1 /*
2 * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
41
42 class LibraryIntrinsic : public InlineCallGenerator {
43 // Extend the set of intrinsics known to the runtime:
44 public:
45 private:
46 bool _is_virtual;
47 bool _is_predicted;
48 vmIntrinsics::ID _intrinsic_id;
49
50 public:
51 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id)
52 : InlineCallGenerator(m),
53 _is_virtual(is_virtual),
54 _is_predicted(is_predicted),
55 _intrinsic_id(id)
56 {
57 }
58 virtual bool is_intrinsic() const { return true; }
59 virtual bool is_virtual() const { return _is_virtual; }
60 virtual bool is_predicted() const { return _is_predicted; }
61 virtual JVMState* generate(JVMState* jvms);
62 virtual Node* generate_predicate(JVMState* jvms);
63 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
64 };
65
66
67 // Local helper class for LibraryIntrinsic:
68 class LibraryCallKit : public GraphKit {
69 private:
70 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
71 Node* _result; // the result node, if any
72 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
73
74 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
75
76 public:
77 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
78 : GraphKit(jvms),
79 _intrinsic(intrinsic),
80 _result(NULL)
81 {
82 // Check if this is a root compile. In that case we don't have a caller.
83 if (!jvms->has_method()) {
84 _reexecute_sp = sp();
85 } else {
86 // Find out how many arguments the interpreter needs when deoptimizing
87 // and save the stack pointer value so it can used by uncommon_trap.
88 // We find the argument count by looking at the declared signature.
89 bool ignored_will_link;
90 ciSignature* declared_signature = NULL;
91 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
92 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
93 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
94 }
95 }
96
97 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
98
99 ciMethod* caller() const { return jvms()->method(); }
100 int bci() const { return jvms()->bci(); }
101 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
102 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
103 ciMethod* callee() const { return _intrinsic->method(); }
104
105 bool try_to_inline();
106 Node* try_to_predicate();
107
108 void push_result() {
109 // Push the result onto the stack.
110 if (!stopped() && result() != NULL) {
111 BasicType bt = result()->bottom_type()->basic_type();
112 push_node(bt, result());
113 }
114 }
115
116 private:
117 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
118 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
119 }
120
121 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
122 void set_result(RegionNode* region, PhiNode* value);
123 Node* result() { return _result; }
124
125 virtual int reexecute_sp() { return _reexecute_sp; }
126
127 // Helper functions to inline natives
128 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
129 Node* generate_slow_guard(Node* test, RegionNode* region);
130 Node* generate_fair_guard(Node* test, RegionNode* region);
131 Node* generate_negative_guard(Node* index, RegionNode* region,
132 // resulting CastII of index:
133 Node* *pos_index = NULL);
134 Node* generate_nonpositive_guard(Node* index, bool never_negative,
135 // resulting CastII of index:
136 Node* *pos_index = NULL);
137 Node* generate_limit_guard(Node* offset, Node* subseq_length,
138 Node* array_length,
139 RegionNode* region);
140 Node* generate_current_thread(Node* &tls_output);
141 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
142 bool disjoint_bases, const char* &name, bool dest_uninitialized);
143 Node* load_mirror_from_klass(Node* klass);
144 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
145 RegionNode* region, int null_path,
146 int offset);
147 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
148 RegionNode* region, int null_path) {
149 int offset = java_lang_Class::klass_offset_in_bytes();
150 return load_klass_from_mirror_common(mirror, never_see_null,
151 region, null_path,
152 offset);
153 }
154 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path) {
156 int offset = java_lang_Class::array_klass_offset_in_bytes();
157 return load_klass_from_mirror_common(mirror, never_see_null,
158 region, null_path,
159 offset);
160 }
161 Node* generate_access_flags_guard(Node* kls,
162 int modifier_mask, int modifier_bits,
163 RegionNode* region);
164 Node* generate_interface_guard(Node* kls, RegionNode* region);
165 Node* generate_array_guard(Node* kls, RegionNode* region) {
166 return generate_array_guard_common(kls, region, false, false);
167 }
168 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
169 return generate_array_guard_common(kls, region, false, true);
170 }
171 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
172 return generate_array_guard_common(kls, region, true, false);
173 }
174 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
175 return generate_array_guard_common(kls, region, true, true);
176 }
177 Node* generate_array_guard_common(Node* kls, RegionNode* region,
178 bool obj_array, bool not_array);
179 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
180 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
181 bool is_virtual = false, bool is_static = false);
182 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
183 return generate_method_call(method_id, false, true);
184 }
185 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
186 return generate_method_call(method_id, true, false);
187 }
188 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
189
190 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
191 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
192 bool inline_string_compareTo();
193 bool inline_string_indexOf();
194 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
195 bool inline_string_equals();
196 Node* round_double_node(Node* n);
197 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
198 bool inline_math_native(vmIntrinsics::ID id);
199 bool inline_trig(vmIntrinsics::ID id);
200 bool inline_math(vmIntrinsics::ID id);
201 bool inline_exp();
202 bool inline_pow();
203 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
204 bool inline_min_max(vmIntrinsics::ID id);
205 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
206 // This returns Type::AnyPtr, RawPtr, or OopPtr.
207 int classify_unsafe_addr(Node* &base, Node* &offset);
208 Node* make_unsafe_address(Node* base, Node* offset);
209 // Helper for inline_unsafe_access.
210 // Generates the guards that check whether the result of
211 // Unsafe.getObject should be recorded in an SATB log buffer.
212 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
213 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
214 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
215 bool inline_unsafe_allocate();
216 bool inline_unsafe_copyMemory();
217 bool inline_native_currentThread();
218 #ifdef TRACE_HAVE_INTRINSICS
219 bool inline_native_classID();
220 bool inline_native_threadID();
221 #endif
222 bool inline_native_time_funcs(address method, const char* funcName);
223 bool inline_native_isInterrupted();
224 bool inline_native_Class_query(vmIntrinsics::ID id);
225 bool inline_native_subtype_check();
226
227 bool inline_native_newArray();
228 bool inline_native_getLength();
229 bool inline_array_copyOf(bool is_copyOfRange);
230 bool inline_array_equals();
231 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
232 bool inline_native_clone(bool is_virtual);
233 bool inline_native_Reflection_getCallerClass();
234 // Helper function for inlining native object hash method
235 bool inline_native_hashcode(bool is_virtual, bool is_static);
236 bool inline_native_getClass();
237
238 // Helper functions for inlining arraycopy
239 bool inline_arraycopy();
240 void generate_arraycopy(const TypePtr* adr_type,
241 BasicType basic_elem_type,
242 Node* src, Node* src_offset,
243 Node* dest, Node* dest_offset,
244 Node* copy_length,
245 bool disjoint_bases = false,
246 bool length_never_negative = false,
247 RegionNode* slow_region = NULL);
248 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
249 RegionNode* slow_region);
250 void generate_clear_array(const TypePtr* adr_type,
251 Node* dest,
252 BasicType basic_elem_type,
253 Node* slice_off,
254 Node* slice_len,
255 Node* slice_end);
256 bool generate_block_arraycopy(const TypePtr* adr_type,
257 BasicType basic_elem_type,
258 AllocateNode* alloc,
259 Node* src, Node* src_offset,
260 Node* dest, Node* dest_offset,
261 Node* dest_size, bool dest_uninitialized);
262 void generate_slow_arraycopy(const TypePtr* adr_type,
263 Node* src, Node* src_offset,
264 Node* dest, Node* dest_offset,
265 Node* copy_length, bool dest_uninitialized);
266 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
267 Node* dest_elem_klass,
268 Node* src, Node* src_offset,
269 Node* dest, Node* dest_offset,
270 Node* copy_length, bool dest_uninitialized);
271 Node* generate_generic_arraycopy(const TypePtr* adr_type,
272 Node* src, Node* src_offset,
273 Node* dest, Node* dest_offset,
274 Node* copy_length, bool dest_uninitialized);
275 void generate_unchecked_arraycopy(const TypePtr* adr_type,
276 BasicType basic_elem_type,
277 bool disjoint_bases,
278 Node* src, Node* src_offset,
279 Node* dest, Node* dest_offset,
280 Node* copy_length, bool dest_uninitialized);
281 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
282 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
283 bool inline_unsafe_ordered_store(BasicType type);
284 bool inline_unsafe_fence(vmIntrinsics::ID id);
285 bool inline_fp_conversions(vmIntrinsics::ID id);
286 bool inline_number_methods(vmIntrinsics::ID id);
287 bool inline_reference_get();
288 bool inline_aescrypt_Block(vmIntrinsics::ID id);
289 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
290 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
291 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
292 bool inline_encodeISOArray();
293 };
294
295
296 //---------------------------make_vm_intrinsic----------------------------
297 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
298 vmIntrinsics::ID id = m->intrinsic_id();
299 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
300
301 if (DisableIntrinsic[0] != '\0'
302 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
303 // disabled by a user request on the command line:
304 // example: -XX:DisableIntrinsic=_hashCode,_getClass
305 return NULL;
306 }
307
308 if (!m->is_loaded()) {
309 // do not attempt to inline unloaded methods
310 return NULL;
311 }
312
313 // Only a few intrinsics implement a virtual dispatch.
314 // They are expensive calls which are also frequently overridden.
315 if (is_virtual) {
316 switch (id) {
317 case vmIntrinsics::_hashCode:
318 case vmIntrinsics::_clone:
319 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
320 break;
321 default:
322 return NULL;
323 }
324 }
325
326 // -XX:-InlineNatives disables nearly all intrinsics:
327 if (!InlineNatives) {
328 switch (id) {
329 case vmIntrinsics::_indexOf:
330 case vmIntrinsics::_compareTo:
331 case vmIntrinsics::_equals:
332 case vmIntrinsics::_equalsC:
333 case vmIntrinsics::_getAndAddInt:
334 case vmIntrinsics::_getAndAddLong:
335 case vmIntrinsics::_getAndSetInt:
336 case vmIntrinsics::_getAndSetLong:
337 case vmIntrinsics::_getAndSetObject:
338 case vmIntrinsics::_loadFence:
339 case vmIntrinsics::_storeFence:
340 case vmIntrinsics::_fullFence:
341 break; // InlineNatives does not control String.compareTo
342 case vmIntrinsics::_Reference_get:
343 break; // InlineNatives does not control Reference.get
344 default:
345 return NULL;
346 }
347 }
348
349 bool is_predicted = false;
350
351 switch (id) {
352 case vmIntrinsics::_compareTo:
353 if (!SpecialStringCompareTo) return NULL;
354 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
355 break;
356 case vmIntrinsics::_indexOf:
357 if (!SpecialStringIndexOf) return NULL;
358 break;
359 case vmIntrinsics::_equals:
360 if (!SpecialStringEquals) return NULL;
361 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
362 break;
363 case vmIntrinsics::_equalsC:
364 if (!SpecialArraysEquals) return NULL;
365 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
366 break;
367 case vmIntrinsics::_arraycopy:
368 if (!InlineArrayCopy) return NULL;
369 break;
370 case vmIntrinsics::_copyMemory:
371 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
372 if (!InlineArrayCopy) return NULL;
373 break;
374 case vmIntrinsics::_hashCode:
375 if (!InlineObjectHash) return NULL;
376 break;
377 case vmIntrinsics::_clone:
378 case vmIntrinsics::_copyOf:
379 case vmIntrinsics::_copyOfRange:
380 if (!InlineObjectCopy) return NULL;
381 // These also use the arraycopy intrinsic mechanism:
382 if (!InlineArrayCopy) return NULL;
383 break;
384 case vmIntrinsics::_encodeISOArray:
385 if (!SpecialEncodeISOArray) return NULL;
386 if (!Matcher::match_rule_supported(Op_EncodeISOArray)) return NULL;
387 break;
388 case vmIntrinsics::_checkIndex:
389 // We do not intrinsify this. The optimizer does fine with it.
390 return NULL;
391
392 case vmIntrinsics::_getCallerClass:
393 if (!UseNewReflection) return NULL;
394 if (!InlineReflectionGetCallerClass) return NULL;
395 if (SystemDictionary::reflect_CallerSensitive_klass() == NULL) return NULL;
396 break;
397
398 case vmIntrinsics::_bitCount_i:
399 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
400 break;
401
402 case vmIntrinsics::_bitCount_l:
403 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
404 break;
405
406 case vmIntrinsics::_numberOfLeadingZeros_i:
407 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
408 break;
409
410 case vmIntrinsics::_numberOfLeadingZeros_l:
411 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
412 break;
413
414 case vmIntrinsics::_numberOfTrailingZeros_i:
415 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
416 break;
417
418 case vmIntrinsics::_numberOfTrailingZeros_l:
419 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
420 break;
421
422 case vmIntrinsics::_reverseBytes_c:
423 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
424 break;
425 case vmIntrinsics::_reverseBytes_s:
426 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
427 break;
428 case vmIntrinsics::_reverseBytes_i:
429 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
430 break;
431 case vmIntrinsics::_reverseBytes_l:
432 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
433 break;
434
435 case vmIntrinsics::_Reference_get:
436 // Use the intrinsic version of Reference.get() so that the value in
437 // the referent field can be registered by the G1 pre-barrier code.
438 // Also add memory barrier to prevent commoning reads from this field
439 // across safepoint since GC can change it value.
440 break;
441
442 case vmIntrinsics::_compareAndSwapObject:
443 #ifdef _LP64
444 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
445 #endif
446 break;
447
448 case vmIntrinsics::_compareAndSwapLong:
449 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
450 break;
451
452 case vmIntrinsics::_getAndAddInt:
453 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
454 break;
455
456 case vmIntrinsics::_getAndAddLong:
457 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
458 break;
459
460 case vmIntrinsics::_getAndSetInt:
461 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
462 break;
463
464 case vmIntrinsics::_getAndSetLong:
465 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
466 break;
467
468 case vmIntrinsics::_getAndSetObject:
469 #ifdef _LP64
470 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
471 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
472 break;
473 #else
474 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
475 break;
476 #endif
477
478 case vmIntrinsics::_aescrypt_encryptBlock:
479 case vmIntrinsics::_aescrypt_decryptBlock:
480 if (!UseAESIntrinsics) return NULL;
481 break;
482
483 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
484 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
485 if (!UseAESIntrinsics) return NULL;
486 // these two require the predicated logic
487 is_predicted = true;
488 break;
489
490 default:
491 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
492 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
493 break;
494 }
495
496 // -XX:-InlineClassNatives disables natives from the Class class.
497 // The flag applies to all reflective calls, notably Array.newArray
498 // (visible to Java programmers as Array.newInstance).
499 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
500 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
501 if (!InlineClassNatives) return NULL;
502 }
503
504 // -XX:-InlineThreadNatives disables natives from the Thread class.
505 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
506 if (!InlineThreadNatives) return NULL;
507 }
508
509 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
510 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
511 m->holder()->name() == ciSymbol::java_lang_Float() ||
512 m->holder()->name() == ciSymbol::java_lang_Double()) {
513 if (!InlineMathNatives) return NULL;
514 }
515
516 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
517 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
518 if (!InlineUnsafeOps) return NULL;
519 }
520
521 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
522 }
523
524 //----------------------register_library_intrinsics-----------------------
525 // Initialize this file's data structures, for each Compile instance.
526 void Compile::register_library_intrinsics() {
527 // Nothing to do here.
528 }
529
530 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
531 LibraryCallKit kit(jvms, this);
532 Compile* C = kit.C;
533 int nodes = C->unique();
534 #ifndef PRODUCT
535 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
536 char buf[1000];
537 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
538 tty->print_cr("Intrinsic %s", str);
539 }
540 #endif
541 ciMethod* callee = kit.callee();
542 const int bci = kit.bci();
543
544 // Try to inline the intrinsic.
545 if (kit.try_to_inline()) {
546 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
547 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
548 }
549 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
550 if (C->log()) {
551 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
552 vmIntrinsics::name_at(intrinsic_id()),
553 (is_virtual() ? " virtual='1'" : ""),
554 C->unique() - nodes);
555 }
556 // Push the result from the inlined method onto the stack.
557 kit.push_result();
558 return kit.transfer_exceptions_into_jvms();
559 }
560
561 // The intrinsic bailed out
562 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
563 if (jvms->has_method()) {
564 // Not a root compile.
565 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
566 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
567 } else {
568 // Root compile
569 tty->print("Did not generate intrinsic %s%s at bci:%d in",
570 vmIntrinsics::name_at(intrinsic_id()),
571 (is_virtual() ? " (virtual)" : ""), bci);
572 }
573 }
574 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
575 return NULL;
576 }
577
578 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
579 LibraryCallKit kit(jvms, this);
580 Compile* C = kit.C;
581 int nodes = C->unique();
582 #ifndef PRODUCT
583 assert(is_predicted(), "sanity");
584 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
585 char buf[1000];
586 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
587 tty->print_cr("Predicate for intrinsic %s", str);
588 }
589 #endif
590 ciMethod* callee = kit.callee();
591 const int bci = kit.bci();
592
593 Node* slow_ctl = kit.try_to_predicate();
594 if (!kit.failing()) {
595 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
596 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
597 }
598 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
599 if (C->log()) {
600 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
601 vmIntrinsics::name_at(intrinsic_id()),
602 (is_virtual() ? " virtual='1'" : ""),
603 C->unique() - nodes);
604 }
605 return slow_ctl; // Could be NULL if the check folds.
606 }
607
608 // The intrinsic bailed out
609 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
610 if (jvms->has_method()) {
611 // Not a root compile.
612 const char* msg = "failed to generate predicate for intrinsic";
613 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
614 } else {
615 // Root compile
616 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
617 vmIntrinsics::name_at(intrinsic_id()),
618 (is_virtual() ? " (virtual)" : ""), bci);
619 }
620 }
621 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
622 return NULL;
623 }
624
625 bool LibraryCallKit::try_to_inline() {
626 // Handle symbolic names for otherwise undistinguished boolean switches:
627 const bool is_store = true;
628 const bool is_native_ptr = true;
629 const bool is_static = true;
630 const bool is_volatile = true;
631
632 if (!jvms()->has_method()) {
633 // Root JVMState has a null method.
634 assert(map()->memory()->Opcode() == Op_Parm, "");
635 // Insert the memory aliasing node
636 set_all_memory(reset_memory());
637 }
638 assert(merged_memory(), "");
639
640
641 switch (intrinsic_id()) {
642 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
643 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
644 case vmIntrinsics::_getClass: return inline_native_getClass();
645
646 case vmIntrinsics::_dsin:
647 case vmIntrinsics::_dcos:
648 case vmIntrinsics::_dtan:
649 case vmIntrinsics::_dabs:
650 case vmIntrinsics::_datan2:
651 case vmIntrinsics::_dsqrt:
652 case vmIntrinsics::_dexp:
653 case vmIntrinsics::_dlog:
654 case vmIntrinsics::_dlog10:
655 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
656
657 case vmIntrinsics::_min:
658 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
659
660 case vmIntrinsics::_arraycopy: return inline_arraycopy();
661
662 case vmIntrinsics::_compareTo: return inline_string_compareTo();
663 case vmIntrinsics::_indexOf: return inline_string_indexOf();
664 case vmIntrinsics::_equals: return inline_string_equals();
665
666 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
667 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
668 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
669 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
670 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
671 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
672 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
673 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
674 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
675
676 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
677 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
678 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
679 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
680 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
681 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
682 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
683 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
684 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
685
686 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
687 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
688 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
689 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
690 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
691 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
692 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
693 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
694
695 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
696 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
697 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
698 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
699 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
700 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
701 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
702 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
703
704 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
705 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
706 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
707 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
708 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
709 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
710 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
711 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
712 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
713
714 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
715 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
716 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
717 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
718 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
719 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
720 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
721 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
722 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
723
724 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
725 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
726 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
727 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
728
729 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
730 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
731 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
732
733 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
734 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
735 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
736
737 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
738 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
739 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
740 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
741 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
742
743 case vmIntrinsics::_loadFence:
744 case vmIntrinsics::_storeFence:
745 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
746
747 case vmIntrinsics::_currentThread: return inline_native_currentThread();
748 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
749
750 #ifdef TRACE_HAVE_INTRINSICS
751 case vmIntrinsics::_classID: return inline_native_classID();
752 case vmIntrinsics::_threadID: return inline_native_threadID();
753 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
754 #endif
755 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
756 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
757 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
758 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
759 case vmIntrinsics::_newArray: return inline_native_newArray();
760 case vmIntrinsics::_getLength: return inline_native_getLength();
761 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
762 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
763 case vmIntrinsics::_equalsC: return inline_array_equals();
764 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
765
766 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
767
768 case vmIntrinsics::_isInstance:
769 case vmIntrinsics::_getModifiers:
770 case vmIntrinsics::_isInterface:
771 case vmIntrinsics::_isArray:
772 case vmIntrinsics::_isPrimitive:
773 case vmIntrinsics::_getSuperclass:
774 case vmIntrinsics::_getComponentType:
775 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
776
777 case vmIntrinsics::_floatToRawIntBits:
778 case vmIntrinsics::_floatToIntBits:
779 case vmIntrinsics::_intBitsToFloat:
780 case vmIntrinsics::_doubleToRawLongBits:
781 case vmIntrinsics::_doubleToLongBits:
782 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
783
784 case vmIntrinsics::_numberOfLeadingZeros_i:
785 case vmIntrinsics::_numberOfLeadingZeros_l:
786 case vmIntrinsics::_numberOfTrailingZeros_i:
787 case vmIntrinsics::_numberOfTrailingZeros_l:
788 case vmIntrinsics::_bitCount_i:
789 case vmIntrinsics::_bitCount_l:
790 case vmIntrinsics::_reverseBytes_i:
791 case vmIntrinsics::_reverseBytes_l:
792 case vmIntrinsics::_reverseBytes_s:
793 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
794
795 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
796
797 case vmIntrinsics::_Reference_get: return inline_reference_get();
798
799 case vmIntrinsics::_aescrypt_encryptBlock:
800 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
801
802 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
803 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
804 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
805
806 case vmIntrinsics::_encodeISOArray:
807 return inline_encodeISOArray();
808
809 default:
810 // If you get here, it may be that someone has added a new intrinsic
811 // to the list in vmSymbols.hpp without implementing it here.
812 #ifndef PRODUCT
813 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
814 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
815 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
816 }
817 #endif
818 return false;
819 }
820 }
821
822 Node* LibraryCallKit::try_to_predicate() {
823 if (!jvms()->has_method()) {
824 // Root JVMState has a null method.
825 assert(map()->memory()->Opcode() == Op_Parm, "");
826 // Insert the memory aliasing node
827 set_all_memory(reset_memory());
828 }
829 assert(merged_memory(), "");
830
831 switch (intrinsic_id()) {
832 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
833 return inline_cipherBlockChaining_AESCrypt_predicate(false);
834 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
835 return inline_cipherBlockChaining_AESCrypt_predicate(true);
836
837 default:
838 // If you get here, it may be that someone has added a new intrinsic
839 // to the list in vmSymbols.hpp without implementing it here.
840 #ifndef PRODUCT
841 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
842 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
843 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
844 }
845 #endif
846 Node* slow_ctl = control();
847 set_control(top()); // No fast path instrinsic
848 return slow_ctl;
849 }
850 }
851
852 //------------------------------set_result-------------------------------
853 // Helper function for finishing intrinsics.
854 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
855 record_for_igvn(region);
856 set_control(_gvn.transform(region));
857 set_result( _gvn.transform(value));
858 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
859 }
860
861 //------------------------------generate_guard---------------------------
862 // Helper function for generating guarded fast-slow graph structures.
863 // The given 'test', if true, guards a slow path. If the test fails
864 // then a fast path can be taken. (We generally hope it fails.)
865 // In all cases, GraphKit::control() is updated to the fast path.
866 // The returned value represents the control for the slow path.
867 // The return value is never 'top'; it is either a valid control
868 // or NULL if it is obvious that the slow path can never be taken.
869 // Also, if region and the slow control are not NULL, the slow edge
870 // is appended to the region.
871 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
872 if (stopped()) {
873 // Already short circuited.
874 return NULL;
875 }
876
877 // Build an if node and its projections.
878 // If test is true we take the slow path, which we assume is uncommon.
879 if (_gvn.type(test) == TypeInt::ZERO) {
880 // The slow branch is never taken. No need to build this guard.
881 return NULL;
882 }
883
884 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
885
886 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
887 if (if_slow == top()) {
888 // The slow branch is never taken. No need to build this guard.
889 return NULL;
890 }
891
892 if (region != NULL)
893 region->add_req(if_slow);
894
895 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
896 set_control(if_fast);
897
898 return if_slow;
899 }
900
901 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
902 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
903 }
904 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
905 return generate_guard(test, region, PROB_FAIR);
906 }
907
908 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
909 Node* *pos_index) {
910 if (stopped())
911 return NULL; // already stopped
912 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
913 return NULL; // index is already adequately typed
914 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
915 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
916 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
917 if (is_neg != NULL && pos_index != NULL) {
918 // Emulate effect of Parse::adjust_map_after_if.
919 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
920 ccast->set_req(0, control());
921 (*pos_index) = _gvn.transform(ccast);
922 }
923 return is_neg;
924 }
925
926 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
927 Node* *pos_index) {
928 if (stopped())
929 return NULL; // already stopped
930 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
931 return NULL; // index is already adequately typed
932 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
933 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
934 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
935 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
936 if (is_notp != NULL && pos_index != NULL) {
937 // Emulate effect of Parse::adjust_map_after_if.
938 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
939 ccast->set_req(0, control());
940 (*pos_index) = _gvn.transform(ccast);
941 }
942 return is_notp;
943 }
944
945 // Make sure that 'position' is a valid limit index, in [0..length].
946 // There are two equivalent plans for checking this:
947 // A. (offset + copyLength) unsigned<= arrayLength
948 // B. offset <= (arrayLength - copyLength)
949 // We require that all of the values above, except for the sum and
950 // difference, are already known to be non-negative.
951 // Plan A is robust in the face of overflow, if offset and copyLength
952 // are both hugely positive.
953 //
954 // Plan B is less direct and intuitive, but it does not overflow at
955 // all, since the difference of two non-negatives is always
956 // representable. Whenever Java methods must perform the equivalent
957 // check they generally use Plan B instead of Plan A.
958 // For the moment we use Plan A.
959 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
960 Node* subseq_length,
961 Node* array_length,
962 RegionNode* region) {
963 if (stopped())
964 return NULL; // already stopped
965 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
966 if (zero_offset && subseq_length->eqv_uncast(array_length))
967 return NULL; // common case of whole-array copy
968 Node* last = subseq_length;
969 if (!zero_offset) // last += offset
970 last = _gvn.transform( new (C) AddINode(last, offset));
971 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
972 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
973 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
974 return is_over;
975 }
976
977
978 //--------------------------generate_current_thread--------------------
979 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
980 ciKlass* thread_klass = env()->Thread_klass();
981 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
982 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
983 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
984 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
985 tls_output = thread;
986 return threadObj;
987 }
988
989
990 //------------------------------make_string_method_node------------------------
991 // Helper method for String intrinsic functions. This version is called
992 // with str1 and str2 pointing to String object nodes.
993 //
994 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
995 Node* no_ctrl = NULL;
996
997 // Get start addr of string
998 Node* str1_value = load_String_value(no_ctrl, str1);
999 Node* str1_offset = load_String_offset(no_ctrl, str1);
1000 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1001
1002 // Get length of string 1
1003 Node* str1_len = load_String_length(no_ctrl, str1);
1004
1005 Node* str2_value = load_String_value(no_ctrl, str2);
1006 Node* str2_offset = load_String_offset(no_ctrl, str2);
1007 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1008
1009 Node* str2_len = NULL;
1010 Node* result = NULL;
1011
1012 switch (opcode) {
1013 case Op_StrIndexOf:
1014 // Get length of string 2
1015 str2_len = load_String_length(no_ctrl, str2);
1016
1017 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1018 str1_start, str1_len, str2_start, str2_len);
1019 break;
1020 case Op_StrComp:
1021 // Get length of string 2
1022 str2_len = load_String_length(no_ctrl, str2);
1023
1024 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1025 str1_start, str1_len, str2_start, str2_len);
1026 break;
1027 case Op_StrEquals:
1028 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1029 str1_start, str2_start, str1_len);
1030 break;
1031 default:
1032 ShouldNotReachHere();
1033 return NULL;
1034 }
1035
1036 // All these intrinsics have checks.
1037 C->set_has_split_ifs(true); // Has chance for split-if optimization
1038
1039 return _gvn.transform(result);
1040 }
1041
1042 // Helper method for String intrinsic functions. This version is called
1043 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1044 // to Int nodes containing the lenghts of str1 and str2.
1045 //
1046 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1047 Node* result = NULL;
1048 switch (opcode) {
1049 case Op_StrIndexOf:
1050 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1051 str1_start, cnt1, str2_start, cnt2);
1052 break;
1053 case Op_StrComp:
1054 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1055 str1_start, cnt1, str2_start, cnt2);
1056 break;
1057 case Op_StrEquals:
1058 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1059 str1_start, str2_start, cnt1);
1060 break;
1061 default:
1062 ShouldNotReachHere();
1063 return NULL;
1064 }
1065
1066 // All these intrinsics have checks.
1067 C->set_has_split_ifs(true); // Has chance for split-if optimization
1068
1069 return _gvn.transform(result);
1070 }
1071
1072 //------------------------------inline_string_compareTo------------------------
1073 // public int java.lang.String.compareTo(String anotherString);
1074 bool LibraryCallKit::inline_string_compareTo() {
1075 Node* receiver = null_check(argument(0));
1076 Node* arg = null_check(argument(1));
1077 if (stopped()) {
1078 return true;
1079 }
1080 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1081 return true;
1082 }
1083
1084 //------------------------------inline_string_equals------------------------
1085 bool LibraryCallKit::inline_string_equals() {
1086 Node* receiver = null_check_receiver();
1087 // NOTE: Do not null check argument for String.equals() because spec
1088 // allows to specify NULL as argument.
1089 Node* argument = this->argument(1);
1090 if (stopped()) {
1091 return true;
1092 }
1093
1094 // paths (plus control) merge
1095 RegionNode* region = new (C) RegionNode(5);
1096 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1097
1098 // does source == target string?
1099 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1100 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1101
1102 Node* if_eq = generate_slow_guard(bol, NULL);
1103 if (if_eq != NULL) {
1104 // receiver == argument
1105 phi->init_req(2, intcon(1));
1106 region->init_req(2, if_eq);
1107 }
1108
1109 // get String klass for instanceOf
1110 ciInstanceKlass* klass = env()->String_klass();
1111
1112 if (!stopped()) {
1113 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1114 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1115 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1116
1117 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1118 //instanceOf == true, fallthrough
1119
1120 if (inst_false != NULL) {
1121 phi->init_req(3, intcon(0));
1122 region->init_req(3, inst_false);
1123 }
1124 }
1125
1126 if (!stopped()) {
1127 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1128
1129 // Properly cast the argument to String
1130 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1131 // This path is taken only when argument's type is String:NotNull.
1132 argument = cast_not_null(argument, false);
1133
1134 Node* no_ctrl = NULL;
1135
1136 // Get start addr of receiver
1137 Node* receiver_val = load_String_value(no_ctrl, receiver);
1138 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1139 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1140
1141 // Get length of receiver
1142 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1143
1144 // Get start addr of argument
1145 Node* argument_val = load_String_value(no_ctrl, argument);
1146 Node* argument_offset = load_String_offset(no_ctrl, argument);
1147 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1148
1149 // Get length of argument
1150 Node* argument_cnt = load_String_length(no_ctrl, argument);
1151
1152 // Check for receiver count != argument count
1153 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1154 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1155 Node* if_ne = generate_slow_guard(bol, NULL);
1156 if (if_ne != NULL) {
1157 phi->init_req(4, intcon(0));
1158 region->init_req(4, if_ne);
1159 }
1160
1161 // Check for count == 0 is done by assembler code for StrEquals.
1162
1163 if (!stopped()) {
1164 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1165 phi->init_req(1, equals);
1166 region->init_req(1, control());
1167 }
1168 }
1169
1170 // post merge
1171 set_control(_gvn.transform(region));
1172 record_for_igvn(region);
1173
1174 set_result(_gvn.transform(phi));
1175 return true;
1176 }
1177
1178 //------------------------------inline_array_equals----------------------------
1179 bool LibraryCallKit::inline_array_equals() {
1180 Node* arg1 = argument(0);
1181 Node* arg2 = argument(1);
1182 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1183 return true;
1184 }
1185
1186 // Java version of String.indexOf(constant string)
1187 // class StringDecl {
1188 // StringDecl(char[] ca) {
1189 // offset = 0;
1190 // count = ca.length;
1191 // value = ca;
1192 // }
1193 // int offset;
1194 // int count;
1195 // char[] value;
1196 // }
1197 //
1198 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1199 // int targetOffset, int cache_i, int md2) {
1200 // int cache = cache_i;
1201 // int sourceOffset = string_object.offset;
1202 // int sourceCount = string_object.count;
1203 // int targetCount = target_object.length;
1204 //
1205 // int targetCountLess1 = targetCount - 1;
1206 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1207 //
1208 // char[] source = string_object.value;
1209 // char[] target = target_object;
1210 // int lastChar = target[targetCountLess1];
1211 //
1212 // outer_loop:
1213 // for (int i = sourceOffset; i < sourceEnd; ) {
1214 // int src = source[i + targetCountLess1];
1215 // if (src == lastChar) {
1216 // // With random strings and a 4-character alphabet,
1217 // // reverse matching at this point sets up 0.8% fewer
1218 // // frames, but (paradoxically) makes 0.3% more probes.
1219 // // Since those probes are nearer the lastChar probe,
1220 // // there is may be a net D$ win with reverse matching.
1221 // // But, reversing loop inhibits unroll of inner loop
1222 // // for unknown reason. So, does running outer loop from
1223 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1224 // for (int j = 0; j < targetCountLess1; j++) {
1225 // if (target[targetOffset + j] != source[i+j]) {
1226 // if ((cache & (1 << source[i+j])) == 0) {
1227 // if (md2 < j+1) {
1228 // i += j+1;
1229 // continue outer_loop;
1230 // }
1231 // }
1232 // i += md2;
1233 // continue outer_loop;
1234 // }
1235 // }
1236 // return i - sourceOffset;
1237 // }
1238 // if ((cache & (1 << src)) == 0) {
1239 // i += targetCountLess1;
1240 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1241 // i++;
1242 // }
1243 // return -1;
1244 // }
1245
1246 //------------------------------string_indexOf------------------------
1247 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1248 jint cache_i, jint md2_i) {
1249
1250 Node* no_ctrl = NULL;
1251 float likely = PROB_LIKELY(0.9);
1252 float unlikely = PROB_UNLIKELY(0.9);
1253
1254 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1255
1256 Node* source = load_String_value(no_ctrl, string_object);
1257 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1258 Node* sourceCount = load_String_length(no_ctrl, string_object);
1259
1260 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1261 jint target_length = target_array->length();
1262 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1263 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1264
1265 IdealKit kit(this, false, true);
1266 #define __ kit.
1267 Node* zero = __ ConI(0);
1268 Node* one = __ ConI(1);
1269 Node* cache = __ ConI(cache_i);
1270 Node* md2 = __ ConI(md2_i);
1271 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1272 Node* targetCount = __ ConI(target_length);
1273 Node* targetCountLess1 = __ ConI(target_length - 1);
1274 Node* targetOffset = __ ConI(targetOffset_i);
1275 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1276
1277 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1278 Node* outer_loop = __ make_label(2 /* goto */);
1279 Node* return_ = __ make_label(1);
1280
1281 __ set(rtn,__ ConI(-1));
1282 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1283 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1284 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1285 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1286 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1287 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1288 Node* tpj = __ AddI(targetOffset, __ value(j));
1289 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1290 Node* ipj = __ AddI(__ value(i), __ value(j));
1291 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1292 __ if_then(targ, BoolTest::ne, src2); {
1293 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1294 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1295 __ increment(i, __ AddI(__ value(j), one));
1296 __ goto_(outer_loop);
1297 } __ end_if(); __ dead(j);
1298 }__ end_if(); __ dead(j);
1299 __ increment(i, md2);
1300 __ goto_(outer_loop);
1301 }__ end_if();
1302 __ increment(j, one);
1303 }__ end_loop(); __ dead(j);
1304 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1305 __ goto_(return_);
1306 }__ end_if();
1307 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1308 __ increment(i, targetCountLess1);
1309 }__ end_if();
1310 __ increment(i, one);
1311 __ bind(outer_loop);
1312 }__ end_loop(); __ dead(i);
1313 __ bind(return_);
1314
1315 // Final sync IdealKit and GraphKit.
1316 final_sync(kit);
1317 Node* result = __ value(rtn);
1318 #undef __
1319 C->set_has_loops(true);
1320 return result;
1321 }
1322
1323 //------------------------------inline_string_indexOf------------------------
1324 bool LibraryCallKit::inline_string_indexOf() {
1325 Node* receiver = argument(0);
1326 Node* arg = argument(1);
1327
1328 Node* result;
1329 // Disable the use of pcmpestri until it can be guaranteed that
1330 // the load doesn't cross into the uncommited space.
1331 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1332 UseSSE42Intrinsics) {
1333 // Generate SSE4.2 version of indexOf
1334 // We currently only have match rules that use SSE4.2
1335
1336 receiver = null_check(receiver);
1337 arg = null_check(arg);
1338 if (stopped()) {
1339 return true;
1340 }
1341
1342 ciInstanceKlass* str_klass = env()->String_klass();
1343 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1344
1345 // Make the merge point
1346 RegionNode* result_rgn = new (C) RegionNode(4);
1347 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1348 Node* no_ctrl = NULL;
1349
1350 // Get start addr of source string
1351 Node* source = load_String_value(no_ctrl, receiver);
1352 Node* source_offset = load_String_offset(no_ctrl, receiver);
1353 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1354
1355 // Get length of source string
1356 Node* source_cnt = load_String_length(no_ctrl, receiver);
1357
1358 // Get start addr of substring
1359 Node* substr = load_String_value(no_ctrl, arg);
1360 Node* substr_offset = load_String_offset(no_ctrl, arg);
1361 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1362
1363 // Get length of source string
1364 Node* substr_cnt = load_String_length(no_ctrl, arg);
1365
1366 // Check for substr count > string count
1367 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1368 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1369 Node* if_gt = generate_slow_guard(bol, NULL);
1370 if (if_gt != NULL) {
1371 result_phi->init_req(2, intcon(-1));
1372 result_rgn->init_req(2, if_gt);
1373 }
1374
1375 if (!stopped()) {
1376 // Check for substr count == 0
1377 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1378 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1379 Node* if_zero = generate_slow_guard(bol, NULL);
1380 if (if_zero != NULL) {
1381 result_phi->init_req(3, intcon(0));
1382 result_rgn->init_req(3, if_zero);
1383 }
1384 }
1385
1386 if (!stopped()) {
1387 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1388 result_phi->init_req(1, result);
1389 result_rgn->init_req(1, control());
1390 }
1391 set_control(_gvn.transform(result_rgn));
1392 record_for_igvn(result_rgn);
1393 result = _gvn.transform(result_phi);
1394
1395 } else { // Use LibraryCallKit::string_indexOf
1396 // don't intrinsify if argument isn't a constant string.
1397 if (!arg->is_Con()) {
1398 return false;
1399 }
1400 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1401 if (str_type == NULL) {
1402 return false;
1403 }
1404 ciInstanceKlass* klass = env()->String_klass();
1405 ciObject* str_const = str_type->const_oop();
1406 if (str_const == NULL || str_const->klass() != klass) {
1407 return false;
1408 }
1409 ciInstance* str = str_const->as_instance();
1410 assert(str != NULL, "must be instance");
1411
1412 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1413 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1414
1415 int o;
1416 int c;
1417 if (java_lang_String::has_offset_field()) {
1418 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1419 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1420 } else {
1421 o = 0;
1422 c = pat->length();
1423 }
1424
1425 // constant strings have no offset and count == length which
1426 // simplifies the resulting code somewhat so lets optimize for that.
1427 if (o != 0 || c != pat->length()) {
1428 return false;
1429 }
1430
1431 receiver = null_check(receiver, T_OBJECT);
1432 // NOTE: No null check on the argument is needed since it's a constant String oop.
1433 if (stopped()) {
1434 return true;
1435 }
1436
1437 // The null string as a pattern always returns 0 (match at beginning of string)
1438 if (c == 0) {
1439 set_result(intcon(0));
1440 return true;
1441 }
1442
1443 // Generate default indexOf
1444 jchar lastChar = pat->char_at(o + (c - 1));
1445 int cache = 0;
1446 int i;
1447 for (i = 0; i < c - 1; i++) {
1448 assert(i < pat->length(), "out of range");
1449 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1450 }
1451
1452 int md2 = c;
1453 for (i = 0; i < c - 1; i++) {
1454 assert(i < pat->length(), "out of range");
1455 if (pat->char_at(o + i) == lastChar) {
1456 md2 = (c - 1) - i;
1457 }
1458 }
1459
1460 result = string_indexOf(receiver, pat, o, cache, md2);
1461 }
1462 set_result(result);
1463 return true;
1464 }
1465
1466 //--------------------------round_double_node--------------------------------
1467 // Round a double node if necessary.
1468 Node* LibraryCallKit::round_double_node(Node* n) {
1469 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1470 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1471 return n;
1472 }
1473
1474 //------------------------------inline_math-----------------------------------
1475 // public static double Math.abs(double)
1476 // public static double Math.sqrt(double)
1477 // public static double Math.log(double)
1478 // public static double Math.log10(double)
1479 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1480 Node* arg = round_double_node(argument(0));
1481 Node* n;
1482 switch (id) {
1483 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1484 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1485 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1486 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1487 default: fatal_unexpected_iid(id); break;
1488 }
1489 set_result(_gvn.transform(n));
1490 return true;
1491 }
1492
1493 //------------------------------inline_trig----------------------------------
1494 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1495 // argument reduction which will turn into a fast/slow diamond.
1496 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1497 Node* arg = round_double_node(argument(0));
1498 Node* n = NULL;
1499
1500 switch (id) {
1501 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1502 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1503 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1504 default: fatal_unexpected_iid(id); break;
1505 }
1506 n = _gvn.transform(n);
1507
1508 // Rounding required? Check for argument reduction!
1509 if (Matcher::strict_fp_requires_explicit_rounding) {
1510 static const double pi_4 = 0.7853981633974483;
1511 static const double neg_pi_4 = -0.7853981633974483;
1512 // pi/2 in 80-bit extended precision
1513 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1514 // -pi/2 in 80-bit extended precision
1515 // 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};
1516 // Cutoff value for using this argument reduction technique
1517 //static const double pi_2_minus_epsilon = 1.564660403643354;
1518 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1519
1520 // Pseudocode for sin:
1521 // if (x <= Math.PI / 4.0) {
1522 // if (x >= -Math.PI / 4.0) return fsin(x);
1523 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1524 // } else {
1525 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1526 // }
1527 // return StrictMath.sin(x);
1528
1529 // Pseudocode for cos:
1530 // if (x <= Math.PI / 4.0) {
1531 // if (x >= -Math.PI / 4.0) return fcos(x);
1532 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1533 // } else {
1534 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1535 // }
1536 // return StrictMath.cos(x);
1537
1538 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1539 // requires a special machine instruction to load it. Instead we'll try
1540 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1541 // probably do the math inside the SIN encoding.
1542
1543 // Make the merge point
1544 RegionNode* r = new (C) RegionNode(3);
1545 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1546
1547 // Flatten arg so we need only 1 test
1548 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1549 // Node for PI/4 constant
1550 Node *pi4 = makecon(TypeD::make(pi_4));
1551 // Check PI/4 : abs(arg)
1552 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1553 // Check: If PI/4 < abs(arg) then go slow
1554 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1555 // Branch either way
1556 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1557 set_control(opt_iff(r,iff));
1558
1559 // Set fast path result
1560 phi->init_req(2, n);
1561
1562 // Slow path - non-blocking leaf call
1563 Node* call = NULL;
1564 switch (id) {
1565 case vmIntrinsics::_dsin:
1566 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1567 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1568 "Sin", NULL, arg, top());
1569 break;
1570 case vmIntrinsics::_dcos:
1571 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1572 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1573 "Cos", NULL, arg, top());
1574 break;
1575 case vmIntrinsics::_dtan:
1576 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1577 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1578 "Tan", NULL, arg, top());
1579 break;
1580 }
1581 assert(control()->in(0) == call, "");
1582 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1583 r->init_req(1, control());
1584 phi->init_req(1, slow_result);
1585
1586 // Post-merge
1587 set_control(_gvn.transform(r));
1588 record_for_igvn(r);
1589 n = _gvn.transform(phi);
1590
1591 C->set_has_split_ifs(true); // Has chance for split-if optimization
1592 }
1593 set_result(n);
1594 return true;
1595 }
1596
1597 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1598 //-------------------
1599 //result=(result.isNaN())? funcAddr():result;
1600 // Check: If isNaN() by checking result!=result? then either trap
1601 // or go to runtime
1602 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1603 // Build the boolean node
1604 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1605
1606 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1607 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1608 // The pow or exp intrinsic returned a NaN, which requires a call
1609 // to the runtime. Recompile with the runtime call.
1610 uncommon_trap(Deoptimization::Reason_intrinsic,
1611 Deoptimization::Action_make_not_entrant);
1612 }
1613 set_result(result);
1614 } else {
1615 // If this inlining ever returned NaN in the past, we compile a call
1616 // to the runtime to properly handle corner cases
1617
1618 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1619 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1620 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1621
1622 if (!if_slow->is_top()) {
1623 RegionNode* result_region = new (C) RegionNode(3);
1624 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1625
1626 result_region->init_req(1, if_fast);
1627 result_val->init_req(1, result);
1628
1629 set_control(if_slow);
1630
1631 const TypePtr* no_memory_effects = NULL;
1632 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1633 no_memory_effects,
1634 x, top(), y, y ? top() : NULL);
1635 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1636 #ifdef ASSERT
1637 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1638 assert(value_top == top(), "second value must be top");
1639 #endif
1640
1641 result_region->init_req(2, control());
1642 result_val->init_req(2, value);
1643 set_result(result_region, result_val);
1644 } else {
1645 set_result(result);
1646 }
1647 }
1648 }
1649
1650 //------------------------------inline_exp-------------------------------------
1651 // Inline exp instructions, if possible. The Intel hardware only misses
1652 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1653 bool LibraryCallKit::inline_exp() {
1654 Node* arg = round_double_node(argument(0));
1655 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1656
1657 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1658
1659 C->set_has_split_ifs(true); // Has chance for split-if optimization
1660 return true;
1661 }
1662
1663 //------------------------------inline_pow-------------------------------------
1664 // Inline power instructions, if possible.
1665 bool LibraryCallKit::inline_pow() {
1666 // Pseudocode for pow
1667 // if (x <= 0.0) {
1668 // long longy = (long)y;
1669 // if ((double)longy == y) { // if y is long
1670 // if (y + 1 == y) longy = 0; // huge number: even
1671 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1672 // } else {
1673 // result = NaN;
1674 // }
1675 // } else {
1676 // result = DPow(x,y);
1677 // }
1678 // if (result != result)? {
1679 // result = uncommon_trap() or runtime_call();
1680 // }
1681 // return result;
1682
1683 Node* x = round_double_node(argument(0));
1684 Node* y = round_double_node(argument(2));
1685
1686 Node* result = NULL;
1687
1688 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1689 // Short form: skip the fancy tests and just check for NaN result.
1690 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1691 } else {
1692 // If this inlining ever returned NaN in the past, include all
1693 // checks + call to the runtime.
1694
1695 // Set the merge point for If node with condition of (x <= 0.0)
1696 // There are four possible paths to region node and phi node
1697 RegionNode *r = new (C) RegionNode(4);
1698 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1699
1700 // Build the first if node: if (x <= 0.0)
1701 // Node for 0 constant
1702 Node *zeronode = makecon(TypeD::ZERO);
1703 // Check x:0
1704 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1705 // Check: If (x<=0) then go complex path
1706 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1707 // Branch either way
1708 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1709 // Fast path taken; set region slot 3
1710 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1711 r->init_req(3,fast_taken); // Capture fast-control
1712
1713 // Fast path not-taken, i.e. slow path
1714 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1715
1716 // Set fast path result
1717 Node *fast_result = _gvn.transform( new (C) PowDNode(C, control(), x, y) );
1718 phi->init_req(3, fast_result);
1719
1720 // Complex path
1721 // Build the second if node (if y is long)
1722 // Node for (long)y
1723 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1724 // Node for (double)((long) y)
1725 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1726 // Check (double)((long) y) : y
1727 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1728 // Check if (y isn't long) then go to slow path
1729
1730 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1731 // Branch either way
1732 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1733 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1734
1735 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1736
1737 // Calculate DPow(abs(x), y)*(1 & (long)y)
1738 // Node for constant 1
1739 Node *conone = longcon(1);
1740 // 1& (long)y
1741 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1742
1743 // A huge number is always even. Detect a huge number by checking
1744 // if y + 1 == y and set integer to be tested for parity to 0.
1745 // Required for corner case:
1746 // (long)9.223372036854776E18 = max_jlong
1747 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1748 // max_jlong is odd but 9.223372036854776E18 is even
1749 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1750 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1751 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1752 Node* correctedsign = NULL;
1753 if (ConditionalMoveLimit != 0) {
1754 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1755 } else {
1756 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1757 RegionNode *r = new (C) RegionNode(3);
1758 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1759 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1760 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1761 phi->init_req(1, signnode);
1762 phi->init_req(2, longcon(0));
1763 correctedsign = _gvn.transform(phi);
1764 ylong_path = _gvn.transform(r);
1765 record_for_igvn(r);
1766 }
1767
1768 // zero node
1769 Node *conzero = longcon(0);
1770 // Check (1&(long)y)==0?
1771 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1772 // Check if (1&(long)y)!=0?, if so the result is negative
1773 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1774 // abs(x)
1775 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1776 // abs(x)^y
1777 Node *absxpowy = _gvn.transform( new (C) PowDNode(C, control(), absx, y) );
1778 // -abs(x)^y
1779 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1780 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1781 Node *signresult = NULL;
1782 if (ConditionalMoveLimit != 0) {
1783 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1784 } else {
1785 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1786 RegionNode *r = new (C) RegionNode(3);
1787 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1788 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1789 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1790 phi->init_req(1, absxpowy);
1791 phi->init_req(2, negabsxpowy);
1792 signresult = _gvn.transform(phi);
1793 ylong_path = _gvn.transform(r);
1794 record_for_igvn(r);
1795 }
1796 // Set complex path fast result
1797 r->init_req(2, ylong_path);
1798 phi->init_req(2, signresult);
1799
1800 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1801 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1802 r->init_req(1,slow_path);
1803 phi->init_req(1,slow_result);
1804
1805 // Post merge
1806 set_control(_gvn.transform(r));
1807 record_for_igvn(r);
1808 result = _gvn.transform(phi);
1809 }
1810
1811 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1812
1813 C->set_has_split_ifs(true); // Has chance for split-if optimization
1814 return true;
1815 }
1816
1817 //------------------------------runtime_math-----------------------------
1818 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1819 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1820 "must be (DD)D or (D)D type");
1821
1822 // Inputs
1823 Node* a = round_double_node(argument(0));
1824 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1825
1826 const TypePtr* no_memory_effects = NULL;
1827 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1828 no_memory_effects,
1829 a, top(), b, b ? top() : NULL);
1830 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1831 #ifdef ASSERT
1832 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1833 assert(value_top == top(), "second value must be top");
1834 #endif
1835
1836 set_result(value);
1837 return true;
1838 }
1839
1840 //------------------------------inline_math_native-----------------------------
1841 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1842 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1843 switch (id) {
1844 // These intrinsics are not properly supported on all hardware
1845 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1846 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1847 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1848 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1849 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1850 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1851
1852 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1853 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1854 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1855 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1856
1857 // These intrinsics are supported on all hardware
1858 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1859 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1860
1861 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1862 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1863 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1864 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1865 #undef FN_PTR
1866
1867 // These intrinsics are not yet correctly implemented
1868 case vmIntrinsics::_datan2:
1869 return false;
1870
1871 default:
1872 fatal_unexpected_iid(id);
1873 return false;
1874 }
1875 }
1876
1877 static bool is_simple_name(Node* n) {
1878 return (n->req() == 1 // constant
1879 || (n->is_Type() && n->as_Type()->type()->singleton())
1880 || n->is_Proj() // parameter or return value
1881 || n->is_Phi() // local of some sort
1882 );
1883 }
1884
1885 //----------------------------inline_min_max-----------------------------------
1886 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1887 set_result(generate_min_max(id, argument(0), argument(1)));
1888 return true;
1889 }
1890
1891 Node*
1892 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1893 // These are the candidate return value:
1894 Node* xvalue = x0;
1895 Node* yvalue = y0;
1896
1897 if (xvalue == yvalue) {
1898 return xvalue;
1899 }
1900
1901 bool want_max = (id == vmIntrinsics::_max);
1902
1903 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1904 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1905 if (txvalue == NULL || tyvalue == NULL) return top();
1906 // This is not really necessary, but it is consistent with a
1907 // hypothetical MaxINode::Value method:
1908 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1909
1910 // %%% This folding logic should (ideally) be in a different place.
1911 // Some should be inside IfNode, and there to be a more reliable
1912 // transformation of ?: style patterns into cmoves. We also want
1913 // more powerful optimizations around cmove and min/max.
1914
1915 // Try to find a dominating comparison of these guys.
1916 // It can simplify the index computation for Arrays.copyOf
1917 // and similar uses of System.arraycopy.
1918 // First, compute the normalized version of CmpI(x, y).
1919 int cmp_op = Op_CmpI;
1920 Node* xkey = xvalue;
1921 Node* ykey = yvalue;
1922 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1923 if (ideal_cmpxy->is_Cmp()) {
1924 // E.g., if we have CmpI(length - offset, count),
1925 // it might idealize to CmpI(length, count + offset)
1926 cmp_op = ideal_cmpxy->Opcode();
1927 xkey = ideal_cmpxy->in(1);
1928 ykey = ideal_cmpxy->in(2);
1929 }
1930
1931 // Start by locating any relevant comparisons.
1932 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1933 Node* cmpxy = NULL;
1934 Node* cmpyx = NULL;
1935 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1936 Node* cmp = start_from->fast_out(k);
1937 if (cmp->outcnt() > 0 && // must have prior uses
1938 cmp->in(0) == NULL && // must be context-independent
1939 cmp->Opcode() == cmp_op) { // right kind of compare
1940 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1941 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1942 }
1943 }
1944
1945 const int NCMPS = 2;
1946 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1947 int cmpn;
1948 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1949 if (cmps[cmpn] != NULL) break; // find a result
1950 }
1951 if (cmpn < NCMPS) {
1952 // Look for a dominating test that tells us the min and max.
1953 int depth = 0; // Limit search depth for speed
1954 Node* dom = control();
1955 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1956 if (++depth >= 100) break;
1957 Node* ifproj = dom;
1958 if (!ifproj->is_Proj()) continue;
1959 Node* iff = ifproj->in(0);
1960 if (!iff->is_If()) continue;
1961 Node* bol = iff->in(1);
1962 if (!bol->is_Bool()) continue;
1963 Node* cmp = bol->in(1);
1964 if (cmp == NULL) continue;
1965 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1966 if (cmps[cmpn] == cmp) break;
1967 if (cmpn == NCMPS) continue;
1968 BoolTest::mask btest = bol->as_Bool()->_test._test;
1969 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1970 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1971 // At this point, we know that 'x btest y' is true.
1972 switch (btest) {
1973 case BoolTest::eq:
1974 // They are proven equal, so we can collapse the min/max.
1975 // Either value is the answer. Choose the simpler.
1976 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1977 return yvalue;
1978 return xvalue;
1979 case BoolTest::lt: // x < y
1980 case BoolTest::le: // x <= y
1981 return (want_max ? yvalue : xvalue);
1982 case BoolTest::gt: // x > y
1983 case BoolTest::ge: // x >= y
1984 return (want_max ? xvalue : yvalue);
1985 }
1986 }
1987 }
1988
1989 // We failed to find a dominating test.
1990 // Let's pick a test that might GVN with prior tests.
1991 Node* best_bol = NULL;
1992 BoolTest::mask best_btest = BoolTest::illegal;
1993 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1994 Node* cmp = cmps[cmpn];
1995 if (cmp == NULL) continue;
1996 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1997 Node* bol = cmp->fast_out(j);
1998 if (!bol->is_Bool()) continue;
1999 BoolTest::mask btest = bol->as_Bool()->_test._test;
2000 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2001 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2002 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2003 best_bol = bol->as_Bool();
2004 best_btest = btest;
2005 }
2006 }
2007 }
2008
2009 Node* answer_if_true = NULL;
2010 Node* answer_if_false = NULL;
2011 switch (best_btest) {
2012 default:
2013 if (cmpxy == NULL)
2014 cmpxy = ideal_cmpxy;
2015 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2016 // and fall through:
2017 case BoolTest::lt: // x < y
2018 case BoolTest::le: // x <= y
2019 answer_if_true = (want_max ? yvalue : xvalue);
2020 answer_if_false = (want_max ? xvalue : yvalue);
2021 break;
2022 case BoolTest::gt: // x > y
2023 case BoolTest::ge: // x >= y
2024 answer_if_true = (want_max ? xvalue : yvalue);
2025 answer_if_false = (want_max ? yvalue : xvalue);
2026 break;
2027 }
2028
2029 jint hi, lo;
2030 if (want_max) {
2031 // We can sharpen the minimum.
2032 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2033 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2034 } else {
2035 // We can sharpen the maximum.
2036 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2037 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2038 }
2039
2040 // Use a flow-free graph structure, to avoid creating excess control edges
2041 // which could hinder other optimizations.
2042 // Since Math.min/max is often used with arraycopy, we want
2043 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2044 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2045 answer_if_false, answer_if_true,
2046 TypeInt::make(lo, hi, widen));
2047
2048 return _gvn.transform(cmov);
2049
2050 /*
2051 // This is not as desirable as it may seem, since Min and Max
2052 // nodes do not have a full set of optimizations.
2053 // And they would interfere, anyway, with 'if' optimizations
2054 // and with CMoveI canonical forms.
2055 switch (id) {
2056 case vmIntrinsics::_min:
2057 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2058 case vmIntrinsics::_max:
2059 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2060 default:
2061 ShouldNotReachHere();
2062 }
2063 */
2064 }
2065
2066 inline int
2067 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2068 const TypePtr* base_type = TypePtr::NULL_PTR;
2069 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2070 if (base_type == NULL) {
2071 // Unknown type.
2072 return Type::AnyPtr;
2073 } else if (base_type == TypePtr::NULL_PTR) {
2074 // Since this is a NULL+long form, we have to switch to a rawptr.
2075 base = _gvn.transform( new (C) CastX2PNode(offset) );
2076 offset = MakeConX(0);
2077 return Type::RawPtr;
2078 } else if (base_type->base() == Type::RawPtr) {
2079 return Type::RawPtr;
2080 } else if (base_type->isa_oopptr()) {
2081 // Base is never null => always a heap address.
2082 if (base_type->ptr() == TypePtr::NotNull) {
2083 return Type::OopPtr;
2084 }
2085 // Offset is small => always a heap address.
2086 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2087 if (offset_type != NULL &&
2088 base_type->offset() == 0 && // (should always be?)
2089 offset_type->_lo >= 0 &&
2090 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2091 return Type::OopPtr;
2092 }
2093 // Otherwise, it might either be oop+off or NULL+addr.
2094 return Type::AnyPtr;
2095 } else {
2096 // No information:
2097 return Type::AnyPtr;
2098 }
2099 }
2100
2101 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2102 int kind = classify_unsafe_addr(base, offset);
2103 if (kind == Type::RawPtr) {
2104 return basic_plus_adr(top(), base, offset);
2105 } else {
2106 return basic_plus_adr(base, offset);
2107 }
2108 }
2109
2110 //--------------------------inline_number_methods-----------------------------
2111 // inline int Integer.numberOfLeadingZeros(int)
2112 // inline int Long.numberOfLeadingZeros(long)
2113 //
2114 // inline int Integer.numberOfTrailingZeros(int)
2115 // inline int Long.numberOfTrailingZeros(long)
2116 //
2117 // inline int Integer.bitCount(int)
2118 // inline int Long.bitCount(long)
2119 //
2120 // inline char Character.reverseBytes(char)
2121 // inline short Short.reverseBytes(short)
2122 // inline int Integer.reverseBytes(int)
2123 // inline long Long.reverseBytes(long)
2124 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2125 Node* arg = argument(0);
2126 Node* n;
2127 switch (id) {
2128 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2129 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2130 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2131 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2132 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2133 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2134 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2135 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2136 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2137 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2138 default: fatal_unexpected_iid(id); break;
2139 }
2140 set_result(_gvn.transform(n));
2141 return true;
2142 }
2143
2144 //----------------------------inline_unsafe_access----------------------------
2145
2146 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2147
2148 // Helper that guards and inserts a pre-barrier.
2149 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2150 Node* pre_val, bool need_mem_bar) {
2151 // We could be accessing the referent field of a reference object. If so, when G1
2152 // is enabled, we need to log the value in the referent field in an SATB buffer.
2153 // This routine performs some compile time filters and generates suitable
2154 // runtime filters that guard the pre-barrier code.
2155 // Also add memory barrier for non volatile load from the referent field
2156 // to prevent commoning of loads across safepoint.
2157 if (!UseG1GC && !need_mem_bar)
2158 return;
2159
2160 // Some compile time checks.
2161
2162 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2163 const TypeX* otype = offset->find_intptr_t_type();
2164 if (otype != NULL && otype->is_con() &&
2165 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2166 // Constant offset but not the reference_offset so just return
2167 return;
2168 }
2169
2170 // We only need to generate the runtime guards for instances.
2171 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2172 if (btype != NULL) {
2173 if (btype->isa_aryptr()) {
2174 // Array type so nothing to do
2175 return;
2176 }
2177
2178 const TypeInstPtr* itype = btype->isa_instptr();
2179 if (itype != NULL) {
2180 // Can the klass of base_oop be statically determined to be
2181 // _not_ a sub-class of Reference and _not_ Object?
2182 ciKlass* klass = itype->klass();
2183 if ( klass->is_loaded() &&
2184 !klass->is_subtype_of(env()->Reference_klass()) &&
2185 !env()->Object_klass()->is_subtype_of(klass)) {
2186 return;
2187 }
2188 }
2189 }
2190
2191 // The compile time filters did not reject base_oop/offset so
2192 // we need to generate the following runtime filters
2193 //
2194 // if (offset == java_lang_ref_Reference::_reference_offset) {
2195 // if (instance_of(base, java.lang.ref.Reference)) {
2196 // pre_barrier(_, pre_val, ...);
2197 // }
2198 // }
2199
2200 float likely = PROB_LIKELY( 0.999);
2201 float unlikely = PROB_UNLIKELY(0.999);
2202
2203 IdealKit ideal(this);
2204 #define __ ideal.
2205
2206 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2207
2208 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2209 // Update graphKit memory and control from IdealKit.
2210 sync_kit(ideal);
2211
2212 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2213 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2214
2215 // Update IdealKit memory and control from graphKit.
2216 __ sync_kit(this);
2217
2218 Node* one = __ ConI(1);
2219 // is_instof == 0 if base_oop == NULL
2220 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2221
2222 // Update graphKit from IdeakKit.
2223 sync_kit(ideal);
2224
2225 // Use the pre-barrier to record the value in the referent field
2226 pre_barrier(false /* do_load */,
2227 __ ctrl(),
2228 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2229 pre_val /* pre_val */,
2230 T_OBJECT);
2231 if (need_mem_bar) {
2232 // Add memory barrier to prevent commoning reads from this field
2233 // across safepoint since GC can change its value.
2234 insert_mem_bar(Op_MemBarCPUOrder);
2235 }
2236 // Update IdealKit from graphKit.
2237 __ sync_kit(this);
2238
2239 } __ end_if(); // _ref_type != ref_none
2240 } __ end_if(); // offset == referent_offset
2241
2242 // Final sync IdealKit and GraphKit.
2243 final_sync(ideal);
2244 #undef __
2245 }
2246
2247
2248 // Interpret Unsafe.fieldOffset cookies correctly:
2249 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2250
2251 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2252 // Attempt to infer a sharper value type from the offset and base type.
2253 ciKlass* sharpened_klass = NULL;
2254
2255 // See if it is an instance field, with an object type.
2256 if (alias_type->field() != NULL) {
2257 assert(!is_native_ptr, "native pointer op cannot use a java address");
2258 if (alias_type->field()->type()->is_klass()) {
2259 sharpened_klass = alias_type->field()->type()->as_klass();
2260 }
2261 }
2262
2263 // See if it is a narrow oop array.
2264 if (adr_type->isa_aryptr()) {
2265 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2266 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2267 if (elem_type != NULL) {
2268 sharpened_klass = elem_type->klass();
2269 }
2270 }
2271 }
2272
2273 // The sharpened class might be unloaded if there is no class loader
2274 // contraint in place.
2275 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2276 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2277
2278 #ifndef PRODUCT
2279 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2280 tty->print(" from base type: "); adr_type->dump();
2281 tty->print(" sharpened value: "); tjp->dump();
2282 }
2283 #endif
2284 // Sharpen the value type.
2285 return tjp;
2286 }
2287 return NULL;
2288 }
2289
2290 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2291 if (callee()->is_static()) return false; // caller must have the capability!
2292
2293 #ifndef PRODUCT
2294 {
2295 ResourceMark rm;
2296 // Check the signatures.
2297 ciSignature* sig = callee()->signature();
2298 #ifdef ASSERT
2299 if (!is_store) {
2300 // Object getObject(Object base, int/long offset), etc.
2301 BasicType rtype = sig->return_type()->basic_type();
2302 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2303 rtype = T_ADDRESS; // it is really a C void*
2304 assert(rtype == type, "getter must return the expected value");
2305 if (!is_native_ptr) {
2306 assert(sig->count() == 2, "oop getter has 2 arguments");
2307 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2308 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2309 } else {
2310 assert(sig->count() == 1, "native getter has 1 argument");
2311 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2312 }
2313 } else {
2314 // void putObject(Object base, int/long offset, Object x), etc.
2315 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2316 if (!is_native_ptr) {
2317 assert(sig->count() == 3, "oop putter has 3 arguments");
2318 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2319 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2320 } else {
2321 assert(sig->count() == 2, "native putter has 2 arguments");
2322 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2323 }
2324 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2325 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2326 vtype = T_ADDRESS; // it is really a C void*
2327 assert(vtype == type, "putter must accept the expected value");
2328 }
2329 #endif // ASSERT
2330 }
2331 #endif //PRODUCT
2332
2333 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2334
2335 Node* receiver = argument(0); // type: oop
2336
2337 // Build address expression. See the code in inline_unsafe_prefetch.
2338 Node* adr;
2339 Node* heap_base_oop = top();
2340 Node* offset = top();
2341 Node* val;
2342
2343 if (!is_native_ptr) {
2344 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2345 Node* base = argument(1); // type: oop
2346 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2347 offset = argument(2); // type: long
2348 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2349 // to be plain byte offsets, which are also the same as those accepted
2350 // by oopDesc::field_base.
2351 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2352 "fieldOffset must be byte-scaled");
2353 // 32-bit machines ignore the high half!
2354 offset = ConvL2X(offset);
2355 adr = make_unsafe_address(base, offset);
2356 heap_base_oop = base;
2357 val = is_store ? argument(4) : NULL;
2358 } else {
2359 Node* ptr = argument(1); // type: long
2360 ptr = ConvL2X(ptr); // adjust Java long to machine word
2361 adr = make_unsafe_address(NULL, ptr);
2362 val = is_store ? argument(3) : NULL;
2363 }
2364
2365 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2366
2367 // First guess at the value type.
2368 const Type *value_type = Type::get_const_basic_type(type);
2369
2370 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2371 // there was not enough information to nail it down.
2372 Compile::AliasType* alias_type = C->alias_type(adr_type);
2373 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2374
2375 // We will need memory barriers unless we can determine a unique
2376 // alias category for this reference. (Note: If for some reason
2377 // the barriers get omitted and the unsafe reference begins to "pollute"
2378 // the alias analysis of the rest of the graph, either Compile::can_alias
2379 // or Compile::must_alias will throw a diagnostic assert.)
2380 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2381
2382 // If we are reading the value of the referent field of a Reference
2383 // object (either by using Unsafe directly or through reflection)
2384 // then, if G1 is enabled, we need to record the referent in an
2385 // SATB log buffer using the pre-barrier mechanism.
2386 // Also we need to add memory barrier to prevent commoning reads
2387 // from this field across safepoint since GC can change its value.
2388 bool need_read_barrier = !is_native_ptr && !is_store &&
2389 offset != top() && heap_base_oop != top();
2390
2391 if (!is_store && type == T_OBJECT) {
2392 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2393 if (tjp != NULL) {
2394 value_type = tjp;
2395 }
2396 }
2397
2398 receiver = null_check(receiver);
2399 if (stopped()) {
2400 return true;
2401 }
2402 // Heap pointers get a null-check from the interpreter,
2403 // as a courtesy. However, this is not guaranteed by Unsafe,
2404 // and it is not possible to fully distinguish unintended nulls
2405 // from intended ones in this API.
2406
2407 if (is_volatile) {
2408 // We need to emit leading and trailing CPU membars (see below) in
2409 // addition to memory membars when is_volatile. This is a little
2410 // too strong, but avoids the need to insert per-alias-type
2411 // volatile membars (for stores; compare Parse::do_put_xxx), which
2412 // we cannot do effectively here because we probably only have a
2413 // rough approximation of type.
2414 need_mem_bar = true;
2415 // For Stores, place a memory ordering barrier now.
2416 if (is_store)
2417 insert_mem_bar(Op_MemBarRelease);
2418 }
2419
2420 // Memory barrier to prevent normal and 'unsafe' accesses from
2421 // bypassing each other. Happens after null checks, so the
2422 // exception paths do not take memory state from the memory barrier,
2423 // so there's no problems making a strong assert about mixing users
2424 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2425 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2426 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2427
2428 if (!is_store) {
2429 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2430 // load value
2431 switch (type) {
2432 case T_BOOLEAN:
2433 case T_CHAR:
2434 case T_BYTE:
2435 case T_SHORT:
2436 case T_INT:
2437 case T_LONG:
2438 case T_FLOAT:
2439 case T_DOUBLE:
2440 break;
2441 case T_OBJECT:
2442 if (need_read_barrier) {
2443 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2444 }
2445 break;
2446 case T_ADDRESS:
2447 // Cast to an int type.
2448 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2449 p = ConvX2L(p);
2450 break;
2451 default:
2452 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2453 break;
2454 }
2455 // The load node has the control of the preceding MemBarCPUOrder. All
2456 // following nodes will have the control of the MemBarCPUOrder inserted at
2457 // the end of this method. So, pushing the load onto the stack at a later
2458 // point is fine.
2459 set_result(p);
2460 } else {
2461 // place effect of store into memory
2462 switch (type) {
2463 case T_DOUBLE:
2464 val = dstore_rounding(val);
2465 break;
2466 case T_ADDRESS:
2467 // Repackage the long as a pointer.
2468 val = ConvL2X(val);
2469 val = _gvn.transform( new (C) CastX2PNode(val) );
2470 break;
2471 }
2472
2473 if (type != T_OBJECT ) {
2474 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2475 } else {
2476 // Possibly an oop being stored to Java heap or native memory
2477 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2478 // oop to Java heap.
2479 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2480 } else {
2481 // We can't tell at compile time if we are storing in the Java heap or outside
2482 // of it. So we need to emit code to conditionally do the proper type of
2483 // store.
2484
2485 IdealKit ideal(this);
2486 #define __ ideal.
2487 // QQQ who knows what probability is here??
2488 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2489 // Sync IdealKit and graphKit.
2490 sync_kit(ideal);
2491 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2492 // Update IdealKit memory.
2493 __ sync_kit(this);
2494 } __ else_(); {
2495 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2496 } __ end_if();
2497 // Final sync IdealKit and GraphKit.
2498 final_sync(ideal);
2499 #undef __
2500 }
2501 }
2502 }
2503
2504 if (is_volatile) {
2505 if (!is_store)
2506 insert_mem_bar(Op_MemBarAcquire);
2507 else
2508 insert_mem_bar(Op_MemBarVolatile);
2509 }
2510
2511 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2512
2513 return true;
2514 }
2515
2516 //----------------------------inline_unsafe_prefetch----------------------------
2517
2518 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2519 #ifndef PRODUCT
2520 {
2521 ResourceMark rm;
2522 // Check the signatures.
2523 ciSignature* sig = callee()->signature();
2524 #ifdef ASSERT
2525 // Object getObject(Object base, int/long offset), etc.
2526 BasicType rtype = sig->return_type()->basic_type();
2527 if (!is_native_ptr) {
2528 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2529 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2530 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2531 } else {
2532 assert(sig->count() == 1, "native prefetch has 1 argument");
2533 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2534 }
2535 #endif // ASSERT
2536 }
2537 #endif // !PRODUCT
2538
2539 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2540
2541 const int idx = is_static ? 0 : 1;
2542 if (!is_static) {
2543 null_check_receiver();
2544 if (stopped()) {
2545 return true;
2546 }
2547 }
2548
2549 // Build address expression. See the code in inline_unsafe_access.
2550 Node *adr;
2551 if (!is_native_ptr) {
2552 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2553 Node* base = argument(idx + 0); // type: oop
2554 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2555 Node* offset = argument(idx + 1); // type: long
2556 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2557 // to be plain byte offsets, which are also the same as those accepted
2558 // by oopDesc::field_base.
2559 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2560 "fieldOffset must be byte-scaled");
2561 // 32-bit machines ignore the high half!
2562 offset = ConvL2X(offset);
2563 adr = make_unsafe_address(base, offset);
2564 } else {
2565 Node* ptr = argument(idx + 0); // type: long
2566 ptr = ConvL2X(ptr); // adjust Java long to machine word
2567 adr = make_unsafe_address(NULL, ptr);
2568 }
2569
2570 // Generate the read or write prefetch
2571 Node *prefetch;
2572 if (is_store) {
2573 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2574 } else {
2575 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2576 }
2577 prefetch->init_req(0, control());
2578 set_i_o(_gvn.transform(prefetch));
2579
2580 return true;
2581 }
2582
2583 //----------------------------inline_unsafe_load_store----------------------------
2584 // This method serves a couple of different customers (depending on LoadStoreKind):
2585 //
2586 // LS_cmpxchg:
2587 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2588 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2589 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2590 //
2591 // LS_xadd:
2592 // public int getAndAddInt( Object o, long offset, int delta)
2593 // public long getAndAddLong(Object o, long offset, long delta)
2594 //
2595 // LS_xchg:
2596 // int getAndSet(Object o, long offset, int newValue)
2597 // long getAndSet(Object o, long offset, long newValue)
2598 // Object getAndSet(Object o, long offset, Object newValue)
2599 //
2600 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2601 // This basic scheme here is the same as inline_unsafe_access, but
2602 // differs in enough details that combining them would make the code
2603 // overly confusing. (This is a true fact! I originally combined
2604 // them, but even I was confused by it!) As much code/comments as
2605 // possible are retained from inline_unsafe_access though to make
2606 // the correspondences clearer. - dl
2607
2608 if (callee()->is_static()) return false; // caller must have the capability!
2609
2610 #ifndef PRODUCT
2611 BasicType rtype;
2612 {
2613 ResourceMark rm;
2614 // Check the signatures.
2615 ciSignature* sig = callee()->signature();
2616 rtype = sig->return_type()->basic_type();
2617 if (kind == LS_xadd || kind == LS_xchg) {
2618 // Check the signatures.
2619 #ifdef ASSERT
2620 assert(rtype == type, "get and set must return the expected type");
2621 assert(sig->count() == 3, "get and set has 3 arguments");
2622 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2623 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2624 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2625 #endif // ASSERT
2626 } else if (kind == LS_cmpxchg) {
2627 // Check the signatures.
2628 #ifdef ASSERT
2629 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2630 assert(sig->count() == 4, "CAS has 4 arguments");
2631 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2632 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2633 #endif // ASSERT
2634 } else {
2635 ShouldNotReachHere();
2636 }
2637 }
2638 #endif //PRODUCT
2639
2640 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2641
2642 // Get arguments:
2643 Node* receiver = NULL;
2644 Node* base = NULL;
2645 Node* offset = NULL;
2646 Node* oldval = NULL;
2647 Node* newval = NULL;
2648 if (kind == LS_cmpxchg) {
2649 const bool two_slot_type = type2size[type] == 2;
2650 receiver = argument(0); // type: oop
2651 base = argument(1); // type: oop
2652 offset = argument(2); // type: long
2653 oldval = argument(4); // type: oop, int, or long
2654 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2655 } else if (kind == LS_xadd || kind == LS_xchg){
2656 receiver = argument(0); // type: oop
2657 base = argument(1); // type: oop
2658 offset = argument(2); // type: long
2659 oldval = NULL;
2660 newval = argument(4); // type: oop, int, or long
2661 }
2662
2663 // Null check receiver.
2664 receiver = null_check(receiver);
2665 if (stopped()) {
2666 return true;
2667 }
2668
2669 // Build field offset expression.
2670 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2671 // to be plain byte offsets, which are also the same as those accepted
2672 // by oopDesc::field_base.
2673 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2674 // 32-bit machines ignore the high half of long offsets
2675 offset = ConvL2X(offset);
2676 Node* adr = make_unsafe_address(base, offset);
2677 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2678
2679 // For CAS, unlike inline_unsafe_access, there seems no point in
2680 // trying to refine types. Just use the coarse types here.
2681 const Type *value_type = Type::get_const_basic_type(type);
2682 Compile::AliasType* alias_type = C->alias_type(adr_type);
2683 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2684
2685 if (kind == LS_xchg && type == T_OBJECT) {
2686 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2687 if (tjp != NULL) {
2688 value_type = tjp;
2689 }
2690 }
2691
2692 int alias_idx = C->get_alias_index(adr_type);
2693
2694 // Memory-model-wise, a LoadStore acts like a little synchronized
2695 // block, so needs barriers on each side. These don't translate
2696 // into actual barriers on most machines, but we still need rest of
2697 // compiler to respect ordering.
2698
2699 insert_mem_bar(Op_MemBarRelease);
2700 insert_mem_bar(Op_MemBarCPUOrder);
2701
2702 // 4984716: MemBars must be inserted before this
2703 // memory node in order to avoid a false
2704 // dependency which will confuse the scheduler.
2705 Node *mem = memory(alias_idx);
2706
2707 // For now, we handle only those cases that actually exist: ints,
2708 // longs, and Object. Adding others should be straightforward.
2709 Node* load_store;
2710 switch(type) {
2711 case T_INT:
2712 if (kind == LS_xadd) {
2713 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2714 } else if (kind == LS_xchg) {
2715 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2716 } else if (kind == LS_cmpxchg) {
2717 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2718 } else {
2719 ShouldNotReachHere();
2720 }
2721 break;
2722 case T_LONG:
2723 if (kind == LS_xadd) {
2724 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2725 } else if (kind == LS_xchg) {
2726 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2727 } else if (kind == LS_cmpxchg) {
2728 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2729 } else {
2730 ShouldNotReachHere();
2731 }
2732 break;
2733 case T_OBJECT:
2734 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2735 // could be delayed during Parse (for example, in adjust_map_after_if()).
2736 // Execute transformation here to avoid barrier generation in such case.
2737 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2738 newval = _gvn.makecon(TypePtr::NULL_PTR);
2739
2740 // Reference stores need a store barrier.
2741 pre_barrier(true /* do_load*/,
2742 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2743 NULL /* pre_val*/,
2744 T_OBJECT);
2745 #ifdef _LP64
2746 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2747 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2748 if (kind == LS_xchg) {
2749 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2750 newval_enc, adr_type, value_type->make_narrowoop()));
2751 } else {
2752 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2753 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2754 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2755 newval_enc, oldval_enc));
2756 }
2757 } else
2758 #endif
2759 {
2760 if (kind == LS_xchg) {
2761 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2762 } else {
2763 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2764 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2765 }
2766 }
2767 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2768 break;
2769 default:
2770 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2771 break;
2772 }
2773
2774 // SCMemProjNodes represent the memory state of a LoadStore. Their
2775 // main role is to prevent LoadStore nodes from being optimized away
2776 // when their results aren't used.
2777 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2778 set_memory(proj, alias_idx);
2779
2780 // Add the trailing membar surrounding the access
2781 insert_mem_bar(Op_MemBarCPUOrder);
2782 insert_mem_bar(Op_MemBarAcquire);
2783
2784 #ifdef _LP64
2785 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2786 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
2787 }
2788 #endif
2789
2790 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2791 set_result(load_store);
2792 return true;
2793 }
2794
2795 //----------------------------inline_unsafe_ordered_store----------------------
2796 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2797 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2798 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2799 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2800 // This is another variant of inline_unsafe_access, differing in
2801 // that it always issues store-store ("release") barrier and ensures
2802 // store-atomicity (which only matters for "long").
2803
2804 if (callee()->is_static()) return false; // caller must have the capability!
2805
2806 #ifndef PRODUCT
2807 {
2808 ResourceMark rm;
2809 // Check the signatures.
2810 ciSignature* sig = callee()->signature();
2811 #ifdef ASSERT
2812 BasicType rtype = sig->return_type()->basic_type();
2813 assert(rtype == T_VOID, "must return void");
2814 assert(sig->count() == 3, "has 3 arguments");
2815 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2816 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2817 #endif // ASSERT
2818 }
2819 #endif //PRODUCT
2820
2821 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2822
2823 // Get arguments:
2824 Node* receiver = argument(0); // type: oop
2825 Node* base = argument(1); // type: oop
2826 Node* offset = argument(2); // type: long
2827 Node* val = argument(4); // type: oop, int, or long
2828
2829 // Null check receiver.
2830 receiver = null_check(receiver);
2831 if (stopped()) {
2832 return true;
2833 }
2834
2835 // Build field offset expression.
2836 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2837 // 32-bit machines ignore the high half of long offsets
2838 offset = ConvL2X(offset);
2839 Node* adr = make_unsafe_address(base, offset);
2840 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2841 const Type *value_type = Type::get_const_basic_type(type);
2842 Compile::AliasType* alias_type = C->alias_type(adr_type);
2843
2844 insert_mem_bar(Op_MemBarRelease);
2845 insert_mem_bar(Op_MemBarCPUOrder);
2846 // Ensure that the store is atomic for longs:
2847 const bool require_atomic_access = true;
2848 Node* store;
2849 if (type == T_OBJECT) // reference stores need a store barrier.
2850 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2851 else {
2852 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2853 }
2854 insert_mem_bar(Op_MemBarCPUOrder);
2855 return true;
2856 }
2857
2858 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2859 // Regardless of form, don't allow previous ld/st to move down,
2860 // then issue acquire, release, or volatile mem_bar.
2861 insert_mem_bar(Op_MemBarCPUOrder);
2862 switch(id) {
2863 case vmIntrinsics::_loadFence:
2864 insert_mem_bar(Op_MemBarAcquire);
2865 return true;
2866 case vmIntrinsics::_storeFence:
2867 insert_mem_bar(Op_MemBarRelease);
2868 return true;
2869 case vmIntrinsics::_fullFence:
2870 insert_mem_bar(Op_MemBarVolatile);
2871 return true;
2872 default:
2873 fatal_unexpected_iid(id);
2874 return false;
2875 }
2876 }
2877
2878 //----------------------------inline_unsafe_allocate---------------------------
2879 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2880 bool LibraryCallKit::inline_unsafe_allocate() {
2881 if (callee()->is_static()) return false; // caller must have the capability!
2882
2883 null_check_receiver(); // null-check, then ignore
2884 Node* cls = null_check(argument(1));
2885 if (stopped()) return true;
2886
2887 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2888 kls = null_check(kls);
2889 if (stopped()) return true; // argument was like int.class
2890
2891 // Note: The argument might still be an illegal value like
2892 // Serializable.class or Object[].class. The runtime will handle it.
2893 // But we must make an explicit check for initialization.
2894 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2895 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2896 // can generate code to load it as unsigned byte.
2897 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2898 Node* bits = intcon(InstanceKlass::fully_initialized);
2899 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2900 // The 'test' is non-zero if we need to take a slow path.
2901
2902 Node* obj = new_instance(kls, test);
2903 set_result(obj);
2904 return true;
2905 }
2906
2907 #ifdef TRACE_HAVE_INTRINSICS
2908 /*
2909 * oop -> myklass
2910 * myklass->trace_id |= USED
2911 * return myklass->trace_id & ~0x3
2912 */
2913 bool LibraryCallKit::inline_native_classID() {
2914 null_check_receiver(); // null-check, then ignore
2915 Node* cls = null_check(argument(1), T_OBJECT);
2916 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2917 kls = null_check(kls, T_OBJECT);
2918 ByteSize offset = TRACE_ID_OFFSET;
2919 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2920 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2921 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2922 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2923 Node* clsused = longcon(0x01l); // set the class bit
2924 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2925
2926 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2927 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2928 set_result(andl);
2929 return true;
2930 }
2931
2932 bool LibraryCallKit::inline_native_threadID() {
2933 Node* tls_ptr = NULL;
2934 Node* cur_thr = generate_current_thread(tls_ptr);
2935 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2936 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2937 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2938
2939 Node* threadid = NULL;
2940 size_t thread_id_size = OSThread::thread_id_size();
2941 if (thread_id_size == (size_t) BytesPerLong) {
2942 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2943 } else if (thread_id_size == (size_t) BytesPerInt) {
2944 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2945 } else {
2946 ShouldNotReachHere();
2947 }
2948 set_result(threadid);
2949 return true;
2950 }
2951 #endif
2952
2953 //------------------------inline_native_time_funcs--------------
2954 // inline code for System.currentTimeMillis() and System.nanoTime()
2955 // these have the same type and signature
2956 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2957 const TypeFunc* tf = OptoRuntime::void_long_Type();
2958 const TypePtr* no_memory_effects = NULL;
2959 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2960 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2961 #ifdef ASSERT
2962 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2963 assert(value_top == top(), "second value must be top");
2964 #endif
2965 set_result(value);
2966 return true;
2967 }
2968
2969 //------------------------inline_native_currentThread------------------
2970 bool LibraryCallKit::inline_native_currentThread() {
2971 Node* junk = NULL;
2972 set_result(generate_current_thread(junk));
2973 return true;
2974 }
2975
2976 //------------------------inline_native_isInterrupted------------------
2977 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2978 bool LibraryCallKit::inline_native_isInterrupted() {
2979 // Add a fast path to t.isInterrupted(clear_int):
2980 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2981 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2982 // So, in the common case that the interrupt bit is false,
2983 // we avoid making a call into the VM. Even if the interrupt bit
2984 // is true, if the clear_int argument is false, we avoid the VM call.
2985 // However, if the receiver is not currentThread, we must call the VM,
2986 // because there must be some locking done around the operation.
2987
2988 // We only go to the fast case code if we pass two guards.
2989 // Paths which do not pass are accumulated in the slow_region.
2990
2991 enum {
2992 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2993 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2994 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2995 PATH_LIMIT
2996 };
2997
2998 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
2999 // out of the function.
3000 insert_mem_bar(Op_MemBarCPUOrder);
3001
3002 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
3003 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3004
3005 RegionNode* slow_region = new (C) RegionNode(1);
3006 record_for_igvn(slow_region);
3007
3008 // (a) Receiving thread must be the current thread.
3009 Node* rec_thr = argument(0);
3010 Node* tls_ptr = NULL;
3011 Node* cur_thr = generate_current_thread(tls_ptr);
3012 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3013 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3014
3015 generate_slow_guard(bol_thr, slow_region);
3016
3017 // (b) Interrupt bit on TLS must be false.
3018 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3019 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3020 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3021
3022 // Set the control input on the field _interrupted read to prevent it floating up.
3023 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3024 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3025 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3026
3027 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3028
3029 // First fast path: if (!TLS._interrupted) return false;
3030 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3031 result_rgn->init_req(no_int_result_path, false_bit);
3032 result_val->init_req(no_int_result_path, intcon(0));
3033
3034 // drop through to next case
3035 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3036
3037 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3038 Node* clr_arg = argument(1);
3039 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3040 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3041 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3042
3043 // Second fast path: ... else if (!clear_int) return true;
3044 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3045 result_rgn->init_req(no_clear_result_path, false_arg);
3046 result_val->init_req(no_clear_result_path, intcon(1));
3047
3048 // drop through to next case
3049 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3050
3051 // (d) Otherwise, go to the slow path.
3052 slow_region->add_req(control());
3053 set_control( _gvn.transform(slow_region) );
3054
3055 if (stopped()) {
3056 // There is no slow path.
3057 result_rgn->init_req(slow_result_path, top());
3058 result_val->init_req(slow_result_path, top());
3059 } else {
3060 // non-virtual because it is a private non-static
3061 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3062
3063 Node* slow_val = set_results_for_java_call(slow_call);
3064 // this->control() comes from set_results_for_java_call
3065
3066 Node* fast_io = slow_call->in(TypeFunc::I_O);
3067 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3068
3069 // These two phis are pre-filled with copies of of the fast IO and Memory
3070 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3071 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3072
3073 result_rgn->init_req(slow_result_path, control());
3074 result_io ->init_req(slow_result_path, i_o());
3075 result_mem->init_req(slow_result_path, reset_memory());
3076 result_val->init_req(slow_result_path, slow_val);
3077
3078 set_all_memory(_gvn.transform(result_mem));
3079 set_i_o( _gvn.transform(result_io));
3080 }
3081
3082 C->set_has_split_ifs(true); // Has chance for split-if optimization
3083 set_result(result_rgn, result_val);
3084 return true;
3085 }
3086
3087 //---------------------------load_mirror_from_klass----------------------------
3088 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3089 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3090 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3091 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3092 }
3093
3094 //-----------------------load_klass_from_mirror_common-------------------------
3095 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3096 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3097 // and branch to the given path on the region.
3098 // If never_see_null, take an uncommon trap on null, so we can optimistically
3099 // compile for the non-null case.
3100 // If the region is NULL, force never_see_null = true.
3101 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3102 bool never_see_null,
3103 RegionNode* region,
3104 int null_path,
3105 int offset) {
3106 if (region == NULL) never_see_null = true;
3107 Node* p = basic_plus_adr(mirror, offset);
3108 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3109 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3110 Node* null_ctl = top();
3111 kls = null_check_oop(kls, &null_ctl, never_see_null);
3112 if (region != NULL) {
3113 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3114 region->init_req(null_path, null_ctl);
3115 } else {
3116 assert(null_ctl == top(), "no loose ends");
3117 }
3118 return kls;
3119 }
3120
3121 //--------------------(inline_native_Class_query helpers)---------------------
3122 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3123 // Fall through if (mods & mask) == bits, take the guard otherwise.
3124 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3125 // Branch around if the given klass has the given modifier bit set.
3126 // Like generate_guard, adds a new path onto the region.
3127 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3128 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3129 Node* mask = intcon(modifier_mask);
3130 Node* bits = intcon(modifier_bits);
3131 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3132 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3133 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3134 return generate_fair_guard(bol, region);
3135 }
3136 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3137 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3138 }
3139
3140 //-------------------------inline_native_Class_query-------------------
3141 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3142 const Type* return_type = TypeInt::BOOL;
3143 Node* prim_return_value = top(); // what happens if it's a primitive class?
3144 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3145 bool expect_prim = false; // most of these guys expect to work on refs
3146
3147 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3148
3149 Node* mirror = argument(0);
3150 Node* obj = top();
3151
3152 switch (id) {
3153 case vmIntrinsics::_isInstance:
3154 // nothing is an instance of a primitive type
3155 prim_return_value = intcon(0);
3156 obj = argument(1);
3157 break;
3158 case vmIntrinsics::_getModifiers:
3159 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3160 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3161 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3162 break;
3163 case vmIntrinsics::_isInterface:
3164 prim_return_value = intcon(0);
3165 break;
3166 case vmIntrinsics::_isArray:
3167 prim_return_value = intcon(0);
3168 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3169 break;
3170 case vmIntrinsics::_isPrimitive:
3171 prim_return_value = intcon(1);
3172 expect_prim = true; // obviously
3173 break;
3174 case vmIntrinsics::_getSuperclass:
3175 prim_return_value = null();
3176 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3177 break;
3178 case vmIntrinsics::_getComponentType:
3179 prim_return_value = null();
3180 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3181 break;
3182 case vmIntrinsics::_getClassAccessFlags:
3183 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3184 return_type = TypeInt::INT; // not bool! 6297094
3185 break;
3186 default:
3187 fatal_unexpected_iid(id);
3188 break;
3189 }
3190
3191 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3192 if (mirror_con == NULL) return false; // cannot happen?
3193
3194 #ifndef PRODUCT
3195 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3196 ciType* k = mirror_con->java_mirror_type();
3197 if (k) {
3198 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3199 k->print_name();
3200 tty->cr();
3201 }
3202 }
3203 #endif
3204
3205 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3206 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3207 record_for_igvn(region);
3208 PhiNode* phi = new (C) PhiNode(region, return_type);
3209
3210 // The mirror will never be null of Reflection.getClassAccessFlags, however
3211 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3212 // if it is. See bug 4774291.
3213
3214 // For Reflection.getClassAccessFlags(), the null check occurs in
3215 // the wrong place; see inline_unsafe_access(), above, for a similar
3216 // situation.
3217 mirror = null_check(mirror);
3218 // If mirror or obj is dead, only null-path is taken.
3219 if (stopped()) return true;
3220
3221 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3222
3223 // Now load the mirror's klass metaobject, and null-check it.
3224 // Side-effects region with the control path if the klass is null.
3225 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3226 // If kls is null, we have a primitive mirror.
3227 phi->init_req(_prim_path, prim_return_value);
3228 if (stopped()) { set_result(region, phi); return true; }
3229
3230 Node* p; // handy temp
3231 Node* null_ctl;
3232
3233 // Now that we have the non-null klass, we can perform the real query.
3234 // For constant classes, the query will constant-fold in LoadNode::Value.
3235 Node* query_value = top();
3236 switch (id) {
3237 case vmIntrinsics::_isInstance:
3238 // nothing is an instance of a primitive type
3239 query_value = gen_instanceof(obj, kls);
3240 break;
3241
3242 case vmIntrinsics::_getModifiers:
3243 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3244 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3245 break;
3246
3247 case vmIntrinsics::_isInterface:
3248 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3249 if (generate_interface_guard(kls, region) != NULL)
3250 // A guard was added. If the guard is taken, it was an interface.
3251 phi->add_req(intcon(1));
3252 // If we fall through, it's a plain class.
3253 query_value = intcon(0);
3254 break;
3255
3256 case vmIntrinsics::_isArray:
3257 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3258 if (generate_array_guard(kls, region) != NULL)
3259 // A guard was added. If the guard is taken, it was an array.
3260 phi->add_req(intcon(1));
3261 // If we fall through, it's a plain class.
3262 query_value = intcon(0);
3263 break;
3264
3265 case vmIntrinsics::_isPrimitive:
3266 query_value = intcon(0); // "normal" path produces false
3267 break;
3268
3269 case vmIntrinsics::_getSuperclass:
3270 // The rules here are somewhat unfortunate, but we can still do better
3271 // with random logic than with a JNI call.
3272 // Interfaces store null or Object as _super, but must report null.
3273 // Arrays store an intermediate super as _super, but must report Object.
3274 // Other types can report the actual _super.
3275 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3276 if (generate_interface_guard(kls, region) != NULL)
3277 // A guard was added. If the guard is taken, it was an interface.
3278 phi->add_req(null());
3279 if (generate_array_guard(kls, region) != NULL)
3280 // A guard was added. If the guard is taken, it was an array.
3281 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3282 // If we fall through, it's a plain class. Get its _super.
3283 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3284 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3285 null_ctl = top();
3286 kls = null_check_oop(kls, &null_ctl);
3287 if (null_ctl != top()) {
3288 // If the guard is taken, Object.superClass is null (both klass and mirror).
3289 region->add_req(null_ctl);
3290 phi ->add_req(null());
3291 }
3292 if (!stopped()) {
3293 query_value = load_mirror_from_klass(kls);
3294 }
3295 break;
3296
3297 case vmIntrinsics::_getComponentType:
3298 if (generate_array_guard(kls, region) != NULL) {
3299 // Be sure to pin the oop load to the guard edge just created:
3300 Node* is_array_ctrl = region->in(region->req()-1);
3301 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3302 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3303 phi->add_req(cmo);
3304 }
3305 query_value = null(); // non-array case is null
3306 break;
3307
3308 case vmIntrinsics::_getClassAccessFlags:
3309 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3310 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3311 break;
3312
3313 default:
3314 fatal_unexpected_iid(id);
3315 break;
3316 }
3317
3318 // Fall-through is the normal case of a query to a real class.
3319 phi->init_req(1, query_value);
3320 region->init_req(1, control());
3321
3322 C->set_has_split_ifs(true); // Has chance for split-if optimization
3323 set_result(region, phi);
3324 return true;
3325 }
3326
3327 //--------------------------inline_native_subtype_check------------------------
3328 // This intrinsic takes the JNI calls out of the heart of
3329 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3330 bool LibraryCallKit::inline_native_subtype_check() {
3331 // Pull both arguments off the stack.
3332 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3333 args[0] = argument(0);
3334 args[1] = argument(1);
3335 Node* klasses[2]; // corresponding Klasses: superk, subk
3336 klasses[0] = klasses[1] = top();
3337
3338 enum {
3339 // A full decision tree on {superc is prim, subc is prim}:
3340 _prim_0_path = 1, // {P,N} => false
3341 // {P,P} & superc!=subc => false
3342 _prim_same_path, // {P,P} & superc==subc => true
3343 _prim_1_path, // {N,P} => false
3344 _ref_subtype_path, // {N,N} & subtype check wins => true
3345 _both_ref_path, // {N,N} & subtype check loses => false
3346 PATH_LIMIT
3347 };
3348
3349 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3350 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3351 record_for_igvn(region);
3352
3353 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3354 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3355 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3356
3357 // First null-check both mirrors and load each mirror's klass metaobject.
3358 int which_arg;
3359 for (which_arg = 0; which_arg <= 1; which_arg++) {
3360 Node* arg = args[which_arg];
3361 arg = null_check(arg);
3362 if (stopped()) break;
3363 args[which_arg] = arg;
3364
3365 Node* p = basic_plus_adr(arg, class_klass_offset);
3366 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3367 klasses[which_arg] = _gvn.transform(kls);
3368 }
3369
3370 // Having loaded both klasses, test each for null.
3371 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3372 for (which_arg = 0; which_arg <= 1; which_arg++) {
3373 Node* kls = klasses[which_arg];
3374 Node* null_ctl = top();
3375 kls = null_check_oop(kls, &null_ctl, never_see_null);
3376 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3377 region->init_req(prim_path, null_ctl);
3378 if (stopped()) break;
3379 klasses[which_arg] = kls;
3380 }
3381
3382 if (!stopped()) {
3383 // now we have two reference types, in klasses[0..1]
3384 Node* subk = klasses[1]; // the argument to isAssignableFrom
3385 Node* superk = klasses[0]; // the receiver
3386 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3387 // now we have a successful reference subtype check
3388 region->set_req(_ref_subtype_path, control());
3389 }
3390
3391 // If both operands are primitive (both klasses null), then
3392 // we must return true when they are identical primitives.
3393 // It is convenient to test this after the first null klass check.
3394 set_control(region->in(_prim_0_path)); // go back to first null check
3395 if (!stopped()) {
3396 // Since superc is primitive, make a guard for the superc==subc case.
3397 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3398 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3399 generate_guard(bol_eq, region, PROB_FAIR);
3400 if (region->req() == PATH_LIMIT+1) {
3401 // A guard was added. If the added guard is taken, superc==subc.
3402 region->swap_edges(PATH_LIMIT, _prim_same_path);
3403 region->del_req(PATH_LIMIT);
3404 }
3405 region->set_req(_prim_0_path, control()); // Not equal after all.
3406 }
3407
3408 // these are the only paths that produce 'true':
3409 phi->set_req(_prim_same_path, intcon(1));
3410 phi->set_req(_ref_subtype_path, intcon(1));
3411
3412 // pull together the cases:
3413 assert(region->req() == PATH_LIMIT, "sane region");
3414 for (uint i = 1; i < region->req(); i++) {
3415 Node* ctl = region->in(i);
3416 if (ctl == NULL || ctl == top()) {
3417 region->set_req(i, top());
3418 phi ->set_req(i, top());
3419 } else if (phi->in(i) == NULL) {
3420 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3421 }
3422 }
3423
3424 set_control(_gvn.transform(region));
3425 set_result(_gvn.transform(phi));
3426 return true;
3427 }
3428
3429 //---------------------generate_array_guard_common------------------------
3430 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3431 bool obj_array, bool not_array) {
3432 // If obj_array/non_array==false/false:
3433 // Branch around if the given klass is in fact an array (either obj or prim).
3434 // If obj_array/non_array==false/true:
3435 // Branch around if the given klass is not an array klass of any kind.
3436 // If obj_array/non_array==true/true:
3437 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3438 // If obj_array/non_array==true/false:
3439 // Branch around if the kls is an oop array (Object[] or subtype)
3440 //
3441 // Like generate_guard, adds a new path onto the region.
3442 jint layout_con = 0;
3443 Node* layout_val = get_layout_helper(kls, layout_con);
3444 if (layout_val == NULL) {
3445 bool query = (obj_array
3446 ? Klass::layout_helper_is_objArray(layout_con)
3447 : Klass::layout_helper_is_array(layout_con));
3448 if (query == not_array) {
3449 return NULL; // never a branch
3450 } else { // always a branch
3451 Node* always_branch = control();
3452 if (region != NULL)
3453 region->add_req(always_branch);
3454 set_control(top());
3455 return always_branch;
3456 }
3457 }
3458 // Now test the correct condition.
3459 jint nval = (obj_array
3460 ? ((jint)Klass::_lh_array_tag_type_value
3461 << Klass::_lh_array_tag_shift)
3462 : Klass::_lh_neutral_value);
3463 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3464 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3465 // invert the test if we are looking for a non-array
3466 if (not_array) btest = BoolTest(btest).negate();
3467 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3468 return generate_fair_guard(bol, region);
3469 }
3470
3471
3472 //-----------------------inline_native_newArray--------------------------
3473 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3474 bool LibraryCallKit::inline_native_newArray() {
3475 Node* mirror = argument(0);
3476 Node* count_val = argument(1);
3477
3478 mirror = null_check(mirror);
3479 // If mirror or obj is dead, only null-path is taken.
3480 if (stopped()) return true;
3481
3482 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3483 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3484 PhiNode* result_val = new(C) PhiNode(result_reg,
3485 TypeInstPtr::NOTNULL);
3486 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3487 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3488 TypePtr::BOTTOM);
3489
3490 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3491 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3492 result_reg, _slow_path);
3493 Node* normal_ctl = control();
3494 Node* no_array_ctl = result_reg->in(_slow_path);
3495
3496 // Generate code for the slow case. We make a call to newArray().
3497 set_control(no_array_ctl);
3498 if (!stopped()) {
3499 // Either the input type is void.class, or else the
3500 // array klass has not yet been cached. Either the
3501 // ensuing call will throw an exception, or else it
3502 // will cache the array klass for next time.
3503 PreserveJVMState pjvms(this);
3504 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3505 Node* slow_result = set_results_for_java_call(slow_call);
3506 // this->control() comes from set_results_for_java_call
3507 result_reg->set_req(_slow_path, control());
3508 result_val->set_req(_slow_path, slow_result);
3509 result_io ->set_req(_slow_path, i_o());
3510 result_mem->set_req(_slow_path, reset_memory());
3511 }
3512
3513 set_control(normal_ctl);
3514 if (!stopped()) {
3515 // Normal case: The array type has been cached in the java.lang.Class.
3516 // The following call works fine even if the array type is polymorphic.
3517 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3518 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3519 result_reg->init_req(_normal_path, control());
3520 result_val->init_req(_normal_path, obj);
3521 result_io ->init_req(_normal_path, i_o());
3522 result_mem->init_req(_normal_path, reset_memory());
3523 }
3524
3525 // Return the combined state.
3526 set_i_o( _gvn.transform(result_io) );
3527 set_all_memory( _gvn.transform(result_mem) );
3528
3529 C->set_has_split_ifs(true); // Has chance for split-if optimization
3530 set_result(result_reg, result_val);
3531 return true;
3532 }
3533
3534 //----------------------inline_native_getLength--------------------------
3535 // public static native int java.lang.reflect.Array.getLength(Object array);
3536 bool LibraryCallKit::inline_native_getLength() {
3537 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3538
3539 Node* array = null_check(argument(0));
3540 // If array is dead, only null-path is taken.
3541 if (stopped()) return true;
3542
3543 // Deoptimize if it is a non-array.
3544 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3545
3546 if (non_array != NULL) {
3547 PreserveJVMState pjvms(this);
3548 set_control(non_array);
3549 uncommon_trap(Deoptimization::Reason_intrinsic,
3550 Deoptimization::Action_maybe_recompile);
3551 }
3552
3553 // If control is dead, only non-array-path is taken.
3554 if (stopped()) return true;
3555
3556 // The works fine even if the array type is polymorphic.
3557 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3558 Node* result = load_array_length(array);
3559
3560 C->set_has_split_ifs(true); // Has chance for split-if optimization
3561 set_result(result);
3562 return true;
3563 }
3564
3565 //------------------------inline_array_copyOf----------------------------
3566 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3567 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3568 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3569 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3570
3571 // Get the arguments.
3572 Node* original = argument(0);
3573 Node* start = is_copyOfRange? argument(1): intcon(0);
3574 Node* end = is_copyOfRange? argument(2): argument(1);
3575 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3576
3577 Node* newcopy;
3578
3579 // Set the original stack and the reexecute bit for the interpreter to reexecute
3580 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3581 { PreserveReexecuteState preexecs(this);
3582 jvms()->set_should_reexecute(true);
3583
3584 array_type_mirror = null_check(array_type_mirror);
3585 original = null_check(original);
3586
3587 // Check if a null path was taken unconditionally.
3588 if (stopped()) return true;
3589
3590 Node* orig_length = load_array_length(original);
3591
3592 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3593 klass_node = null_check(klass_node);
3594
3595 RegionNode* bailout = new (C) RegionNode(1);
3596 record_for_igvn(bailout);
3597
3598 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3599 // Bail out if that is so.
3600 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3601 if (not_objArray != NULL) {
3602 // Improve the klass node's type from the new optimistic assumption:
3603 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3604 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3605 Node* cast = new (C) CastPPNode(klass_node, akls);
3606 cast->init_req(0, control());
3607 klass_node = _gvn.transform(cast);
3608 }
3609
3610 // Bail out if either start or end is negative.
3611 generate_negative_guard(start, bailout, &start);
3612 generate_negative_guard(end, bailout, &end);
3613
3614 Node* length = end;
3615 if (_gvn.type(start) != TypeInt::ZERO) {
3616 length = _gvn.transform(new (C) SubINode(end, start));
3617 }
3618
3619 // Bail out if length is negative.
3620 // Without this the new_array would throw
3621 // NegativeArraySizeException but IllegalArgumentException is what
3622 // should be thrown
3623 generate_negative_guard(length, bailout, &length);
3624
3625 if (bailout->req() > 1) {
3626 PreserveJVMState pjvms(this);
3627 set_control(_gvn.transform(bailout));
3628 uncommon_trap(Deoptimization::Reason_intrinsic,
3629 Deoptimization::Action_maybe_recompile);
3630 }
3631
3632 if (!stopped()) {
3633 // How many elements will we copy from the original?
3634 // The answer is MinI(orig_length - start, length).
3635 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3636 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3637
3638 newcopy = new_array(klass_node, length, 0); // no argments to push
3639
3640 // Generate a direct call to the right arraycopy function(s).
3641 // We know the copy is disjoint but we might not know if the
3642 // oop stores need checking.
3643 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3644 // This will fail a store-check if x contains any non-nulls.
3645 bool disjoint_bases = true;
3646 // if start > orig_length then the length of the copy may be
3647 // negative.
3648 bool length_never_negative = !is_copyOfRange;
3649 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3650 original, start, newcopy, intcon(0), moved,
3651 disjoint_bases, length_never_negative);
3652 }
3653 } // original reexecute is set back here
3654
3655 C->set_has_split_ifs(true); // Has chance for split-if optimization
3656 if (!stopped()) {
3657 set_result(newcopy);
3658 }
3659 return true;
3660 }
3661
3662
3663 //----------------------generate_virtual_guard---------------------------
3664 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3665 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3666 RegionNode* slow_region) {
3667 ciMethod* method = callee();
3668 int vtable_index = method->vtable_index();
3669 // Get the Method* out of the appropriate vtable entry.
3670 int entry_offset = (InstanceKlass::vtable_start_offset() +
3671 vtable_index*vtableEntry::size()) * wordSize +
3672 vtableEntry::method_offset_in_bytes();
3673 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3674 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3675
3676 // Compare the target method with the expected method (e.g., Object.hashCode).
3677 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3678
3679 Node* native_call = makecon(native_call_addr);
3680 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3681 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3682
3683 return generate_slow_guard(test_native, slow_region);
3684 }
3685
3686 //-----------------------generate_method_call----------------------------
3687 // Use generate_method_call to make a slow-call to the real
3688 // method if the fast path fails. An alternative would be to
3689 // use a stub like OptoRuntime::slow_arraycopy_Java.
3690 // This only works for expanding the current library call,
3691 // not another intrinsic. (E.g., don't use this for making an
3692 // arraycopy call inside of the copyOf intrinsic.)
3693 CallJavaNode*
3694 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3695 // When compiling the intrinsic method itself, do not use this technique.
3696 guarantee(callee() != C->method(), "cannot make slow-call to self");
3697
3698 ciMethod* method = callee();
3699 // ensure the JVMS we have will be correct for this call
3700 guarantee(method_id == method->intrinsic_id(), "must match");
3701
3702 const TypeFunc* tf = TypeFunc::make(method);
3703 CallJavaNode* slow_call;
3704 if (is_static) {
3705 assert(!is_virtual, "");
3706 slow_call = new(C) CallStaticJavaNode(tf,
3707 SharedRuntime::get_resolve_static_call_stub(),
3708 method, bci());
3709 } else if (is_virtual) {
3710 null_check_receiver();
3711 int vtable_index = Method::invalid_vtable_index;
3712 if (UseInlineCaches) {
3713 // Suppress the vtable call
3714 } else {
3715 // hashCode and clone are not a miranda methods,
3716 // so the vtable index is fixed.
3717 // No need to use the linkResolver to get it.
3718 vtable_index = method->vtable_index();
3719 }
3720 slow_call = new(C) CallDynamicJavaNode(tf,
3721 SharedRuntime::get_resolve_virtual_call_stub(),
3722 method, vtable_index, bci());
3723 } else { // neither virtual nor static: opt_virtual
3724 null_check_receiver();
3725 slow_call = new(C) CallStaticJavaNode(tf,
3726 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3727 method, bci());
3728 slow_call->set_optimized_virtual(true);
3729 }
3730 set_arguments_for_java_call(slow_call);
3731 set_edges_for_java_call(slow_call);
3732 return slow_call;
3733 }
3734
3735
3736 //------------------------------inline_native_hashcode--------------------
3737 // Build special case code for calls to hashCode on an object.
3738 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3739 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3740 assert(!(is_virtual && is_static), "either virtual, special, or static");
3741
3742 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3743
3744 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3745 PhiNode* result_val = new(C) PhiNode(result_reg,
3746 TypeInt::INT);
3747 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3748 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3749 TypePtr::BOTTOM);
3750 Node* obj = NULL;
3751 if (!is_static) {
3752 // Check for hashing null object
3753 obj = null_check_receiver();
3754 if (stopped()) return true; // unconditionally null
3755 result_reg->init_req(_null_path, top());
3756 result_val->init_req(_null_path, top());
3757 } else {
3758 // Do a null check, and return zero if null.
3759 // System.identityHashCode(null) == 0
3760 obj = argument(0);
3761 Node* null_ctl = top();
3762 obj = null_check_oop(obj, &null_ctl);
3763 result_reg->init_req(_null_path, null_ctl);
3764 result_val->init_req(_null_path, _gvn.intcon(0));
3765 }
3766
3767 // Unconditionally null? Then return right away.
3768 if (stopped()) {
3769 set_control( result_reg->in(_null_path));
3770 if (!stopped())
3771 set_result(result_val->in(_null_path));
3772 return true;
3773 }
3774
3775 // After null check, get the object's klass.
3776 Node* obj_klass = load_object_klass(obj);
3777
3778 // This call may be virtual (invokevirtual) or bound (invokespecial).
3779 // For each case we generate slightly different code.
3780
3781 // We only go to the fast case code if we pass a number of guards. The
3782 // paths which do not pass are accumulated in the slow_region.
3783 RegionNode* slow_region = new (C) RegionNode(1);
3784 record_for_igvn(slow_region);
3785
3786 // If this is a virtual call, we generate a funny guard. We pull out
3787 // the vtable entry corresponding to hashCode() from the target object.
3788 // If the target method which we are calling happens to be the native
3789 // Object hashCode() method, we pass the guard. We do not need this
3790 // guard for non-virtual calls -- the caller is known to be the native
3791 // Object hashCode().
3792 if (is_virtual) {
3793 generate_virtual_guard(obj_klass, slow_region);
3794 }
3795
3796 // Get the header out of the object, use LoadMarkNode when available
3797 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3798 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3799
3800 // Test the header to see if it is unlocked.
3801 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3802 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3803 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3804 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3805 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3806
3807 generate_slow_guard(test_unlocked, slow_region);
3808
3809 // Get the hash value and check to see that it has been properly assigned.
3810 // We depend on hash_mask being at most 32 bits and avoid the use of
3811 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3812 // vm: see markOop.hpp.
3813 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3814 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3815 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3816 // This hack lets the hash bits live anywhere in the mark object now, as long
3817 // as the shift drops the relevant bits into the low 32 bits. Note that
3818 // Java spec says that HashCode is an int so there's no point in capturing
3819 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3820 hshifted_header = ConvX2I(hshifted_header);
3821 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3822
3823 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3824 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3825 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3826
3827 generate_slow_guard(test_assigned, slow_region);
3828
3829 Node* init_mem = reset_memory();
3830 // fill in the rest of the null path:
3831 result_io ->init_req(_null_path, i_o());
3832 result_mem->init_req(_null_path, init_mem);
3833
3834 result_val->init_req(_fast_path, hash_val);
3835 result_reg->init_req(_fast_path, control());
3836 result_io ->init_req(_fast_path, i_o());
3837 result_mem->init_req(_fast_path, init_mem);
3838
3839 // Generate code for the slow case. We make a call to hashCode().
3840 set_control(_gvn.transform(slow_region));
3841 if (!stopped()) {
3842 // No need for PreserveJVMState, because we're using up the present state.
3843 set_all_memory(init_mem);
3844 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3845 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3846 Node* slow_result = set_results_for_java_call(slow_call);
3847 // this->control() comes from set_results_for_java_call
3848 result_reg->init_req(_slow_path, control());
3849 result_val->init_req(_slow_path, slow_result);
3850 result_io ->set_req(_slow_path, i_o());
3851 result_mem ->set_req(_slow_path, reset_memory());
3852 }
3853
3854 // Return the combined state.
3855 set_i_o( _gvn.transform(result_io) );
3856 set_all_memory( _gvn.transform(result_mem) );
3857
3858 set_result(result_reg, result_val);
3859 return true;
3860 }
3861
3862 //---------------------------inline_native_getClass----------------------------
3863 // public final native Class<?> java.lang.Object.getClass();
3864 //
3865 // Build special case code for calls to getClass on an object.
3866 bool LibraryCallKit::inline_native_getClass() {
3867 Node* obj = null_check_receiver();
3868 if (stopped()) return true;
3869 set_result(load_mirror_from_klass(load_object_klass(obj)));
3870 return true;
3871 }
3872
3873 //-----------------inline_native_Reflection_getCallerClass---------------------
3874 // public static native Class<?> sun.reflect.Reflection.getCallerClass();
3875 //
3876 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3877 //
3878 // NOTE: This code must perform the same logic as JVM_GetCallerClass
3879 // in that it must skip particular security frames and checks for
3880 // caller sensitive methods.
3881 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3882 #ifndef PRODUCT
3883 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3884 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3885 }
3886 #endif
3887
3888 if (!jvms()->has_method()) {
3889 #ifndef PRODUCT
3890 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3891 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3892 }
3893 #endif
3894 return false;
3895 }
3896
3897 // Walk back up the JVM state to find the caller at the required
3898 // depth.
3899 JVMState* caller_jvms = jvms();
3900
3901 // Cf. JVM_GetCallerClass
3902 // NOTE: Start the loop at depth 1 because the current JVM state does
3903 // not include the Reflection.getCallerClass() frame.
3904 for (int n = 1; caller_jvms != NULL; caller_jvms = caller_jvms->caller(), n++) {
3905 ciMethod* m = caller_jvms->method();
3906 switch (n) {
3907 case 0:
3908 fatal("current JVM state does not include the Reflection.getCallerClass frame");
3909 break;
3910 case 1:
3911 // Frame 0 and 1 must be caller sensitive (see JVM_GetCallerClass).
3912 if (!m->caller_sensitive()) {
3913 #ifndef PRODUCT
3914 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3915 tty->print_cr(" Bailing out: CallerSensitive annotation expected at frame %d", n);
3916 }
3917 #endif
3918 return false; // bail-out; let JVM_GetCallerClass do the work
3919 }
3920 break;
3921 default:
3922 if (!m->is_ignored_by_security_stack_walk()) {
3923 // We have reached the desired frame; return the holder class.
3924 // Acquire method holder as java.lang.Class and push as constant.
3925 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3926 ciInstance* caller_mirror = caller_klass->java_mirror();
3927 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3928
3929 #ifndef PRODUCT
3930 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3931 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());
3932 tty->print_cr(" JVM state at this point:");
3933 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
3934 ciMethod* m = jvms()->of_depth(i)->method();
3935 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3936 }
3937 }
3938 #endif
3939 return true;
3940 }
3941 break;
3942 }
3943 }
3944
3945 #ifndef PRODUCT
3946 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3947 tty->print_cr(" Bailing out because caller depth exceeded inlining depth = %d", jvms()->depth());
3948 tty->print_cr(" JVM state at this point:");
3949 for (int i = jvms()->depth(), n = 1; i >= 1; i--, n++) {
3950 ciMethod* m = jvms()->of_depth(i)->method();
3951 tty->print_cr(" %d) %s.%s", n, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3952 }
3953 }
3954 #endif
3955
3956 return false; // bail-out; let JVM_GetCallerClass do the work
3957 }
3958
3959 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3960 Node* arg = argument(0);
3961 Node* result;
3962
3963 switch (id) {
3964 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
3965 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
3966 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
3967 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
3968
3969 case vmIntrinsics::_doubleToLongBits: {
3970 // two paths (plus control) merge in a wood
3971 RegionNode *r = new (C) RegionNode(3);
3972 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
3973
3974 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
3975 // Build the boolean node
3976 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
3977
3978 // Branch either way.
3979 // NaN case is less traveled, which makes all the difference.
3980 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3981 Node *opt_isnan = _gvn.transform(ifisnan);
3982 assert( opt_isnan->is_If(), "Expect an IfNode");
3983 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
3984 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
3985
3986 set_control(iftrue);
3987
3988 static const jlong nan_bits = CONST64(0x7ff8000000000000);
3989 Node *slow_result = longcon(nan_bits); // return NaN
3990 phi->init_req(1, _gvn.transform( slow_result ));
3991 r->init_req(1, iftrue);
3992
3993 // Else fall through
3994 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
3995 set_control(iffalse);
3996
3997 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
3998 r->init_req(2, iffalse);
3999
4000 // Post merge
4001 set_control(_gvn.transform(r));
4002 record_for_igvn(r);
4003
4004 C->set_has_split_ifs(true); // Has chance for split-if optimization
4005 result = phi;
4006 assert(result->bottom_type()->isa_long(), "must be");
4007 break;
4008 }
4009
4010 case vmIntrinsics::_floatToIntBits: {
4011 // two paths (plus control) merge in a wood
4012 RegionNode *r = new (C) RegionNode(3);
4013 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4014
4015 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4016 // Build the boolean node
4017 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4018
4019 // Branch either way.
4020 // NaN case is less traveled, which makes all the difference.
4021 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4022 Node *opt_isnan = _gvn.transform(ifisnan);
4023 assert( opt_isnan->is_If(), "Expect an IfNode");
4024 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4025 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4026
4027 set_control(iftrue);
4028
4029 static const jint nan_bits = 0x7fc00000;
4030 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4031 phi->init_req(1, _gvn.transform( slow_result ));
4032 r->init_req(1, iftrue);
4033
4034 // Else fall through
4035 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4036 set_control(iffalse);
4037
4038 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4039 r->init_req(2, iffalse);
4040
4041 // Post merge
4042 set_control(_gvn.transform(r));
4043 record_for_igvn(r);
4044
4045 C->set_has_split_ifs(true); // Has chance for split-if optimization
4046 result = phi;
4047 assert(result->bottom_type()->isa_int(), "must be");
4048 break;
4049 }
4050
4051 default:
4052 fatal_unexpected_iid(id);
4053 break;
4054 }
4055 set_result(_gvn.transform(result));
4056 return true;
4057 }
4058
4059 #ifdef _LP64
4060 #define XTOP ,top() /*additional argument*/
4061 #else //_LP64
4062 #define XTOP /*no additional argument*/
4063 #endif //_LP64
4064
4065 //----------------------inline_unsafe_copyMemory-------------------------
4066 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4067 bool LibraryCallKit::inline_unsafe_copyMemory() {
4068 if (callee()->is_static()) return false; // caller must have the capability!
4069 null_check_receiver(); // null-check receiver
4070 if (stopped()) return true;
4071
4072 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4073
4074 Node* src_ptr = argument(1); // type: oop
4075 Node* src_off = ConvL2X(argument(2)); // type: long
4076 Node* dst_ptr = argument(4); // type: oop
4077 Node* dst_off = ConvL2X(argument(5)); // type: long
4078 Node* size = ConvL2X(argument(7)); // type: long
4079
4080 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4081 "fieldOffset must be byte-scaled");
4082
4083 Node* src = make_unsafe_address(src_ptr, src_off);
4084 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4085
4086 // Conservatively insert a memory barrier on all memory slices.
4087 // Do not let writes of the copy source or destination float below the copy.
4088 insert_mem_bar(Op_MemBarCPUOrder);
4089
4090 // Call it. Note that the length argument is not scaled.
4091 make_runtime_call(RC_LEAF|RC_NO_FP,
4092 OptoRuntime::fast_arraycopy_Type(),
4093 StubRoutines::unsafe_arraycopy(),
4094 "unsafe_arraycopy",
4095 TypeRawPtr::BOTTOM,
4096 src, dst, size XTOP);
4097
4098 // Do not let reads of the copy destination float above the copy.
4099 insert_mem_bar(Op_MemBarCPUOrder);
4100
4101 return true;
4102 }
4103
4104 //------------------------clone_coping-----------------------------------
4105 // Helper function for inline_native_clone.
4106 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4107 assert(obj_size != NULL, "");
4108 Node* raw_obj = alloc_obj->in(1);
4109 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4110
4111 AllocateNode* alloc = NULL;
4112 if (ReduceBulkZeroing) {
4113 // We will be completely responsible for initializing this object -
4114 // mark Initialize node as complete.
4115 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4116 // The object was just allocated - there should be no any stores!
4117 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4118 // Mark as complete_with_arraycopy so that on AllocateNode
4119 // expansion, we know this AllocateNode is initialized by an array
4120 // copy and a StoreStore barrier exists after the array copy.
4121 alloc->initialization()->set_complete_with_arraycopy();
4122 }
4123
4124 // Copy the fastest available way.
4125 // TODO: generate fields copies for small objects instead.
4126 Node* src = obj;
4127 Node* dest = alloc_obj;
4128 Node* size = _gvn.transform(obj_size);
4129
4130 // Exclude the header but include array length to copy by 8 bytes words.
4131 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4132 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4133 instanceOopDesc::base_offset_in_bytes();
4134 // base_off:
4135 // 8 - 32-bit VM
4136 // 12 - 64-bit VM, compressed klass
4137 // 16 - 64-bit VM, normal klass
4138 if (base_off % BytesPerLong != 0) {
4139 assert(UseCompressedKlassPointers, "");
4140 if (is_array) {
4141 // Exclude length to copy by 8 bytes words.
4142 base_off += sizeof(int);
4143 } else {
4144 // Include klass to copy by 8 bytes words.
4145 base_off = instanceOopDesc::klass_offset_in_bytes();
4146 }
4147 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4148 }
4149 src = basic_plus_adr(src, base_off);
4150 dest = basic_plus_adr(dest, base_off);
4151
4152 // Compute the length also, if needed:
4153 Node* countx = size;
4154 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4155 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4156
4157 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4158 bool disjoint_bases = true;
4159 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4160 src, NULL, dest, NULL, countx,
4161 /*dest_uninitialized*/true);
4162
4163 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4164 if (card_mark) {
4165 assert(!is_array, "");
4166 // Put in store barrier for any and all oops we are sticking
4167 // into this object. (We could avoid this if we could prove
4168 // that the object type contains no oop fields at all.)
4169 Node* no_particular_value = NULL;
4170 Node* no_particular_field = NULL;
4171 int raw_adr_idx = Compile::AliasIdxRaw;
4172 post_barrier(control(),
4173 memory(raw_adr_type),
4174 alloc_obj,
4175 no_particular_field,
4176 raw_adr_idx,
4177 no_particular_value,
4178 T_OBJECT,
4179 false);
4180 }
4181
4182 // Do not let reads from the cloned object float above the arraycopy.
4183 if (alloc != NULL) {
4184 // Do not let stores that initialize this object be reordered with
4185 // a subsequent store that would make this object accessible by
4186 // other threads.
4187 // Record what AllocateNode this StoreStore protects so that
4188 // escape analysis can go from the MemBarStoreStoreNode to the
4189 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4190 // based on the escape status of the AllocateNode.
4191 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4192 } else {
4193 insert_mem_bar(Op_MemBarCPUOrder);
4194 }
4195 }
4196
4197 //------------------------inline_native_clone----------------------------
4198 // protected native Object java.lang.Object.clone();
4199 //
4200 // Here are the simple edge cases:
4201 // null receiver => normal trap
4202 // virtual and clone was overridden => slow path to out-of-line clone
4203 // not cloneable or finalizer => slow path to out-of-line Object.clone
4204 //
4205 // The general case has two steps, allocation and copying.
4206 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4207 //
4208 // Copying also has two cases, oop arrays and everything else.
4209 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4210 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4211 //
4212 // These steps fold up nicely if and when the cloned object's klass
4213 // can be sharply typed as an object array, a type array, or an instance.
4214 //
4215 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4216 PhiNode* result_val;
4217
4218 // Set the reexecute bit for the interpreter to reexecute
4219 // the bytecode that invokes Object.clone if deoptimization happens.
4220 { PreserveReexecuteState preexecs(this);
4221 jvms()->set_should_reexecute(true);
4222
4223 Node* obj = null_check_receiver();
4224 if (stopped()) return true;
4225
4226 Node* obj_klass = load_object_klass(obj);
4227 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4228 const TypeOopPtr* toop = ((tklass != NULL)
4229 ? tklass->as_instance_type()
4230 : TypeInstPtr::NOTNULL);
4231
4232 // Conservatively insert a memory barrier on all memory slices.
4233 // Do not let writes into the original float below the clone.
4234 insert_mem_bar(Op_MemBarCPUOrder);
4235
4236 // paths into result_reg:
4237 enum {
4238 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4239 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4240 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4241 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4242 PATH_LIMIT
4243 };
4244 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4245 result_val = new(C) PhiNode(result_reg,
4246 TypeInstPtr::NOTNULL);
4247 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4248 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4249 TypePtr::BOTTOM);
4250 record_for_igvn(result_reg);
4251
4252 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4253 int raw_adr_idx = Compile::AliasIdxRaw;
4254
4255 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4256 if (array_ctl != NULL) {
4257 // It's an array.
4258 PreserveJVMState pjvms(this);
4259 set_control(array_ctl);
4260 Node* obj_length = load_array_length(obj);
4261 Node* obj_size = NULL;
4262 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4263
4264 if (!use_ReduceInitialCardMarks()) {
4265 // If it is an oop array, it requires very special treatment,
4266 // because card marking is required on each card of the array.
4267 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4268 if (is_obja != NULL) {
4269 PreserveJVMState pjvms2(this);
4270 set_control(is_obja);
4271 // Generate a direct call to the right arraycopy function(s).
4272 bool disjoint_bases = true;
4273 bool length_never_negative = true;
4274 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4275 obj, intcon(0), alloc_obj, intcon(0),
4276 obj_length,
4277 disjoint_bases, length_never_negative);
4278 result_reg->init_req(_objArray_path, control());
4279 result_val->init_req(_objArray_path, alloc_obj);
4280 result_i_o ->set_req(_objArray_path, i_o());
4281 result_mem ->set_req(_objArray_path, reset_memory());
4282 }
4283 }
4284 // Otherwise, there are no card marks to worry about.
4285 // (We can dispense with card marks if we know the allocation
4286 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4287 // causes the non-eden paths to take compensating steps to
4288 // simulate a fresh allocation, so that no further
4289 // card marks are required in compiled code to initialize
4290 // the object.)
4291
4292 if (!stopped()) {
4293 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4294
4295 // Present the results of the copy.
4296 result_reg->init_req(_array_path, control());
4297 result_val->init_req(_array_path, alloc_obj);
4298 result_i_o ->set_req(_array_path, i_o());
4299 result_mem ->set_req(_array_path, reset_memory());
4300 }
4301 }
4302
4303 // We only go to the instance fast case code if we pass a number of guards.
4304 // The paths which do not pass are accumulated in the slow_region.
4305 RegionNode* slow_region = new (C) RegionNode(1);
4306 record_for_igvn(slow_region);
4307 if (!stopped()) {
4308 // It's an instance (we did array above). Make the slow-path tests.
4309 // If this is a virtual call, we generate a funny guard. We grab
4310 // the vtable entry corresponding to clone() from the target object.
4311 // If the target method which we are calling happens to be the
4312 // Object clone() method, we pass the guard. We do not need this
4313 // guard for non-virtual calls; the caller is known to be the native
4314 // Object clone().
4315 if (is_virtual) {
4316 generate_virtual_guard(obj_klass, slow_region);
4317 }
4318
4319 // The object must be cloneable and must not have a finalizer.
4320 // Both of these conditions may be checked in a single test.
4321 // We could optimize the cloneable test further, but we don't care.
4322 generate_access_flags_guard(obj_klass,
4323 // Test both conditions:
4324 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4325 // Must be cloneable but not finalizer:
4326 JVM_ACC_IS_CLONEABLE,
4327 slow_region);
4328 }
4329
4330 if (!stopped()) {
4331 // It's an instance, and it passed the slow-path tests.
4332 PreserveJVMState pjvms(this);
4333 Node* obj_size = NULL;
4334 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4335
4336 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4337
4338 // Present the results of the slow call.
4339 result_reg->init_req(_instance_path, control());
4340 result_val->init_req(_instance_path, alloc_obj);
4341 result_i_o ->set_req(_instance_path, i_o());
4342 result_mem ->set_req(_instance_path, reset_memory());
4343 }
4344
4345 // Generate code for the slow case. We make a call to clone().
4346 set_control(_gvn.transform(slow_region));
4347 if (!stopped()) {
4348 PreserveJVMState pjvms(this);
4349 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4350 Node* slow_result = set_results_for_java_call(slow_call);
4351 // this->control() comes from set_results_for_java_call
4352 result_reg->init_req(_slow_path, control());
4353 result_val->init_req(_slow_path, slow_result);
4354 result_i_o ->set_req(_slow_path, i_o());
4355 result_mem ->set_req(_slow_path, reset_memory());
4356 }
4357
4358 // Return the combined state.
4359 set_control( _gvn.transform(result_reg) );
4360 set_i_o( _gvn.transform(result_i_o) );
4361 set_all_memory( _gvn.transform(result_mem) );
4362 } // original reexecute is set back here
4363
4364 set_result(_gvn.transform(result_val));
4365 return true;
4366 }
4367
4368 //------------------------------basictype2arraycopy----------------------------
4369 address LibraryCallKit::basictype2arraycopy(BasicType t,
4370 Node* src_offset,
4371 Node* dest_offset,
4372 bool disjoint_bases,
4373 const char* &name,
4374 bool dest_uninitialized) {
4375 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4376 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4377
4378 bool aligned = false;
4379 bool disjoint = disjoint_bases;
4380
4381 // if the offsets are the same, we can treat the memory regions as
4382 // disjoint, because either the memory regions are in different arrays,
4383 // or they are identical (which we can treat as disjoint.) We can also
4384 // treat a copy with a destination index less that the source index
4385 // as disjoint since a low->high copy will work correctly in this case.
4386 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4387 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4388 // both indices are constants
4389 int s_offs = src_offset_inttype->get_con();
4390 int d_offs = dest_offset_inttype->get_con();
4391 int element_size = type2aelembytes(t);
4392 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4393 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4394 if (s_offs >= d_offs) disjoint = true;
4395 } else if (src_offset == dest_offset && src_offset != NULL) {
4396 // This can occur if the offsets are identical non-constants.
4397 disjoint = true;
4398 }
4399
4400 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4401 }
4402
4403
4404 //------------------------------inline_arraycopy-----------------------
4405 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4406 // Object dest, int destPos,
4407 // int length);
4408 bool LibraryCallKit::inline_arraycopy() {
4409 // Get the arguments.
4410 Node* src = argument(0); // type: oop
4411 Node* src_offset = argument(1); // type: int
4412 Node* dest = argument(2); // type: oop
4413 Node* dest_offset = argument(3); // type: int
4414 Node* length = argument(4); // type: int
4415
4416 // Compile time checks. If any of these checks cannot be verified at compile time,
4417 // we do not make a fast path for this call. Instead, we let the call remain as it
4418 // is. The checks we choose to mandate at compile time are:
4419 //
4420 // (1) src and dest are arrays.
4421 const Type* src_type = src->Value(&_gvn);
4422 const Type* dest_type = dest->Value(&_gvn);
4423 const TypeAryPtr* top_src = src_type->isa_aryptr();
4424 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4425 if (top_src == NULL || top_src->klass() == NULL ||
4426 top_dest == NULL || top_dest->klass() == NULL) {
4427 // Conservatively insert a memory barrier on all memory slices.
4428 // Do not let writes into the source float below the arraycopy.
4429 insert_mem_bar(Op_MemBarCPUOrder);
4430
4431 // Call StubRoutines::generic_arraycopy stub.
4432 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4433 src, src_offset, dest, dest_offset, length);
4434
4435 // Do not let reads from the destination float above the arraycopy.
4436 // Since we cannot type the arrays, we don't know which slices
4437 // might be affected. We could restrict this barrier only to those
4438 // memory slices which pertain to array elements--but don't bother.
4439 if (!InsertMemBarAfterArraycopy)
4440 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4441 insert_mem_bar(Op_MemBarCPUOrder);
4442 return true;
4443 }
4444
4445 // (2) src and dest arrays must have elements of the same BasicType
4446 // Figure out the size and type of the elements we will be copying.
4447 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4448 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4449 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4450 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4451
4452 if (src_elem != dest_elem || dest_elem == T_VOID) {
4453 // The component types are not the same or are not recognized. Punt.
4454 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4455 generate_slow_arraycopy(TypePtr::BOTTOM,
4456 src, src_offset, dest, dest_offset, length,
4457 /*dest_uninitialized*/false);
4458 return true;
4459 }
4460
4461 //---------------------------------------------------------------------------
4462 // We will make a fast path for this call to arraycopy.
4463
4464 // We have the following tests left to perform:
4465 //
4466 // (3) src and dest must not be null.
4467 // (4) src_offset must not be negative.
4468 // (5) dest_offset must not be negative.
4469 // (6) length must not be negative.
4470 // (7) src_offset + length must not exceed length of src.
4471 // (8) dest_offset + length must not exceed length of dest.
4472 // (9) each element of an oop array must be assignable
4473
4474 RegionNode* slow_region = new (C) RegionNode(1);
4475 record_for_igvn(slow_region);
4476
4477 // (3) operands must not be null
4478 // We currently perform our null checks with the null_check routine.
4479 // This means that the null exceptions will be reported in the caller
4480 // rather than (correctly) reported inside of the native arraycopy call.
4481 // This should be corrected, given time. We do our null check with the
4482 // stack pointer restored.
4483 src = null_check(src, T_ARRAY);
4484 dest = null_check(dest, T_ARRAY);
4485
4486 // (4) src_offset must not be negative.
4487 generate_negative_guard(src_offset, slow_region);
4488
4489 // (5) dest_offset must not be negative.
4490 generate_negative_guard(dest_offset, slow_region);
4491
4492 // (6) length must not be negative (moved to generate_arraycopy()).
4493 // generate_negative_guard(length, slow_region);
4494
4495 // (7) src_offset + length must not exceed length of src.
4496 generate_limit_guard(src_offset, length,
4497 load_array_length(src),
4498 slow_region);
4499
4500 // (8) dest_offset + length must not exceed length of dest.
4501 generate_limit_guard(dest_offset, length,
4502 load_array_length(dest),
4503 slow_region);
4504
4505 // (9) each element of an oop array must be assignable
4506 // The generate_arraycopy subroutine checks this.
4507
4508 // This is where the memory effects are placed:
4509 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4510 generate_arraycopy(adr_type, dest_elem,
4511 src, src_offset, dest, dest_offset, length,
4512 false, false, slow_region);
4513
4514 return true;
4515 }
4516
4517 //-----------------------------generate_arraycopy----------------------
4518 // Generate an optimized call to arraycopy.
4519 // Caller must guard against non-arrays.
4520 // Caller must determine a common array basic-type for both arrays.
4521 // Caller must validate offsets against array bounds.
4522 // The slow_region has already collected guard failure paths
4523 // (such as out of bounds length or non-conformable array types).
4524 // The generated code has this shape, in general:
4525 //
4526 // if (length == 0) return // via zero_path
4527 // slowval = -1
4528 // if (types unknown) {
4529 // slowval = call generic copy loop
4530 // if (slowval == 0) return // via checked_path
4531 // } else if (indexes in bounds) {
4532 // if ((is object array) && !(array type check)) {
4533 // slowval = call checked copy loop
4534 // if (slowval == 0) return // via checked_path
4535 // } else {
4536 // call bulk copy loop
4537 // return // via fast_path
4538 // }
4539 // }
4540 // // adjust params for remaining work:
4541 // if (slowval != -1) {
4542 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4543 // }
4544 // slow_region:
4545 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4546 // return // via slow_call_path
4547 //
4548 // This routine is used from several intrinsics: System.arraycopy,
4549 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4550 //
4551 void
4552 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4553 BasicType basic_elem_type,
4554 Node* src, Node* src_offset,
4555 Node* dest, Node* dest_offset,
4556 Node* copy_length,
4557 bool disjoint_bases,
4558 bool length_never_negative,
4559 RegionNode* slow_region) {
4560
4561 if (slow_region == NULL) {
4562 slow_region = new(C) RegionNode(1);
4563 record_for_igvn(slow_region);
4564 }
4565
4566 Node* original_dest = dest;
4567 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4568 bool dest_uninitialized = false;
4569
4570 // See if this is the initialization of a newly-allocated array.
4571 // If so, we will take responsibility here for initializing it to zero.
4572 // (Note: Because tightly_coupled_allocation performs checks on the
4573 // out-edges of the dest, we need to avoid making derived pointers
4574 // from it until we have checked its uses.)
4575 if (ReduceBulkZeroing
4576 && !ZeroTLAB // pointless if already zeroed
4577 && basic_elem_type != T_CONFLICT // avoid corner case
4578 && !src->eqv_uncast(dest)
4579 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4580 != NULL)
4581 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4582 && alloc->maybe_set_complete(&_gvn)) {
4583 // "You break it, you buy it."
4584 InitializeNode* init = alloc->initialization();
4585 assert(init->is_complete(), "we just did this");
4586 init->set_complete_with_arraycopy();
4587 assert(dest->is_CheckCastPP(), "sanity");
4588 assert(dest->in(0)->in(0) == init, "dest pinned");
4589 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4590 // From this point on, every exit path is responsible for
4591 // initializing any non-copied parts of the object to zero.
4592 // Also, if this flag is set we make sure that arraycopy interacts properly
4593 // with G1, eliding pre-barriers. See CR 6627983.
4594 dest_uninitialized = true;
4595 } else {
4596 // No zeroing elimination here.
4597 alloc = NULL;
4598 //original_dest = dest;
4599 //dest_uninitialized = false;
4600 }
4601
4602 // Results are placed here:
4603 enum { fast_path = 1, // normal void-returning assembly stub
4604 checked_path = 2, // special assembly stub with cleanup
4605 slow_call_path = 3, // something went wrong; call the VM
4606 zero_path = 4, // bypass when length of copy is zero
4607 bcopy_path = 5, // copy primitive array by 64-bit blocks
4608 PATH_LIMIT = 6
4609 };
4610 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4611 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4612 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4613 record_for_igvn(result_region);
4614 _gvn.set_type_bottom(result_i_o);
4615 _gvn.set_type_bottom(result_memory);
4616 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4617
4618 // The slow_control path:
4619 Node* slow_control;
4620 Node* slow_i_o = i_o();
4621 Node* slow_mem = memory(adr_type);
4622 debug_only(slow_control = (Node*) badAddress);
4623
4624 // Checked control path:
4625 Node* checked_control = top();
4626 Node* checked_mem = NULL;
4627 Node* checked_i_o = NULL;
4628 Node* checked_value = NULL;
4629
4630 if (basic_elem_type == T_CONFLICT) {
4631 assert(!dest_uninitialized, "");
4632 Node* cv = generate_generic_arraycopy(adr_type,
4633 src, src_offset, dest, dest_offset,
4634 copy_length, dest_uninitialized);
4635 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4636 checked_control = control();
4637 checked_i_o = i_o();
4638 checked_mem = memory(adr_type);
4639 checked_value = cv;
4640 set_control(top()); // no fast path
4641 }
4642
4643 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4644 if (not_pos != NULL) {
4645 PreserveJVMState pjvms(this);
4646 set_control(not_pos);
4647
4648 // (6) length must not be negative.
4649 if (!length_never_negative) {
4650 generate_negative_guard(copy_length, slow_region);
4651 }
4652
4653 // copy_length is 0.
4654 if (!stopped() && dest_uninitialized) {
4655 Node* dest_length = alloc->in(AllocateNode::ALength);
4656 if (copy_length->eqv_uncast(dest_length)
4657 || _gvn.find_int_con(dest_length, 1) <= 0) {
4658 // There is no zeroing to do. No need for a secondary raw memory barrier.
4659 } else {
4660 // Clear the whole thing since there are no source elements to copy.
4661 generate_clear_array(adr_type, dest, basic_elem_type,
4662 intcon(0), NULL,
4663 alloc->in(AllocateNode::AllocSize));
4664 // Use a secondary InitializeNode as raw memory barrier.
4665 // Currently it is needed only on this path since other
4666 // paths have stub or runtime calls as raw memory barriers.
4667 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4668 Compile::AliasIdxRaw,
4669 top())->as_Initialize();
4670 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4671 }
4672 }
4673
4674 // Present the results of the fast call.
4675 result_region->init_req(zero_path, control());
4676 result_i_o ->init_req(zero_path, i_o());
4677 result_memory->init_req(zero_path, memory(adr_type));
4678 }
4679
4680 if (!stopped() && dest_uninitialized) {
4681 // We have to initialize the *uncopied* part of the array to zero.
4682 // The copy destination is the slice dest[off..off+len]. The other slices
4683 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4684 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4685 Node* dest_length = alloc->in(AllocateNode::ALength);
4686 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4687 copy_length) );
4688
4689 // If there is a head section that needs zeroing, do it now.
4690 if (find_int_con(dest_offset, -1) != 0) {
4691 generate_clear_array(adr_type, dest, basic_elem_type,
4692 intcon(0), dest_offset,
4693 NULL);
4694 }
4695
4696 // Next, perform a dynamic check on the tail length.
4697 // It is often zero, and we can win big if we prove this.
4698 // There are two wins: Avoid generating the ClearArray
4699 // with its attendant messy index arithmetic, and upgrade
4700 // the copy to a more hardware-friendly word size of 64 bits.
4701 Node* tail_ctl = NULL;
4702 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4703 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4704 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4705 tail_ctl = generate_slow_guard(bol_lt, NULL);
4706 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4707 }
4708
4709 // At this point, let's assume there is no tail.
4710 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4711 // There is no tail. Try an upgrade to a 64-bit copy.
4712 bool didit = false;
4713 { PreserveJVMState pjvms(this);
4714 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4715 src, src_offset, dest, dest_offset,
4716 dest_size, dest_uninitialized);
4717 if (didit) {
4718 // Present the results of the block-copying fast call.
4719 result_region->init_req(bcopy_path, control());
4720 result_i_o ->init_req(bcopy_path, i_o());
4721 result_memory->init_req(bcopy_path, memory(adr_type));
4722 }
4723 }
4724 if (didit)
4725 set_control(top()); // no regular fast path
4726 }
4727
4728 // Clear the tail, if any.
4729 if (tail_ctl != NULL) {
4730 Node* notail_ctl = stopped() ? NULL : control();
4731 set_control(tail_ctl);
4732 if (notail_ctl == NULL) {
4733 generate_clear_array(adr_type, dest, basic_elem_type,
4734 dest_tail, NULL,
4735 dest_size);
4736 } else {
4737 // Make a local merge.
4738 Node* done_ctl = new(C) RegionNode(3);
4739 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4740 done_ctl->init_req(1, notail_ctl);
4741 done_mem->init_req(1, memory(adr_type));
4742 generate_clear_array(adr_type, dest, basic_elem_type,
4743 dest_tail, NULL,
4744 dest_size);
4745 done_ctl->init_req(2, control());
4746 done_mem->init_req(2, memory(adr_type));
4747 set_control( _gvn.transform(done_ctl) );
4748 set_memory( _gvn.transform(done_mem), adr_type );
4749 }
4750 }
4751 }
4752
4753 BasicType copy_type = basic_elem_type;
4754 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4755 if (!stopped() && copy_type == T_OBJECT) {
4756 // If src and dest have compatible element types, we can copy bits.
4757 // Types S[] and D[] are compatible if D is a supertype of S.
4758 //
4759 // If they are not, we will use checked_oop_disjoint_arraycopy,
4760 // which performs a fast optimistic per-oop check, and backs off
4761 // further to JVM_ArrayCopy on the first per-oop check that fails.
4762 // (Actually, we don't move raw bits only; the GC requires card marks.)
4763
4764 // Get the Klass* for both src and dest
4765 Node* src_klass = load_object_klass(src);
4766 Node* dest_klass = load_object_klass(dest);
4767
4768 // Generate the subtype check.
4769 // This might fold up statically, or then again it might not.
4770 //
4771 // Non-static example: Copying List<String>.elements to a new String[].
4772 // The backing store for a List<String> is always an Object[],
4773 // but its elements are always type String, if the generic types
4774 // are correct at the source level.
4775 //
4776 // Test S[] against D[], not S against D, because (probably)
4777 // the secondary supertype cache is less busy for S[] than S.
4778 // This usually only matters when D is an interface.
4779 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4780 // Plug failing path into checked_oop_disjoint_arraycopy
4781 if (not_subtype_ctrl != top()) {
4782 PreserveJVMState pjvms(this);
4783 set_control(not_subtype_ctrl);
4784 // (At this point we can assume disjoint_bases, since types differ.)
4785 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
4786 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4787 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4788 Node* dest_elem_klass = _gvn.transform(n1);
4789 Node* cv = generate_checkcast_arraycopy(adr_type,
4790 dest_elem_klass,
4791 src, src_offset, dest, dest_offset,
4792 ConvI2X(copy_length), dest_uninitialized);
4793 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4794 checked_control = control();
4795 checked_i_o = i_o();
4796 checked_mem = memory(adr_type);
4797 checked_value = cv;
4798 }
4799 // At this point we know we do not need type checks on oop stores.
4800
4801 // Let's see if we need card marks:
4802 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4803 // If we do not need card marks, copy using the jint or jlong stub.
4804 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4805 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4806 "sizes agree");
4807 }
4808 }
4809
4810 if (!stopped()) {
4811 // Generate the fast path, if possible.
4812 PreserveJVMState pjvms(this);
4813 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4814 src, src_offset, dest, dest_offset,
4815 ConvI2X(copy_length), dest_uninitialized);
4816
4817 // Present the results of the fast call.
4818 result_region->init_req(fast_path, control());
4819 result_i_o ->init_req(fast_path, i_o());
4820 result_memory->init_req(fast_path, memory(adr_type));
4821 }
4822
4823 // Here are all the slow paths up to this point, in one bundle:
4824 slow_control = top();
4825 if (slow_region != NULL)
4826 slow_control = _gvn.transform(slow_region);
4827 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4828
4829 set_control(checked_control);
4830 if (!stopped()) {
4831 // Clean up after the checked call.
4832 // The returned value is either 0 or -1^K,
4833 // where K = number of partially transferred array elements.
4834 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4835 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4836 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4837
4838 // If it is 0, we are done, so transfer to the end.
4839 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4840 result_region->init_req(checked_path, checks_done);
4841 result_i_o ->init_req(checked_path, checked_i_o);
4842 result_memory->init_req(checked_path, checked_mem);
4843
4844 // If it is not zero, merge into the slow call.
4845 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4846 RegionNode* slow_reg2 = new(C) RegionNode(3);
4847 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4848 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4849 record_for_igvn(slow_reg2);
4850 slow_reg2 ->init_req(1, slow_control);
4851 slow_i_o2 ->init_req(1, slow_i_o);
4852 slow_mem2 ->init_req(1, slow_mem);
4853 slow_reg2 ->init_req(2, control());
4854 slow_i_o2 ->init_req(2, checked_i_o);
4855 slow_mem2 ->init_req(2, checked_mem);
4856
4857 slow_control = _gvn.transform(slow_reg2);
4858 slow_i_o = _gvn.transform(slow_i_o2);
4859 slow_mem = _gvn.transform(slow_mem2);
4860
4861 if (alloc != NULL) {
4862 // We'll restart from the very beginning, after zeroing the whole thing.
4863 // This can cause double writes, but that's OK since dest is brand new.
4864 // So we ignore the low 31 bits of the value returned from the stub.
4865 } else {
4866 // We must continue the copy exactly where it failed, or else
4867 // another thread might see the wrong number of writes to dest.
4868 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4869 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4870 slow_offset->init_req(1, intcon(0));
4871 slow_offset->init_req(2, checked_offset);
4872 slow_offset = _gvn.transform(slow_offset);
4873
4874 // Adjust the arguments by the conditionally incoming offset.
4875 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4876 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4877 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4878
4879 // Tweak the node variables to adjust the code produced below:
4880 src_offset = src_off_plus;
4881 dest_offset = dest_off_plus;
4882 copy_length = length_minus;
4883 }
4884 }
4885
4886 set_control(slow_control);
4887 if (!stopped()) {
4888 // Generate the slow path, if needed.
4889 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4890
4891 set_memory(slow_mem, adr_type);
4892 set_i_o(slow_i_o);
4893
4894 if (dest_uninitialized) {
4895 generate_clear_array(adr_type, dest, basic_elem_type,
4896 intcon(0), NULL,
4897 alloc->in(AllocateNode::AllocSize));
4898 }
4899
4900 generate_slow_arraycopy(adr_type,
4901 src, src_offset, dest, dest_offset,
4902 copy_length, /*dest_uninitialized*/false);
4903
4904 result_region->init_req(slow_call_path, control());
4905 result_i_o ->init_req(slow_call_path, i_o());
4906 result_memory->init_req(slow_call_path, memory(adr_type));
4907 }
4908
4909 // Remove unused edges.
4910 for (uint i = 1; i < result_region->req(); i++) {
4911 if (result_region->in(i) == NULL)
4912 result_region->init_req(i, top());
4913 }
4914
4915 // Finished; return the combined state.
4916 set_control( _gvn.transform(result_region) );
4917 set_i_o( _gvn.transform(result_i_o) );
4918 set_memory( _gvn.transform(result_memory), adr_type );
4919
4920 // The memory edges above are precise in order to model effects around
4921 // array copies accurately to allow value numbering of field loads around
4922 // arraycopy. Such field loads, both before and after, are common in Java
4923 // collections and similar classes involving header/array data structures.
4924 //
4925 // But with low number of register or when some registers are used or killed
4926 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4927 // The next memory barrier is added to avoid it. If the arraycopy can be
4928 // optimized away (which it can, sometimes) then we can manually remove
4929 // the membar also.
4930 //
4931 // Do not let reads from the cloned object float above the arraycopy.
4932 if (alloc != NULL) {
4933 // Do not let stores that initialize this object be reordered with
4934 // a subsequent store that would make this object accessible by
4935 // other threads.
4936 // Record what AllocateNode this StoreStore protects so that
4937 // escape analysis can go from the MemBarStoreStoreNode to the
4938 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4939 // based on the escape status of the AllocateNode.
4940 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4941 } else if (InsertMemBarAfterArraycopy)
4942 insert_mem_bar(Op_MemBarCPUOrder);
4943 }
4944
4945
4946 // Helper function which determines if an arraycopy immediately follows
4947 // an allocation, with no intervening tests or other escapes for the object.
4948 AllocateArrayNode*
4949 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4950 RegionNode* slow_region) {
4951 if (stopped()) return NULL; // no fast path
4952 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4953
4954 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4955 if (alloc == NULL) return NULL;
4956
4957 Node* rawmem = memory(Compile::AliasIdxRaw);
4958 // Is the allocation's memory state untouched?
4959 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4960 // Bail out if there have been raw-memory effects since the allocation.
4961 // (Example: There might have been a call or safepoint.)
4962 return NULL;
4963 }
4964 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4965 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4966 return NULL;
4967 }
4968
4969 // There must be no unexpected observers of this allocation.
4970 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4971 Node* obs = ptr->fast_out(i);
4972 if (obs != this->map()) {
4973 return NULL;
4974 }
4975 }
4976
4977 // This arraycopy must unconditionally follow the allocation of the ptr.
4978 Node* alloc_ctl = ptr->in(0);
4979 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4980
4981 Node* ctl = control();
4982 while (ctl != alloc_ctl) {
4983 // There may be guards which feed into the slow_region.
4984 // Any other control flow means that we might not get a chance
4985 // to finish initializing the allocated object.
4986 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
4987 IfNode* iff = ctl->in(0)->as_If();
4988 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
4989 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
4990 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
4991 ctl = iff->in(0); // This test feeds the known slow_region.
4992 continue;
4993 }
4994 // One more try: Various low-level checks bottom out in
4995 // uncommon traps. If the debug-info of the trap omits
4996 // any reference to the allocation, as we've already
4997 // observed, then there can be no objection to the trap.
4998 bool found_trap = false;
4999 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5000 Node* obs = not_ctl->fast_out(j);
5001 if (obs->in(0) == not_ctl && obs->is_Call() &&
5002 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5003 found_trap = true; break;
5004 }
5005 }
5006 if (found_trap) {
5007 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5008 continue;
5009 }
5010 }
5011 return NULL;
5012 }
5013
5014 // If we get this far, we have an allocation which immediately
5015 // precedes the arraycopy, and we can take over zeroing the new object.
5016 // The arraycopy will finish the initialization, and provide
5017 // a new control state to which we will anchor the destination pointer.
5018
5019 return alloc;
5020 }
5021
5022 // Helper for initialization of arrays, creating a ClearArray.
5023 // It writes zero bits in [start..end), within the body of an array object.
5024 // The memory effects are all chained onto the 'adr_type' alias category.
5025 //
5026 // Since the object is otherwise uninitialized, we are free
5027 // to put a little "slop" around the edges of the cleared area,
5028 // as long as it does not go back into the array's header,
5029 // or beyond the array end within the heap.
5030 //
5031 // The lower edge can be rounded down to the nearest jint and the
5032 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5033 //
5034 // Arguments:
5035 // adr_type memory slice where writes are generated
5036 // dest oop of the destination array
5037 // basic_elem_type element type of the destination
5038 // slice_idx array index of first element to store
5039 // slice_len number of elements to store (or NULL)
5040 // dest_size total size in bytes of the array object
5041 //
5042 // Exactly one of slice_len or dest_size must be non-NULL.
5043 // If dest_size is non-NULL, zeroing extends to the end of the object.
5044 // If slice_len is non-NULL, the slice_idx value must be a constant.
5045 void
5046 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5047 Node* dest,
5048 BasicType basic_elem_type,
5049 Node* slice_idx,
5050 Node* slice_len,
5051 Node* dest_size) {
5052 // one or the other but not both of slice_len and dest_size:
5053 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5054 if (slice_len == NULL) slice_len = top();
5055 if (dest_size == NULL) dest_size = top();
5056
5057 // operate on this memory slice:
5058 Node* mem = memory(adr_type); // memory slice to operate on
5059
5060 // scaling and rounding of indexes:
5061 int scale = exact_log2(type2aelembytes(basic_elem_type));
5062 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5063 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5064 int bump_bit = (-1 << scale) & BytesPerInt;
5065
5066 // determine constant starts and ends
5067 const intptr_t BIG_NEG = -128;
5068 assert(BIG_NEG + 2*abase < 0, "neg enough");
5069 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5070 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5071 if (slice_len_con == 0) {
5072 return; // nothing to do here
5073 }
5074 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5075 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5076 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5077 assert(end_con < 0, "not two cons");
5078 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5079 BytesPerLong);
5080 }
5081
5082 if (start_con >= 0 && end_con >= 0) {
5083 // Constant start and end. Simple.
5084 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5085 start_con, end_con, &_gvn);
5086 } else if (start_con >= 0 && dest_size != top()) {
5087 // Constant start, pre-rounded end after the tail of the array.
5088 Node* end = dest_size;
5089 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5090 start_con, end, &_gvn);
5091 } else if (start_con >= 0 && slice_len != top()) {
5092 // Constant start, non-constant end. End needs rounding up.
5093 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5094 intptr_t end_base = abase + (slice_idx_con << scale);
5095 int end_round = (-1 << scale) & (BytesPerLong - 1);
5096 Node* end = ConvI2X(slice_len);
5097 if (scale != 0)
5098 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5099 end_base += end_round;
5100 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5101 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5102 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5103 start_con, end, &_gvn);
5104 } else if (start_con < 0 && dest_size != top()) {
5105 // Non-constant start, pre-rounded end after the tail of the array.
5106 // This is almost certainly a "round-to-end" operation.
5107 Node* start = slice_idx;
5108 start = ConvI2X(start);
5109 if (scale != 0)
5110 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5111 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5112 if ((bump_bit | clear_low) != 0) {
5113 int to_clear = (bump_bit | clear_low);
5114 // Align up mod 8, then store a jint zero unconditionally
5115 // just before the mod-8 boundary.
5116 if (((abase + bump_bit) & ~to_clear) - bump_bit
5117 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5118 bump_bit = 0;
5119 assert((abase & to_clear) == 0, "array base must be long-aligned");
5120 } else {
5121 // Bump 'start' up to (or past) the next jint boundary:
5122 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5123 assert((abase & clear_low) == 0, "array base must be int-aligned");
5124 }
5125 // Round bumped 'start' down to jlong boundary in body of array.
5126 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5127 if (bump_bit != 0) {
5128 // Store a zero to the immediately preceding jint:
5129 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5130 Node* p1 = basic_plus_adr(dest, x1);
5131 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5132 mem = _gvn.transform(mem);
5133 }
5134 }
5135 Node* end = dest_size; // pre-rounded
5136 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5137 start, end, &_gvn);
5138 } else {
5139 // Non-constant start, unrounded non-constant end.
5140 // (Nobody zeroes a random midsection of an array using this routine.)
5141 ShouldNotReachHere(); // fix caller
5142 }
5143
5144 // Done.
5145 set_memory(mem, adr_type);
5146 }
5147
5148
5149 bool
5150 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5151 BasicType basic_elem_type,
5152 AllocateNode* alloc,
5153 Node* src, Node* src_offset,
5154 Node* dest, Node* dest_offset,
5155 Node* dest_size, bool dest_uninitialized) {
5156 // See if there is an advantage from block transfer.
5157 int scale = exact_log2(type2aelembytes(basic_elem_type));
5158 if (scale >= LogBytesPerLong)
5159 return false; // it is already a block transfer
5160
5161 // Look at the alignment of the starting offsets.
5162 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5163
5164 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5165 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5166 if (src_off_con < 0 || dest_off_con < 0)
5167 // At present, we can only understand constants.
5168 return false;
5169
5170 intptr_t src_off = abase + (src_off_con << scale);
5171 intptr_t dest_off = abase + (dest_off_con << scale);
5172
5173 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5174 // Non-aligned; too bad.
5175 // One more chance: Pick off an initial 32-bit word.
5176 // This is a common case, since abase can be odd mod 8.
5177 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5178 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5179 Node* sptr = basic_plus_adr(src, src_off);
5180 Node* dptr = basic_plus_adr(dest, dest_off);
5181 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5182 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5183 src_off += BytesPerInt;
5184 dest_off += BytesPerInt;
5185 } else {
5186 return false;
5187 }
5188 }
5189 assert(src_off % BytesPerLong == 0, "");
5190 assert(dest_off % BytesPerLong == 0, "");
5191
5192 // Do this copy by giant steps.
5193 Node* sptr = basic_plus_adr(src, src_off);
5194 Node* dptr = basic_plus_adr(dest, dest_off);
5195 Node* countx = dest_size;
5196 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5197 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5198
5199 bool disjoint_bases = true; // since alloc != NULL
5200 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5201 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5202
5203 return true;
5204 }
5205
5206
5207 // Helper function; generates code for the slow case.
5208 // We make a call to a runtime method which emulates the native method,
5209 // but without the native wrapper overhead.
5210 void
5211 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5212 Node* src, Node* src_offset,
5213 Node* dest, Node* dest_offset,
5214 Node* copy_length, bool dest_uninitialized) {
5215 assert(!dest_uninitialized, "Invariant");
5216 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5217 OptoRuntime::slow_arraycopy_Type(),
5218 OptoRuntime::slow_arraycopy_Java(),
5219 "slow_arraycopy", adr_type,
5220 src, src_offset, dest, dest_offset,
5221 copy_length);
5222
5223 // Handle exceptions thrown by this fellow:
5224 make_slow_call_ex(call, env()->Throwable_klass(), false);
5225 }
5226
5227 // Helper function; generates code for cases requiring runtime checks.
5228 Node*
5229 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5230 Node* dest_elem_klass,
5231 Node* src, Node* src_offset,
5232 Node* dest, Node* dest_offset,
5233 Node* copy_length, bool dest_uninitialized) {
5234 if (stopped()) return NULL;
5235
5236 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5237 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5238 return NULL;
5239 }
5240
5241 // Pick out the parameters required to perform a store-check
5242 // for the target array. This is an optimistic check. It will
5243 // look in each non-null element's class, at the desired klass's
5244 // super_check_offset, for the desired klass.
5245 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5246 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5247 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5248 Node* check_offset = ConvI2X(_gvn.transform(n3));
5249 Node* check_value = dest_elem_klass;
5250
5251 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5252 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5253
5254 // (We know the arrays are never conjoint, because their types differ.)
5255 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5256 OptoRuntime::checkcast_arraycopy_Type(),
5257 copyfunc_addr, "checkcast_arraycopy", adr_type,
5258 // five arguments, of which two are
5259 // intptr_t (jlong in LP64)
5260 src_start, dest_start,
5261 copy_length XTOP,
5262 check_offset XTOP,
5263 check_value);
5264
5265 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5266 }
5267
5268
5269 // Helper function; generates code for cases requiring runtime checks.
5270 Node*
5271 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5272 Node* src, Node* src_offset,
5273 Node* dest, Node* dest_offset,
5274 Node* copy_length, bool dest_uninitialized) {
5275 assert(!dest_uninitialized, "Invariant");
5276 if (stopped()) return NULL;
5277 address copyfunc_addr = StubRoutines::generic_arraycopy();
5278 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5279 return NULL;
5280 }
5281
5282 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5283 OptoRuntime::generic_arraycopy_Type(),
5284 copyfunc_addr, "generic_arraycopy", adr_type,
5285 src, src_offset, dest, dest_offset, copy_length);
5286
5287 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5288 }
5289
5290 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5291 void
5292 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5293 BasicType basic_elem_type,
5294 bool disjoint_bases,
5295 Node* src, Node* src_offset,
5296 Node* dest, Node* dest_offset,
5297 Node* copy_length, bool dest_uninitialized) {
5298 if (stopped()) return; // nothing to do
5299
5300 Node* src_start = src;
5301 Node* dest_start = dest;
5302 if (src_offset != NULL || dest_offset != NULL) {
5303 assert(src_offset != NULL && dest_offset != NULL, "");
5304 src_start = array_element_address(src, src_offset, basic_elem_type);
5305 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5306 }
5307
5308 // Figure out which arraycopy runtime method to call.
5309 const char* copyfunc_name = "arraycopy";
5310 address copyfunc_addr =
5311 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5312 disjoint_bases, copyfunc_name, dest_uninitialized);
5313
5314 // Call it. Note that the count_ix value is not scaled to a byte-size.
5315 make_runtime_call(RC_LEAF|RC_NO_FP,
5316 OptoRuntime::fast_arraycopy_Type(),
5317 copyfunc_addr, copyfunc_name, adr_type,
5318 src_start, dest_start, copy_length XTOP);
5319 }
5320
5321 //-------------inline_encodeISOArray-----------------------------------
5322 // encode char[] to byte[] in ISO_8859_1
5323 bool LibraryCallKit::inline_encodeISOArray() {
5324 assert(callee()->signature()->size() == 5, "encodeISOArray has 5 parameters");
5325 // no receiver since it is static method
5326 Node *src = argument(0);
5327 Node *src_offset = argument(1);
5328 Node *dst = argument(2);
5329 Node *dst_offset = argument(3);
5330 Node *length = argument(4);
5331
5332 const Type* src_type = src->Value(&_gvn);
5333 const Type* dst_type = dst->Value(&_gvn);
5334 const TypeAryPtr* top_src = src_type->isa_aryptr();
5335 const TypeAryPtr* top_dest = dst_type->isa_aryptr();
5336 if (top_src == NULL || top_src->klass() == NULL ||
5337 top_dest == NULL || top_dest->klass() == NULL) {
5338 // failed array check
5339 return false;
5340 }
5341
5342 // Figure out the size and type of the elements we will be copying.
5343 BasicType src_elem = src_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5344 BasicType dst_elem = dst_type->isa_aryptr()->klass()->as_array_klass()->element_type()->basic_type();
5345 if (src_elem != T_CHAR || dst_elem != T_BYTE) {
5346 return false;
5347 }
5348 Node* src_start = array_element_address(src, src_offset, src_elem);
5349 Node* dst_start = array_element_address(dst, dst_offset, dst_elem);
5350 // 'src_start' points to src array + scaled offset
5351 // 'dst_start' points to dst array + scaled offset
5352
5353 const TypeAryPtr* mtype = TypeAryPtr::BYTES;
5354 Node* enc = new (C) EncodeISOArrayNode(control(), memory(mtype), src_start, dst_start, length);
5355 enc = _gvn.transform(enc);
5356 Node* res_mem = _gvn.transform(new (C) SCMemProjNode(enc));
5357 set_memory(res_mem, mtype);
5358 set_result(enc);
5359 return true;
5360 }
5361
5362 //----------------------------inline_reference_get----------------------------
5363 // public T java.lang.ref.Reference.get();
5364 bool LibraryCallKit::inline_reference_get() {
5365 const int referent_offset = java_lang_ref_Reference::referent_offset;
5366 guarantee(referent_offset > 0, "should have already been set");
5367
5368 // Get the argument:
5369 Node* reference_obj = null_check_receiver();
5370 if (stopped()) return true;
5371
5372 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5373
5374 ciInstanceKlass* klass = env()->Object_klass();
5375 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5376
5377 Node* no_ctrl = NULL;
5378 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5379
5380 // Use the pre-barrier to record the value in the referent field
5381 pre_barrier(false /* do_load */,
5382 control(),
5383 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5384 result /* pre_val */,
5385 T_OBJECT);
5386
5387 // Add memory barrier to prevent commoning reads from this field
5388 // across safepoint since GC can change its value.
5389 insert_mem_bar(Op_MemBarCPUOrder);
5390
5391 set_result(result);
5392 return true;
5393 }
5394
5395
5396 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5397 bool is_exact=true, bool is_static=false) {
5398
5399 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5400 assert(tinst != NULL, "obj is null");
5401 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5402 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5403
5404 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5405 ciSymbol::make(fieldTypeString),
5406 is_static);
5407 if (field == NULL) return (Node *) NULL;
5408 assert (field != NULL, "undefined field");
5409
5410 // Next code copied from Parse::do_get_xxx():
5411
5412 // Compute address and memory type.
5413 int offset = field->offset_in_bytes();
5414 bool is_vol = field->is_volatile();
5415 ciType* field_klass = field->type();
5416 assert(field_klass->is_loaded(), "should be loaded");
5417 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5418 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5419 BasicType bt = field->layout_type();
5420
5421 // Build the resultant type of the load
5422 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5423
5424 // Build the load.
5425 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5426 return loadedField;
5427 }
5428
5429
5430 //------------------------------inline_aescrypt_Block-----------------------
5431 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5432 address stubAddr;
5433 const char *stubName;
5434 assert(UseAES, "need AES instruction support");
5435
5436 switch(id) {
5437 case vmIntrinsics::_aescrypt_encryptBlock:
5438 stubAddr = StubRoutines::aescrypt_encryptBlock();
5439 stubName = "aescrypt_encryptBlock";
5440 break;
5441 case vmIntrinsics::_aescrypt_decryptBlock:
5442 stubAddr = StubRoutines::aescrypt_decryptBlock();
5443 stubName = "aescrypt_decryptBlock";
5444 break;
5445 }
5446 if (stubAddr == NULL) return false;
5447
5448 Node* aescrypt_object = argument(0);
5449 Node* src = argument(1);
5450 Node* src_offset = argument(2);
5451 Node* dest = argument(3);
5452 Node* dest_offset = argument(4);
5453
5454 // (1) src and dest are arrays.
5455 const Type* src_type = src->Value(&_gvn);
5456 const Type* dest_type = dest->Value(&_gvn);
5457 const TypeAryPtr* top_src = src_type->isa_aryptr();
5458 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5459 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5460
5461 // for the quick and dirty code we will skip all the checks.
5462 // we are just trying to get the call to be generated.
5463 Node* src_start = src;
5464 Node* dest_start = dest;
5465 if (src_offset != NULL || dest_offset != NULL) {
5466 assert(src_offset != NULL && dest_offset != NULL, "");
5467 src_start = array_element_address(src, src_offset, T_BYTE);
5468 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5469 }
5470
5471 // now need to get the start of its expanded key array
5472 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5473 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5474 if (k_start == NULL) return false;
5475
5476 // Call the stub.
5477 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5478 stubAddr, stubName, TypePtr::BOTTOM,
5479 src_start, dest_start, k_start);
5480
5481 return true;
5482 }
5483
5484 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5485 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5486 address stubAddr;
5487 const char *stubName;
5488
5489 assert(UseAES, "need AES instruction support");
5490
5491 switch(id) {
5492 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5493 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5494 stubName = "cipherBlockChaining_encryptAESCrypt";
5495 break;
5496 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5497 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5498 stubName = "cipherBlockChaining_decryptAESCrypt";
5499 break;
5500 }
5501 if (stubAddr == NULL) return false;
5502
5503 Node* cipherBlockChaining_object = argument(0);
5504 Node* src = argument(1);
5505 Node* src_offset = argument(2);
5506 Node* len = argument(3);
5507 Node* dest = argument(4);
5508 Node* dest_offset = argument(5);
5509
5510 // (1) src and dest are arrays.
5511 const Type* src_type = src->Value(&_gvn);
5512 const Type* dest_type = dest->Value(&_gvn);
5513 const TypeAryPtr* top_src = src_type->isa_aryptr();
5514 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5515 assert (top_src != NULL && top_src->klass() != NULL
5516 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5517
5518 // checks are the responsibility of the caller
5519 Node* src_start = src;
5520 Node* dest_start = dest;
5521 if (src_offset != NULL || dest_offset != NULL) {
5522 assert(src_offset != NULL && dest_offset != NULL, "");
5523 src_start = array_element_address(src, src_offset, T_BYTE);
5524 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5525 }
5526
5527 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5528 // (because of the predicated logic executed earlier).
5529 // so we cast it here safely.
5530 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5531
5532 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5533 if (embeddedCipherObj == NULL) return false;
5534
5535 // cast it to what we know it will be at runtime
5536 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5537 assert(tinst != NULL, "CBC obj is null");
5538 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5539 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5540 if (!klass_AESCrypt->is_loaded()) return false;
5541
5542 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5543 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5544 const TypeOopPtr* xtype = aklass->as_instance_type();
5545 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5546 aescrypt_object = _gvn.transform(aescrypt_object);
5547
5548 // we need to get the start of the aescrypt_object's expanded key array
5549 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5550 if (k_start == NULL) return false;
5551
5552 // similarly, get the start address of the r vector
5553 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5554 if (objRvec == NULL) return false;
5555 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5556
5557 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5558 make_runtime_call(RC_LEAF|RC_NO_FP,
5559 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5560 stubAddr, stubName, TypePtr::BOTTOM,
5561 src_start, dest_start, k_start, r_start, len);
5562
5563 // return is void so no result needs to be pushed
5564
5565 return true;
5566 }
5567
5568 //------------------------------get_key_start_from_aescrypt_object-----------------------
5569 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5570 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5571 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5572 if (objAESCryptKey == NULL) return (Node *) NULL;
5573
5574 // now have the array, need to get the start address of the K array
5575 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5576 return k_start;
5577 }
5578
5579 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5580 // Return node representing slow path of predicate check.
5581 // the pseudo code we want to emulate with this predicate is:
5582 // for encryption:
5583 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5584 // for decryption:
5585 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5586 // note cipher==plain is more conservative than the original java code but that's OK
5587 //
5588 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5589 // First, check receiver for NULL since it is virtual method.
5590 Node* objCBC = argument(0);
5591 objCBC = null_check(objCBC);
5592
5593 if (stopped()) return NULL; // Always NULL
5594
5595 // Load embeddedCipher field of CipherBlockChaining object.
5596 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5597
5598 // get AESCrypt klass for instanceOf check
5599 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5600 // will have same classloader as CipherBlockChaining object
5601 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5602 assert(tinst != NULL, "CBCobj is null");
5603 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5604
5605 // we want to do an instanceof comparison against the AESCrypt class
5606 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5607 if (!klass_AESCrypt->is_loaded()) {
5608 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5609 Node* ctrl = control();
5610 set_control(top()); // no regular fast path
5611 return ctrl;
5612 }
5613 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5614
5615 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5616 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5617 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5618
5619 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5620
5621 // for encryption, we are done
5622 if (!decrypting)
5623 return instof_false; // even if it is NULL
5624
5625 // for decryption, we need to add a further check to avoid
5626 // taking the intrinsic path when cipher and plain are the same
5627 // see the original java code for why.
5628 RegionNode* region = new(C) RegionNode(3);
5629 region->init_req(1, instof_false);
5630 Node* src = argument(1);
5631 Node* dest = argument(4);
5632 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5633 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5634 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5635 region->init_req(2, src_dest_conjoint);
5636
5637 record_for_igvn(region);
5638 return _gvn.transform(region);
5639 }