no argments to push
3841 // Generate a direct call to the right arraycopy function(s).
3842 // We know the copy is disjoint but we might not know if the
3843 // oop stores need checking.
3844 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3845 // This will fail a store-check if x contains any non-nulls.
3846 bool disjoint_bases = true;
3847 // if start > orig_length then the length of the copy may be
3848 // negative.
3849 bool length_never_negative = !is_copyOfRange;
3850 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3851 original, start, newcopy, intcon(0), moved,
3852 disjoint_bases, length_never_negative);
3853 }
3854 } // original reexecute is set back here
3856 C->set_has_split_ifs(true); // Has chance for split-if optimization
3857 if (!stopped()) {
3858 set_result(newcopy);
3859 }
3860 return true;
3861 }
3864 //----------------------generate_virtual_guard---------------------------
3865 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3866 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3867 RegionNode* slow_region) {
3868 ciMethod* method = callee();
3869 int vtable_index = method->vtable_index();
3870 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3871 err_msg_res("bad index %d", vtable_index));
3872 // Get the Method* out of the appropriate vtable entry.
3873 int entry_offset = (InstanceKlass::vtable_start_offset() +
3874 vtable_index*vtableEntry::size()) * wordSize +
3875 vtableEntry::method_offset_in_bytes();
3876 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3877 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3879 // Compare the target method with the expected method (e.g., Object.hashCode).
3880 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3882 Node* native_call = makecon(native_call_addr);
3883 Node* chk_native = _gvn.transform(new(C) CmpPNode(target_call, native_call));
3884 Node* test_native = _gvn.transform(new(C) BoolNode(chk_native, BoolTest::ne));
3886 return generate_slow_guard(test_native, slow_region);
3887 }
3889 //-----------------------generate_method_call----------------------------
3890 // Use generate_method_call to make a slow-call to the real
3891 // method if the fast path fails. An alternative would be to
3892 // use a stub like OptoRuntime::slow_arraycopy_Java.
3893 // This only works for expanding the current library call,
3894 // not another intrinsic. (E.g., don't use this for making an
3895 // arraycopy call inside of the copyOf intrinsic.)
3896 CallJavaNode*
3897 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3898 // When compiling the intrinsic method itself, do not use this technique.
3899 guarantee(callee() != C->method(), "cannot make slow-call to self");
3901 ciMethod* method = callee();
3902 // ensure the JVMS we have will be correct for this call
3903 guarantee(method_id == method->intrinsic_id(), "must match");
3905 const TypeFunc* tf = TypeFunc::make(method);
3906 CallJavaNode* slow_call;
3907 if (is_static) {
3908 assert(!is_virtual, "");
3909 slow_call = new(C) CallStaticJavaNode(C, tf,
3910 SharedRuntime::get_resolve_static_call_stub(),
3911 method, bci());
3912 } else if (is_virtual) {
3913 null_check_receiver();
3914 int vtable_index = Method::invalid_vtable_index;
3915 if (UseInlineCaches) {
3916 // Suppress the vtable call
3917 } else {
3918 // hashCode and clone are not a miranda methods,
3919 // so the vtable index is fixed.
3920 // No need to use the linkResolver to get it.
3921 vtable_index = method->vtable_index();
3922 assert(vtable_index >= 0 || vtable_index == Method::nonvirtual_vtable_index,
3923 err_msg_res("bad index %d", vtable_index));
3924 }
3925 slow_call = new(C) CallDynamicJavaNode(tf,
3926 SharedRuntime::get_resolve_virtual_call_stub(),
3927 method, vtable_index, bci());
3928 } else { // neither virtual nor static: opt_virtual
3929 null_check_receiver();
3930 slow_call = new(C) CallStaticJavaNode(C, tf,
3931 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3932 method, bci());
3933 slow_call->set_optimized_virtual(true);
3934 }
3935 set_arguments_for_java_call(slow_call);
3936 set_edges_for_java_call(slow_call);
3937 return slow_call;
3938 }
3941 //------------------------------inline_native_hashcode--------------------
3942 // Build special case code for calls to hashCode on an object.
3943 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3944 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3945 assert(!(is_virtual && is_static), "either virtual, special, or static");
3947 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3949 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3950 PhiNode* result_val = new(C) PhiNode(result_reg,
3951 TypeInt::INT);
3952 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3953 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3954 TypePtr::BOTTOM);
3955 Node* obj = NULL;
3956 if (!is_static) {
3957 // Check for hashing null object
3958 obj = null_check_receiver();
3959 if (stopped()) return true; // unconditionally null
3960 result_reg->init_req(_null_path, top());
3961 result_val->init_req(_null_path, top());
3962 } else {
3963 // Do a null check, and return zero if null.
3964 // System.identityHashCode(null) == 0
3965 obj = argument(0);
3966 Node* null_ctl = top();
3967 obj = null_check_oop(obj, &null_ctl);
3968 result_reg->init_req(_null_path, null_ctl);
3969 result_val->init_req(_null_path, _gvn.intcon(0));
3970 }
3972 // Unconditionally null? Then return right away.
3973 if (stopped()) {
3974 set_control( result_reg->in(_null_path));
3975 if (!stopped())
3976 set_result(result_val->in(_null_path));
3977 return true;
3978 }
3980 // After null check, get the object's klass.
3981 Node* obj_klass = load_object_klass(obj);
3983 // This call may be virtual (invokevirtual) or bound (invokespecial).
3984 // For each case we generate slightly different code.
3986 // We only go to the fast case code if we pass a number of guards. The
3987 // paths which do not pass are accumulated in the slow_region.
3988 RegionNode* slow_region = new (C) RegionNode(1);
3989 record_for_igvn(slow_region);
3991 // If this is a virtual call, we generate a funny guard. We pull out
3992 // the vtable entry corresponding to hashCode() from the target object.
3993 // If the target method which we are calling happens to be the native
3994 // Object hashCode() method, we pass the guard. We do not need this
3995 // guard for non-virtual calls -- the caller is known to be the native
3996 // Object hashCode().
3997 if (is_virtual) {
3998 generate_virtual_guard(obj_klass, slow_region);
3999 }
4001 // Get the header out of the object, use LoadMarkNode when available
4002 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4003 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
4005 // Test the header to see if it is unlocked.
4006 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4007 Node *lmasked_header = _gvn.transform(new (C) AndXNode(header, lock_mask));
4008 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4009 Node *chk_unlocked = _gvn.transform(new (C) CmpXNode( lmasked_header, unlocked_val));
4010 Node *test_unlocked = _gvn.transform(new (C) BoolNode( chk_unlocked, BoolTest::ne));
4012 generate_slow_guard(test_unlocked, slow_region);
4014 // Get the hash value and check to see that it has been properly assigned.
4015 // We depend on hash_mask being at most 32 bits and avoid the use of
4016 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4017 // vm: see markOop.hpp.
4018 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4019 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4020 Node *hshifted_header= _gvn.transform(new (C) URShiftXNode(header, hash_shift));
4021 // This hack lets the hash bits live anywhere in the mark object now, as long
4022 // as the shift drops the relevant bits into the low 32 bits. Note that
4023 // Java spec says that HashCode is an int so there's no point in capturing
4024 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4025 hshifted_header = ConvX2I(hshifted_header);
4026 Node *hash_val = _gvn.transform(new (C) AndINode(hshifted_header, hash_mask));
4028 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4029 Node *chk_assigned = _gvn.transform(new (C) CmpINode( hash_val, no_hash_val));
4030 Node *test_assigned = _gvn.transform(new (C) BoolNode( chk_assigned, BoolTest::eq));
4032 generate_slow_guard(test_assigned, slow_region);
4034 Node* init_mem = reset_memory();
4035 // fill in the rest of the null path:
4036 result_io ->init_req(_null_path, i_o());
4037 result_mem->init_req(_null_path, init_mem);
4039 result_val->init_req(_fast_path, hash_val);
4040 result_reg->init_req(_fast_path, control());
4041 result_io ->init_req(_fast_path, i_o());
4042 result_mem->init_req(_fast_path, init_mem);
4044 // Generate code for the slow case. We make a call to hashCode().
4045 set_control(_gvn.transform(slow_region));
4046 if (!stopped()) {
4047 // No need for PreserveJVMState, because we're using up the present state.
4048 set_all_memory(init_mem);
4049 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
4050 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4051 Node* slow_result = set_results_for_java_call(slow_call);
4052 // this->control() comes from set_results_for_java_call
4053 result_reg->init_req(_slow_path, control());
4054 result_val->init_req(_slow_path, slow_result);
4055 result_io ->set_req(_slow_path, i_o());
4056 result_mem ->set_req(_slow_path, reset_memory());
4057 }
4059 // Return the combined state.
4060 set_i_o( _gvn.transform(result_io) );
4061 set_all_memory( _gvn.transform(result_mem));
4063 set_result(result_reg, result_val);
4064 return true;
4065 }
4067 //---------------------------inline_native_getClass----------------------------
4068 // public final native Class<?> java.lang.Object.getClass();
4069 //
4070 // Build special case code for calls to getClass on an object.
4071 bool LibraryCallKit::inline_native_getClass() {
4072 Node* obj = null_check_receiver();
4073 if (stopped()) return true;
4074 set_result(load_mirror_from_klass(load_object_klass(obj)));
4075 return true;
4076 }
4078 //-----------------inline_native_Reflection_getCallerClass---------------------
4079 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
4080 //
4081 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4082 //
4083 // NOTE: This code must perform the same logic as JVM_GetCallerClass
4084 // in that it must skip particular security frames and checks for
4085 // caller sensitive methods.
4086 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4087 #ifndef PRODUCT
4088 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4089 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4090 }
4091 #endif
4093 if (!jvms()->has_method()) {
4094 #ifndef PRODUCT
4095 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4096 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4097 }
4098 #endif
4099 return false;
4100 }
4102 // Walk back up the JVM state to find the caller at the required
4103 // depth.
4104 JVMState* caller_jvms = jvms();
4106 // Cf. JVM_GetCallerClass
4107 // NOTE: Start the loop at depth 1 because the current JVM state does
4108 // not include the Reflection.getCallerClass() frame.
4109 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
4110 ciMethod* m = caller_jvms->method();
4111 switch (n) {
4112 case 0:
4113 fatal("current JVM state does not include the Reflection.getCallerClass frame");
4114 break;
4115 case 1:
4116 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
4117 if (!m->caller_sensitive()) {
4118 #ifndef PRODUCT
4119 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4120 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
4121 }
4122 #endif
4123 return false; // bail-out; let JVM_GetCallerClass do the work
4124 }
4125 break;
4126 default:
4127 if (!m->is_ignored_by_security_stack_walk()) {
4128 // We have reached the desired frame; return the holder class.
4129 // Acquire method holder as java.lang.Class and push as constant.
4130 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4131 ciInstance* caller_mirror = caller_klass->java_mirror();
4132 set_result(makecon(TypeInstPtr::make(caller_mirror)));
4134 #ifndef PRODUCT
4135 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4136 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());
4137 tty->print_cr(" JVM state at this point:");
4138 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4139 ciMethod* m = jvms()->of_depth(i)->method();
4140 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4141 }
4142 }
4143 #endif
4144 return true;
4145 }
4146 break;
4147 }
4148 }
4150 #ifndef PRODUCT
4151 if ((C->print_intrinsics() || C->print_inlining()) && Verbose) {
4152 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
4153 tty->print_cr(" JVM state at this point:");
4154 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
4155 ciMethod* m = jvms()->of_depth(i)->method();
4156 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
4157 }
4158 }
4159 #endif
4161 return false; // bail-out; let JVM_GetCallerClass do the work
4162 }
4164 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4165 Node* arg = argument(0);
4166 Node* result;
4168 switch (id) {
4169 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4170 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4171 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4172 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4174 case vmIntrinsics::_doubleToLongBits: {
4175 // two paths (plus control) merge in a wood
4176 RegionNode *r = new (C) RegionNode(3);
4177 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4179 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4180 // Build the boolean node
4181 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4183 // Branch either way.
4184 // NaN case is less traveled, which makes all the difference.
4185 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4186 Node *opt_isnan = _gvn.transform(ifisnan);
4187 assert( opt_isnan->is_If(), "Expect an IfNode");
4188 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4189 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4191 set_control(iftrue);
4193 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4194 Node *slow_result = longcon(nan_bits); // return NaN
4195 phi->init_req(1, _gvn.transform( slow_result ));
4196 r->init_req(1, iftrue);
4198 // Else fall through
4199 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4200 set_control(iffalse);
4202 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4203 r->init_req(2, iffalse);
4205 // Post merge
4206 set_control(_gvn.transform(r));
4207 record_for_igvn(r);
4209 C->set_has_split_ifs(true); // Has chance for split-if optimization
4210 result = phi;
4211 assert(result->bottom_type()->isa_long(), "must be");
4212 break;
4213 }
4215 case vmIntrinsics::_floatToIntBits: {
4216 // two paths (plus control) merge in a wood
4217 RegionNode *r = new (C) RegionNode(3);
4218 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4220 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4221 // Build the boolean node
4222 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4224 // Branch either way.
4225 // NaN case is less traveled, which makes all the difference.
4226 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4227 Node *opt_isnan = _gvn.transform(ifisnan);
4228 assert( opt_isnan->is_If(), "Expect an IfNode");
4229 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4230 Node *iftrue = _gvn.transform(new (C) IfTrueNode(opt_ifisnan));
4232 set_control(iftrue);
4234 static const jint nan_bits = 0x7fc00000;
4235 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4236 phi->init_req(1, _gvn.transform( slow_result ));
4237 r->init_req(1, iftrue);
4239 // Else fall through
4240 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4241 set_control(iffalse);
4243 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4244 r->init_req(2, iffalse);
4246 // Post merge
4247 set_control(_gvn.transform(r));
4248 record_for_igvn(r);
4250 C->set_has_split_ifs(true); // Has chance for split-if optimization
4251 result = phi;
4252 assert(result->bottom_type()->isa_int(), "must be");
4253 break;
4254 }
4256 default:
4257 fatal_unexpected_iid(id);
4258 break;
4259 }
4260 set_result(_gvn.transform(result));
4261 return true;
4262 }
4264 #ifdef _LP64
4265 #define XTOP ,top() /*additional argument*/
4266 #else //_LP64
4267 #define XTOP /*no additional argument*/
4268 #endif //_LP64
4270 //----------------------inline_unsafe_copyMemory-------------------------
4271 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4272 bool LibraryCallKit::inline_unsafe_copyMemory() {
4273 if (callee()->is_static()) return false; // caller must have the capability!
4274 null_check_receiver(); // null-check receiver
4275 if (stopped()) return true;
4277 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4279 Node* src_ptr = argument(1); // type: oop
4280 Node* src_off = ConvL2X(argument(2)); // type: long
4281 Node* dst_ptr = argument(4); // type: oop
4282 Node* dst_off = ConvL2X(argument(5)); // type: long
4283 Node* size = ConvL2X(argument(7)); // type: long
4285 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4286 "fieldOffset must be byte-scaled");
4288 Node* src = make_unsafe_address(src_ptr, src_off);
4289 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4291 // Conservatively insert a memory barrier on all memory slices.
4292 // Do not let writes of the copy source or destination float below the copy.
4293 insert_mem_bar(Op_MemBarCPUOrder);
4295 // Call it. Note that the length argument is not scaled.
4296 make_runtime_call(RC_LEAF|RC_NO_FP,
4297 OptoRuntime::fast_arraycopy_Type(),
4298 StubRoutines::unsafe_arraycopy(),
4299 "unsafe_arraycopy",
4300 TypeRawPtr::BOTTOM,
4301 src, dst, size XTOP);
4303 // Do not let reads of the copy destination float above the copy.
4304 insert_mem_bar(Op_MemBarCPUOrder);
4306 return true;
4307 }
4309 //------------------------clone_coping-----------------------------------
4310 // Helper function for inline_native_clone.
4311 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4312 assert(obj_size != NULL, "");
4313 Node* raw_obj = alloc_obj->in(1);
4314 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4316 AllocateNode* alloc = NULL;
4317 if (ReduceBulkZeroing) {
4318 // We will be completely responsible for initializing this object -
4319 // mark Initialize node as complete.
4320 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4321 // The object was just allocated - there should be no any stores!
4322 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4323 // Mark as complete_with_arraycopy so that on AllocateNode
4324 // expansion, we know this AllocateNode is initialized by an array
4325 // copy and a StoreStore barrier exists after the array copy.
4326 alloc->initialization()->set_complete_with_arraycopy();
4327 }
4329 // Copy the fastest available way.
4330 // TODO: generate fields copies for small objects instead.
4331 Node* src = obj;
4332 Node* dest = alloc_obj;
4333 Node* size = _gvn.transform(obj_size);
4335 // Exclude the header but include array length to copy by 8 bytes words.
4336 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4337 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4338 instanceOopDesc::base_offset_in_bytes();
4339 // base_off:
4340 // 8 - 32-bit VM
4341 // 12 - 64-bit VM, compressed klass
4342 // 16 - 64-bit VM, normal klass
4343 if (base_off % BytesPerLong != 0) {
4344 assert(UseCompressedClassPointers, "");
4345 if (is_array) {
4346 // Exclude length to copy by 8 bytes words.
4347 base_off += sizeof(int);
4348 } else {
4349 // Include klass to copy by 8 bytes words.
4350 base_off = instanceOopDesc::klass_offset_in_bytes();
4351 }
4352 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4353 }
4354 src = basic_plus_adr(src, base_off);
4355 dest = basic_plus_adr(dest, base_off);
4357 // Compute the length also, if needed:
4358 Node* countx = size;
4359 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(base_off)));
4360 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4362 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4363 bool disjoint_bases = true;
4364 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4365 src, NULL, dest, NULL, countx,
4366 /*dest_uninitialized*/true);
4368 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4369 if (card_mark) {
4370 assert(!is_array, "");
4371 // Put in store barrier for any and all oops we are sticking
4372 // into this object. (We could avoid this if we could prove
4373 // that the object type contains no oop fields at all.)
4374 Node* no_particular_value = NULL;
4375 Node* no_particular_field = NULL;
4376 int raw_adr_idx = Compile::AliasIdxRaw;
4377 post_barrier(control(),
4378 memory(raw_adr_type),
4379 alloc_obj,
4380 no_particular_field,
4381 raw_adr_idx,
4382 no_particular_value,
4383 T_OBJECT,
4384 false);
4385 }
4387 // Do not let reads from the cloned object float above the arraycopy.
4388 if (alloc != NULL) {
4389 // Do not let stores that initialize this object be reordered with
4390 // a subsequent store that would make this object accessible by
4391 // other threads.
4392 // Record what AllocateNode this StoreStore protects so that
4393 // escape analysis can go from the MemBarStoreStoreNode to the
4394 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4395 // based on the escape status of the AllocateNode.
4396 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4397 } else {
4398 insert_mem_bar(Op_MemBarCPUOrder);
4399 }
4400 }
4402 //------------------------inline_native_clone----------------------------
4403 // protected native Object java.lang.Object.clone();
4404 //
4405 // Here are the simple edge cases:
4406 // null receiver => normal trap
4407 // virtual and clone was overridden => slow path to out-of-line clone
4408 // not cloneable or finalizer => slow path to out-of-line Object.clone
4409 //
4410 // The general case has two steps, allocation and copying.
4411 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4412 //
4413 // Copying also has two cases, oop arrays and everything else.
4414 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4415 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4416 //
4417 // These steps fold up nicely if and when the cloned object's klass
4418 // can be sharply typed as an object array, a type array, or an instance.
4419 //
4420 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4421 PhiNode* result_val;
4423 // Set the reexecute bit for the interpreter to reexecute
4424 // the bytecode that invokes Object.clone if deoptimization happens.
4425 { PreserveReexecuteState preexecs(this);
4426 jvms()->set_should_reexecute(true);
4428 Node* obj = null_check_receiver();
4429 if (stopped()) return true;
4431 Node* obj_klass = load_object_klass(obj);
4432 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4433 const TypeOopPtr* toop = ((tklass != NULL)
4434 ? tklass->as_instance_type()
4435 : TypeInstPtr::NOTNULL);
4437 // Conservatively insert a memory barrier on all memory slices.
4438 // Do not let writes into the original float below the clone.
4439 insert_mem_bar(Op_MemBarCPUOrder);
4441 // paths into result_reg:
4442 enum {
4443 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4444 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4445 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4446 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4447 PATH_LIMIT
4448 };
4449 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4450 result_val = new(C) PhiNode(result_reg,
4451 TypeInstPtr::NOTNULL);
4452 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4453 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4454 TypePtr::BOTTOM);
4455 record_for_igvn(result_reg);
4457 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4458 int raw_adr_idx = Compile::AliasIdxRaw;
4460 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4461 if (array_ctl != NULL) {
4462 // It's an array.
4463 PreserveJVMState pjvms(this);
4464 set_control(array_ctl);
4465 Node* obj_length = load_array_length(obj);
4466 Node* obj_size = NULL;
4467 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4469 if (!use_ReduceInitialCardMarks()) {
4470 // If it is an oop array, it requires very special treatment,
4471 // because card marking is required on each card of the array.
4472 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4473 if (is_obja != NULL) {
4474 PreserveJVMState pjvms2(this);
4475 set_control(is_obja);
4476 // Generate a direct call to the right arraycopy function(s).
4477 bool disjoint_bases = true;
4478 bool length_never_negative = true;
4479 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4480 obj, intcon(0), alloc_obj, intcon(0),
4481 obj_length,
4482 disjoint_bases, length_never_negative);
4483 result_reg->init_req(_objArray_path, control());
4484 result_val->init_req(_objArray_path, alloc_obj);
4485 result_i_o ->set_req(_objArray_path, i_o());
4486 result_mem ->set_req(_objArray_path, reset_memory());
4487 }
4488 }
4489 // Otherwise, there are no card marks to worry about.
4490 // (We can dispense with card marks if we know the allocation
4491 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4492 // causes the non-eden paths to take compensating steps to
4493 // simulate a fresh allocation, so that no further
4494 // card marks are required in compiled code to initialize
4495 // the object.)
4497 if (!stopped()) {
4498 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4500 // Present the results of the copy.
4501 result_reg->init_req(_array_path, control());
4502 result_val->init_req(_array_path, alloc_obj);
4503 result_i_o ->set_req(_array_path, i_o());
4504 result_mem ->set_req(_array_path, reset_memory());
4505 }
4506 }
4508 // We only go to the instance fast case code if we pass a number of guards.
4509 // The paths which do not pass are accumulated in the slow_region.
4510 RegionNode* slow_region = new (C) RegionNode(1);
4511 record_for_igvn(slow_region);
4512 if (!stopped()) {
4513 // It's an instance (we did array above). Make the slow-path tests.
4514 // If this is a virtual call, we generate a funny guard. We grab
4515 // the vtable entry corresponding to clone() from the target object.
4516 // If the target method which we are calling happens to be the
4517 // Object clone() method, we pass the guard. We do not need this
4518 // guard for non-virtual calls; the caller is known to be the native
4519 // Object clone().
4520 if (is_virtual) {
4521 generate_virtual_guard(obj_klass, slow_region);
4522 }
4524 // The object must be cloneable and must not have a finalizer.
4525 // Both of these conditions may be checked in a single test.
4526 // We could optimize the cloneable test further, but we don't care.
4527 generate_access_flags_guard(obj_klass,
4528 // Test both conditions:
4529 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4530 // Must be cloneable but not finalizer:
4531 JVM_ACC_IS_CLONEABLE,
4532 slow_region);
4533 }
4535 if (!stopped()) {
4536 // It's an instance, and it passed the slow-path tests.
4537 PreserveJVMState pjvms(this);
4538 Node* obj_size = NULL;
4539 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4541 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4543 // Present the results of the slow call.
4544 result_reg->init_req(_instance_path, control());
4545 result_val->init_req(_instance_path, alloc_obj);
4546 result_i_o ->set_req(_instance_path, i_o());
4547 result_mem ->set_req(_instance_path, reset_memory());
4548 }
4550 // Generate code for the slow case. We make a call to clone().
4551 set_control(_gvn.transform(slow_region));
4552 if (!stopped()) {
4553 PreserveJVMState pjvms(this);
4554 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4555 Node* slow_result = set_results_for_java_call(slow_call);
4556 // this->control() comes from set_results_for_java_call
4557 result_reg->init_req(_slow_path, control());
4558 result_val->init_req(_slow_path, slow_result);
4559 result_i_o ->set_req(_slow_path, i_o());
4560 result_mem ->set_req(_slow_path, reset_memory());
4561 }
4563 // Return the combined state.
4564 set_control( _gvn.transform(result_reg));
4565 set_i_o( _gvn.transform(result_i_o));
4566 set_all_memory( _gvn.transform(result_mem));
4567 } // original reexecute is set back here
4569 set_result(_gvn.transform(result_val));
4570 return true;
4571 }
4573 //------------------------------basictype2arraycopy----------------------------
4574 address LibraryCallKit::basictype2arraycopy(BasicType t,
4575 Node* src_offset,
4576 Node* dest_offset,
4577 bool disjoint_bases,
4578 const char* &name,
4579 bool dest_uninitialized) {
4580 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4581 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4583 bool aligned = false;
4584 bool disjoint = disjoint_bases;
4586 // if the offsets are the same, we can treat the memory regions as
4587 // disjoint, because either the memory regions are in different arrays,
4588 // or they are identical (which we can treat as disjoint.) We can also
4589 // treat a copy with a destination index less that the source index
4590 // as disjoint since a low->high copy will work correctly in this case.
4591 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4592 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4593 // both indices are constants
4594 int s_offs = src_offset_inttype->get_con();
4595 int d_offs = dest_offset_inttype->get_con();
4596 int element_size = type2aelembytes(t);
4597 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4598 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4599 if (s_offs >= d_offs) disjoint = true;
4600 } else if (src_offset == dest_offset && src_offset != NULL) {
4601 // This can occur if the offsets are identical non-constants.
4602 disjoint = true;
4603 }
4605 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4606 }
4609 //------------------------------inline_arraycopy-----------------------
4610 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4611 // Object dest, int destPos,
4612 // int length);
4613 bool LibraryCallKit::inline_arraycopy() {
4614 // Get the arguments.
4615 Node* src = argument(0); // type: oop
4616 Node* src_offset = argument(1); // type: int
4617 Node* dest = argument(2); // type: oop
4618 Node* dest_offset = argument(3); // type: int
4619 Node* length = argument(4); // type: int
4621 // Compile time checks. If any of these checks cannot be verified at compile time,
4622 // we do not make a fast path for this call. Instead, we let the call remain as it
4623 // is. The checks we choose to mandate at compile time are:
4624 //
4625 // (1) src and dest are arrays.
4626 const Type* src_type = src->Value(&_gvn);
4627 const Type* dest_type = dest->Value(&_gvn);
4628 const TypeAryPtr* top_src = src_type->isa_aryptr();
4629 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4631 // Do we have the type of src?
4632 bool has_src = (top_src != NULL && top_src->klass() != NULL);
4633 // Do we have the type of dest?
4634 bool has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4635 // Is the type for src from speculation?
4636 bool src_spec = false;
4637 // Is the type for dest from speculation?
4638 bool dest_spec = false;
4640 if (!has_src || !has_dest) {
4641 // We don't have sufficient type information, let's see if
4642 // speculative types can help. We need to have types for both src
4643 // and dest so that it pays off.
4645 // Do we already have or could we have type information for src
4646 bool could_have_src = has_src;
4647 // Do we already have or could we have type information for dest
4648 bool could_have_dest = has_dest;
4650 ciKlass* src_k = NULL;
4651 if (!has_src) {
4652 src_k = src_type->speculative_type();
4653 if (src_k != NULL && src_k->is_array_klass()) {
4654 could_have_src = true;
4655 }
4656 }
4658 ciKlass* dest_k = NULL;
4659 if (!has_dest) {
4660 dest_k = dest_type->speculative_type();
4661 if (dest_k != NULL && dest_k->is_array_klass()) {
4662 could_have_dest = true;
4663 }
4664 }
4666 if (could_have_src && could_have_dest) {
4667 // This is going to pay off so emit the required guards
4668 if (!has_src) {
4669 src = maybe_cast_profiled_obj(src, src_k);
4670 src_type = _gvn.type(src);
4671 top_src = src_type->isa_aryptr();
4672 has_src = (top_src != NULL && top_src->klass() != NULL);
4673 src_spec = true;
4674 }
4675 if (!has_dest) {
4676 dest = maybe_cast_profiled_obj(dest, dest_k);
4677 dest_type = _gvn.type(dest);
4678 top_dest = dest_type->isa_aryptr();
4679 has_dest = (top_dest != NULL && top_dest->klass() != NULL);
4680 dest_spec = true;
4681 }
4682 }
4683 }
4685 if (!has_src || !has_dest) {
4686 // Conservatively insert a memory barrier on all memory slices.
4687 // Do not let writes into the source float below the arraycopy.
4688 insert_mem_bar(Op_MemBarCPUOrder);
4690 // Call StubRoutines::generic_arraycopy stub.
4691 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4692 src, src_offset, dest, dest_offset, length);
4694 // Do not let reads from the destination float above the arraycopy.
4695 // Since we cannot type the arrays, we don't know which slices
4696 // might be affected. We could restrict this barrier only to those
4697 // memory slices which pertain to array elements--but don't bother.
4698 if (!InsertMemBarAfterArraycopy)
4699 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4700 insert_mem_bar(Op_MemBarCPUOrder);
4701 return true;
4702 }
4704 // (2) src and dest arrays must have elements of the same BasicType
4705 // Figure out the size and type of the elements we will be copying.
4706 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4707 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4708 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4709 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4711 if (src_elem != dest_elem || dest_elem == T_VOID) {
4712 // The component types are not the same or are not recognized. Punt.
4713 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4714 generate_slow_arraycopy(TypePtr::BOTTOM,
4715 src, src_offset, dest, dest_offset, length,
4716 /*dest_uninitialized*/false);
4717 return true;
4718 }
4720 if (src_elem == T_OBJECT) {
4721 // If both arrays are object arrays then having the exact types
4722 // for both will remove the need for a subtype check at runtime
4723 // before the call and may make it possible to pick a faster copy
4724 // routine (without a subtype check on every element)
4725 // Do we have the exact type of src?
4726 bool could_have_src = src_spec;
4727 // Do we have the exact type of dest?
4728 bool could_have_dest = dest_spec;
4729 ciKlass* src_k = top_src->klass();
4730 ciKlass* dest_k = top_dest->klass();
4731 if (!src_spec) {
4732 src_k = src_type->speculative_type();
4733 if (src_k != NULL && src_k->is_array_klass()) {
4734 could_have_src = true;
4735 }
4736 }
4737 if (!dest_spec) {
4738 dest_k = dest_type->speculative_type();
4739 if (dest_k != NULL && dest_k->is_array_klass()) {
4740 could_have_dest = true;
4741 }
4742 }
4743 if (could_have_src && could_have_dest) {
4744 // If we can have both exact types, emit the missing guards
4745 if (could_have_src && !src_spec) {
4746 src = maybe_cast_profiled_obj(src, src_k);
4747 }
4748 if (could_have_dest && !dest_spec) {
4749 dest = maybe_cast_profiled_obj(dest, dest_k);
4750 }
4751 }
4752 }
4754 //---------------------------------------------------------------------------
4755 // We will make a fast path for this call to arraycopy.
4757 // We have the following tests left to perform:
4758 //
4759 // (3) src and dest must not be null.
4760 // (4) src_offset must not be negative.
4761 // (5) dest_offset must not be negative.
4762 // (6) length must not be negative.
4763 // (7) src_offset + length must not exceed length of src.
4764 // (8) dest_offset + length must not exceed length of dest.
4765 // (9) each element of an oop array must be assignable
4767 RegionNode* slow_region = new (C) RegionNode(1);
4768 record_for_igvn(slow_region);
4770 // (3) operands must not be null
4771 // We currently perform our null checks with the null_check routine.
4772 // This means that the null exceptions will be reported in the caller
4773 // rather than (correctly) reported inside of the native arraycopy call.
4774 // This should be corrected, given time. We do our null check with the
4775 // stack pointer restored.
4776 src = null_check(src, T_ARRAY);
4777 dest = null_check(dest, T_ARRAY);
4779 // (4) src_offset must not be negative.
4780 generate_negative_guard(src_offset, slow_region);
4782 // (5) dest_offset must not be negative.
4783 generate_negative_guard(dest_offset, slow_region);
4785 // (6) length must not be negative (moved to generate_arraycopy()).
4786 // generate_negative_guard(length, slow_region);
4788 // (7) src_offset + length must not exceed length of src.
4789 generate_limit_guard(src_offset, length,
4790 load_array_length(src),
4791 slow_region);
4793 // (8) dest_offset + length must not exceed length of dest.
4794 generate_limit_guard(dest_offset, length,
4795 load_array_length(dest),
4796 slow_region);
4798 // (9) each element of an oop array must be assignable
4799 // The generate_arraycopy subroutine checks this.
4801 // This is where the memory effects are placed:
4802 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4803 generate_arraycopy(adr_type, dest_elem,
4804 src, src_offset, dest, dest_offset, length,
4805 false, false, slow_region);
4807 return true;
4808 }
4810 //-----------------------------generate_arraycopy----------------------
4811 // Generate an optimized call to arraycopy.
4812 // Caller must guard against non-arrays.
4813 // Caller must determine a common array basic-type for both arrays.
4814 // Caller must validate offsets against array bounds.
4815 // The slow_region has already collected guard failure paths
4816 // (such as out of bounds length or non-conformable array types).
4817 // The generated code has this shape, in general:
4818 //
4819 // if (length == 0) return // via zero_path
4820 // slowval = -1
4821 // if (types unknown) {
4822 // slowval = call generic copy loop
4823 // if (slowval == 0) return // via checked_path
4824 // } else if (indexes in bounds) {
4825 // if ((is object array) && !(array type check)) {
4826 // slowval = call checked copy loop
4827 // if (slowval == 0) return // via checked_path
4828 // } else {
4829 // call bulk copy loop
4830 // return // via fast_path
4831 // }
4832 // }
4833 // // adjust params for remaining work:
4834 // if (slowval != -1) {
4835 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4836 // }
4837 // slow_region:
4838 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4839 // return // via slow_call_path
4840 //
4841 // This routine is used from several intrinsics: System.arraycopy,
4842 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4843 //
4844 void
4845 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4846 BasicType basic_elem_type,
4847 Node* src, Node* src_offset,
4848 Node* dest, Node* dest_offset,
4849 Node* copy_length,
4850 bool disjoint_bases,
4851 bool length_never_negative,
4852 RegionNode* slow_region) {
4854 if (slow_region == NULL) {
4855 slow_region = new(C) RegionNode(1);
4856 record_for_igvn(slow_region);
4857 }
4859 Node* original_dest = dest;
4860 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4861 bool dest_uninitialized = false;
4863 // See if this is the initialization of a newly-allocated array.
4864 // If so, we will take responsibility here for initializing it to zero.
4865 // (Note: Because tightly_coupled_allocation performs checks on the
4866 // out-edges of the dest, we need to avoid making derived pointers
4867 // from it until we have checked its uses.)
4868 if (ReduceBulkZeroing
4869 && !ZeroTLAB // pointless if already zeroed
4870 && basic_elem_type != T_CONFLICT // avoid corner case
4871 && !src->eqv_uncast(dest)
4872 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4873 != NULL)
4874 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4875 && alloc->maybe_set_complete(&_gvn)) {
4876 // "You break it, you buy it."
4877 InitializeNode* init = alloc->initialization();
4878 assert(init->is_complete(), "we just did this");
4879 init->set_complete_with_arraycopy();
4880 assert(dest->is_CheckCastPP(), "sanity");
4881 assert(dest->in(0)->in(0) == init, "dest pinned");
4882 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4883 // From this point on, every exit path is responsible for
4884 // initializing any non-copied parts of the object to zero.
4885 // Also, if this flag is set we make sure that arraycopy interacts properly
4886 // with G1, eliding pre-barriers. See CR 6627983.
4887 dest_uninitialized = true;
4888 } else {
4889 // No zeroing elimination here.
4890 alloc = NULL;
4891 //original_dest = dest;
4892 //dest_uninitialized = false;
4893 }
4895 // Results are placed here:
4896 enum { fast_path = 1, // normal void-returning assembly stub
4897 checked_path = 2, // special assembly stub with cleanup
4898 slow_call_path = 3, // something went wrong; call the VM
4899 zero_path = 4, // bypass when length of copy is zero
4900 bcopy_path = 5, // copy primitive array by 64-bit blocks
4901 PATH_LIMIT = 6
4902 };
4903 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4904 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4905 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4906 record_for_igvn(result_region);
4907 _gvn.set_type_bottom(result_i_o);
4908 _gvn.set_type_bottom(result_memory);
4909 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4911 // The slow_control path:
4912 Node* slow_control;
4913 Node* slow_i_o = i_o();
4914 Node* slow_mem = memory(adr_type);
4915 debug_only(slow_control = (Node*) badAddress);
4917 // Checked control path:
4918 Node* checked_control = top();
4919 Node* checked_mem = NULL;
4920 Node* checked_i_o = NULL;
4921 Node* checked_value = NULL;
4923 if (basic_elem_type == T_CONFLICT) {
4924 assert(!dest_uninitialized, "");
4925 Node* cv = generate_generic_arraycopy(adr_type,
4926 src, src_offset, dest, dest_offset,
4927 copy_length, dest_uninitialized);
4928 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4929 checked_control = control();
4930 checked_i_o = i_o();
4931 checked_mem = memory(adr_type);
4932 checked_value = cv;
4933 set_control(top()); // no fast path
4934 }
4936 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4937 if (not_pos != NULL) {
4938 PreserveJVMState pjvms(this);
4939 set_control(not_pos);
4941 // (6) length must not be negative.
4942 if (!length_never_negative) {
4943 generate_negative_guard(copy_length, slow_region);
4944 }
4946 // copy_length is 0.
4947 if (!stopped() && dest_uninitialized) {
4948 Node* dest_length = alloc->in(AllocateNode::ALength);
4949 if (copy_length->eqv_uncast(dest_length)
4950 || _gvn.find_int_con(dest_length, 1) <= 0) {
4951 // There is no zeroing to do. No need for a secondary raw memory barrier.
4952 } else {
4953 // Clear the whole thing since there are no source elements to copy.
4954 generate_clear_array(adr_type, dest, basic_elem_type,
4955 intcon(0), NULL,
4956 alloc->in(AllocateNode::AllocSize));
4957 // Use a secondary InitializeNode as raw memory barrier.
4958 // Currently it is needed only on this path since other
4959 // paths have stub or runtime calls as raw memory barriers.
4960 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4961 Compile::AliasIdxRaw,
4962 top())->as_Initialize();
4963 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4964 }
4965 }
4967 // Present the results of the fast call.
4968 result_region->init_req(zero_path, control());
4969 result_i_o ->init_req(zero_path, i_o());
4970 result_memory->init_req(zero_path, memory(adr_type));
4971 }
4973 if (!stopped() && dest_uninitialized) {
4974 // We have to initialize the *uncopied* part of the array to zero.
4975 // The copy destination is the slice dest[off..off+len]. The other slices
4976 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4977 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4978 Node* dest_length = alloc->in(AllocateNode::ALength);
4979 Node* dest_tail = _gvn.transform(new(C) AddINode(dest_offset,
4980 copy_length));
4982 // If there is a head section that needs zeroing, do it now.
4983 if (find_int_con(dest_offset, -1) != 0) {
4984 generate_clear_array(adr_type, dest, basic_elem_type,
4985 intcon(0), dest_offset,
4986 NULL);
4987 }
4989 // Next, perform a dynamic check on the tail length.
4990 // It is often zero, and we can win big if we prove this.
4991 // There are two wins: Avoid generating the ClearArray
4992 // with its attendant messy index arithmetic, and upgrade
4993 // the copy to a more hardware-friendly word size of 64 bits.
4994 Node* tail_ctl = NULL;
4995 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4996 Node* cmp_lt = _gvn.transform(new(C) CmpINode(dest_tail, dest_length));
4997 Node* bol_lt = _gvn.transform(new(C) BoolNode(cmp_lt, BoolTest::lt));
4998 tail_ctl = generate_slow_guard(bol_lt, NULL);
4999 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5000 }
5002 // At this point, let's assume there is no tail.
5003 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5004 // There is no tail. Try an upgrade to a 64-bit copy.
5005 bool didit = false;
5006 { PreserveJVMState pjvms(this);
5007 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5008 src, src_offset, dest, dest_offset,
5009 dest_size, dest_uninitialized);
5010 if (didit) {
5011 // Present the results of the block-copying fast call.
5012 result_region->init_req(bcopy_path, control());
5013 result_i_o ->init_req(bcopy_path, i_o());
5014 result_memory->init_req(bcopy_path, memory(adr_type));
5015 }
5016 }
5017 if (didit)
5018 set_control(top()); // no regular fast path
5019 }
5021 // Clear the tail, if any.
5022 if (tail_ctl != NULL) {
5023 Node* notail_ctl = stopped() ? NULL : control();
5024 set_control(tail_ctl);
5025 if (notail_ctl == NULL) {
5026 generate_clear_array(adr_type, dest, basic_elem_type,
5027 dest_tail, NULL,
5028 dest_size);
5029 } else {
5030 // Make a local merge.
5031 Node* done_ctl = new(C) RegionNode(3);
5032 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5033 done_ctl->init_req(1, notail_ctl);
5034 done_mem->init_req(1, memory(adr_type));
5035 generate_clear_array(adr_type, dest, basic_elem_type,
5036 dest_tail, NULL,
5037 dest_size);
5038 done_ctl->init_req(2, control());
5039 done_mem->init_req(2, memory(adr_type));
5040 set_control( _gvn.transform(done_ctl));
5041 set_memory( _gvn.transform(done_mem), adr_type );
5042 }
5043 }
5044 }
5046 BasicType copy_type = basic_elem_type;
5047 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5048 if (!stopped() && copy_type == T_OBJECT) {
5049 // If src and dest have compatible element types, we can copy bits.
5050 // Types S[] and D[] are compatible if D is a supertype of S.
5051 //
5052 // If they are not, we will use checked_oop_disjoint_arraycopy,
5053 // which performs a fast optimistic per-oop check, and backs off
5054 // further to JVM_ArrayCopy on the first per-oop check that fails.
5055 // (Actually, we don't move raw bits only; the GC requires card marks.)
5057 // Get the Klass* for both src and dest
5058 Node* src_klass = load_object_klass(src);
5059 Node* dest_klass = load_object_klass(dest);
5061 // Generate the subtype check.
5062 // This might fold up statically, or then again it might not.
5063 //
5064 // Non-static example: Copying List<String>.elements to a new String[].
5065 // The backing store for a List<String> is always an Object[],
5066 // but its elements are always type String, if the generic types
5067 // are correct at the source level.
5068 //
5069 // Test S[] against D[], not S against D, because (probably)
5070 // the secondary supertype cache is less busy for S[] than S.
5071 // This usually only matters when D is an interface.
5072 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5073 // Plug failing path into checked_oop_disjoint_arraycopy
5074 if (not_subtype_ctrl != top()) {
5075 PreserveJVMState pjvms(this);
5076 set_control(not_subtype_ctrl);
5077 // (At this point we can assume disjoint_bases, since types differ.)
5078 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5079 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5080 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5081 Node* dest_elem_klass = _gvn.transform(n1);
5082 Node* cv = generate_checkcast_arraycopy(adr_type,
5083 dest_elem_klass,
5084 src, src_offset, dest, dest_offset,
5085 ConvI2X(copy_length), dest_uninitialized);
5086 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5087 checked_control = control();
5088 checked_i_o = i_o();
5089 checked_mem = memory(adr_type);
5090 checked_value = cv;
5091 }
5092 // At this point we know we do not need type checks on oop stores.
5094 // Let's see if we need card marks:
5095 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5096 // If we do not need card marks, copy using the jint or jlong stub.
5097 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5098 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5099 "sizes agree");
5100 }
5101 }
5103 if (!stopped()) {
5104 // Generate the fast path, if possible.
5105 PreserveJVMState pjvms(this);
5106 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5107 src, src_offset, dest, dest_offset,
5108 ConvI2X(copy_length), dest_uninitialized);
5110 // Present the results of the fast call.
5111 result_region->init_req(fast_path, control());
5112 result_i_o ->init_req(fast_path, i_o());
5113 result_memory->init_req(fast_path, memory(adr_type));
5114 }
5116 // Here are all the slow paths up to this point, in one bundle:
5117 slow_control = top();
5118 if (slow_region != NULL)
5119 slow_control = _gvn.transform(slow_region);
5120 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
5122 set_control(checked_control);
5123 if (!stopped()) {
5124 // Clean up after the checked call.
5125 // The returned value is either 0 or -1^K,
5126 // where K = number of partially transferred array elements.
5127 Node* cmp = _gvn.transform(new(C) CmpINode(checked_value, intcon(0)));
5128 Node* bol = _gvn.transform(new(C) BoolNode(cmp, BoolTest::eq));
5129 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5131 // If it is 0, we are done, so transfer to the end.
5132 Node* checks_done = _gvn.transform(new(C) IfTrueNode(iff));
5133 result_region->init_req(checked_path, checks_done);
5134 result_i_o ->init_req(checked_path, checked_i_o);
5135 result_memory->init_req(checked_path, checked_mem);
5137 // If it is not zero, merge into the slow call.
5138 set_control( _gvn.transform(new(C) IfFalseNode(iff) ));
5139 RegionNode* slow_reg2 = new(C) RegionNode(3);
5140 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5141 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5142 record_for_igvn(slow_reg2);
5143 slow_reg2 ->init_req(1, slow_control);
5144 slow_i_o2 ->init_req(1, slow_i_o);
5145 slow_mem2 ->init_req(1, slow_mem);
5146 slow_reg2 ->init_req(2, control());
5147 slow_i_o2 ->init_req(2, checked_i_o);
5148 slow_mem2 ->init_req(2, checked_mem);
5150 slow_control = _gvn.transform(slow_reg2);
5151 slow_i_o = _gvn.transform(slow_i_o2);
5152 slow_mem = _gvn.transform(slow_mem2);
5154 if (alloc != NULL) {
5155 // We'll restart from the very beginning, after zeroing the whole thing.
5156 // This can cause double writes, but that's OK since dest is brand new.
5157 // So we ignore the low 31 bits of the value returned from the stub.
5158 } else {
5159 // We must continue the copy exactly where it failed, or else
5160 // another thread might see the wrong number of writes to dest.
5161 Node* checked_offset = _gvn.transform(new(C) XorINode(checked_value, intcon(-1)));
5162 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5163 slow_offset->init_req(1, intcon(0));
5164 slow_offset->init_req(2, checked_offset);
5165 slow_offset = _gvn.transform(slow_offset);
5167 // Adjust the arguments by the conditionally incoming offset.
5168 Node* src_off_plus = _gvn.transform(new(C) AddINode(src_offset, slow_offset));
5169 Node* dest_off_plus = _gvn.transform(new(C) AddINode(dest_offset, slow_offset));
5170 Node* length_minus = _gvn.transform(new(C) SubINode(copy_length, slow_offset));
5172 // Tweak the node variables to adjust the code produced below:
5173 src_offset = src_off_plus;
5174 dest_offset = dest_off_plus;
5175 copy_length = length_minus;
5176 }
5177 }
5179 set_control(slow_control);
5180 if (!stopped()) {
5181 // Generate the slow path, if needed.
5182 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5184 set_memory(slow_mem, adr_type);
5185 set_i_o(slow_i_o);
5187 if (dest_uninitialized) {
5188 generate_clear_array(adr_type, dest, basic_elem_type,
5189 intcon(0), NULL,
5190 alloc->in(AllocateNode::AllocSize));
5191 }
5193 generate_slow_arraycopy(adr_type,
5194 src, src_offset, dest, dest_offset,
5195 copy_length, /*dest_uninitialized*/false);
5197 result_region->init_req(slow_call_path, control());
5198 result_i_o ->init_req(slow_call_path, i_o());
5199 result_memory->init_req(slow_call_path, memory(adr_type));
5200 }
5202 // Remove unused edges.
5203 for (uint i = 1; i < result_region->req(); i++) {
5204 if (result_region->in(i) == NULL)
5205 result_region->init_req(i, top());
5206 }
5208 // Finished; return the combined state.
5209 set_control( _gvn.transform(result_region));
5210 set_i_o( _gvn.transform(result_i_o) );
5211 set_memory( _gvn.transform(result_memory), adr_type );
5213 // The memory edges above are precise in order to model effects around
5214 // array copies accurately to allow value numbering of field loads around
5215 // arraycopy. Such field loads, both before and after, are common in Java
5216 // collections and similar classes involving header/array data structures.
5217 //
5218 // But with low number of register or when some registers are used or killed
5219 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5220 // The next memory barrier is added to avoid it. If the arraycopy can be
5221 // optimized away (which it can, sometimes) then we can manually remove
5222 // the membar also.
5223 //
5224 // Do not let reads from the cloned object float above the arraycopy.
5225 if (alloc != NULL) {
5226 // Do not let stores that initialize this object be reordered with
5227 // a subsequent store that would make this object accessible by
5228 // other threads.
5229 // Record what AllocateNode this StoreStore protects so that
5230 // escape analysis can go from the MemBarStoreStoreNode to the
5231 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5232 // based on the escape status of the AllocateNode.
5233 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5234 } else if (InsertMemBarAfterArraycopy)
5235 insert_mem_bar(Op_MemBarCPUOrder);
5236 }
5239 // Helper function which determines if an arraycopy immediately follows
5240 // an allocation, with no intervening tests or other escapes for the object.
5241 AllocateArrayNode*
5242 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5243 RegionNode* slow_region) {
5244 if (stopped()) return NULL; // no fast path
5245 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5247 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5248 if (alloc == NULL) return NULL;
5250 Node* rawmem = memory(Compile::AliasIdxRaw);
5251 // Is the allocation's memory state untouched?
5252 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5253 // Bail out if there have been raw-memory effects since the allocation.
5254 // (Example: There might have been a call or safepoint.)
5255 return NULL;
5256 }
5257 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5258 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5259 return NULL;
5260 }
5262 // There must be no unexpected observers of this allocation.
5263 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5264 Node* obs = ptr->fast_out(i);
5265 if (obs != this->map()) {
5266 return NULL;
5267 }
5268 }
5270 // This arraycopy must unconditionally follow the allocation of the ptr.
5271 Node* alloc_ctl = ptr->in(0);
5272 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5274 Node* ctl = control();
5275 while (ctl != alloc_ctl) {
5276 // There may be guards which feed into the slow_region.
5277 // Any other control flow means that we might not get a chance
5278 // to finish initializing the allocated object.
5279 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5280 IfNode* iff = ctl->in(0)->as_If();
5281 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5282 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5283 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5284 ctl = iff->in(0); // This test feeds the known slow_region.
5285 continue;
5286 }
5287 // One more try: Various low-level checks bottom out in
5288 // uncommon traps. If the debug-info of the trap omits
5289 // any reference to the allocation, as we've already
5290 // observed, then there can be no objection to the trap.
5291 bool found_trap = false;
5292 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5293 Node* obs = not_ctl->fast_out(j);
5294 if (obs->in(0) == not_ctl && obs->is_Call() &&
5295 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5296 found_trap = true; break;
5297 }
5298 }
5299 if (found_trap) {
5300 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5301 continue;
5302 }
5303 }
5304 return NULL;
5305 }
5307 // If we get this far, we have an allocation which immediately
5308 // precedes the arraycopy, and we can take over zeroing the new object.
5309 // The arraycopy will finish the initialization, and provide
5310 // a new control state to which we will anchor the destination pointer.
5312 return alloc;
5313 }
5315 // Helper for initialization of arrays, creating a ClearArray.
5316 // It writes zero bits in [start..end), within the body of an array object.
5317 // The memory effects are all chained onto the 'adr_type' alias category.
5318 //
5319 // Since the object is otherwise uninitialized, we are free
5320 // to put a little "slop" around the edges of the cleared area,
5321 // as long as it does not go back into the array's header,
5322 // or beyond the array end within the heap.
5323 //
5324 // The lower edge can be rounded down to the nearest jint and the
5325 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5326 //
5327 // Arguments:
5328 // adr_type memory slice where writes are generated
5329 // dest oop of the destination array
5330 // basic_elem_type element type of the destination
5331 // slice_idx array index of first element to store
5332 // slice_len number of elements to store (or NULL)
5333 // dest_size total size in bytes of the array object
5334 //
5335 // Exactly one of slice_len or dest_size must be non-NULL.
5336 // If dest_size is non-NULL, zeroing extends to the end of the object.
5337 // If slice_len is non-NULL, the slice_idx value must be a constant.
5338 void
5339 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5340 Node* dest,
5341 BasicType basic_elem_type,
5342 Node* slice_idx,
5343 Node* slice_len,
5344 Node* dest_size) {
5345 // one or the other but not both of slice_len and dest_size:
5346 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5347 if (slice_len == NULL) slice_len = top();
5348 if (dest_size == NULL) dest_size = top();
5350 // operate on this memory slice:
5351 Node* mem = memory(adr_type); // memory slice to operate on
5353 // scaling and rounding of indexes:
5354 int scale = exact_log2(type2aelembytes(basic_elem_type));
5355 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5356 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5357 int bump_bit = (-1 << scale) & BytesPerInt;
5359 // determine constant starts and ends
5360 const intptr_t BIG_NEG = -128;
5361 assert(BIG_NEG + 2*abase < 0, "neg enough");
5362 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5363 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5364 if (slice_len_con == 0) {
5365 return; // nothing to do here
5366 }
5367 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5368 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5369 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5370 assert(end_con < 0, "not two cons");
5371 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5372 BytesPerLong);
5373 }
5375 if (start_con >= 0 && end_con >= 0) {
5376 // Constant start and end. Simple.
5377 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5378 start_con, end_con, &_gvn);
5379 } else if (start_con >= 0 && dest_size != top()) {
5380 // Constant start, pre-rounded end after the tail of the array.
5381 Node* end = dest_size;
5382 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5383 start_con, end, &_gvn);
5384 } else if (start_con >= 0 && slice_len != top()) {
5385 // Constant start, non-constant end. End needs rounding up.
5386 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5387 intptr_t end_base = abase + (slice_idx_con << scale);
5388 int end_round = (-1 << scale) & (BytesPerLong - 1);
5389 Node* end = ConvI2X(slice_len);
5390 if (scale != 0)
5391 end = _gvn.transform(new(C) LShiftXNode(end, intcon(scale) ));
5392 end_base += end_round;
5393 end = _gvn.transform(new(C) AddXNode(end, MakeConX(end_base)));
5394 end = _gvn.transform(new(C) AndXNode(end, MakeConX(~end_round)));
5395 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5396 start_con, end, &_gvn);
5397 } else if (start_con < 0 && dest_size != top()) {
5398 // Non-constant start, pre-rounded end after the tail of the array.
5399 // This is almost certainly a "round-to-end" operation.
5400 Node* start = slice_idx;
5401 start = ConvI2X(start);
5402 if (scale != 0)
5403 start = _gvn.transform(new(C) LShiftXNode( start, intcon(scale) ));
5404 start = _gvn.transform(new(C) AddXNode(start, MakeConX(abase)));
5405 if ((bump_bit | clear_low) != 0) {
5406 int to_clear = (bump_bit | clear_low);
5407 // Align up mod 8, then store a jint zero unconditionally
5408 // just before the mod-8 boundary.
5409 if (((abase + bump_bit) & ~to_clear) - bump_bit
5410 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5411 bump_bit = 0;
5412 assert((abase & to_clear) == 0, "array base must be long-aligned");
5413 } else {
5414 // Bump 'start' up to (or past) the next jint boundary:
5415 start = _gvn.transform(new(C) AddXNode(start, MakeConX(bump_bit)));
5416 assert((abase & clear_low) == 0, "array base must be int-aligned");
5417 }
5418 // Round bumped 'start' down to jlong boundary in body of array.
5419 start = _gvn.transform(new(C) AndXNode(start, MakeConX(~to_clear)));
5420 if (bump_bit != 0) {
5421 // Store a zero to the immediately preceding jint:
5422 Node* x1 = _gvn.transform(new(C) AddXNode(start, MakeConX(-bump_bit)));
5423 Node* p1 = basic_plus_adr(dest, x1);
5424 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5425 mem = _gvn.transform(mem);
5426 }
5427 }
5428 Node* end = dest_size; // pre-rounded
5429 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5430 start, end, &_gvn);
5431 } else {
5432 // Non-constant start, unrounded non-constant end.
5433 // (Nobody zeroes a random midsection of an array using this routine.)
5434 ShouldNotReachHere(); // fix caller
5435 }
5437 // Done.
5438 set_memory(mem, adr_type);
5439 }
5442 bool
5443 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5444 BasicType basic_elem_type,
5445 AllocateNode* alloc,
5446 Node* src, Node* src_offset,
5447 Node* dest, Node* dest_offset,
5448 Node* dest_size, bool dest_uninitialized) {
5449 // See if there is an advantage from block transfer.
5450 int scale = exact_log2(type2aelembytes(basic_elem_type));
5451 if (scale >= LogBytesPerLong)
5452 return false; // it is already a block transfer
5454 // Look at the alignment of the starting offsets.
5455 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5457 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5458 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5459 if (src_off_con < 0 || dest_off_con < 0)
5460 // At present, we can only understand constants.
5461 return false;
5463 intptr_t src_off = abase + (src_off_con << scale);
5464 intptr_t dest_off = abase + (dest_off_con << scale);
5466 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5467 // Non-aligned; too bad.
5468 // One more chance: Pick off an initial 32-bit word.
5469 // This is a common case, since abase can be odd mod 8.
5470 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5471 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5472 Node* sptr = basic_plus_adr(src, src_off);
5473 Node* dptr = basic_plus_adr(dest, dest_off);
5474 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5475 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5476 src_off += BytesPerInt;
5477 dest_off += BytesPerInt;
5478 } else {
5479 return false;
5480 }
5481 }
5482 assert(src_off % BytesPerLong == 0, "");
5483 assert(dest_off % BytesPerLong == 0, "");
5485 // Do this copy by giant steps.
5486 Node* sptr = basic_plus_adr(src, src_off);
5487 Node* dptr = basic_plus_adr(dest, dest_off);
5488 Node* countx = dest_size;
5489 countx = _gvn.transform(new (C) SubXNode(countx, MakeConX(dest_off)));
5490 countx = _gvn.transform(new (C) URShiftXNode(countx, intcon(LogBytesPerLong)));
5492 bool disjoint_bases = true; // since alloc != NULL
5493 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5494 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5496 return true;
5497 }
5500 // Helper function; generates code for the slow case.
5501 // We make a call to a runtime method which emulates the native method,
5502 // but without the native wrapper overhead.
5503 void
5504 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5505 Node* src, Node* src_offset,
5506 Node* dest, Node* dest_offset,
5507 Node* copy_length, bool dest_uninitialized) {
5508 assert(!dest_uninitialized, "Invariant");
5509 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5510 OptoRuntime::slow_arraycopy_Type(),
5511 OptoRuntime::slow_arraycopy_Java(),
5512 "slow_arraycopy", adr_type,
5513 src, src_offset, dest, dest_offset,
5514 copy_length);
5516 // Handle exceptions thrown by this fellow:
5517 make_slow_call_ex(call, env()->Throwable_klass(), false);
5518 }
5520 // Helper function; generates code for cases requiring runtime checks.
5521 Node*
5522 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5523 Node* dest_elem_klass,
5524 Node* src, Node* src_offset,
5525 Node* dest, Node* dest_offset,
5526 Node* copy_length, bool dest_uninitialized) {
5527 if (stopped()) return NULL;
5529 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5530 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5531 return NULL;
5532 }
5534 // Pick out the parameters required to perform a store-check
5535 // for the target array. This is an optimistic check. It will
5536 // look in each non-null element's class, at the desired klass's
5537 // super_check_offset, for the desired klass.
5538 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5539 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5540 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5541 Node* check_offset = ConvI2X(_gvn.transform(n3));
5542 Node* check_value = dest_elem_klass;
5544 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5545 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5547 // (We know the arrays are never conjoint, because their types differ.)
5548 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5549 OptoRuntime::checkcast_arraycopy_Type(),
5550 copyfunc_addr, "checkcast_arraycopy", adr_type,
5551 // five arguments, of which two are
5552 // intptr_t (jlong in LP64)
5553 src_start, dest_start,
5554 copy_length XTOP,
5555 check_offset XTOP,
5556 check_value);
5558 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5559 }
5562 // Helper function; generates code for cases requiring runtime checks.
5563 Node*
5564 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5565 Node* src, Node* src_offset,
5566 Node* dest, Node* dest_offset,
5567 Node* copy_length, bool dest_uninitialized) {
5568 assert(!dest_uninitialized, "Invariant");
5569 if (stopped()) return NULL;
5570 address copyfunc_addr = StubRoutines::generic_arraycopy();
5571 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5572 return NULL;
5573 }
5575 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5576 OptoRuntime::generic_arraycopy_Type(),
5577 copyfunc_addr, "generic_arraycopy", adr_type,
5578 src, src_offset, dest, dest_offset, copy_length);
5580 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5581 }
5583 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5584 void
5585 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5586 BasicType basic_elem_type,
5587 bool disjoint_bases,
5588 Node* src, Node* src_offset,
5589 Node* dest, Node* dest_offset,
5590 Node* copy_length, bool dest_uninitialized) {
5591 if (stopped()) return; // nothing to do
5593 Node* src_start = src;
5594 Node* dest_start = dest;
5595 if (src_offset != NULL || dest_offset != NULL) {
5596 assert(src_offset != NULL && dest_offset != NULL, "");
5597 src_start = array_element_address(src, src_offset, basic_elem_type);
5598 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5599 }
5601 // Figure out which arraycopy runtime method to call.
5602 const char* copyfunc_name = "arraycopy";
5603 address copyfunc_addr =
5604 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5605 disjoint_bases, copyfunc_name, dest_uninitialized);
5607 // Call it. Note that the count_ix value is not scaled to a byte-size.
5608 make_runtime_call(RC_LEAF|RC_NO_FP,
5609 OptoRuntime::fast_arraycopy_Type(),
5610 copyfunc_addr, copyfunc_name, adr_type,
5611 src_start, dest_start, copy_length XTOP);
5612 }
5614 //-------------inline_encodeISOArray-----------------------------------
5615 // encode char[] to byte[] in ISO_8859_1
5616 bool LibraryCallKit::inline_encodeISOArray() {
5617 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5618 // no receiver since it is static method
5619 Node *src = argument(0);
5620 Node *src_offset = argument(1);
5621 Node *dst = argument(2);
5622 Node *dst_offset = argument(3);
5623 Node *length = argument(4);
5625 const Type* src_type = src->Value(&_gvn);
5626 const Type* dst_type = dst->Value(&_gvn);
5627 const TypeAryPtr* top_src = src_type->isa_aryptr();
5628 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5629 if (top_src == NULL || top_src->klass() == NULL ||
5630 top_dest == NULL || top_dest->klass() == NULL) {
5631 // failed array check
5632 return false;
5633 }
5635 // Figure out the size and type of the elements we will be copying.
5636 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5637 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5638 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5639 return false;
5640 }
5641 Node* src_start = array_element_address(src, src_offset, src_elem);
5642 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5643 // 'src_start' points to src array + scaled offset
5644 // 'dst_start' points to dst array + scaled offset
5646 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5647 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5648 enc = _gvn.transform(enc);
5649 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5650 set_memory(res_mem, mtype);
5651 set_result(enc);
5652 return true;
5653 }
5655 /**
5656 * Calculate CRC32 for byte.
5657 * int java.util.zip.CRC32.update(int crc, int b)
5658 */
5659 bool LibraryCallKit::inline_updateCRC32() {
5660 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5661 assert(callee()->signature()->size() == 2, "update has 2 parameters");
5662 // no receiver since it is static method
5663 Node* crc = argument(0); // type: int
5664 Node* b = argument(1); // type: int
5666 /*
5667 * int c = ~ crc;
5668 * b = timesXtoThe32[(b ^ c) & 0xFF];
5669 * b = b ^ (c >>> 8);
5670 * crc = ~b;
5671 */
5673 Node* M1 = intcon(-1);
5674 crc = _gvn.transform(new (C) XorINode(crc, M1));
5675 Node* result = _gvn.transform(new (C) XorINode(crc, b));
5676 result = _gvn.transform(new (C) AndINode(result, intcon(0xFF)));
5678 Node* base = makecon(TypeRawPtr::make(StubRoutines::crc_table_addr()));
5679 Node* offset = _gvn.transform(new (C) LShiftINode(result, intcon(0x2)));
5680 Node* adr = basic_plus_adr(top(), base, ConvI2X(offset));
5681 result = make_load(control(), adr, TypeInt::INT, T_INT);
5683 crc = _gvn.transform(new (C) URShiftINode(crc, intcon(8)));
5684 result = _gvn.transform(new (C) XorINode(crc, result));
5685 result = _gvn.transform(new (C) XorINode(result, M1));
5686 set_result(result);
5687 return true;
5688 }
5690 /**
5691 * Calculate CRC32 for byte[] array.
5692 * int java.util.zip.CRC32.updateBytes(int crc, byte[] buf, int off, int len)
5693 */
5694 bool LibraryCallKit::inline_updateBytesCRC32() {
5695 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5696 assert(callee()->signature()->size() == 4, "updateBytes has 4 parameters");
5697 // no receiver since it is static method
5698 Node* crc = argument(0); // type: int
5699 Node* src = argument(1); // type: oop
5700 Node* offset = argument(2); // type: int
5701 Node* length = argument(3); // type: int
5703 const Type* src_type = src->Value(&_gvn);
5704 const TypeAryPtr* top_src = src_type->isa_aryptr();
5705 if (top_src == NULL || top_src->klass() == NULL) {
5706 // failed array check
5707 return false;
5708 }
5710 // Figure out the size and type of the elements we will be copying.
5711 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5712 if (src_elem != T_BYTE) {
5713 return false;
5714 }
5716 // 'src_start' points to src array + scaled offset
5717 Node* src_start = array_element_address(src, offset, src_elem);
5719 // We assume that range check is done by caller.
5720 // TODO: generate range check (offset+length < src.length) in debug VM.
5722 // Call the stub.
5723 address stubAddr = StubRoutines::updateBytesCRC32();
5724 const char *stubName = "updateBytesCRC32";
5726 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5727 stubAddr, stubName, TypePtr::BOTTOM,
5728 crc, src_start, length);
5729 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5730 set_result(result);
5731 return true;
5732 }
5734 /**
5735 * Calculate CRC32 for ByteBuffer.
5736 * int java.util.zip.CRC32.updateByteBuffer(int crc, long buf, int off, int len)
5737 */
5738 bool LibraryCallKit::inline_updateByteBufferCRC32() {
5739 assert(UseCRC32Intrinsics, "need AVX and LCMUL instructions support");
5740 assert(callee()->signature()->size() == 5, "updateByteBuffer has 4 parameters and one is long");
5741 // no receiver since it is static method
5742 Node* crc = argument(0); // type: int
5743 Node* src = argument(1); // type: long
5744 Node* offset = argument(3); // type: int
5745 Node* length = argument(4); // type: int
5747 src = ConvL2X(src); // adjust Java long to machine word
5748 Node* base = _gvn.transform(new (C) CastX2PNode(src));
5749 offset = ConvI2X(offset);
5751 // 'src_start' points to src array + scaled offset
5752 Node* src_start = basic_plus_adr(top(), base, offset);
5754 // Call the stub.
5755 address stubAddr = StubRoutines::updateBytesCRC32();
5756 const char *stubName = "updateBytesCRC32";
5758 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::updateBytesCRC32_Type(),
5759 stubAddr, stubName, TypePtr::BOTTOM,
5760 crc, src_start, length);
5761 Node* result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5762 set_result(result);
5763 return true;
5764 }
5766 //----------------------------inline_reference_get----------------------------
5767 // public T java.lang.ref.Reference.get();
5768 bool LibraryCallKit::inline_reference_get() {
5769 const int referent_offset = java_lang_ref_Reference::referent_offset;
5770 guarantee(referent_offset > 0, "should have already been set");
5772 // Get the argument:
5773 Node* reference_obj = null_check_receiver();
5774 if (stopped()) return true;
5776 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5778 ciInstanceKlass* klass = env()->Object_klass();
5779 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5781 Node* no_ctrl = NULL;
5782 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5784 // Use the pre-barrier to record the value in the referent field
5785 pre_barrier(false /* do_load */,
5786 control(),
5787 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5788 result /* pre_val */,
5789 T_OBJECT);
5791 // Add memory barrier to prevent commoning reads from this field
5792 // across safepoint since GC can change its value.
5793 insert_mem_bar(Op_MemBarCPUOrder);
5795 set_result(result);
5796 return true;
5797 }
5800 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5801 bool is_exact=true, bool is_static=false) {
5803 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5804 assert(tinst != NULL, "obj is null");
5805 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5806 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5808 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5809 ciSymbol::make(fieldTypeString),
5810 is_static);
5811 if (field == NULL) return (Node *) NULL;
5812 assert (field != NULL, "undefined field");
5814 // Next code copied from Parse::do_get_xxx():
5816 // Compute address and memory type.
5817 int offset = field->offset_in_bytes();
5818 bool is_vol = field->is_volatile();
5819 ciType* field_klass = field->type();
5820 assert(field_klass->is_loaded(), "should be loaded");
5821 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5822 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5823 BasicType bt = field->layout_type();
5825 // Build the resultant type of the load
5826 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5828 // Build the load.
5829 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5830 return loadedField;
5831 }
5834 //------------------------------inline_aescrypt_Block-----------------------
5835 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5836 address stubAddr;
5837 const char *stubName;
5838 assert(UseAES, "need AES instruction support");
5840 switch(id) {
5841 case vmIntrinsics::_aescrypt_encryptBlock:
5842 stubAddr = StubRoutines::aescrypt_encryptBlock();
5843 stubName = "aescrypt_encryptBlock";
5844 break;
5845 case vmIntrinsics::_aescrypt_decryptBlock:
5846 stubAddr = StubRoutines::aescrypt_decryptBlock();
5847 stubName = "aescrypt_decryptBlock";
5848 break;
5849 }
5850 if (stubAddr == NULL) return false;
5852 Node* aescrypt_object = argument(0);
5853 Node* src = argument(1);
5854 Node* src_offset = argument(2);
5855 Node* dest = argument(3);
5856 Node* dest_offset = argument(4);
5858 // (1) src and dest are arrays.
5859 const Type* src_type = src->Value(&_gvn);
5860 const Type* dest_type = dest->Value(&_gvn);
5861 const TypeAryPtr* top_src = src_type->isa_aryptr();
5862 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5863 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5865 // for the quick and dirty code we will skip all the checks.
5866 // we are just trying to get the call to be generated.
5867 Node* src_start = src;
5868 Node* dest_start = dest;
5869 if (src_offset != NULL || dest_offset != NULL) {
5870 assert(src_offset != NULL && dest_offset != NULL, "");
5871 src_start = array_element_address(src, src_offset, T_BYTE);
5872 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5873 }
5875 // now need to get the start of its expanded key array
5876 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5877 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5878 if (k_start == NULL) return false;
5880 if (Matcher::pass_original_key_for_aes()) {
5881 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5882 // compatibility issues between Java key expansion and SPARC crypto instructions
5883 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5884 if (original_k_start == NULL) return false;
5886 // Call the stub.
5887 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5888 stubAddr, stubName, TypePtr::BOTTOM,
5889 src_start, dest_start, k_start, original_k_start);
5890 } else {
5891 // Call the stub.
5892 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5893 stubAddr, stubName, TypePtr::BOTTOM,
5894 src_start, dest_start, k_start);
5895 }
5897 return true;
5898 }
5900 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5901 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5902 address stubAddr;
5903 const char *stubName;
5905 assert(UseAES, "need AES instruction support");
5907 switch(id) {
5908 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5909 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5910 stubName = "cipherBlockChaining_encryptAESCrypt";
5911 break;
5912 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5913 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5914 stubName = "cipherBlockChaining_decryptAESCrypt";
5915 break;
5916 }
5917 if (stubAddr == NULL) return false;
5919 Node* cipherBlockChaining_object = argument(0);
5920 Node* src = argument(1);
5921 Node* src_offset = argument(2);
5922 Node* len = argument(3);
5923 Node* dest = argument(4);
5924 Node* dest_offset = argument(5);
5926 // (1) src and dest are arrays.
5927 const Type* src_type = src->Value(&_gvn);
5928 const Type* dest_type = dest->Value(&_gvn);
5929 const TypeAryPtr* top_src = src_type->isa_aryptr();
5930 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5931 assert (top_src != NULL && top_src->klass() != NULL
5932 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5934 // checks are the responsibility of the caller
5935 Node* src_start = src;
5936 Node* dest_start = dest;
5937 if (src_offset != NULL || dest_offset != NULL) {
5938 assert(src_offset != NULL && dest_offset != NULL, "");
5939 src_start = array_element_address(src, src_offset, T_BYTE);
5940 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5941 }
5943 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5944 // (because of the predicated logic executed earlier).
5945 // so we cast it here safely.
5946 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5948 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5949 if (embeddedCipherObj == NULL) return false;
5951 // cast it to what we know it will be at runtime
5952 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5953 assert(tinst != NULL, "CBC obj is null");
5954 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5955 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5956 if (!klass_AESCrypt->is_loaded()) return false;
5958 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5959 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5960 const TypeOopPtr* xtype = aklass->as_instance_type();
5961 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5962 aescrypt_object = _gvn.transform(aescrypt_object);
5964 // we need to get the start of the aescrypt_object's expanded key array
5965 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5966 if (k_start == NULL) return false;
5968 // similarly, get the start address of the r vector
5969 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5970 if (objRvec == NULL) return false;
5971 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5973 Node* cbcCrypt;
5974 if (Matcher::pass_original_key_for_aes()) {
5975 // on SPARC we need to pass the original key since key expansion needs to happen in intrinsics due to
5976 // compatibility issues between Java key expansion and SPARC crypto instructions
5977 Node* original_k_start = get_original_key_start_from_aescrypt_object(aescrypt_object);
5978 if (original_k_start == NULL) return false;
5980 // Call the stub, passing src_start, dest_start, k_start, r_start, src_len and original_k_start
5981 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5982 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5983 stubAddr, stubName, TypePtr::BOTTOM,
5984 src_start, dest_start, k_start, r_start, len, original_k_start);
5985 } else {
5986 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5987 cbcCrypt = make_runtime_call(RC_LEAF|RC_NO_FP,
5988 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5989 stubAddr, stubName, TypePtr::BOTTOM,
5990 src_start, dest_start, k_start, r_start, len);
5991 }
5993 // return cipher length (int)
5994 Node* retvalue = _gvn.transform(new (C) ProjNode(cbcCrypt, TypeFunc::Parms));
5995 set_result(retvalue);
5996 return true;
5997 }
5999 //------------------------------get_key_start_from_aescrypt_object-----------------------
6000 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
6001 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
6002 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6003 if (objAESCryptKey == NULL) return (Node *) NULL;
6005 // now have the array, need to get the start address of the K array
6006 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
6007 return k_start;
6008 }
6010 //------------------------------get_original_key_start_from_aescrypt_object-----------------------
6011 Node * LibraryCallKit::get_original_key_start_from_aescrypt_object(Node *aescrypt_object) {
6012 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "lastKey", "[B", /*is_exact*/ false);
6013 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
6014 if (objAESCryptKey == NULL) return (Node *) NULL;
6016 // now have the array, need to get the start address of the lastKey array
6017 Node* original_k_start = array_element_address(objAESCryptKey, intcon(0), T_BYTE);
6018 return original_k_start;
6019 }
6021 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
6022 // Return node representing slow path of predicate check.
6023 // the pseudo code we want to emulate with this predicate is:
6024 // for encryption:
6025 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
6026 // for decryption:
6027 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
6028 // note cipher==plain is more conservative than the original java code but that's OK
6029 //
6030 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
6031 // First, check receiver for NULL since it is virtual method.
6032 Node* objCBC = argument(0);
6033 objCBC = null_check(objCBC);
6035 if (stopped()) return NULL; // Always NULL
6037 // Load embeddedCipher field of CipherBlockChaining object.
6038 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
6040 // get AESCrypt klass for instanceOf check
6041 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
6042 // will have same classloader as CipherBlockChaining object
6043 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
6044 assert(tinst != NULL, "CBCobj is null");
6045 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
6047 // we want to do an instanceof comparison against the AESCrypt class
6048 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
6049 if (!klass_AESCrypt->is_loaded()) {
6050 // if AESCrypt is not even loaded, we never take the intrinsic fast path
6051 Node* ctrl = control();
6052 set_control(top()); // no regular fast path
6053 return ctrl;
6054 }
6055 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
6057 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
6058 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
6059 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
6061 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
6063 // for encryption, we are done
6064 if (!decrypting)
6065 return instof_false; // even if it is NULL
6067 // for decryption, we need to add a further check to avoid
6068 // taking the intrinsic path when cipher and plain are the same
6069 // see the original java code for why.
6070 RegionNode* region = new(C) RegionNode(3);
6071 region->init_req(1, instof_false);
6072 Node* src = argument(1);
6073 Node* dest = argument(4);
6074 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
6075 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
6076 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
6077 region->init_req(2, src_dest_conjoint);
6079 record_for_igvn(region);
6080 return _gvn.transform(region);
6081 }