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 #include "trace/traceMacros.hpp"
42
43 class LibraryIntrinsic : public InlineCallGenerator {
44 // Extend the set of intrinsics known to the runtime:
45 public:
46 private:
47 bool _is_virtual;
48 bool _is_predicted;
49 vmIntrinsics::ID _intrinsic_id;
50
51 public:
52 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id)
53 : InlineCallGenerator(m),
54 _is_virtual(is_virtual),
55 _is_predicted(is_predicted),
56 _intrinsic_id(id)
57 {
58 }
59 virtual bool is_intrinsic() const { return true; }
60 virtual bool is_virtual() const { return _is_virtual; }
61 virtual bool is_predicted() const { return _is_predicted; }
62 virtual JVMState* generate(JVMState* jvms);
63 virtual Node* generate_predicate(JVMState* jvms);
64 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
65 };
66
67
68 // Local helper class for LibraryIntrinsic:
69 class LibraryCallKit : public GraphKit {
70 private:
71 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
72 Node* _result; // the result node, if any
73 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
74
75 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
76
77 public:
78 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
79 : GraphKit(jvms),
80 _intrinsic(intrinsic),
81 _result(NULL)
82 {
83 // Check if this is a root compile. In that case we don't have a caller.
84 if (!jvms->has_method()) {
85 _reexecute_sp = sp();
86 } else {
87 // Find out how many arguments the interpreter needs when deoptimizing
88 // and save the stack pointer value so it can used by uncommon_trap.
89 // We find the argument count by looking at the declared signature.
90 bool ignored_will_link;
91 ciSignature* declared_signature = NULL;
92 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
93 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
94 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
95 }
96 }
97
98 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
99
100 ciMethod* caller() const { return jvms()->method(); }
101 int bci() const { return jvms()->bci(); }
102 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
103 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
104 ciMethod* callee() const { return _intrinsic->method(); }
105
106 bool try_to_inline();
107 Node* try_to_predicate();
108
109 void push_result() {
110 // Push the result onto the stack.
111 if (!stopped() && result() != NULL) {
112 BasicType bt = result()->bottom_type()->basic_type();
113 push_node(bt, result());
114 }
115 }
116
117 private:
118 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
119 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
120 }
121
122 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
123 void set_result(RegionNode* region, PhiNode* value);
124 Node* result() { return _result; }
125
126 virtual int reexecute_sp() { return _reexecute_sp; }
127
128 // Helper functions to inline natives
129 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
130 Node* generate_slow_guard(Node* test, RegionNode* region);
131 Node* generate_fair_guard(Node* test, RegionNode* region);
132 Node* generate_negative_guard(Node* index, RegionNode* region,
133 // resulting CastII of index:
134 Node* *pos_index = NULL);
135 Node* generate_nonpositive_guard(Node* index, bool never_negative,
136 // resulting CastII of index:
137 Node* *pos_index = NULL);
138 Node* generate_limit_guard(Node* offset, Node* subseq_length,
139 Node* array_length,
140 RegionNode* region);
141 Node* generate_current_thread(Node* &tls_output);
142 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
143 bool disjoint_bases, const char* &name, bool dest_uninitialized);
144 Node* load_mirror_from_klass(Node* klass);
145 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
146 RegionNode* region, int null_path,
147 int offset);
148 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
149 RegionNode* region, int null_path) {
150 int offset = java_lang_Class::klass_offset_in_bytes();
151 return load_klass_from_mirror_common(mirror, never_see_null,
152 region, null_path,
153 offset);
154 }
155 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
156 RegionNode* region, int null_path) {
157 int offset = java_lang_Class::array_klass_offset_in_bytes();
158 return load_klass_from_mirror_common(mirror, never_see_null,
159 region, null_path,
160 offset);
161 }
162 Node* generate_access_flags_guard(Node* kls,
163 int modifier_mask, int modifier_bits,
164 RegionNode* region);
165 Node* generate_interface_guard(Node* kls, RegionNode* region);
166 Node* generate_array_guard(Node* kls, RegionNode* region) {
167 return generate_array_guard_common(kls, region, false, false);
168 }
169 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
170 return generate_array_guard_common(kls, region, false, true);
171 }
172 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
173 return generate_array_guard_common(kls, region, true, false);
174 }
175 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
176 return generate_array_guard_common(kls, region, true, true);
177 }
178 Node* generate_array_guard_common(Node* kls, RegionNode* region,
179 bool obj_array, bool not_array);
180 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
181 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
182 bool is_virtual = false, bool is_static = false);
183 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
184 return generate_method_call(method_id, false, true);
185 }
186 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
187 return generate_method_call(method_id, true, false);
188 }
189 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
190
191 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
192 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
193 bool inline_string_compareTo();
194 bool inline_string_indexOf();
195 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
196 bool inline_string_equals();
197 Node* round_double_node(Node* n);
198 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
199 bool inline_math_native(vmIntrinsics::ID id);
200 bool inline_trig(vmIntrinsics::ID id);
201 bool inline_math(vmIntrinsics::ID id);
202 bool inline_exp();
203 bool inline_pow();
204 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
205 bool inline_min_max(vmIntrinsics::ID id);
206 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
207 // This returns Type::AnyPtr, RawPtr, or OopPtr.
208 int classify_unsafe_addr(Node* &base, Node* &offset);
209 Node* make_unsafe_address(Node* base, Node* offset);
210 // Helper for inline_unsafe_access.
211 // Generates the guards that check whether the result of
212 // Unsafe.getObject should be recorded in an SATB log buffer.
213 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
214 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
215 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
216 bool inline_unsafe_allocate();
217 bool inline_unsafe_copyMemory();
218 bool inline_native_currentThread();
219 #ifdef TRACE_HAVE_INTRINSICS
220 bool inline_native_classID();
221 bool inline_native_threadID();
222 #endif
223 bool inline_native_time_funcs(address method, const char* funcName);
224 bool inline_native_isInterrupted();
225 bool inline_native_Class_query(vmIntrinsics::ID id);
226 bool inline_native_subtype_check();
227
228 bool inline_native_newArray();
229 bool inline_native_getLength();
230 bool inline_array_copyOf(bool is_copyOfRange);
231 bool inline_array_equals();
232 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
233 bool inline_native_clone(bool is_virtual);
234 bool inline_native_Reflection_getCallerClass();
235 bool is_method_invoke_or_aux_frame(JVMState* jvms);
236 // Helper function for inlining native object hash method
237 bool inline_native_hashcode(bool is_virtual, bool is_static);
238 bool inline_native_getClass();
239
240 // Helper functions for inlining arraycopy
241 bool inline_arraycopy();
242 void generate_arraycopy(const TypePtr* adr_type,
243 BasicType basic_elem_type,
244 Node* src, Node* src_offset,
245 Node* dest, Node* dest_offset,
246 Node* copy_length,
247 bool disjoint_bases = false,
248 bool length_never_negative = false,
249 RegionNode* slow_region = NULL);
250 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
251 RegionNode* slow_region);
252 void generate_clear_array(const TypePtr* adr_type,
253 Node* dest,
254 BasicType basic_elem_type,
255 Node* slice_off,
256 Node* slice_len,
257 Node* slice_end);
258 bool generate_block_arraycopy(const TypePtr* adr_type,
259 BasicType basic_elem_type,
260 AllocateNode* alloc,
261 Node* src, Node* src_offset,
262 Node* dest, Node* dest_offset,
263 Node* dest_size, bool dest_uninitialized);
264 void generate_slow_arraycopy(const TypePtr* adr_type,
265 Node* src, Node* src_offset,
266 Node* dest, Node* dest_offset,
267 Node* copy_length, bool dest_uninitialized);
268 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
269 Node* dest_elem_klass,
270 Node* src, Node* src_offset,
271 Node* dest, Node* dest_offset,
272 Node* copy_length, bool dest_uninitialized);
273 Node* generate_generic_arraycopy(const TypePtr* adr_type,
274 Node* src, Node* src_offset,
275 Node* dest, Node* dest_offset,
276 Node* copy_length, bool dest_uninitialized);
277 void generate_unchecked_arraycopy(const TypePtr* adr_type,
278 BasicType basic_elem_type,
279 bool disjoint_bases,
280 Node* src, Node* src_offset,
281 Node* dest, Node* dest_offset,
282 Node* copy_length, bool dest_uninitialized);
283 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
284 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
285 bool inline_unsafe_ordered_store(BasicType type);
286 bool inline_fp_conversions(vmIntrinsics::ID id);
287 bool inline_number_methods(vmIntrinsics::ID id);
288 bool inline_reference_get();
289 bool inline_aescrypt_Block(vmIntrinsics::ID id);
290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
292 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
293 };
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 break; // InlineNatives does not control String.compareTo
339 case vmIntrinsics::_Reference_get:
340 break; // InlineNatives does not control Reference.get
341 default:
342 return NULL;
343 }
344 }
345
346 bool is_predicted = false;
347
348 switch (id) {
349 case vmIntrinsics::_compareTo:
350 if (!SpecialStringCompareTo) return NULL;
351 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
352 break;
353 case vmIntrinsics::_indexOf:
354 if (!SpecialStringIndexOf) return NULL;
355 break;
356 case vmIntrinsics::_equals:
357 if (!SpecialStringEquals) return NULL;
358 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
359 break;
360 case vmIntrinsics::_equalsC:
361 if (!SpecialArraysEquals) return NULL;
362 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
363 break;
364 case vmIntrinsics::_arraycopy:
365 if (!InlineArrayCopy) return NULL;
366 break;
367 case vmIntrinsics::_copyMemory:
368 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
369 if (!InlineArrayCopy) return NULL;
370 break;
371 case vmIntrinsics::_hashCode:
372 if (!InlineObjectHash) return NULL;
373 break;
374 case vmIntrinsics::_clone:
375 case vmIntrinsics::_copyOf:
376 case vmIntrinsics::_copyOfRange:
377 if (!InlineObjectCopy) return NULL;
378 // These also use the arraycopy intrinsic mechanism:
379 if (!InlineArrayCopy) return NULL;
380 break;
381 case vmIntrinsics::_checkIndex:
382 // We do not intrinsify this. The optimizer does fine with it.
383 return NULL;
384
385 case vmIntrinsics::_getCallerClass:
386 if (!UseNewReflection) return NULL;
387 if (!InlineReflectionGetCallerClass) return NULL;
388 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
389 break;
390
391 case vmIntrinsics::_bitCount_i:
392 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
393 break;
394
395 case vmIntrinsics::_bitCount_l:
396 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
397 break;
398
399 case vmIntrinsics::_numberOfLeadingZeros_i:
400 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
401 break;
402
403 case vmIntrinsics::_numberOfLeadingZeros_l:
404 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
405 break;
406
407 case vmIntrinsics::_numberOfTrailingZeros_i:
408 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
409 break;
410
411 case vmIntrinsics::_numberOfTrailingZeros_l:
412 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
413 break;
414
415 case vmIntrinsics::_reverseBytes_c:
416 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
417 break;
418 case vmIntrinsics::_reverseBytes_s:
419 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
420 break;
421 case vmIntrinsics::_reverseBytes_i:
422 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
423 break;
424 case vmIntrinsics::_reverseBytes_l:
425 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
426 break;
427
428 case vmIntrinsics::_Reference_get:
429 // Use the intrinsic version of Reference.get() so that the value in
430 // the referent field can be registered by the G1 pre-barrier code.
431 // Also add memory barrier to prevent commoning reads from this field
432 // across safepoint since GC can change it value.
433 break;
434
435 case vmIntrinsics::_compareAndSwapObject:
436 #ifdef _LP64
437 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
438 #endif
439 break;
440
441 case vmIntrinsics::_compareAndSwapLong:
442 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
443 break;
444
445 case vmIntrinsics::_getAndAddInt:
446 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
447 break;
448
449 case vmIntrinsics::_getAndAddLong:
450 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
451 break;
452
453 case vmIntrinsics::_getAndSetInt:
454 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
455 break;
456
457 case vmIntrinsics::_getAndSetLong:
458 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
459 break;
460
461 case vmIntrinsics::_getAndSetObject:
462 #ifdef _LP64
463 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
464 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
465 break;
466 #else
467 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
468 break;
469 #endif
470
471 case vmIntrinsics::_aescrypt_encryptBlock:
472 case vmIntrinsics::_aescrypt_decryptBlock:
473 if (!UseAESIntrinsics) return NULL;
474 break;
475
476 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
477 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
478 if (!UseAESIntrinsics) return NULL;
479 // these two require the predicated logic
480 is_predicted = true;
481 break;
482
483 default:
484 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
485 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
486 break;
487 }
488
489 // -XX:-InlineClassNatives disables natives from the Class class.
490 // The flag applies to all reflective calls, notably Array.newArray
491 // (visible to Java programmers as Array.newInstance).
492 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
493 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
494 if (!InlineClassNatives) return NULL;
495 }
496
497 // -XX:-InlineThreadNatives disables natives from the Thread class.
498 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
499 if (!InlineThreadNatives) return NULL;
500 }
501
502 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
503 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
504 m->holder()->name() == ciSymbol::java_lang_Float() ||
505 m->holder()->name() == ciSymbol::java_lang_Double()) {
506 if (!InlineMathNatives) return NULL;
507 }
508
509 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
510 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
511 if (!InlineUnsafeOps) return NULL;
512 }
513
514 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
515 }
516
517 //----------------------register_library_intrinsics-----------------------
518 // Initialize this file's data structures, for each Compile instance.
519 void Compile::register_library_intrinsics() {
520 // Nothing to do here.
521 }
522
523 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
524 LibraryCallKit kit(jvms, this);
525 Compile* C = kit.C;
526 int nodes = C->unique();
527 #ifndef PRODUCT
528 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
529 char buf[1000];
530 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
531 tty->print_cr("Intrinsic %s", str);
532 }
533 #endif
534 ciMethod* callee = kit.callee();
535 const int bci = kit.bci();
536
537 // Try to inline the intrinsic.
538 if (kit.try_to_inline()) {
539 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
540 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
541 }
542 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
543 if (C->log()) {
544 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
545 vmIntrinsics::name_at(intrinsic_id()),
546 (is_virtual() ? " virtual='1'" : ""),
547 C->unique() - nodes);
548 }
549 // Push the result from the inlined method onto the stack.
550 kit.push_result();
551 return kit.transfer_exceptions_into_jvms();
552 }
553
554 // The intrinsic bailed out
555 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
556 if (jvms->has_method()) {
557 // Not a root compile.
558 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
559 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
560 } else {
561 // Root compile
562 tty->print("Did not generate intrinsic %s%s at bci:%d in",
563 vmIntrinsics::name_at(intrinsic_id()),
564 (is_virtual() ? " (virtual)" : ""), bci);
565 }
566 }
567 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
568 return NULL;
569 }
570
571 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
572 LibraryCallKit kit(jvms, this);
573 Compile* C = kit.C;
574 int nodes = C->unique();
575 #ifndef PRODUCT
576 assert(is_predicted(), "sanity");
577 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
578 char buf[1000];
579 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
580 tty->print_cr("Predicate for intrinsic %s", str);
581 }
582 #endif
583 ciMethod* callee = kit.callee();
584 const int bci = kit.bci();
585
586 Node* slow_ctl = kit.try_to_predicate();
587 if (!kit.failing()) {
588 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
589 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
590 }
591 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
592 if (C->log()) {
593 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
594 vmIntrinsics::name_at(intrinsic_id()),
595 (is_virtual() ? " virtual='1'" : ""),
596 C->unique() - nodes);
597 }
598 return slow_ctl; // Could be NULL if the check folds.
599 }
600
601 // The intrinsic bailed out
602 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
603 if (jvms->has_method()) {
604 // Not a root compile.
605 const char* msg = "failed to generate predicate for intrinsic";
606 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
607 } else {
608 // Root compile
609 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
610 vmIntrinsics::name_at(intrinsic_id()),
611 (is_virtual() ? " (virtual)" : ""), bci);
612 }
613 }
614 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
615 return NULL;
616 }
617
618 bool LibraryCallKit::try_to_inline() {
619 // Handle symbolic names for otherwise undistinguished boolean switches:
620 const bool is_store = true;
621 const bool is_native_ptr = true;
622 const bool is_static = true;
623 const bool is_volatile = true;
624
625 if (!jvms()->has_method()) {
626 // Root JVMState has a null method.
627 assert(map()->memory()->Opcode() == Op_Parm, "");
628 // Insert the memory aliasing node
629 set_all_memory(reset_memory());
630 }
631 assert(merged_memory(), "");
632
633
634 switch (intrinsic_id()) {
635 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
636 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
637 case vmIntrinsics::_getClass: return inline_native_getClass();
638
639 case vmIntrinsics::_dsin:
640 case vmIntrinsics::_dcos:
641 case vmIntrinsics::_dtan:
642 case vmIntrinsics::_dabs:
643 case vmIntrinsics::_datan2:
644 case vmIntrinsics::_dsqrt:
645 case vmIntrinsics::_dexp:
646 case vmIntrinsics::_dlog:
647 case vmIntrinsics::_dlog10:
648 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
649
650 case vmIntrinsics::_min:
651 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
652
653 case vmIntrinsics::_arraycopy: return inline_arraycopy();
654
655 case vmIntrinsics::_compareTo: return inline_string_compareTo();
656 case vmIntrinsics::_indexOf: return inline_string_indexOf();
657 case vmIntrinsics::_equals: return inline_string_equals();
658
659 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
660 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
661 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
662 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
663 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
664 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
665 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
666 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
667 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
668
669 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
670 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
671 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
672 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
673 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
674 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
675 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
676 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
677 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
678
679 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
680 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
681 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
682 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
683 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
684 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
685 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
686 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
687
688 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
689 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
690 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
691 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
692 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
693 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
694 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
695 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
696
697 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
698 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
699 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
700 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
701 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
702 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
703 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
704 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
705 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
706
707 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
708 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
709 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
710 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
711 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
712 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
713 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
714 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
715 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
716
717 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
718 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
719 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
720 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
721
722 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
723 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
724 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
725
726 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
727 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
728 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
729
730 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
731 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
732 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
733 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
734 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
735
736 case vmIntrinsics::_currentThread: return inline_native_currentThread();
737 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
738
739 #ifdef TRACE_HAVE_INTRINSICS
740 case vmIntrinsics::_classID: return inline_native_classID();
741 case vmIntrinsics::_threadID: return inline_native_threadID();
742 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
743 #endif
744 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
745 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
746 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
747 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
748 case vmIntrinsics::_newArray: return inline_native_newArray();
749 case vmIntrinsics::_getLength: return inline_native_getLength();
750 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
751 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
752 case vmIntrinsics::_equalsC: return inline_array_equals();
753 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
754
755 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
756
757 case vmIntrinsics::_isInstance:
758 case vmIntrinsics::_getModifiers:
759 case vmIntrinsics::_isInterface:
760 case vmIntrinsics::_isArray:
761 case vmIntrinsics::_isPrimitive:
762 case vmIntrinsics::_getSuperclass:
763 case vmIntrinsics::_getComponentType:
764 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
765
766 case vmIntrinsics::_floatToRawIntBits:
767 case vmIntrinsics::_floatToIntBits:
768 case vmIntrinsics::_intBitsToFloat:
769 case vmIntrinsics::_doubleToRawLongBits:
770 case vmIntrinsics::_doubleToLongBits:
771 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
772
773 case vmIntrinsics::_numberOfLeadingZeros_i:
774 case vmIntrinsics::_numberOfLeadingZeros_l:
775 case vmIntrinsics::_numberOfTrailingZeros_i:
776 case vmIntrinsics::_numberOfTrailingZeros_l:
777 case vmIntrinsics::_bitCount_i:
778 case vmIntrinsics::_bitCount_l:
779 case vmIntrinsics::_reverseBytes_i:
780 case vmIntrinsics::_reverseBytes_l:
781 case vmIntrinsics::_reverseBytes_s:
782 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
783
784 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
785
786 case vmIntrinsics::_Reference_get: return inline_reference_get();
787
788 case vmIntrinsics::_aescrypt_encryptBlock:
789 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
790
791 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
792 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
793 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
794
795 default:
796 // If you get here, it may be that someone has added a new intrinsic
797 // to the list in vmSymbols.hpp without implementing it here.
798 #ifndef PRODUCT
799 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
800 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
801 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
802 }
803 #endif
804 return false;
805 }
806 }
807
808 Node* LibraryCallKit::try_to_predicate() {
809 if (!jvms()->has_method()) {
810 // Root JVMState has a null method.
811 assert(map()->memory()->Opcode() == Op_Parm, "");
812 // Insert the memory aliasing node
813 set_all_memory(reset_memory());
814 }
815 assert(merged_memory(), "");
816
817 switch (intrinsic_id()) {
818 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
819 return inline_cipherBlockChaining_AESCrypt_predicate(false);
820 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
821 return inline_cipherBlockChaining_AESCrypt_predicate(true);
822
823 default:
824 // If you get here, it may be that someone has added a new intrinsic
825 // to the list in vmSymbols.hpp without implementing it here.
826 #ifndef PRODUCT
827 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
828 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
829 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
830 }
831 #endif
832 Node* slow_ctl = control();
833 set_control(top()); // No fast path instrinsic
834 return slow_ctl;
835 }
836 }
837
838 //------------------------------set_result-------------------------------
839 // Helper function for finishing intrinsics.
840 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
841 record_for_igvn(region);
842 set_control(_gvn.transform(region));
843 set_result( _gvn.transform(value));
844 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
845 }
846
847 //------------------------------generate_guard---------------------------
848 // Helper function for generating guarded fast-slow graph structures.
849 // The given 'test', if true, guards a slow path. If the test fails
850 // then a fast path can be taken. (We generally hope it fails.)
851 // In all cases, GraphKit::control() is updated to the fast path.
852 // The returned value represents the control for the slow path.
853 // The return value is never 'top'; it is either a valid control
854 // or NULL if it is obvious that the slow path can never be taken.
855 // Also, if region and the slow control are not NULL, the slow edge
856 // is appended to the region.
857 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
858 if (stopped()) {
859 // Already short circuited.
860 return NULL;
861 }
862
863 // Build an if node and its projections.
864 // If test is true we take the slow path, which we assume is uncommon.
865 if (_gvn.type(test) == TypeInt::ZERO) {
866 // The slow branch is never taken. No need to build this guard.
867 return NULL;
868 }
869
870 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
871
872 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
873 if (if_slow == top()) {
874 // The slow branch is never taken. No need to build this guard.
875 return NULL;
876 }
877
878 if (region != NULL)
879 region->add_req(if_slow);
880
881 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
882 set_control(if_fast);
883
884 return if_slow;
885 }
886
887 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
888 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
889 }
890 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
891 return generate_guard(test, region, PROB_FAIR);
892 }
893
894 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
895 Node* *pos_index) {
896 if (stopped())
897 return NULL; // already stopped
898 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
899 return NULL; // index is already adequately typed
900 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
901 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
902 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
903 if (is_neg != NULL && pos_index != NULL) {
904 // Emulate effect of Parse::adjust_map_after_if.
905 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
906 ccast->set_req(0, control());
907 (*pos_index) = _gvn.transform(ccast);
908 }
909 return is_neg;
910 }
911
912 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
913 Node* *pos_index) {
914 if (stopped())
915 return NULL; // already stopped
916 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
917 return NULL; // index is already adequately typed
918 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
919 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
920 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
921 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
922 if (is_notp != NULL && pos_index != NULL) {
923 // Emulate effect of Parse::adjust_map_after_if.
924 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
925 ccast->set_req(0, control());
926 (*pos_index) = _gvn.transform(ccast);
927 }
928 return is_notp;
929 }
930
931 // Make sure that 'position' is a valid limit index, in [0..length].
932 // There are two equivalent plans for checking this:
933 // A. (offset + copyLength) unsigned<= arrayLength
934 // B. offset <= (arrayLength - copyLength)
935 // We require that all of the values above, except for the sum and
936 // difference, are already known to be non-negative.
937 // Plan A is robust in the face of overflow, if offset and copyLength
938 // are both hugely positive.
939 //
940 // Plan B is less direct and intuitive, but it does not overflow at
941 // all, since the difference of two non-negatives is always
942 // representable. Whenever Java methods must perform the equivalent
943 // check they generally use Plan B instead of Plan A.
944 // For the moment we use Plan A.
945 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
946 Node* subseq_length,
947 Node* array_length,
948 RegionNode* region) {
949 if (stopped())
950 return NULL; // already stopped
951 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
952 if (zero_offset && subseq_length->eqv_uncast(array_length))
953 return NULL; // common case of whole-array copy
954 Node* last = subseq_length;
955 if (!zero_offset) // last += offset
956 last = _gvn.transform( new (C) AddINode(last, offset));
957 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
958 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
959 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
960 return is_over;
961 }
962
963
964 //--------------------------generate_current_thread--------------------
965 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
966 ciKlass* thread_klass = env()->Thread_klass();
967 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
968 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
969 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
970 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
971 tls_output = thread;
972 return threadObj;
973 }
974
975
976 //------------------------------make_string_method_node------------------------
977 // Helper method for String intrinsic functions. This version is called
978 // with str1 and str2 pointing to String object nodes.
979 //
980 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
981 Node* no_ctrl = NULL;
982
983 // Get start addr of string
984 Node* str1_value = load_String_value(no_ctrl, str1);
985 Node* str1_offset = load_String_offset(no_ctrl, str1);
986 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
987
988 // Get length of string 1
989 Node* str1_len = load_String_length(no_ctrl, str1);
990
991 Node* str2_value = load_String_value(no_ctrl, str2);
992 Node* str2_offset = load_String_offset(no_ctrl, str2);
993 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
994
995 Node* str2_len = NULL;
996 Node* result = NULL;
997
998 switch (opcode) {
999 case Op_StrIndexOf:
1000 // Get length of string 2
1001 str2_len = load_String_length(no_ctrl, str2);
1002
1003 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1004 str1_start, str1_len, str2_start, str2_len);
1005 break;
1006 case Op_StrComp:
1007 // Get length of string 2
1008 str2_len = load_String_length(no_ctrl, str2);
1009
1010 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1011 str1_start, str1_len, str2_start, str2_len);
1012 break;
1013 case Op_StrEquals:
1014 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1015 str1_start, str2_start, str1_len);
1016 break;
1017 default:
1018 ShouldNotReachHere();
1019 return NULL;
1020 }
1021
1022 // All these intrinsics have checks.
1023 C->set_has_split_ifs(true); // Has chance for split-if optimization
1024
1025 return _gvn.transform(result);
1026 }
1027
1028 // Helper method for String intrinsic functions. This version is called
1029 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1030 // to Int nodes containing the lenghts of str1 and str2.
1031 //
1032 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1033 Node* result = NULL;
1034 switch (opcode) {
1035 case Op_StrIndexOf:
1036 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1037 str1_start, cnt1, str2_start, cnt2);
1038 break;
1039 case Op_StrComp:
1040 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1041 str1_start, cnt1, str2_start, cnt2);
1042 break;
1043 case Op_StrEquals:
1044 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1045 str1_start, str2_start, cnt1);
1046 break;
1047 default:
1048 ShouldNotReachHere();
1049 return NULL;
1050 }
1051
1052 // All these intrinsics have checks.
1053 C->set_has_split_ifs(true); // Has chance for split-if optimization
1054
1055 return _gvn.transform(result);
1056 }
1057
1058 //------------------------------inline_string_compareTo------------------------
1059 // public int java.lang.String.compareTo(String anotherString);
1060 bool LibraryCallKit::inline_string_compareTo() {
1061 Node* receiver = null_check(argument(0));
1062 Node* arg = null_check(argument(1));
1063 if (stopped()) {
1064 return true;
1065 }
1066 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1067 return true;
1068 }
1069
1070 //------------------------------inline_string_equals------------------------
1071 bool LibraryCallKit::inline_string_equals() {
1072 Node* receiver = null_check_receiver();
1073 // NOTE: Do not null check argument for String.equals() because spec
1074 // allows to specify NULL as argument.
1075 Node* argument = this->argument(1);
1076 if (stopped()) {
1077 return true;
1078 }
1079
1080 // paths (plus control) merge
1081 RegionNode* region = new (C) RegionNode(5);
1082 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1083
1084 // does source == target string?
1085 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1086 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1087
1088 Node* if_eq = generate_slow_guard(bol, NULL);
1089 if (if_eq != NULL) {
1090 // receiver == argument
1091 phi->init_req(2, intcon(1));
1092 region->init_req(2, if_eq);
1093 }
1094
1095 // get String klass for instanceOf
1096 ciInstanceKlass* klass = env()->String_klass();
1097
1098 if (!stopped()) {
1099 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1100 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1101 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1102
1103 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1104 //instanceOf == true, fallthrough
1105
1106 if (inst_false != NULL) {
1107 phi->init_req(3, intcon(0));
1108 region->init_req(3, inst_false);
1109 }
1110 }
1111
1112 if (!stopped()) {
1113 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1114
1115 // Properly cast the argument to String
1116 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1117 // This path is taken only when argument's type is String:NotNull.
1118 argument = cast_not_null(argument, false);
1119
1120 Node* no_ctrl = NULL;
1121
1122 // Get start addr of receiver
1123 Node* receiver_val = load_String_value(no_ctrl, receiver);
1124 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1125 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1126
1127 // Get length of receiver
1128 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1129
1130 // Get start addr of argument
1131 Node* argument_val = load_String_value(no_ctrl, argument);
1132 Node* argument_offset = load_String_offset(no_ctrl, argument);
1133 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1134
1135 // Get length of argument
1136 Node* argument_cnt = load_String_length(no_ctrl, argument);
1137
1138 // Check for receiver count != argument count
1139 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1140 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1141 Node* if_ne = generate_slow_guard(bol, NULL);
1142 if (if_ne != NULL) {
1143 phi->init_req(4, intcon(0));
1144 region->init_req(4, if_ne);
1145 }
1146
1147 // Check for count == 0 is done by assembler code for StrEquals.
1148
1149 if (!stopped()) {
1150 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1151 phi->init_req(1, equals);
1152 region->init_req(1, control());
1153 }
1154 }
1155
1156 // post merge
1157 set_control(_gvn.transform(region));
1158 record_for_igvn(region);
1159
1160 set_result(_gvn.transform(phi));
1161 return true;
1162 }
1163
1164 //------------------------------inline_array_equals----------------------------
1165 bool LibraryCallKit::inline_array_equals() {
1166 Node* arg1 = argument(0);
1167 Node* arg2 = argument(1);
1168 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1169 return true;
1170 }
1171
1172 // Java version of String.indexOf(constant string)
1173 // class StringDecl {
1174 // StringDecl(char[] ca) {
1175 // offset = 0;
1176 // count = ca.length;
1177 // value = ca;
1178 // }
1179 // int offset;
1180 // int count;
1181 // char[] value;
1182 // }
1183 //
1184 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1185 // int targetOffset, int cache_i, int md2) {
1186 // int cache = cache_i;
1187 // int sourceOffset = string_object.offset;
1188 // int sourceCount = string_object.count;
1189 // int targetCount = target_object.length;
1190 //
1191 // int targetCountLess1 = targetCount - 1;
1192 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1193 //
1194 // char[] source = string_object.value;
1195 // char[] target = target_object;
1196 // int lastChar = target[targetCountLess1];
1197 //
1198 // outer_loop:
1199 // for (int i = sourceOffset; i < sourceEnd; ) {
1200 // int src = source[i + targetCountLess1];
1201 // if (src == lastChar) {
1202 // // With random strings and a 4-character alphabet,
1203 // // reverse matching at this point sets up 0.8% fewer
1204 // // frames, but (paradoxically) makes 0.3% more probes.
1205 // // Since those probes are nearer the lastChar probe,
1206 // // there is may be a net D$ win with reverse matching.
1207 // // But, reversing loop inhibits unroll of inner loop
1208 // // for unknown reason. So, does running outer loop from
1209 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1210 // for (int j = 0; j < targetCountLess1; j++) {
1211 // if (target[targetOffset + j] != source[i+j]) {
1212 // if ((cache & (1 << source[i+j])) == 0) {
1213 // if (md2 < j+1) {
1214 // i += j+1;
1215 // continue outer_loop;
1216 // }
1217 // }
1218 // i += md2;
1219 // continue outer_loop;
1220 // }
1221 // }
1222 // return i - sourceOffset;
1223 // }
1224 // if ((cache & (1 << src)) == 0) {
1225 // i += targetCountLess1;
1226 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1227 // i++;
1228 // }
1229 // return -1;
1230 // }
1231
1232 //------------------------------string_indexOf------------------------
1233 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1234 jint cache_i, jint md2_i) {
1235
1236 Node* no_ctrl = NULL;
1237 float likely = PROB_LIKELY(0.9);
1238 float unlikely = PROB_UNLIKELY(0.9);
1239
1240 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1241
1242 Node* source = load_String_value(no_ctrl, string_object);
1243 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1244 Node* sourceCount = load_String_length(no_ctrl, string_object);
1245
1246 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1247 jint target_length = target_array->length();
1248 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1249 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1250
1251 IdealKit kit(this, false, true);
1252 #define __ kit.
1253 Node* zero = __ ConI(0);
1254 Node* one = __ ConI(1);
1255 Node* cache = __ ConI(cache_i);
1256 Node* md2 = __ ConI(md2_i);
1257 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1258 Node* targetCount = __ ConI(target_length);
1259 Node* targetCountLess1 = __ ConI(target_length - 1);
1260 Node* targetOffset = __ ConI(targetOffset_i);
1261 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1262
1263 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1264 Node* outer_loop = __ make_label(2 /* goto */);
1265 Node* return_ = __ make_label(1);
1266
1267 __ set(rtn,__ ConI(-1));
1268 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1269 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1270 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1271 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1272 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1273 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1274 Node* tpj = __ AddI(targetOffset, __ value(j));
1275 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1276 Node* ipj = __ AddI(__ value(i), __ value(j));
1277 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1278 __ if_then(targ, BoolTest::ne, src2); {
1279 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1280 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1281 __ increment(i, __ AddI(__ value(j), one));
1282 __ goto_(outer_loop);
1283 } __ end_if(); __ dead(j);
1284 }__ end_if(); __ dead(j);
1285 __ increment(i, md2);
1286 __ goto_(outer_loop);
1287 }__ end_if();
1288 __ increment(j, one);
1289 }__ end_loop(); __ dead(j);
1290 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1291 __ goto_(return_);
1292 }__ end_if();
1293 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1294 __ increment(i, targetCountLess1);
1295 }__ end_if();
1296 __ increment(i, one);
1297 __ bind(outer_loop);
1298 }__ end_loop(); __ dead(i);
1299 __ bind(return_);
1300
1301 // Final sync IdealKit and GraphKit.
1302 final_sync(kit);
1303 Node* result = __ value(rtn);
1304 #undef __
1305 C->set_has_loops(true);
1306 return result;
1307 }
1308
1309 //------------------------------inline_string_indexOf------------------------
1310 bool LibraryCallKit::inline_string_indexOf() {
1311 Node* receiver = argument(0);
1312 Node* arg = argument(1);
1313
1314 Node* result;
1315 // Disable the use of pcmpestri until it can be guaranteed that
1316 // the load doesn't cross into the uncommited space.
1317 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1318 UseSSE42Intrinsics) {
1319 // Generate SSE4.2 version of indexOf
1320 // We currently only have match rules that use SSE4.2
1321
1322 receiver = null_check(receiver);
1323 arg = null_check(arg);
1324 if (stopped()) {
1325 return true;
1326 }
1327
1328 ciInstanceKlass* str_klass = env()->String_klass();
1329 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1330
1331 // Make the merge point
1332 RegionNode* result_rgn = new (C) RegionNode(4);
1333 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1334 Node* no_ctrl = NULL;
1335
1336 // Get start addr of source string
1337 Node* source = load_String_value(no_ctrl, receiver);
1338 Node* source_offset = load_String_offset(no_ctrl, receiver);
1339 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1340
1341 // Get length of source string
1342 Node* source_cnt = load_String_length(no_ctrl, receiver);
1343
1344 // Get start addr of substring
1345 Node* substr = load_String_value(no_ctrl, arg);
1346 Node* substr_offset = load_String_offset(no_ctrl, arg);
1347 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1348
1349 // Get length of source string
1350 Node* substr_cnt = load_String_length(no_ctrl, arg);
1351
1352 // Check for substr count > string count
1353 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1354 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1355 Node* if_gt = generate_slow_guard(bol, NULL);
1356 if (if_gt != NULL) {
1357 result_phi->init_req(2, intcon(-1));
1358 result_rgn->init_req(2, if_gt);
1359 }
1360
1361 if (!stopped()) {
1362 // Check for substr count == 0
1363 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1364 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1365 Node* if_zero = generate_slow_guard(bol, NULL);
1366 if (if_zero != NULL) {
1367 result_phi->init_req(3, intcon(0));
1368 result_rgn->init_req(3, if_zero);
1369 }
1370 }
1371
1372 if (!stopped()) {
1373 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1374 result_phi->init_req(1, result);
1375 result_rgn->init_req(1, control());
1376 }
1377 set_control(_gvn.transform(result_rgn));
1378 record_for_igvn(result_rgn);
1379 result = _gvn.transform(result_phi);
1380
1381 } else { // Use LibraryCallKit::string_indexOf
1382 // don't intrinsify if argument isn't a constant string.
1383 if (!arg->is_Con()) {
1384 return false;
1385 }
1386 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1387 if (str_type == NULL) {
1388 return false;
1389 }
1390 ciInstanceKlass* klass = env()->String_klass();
1391 ciObject* str_const = str_type->const_oop();
1392 if (str_const == NULL || str_const->klass() != klass) {
1393 return false;
1394 }
1395 ciInstance* str = str_const->as_instance();
1396 assert(str != NULL, "must be instance");
1397
1398 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1399 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1400
1401 int o;
1402 int c;
1403 if (java_lang_String::has_offset_field()) {
1404 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1405 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1406 } else {
1407 o = 0;
1408 c = pat->length();
1409 }
1410
1411 // constant strings have no offset and count == length which
1412 // simplifies the resulting code somewhat so lets optimize for that.
1413 if (o != 0 || c != pat->length()) {
1414 return false;
1415 }
1416
1417 receiver = null_check(receiver, T_OBJECT);
1418 // NOTE: No null check on the argument is needed since it's a constant String oop.
1419 if (stopped()) {
1420 return true;
1421 }
1422
1423 // The null string as a pattern always returns 0 (match at beginning of string)
1424 if (c == 0) {
1425 set_result(intcon(0));
1426 return true;
1427 }
1428
1429 // Generate default indexOf
1430 jchar lastChar = pat->char_at(o + (c - 1));
1431 int cache = 0;
1432 int i;
1433 for (i = 0; i < c - 1; i++) {
1434 assert(i < pat->length(), "out of range");
1435 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1436 }
1437
1438 int md2 = c;
1439 for (i = 0; i < c - 1; i++) {
1440 assert(i < pat->length(), "out of range");
1441 if (pat->char_at(o + i) == lastChar) {
1442 md2 = (c - 1) - i;
1443 }
1444 }
1445
1446 result = string_indexOf(receiver, pat, o, cache, md2);
1447 }
1448 set_result(result);
1449 return true;
1450 }
1451
1452 //--------------------------round_double_node--------------------------------
1453 // Round a double node if necessary.
1454 Node* LibraryCallKit::round_double_node(Node* n) {
1455 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1456 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1457 return n;
1458 }
1459
1460 //------------------------------inline_math-----------------------------------
1461 // public static double Math.abs(double)
1462 // public static double Math.sqrt(double)
1463 // public static double Math.log(double)
1464 // public static double Math.log10(double)
1465 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1466 Node* arg = round_double_node(argument(0));
1467 Node* n;
1468 switch (id) {
1469 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1470 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(C, control(), arg); break;
1471 case vmIntrinsics::_dlog: n = new (C) LogDNode(C, control(), arg); break;
1472 case vmIntrinsics::_dlog10: n = new (C) Log10DNode(C, control(), arg); break;
1473 default: fatal_unexpected_iid(id); break;
1474 }
1475 set_result(_gvn.transform(n));
1476 return true;
1477 }
1478
1479 //------------------------------inline_trig----------------------------------
1480 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1481 // argument reduction which will turn into a fast/slow diamond.
1482 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1483 Node* arg = round_double_node(argument(0));
1484 Node* n = NULL;
1485
1486 switch (id) {
1487 case vmIntrinsics::_dsin: n = new (C) SinDNode(C, control(), arg); break;
1488 case vmIntrinsics::_dcos: n = new (C) CosDNode(C, control(), arg); break;
1489 case vmIntrinsics::_dtan: n = new (C) TanDNode(C, control(), arg); break;
1490 default: fatal_unexpected_iid(id); break;
1491 }
1492 n = _gvn.transform(n);
1493
1494 // Rounding required? Check for argument reduction!
1495 if (Matcher::strict_fp_requires_explicit_rounding) {
1496 static const double pi_4 = 0.7853981633974483;
1497 static const double neg_pi_4 = -0.7853981633974483;
1498 // pi/2 in 80-bit extended precision
1499 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1500 // -pi/2 in 80-bit extended precision
1501 // 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};
1502 // Cutoff value for using this argument reduction technique
1503 //static const double pi_2_minus_epsilon = 1.564660403643354;
1504 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1505
1506 // Pseudocode for sin:
1507 // if (x <= Math.PI / 4.0) {
1508 // if (x >= -Math.PI / 4.0) return fsin(x);
1509 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1510 // } else {
1511 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1512 // }
1513 // return StrictMath.sin(x);
1514
1515 // Pseudocode for cos:
1516 // if (x <= Math.PI / 4.0) {
1517 // if (x >= -Math.PI / 4.0) return fcos(x);
1518 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1519 // } else {
1520 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1521 // }
1522 // return StrictMath.cos(x);
1523
1524 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1525 // requires a special machine instruction to load it. Instead we'll try
1526 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1527 // probably do the math inside the SIN encoding.
1528
1529 // Make the merge point
1530 RegionNode* r = new (C) RegionNode(3);
1531 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1532
1533 // Flatten arg so we need only 1 test
1534 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1535 // Node for PI/4 constant
1536 Node *pi4 = makecon(TypeD::make(pi_4));
1537 // Check PI/4 : abs(arg)
1538 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1539 // Check: If PI/4 < abs(arg) then go slow
1540 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1541 // Branch either way
1542 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1543 set_control(opt_iff(r,iff));
1544
1545 // Set fast path result
1546 phi->init_req(2, n);
1547
1548 // Slow path - non-blocking leaf call
1549 Node* call = NULL;
1550 switch (id) {
1551 case vmIntrinsics::_dsin:
1552 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1553 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1554 "Sin", NULL, arg, top());
1555 break;
1556 case vmIntrinsics::_dcos:
1557 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1558 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1559 "Cos", NULL, arg, top());
1560 break;
1561 case vmIntrinsics::_dtan:
1562 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1563 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1564 "Tan", NULL, arg, top());
1565 break;
1566 }
1567 assert(control()->in(0) == call, "");
1568 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1569 r->init_req(1, control());
1570 phi->init_req(1, slow_result);
1571
1572 // Post-merge
1573 set_control(_gvn.transform(r));
1574 record_for_igvn(r);
1575 n = _gvn.transform(phi);
1576
1577 C->set_has_split_ifs(true); // Has chance for split-if optimization
1578 }
1579 set_result(n);
1580 return true;
1581 }
1582
1583 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1584 //-------------------
1585 //result=(result.isNaN())? funcAddr():result;
1586 // Check: If isNaN() by checking result!=result? then either trap
1587 // or go to runtime
1588 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1589 // Build the boolean node
1590 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1591
1592 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1593 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1594 // The pow or exp intrinsic returned a NaN, which requires a call
1595 // to the runtime. Recompile with the runtime call.
1596 uncommon_trap(Deoptimization::Reason_intrinsic,
1597 Deoptimization::Action_make_not_entrant);
1598 }
1599 set_result(result);
1600 } else {
1601 // If this inlining ever returned NaN in the past, we compile a call
1602 // to the runtime to properly handle corner cases
1603
1604 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1605 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1606 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1607
1608 if (!if_slow->is_top()) {
1609 RegionNode* result_region = new (C) RegionNode(3);
1610 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1611
1612 result_region->init_req(1, if_fast);
1613 result_val->init_req(1, result);
1614
1615 set_control(if_slow);
1616
1617 const TypePtr* no_memory_effects = NULL;
1618 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1619 no_memory_effects,
1620 x, top(), y, y ? top() : NULL);
1621 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1622 #ifdef ASSERT
1623 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1624 assert(value_top == top(), "second value must be top");
1625 #endif
1626
1627 result_region->init_req(2, control());
1628 result_val->init_req(2, value);
1629 set_result(result_region, result_val);
1630 } else {
1631 set_result(result);
1632 }
1633 }
1634 }
1635
1636 //------------------------------inline_exp-------------------------------------
1637 // Inline exp instructions, if possible. The Intel hardware only misses
1638 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1639 bool LibraryCallKit::inline_exp() {
1640 Node* arg = round_double_node(argument(0));
1641 Node* n = _gvn.transform(new (C) ExpDNode(C, control(), arg));
1642
1643 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1644
1645 C->set_has_split_ifs(true); // Has chance for split-if optimization
1646 return true;
1647 }
1648
1649 //------------------------------inline_pow-------------------------------------
1650 // Inline power instructions, if possible.
1651 bool LibraryCallKit::inline_pow() {
1652 // Pseudocode for pow
1653 // if (x <= 0.0) {
1654 // long longy = (long)y;
1655 // if ((double)longy == y) { // if y is long
1656 // if (y + 1 == y) longy = 0; // huge number: even
1657 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1658 // } else {
1659 // result = NaN;
1660 // }
1661 // } else {
1662 // result = DPow(x,y);
1663 // }
1664 // if (result != result)? {
1665 // result = uncommon_trap() or runtime_call();
1666 // }
1667 // return result;
1668
1669 Node* x = round_double_node(argument(0));
1670 Node* y = round_double_node(argument(2));
1671
1672 Node* result = NULL;
1673
1674 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1675 // Short form: skip the fancy tests and just check for NaN result.
1676 result = _gvn.transform(new (C) PowDNode(C, control(), x, y));
1677 } else {
1678 // If this inlining ever returned NaN in the past, include all
1679 // checks + call to the runtime.
1680
1681 // Set the merge point for If node with condition of (x <= 0.0)
1682 // There are four possible paths to region node and phi node
1683 RegionNode *r = new (C) RegionNode(4);
1684 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1685
1686 // Build the first if node: if (x <= 0.0)
1687 // Node for 0 constant
1688 Node *zeronode = makecon(TypeD::ZERO);
1689 // Check x:0
1690 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1691 // Check: If (x<=0) then go complex path
1692 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1693 // Branch either way
1694 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1695 // Fast path taken; set region slot 3
1696 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1697 r->init_req(3,fast_taken); // Capture fast-control
1698
1699 // Fast path not-taken, i.e. slow path
1700 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1701
1702 // Set fast path result
1703 Node *fast_result = _gvn.transform( new (C) PowDNode(C, control(), x, y) );
1704 phi->init_req(3, fast_result);
1705
1706 // Complex path
1707 // Build the second if node (if y is long)
1708 // Node for (long)y
1709 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1710 // Node for (double)((long) y)
1711 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1712 // Check (double)((long) y) : y
1713 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1714 // Check if (y isn't long) then go to slow path
1715
1716 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1717 // Branch either way
1718 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1719 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1720
1721 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1722
1723 // Calculate DPow(abs(x), y)*(1 & (long)y)
1724 // Node for constant 1
1725 Node *conone = longcon(1);
1726 // 1& (long)y
1727 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1728
1729 // A huge number is always even. Detect a huge number by checking
1730 // if y + 1 == y and set integer to be tested for parity to 0.
1731 // Required for corner case:
1732 // (long)9.223372036854776E18 = max_jlong
1733 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1734 // max_jlong is odd but 9.223372036854776E18 is even
1735 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1736 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1737 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1738 Node* correctedsign = NULL;
1739 if (ConditionalMoveLimit != 0) {
1740 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1741 } else {
1742 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1743 RegionNode *r = new (C) RegionNode(3);
1744 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1745 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1746 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1747 phi->init_req(1, signnode);
1748 phi->init_req(2, longcon(0));
1749 correctedsign = _gvn.transform(phi);
1750 ylong_path = _gvn.transform(r);
1751 record_for_igvn(r);
1752 }
1753
1754 // zero node
1755 Node *conzero = longcon(0);
1756 // Check (1&(long)y)==0?
1757 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1758 // Check if (1&(long)y)!=0?, if so the result is negative
1759 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1760 // abs(x)
1761 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1762 // abs(x)^y
1763 Node *absxpowy = _gvn.transform( new (C) PowDNode(C, control(), absx, y) );
1764 // -abs(x)^y
1765 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1766 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1767 Node *signresult = NULL;
1768 if (ConditionalMoveLimit != 0) {
1769 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1770 } else {
1771 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1772 RegionNode *r = new (C) RegionNode(3);
1773 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1774 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1775 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1776 phi->init_req(1, absxpowy);
1777 phi->init_req(2, negabsxpowy);
1778 signresult = _gvn.transform(phi);
1779 ylong_path = _gvn.transform(r);
1780 record_for_igvn(r);
1781 }
1782 // Set complex path fast result
1783 r->init_req(2, ylong_path);
1784 phi->init_req(2, signresult);
1785
1786 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1787 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1788 r->init_req(1,slow_path);
1789 phi->init_req(1,slow_result);
1790
1791 // Post merge
1792 set_control(_gvn.transform(r));
1793 record_for_igvn(r);
1794 result = _gvn.transform(phi);
1795 }
1796
1797 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1798
1799 C->set_has_split_ifs(true); // Has chance for split-if optimization
1800 return true;
1801 }
1802
1803 //------------------------------runtime_math-----------------------------
1804 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1805 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1806 "must be (DD)D or (D)D type");
1807
1808 // Inputs
1809 Node* a = round_double_node(argument(0));
1810 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1811
1812 const TypePtr* no_memory_effects = NULL;
1813 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1814 no_memory_effects,
1815 a, top(), b, b ? top() : NULL);
1816 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1817 #ifdef ASSERT
1818 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1819 assert(value_top == top(), "second value must be top");
1820 #endif
1821
1822 set_result(value);
1823 return true;
1824 }
1825
1826 //------------------------------inline_math_native-----------------------------
1827 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1828 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1829 switch (id) {
1830 // These intrinsics are not properly supported on all hardware
1831 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1832 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1833 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1834 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1835 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1836 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1837
1838 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1839 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1840 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1841 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1842
1843 // These intrinsics are supported on all hardware
1844 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1845 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1846
1847 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1848 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1849 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1850 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1851 #undef FN_PTR
1852
1853 // These intrinsics are not yet correctly implemented
1854 case vmIntrinsics::_datan2:
1855 return false;
1856
1857 default:
1858 fatal_unexpected_iid(id);
1859 return false;
1860 }
1861 }
1862
1863 static bool is_simple_name(Node* n) {
1864 return (n->req() == 1 // constant
1865 || (n->is_Type() && n->as_Type()->type()->singleton())
1866 || n->is_Proj() // parameter or return value
1867 || n->is_Phi() // local of some sort
1868 );
1869 }
1870
1871 //----------------------------inline_min_max-----------------------------------
1872 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1873 set_result(generate_min_max(id, argument(0), argument(1)));
1874 return true;
1875 }
1876
1877 Node*
1878 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1879 // These are the candidate return value:
1880 Node* xvalue = x0;
1881 Node* yvalue = y0;
1882
1883 if (xvalue == yvalue) {
1884 return xvalue;
1885 }
1886
1887 bool want_max = (id == vmIntrinsics::_max);
1888
1889 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1890 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1891 if (txvalue == NULL || tyvalue == NULL) return top();
1892 // This is not really necessary, but it is consistent with a
1893 // hypothetical MaxINode::Value method:
1894 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1895
1896 // %%% This folding logic should (ideally) be in a different place.
1897 // Some should be inside IfNode, and there to be a more reliable
1898 // transformation of ?: style patterns into cmoves. We also want
1899 // more powerful optimizations around cmove and min/max.
1900
1901 // Try to find a dominating comparison of these guys.
1902 // It can simplify the index computation for Arrays.copyOf
1903 // and similar uses of System.arraycopy.
1904 // First, compute the normalized version of CmpI(x, y).
1905 int cmp_op = Op_CmpI;
1906 Node* xkey = xvalue;
1907 Node* ykey = yvalue;
1908 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1909 if (ideal_cmpxy->is_Cmp()) {
1910 // E.g., if we have CmpI(length - offset, count),
1911 // it might idealize to CmpI(length, count + offset)
1912 cmp_op = ideal_cmpxy->Opcode();
1913 xkey = ideal_cmpxy->in(1);
1914 ykey = ideal_cmpxy->in(2);
1915 }
1916
1917 // Start by locating any relevant comparisons.
1918 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1919 Node* cmpxy = NULL;
1920 Node* cmpyx = NULL;
1921 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1922 Node* cmp = start_from->fast_out(k);
1923 if (cmp->outcnt() > 0 && // must have prior uses
1924 cmp->in(0) == NULL && // must be context-independent
1925 cmp->Opcode() == cmp_op) { // right kind of compare
1926 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1927 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1928 }
1929 }
1930
1931 const int NCMPS = 2;
1932 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1933 int cmpn;
1934 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1935 if (cmps[cmpn] != NULL) break; // find a result
1936 }
1937 if (cmpn < NCMPS) {
1938 // Look for a dominating test that tells us the min and max.
1939 int depth = 0; // Limit search depth for speed
1940 Node* dom = control();
1941 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1942 if (++depth >= 100) break;
1943 Node* ifproj = dom;
1944 if (!ifproj->is_Proj()) continue;
1945 Node* iff = ifproj->in(0);
1946 if (!iff->is_If()) continue;
1947 Node* bol = iff->in(1);
1948 if (!bol->is_Bool()) continue;
1949 Node* cmp = bol->in(1);
1950 if (cmp == NULL) continue;
1951 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1952 if (cmps[cmpn] == cmp) break;
1953 if (cmpn == NCMPS) continue;
1954 BoolTest::mask btest = bol->as_Bool()->_test._test;
1955 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1956 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1957 // At this point, we know that 'x btest y' is true.
1958 switch (btest) {
1959 case BoolTest::eq:
1960 // They are proven equal, so we can collapse the min/max.
1961 // Either value is the answer. Choose the simpler.
1962 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1963 return yvalue;
1964 return xvalue;
1965 case BoolTest::lt: // x < y
1966 case BoolTest::le: // x <= y
1967 return (want_max ? yvalue : xvalue);
1968 case BoolTest::gt: // x > y
1969 case BoolTest::ge: // x >= y
1970 return (want_max ? xvalue : yvalue);
1971 }
1972 }
1973 }
1974
1975 // We failed to find a dominating test.
1976 // Let's pick a test that might GVN with prior tests.
1977 Node* best_bol = NULL;
1978 BoolTest::mask best_btest = BoolTest::illegal;
1979 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1980 Node* cmp = cmps[cmpn];
1981 if (cmp == NULL) continue;
1982 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1983 Node* bol = cmp->fast_out(j);
1984 if (!bol->is_Bool()) continue;
1985 BoolTest::mask btest = bol->as_Bool()->_test._test;
1986 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1987 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1988 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1989 best_bol = bol->as_Bool();
1990 best_btest = btest;
1991 }
1992 }
1993 }
1994
1995 Node* answer_if_true = NULL;
1996 Node* answer_if_false = NULL;
1997 switch (best_btest) {
1998 default:
1999 if (cmpxy == NULL)
2000 cmpxy = ideal_cmpxy;
2001 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2002 // and fall through:
2003 case BoolTest::lt: // x < y
2004 case BoolTest::le: // x <= y
2005 answer_if_true = (want_max ? yvalue : xvalue);
2006 answer_if_false = (want_max ? xvalue : yvalue);
2007 break;
2008 case BoolTest::gt: // x > y
2009 case BoolTest::ge: // x >= y
2010 answer_if_true = (want_max ? xvalue : yvalue);
2011 answer_if_false = (want_max ? yvalue : xvalue);
2012 break;
2013 }
2014
2015 jint hi, lo;
2016 if (want_max) {
2017 // We can sharpen the minimum.
2018 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2019 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2020 } else {
2021 // We can sharpen the maximum.
2022 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2023 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2024 }
2025
2026 // Use a flow-free graph structure, to avoid creating excess control edges
2027 // which could hinder other optimizations.
2028 // Since Math.min/max is often used with arraycopy, we want
2029 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2030 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2031 answer_if_false, answer_if_true,
2032 TypeInt::make(lo, hi, widen));
2033
2034 return _gvn.transform(cmov);
2035
2036 /*
2037 // This is not as desirable as it may seem, since Min and Max
2038 // nodes do not have a full set of optimizations.
2039 // And they would interfere, anyway, with 'if' optimizations
2040 // and with CMoveI canonical forms.
2041 switch (id) {
2042 case vmIntrinsics::_min:
2043 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2044 case vmIntrinsics::_max:
2045 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2046 default:
2047 ShouldNotReachHere();
2048 }
2049 */
2050 }
2051
2052 inline int
2053 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2054 const TypePtr* base_type = TypePtr::NULL_PTR;
2055 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2056 if (base_type == NULL) {
2057 // Unknown type.
2058 return Type::AnyPtr;
2059 } else if (base_type == TypePtr::NULL_PTR) {
2060 // Since this is a NULL+long form, we have to switch to a rawptr.
2061 base = _gvn.transform( new (C) CastX2PNode(offset) );
2062 offset = MakeConX(0);
2063 return Type::RawPtr;
2064 } else if (base_type->base() == Type::RawPtr) {
2065 return Type::RawPtr;
2066 } else if (base_type->isa_oopptr()) {
2067 // Base is never null => always a heap address.
2068 if (base_type->ptr() == TypePtr::NotNull) {
2069 return Type::OopPtr;
2070 }
2071 // Offset is small => always a heap address.
2072 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2073 if (offset_type != NULL &&
2074 base_type->offset() == 0 && // (should always be?)
2075 offset_type->_lo >= 0 &&
2076 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2077 return Type::OopPtr;
2078 }
2079 // Otherwise, it might either be oop+off or NULL+addr.
2080 return Type::AnyPtr;
2081 } else {
2082 // No information:
2083 return Type::AnyPtr;
2084 }
2085 }
2086
2087 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2088 int kind = classify_unsafe_addr(base, offset);
2089 if (kind == Type::RawPtr) {
2090 return basic_plus_adr(top(), base, offset);
2091 } else {
2092 return basic_plus_adr(base, offset);
2093 }
2094 }
2095
2096 //--------------------------inline_number_methods-----------------------------
2097 // inline int Integer.numberOfLeadingZeros(int)
2098 // inline int Long.numberOfLeadingZeros(long)
2099 //
2100 // inline int Integer.numberOfTrailingZeros(int)
2101 // inline int Long.numberOfTrailingZeros(long)
2102 //
2103 // inline int Integer.bitCount(int)
2104 // inline int Long.bitCount(long)
2105 //
2106 // inline char Character.reverseBytes(char)
2107 // inline short Short.reverseBytes(short)
2108 // inline int Integer.reverseBytes(int)
2109 // inline long Long.reverseBytes(long)
2110 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2111 Node* arg = argument(0);
2112 Node* n;
2113 switch (id) {
2114 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2115 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2116 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2117 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2118 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2119 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2120 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2121 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2122 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2123 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2124 default: fatal_unexpected_iid(id); break;
2125 }
2126 set_result(_gvn.transform(n));
2127 return true;
2128 }
2129
2130 //----------------------------inline_unsafe_access----------------------------
2131
2132 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2133
2134 // Helper that guards and inserts a pre-barrier.
2135 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2136 Node* pre_val, bool need_mem_bar) {
2137 // We could be accessing the referent field of a reference object. If so, when G1
2138 // is enabled, we need to log the value in the referent field in an SATB buffer.
2139 // This routine performs some compile time filters and generates suitable
2140 // runtime filters that guard the pre-barrier code.
2141 // Also add memory barrier for non volatile load from the referent field
2142 // to prevent commoning of loads across safepoint.
2143 if (!UseG1GC && !need_mem_bar)
2144 return;
2145
2146 // Some compile time checks.
2147
2148 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2149 const TypeX* otype = offset->find_intptr_t_type();
2150 if (otype != NULL && otype->is_con() &&
2151 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2152 // Constant offset but not the reference_offset so just return
2153 return;
2154 }
2155
2156 // We only need to generate the runtime guards for instances.
2157 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2158 if (btype != NULL) {
2159 if (btype->isa_aryptr()) {
2160 // Array type so nothing to do
2161 return;
2162 }
2163
2164 const TypeInstPtr* itype = btype->isa_instptr();
2165 if (itype != NULL) {
2166 // Can the klass of base_oop be statically determined to be
2167 // _not_ a sub-class of Reference and _not_ Object?
2168 ciKlass* klass = itype->klass();
2169 if ( klass->is_loaded() &&
2170 !klass->is_subtype_of(env()->Reference_klass()) &&
2171 !env()->Object_klass()->is_subtype_of(klass)) {
2172 return;
2173 }
2174 }
2175 }
2176
2177 // The compile time filters did not reject base_oop/offset so
2178 // we need to generate the following runtime filters
2179 //
2180 // if (offset == java_lang_ref_Reference::_reference_offset) {
2181 // if (instance_of(base, java.lang.ref.Reference)) {
2182 // pre_barrier(_, pre_val, ...);
2183 // }
2184 // }
2185
2186 float likely = PROB_LIKELY( 0.999);
2187 float unlikely = PROB_UNLIKELY(0.999);
2188
2189 IdealKit ideal(this);
2190 #define __ ideal.
2191
2192 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2193
2194 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2195 // Update graphKit memory and control from IdealKit.
2196 sync_kit(ideal);
2197
2198 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2199 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2200
2201 // Update IdealKit memory and control from graphKit.
2202 __ sync_kit(this);
2203
2204 Node* one = __ ConI(1);
2205 // is_instof == 0 if base_oop == NULL
2206 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2207
2208 // Update graphKit from IdeakKit.
2209 sync_kit(ideal);
2210
2211 // Use the pre-barrier to record the value in the referent field
2212 pre_barrier(false /* do_load */,
2213 __ ctrl(),
2214 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2215 pre_val /* pre_val */,
2216 T_OBJECT);
2217 if (need_mem_bar) {
2218 // Add memory barrier to prevent commoning reads from this field
2219 // across safepoint since GC can change its value.
2220 insert_mem_bar(Op_MemBarCPUOrder);
2221 }
2222 // Update IdealKit from graphKit.
2223 __ sync_kit(this);
2224
2225 } __ end_if(); // _ref_type != ref_none
2226 } __ end_if(); // offset == referent_offset
2227
2228 // Final sync IdealKit and GraphKit.
2229 final_sync(ideal);
2230 #undef __
2231 }
2232
2233
2234 // Interpret Unsafe.fieldOffset cookies correctly:
2235 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2236
2237 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2238 // Attempt to infer a sharper value type from the offset and base type.
2239 ciKlass* sharpened_klass = NULL;
2240
2241 // See if it is an instance field, with an object type.
2242 if (alias_type->field() != NULL) {
2243 assert(!is_native_ptr, "native pointer op cannot use a java address");
2244 if (alias_type->field()->type()->is_klass()) {
2245 sharpened_klass = alias_type->field()->type()->as_klass();
2246 }
2247 }
2248
2249 // See if it is a narrow oop array.
2250 if (adr_type->isa_aryptr()) {
2251 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2252 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2253 if (elem_type != NULL) {
2254 sharpened_klass = elem_type->klass();
2255 }
2256 }
2257 }
2258
2259 // The sharpened class might be unloaded if there is no class loader
2260 // contraint in place.
2261 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2262 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2263
2264 #ifndef PRODUCT
2265 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2266 tty->print(" from base type: "); adr_type->dump();
2267 tty->print(" sharpened value: "); tjp->dump();
2268 }
2269 #endif
2270 // Sharpen the value type.
2271 return tjp;
2272 }
2273 return NULL;
2274 }
2275
2276 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2277 if (callee()->is_static()) return false; // caller must have the capability!
2278
2279 #ifndef PRODUCT
2280 {
2281 ResourceMark rm;
2282 // Check the signatures.
2283 ciSignature* sig = callee()->signature();
2284 #ifdef ASSERT
2285 if (!is_store) {
2286 // Object getObject(Object base, int/long offset), etc.
2287 BasicType rtype = sig->return_type()->basic_type();
2288 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2289 rtype = T_ADDRESS; // it is really a C void*
2290 assert(rtype == type, "getter must return the expected value");
2291 if (!is_native_ptr) {
2292 assert(sig->count() == 2, "oop getter has 2 arguments");
2293 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2294 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2295 } else {
2296 assert(sig->count() == 1, "native getter has 1 argument");
2297 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2298 }
2299 } else {
2300 // void putObject(Object base, int/long offset, Object x), etc.
2301 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2302 if (!is_native_ptr) {
2303 assert(sig->count() == 3, "oop putter has 3 arguments");
2304 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2305 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2306 } else {
2307 assert(sig->count() == 2, "native putter has 2 arguments");
2308 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2309 }
2310 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2311 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2312 vtype = T_ADDRESS; // it is really a C void*
2313 assert(vtype == type, "putter must accept the expected value");
2314 }
2315 #endif // ASSERT
2316 }
2317 #endif //PRODUCT
2318
2319 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2320
2321 Node* receiver = argument(0); // type: oop
2322
2323 // Build address expression. See the code in inline_unsafe_prefetch.
2324 Node* adr;
2325 Node* heap_base_oop = top();
2326 Node* offset = top();
2327 Node* val;
2328
2329 if (!is_native_ptr) {
2330 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2331 Node* base = argument(1); // type: oop
2332 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2333 offset = argument(2); // type: long
2334 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2335 // to be plain byte offsets, which are also the same as those accepted
2336 // by oopDesc::field_base.
2337 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2338 "fieldOffset must be byte-scaled");
2339 // 32-bit machines ignore the high half!
2340 offset = ConvL2X(offset);
2341 adr = make_unsafe_address(base, offset);
2342 heap_base_oop = base;
2343 val = is_store ? argument(4) : NULL;
2344 } else {
2345 Node* ptr = argument(1); // type: long
2346 ptr = ConvL2X(ptr); // adjust Java long to machine word
2347 adr = make_unsafe_address(NULL, ptr);
2348 val = is_store ? argument(3) : NULL;
2349 }
2350
2351 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2352
2353 // First guess at the value type.
2354 const Type *value_type = Type::get_const_basic_type(type);
2355
2356 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2357 // there was not enough information to nail it down.
2358 Compile::AliasType* alias_type = C->alias_type(adr_type);
2359 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2360
2361 // We will need memory barriers unless we can determine a unique
2362 // alias category for this reference. (Note: If for some reason
2363 // the barriers get omitted and the unsafe reference begins to "pollute"
2364 // the alias analysis of the rest of the graph, either Compile::can_alias
2365 // or Compile::must_alias will throw a diagnostic assert.)
2366 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2367
2368 // If we are reading the value of the referent field of a Reference
2369 // object (either by using Unsafe directly or through reflection)
2370 // then, if G1 is enabled, we need to record the referent in an
2371 // SATB log buffer using the pre-barrier mechanism.
2372 // Also we need to add memory barrier to prevent commoning reads
2373 // from this field across safepoint since GC can change its value.
2374 bool need_read_barrier = !is_native_ptr && !is_store &&
2375 offset != top() && heap_base_oop != top();
2376
2377 if (!is_store && type == T_OBJECT) {
2378 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2379 if (tjp != NULL) {
2380 value_type = tjp;
2381 }
2382 }
2383
2384 receiver = null_check(receiver);
2385 if (stopped()) {
2386 return true;
2387 }
2388 // Heap pointers get a null-check from the interpreter,
2389 // as a courtesy. However, this is not guaranteed by Unsafe,
2390 // and it is not possible to fully distinguish unintended nulls
2391 // from intended ones in this API.
2392
2393 if (is_volatile) {
2394 // We need to emit leading and trailing CPU membars (see below) in
2395 // addition to memory membars when is_volatile. This is a little
2396 // too strong, but avoids the need to insert per-alias-type
2397 // volatile membars (for stores; compare Parse::do_put_xxx), which
2398 // we cannot do effectively here because we probably only have a
2399 // rough approximation of type.
2400 need_mem_bar = true;
2401 // For Stores, place a memory ordering barrier now.
2402 if (is_store)
2403 insert_mem_bar(Op_MemBarRelease);
2404 }
2405
2406 // Memory barrier to prevent normal and 'unsafe' accesses from
2407 // bypassing each other. Happens after null checks, so the
2408 // exception paths do not take memory state from the memory barrier,
2409 // so there's no problems making a strong assert about mixing users
2410 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2411 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2412 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2413
2414 if (!is_store) {
2415 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2416 // load value
2417 switch (type) {
2418 case T_BOOLEAN:
2419 case T_CHAR:
2420 case T_BYTE:
2421 case T_SHORT:
2422 case T_INT:
2423 case T_LONG:
2424 case T_FLOAT:
2425 case T_DOUBLE:
2426 break;
2427 case T_OBJECT:
2428 if (need_read_barrier) {
2429 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2430 }
2431 break;
2432 case T_ADDRESS:
2433 // Cast to an int type.
2434 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2435 p = ConvX2L(p);
2436 break;
2437 default:
2438 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2439 break;
2440 }
2441 // The load node has the control of the preceding MemBarCPUOrder. All
2442 // following nodes will have the control of the MemBarCPUOrder inserted at
2443 // the end of this method. So, pushing the load onto the stack at a later
2444 // point is fine.
2445 set_result(p);
2446 } else {
2447 // place effect of store into memory
2448 switch (type) {
2449 case T_DOUBLE:
2450 val = dstore_rounding(val);
2451 break;
2452 case T_ADDRESS:
2453 // Repackage the long as a pointer.
2454 val = ConvL2X(val);
2455 val = _gvn.transform( new (C) CastX2PNode(val) );
2456 break;
2457 }
2458
2459 if (type != T_OBJECT ) {
2460 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2461 } else {
2462 // Possibly an oop being stored to Java heap or native memory
2463 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2464 // oop to Java heap.
2465 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2466 } else {
2467 // We can't tell at compile time if we are storing in the Java heap or outside
2468 // of it. So we need to emit code to conditionally do the proper type of
2469 // store.
2470
2471 IdealKit ideal(this);
2472 #define __ ideal.
2473 // QQQ who knows what probability is here??
2474 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2475 // Sync IdealKit and graphKit.
2476 sync_kit(ideal);
2477 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2478 // Update IdealKit memory.
2479 __ sync_kit(this);
2480 } __ else_(); {
2481 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2482 } __ end_if();
2483 // Final sync IdealKit and GraphKit.
2484 final_sync(ideal);
2485 #undef __
2486 }
2487 }
2488 }
2489
2490 if (is_volatile) {
2491 if (!is_store)
2492 insert_mem_bar(Op_MemBarAcquire);
2493 else
2494 insert_mem_bar(Op_MemBarVolatile);
2495 }
2496
2497 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2498
2499 return true;
2500 }
2501
2502 //----------------------------inline_unsafe_prefetch----------------------------
2503
2504 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2505 #ifndef PRODUCT
2506 {
2507 ResourceMark rm;
2508 // Check the signatures.
2509 ciSignature* sig = callee()->signature();
2510 #ifdef ASSERT
2511 // Object getObject(Object base, int/long offset), etc.
2512 BasicType rtype = sig->return_type()->basic_type();
2513 if (!is_native_ptr) {
2514 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2515 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2516 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2517 } else {
2518 assert(sig->count() == 1, "native prefetch has 1 argument");
2519 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2520 }
2521 #endif // ASSERT
2522 }
2523 #endif // !PRODUCT
2524
2525 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2526
2527 const int idx = is_static ? 0 : 1;
2528 if (!is_static) {
2529 null_check_receiver();
2530 if (stopped()) {
2531 return true;
2532 }
2533 }
2534
2535 // Build address expression. See the code in inline_unsafe_access.
2536 Node *adr;
2537 if (!is_native_ptr) {
2538 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2539 Node* base = argument(idx + 0); // type: oop
2540 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2541 Node* offset = argument(idx + 1); // type: long
2542 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2543 // to be plain byte offsets, which are also the same as those accepted
2544 // by oopDesc::field_base.
2545 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2546 "fieldOffset must be byte-scaled");
2547 // 32-bit machines ignore the high half!
2548 offset = ConvL2X(offset);
2549 adr = make_unsafe_address(base, offset);
2550 } else {
2551 Node* ptr = argument(idx + 0); // type: long
2552 ptr = ConvL2X(ptr); // adjust Java long to machine word
2553 adr = make_unsafe_address(NULL, ptr);
2554 }
2555
2556 // Generate the read or write prefetch
2557 Node *prefetch;
2558 if (is_store) {
2559 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2560 } else {
2561 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2562 }
2563 prefetch->init_req(0, control());
2564 set_i_o(_gvn.transform(prefetch));
2565
2566 return true;
2567 }
2568
2569 //----------------------------inline_unsafe_load_store----------------------------
2570 // This method serves a couple of different customers (depending on LoadStoreKind):
2571 //
2572 // LS_cmpxchg:
2573 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2574 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2575 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2576 //
2577 // LS_xadd:
2578 // public int getAndAddInt( Object o, long offset, int delta)
2579 // public long getAndAddLong(Object o, long offset, long delta)
2580 //
2581 // LS_xchg:
2582 // int getAndSet(Object o, long offset, int newValue)
2583 // long getAndSet(Object o, long offset, long newValue)
2584 // Object getAndSet(Object o, long offset, Object newValue)
2585 //
2586 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2587 // This basic scheme here is the same as inline_unsafe_access, but
2588 // differs in enough details that combining them would make the code
2589 // overly confusing. (This is a true fact! I originally combined
2590 // them, but even I was confused by it!) As much code/comments as
2591 // possible are retained from inline_unsafe_access though to make
2592 // the correspondences clearer. - dl
2593
2594 if (callee()->is_static()) return false; // caller must have the capability!
2595
2596 #ifndef PRODUCT
2597 BasicType rtype;
2598 {
2599 ResourceMark rm;
2600 // Check the signatures.
2601 ciSignature* sig = callee()->signature();
2602 rtype = sig->return_type()->basic_type();
2603 if (kind == LS_xadd || kind == LS_xchg) {
2604 // Check the signatures.
2605 #ifdef ASSERT
2606 assert(rtype == type, "get and set must return the expected type");
2607 assert(sig->count() == 3, "get and set has 3 arguments");
2608 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2609 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2610 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2611 #endif // ASSERT
2612 } else if (kind == LS_cmpxchg) {
2613 // Check the signatures.
2614 #ifdef ASSERT
2615 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2616 assert(sig->count() == 4, "CAS has 4 arguments");
2617 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2618 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2619 #endif // ASSERT
2620 } else {
2621 ShouldNotReachHere();
2622 }
2623 }
2624 #endif //PRODUCT
2625
2626 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2627
2628 // Get arguments:
2629 Node* receiver = NULL;
2630 Node* base = NULL;
2631 Node* offset = NULL;
2632 Node* oldval = NULL;
2633 Node* newval = NULL;
2634 if (kind == LS_cmpxchg) {
2635 const bool two_slot_type = type2size[type] == 2;
2636 receiver = argument(0); // type: oop
2637 base = argument(1); // type: oop
2638 offset = argument(2); // type: long
2639 oldval = argument(4); // type: oop, int, or long
2640 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2641 } else if (kind == LS_xadd || kind == LS_xchg){
2642 receiver = argument(0); // type: oop
2643 base = argument(1); // type: oop
2644 offset = argument(2); // type: long
2645 oldval = NULL;
2646 newval = argument(4); // type: oop, int, or long
2647 }
2648
2649 // Null check receiver.
2650 receiver = null_check(receiver);
2651 if (stopped()) {
2652 return true;
2653 }
2654
2655 // Build field offset expression.
2656 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2657 // to be plain byte offsets, which are also the same as those accepted
2658 // by oopDesc::field_base.
2659 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2660 // 32-bit machines ignore the high half of long offsets
2661 offset = ConvL2X(offset);
2662 Node* adr = make_unsafe_address(base, offset);
2663 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2664
2665 // For CAS, unlike inline_unsafe_access, there seems no point in
2666 // trying to refine types. Just use the coarse types here.
2667 const Type *value_type = Type::get_const_basic_type(type);
2668 Compile::AliasType* alias_type = C->alias_type(adr_type);
2669 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2670
2671 if (kind == LS_xchg && type == T_OBJECT) {
2672 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2673 if (tjp != NULL) {
2674 value_type = tjp;
2675 }
2676 }
2677
2678 int alias_idx = C->get_alias_index(adr_type);
2679
2680 // Memory-model-wise, a LoadStore acts like a little synchronized
2681 // block, so needs barriers on each side. These don't translate
2682 // into actual barriers on most machines, but we still need rest of
2683 // compiler to respect ordering.
2684
2685 insert_mem_bar(Op_MemBarRelease);
2686 insert_mem_bar(Op_MemBarCPUOrder);
2687
2688 // 4984716: MemBars must be inserted before this
2689 // memory node in order to avoid a false
2690 // dependency which will confuse the scheduler.
2691 Node *mem = memory(alias_idx);
2692
2693 // For now, we handle only those cases that actually exist: ints,
2694 // longs, and Object. Adding others should be straightforward.
2695 Node* load_store;
2696 switch(type) {
2697 case T_INT:
2698 if (kind == LS_xadd) {
2699 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2700 } else if (kind == LS_xchg) {
2701 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2702 } else if (kind == LS_cmpxchg) {
2703 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2704 } else {
2705 ShouldNotReachHere();
2706 }
2707 break;
2708 case T_LONG:
2709 if (kind == LS_xadd) {
2710 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2711 } else if (kind == LS_xchg) {
2712 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2713 } else if (kind == LS_cmpxchg) {
2714 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2715 } else {
2716 ShouldNotReachHere();
2717 }
2718 break;
2719 case T_OBJECT:
2720 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2721 // could be delayed during Parse (for example, in adjust_map_after_if()).
2722 // Execute transformation here to avoid barrier generation in such case.
2723 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2724 newval = _gvn.makecon(TypePtr::NULL_PTR);
2725
2726 // Reference stores need a store barrier.
2727 pre_barrier(true /* do_load*/,
2728 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2729 NULL /* pre_val*/,
2730 T_OBJECT);
2731 #ifdef _LP64
2732 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2733 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2734 if (kind == LS_xchg) {
2735 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2736 newval_enc, adr_type, value_type->make_narrowoop()));
2737 } else {
2738 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2739 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2740 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2741 newval_enc, oldval_enc));
2742 }
2743 } else
2744 #endif
2745 {
2746 if (kind == LS_xchg) {
2747 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2748 } else {
2749 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2750 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2751 }
2752 }
2753 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2754 break;
2755 default:
2756 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2757 break;
2758 }
2759
2760 // SCMemProjNodes represent the memory state of a LoadStore. Their
2761 // main role is to prevent LoadStore nodes from being optimized away
2762 // when their results aren't used.
2763 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2764 set_memory(proj, alias_idx);
2765
2766 // Add the trailing membar surrounding the access
2767 insert_mem_bar(Op_MemBarCPUOrder);
2768 insert_mem_bar(Op_MemBarAcquire);
2769
2770 #ifdef _LP64
2771 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2772 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->get_ptr_type()));
2773 }
2774 #endif
2775
2776 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2777 set_result(load_store);
2778 return true;
2779 }
2780
2781 //----------------------------inline_unsafe_ordered_store----------------------
2782 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2783 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2784 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2785 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2786 // This is another variant of inline_unsafe_access, differing in
2787 // that it always issues store-store ("release") barrier and ensures
2788 // store-atomicity (which only matters for "long").
2789
2790 if (callee()->is_static()) return false; // caller must have the capability!
2791
2792 #ifndef PRODUCT
2793 {
2794 ResourceMark rm;
2795 // Check the signatures.
2796 ciSignature* sig = callee()->signature();
2797 #ifdef ASSERT
2798 BasicType rtype = sig->return_type()->basic_type();
2799 assert(rtype == T_VOID, "must return void");
2800 assert(sig->count() == 3, "has 3 arguments");
2801 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2802 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2803 #endif // ASSERT
2804 }
2805 #endif //PRODUCT
2806
2807 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2808
2809 // Get arguments:
2810 Node* receiver = argument(0); // type: oop
2811 Node* base = argument(1); // type: oop
2812 Node* offset = argument(2); // type: long
2813 Node* val = argument(4); // type: oop, int, or long
2814
2815 // Null check receiver.
2816 receiver = null_check(receiver);
2817 if (stopped()) {
2818 return true;
2819 }
2820
2821 // Build field offset expression.
2822 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2823 // 32-bit machines ignore the high half of long offsets
2824 offset = ConvL2X(offset);
2825 Node* adr = make_unsafe_address(base, offset);
2826 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2827 const Type *value_type = Type::get_const_basic_type(type);
2828 Compile::AliasType* alias_type = C->alias_type(adr_type);
2829
2830 insert_mem_bar(Op_MemBarRelease);
2831 insert_mem_bar(Op_MemBarCPUOrder);
2832 // Ensure that the store is atomic for longs:
2833 const bool require_atomic_access = true;
2834 Node* store;
2835 if (type == T_OBJECT) // reference stores need a store barrier.
2836 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2837 else {
2838 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2839 }
2840 insert_mem_bar(Op_MemBarCPUOrder);
2841 return true;
2842 }
2843
2844 //----------------------------inline_unsafe_allocate---------------------------
2845 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2846 bool LibraryCallKit::inline_unsafe_allocate() {
2847 if (callee()->is_static()) return false; // caller must have the capability!
2848
2849 null_check_receiver(); // null-check, then ignore
2850 Node* cls = null_check(argument(1));
2851 if (stopped()) return true;
2852
2853 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2854 kls = null_check(kls);
2855 if (stopped()) return true; // argument was like int.class
2856
2857 // Note: The argument might still be an illegal value like
2858 // Serializable.class or Object[].class. The runtime will handle it.
2859 // But we must make an explicit check for initialization.
2860 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2861 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2862 // can generate code to load it as unsigned byte.
2863 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2864 Node* bits = intcon(instanceKlass::fully_initialized);
2865 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2866 // The 'test' is non-zero if we need to take a slow path.
2867
2868 Node* obj = new_instance(kls, test);
2869 set_result(obj);
2870 return true;
2871 }
2872
2873 #ifdef TRACE_HAVE_INTRINSICS
2874 /*
2875 * oop -> myklass
2876 * myklass->trace_id |= USED
2877 * return myklass->trace_id & ~0x3
2878 */
2879 bool LibraryCallKit::inline_native_classID() {
2880 null_check_receiver(); // null-check, then ignore
2881 Node* cls = null_check(argument(1), T_OBJECT);
2882 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2883 kls = null_check(kls, T_OBJECT);
2884 ByteSize offset = TRACE_ID_OFFSET;
2885 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2886 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2887 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2888 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2889 Node* clsused = longcon(0x01l); // set the class bit
2890 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2891
2892 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2893 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2894 set_result(andl);
2895 return true;
2896 }
2897
2898 bool LibraryCallKit::inline_native_threadID() {
2899 Node* tls_ptr = NULL;
2900 Node* cur_thr = generate_current_thread(tls_ptr);
2901 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2902 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2903 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2904
2905 Node* threadid = NULL;
2906 size_t thread_id_size = OSThread::thread_id_size();
2907 if (thread_id_size == (size_t) BytesPerLong) {
2908 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2909 } else if (thread_id_size == (size_t) BytesPerInt) {
2910 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2911 } else {
2912 ShouldNotReachHere();
2913 }
2914 set_result(threadid);
2915 return true;
2916 }
2917 #endif
2918
2919 //------------------------inline_native_time_funcs--------------
2920 // inline code for System.currentTimeMillis() and System.nanoTime()
2921 // these have the same type and signature
2922 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2923 const TypeFunc* tf = OptoRuntime::void_long_Type();
2924 const TypePtr* no_memory_effects = NULL;
2925 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2926 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2927 #ifdef ASSERT
2928 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2929 assert(value_top == top(), "second value must be top");
2930 #endif
2931 set_result(value);
2932 return true;
2933 }
2934
2935 //------------------------inline_native_currentThread------------------
2936 bool LibraryCallKit::inline_native_currentThread() {
2937 Node* junk = NULL;
2938 set_result(generate_current_thread(junk));
2939 return true;
2940 }
2941
2942 //------------------------inline_native_isInterrupted------------------
2943 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2944 bool LibraryCallKit::inline_native_isInterrupted() {
2945 // Add a fast path to t.isInterrupted(clear_int):
2946 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2947 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2948 // So, in the common case that the interrupt bit is false,
2949 // we avoid making a call into the VM. Even if the interrupt bit
2950 // is true, if the clear_int argument is false, we avoid the VM call.
2951 // However, if the receiver is not currentThread, we must call the VM,
2952 // because there must be some locking done around the operation.
2953
2954 // We only go to the fast case code if we pass two guards.
2955 // Paths which do not pass are accumulated in the slow_region.
2956
2957 enum {
2958 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2959 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2960 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2961 PATH_LIMIT
2962 };
2963
2964 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
2965 // out of the function.
2966 insert_mem_bar(Op_MemBarCPUOrder);
2967
2968 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
2969 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
2970
2971 RegionNode* slow_region = new (C) RegionNode(1);
2972 record_for_igvn(slow_region);
2973
2974 // (a) Receiving thread must be the current thread.
2975 Node* rec_thr = argument(0);
2976 Node* tls_ptr = NULL;
2977 Node* cur_thr = generate_current_thread(tls_ptr);
2978 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
2979 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
2980
2981 generate_slow_guard(bol_thr, slow_region);
2982
2983 // (b) Interrupt bit on TLS must be false.
2984 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2985 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2986 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2987
2988 // Set the control input on the field _interrupted read to prevent it floating up.
2989 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2990 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
2991 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
2992
2993 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2994
2995 // First fast path: if (!TLS._interrupted) return false;
2996 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
2997 result_rgn->init_req(no_int_result_path, false_bit);
2998 result_val->init_req(no_int_result_path, intcon(0));
2999
3000 // drop through to next case
3001 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3002
3003 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3004 Node* clr_arg = argument(1);
3005 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3006 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3007 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3008
3009 // Second fast path: ... else if (!clear_int) return true;
3010 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3011 result_rgn->init_req(no_clear_result_path, false_arg);
3012 result_val->init_req(no_clear_result_path, intcon(1));
3013
3014 // drop through to next case
3015 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3016
3017 // (d) Otherwise, go to the slow path.
3018 slow_region->add_req(control());
3019 set_control( _gvn.transform(slow_region) );
3020
3021 if (stopped()) {
3022 // There is no slow path.
3023 result_rgn->init_req(slow_result_path, top());
3024 result_val->init_req(slow_result_path, top());
3025 } else {
3026 // non-virtual because it is a private non-static
3027 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3028
3029 Node* slow_val = set_results_for_java_call(slow_call);
3030 // this->control() comes from set_results_for_java_call
3031
3032 Node* fast_io = slow_call->in(TypeFunc::I_O);
3033 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3034
3035 // These two phis are pre-filled with copies of of the fast IO and Memory
3036 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3037 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3038
3039 result_rgn->init_req(slow_result_path, control());
3040 result_io ->init_req(slow_result_path, i_o());
3041 result_mem->init_req(slow_result_path, reset_memory());
3042 result_val->init_req(slow_result_path, slow_val);
3043
3044 set_all_memory(_gvn.transform(result_mem));
3045 set_i_o( _gvn.transform(result_io));
3046 }
3047
3048 C->set_has_split_ifs(true); // Has chance for split-if optimization
3049 set_result(result_rgn, result_val);
3050 return true;
3051 }
3052
3053 //---------------------------load_mirror_from_klass----------------------------
3054 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3055 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3056 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3057 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3058 }
3059
3060 //-----------------------load_klass_from_mirror_common-------------------------
3061 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3062 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3063 // and branch to the given path on the region.
3064 // If never_see_null, take an uncommon trap on null, so we can optimistically
3065 // compile for the non-null case.
3066 // If the region is NULL, force never_see_null = true.
3067 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3068 bool never_see_null,
3069 RegionNode* region,
3070 int null_path,
3071 int offset) {
3072 if (region == NULL) never_see_null = true;
3073 Node* p = basic_plus_adr(mirror, offset);
3074 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3075 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3076 Node* null_ctl = top();
3077 kls = null_check_oop(kls, &null_ctl, never_see_null);
3078 if (region != NULL) {
3079 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3080 region->init_req(null_path, null_ctl);
3081 } else {
3082 assert(null_ctl == top(), "no loose ends");
3083 }
3084 return kls;
3085 }
3086
3087 //--------------------(inline_native_Class_query helpers)---------------------
3088 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3089 // Fall through if (mods & mask) == bits, take the guard otherwise.
3090 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3091 // Branch around if the given klass has the given modifier bit set.
3092 // Like generate_guard, adds a new path onto the region.
3093 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3094 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3095 Node* mask = intcon(modifier_mask);
3096 Node* bits = intcon(modifier_bits);
3097 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3098 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3099 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3100 return generate_fair_guard(bol, region);
3101 }
3102 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3103 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3104 }
3105
3106 //-------------------------inline_native_Class_query-------------------
3107 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3108 const Type* return_type = TypeInt::BOOL;
3109 Node* prim_return_value = top(); // what happens if it's a primitive class?
3110 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3111 bool expect_prim = false; // most of these guys expect to work on refs
3112
3113 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3114
3115 Node* mirror = argument(0);
3116 Node* obj = top();
3117
3118 switch (id) {
3119 case vmIntrinsics::_isInstance:
3120 // nothing is an instance of a primitive type
3121 prim_return_value = intcon(0);
3122 obj = argument(1);
3123 break;
3124 case vmIntrinsics::_getModifiers:
3125 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3126 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3127 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3128 break;
3129 case vmIntrinsics::_isInterface:
3130 prim_return_value = intcon(0);
3131 break;
3132 case vmIntrinsics::_isArray:
3133 prim_return_value = intcon(0);
3134 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3135 break;
3136 case vmIntrinsics::_isPrimitive:
3137 prim_return_value = intcon(1);
3138 expect_prim = true; // obviously
3139 break;
3140 case vmIntrinsics::_getSuperclass:
3141 prim_return_value = null();
3142 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3143 break;
3144 case vmIntrinsics::_getComponentType:
3145 prim_return_value = null();
3146 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3147 break;
3148 case vmIntrinsics::_getClassAccessFlags:
3149 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3150 return_type = TypeInt::INT; // not bool! 6297094
3151 break;
3152 default:
3153 fatal_unexpected_iid(id);
3154 break;
3155 }
3156
3157 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3158 if (mirror_con == NULL) return false; // cannot happen?
3159
3160 #ifndef PRODUCT
3161 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3162 ciType* k = mirror_con->java_mirror_type();
3163 if (k) {
3164 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3165 k->print_name();
3166 tty->cr();
3167 }
3168 }
3169 #endif
3170
3171 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3172 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3173 record_for_igvn(region);
3174 PhiNode* phi = new (C) PhiNode(region, return_type);
3175
3176 // The mirror will never be null of Reflection.getClassAccessFlags, however
3177 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3178 // if it is. See bug 4774291.
3179
3180 // For Reflection.getClassAccessFlags(), the null check occurs in
3181 // the wrong place; see inline_unsafe_access(), above, for a similar
3182 // situation.
3183 mirror = null_check(mirror);
3184 // If mirror or obj is dead, only null-path is taken.
3185 if (stopped()) return true;
3186
3187 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3188
3189 // Now load the mirror's klass metaobject, and null-check it.
3190 // Side-effects region with the control path if the klass is null.
3191 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3192 // If kls is null, we have a primitive mirror.
3193 phi->init_req(_prim_path, prim_return_value);
3194 if (stopped()) { set_result(region, phi); return true; }
3195
3196 Node* p; // handy temp
3197 Node* null_ctl;
3198
3199 // Now that we have the non-null klass, we can perform the real query.
3200 // For constant classes, the query will constant-fold in LoadNode::Value.
3201 Node* query_value = top();
3202 switch (id) {
3203 case vmIntrinsics::_isInstance:
3204 // nothing is an instance of a primitive type
3205 query_value = gen_instanceof(obj, kls);
3206 break;
3207
3208 case vmIntrinsics::_getModifiers:
3209 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3210 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3211 break;
3212
3213 case vmIntrinsics::_isInterface:
3214 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3215 if (generate_interface_guard(kls, region) != NULL)
3216 // A guard was added. If the guard is taken, it was an interface.
3217 phi->add_req(intcon(1));
3218 // If we fall through, it's a plain class.
3219 query_value = intcon(0);
3220 break;
3221
3222 case vmIntrinsics::_isArray:
3223 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3224 if (generate_array_guard(kls, region) != NULL)
3225 // A guard was added. If the guard is taken, it was an array.
3226 phi->add_req(intcon(1));
3227 // If we fall through, it's a plain class.
3228 query_value = intcon(0);
3229 break;
3230
3231 case vmIntrinsics::_isPrimitive:
3232 query_value = intcon(0); // "normal" path produces false
3233 break;
3234
3235 case vmIntrinsics::_getSuperclass:
3236 // The rules here are somewhat unfortunate, but we can still do better
3237 // with random logic than with a JNI call.
3238 // Interfaces store null or Object as _super, but must report null.
3239 // Arrays store an intermediate super as _super, but must report Object.
3240 // Other types can report the actual _super.
3241 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3242 if (generate_interface_guard(kls, region) != NULL)
3243 // A guard was added. If the guard is taken, it was an interface.
3244 phi->add_req(null());
3245 if (generate_array_guard(kls, region) != NULL)
3246 // A guard was added. If the guard is taken, it was an array.
3247 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3248 // If we fall through, it's a plain class. Get its _super.
3249 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3250 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3251 null_ctl = top();
3252 kls = null_check_oop(kls, &null_ctl);
3253 if (null_ctl != top()) {
3254 // If the guard is taken, Object.superClass is null (both klass and mirror).
3255 region->add_req(null_ctl);
3256 phi ->add_req(null());
3257 }
3258 if (!stopped()) {
3259 query_value = load_mirror_from_klass(kls);
3260 }
3261 break;
3262
3263 case vmIntrinsics::_getComponentType:
3264 if (generate_array_guard(kls, region) != NULL) {
3265 // Be sure to pin the oop load to the guard edge just created:
3266 Node* is_array_ctrl = region->in(region->req()-1);
3267 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3268 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3269 phi->add_req(cmo);
3270 }
3271 query_value = null(); // non-array case is null
3272 break;
3273
3274 case vmIntrinsics::_getClassAccessFlags:
3275 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3276 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3277 break;
3278
3279 default:
3280 fatal_unexpected_iid(id);
3281 break;
3282 }
3283
3284 // Fall-through is the normal case of a query to a real class.
3285 phi->init_req(1, query_value);
3286 region->init_req(1, control());
3287
3288 C->set_has_split_ifs(true); // Has chance for split-if optimization
3289 set_result(region, phi);
3290 return true;
3291 }
3292
3293 //--------------------------inline_native_subtype_check------------------------
3294 // This intrinsic takes the JNI calls out of the heart of
3295 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3296 bool LibraryCallKit::inline_native_subtype_check() {
3297 // Pull both arguments off the stack.
3298 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3299 args[0] = argument(0);
3300 args[1] = argument(1);
3301 Node* klasses[2]; // corresponding Klasses: superk, subk
3302 klasses[0] = klasses[1] = top();
3303
3304 enum {
3305 // A full decision tree on {superc is prim, subc is prim}:
3306 _prim_0_path = 1, // {P,N} => false
3307 // {P,P} & superc!=subc => false
3308 _prim_same_path, // {P,P} & superc==subc => true
3309 _prim_1_path, // {N,P} => false
3310 _ref_subtype_path, // {N,N} & subtype check wins => true
3311 _both_ref_path, // {N,N} & subtype check loses => false
3312 PATH_LIMIT
3313 };
3314
3315 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3316 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3317 record_for_igvn(region);
3318
3319 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3320 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3321 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3322
3323 // First null-check both mirrors and load each mirror's klass metaobject.
3324 int which_arg;
3325 for (which_arg = 0; which_arg <= 1; which_arg++) {
3326 Node* arg = args[which_arg];
3327 arg = null_check(arg);
3328 if (stopped()) break;
3329 args[which_arg] = arg;
3330
3331 Node* p = basic_plus_adr(arg, class_klass_offset);
3332 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3333 klasses[which_arg] = _gvn.transform(kls);
3334 }
3335
3336 // Having loaded both klasses, test each for null.
3337 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3338 for (which_arg = 0; which_arg <= 1; which_arg++) {
3339 Node* kls = klasses[which_arg];
3340 Node* null_ctl = top();
3341 kls = null_check_oop(kls, &null_ctl, never_see_null);
3342 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3343 region->init_req(prim_path, null_ctl);
3344 if (stopped()) break;
3345 klasses[which_arg] = kls;
3346 }
3347
3348 if (!stopped()) {
3349 // now we have two reference types, in klasses[0..1]
3350 Node* subk = klasses[1]; // the argument to isAssignableFrom
3351 Node* superk = klasses[0]; // the receiver
3352 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3353 // now we have a successful reference subtype check
3354 region->set_req(_ref_subtype_path, control());
3355 }
3356
3357 // If both operands are primitive (both klasses null), then
3358 // we must return true when they are identical primitives.
3359 // It is convenient to test this after the first null klass check.
3360 set_control(region->in(_prim_0_path)); // go back to first null check
3361 if (!stopped()) {
3362 // Since superc is primitive, make a guard for the superc==subc case.
3363 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3364 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3365 generate_guard(bol_eq, region, PROB_FAIR);
3366 if (region->req() == PATH_LIMIT+1) {
3367 // A guard was added. If the added guard is taken, superc==subc.
3368 region->swap_edges(PATH_LIMIT, _prim_same_path);
3369 region->del_req(PATH_LIMIT);
3370 }
3371 region->set_req(_prim_0_path, control()); // Not equal after all.
3372 }
3373
3374 // these are the only paths that produce 'true':
3375 phi->set_req(_prim_same_path, intcon(1));
3376 phi->set_req(_ref_subtype_path, intcon(1));
3377
3378 // pull together the cases:
3379 assert(region->req() == PATH_LIMIT, "sane region");
3380 for (uint i = 1; i < region->req(); i++) {
3381 Node* ctl = region->in(i);
3382 if (ctl == NULL || ctl == top()) {
3383 region->set_req(i, top());
3384 phi ->set_req(i, top());
3385 } else if (phi->in(i) == NULL) {
3386 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3387 }
3388 }
3389
3390 set_control(_gvn.transform(region));
3391 set_result(_gvn.transform(phi));
3392 return true;
3393 }
3394
3395 //---------------------generate_array_guard_common------------------------
3396 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3397 bool obj_array, bool not_array) {
3398 // If obj_array/non_array==false/false:
3399 // Branch around if the given klass is in fact an array (either obj or prim).
3400 // If obj_array/non_array==false/true:
3401 // Branch around if the given klass is not an array klass of any kind.
3402 // If obj_array/non_array==true/true:
3403 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3404 // If obj_array/non_array==true/false:
3405 // Branch around if the kls is an oop array (Object[] or subtype)
3406 //
3407 // Like generate_guard, adds a new path onto the region.
3408 jint layout_con = 0;
3409 Node* layout_val = get_layout_helper(kls, layout_con);
3410 if (layout_val == NULL) {
3411 bool query = (obj_array
3412 ? Klass::layout_helper_is_objArray(layout_con)
3413 : Klass::layout_helper_is_javaArray(layout_con));
3414 if (query == not_array) {
3415 return NULL; // never a branch
3416 } else { // always a branch
3417 Node* always_branch = control();
3418 if (region != NULL)
3419 region->add_req(always_branch);
3420 set_control(top());
3421 return always_branch;
3422 }
3423 }
3424 // Now test the correct condition.
3425 jint nval = (obj_array
3426 ? ((jint)Klass::_lh_array_tag_type_value
3427 << Klass::_lh_array_tag_shift)
3428 : Klass::_lh_neutral_value);
3429 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3430 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3431 // invert the test if we are looking for a non-array
3432 if (not_array) btest = BoolTest(btest).negate();
3433 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3434 return generate_fair_guard(bol, region);
3435 }
3436
3437
3438 //-----------------------inline_native_newArray--------------------------
3439 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3440 bool LibraryCallKit::inline_native_newArray() {
3441 Node* mirror = argument(0);
3442 Node* count_val = argument(1);
3443
3444 mirror = null_check(mirror);
3445 // If mirror or obj is dead, only null-path is taken.
3446 if (stopped()) return true;
3447
3448 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3449 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3450 PhiNode* result_val = new(C) PhiNode(result_reg,
3451 TypeInstPtr::NOTNULL);
3452 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3453 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3454 TypePtr::BOTTOM);
3455
3456 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3457 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3458 result_reg, _slow_path);
3459 Node* normal_ctl = control();
3460 Node* no_array_ctl = result_reg->in(_slow_path);
3461
3462 // Generate code for the slow case. We make a call to newArray().
3463 set_control(no_array_ctl);
3464 if (!stopped()) {
3465 // Either the input type is void.class, or else the
3466 // array klass has not yet been cached. Either the
3467 // ensuing call will throw an exception, or else it
3468 // will cache the array klass for next time.
3469 PreserveJVMState pjvms(this);
3470 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3471 Node* slow_result = set_results_for_java_call(slow_call);
3472 // this->control() comes from set_results_for_java_call
3473 result_reg->set_req(_slow_path, control());
3474 result_val->set_req(_slow_path, slow_result);
3475 result_io ->set_req(_slow_path, i_o());
3476 result_mem->set_req(_slow_path, reset_memory());
3477 }
3478
3479 set_control(normal_ctl);
3480 if (!stopped()) {
3481 // Normal case: The array type has been cached in the java.lang.Class.
3482 // The following call works fine even if the array type is polymorphic.
3483 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3484 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3485 result_reg->init_req(_normal_path, control());
3486 result_val->init_req(_normal_path, obj);
3487 result_io ->init_req(_normal_path, i_o());
3488 result_mem->init_req(_normal_path, reset_memory());
3489 }
3490
3491 // Return the combined state.
3492 set_i_o( _gvn.transform(result_io) );
3493 set_all_memory( _gvn.transform(result_mem) );
3494
3495 C->set_has_split_ifs(true); // Has chance for split-if optimization
3496 set_result(result_reg, result_val);
3497 return true;
3498 }
3499
3500 //----------------------inline_native_getLength--------------------------
3501 // public static native int java.lang.reflect.Array.getLength(Object array);
3502 bool LibraryCallKit::inline_native_getLength() {
3503 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3504
3505 Node* array = null_check(argument(0));
3506 // If array is dead, only null-path is taken.
3507 if (stopped()) return true;
3508
3509 // Deoptimize if it is a non-array.
3510 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3511
3512 if (non_array != NULL) {
3513 PreserveJVMState pjvms(this);
3514 set_control(non_array);
3515 uncommon_trap(Deoptimization::Reason_intrinsic,
3516 Deoptimization::Action_maybe_recompile);
3517 }
3518
3519 // If control is dead, only non-array-path is taken.
3520 if (stopped()) return true;
3521
3522 // The works fine even if the array type is polymorphic.
3523 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3524 Node* result = load_array_length(array);
3525
3526 C->set_has_split_ifs(true); // Has chance for split-if optimization
3527 set_result(result);
3528 return true;
3529 }
3530
3531 //------------------------inline_array_copyOf----------------------------
3532 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3533 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3534 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3535 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3536
3537 // Get the arguments.
3538 Node* original = argument(0);
3539 Node* start = is_copyOfRange? argument(1): intcon(0);
3540 Node* end = is_copyOfRange? argument(2): argument(1);
3541 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3542
3543 Node* newcopy;
3544
3545 // Set the original stack and the reexecute bit for the interpreter to reexecute
3546 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3547 { PreserveReexecuteState preexecs(this);
3548 jvms()->set_should_reexecute(true);
3549
3550 array_type_mirror = null_check(array_type_mirror);
3551 original = null_check(original);
3552
3553 // Check if a null path was taken unconditionally.
3554 if (stopped()) return true;
3555
3556 Node* orig_length = load_array_length(original);
3557
3558 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3559 klass_node = null_check(klass_node);
3560
3561 RegionNode* bailout = new (C) RegionNode(1);
3562 record_for_igvn(bailout);
3563
3564 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3565 // Bail out if that is so.
3566 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3567 if (not_objArray != NULL) {
3568 // Improve the klass node's type from the new optimistic assumption:
3569 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3570 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3571 Node* cast = new (C) CastPPNode(klass_node, akls);
3572 cast->init_req(0, control());
3573 klass_node = _gvn.transform(cast);
3574 }
3575
3576 // Bail out if either start or end is negative.
3577 generate_negative_guard(start, bailout, &start);
3578 generate_negative_guard(end, bailout, &end);
3579
3580 Node* length = end;
3581 if (_gvn.type(start) != TypeInt::ZERO) {
3582 length = _gvn.transform(new (C) SubINode(end, start));
3583 }
3584
3585 // Bail out if length is negative.
3586 // Without this the new_array would throw
3587 // NegativeArraySizeException but IllegalArgumentException is what
3588 // should be thrown
3589 generate_negative_guard(length, bailout, &length);
3590
3591 if (bailout->req() > 1) {
3592 PreserveJVMState pjvms(this);
3593 set_control(_gvn.transform(bailout));
3594 uncommon_trap(Deoptimization::Reason_intrinsic,
3595 Deoptimization::Action_maybe_recompile);
3596 }
3597
3598 if (!stopped()) {
3599 // How many elements will we copy from the original?
3600 // The answer is MinI(orig_length - start, length).
3601 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3602 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3603
3604 newcopy = new_array(klass_node, length, 0); // no argments to push
3605
3606 // Generate a direct call to the right arraycopy function(s).
3607 // We know the copy is disjoint but we might not know if the
3608 // oop stores need checking.
3609 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3610 // This will fail a store-check if x contains any non-nulls.
3611 bool disjoint_bases = true;
3612 // if start > orig_length then the length of the copy may be
3613 // negative.
3614 bool length_never_negative = !is_copyOfRange;
3615 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3616 original, start, newcopy, intcon(0), moved,
3617 disjoint_bases, length_never_negative);
3618 }
3619 } // original reexecute is set back here
3620
3621 C->set_has_split_ifs(true); // Has chance for split-if optimization
3622 if (!stopped()) {
3623 set_result(newcopy);
3624 }
3625 return true;
3626 }
3627
3628
3629 //----------------------generate_virtual_guard---------------------------
3630 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3631 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3632 RegionNode* slow_region) {
3633 ciMethod* method = callee();
3634 int vtable_index = method->vtable_index();
3635 // Get the methodOop out of the appropriate vtable entry.
3636 int entry_offset = (instanceKlass::vtable_start_offset() +
3637 vtable_index*vtableEntry::size()) * wordSize +
3638 vtableEntry::method_offset_in_bytes();
3639 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3640 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3641
3642 // Compare the target method with the expected method (e.g., Object.hashCode).
3643 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3644
3645 Node* native_call = makecon(native_call_addr);
3646 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3647 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3648
3649 return generate_slow_guard(test_native, slow_region);
3650 }
3651
3652 //-----------------------generate_method_call----------------------------
3653 // Use generate_method_call to make a slow-call to the real
3654 // method if the fast path fails. An alternative would be to
3655 // use a stub like OptoRuntime::slow_arraycopy_Java.
3656 // This only works for expanding the current library call,
3657 // not another intrinsic. (E.g., don't use this for making an
3658 // arraycopy call inside of the copyOf intrinsic.)
3659 CallJavaNode*
3660 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3661 // When compiling the intrinsic method itself, do not use this technique.
3662 guarantee(callee() != C->method(), "cannot make slow-call to self");
3663
3664 ciMethod* method = callee();
3665 // ensure the JVMS we have will be correct for this call
3666 guarantee(method_id == method->intrinsic_id(), "must match");
3667
3668 const TypeFunc* tf = TypeFunc::make(method);
3669 CallJavaNode* slow_call;
3670 if (is_static) {
3671 assert(!is_virtual, "");
3672 slow_call = new(C) CallStaticJavaNode(tf,
3673 SharedRuntime::get_resolve_static_call_stub(),
3674 method, bci());
3675 } else if (is_virtual) {
3676 null_check_receiver();
3677 int vtable_index = methodOopDesc::invalid_vtable_index;
3678 if (UseInlineCaches) {
3679 // Suppress the vtable call
3680 } else {
3681 // hashCode and clone are not a miranda methods,
3682 // so the vtable index is fixed.
3683 // No need to use the linkResolver to get it.
3684 vtable_index = method->vtable_index();
3685 }
3686 slow_call = new(C) CallDynamicJavaNode(tf,
3687 SharedRuntime::get_resolve_virtual_call_stub(),
3688 method, vtable_index, bci());
3689 } else { // neither virtual nor static: opt_virtual
3690 null_check_receiver();
3691 slow_call = new(C) CallStaticJavaNode(tf,
3692 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3693 method, bci());
3694 slow_call->set_optimized_virtual(true);
3695 }
3696 set_arguments_for_java_call(slow_call);
3697 set_edges_for_java_call(slow_call);
3698 return slow_call;
3699 }
3700
3701
3702 //------------------------------inline_native_hashcode--------------------
3703 // Build special case code for calls to hashCode on an object.
3704 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3705 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3706 assert(!(is_virtual && is_static), "either virtual, special, or static");
3707
3708 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3709
3710 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3711 PhiNode* result_val = new(C) PhiNode(result_reg,
3712 TypeInt::INT);
3713 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3714 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3715 TypePtr::BOTTOM);
3716 Node* obj = NULL;
3717 if (!is_static) {
3718 // Check for hashing null object
3719 obj = null_check_receiver();
3720 if (stopped()) return true; // unconditionally null
3721 result_reg->init_req(_null_path, top());
3722 result_val->init_req(_null_path, top());
3723 } else {
3724 // Do a null check, and return zero if null.
3725 // System.identityHashCode(null) == 0
3726 obj = argument(0);
3727 Node* null_ctl = top();
3728 obj = null_check_oop(obj, &null_ctl);
3729 result_reg->init_req(_null_path, null_ctl);
3730 result_val->init_req(_null_path, _gvn.intcon(0));
3731 }
3732
3733 // Unconditionally null? Then return right away.
3734 if (stopped()) {
3735 set_control( result_reg->in(_null_path));
3736 if (!stopped())
3737 set_result(result_val->in(_null_path));
3738 return true;
3739 }
3740
3741 // After null check, get the object's klass.
3742 Node* obj_klass = load_object_klass(obj);
3743
3744 // This call may be virtual (invokevirtual) or bound (invokespecial).
3745 // For each case we generate slightly different code.
3746
3747 // We only go to the fast case code if we pass a number of guards. The
3748 // paths which do not pass are accumulated in the slow_region.
3749 RegionNode* slow_region = new (C) RegionNode(1);
3750 record_for_igvn(slow_region);
3751
3752 // If this is a virtual call, we generate a funny guard. We pull out
3753 // the vtable entry corresponding to hashCode() from the target object.
3754 // If the target method which we are calling happens to be the native
3755 // Object hashCode() method, we pass the guard. We do not need this
3756 // guard for non-virtual calls -- the caller is known to be the native
3757 // Object hashCode().
3758 if (is_virtual) {
3759 generate_virtual_guard(obj_klass, slow_region);
3760 }
3761
3762 // Get the header out of the object, use LoadMarkNode when available
3763 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3764 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3765
3766 // Test the header to see if it is unlocked.
3767 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3768 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3769 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3770 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3771 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3772
3773 generate_slow_guard(test_unlocked, slow_region);
3774
3775 // Get the hash value and check to see that it has been properly assigned.
3776 // We depend on hash_mask being at most 32 bits and avoid the use of
3777 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3778 // vm: see markOop.hpp.
3779 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3780 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3781 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3782 // This hack lets the hash bits live anywhere in the mark object now, as long
3783 // as the shift drops the relevant bits into the low 32 bits. Note that
3784 // Java spec says that HashCode is an int so there's no point in capturing
3785 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3786 hshifted_header = ConvX2I(hshifted_header);
3787 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3788
3789 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3790 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3791 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3792
3793 generate_slow_guard(test_assigned, slow_region);
3794
3795 Node* init_mem = reset_memory();
3796 // fill in the rest of the null path:
3797 result_io ->init_req(_null_path, i_o());
3798 result_mem->init_req(_null_path, init_mem);
3799
3800 result_val->init_req(_fast_path, hash_val);
3801 result_reg->init_req(_fast_path, control());
3802 result_io ->init_req(_fast_path, i_o());
3803 result_mem->init_req(_fast_path, init_mem);
3804
3805 // Generate code for the slow case. We make a call to hashCode().
3806 set_control(_gvn.transform(slow_region));
3807 if (!stopped()) {
3808 // No need for PreserveJVMState, because we're using up the present state.
3809 set_all_memory(init_mem);
3810 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3811 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3812 Node* slow_result = set_results_for_java_call(slow_call);
3813 // this->control() comes from set_results_for_java_call
3814 result_reg->init_req(_slow_path, control());
3815 result_val->init_req(_slow_path, slow_result);
3816 result_io ->set_req(_slow_path, i_o());
3817 result_mem ->set_req(_slow_path, reset_memory());
3818 }
3819
3820 // Return the combined state.
3821 set_i_o( _gvn.transform(result_io) );
3822 set_all_memory( _gvn.transform(result_mem) );
3823
3824 set_result(result_reg, result_val);
3825 return true;
3826 }
3827
3828 //---------------------------inline_native_getClass----------------------------
3829 // public final native Class<?> java.lang.Object.getClass();
3830 //
3831 // Build special case code for calls to getClass on an object.
3832 bool LibraryCallKit::inline_native_getClass() {
3833 Node* obj = null_check_receiver();
3834 if (stopped()) return true;
3835 set_result(load_mirror_from_klass(load_object_klass(obj)));
3836 return true;
3837 }
3838
3839 //-----------------inline_native_Reflection_getCallerClass---------------------
3840 // public static native Class<?> sun.reflect.Reflection.getCallerClass(int realFramesToSkip);
3841 //
3842 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3843 //
3844 // NOTE that this code must perform the same logic as
3845 // vframeStream::security_get_caller_frame in that it must skip
3846 // Method.invoke() and auxiliary frames.
3847 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3848 #ifndef PRODUCT
3849 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3850 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3851 }
3852 #endif
3853
3854 Node* caller_depth_node = argument(0);
3855
3856 // The depth value must be a constant in order for the runtime call
3857 // to be eliminated.
3858 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3859 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3860 #ifndef PRODUCT
3861 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3862 tty->print_cr(" Bailing out because caller depth was not a constant");
3863 }
3864 #endif
3865 return false;
3866 }
3867 // Note that the JVM state at this point does not include the
3868 // getCallerClass() frame which we are trying to inline. The
3869 // semantics of getCallerClass(), however, are that the "first"
3870 // frame is the getCallerClass() frame, so we subtract one from the
3871 // requested depth before continuing. We don't inline requests of
3872 // getCallerClass(0).
3873 int caller_depth = caller_depth_type->get_con() - 1;
3874 if (caller_depth < 0) {
3875 #ifndef PRODUCT
3876 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3877 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3878 }
3879 #endif
3880 return false;
3881 }
3882
3883 if (!jvms()->has_method()) {
3884 #ifndef PRODUCT
3885 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3886 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3887 }
3888 #endif
3889 return false;
3890 }
3891 int _depth = jvms()->depth(); // cache call chain depth
3892
3893 // Walk back up the JVM state to find the caller at the required
3894 // depth. NOTE that this code must perform the same logic as
3895 // vframeStream::security_get_caller_frame in that it must skip
3896 // Method.invoke() and auxiliary frames. Note also that depth is
3897 // 1-based (1 is the bottom of the inlining).
3898 int inlining_depth = _depth;
3899 JVMState* caller_jvms = NULL;
3900
3901 if (inlining_depth > 0) {
3902 caller_jvms = jvms();
3903 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3904 do {
3905 // The following if-tests should be performed in this order
3906 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3907 // Skip a Method.invoke() or auxiliary frame
3908 } else if (caller_depth > 0) {
3909 // Skip real frame
3910 --caller_depth;
3911 } else {
3912 // We're done: reached desired caller after skipping.
3913 break;
3914 }
3915 caller_jvms = caller_jvms->caller();
3916 --inlining_depth;
3917 } while (inlining_depth > 0);
3918 }
3919
3920 if (inlining_depth == 0) {
3921 #ifndef PRODUCT
3922 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3923 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3924 tty->print_cr(" JVM state at this point:");
3925 for (int i = _depth; i >= 1; i--) {
3926 ciMethod* m = jvms()->of_depth(i)->method();
3927 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3928 }
3929 }
3930 #endif
3931 return false; // Reached end of inlining
3932 }
3933
3934 // Acquire method holder as java.lang.Class
3935 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3936 ciInstance* caller_mirror = caller_klass->java_mirror();
3937
3938 // Push this as a constant
3939 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3940
3941 #ifndef PRODUCT
3942 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3943 tty->print_cr(" Succeeded: caller = %s.%s, caller depth = %d, depth = %d", caller_klass->name()->as_utf8(), caller_jvms->method()->name()->as_utf8(), caller_depth_type->get_con(), _depth);
3944 tty->print_cr(" JVM state at this point:");
3945 for (int i = _depth; i >= 1; i--) {
3946 ciMethod* m = jvms()->of_depth(i)->method();
3947 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3948 }
3949 }
3950 #endif
3951 return true;
3952 }
3953
3954 // Helper routine for above
3955 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3956 ciMethod* method = jvms->method();
3957
3958 // Is this the Method.invoke method itself?
3959 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3960 return true;
3961
3962 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3963 ciKlass* k = method->holder();
3964 if (k->is_instance_klass()) {
3965 ciInstanceKlass* ik = k->as_instance_klass();
3966 for (; ik != NULL; ik = ik->super()) {
3967 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3968 ik == env()->find_system_klass(ik->name())) {
3969 return true;
3970 }
3971 }
3972 }
3973
3974 if (method->is_method_handle_intrinsic() ||
3975 method->is_compiled_lambda_form()) {
3976 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3977 return true;
3978 }
3979
3980 return false;
3981 }
3982
3983 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3984 Node* arg = argument(0);
3985 Node* result;
3986
3987 switch (id) {
3988 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
3989 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
3990 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
3991 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
3992
3993 case vmIntrinsics::_doubleToLongBits: {
3994 // two paths (plus control) merge in a wood
3995 RegionNode *r = new (C) RegionNode(3);
3996 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
3997
3998 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
3999 // Build the boolean node
4000 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4001
4002 // Branch either way.
4003 // NaN case is less traveled, which makes all the difference.
4004 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4005 Node *opt_isnan = _gvn.transform(ifisnan);
4006 assert( opt_isnan->is_If(), "Expect an IfNode");
4007 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4008 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4009
4010 set_control(iftrue);
4011
4012 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4013 Node *slow_result = longcon(nan_bits); // return NaN
4014 phi->init_req(1, _gvn.transform( slow_result ));
4015 r->init_req(1, iftrue);
4016
4017 // Else fall through
4018 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4019 set_control(iffalse);
4020
4021 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4022 r->init_req(2, iffalse);
4023
4024 // Post merge
4025 set_control(_gvn.transform(r));
4026 record_for_igvn(r);
4027
4028 C->set_has_split_ifs(true); // Has chance for split-if optimization
4029 result = phi;
4030 assert(result->bottom_type()->isa_long(), "must be");
4031 break;
4032 }
4033
4034 case vmIntrinsics::_floatToIntBits: {
4035 // two paths (plus control) merge in a wood
4036 RegionNode *r = new (C) RegionNode(3);
4037 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4038
4039 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4040 // Build the boolean node
4041 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4042
4043 // Branch either way.
4044 // NaN case is less traveled, which makes all the difference.
4045 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4046 Node *opt_isnan = _gvn.transform(ifisnan);
4047 assert( opt_isnan->is_If(), "Expect an IfNode");
4048 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4049 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4050
4051 set_control(iftrue);
4052
4053 static const jint nan_bits = 0x7fc00000;
4054 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4055 phi->init_req(1, _gvn.transform( slow_result ));
4056 r->init_req(1, iftrue);
4057
4058 // Else fall through
4059 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4060 set_control(iffalse);
4061
4062 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4063 r->init_req(2, iffalse);
4064
4065 // Post merge
4066 set_control(_gvn.transform(r));
4067 record_for_igvn(r);
4068
4069 C->set_has_split_ifs(true); // Has chance for split-if optimization
4070 result = phi;
4071 assert(result->bottom_type()->isa_int(), "must be");
4072 break;
4073 }
4074
4075 default:
4076 fatal_unexpected_iid(id);
4077 break;
4078 }
4079 set_result(_gvn.transform(result));
4080 return true;
4081 }
4082
4083 #ifdef _LP64
4084 #define XTOP ,top() /*additional argument*/
4085 #else //_LP64
4086 #define XTOP /*no additional argument*/
4087 #endif //_LP64
4088
4089 //----------------------inline_unsafe_copyMemory-------------------------
4090 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4091 bool LibraryCallKit::inline_unsafe_copyMemory() {
4092 if (callee()->is_static()) return false; // caller must have the capability!
4093 null_check_receiver(); // null-check receiver
4094 if (stopped()) return true;
4095
4096 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4097
4098 Node* src_ptr = argument(1); // type: oop
4099 Node* src_off = ConvL2X(argument(2)); // type: long
4100 Node* dst_ptr = argument(4); // type: oop
4101 Node* dst_off = ConvL2X(argument(5)); // type: long
4102 Node* size = ConvL2X(argument(7)); // type: long
4103
4104 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4105 "fieldOffset must be byte-scaled");
4106
4107 Node* src = make_unsafe_address(src_ptr, src_off);
4108 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4109
4110 // Conservatively insert a memory barrier on all memory slices.
4111 // Do not let writes of the copy source or destination float below the copy.
4112 insert_mem_bar(Op_MemBarCPUOrder);
4113
4114 // Call it. Note that the length argument is not scaled.
4115 make_runtime_call(RC_LEAF|RC_NO_FP,
4116 OptoRuntime::fast_arraycopy_Type(),
4117 StubRoutines::unsafe_arraycopy(),
4118 "unsafe_arraycopy",
4119 TypeRawPtr::BOTTOM,
4120 src, dst, size XTOP);
4121
4122 // Do not let reads of the copy destination float above the copy.
4123 insert_mem_bar(Op_MemBarCPUOrder);
4124
4125 return true;
4126 }
4127
4128 //------------------------clone_coping-----------------------------------
4129 // Helper function for inline_native_clone.
4130 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4131 assert(obj_size != NULL, "");
4132 Node* raw_obj = alloc_obj->in(1);
4133 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4134
4135 AllocateNode* alloc = NULL;
4136 if (ReduceBulkZeroing) {
4137 // We will be completely responsible for initializing this object -
4138 // mark Initialize node as complete.
4139 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4140 // The object was just allocated - there should be no any stores!
4141 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4142 // Mark as complete_with_arraycopy so that on AllocateNode
4143 // expansion, we know this AllocateNode is initialized by an array
4144 // copy and a StoreStore barrier exists after the array copy.
4145 alloc->initialization()->set_complete_with_arraycopy();
4146 }
4147
4148 // Copy the fastest available way.
4149 // TODO: generate fields copies for small objects instead.
4150 Node* src = obj;
4151 Node* dest = alloc_obj;
4152 Node* size = _gvn.transform(obj_size);
4153
4154 // Exclude the header but include array length to copy by 8 bytes words.
4155 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4156 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4157 instanceOopDesc::base_offset_in_bytes();
4158 // base_off:
4159 // 8 - 32-bit VM
4160 // 12 - 64-bit VM, compressed oops
4161 // 16 - 64-bit VM, normal oops
4162 if (base_off % BytesPerLong != 0) {
4163 assert(UseCompressedOops, "");
4164 if (is_array) {
4165 // Exclude length to copy by 8 bytes words.
4166 base_off += sizeof(int);
4167 } else {
4168 // Include klass to copy by 8 bytes words.
4169 base_off = instanceOopDesc::klass_offset_in_bytes();
4170 }
4171 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4172 }
4173 src = basic_plus_adr(src, base_off);
4174 dest = basic_plus_adr(dest, base_off);
4175
4176 // Compute the length also, if needed:
4177 Node* countx = size;
4178 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4179 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4180
4181 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4182 bool disjoint_bases = true;
4183 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4184 src, NULL, dest, NULL, countx,
4185 /*dest_uninitialized*/true);
4186
4187 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4188 if (card_mark) {
4189 assert(!is_array, "");
4190 // Put in store barrier for any and all oops we are sticking
4191 // into this object. (We could avoid this if we could prove
4192 // that the object type contains no oop fields at all.)
4193 Node* no_particular_value = NULL;
4194 Node* no_particular_field = NULL;
4195 int raw_adr_idx = Compile::AliasIdxRaw;
4196 post_barrier(control(),
4197 memory(raw_adr_type),
4198 alloc_obj,
4199 no_particular_field,
4200 raw_adr_idx,
4201 no_particular_value,
4202 T_OBJECT,
4203 false);
4204 }
4205
4206 // Do not let reads from the cloned object float above the arraycopy.
4207 if (alloc != NULL) {
4208 // Do not let stores that initialize this object be reordered with
4209 // a subsequent store that would make this object accessible by
4210 // other threads.
4211 // Record what AllocateNode this StoreStore protects so that
4212 // escape analysis can go from the MemBarStoreStoreNode to the
4213 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4214 // based on the escape status of the AllocateNode.
4215 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4216 } else {
4217 insert_mem_bar(Op_MemBarCPUOrder);
4218 }
4219 }
4220
4221 //------------------------inline_native_clone----------------------------
4222 // protected native Object java.lang.Object.clone();
4223 //
4224 // Here are the simple edge cases:
4225 // null receiver => normal trap
4226 // virtual and clone was overridden => slow path to out-of-line clone
4227 // not cloneable or finalizer => slow path to out-of-line Object.clone
4228 //
4229 // The general case has two steps, allocation and copying.
4230 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4231 //
4232 // Copying also has two cases, oop arrays and everything else.
4233 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4234 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4235 //
4236 // These steps fold up nicely if and when the cloned object's klass
4237 // can be sharply typed as an object array, a type array, or an instance.
4238 //
4239 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4240 PhiNode* result_val;
4241
4242 // Set the reexecute bit for the interpreter to reexecute
4243 // the bytecode that invokes Object.clone if deoptimization happens.
4244 { PreserveReexecuteState preexecs(this);
4245 jvms()->set_should_reexecute(true);
4246
4247 Node* obj = null_check_receiver();
4248 if (stopped()) return true;
4249
4250 Node* obj_klass = load_object_klass(obj);
4251 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4252 const TypeOopPtr* toop = ((tklass != NULL)
4253 ? tklass->as_instance_type()
4254 : TypeInstPtr::NOTNULL);
4255
4256 // Conservatively insert a memory barrier on all memory slices.
4257 // Do not let writes into the original float below the clone.
4258 insert_mem_bar(Op_MemBarCPUOrder);
4259
4260 // paths into result_reg:
4261 enum {
4262 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4263 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4264 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4265 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4266 PATH_LIMIT
4267 };
4268 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4269 result_val = new(C) PhiNode(result_reg,
4270 TypeInstPtr::NOTNULL);
4271 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4272 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4273 TypePtr::BOTTOM);
4274 record_for_igvn(result_reg);
4275
4276 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4277 int raw_adr_idx = Compile::AliasIdxRaw;
4278
4279 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4280 if (array_ctl != NULL) {
4281 // It's an array.
4282 PreserveJVMState pjvms(this);
4283 set_control(array_ctl);
4284 Node* obj_length = load_array_length(obj);
4285 Node* obj_size = NULL;
4286 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4287
4288 if (!use_ReduceInitialCardMarks()) {
4289 // If it is an oop array, it requires very special treatment,
4290 // because card marking is required on each card of the array.
4291 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4292 if (is_obja != NULL) {
4293 PreserveJVMState pjvms2(this);
4294 set_control(is_obja);
4295 // Generate a direct call to the right arraycopy function(s).
4296 bool disjoint_bases = true;
4297 bool length_never_negative = true;
4298 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4299 obj, intcon(0), alloc_obj, intcon(0),
4300 obj_length,
4301 disjoint_bases, length_never_negative);
4302 result_reg->init_req(_objArray_path, control());
4303 result_val->init_req(_objArray_path, alloc_obj);
4304 result_i_o ->set_req(_objArray_path, i_o());
4305 result_mem ->set_req(_objArray_path, reset_memory());
4306 }
4307 }
4308 // Otherwise, there are no card marks to worry about.
4309 // (We can dispense with card marks if we know the allocation
4310 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4311 // causes the non-eden paths to take compensating steps to
4312 // simulate a fresh allocation, so that no further
4313 // card marks are required in compiled code to initialize
4314 // the object.)
4315
4316 if (!stopped()) {
4317 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4318
4319 // Present the results of the copy.
4320 result_reg->init_req(_array_path, control());
4321 result_val->init_req(_array_path, alloc_obj);
4322 result_i_o ->set_req(_array_path, i_o());
4323 result_mem ->set_req(_array_path, reset_memory());
4324 }
4325 }
4326
4327 // We only go to the instance fast case code if we pass a number of guards.
4328 // The paths which do not pass are accumulated in the slow_region.
4329 RegionNode* slow_region = new (C) RegionNode(1);
4330 record_for_igvn(slow_region);
4331 if (!stopped()) {
4332 // It's an instance (we did array above). Make the slow-path tests.
4333 // If this is a virtual call, we generate a funny guard. We grab
4334 // the vtable entry corresponding to clone() from the target object.
4335 // If the target method which we are calling happens to be the
4336 // Object clone() method, we pass the guard. We do not need this
4337 // guard for non-virtual calls; the caller is known to be the native
4338 // Object clone().
4339 if (is_virtual) {
4340 generate_virtual_guard(obj_klass, slow_region);
4341 }
4342
4343 // The object must be cloneable and must not have a finalizer.
4344 // Both of these conditions may be checked in a single test.
4345 // We could optimize the cloneable test further, but we don't care.
4346 generate_access_flags_guard(obj_klass,
4347 // Test both conditions:
4348 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4349 // Must be cloneable but not finalizer:
4350 JVM_ACC_IS_CLONEABLE,
4351 slow_region);
4352 }
4353
4354 if (!stopped()) {
4355 // It's an instance, and it passed the slow-path tests.
4356 PreserveJVMState pjvms(this);
4357 Node* obj_size = NULL;
4358 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4359
4360 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4361
4362 // Present the results of the slow call.
4363 result_reg->init_req(_instance_path, control());
4364 result_val->init_req(_instance_path, alloc_obj);
4365 result_i_o ->set_req(_instance_path, i_o());
4366 result_mem ->set_req(_instance_path, reset_memory());
4367 }
4368
4369 // Generate code for the slow case. We make a call to clone().
4370 set_control(_gvn.transform(slow_region));
4371 if (!stopped()) {
4372 PreserveJVMState pjvms(this);
4373 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4374 Node* slow_result = set_results_for_java_call(slow_call);
4375 // this->control() comes from set_results_for_java_call
4376 result_reg->init_req(_slow_path, control());
4377 result_val->init_req(_slow_path, slow_result);
4378 result_i_o ->set_req(_slow_path, i_o());
4379 result_mem ->set_req(_slow_path, reset_memory());
4380 }
4381
4382 // Return the combined state.
4383 set_control( _gvn.transform(result_reg) );
4384 set_i_o( _gvn.transform(result_i_o) );
4385 set_all_memory( _gvn.transform(result_mem) );
4386 } // original reexecute is set back here
4387
4388 set_result(_gvn.transform(result_val));
4389 return true;
4390 }
4391
4392 //------------------------------basictype2arraycopy----------------------------
4393 address LibraryCallKit::basictype2arraycopy(BasicType t,
4394 Node* src_offset,
4395 Node* dest_offset,
4396 bool disjoint_bases,
4397 const char* &name,
4398 bool dest_uninitialized) {
4399 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4400 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4401
4402 bool aligned = false;
4403 bool disjoint = disjoint_bases;
4404
4405 // if the offsets are the same, we can treat the memory regions as
4406 // disjoint, because either the memory regions are in different arrays,
4407 // or they are identical (which we can treat as disjoint.) We can also
4408 // treat a copy with a destination index less that the source index
4409 // as disjoint since a low->high copy will work correctly in this case.
4410 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4411 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4412 // both indices are constants
4413 int s_offs = src_offset_inttype->get_con();
4414 int d_offs = dest_offset_inttype->get_con();
4415 int element_size = type2aelembytes(t);
4416 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4417 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4418 if (s_offs >= d_offs) disjoint = true;
4419 } else if (src_offset == dest_offset && src_offset != NULL) {
4420 // This can occur if the offsets are identical non-constants.
4421 disjoint = true;
4422 }
4423
4424 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4425 }
4426
4427
4428 //------------------------------inline_arraycopy-----------------------
4429 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4430 // Object dest, int destPos,
4431 // int length);
4432 bool LibraryCallKit::inline_arraycopy() {
4433 // Get the arguments.
4434 Node* src = argument(0); // type: oop
4435 Node* src_offset = argument(1); // type: int
4436 Node* dest = argument(2); // type: oop
4437 Node* dest_offset = argument(3); // type: int
4438 Node* length = argument(4); // type: int
4439
4440 // Compile time checks. If any of these checks cannot be verified at compile time,
4441 // we do not make a fast path for this call. Instead, we let the call remain as it
4442 // is. The checks we choose to mandate at compile time are:
4443 //
4444 // (1) src and dest are arrays.
4445 const Type* src_type = src->Value(&_gvn);
4446 const Type* dest_type = dest->Value(&_gvn);
4447 const TypeAryPtr* top_src = src_type->isa_aryptr();
4448 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4449 if (top_src == NULL || top_src->klass() == NULL ||
4450 top_dest == NULL || top_dest->klass() == NULL) {
4451 // Conservatively insert a memory barrier on all memory slices.
4452 // Do not let writes into the source float below the arraycopy.
4453 insert_mem_bar(Op_MemBarCPUOrder);
4454
4455 // Call StubRoutines::generic_arraycopy stub.
4456 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4457 src, src_offset, dest, dest_offset, length);
4458
4459 // Do not let reads from the destination float above the arraycopy.
4460 // Since we cannot type the arrays, we don't know which slices
4461 // might be affected. We could restrict this barrier only to those
4462 // memory slices which pertain to array elements--but don't bother.
4463 if (!InsertMemBarAfterArraycopy)
4464 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4465 insert_mem_bar(Op_MemBarCPUOrder);
4466 return true;
4467 }
4468
4469 // (2) src and dest arrays must have elements of the same BasicType
4470 // Figure out the size and type of the elements we will be copying.
4471 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4472 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4473 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4474 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4475
4476 if (src_elem != dest_elem || dest_elem == T_VOID) {
4477 // The component types are not the same or are not recognized. Punt.
4478 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4479 generate_slow_arraycopy(TypePtr::BOTTOM,
4480 src, src_offset, dest, dest_offset, length,
4481 /*dest_uninitialized*/false);
4482 return true;
4483 }
4484
4485 //---------------------------------------------------------------------------
4486 // We will make a fast path for this call to arraycopy.
4487
4488 // We have the following tests left to perform:
4489 //
4490 // (3) src and dest must not be null.
4491 // (4) src_offset must not be negative.
4492 // (5) dest_offset must not be negative.
4493 // (6) length must not be negative.
4494 // (7) src_offset + length must not exceed length of src.
4495 // (8) dest_offset + length must not exceed length of dest.
4496 // (9) each element of an oop array must be assignable
4497
4498 RegionNode* slow_region = new (C) RegionNode(1);
4499 record_for_igvn(slow_region);
4500
4501 // (3) operands must not be null
4502 // We currently perform our null checks with the null_check routine.
4503 // This means that the null exceptions will be reported in the caller
4504 // rather than (correctly) reported inside of the native arraycopy call.
4505 // This should be corrected, given time. We do our null check with the
4506 // stack pointer restored.
4507 src = null_check(src, T_ARRAY);
4508 dest = null_check(dest, T_ARRAY);
4509
4510 // (4) src_offset must not be negative.
4511 generate_negative_guard(src_offset, slow_region);
4512
4513 // (5) dest_offset must not be negative.
4514 generate_negative_guard(dest_offset, slow_region);
4515
4516 // (6) length must not be negative (moved to generate_arraycopy()).
4517 // generate_negative_guard(length, slow_region);
4518
4519 // (7) src_offset + length must not exceed length of src.
4520 generate_limit_guard(src_offset, length,
4521 load_array_length(src),
4522 slow_region);
4523
4524 // (8) dest_offset + length must not exceed length of dest.
4525 generate_limit_guard(dest_offset, length,
4526 load_array_length(dest),
4527 slow_region);
4528
4529 // (9) each element of an oop array must be assignable
4530 // The generate_arraycopy subroutine checks this.
4531
4532 // This is where the memory effects are placed:
4533 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4534 generate_arraycopy(adr_type, dest_elem,
4535 src, src_offset, dest, dest_offset, length,
4536 false, false, slow_region);
4537
4538 return true;
4539 }
4540
4541 //-----------------------------generate_arraycopy----------------------
4542 // Generate an optimized call to arraycopy.
4543 // Caller must guard against non-arrays.
4544 // Caller must determine a common array basic-type for both arrays.
4545 // Caller must validate offsets against array bounds.
4546 // The slow_region has already collected guard failure paths
4547 // (such as out of bounds length or non-conformable array types).
4548 // The generated code has this shape, in general:
4549 //
4550 // if (length == 0) return // via zero_path
4551 // slowval = -1
4552 // if (types unknown) {
4553 // slowval = call generic copy loop
4554 // if (slowval == 0) return // via checked_path
4555 // } else if (indexes in bounds) {
4556 // if ((is object array) && !(array type check)) {
4557 // slowval = call checked copy loop
4558 // if (slowval == 0) return // via checked_path
4559 // } else {
4560 // call bulk copy loop
4561 // return // via fast_path
4562 // }
4563 // }
4564 // // adjust params for remaining work:
4565 // if (slowval != -1) {
4566 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4567 // }
4568 // slow_region:
4569 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4570 // return // via slow_call_path
4571 //
4572 // This routine is used from several intrinsics: System.arraycopy,
4573 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4574 //
4575 void
4576 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4577 BasicType basic_elem_type,
4578 Node* src, Node* src_offset,
4579 Node* dest, Node* dest_offset,
4580 Node* copy_length,
4581 bool disjoint_bases,
4582 bool length_never_negative,
4583 RegionNode* slow_region) {
4584
4585 if (slow_region == NULL) {
4586 slow_region = new(C) RegionNode(1);
4587 record_for_igvn(slow_region);
4588 }
4589
4590 Node* original_dest = dest;
4591 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4592 bool dest_uninitialized = false;
4593
4594 // See if this is the initialization of a newly-allocated array.
4595 // If so, we will take responsibility here for initializing it to zero.
4596 // (Note: Because tightly_coupled_allocation performs checks on the
4597 // out-edges of the dest, we need to avoid making derived pointers
4598 // from it until we have checked its uses.)
4599 if (ReduceBulkZeroing
4600 && !ZeroTLAB // pointless if already zeroed
4601 && basic_elem_type != T_CONFLICT // avoid corner case
4602 && !src->eqv_uncast(dest)
4603 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4604 != NULL)
4605 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4606 && alloc->maybe_set_complete(&_gvn)) {
4607 // "You break it, you buy it."
4608 InitializeNode* init = alloc->initialization();
4609 assert(init->is_complete(), "we just did this");
4610 init->set_complete_with_arraycopy();
4611 assert(dest->is_CheckCastPP(), "sanity");
4612 assert(dest->in(0)->in(0) == init, "dest pinned");
4613 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4614 // From this point on, every exit path is responsible for
4615 // initializing any non-copied parts of the object to zero.
4616 // Also, if this flag is set we make sure that arraycopy interacts properly
4617 // with G1, eliding pre-barriers. See CR 6627983.
4618 dest_uninitialized = true;
4619 } else {
4620 // No zeroing elimination here.
4621 alloc = NULL;
4622 //original_dest = dest;
4623 //dest_uninitialized = false;
4624 }
4625
4626 // Results are placed here:
4627 enum { fast_path = 1, // normal void-returning assembly stub
4628 checked_path = 2, // special assembly stub with cleanup
4629 slow_call_path = 3, // something went wrong; call the VM
4630 zero_path = 4, // bypass when length of copy is zero
4631 bcopy_path = 5, // copy primitive array by 64-bit blocks
4632 PATH_LIMIT = 6
4633 };
4634 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4635 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4636 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4637 record_for_igvn(result_region);
4638 _gvn.set_type_bottom(result_i_o);
4639 _gvn.set_type_bottom(result_memory);
4640 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4641
4642 // The slow_control path:
4643 Node* slow_control;
4644 Node* slow_i_o = i_o();
4645 Node* slow_mem = memory(adr_type);
4646 debug_only(slow_control = (Node*) badAddress);
4647
4648 // Checked control path:
4649 Node* checked_control = top();
4650 Node* checked_mem = NULL;
4651 Node* checked_i_o = NULL;
4652 Node* checked_value = NULL;
4653
4654 if (basic_elem_type == T_CONFLICT) {
4655 assert(!dest_uninitialized, "");
4656 Node* cv = generate_generic_arraycopy(adr_type,
4657 src, src_offset, dest, dest_offset,
4658 copy_length, dest_uninitialized);
4659 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4660 checked_control = control();
4661 checked_i_o = i_o();
4662 checked_mem = memory(adr_type);
4663 checked_value = cv;
4664 set_control(top()); // no fast path
4665 }
4666
4667 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4668 if (not_pos != NULL) {
4669 PreserveJVMState pjvms(this);
4670 set_control(not_pos);
4671
4672 // (6) length must not be negative.
4673 if (!length_never_negative) {
4674 generate_negative_guard(copy_length, slow_region);
4675 }
4676
4677 // copy_length is 0.
4678 if (!stopped() && dest_uninitialized) {
4679 Node* dest_length = alloc->in(AllocateNode::ALength);
4680 if (copy_length->eqv_uncast(dest_length)
4681 || _gvn.find_int_con(dest_length, 1) <= 0) {
4682 // There is no zeroing to do. No need for a secondary raw memory barrier.
4683 } else {
4684 // Clear the whole thing since there are no source elements to copy.
4685 generate_clear_array(adr_type, dest, basic_elem_type,
4686 intcon(0), NULL,
4687 alloc->in(AllocateNode::AllocSize));
4688 // Use a secondary InitializeNode as raw memory barrier.
4689 // Currently it is needed only on this path since other
4690 // paths have stub or runtime calls as raw memory barriers.
4691 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4692 Compile::AliasIdxRaw,
4693 top())->as_Initialize();
4694 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4695 }
4696 }
4697
4698 // Present the results of the fast call.
4699 result_region->init_req(zero_path, control());
4700 result_i_o ->init_req(zero_path, i_o());
4701 result_memory->init_req(zero_path, memory(adr_type));
4702 }
4703
4704 if (!stopped() && dest_uninitialized) {
4705 // We have to initialize the *uncopied* part of the array to zero.
4706 // The copy destination is the slice dest[off..off+len]. The other slices
4707 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4708 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4709 Node* dest_length = alloc->in(AllocateNode::ALength);
4710 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4711 copy_length) );
4712
4713 // If there is a head section that needs zeroing, do it now.
4714 if (find_int_con(dest_offset, -1) != 0) {
4715 generate_clear_array(adr_type, dest, basic_elem_type,
4716 intcon(0), dest_offset,
4717 NULL);
4718 }
4719
4720 // Next, perform a dynamic check on the tail length.
4721 // It is often zero, and we can win big if we prove this.
4722 // There are two wins: Avoid generating the ClearArray
4723 // with its attendant messy index arithmetic, and upgrade
4724 // the copy to a more hardware-friendly word size of 64 bits.
4725 Node* tail_ctl = NULL;
4726 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4727 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4728 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4729 tail_ctl = generate_slow_guard(bol_lt, NULL);
4730 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4731 }
4732
4733 // At this point, let's assume there is no tail.
4734 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4735 // There is no tail. Try an upgrade to a 64-bit copy.
4736 bool didit = false;
4737 { PreserveJVMState pjvms(this);
4738 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4739 src, src_offset, dest, dest_offset,
4740 dest_size, dest_uninitialized);
4741 if (didit) {
4742 // Present the results of the block-copying fast call.
4743 result_region->init_req(bcopy_path, control());
4744 result_i_o ->init_req(bcopy_path, i_o());
4745 result_memory->init_req(bcopy_path, memory(adr_type));
4746 }
4747 }
4748 if (didit)
4749 set_control(top()); // no regular fast path
4750 }
4751
4752 // Clear the tail, if any.
4753 if (tail_ctl != NULL) {
4754 Node* notail_ctl = stopped() ? NULL : control();
4755 set_control(tail_ctl);
4756 if (notail_ctl == NULL) {
4757 generate_clear_array(adr_type, dest, basic_elem_type,
4758 dest_tail, NULL,
4759 dest_size);
4760 } else {
4761 // Make a local merge.
4762 Node* done_ctl = new(C) RegionNode(3);
4763 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4764 done_ctl->init_req(1, notail_ctl);
4765 done_mem->init_req(1, memory(adr_type));
4766 generate_clear_array(adr_type, dest, basic_elem_type,
4767 dest_tail, NULL,
4768 dest_size);
4769 done_ctl->init_req(2, control());
4770 done_mem->init_req(2, memory(adr_type));
4771 set_control( _gvn.transform(done_ctl) );
4772 set_memory( _gvn.transform(done_mem), adr_type );
4773 }
4774 }
4775 }
4776
4777 BasicType copy_type = basic_elem_type;
4778 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4779 if (!stopped() && copy_type == T_OBJECT) {
4780 // If src and dest have compatible element types, we can copy bits.
4781 // Types S[] and D[] are compatible if D is a supertype of S.
4782 //
4783 // If they are not, we will use checked_oop_disjoint_arraycopy,
4784 // which performs a fast optimistic per-oop check, and backs off
4785 // further to JVM_ArrayCopy on the first per-oop check that fails.
4786 // (Actually, we don't move raw bits only; the GC requires card marks.)
4787
4788 // Get the klassOop for both src and dest
4789 Node* src_klass = load_object_klass(src);
4790 Node* dest_klass = load_object_klass(dest);
4791
4792 // Generate the subtype check.
4793 // This might fold up statically, or then again it might not.
4794 //
4795 // Non-static example: Copying List<String>.elements to a new String[].
4796 // The backing store for a List<String> is always an Object[],
4797 // but its elements are always type String, if the generic types
4798 // are correct at the source level.
4799 //
4800 // Test S[] against D[], not S against D, because (probably)
4801 // the secondary supertype cache is less busy for S[] than S.
4802 // This usually only matters when D is an interface.
4803 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4804 // Plug failing path into checked_oop_disjoint_arraycopy
4805 if (not_subtype_ctrl != top()) {
4806 PreserveJVMState pjvms(this);
4807 set_control(not_subtype_ctrl);
4808 // (At this point we can assume disjoint_bases, since types differ.)
4809 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4810 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4811 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4812 Node* dest_elem_klass = _gvn.transform(n1);
4813 Node* cv = generate_checkcast_arraycopy(adr_type,
4814 dest_elem_klass,
4815 src, src_offset, dest, dest_offset,
4816 ConvI2X(copy_length), dest_uninitialized);
4817 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4818 checked_control = control();
4819 checked_i_o = i_o();
4820 checked_mem = memory(adr_type);
4821 checked_value = cv;
4822 }
4823 // At this point we know we do not need type checks on oop stores.
4824
4825 // Let's see if we need card marks:
4826 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4827 // If we do not need card marks, copy using the jint or jlong stub.
4828 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4829 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4830 "sizes agree");
4831 }
4832 }
4833
4834 if (!stopped()) {
4835 // Generate the fast path, if possible.
4836 PreserveJVMState pjvms(this);
4837 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4838 src, src_offset, dest, dest_offset,
4839 ConvI2X(copy_length), dest_uninitialized);
4840
4841 // Present the results of the fast call.
4842 result_region->init_req(fast_path, control());
4843 result_i_o ->init_req(fast_path, i_o());
4844 result_memory->init_req(fast_path, memory(adr_type));
4845 }
4846
4847 // Here are all the slow paths up to this point, in one bundle:
4848 slow_control = top();
4849 if (slow_region != NULL)
4850 slow_control = _gvn.transform(slow_region);
4851 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4852
4853 set_control(checked_control);
4854 if (!stopped()) {
4855 // Clean up after the checked call.
4856 // The returned value is either 0 or -1^K,
4857 // where K = number of partially transferred array elements.
4858 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4859 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4860 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4861
4862 // If it is 0, we are done, so transfer to the end.
4863 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4864 result_region->init_req(checked_path, checks_done);
4865 result_i_o ->init_req(checked_path, checked_i_o);
4866 result_memory->init_req(checked_path, checked_mem);
4867
4868 // If it is not zero, merge into the slow call.
4869 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4870 RegionNode* slow_reg2 = new(C) RegionNode(3);
4871 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4872 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4873 record_for_igvn(slow_reg2);
4874 slow_reg2 ->init_req(1, slow_control);
4875 slow_i_o2 ->init_req(1, slow_i_o);
4876 slow_mem2 ->init_req(1, slow_mem);
4877 slow_reg2 ->init_req(2, control());
4878 slow_i_o2 ->init_req(2, checked_i_o);
4879 slow_mem2 ->init_req(2, checked_mem);
4880
4881 slow_control = _gvn.transform(slow_reg2);
4882 slow_i_o = _gvn.transform(slow_i_o2);
4883 slow_mem = _gvn.transform(slow_mem2);
4884
4885 if (alloc != NULL) {
4886 // We'll restart from the very beginning, after zeroing the whole thing.
4887 // This can cause double writes, but that's OK since dest is brand new.
4888 // So we ignore the low 31 bits of the value returned from the stub.
4889 } else {
4890 // We must continue the copy exactly where it failed, or else
4891 // another thread might see the wrong number of writes to dest.
4892 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4893 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4894 slow_offset->init_req(1, intcon(0));
4895 slow_offset->init_req(2, checked_offset);
4896 slow_offset = _gvn.transform(slow_offset);
4897
4898 // Adjust the arguments by the conditionally incoming offset.
4899 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4900 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4901 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4902
4903 // Tweak the node variables to adjust the code produced below:
4904 src_offset = src_off_plus;
4905 dest_offset = dest_off_plus;
4906 copy_length = length_minus;
4907 }
4908 }
4909
4910 set_control(slow_control);
4911 if (!stopped()) {
4912 // Generate the slow path, if needed.
4913 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4914
4915 set_memory(slow_mem, adr_type);
4916 set_i_o(slow_i_o);
4917
4918 if (dest_uninitialized) {
4919 generate_clear_array(adr_type, dest, basic_elem_type,
4920 intcon(0), NULL,
4921 alloc->in(AllocateNode::AllocSize));
4922 }
4923
4924 generate_slow_arraycopy(adr_type,
4925 src, src_offset, dest, dest_offset,
4926 copy_length, /*dest_uninitialized*/false);
4927
4928 result_region->init_req(slow_call_path, control());
4929 result_i_o ->init_req(slow_call_path, i_o());
4930 result_memory->init_req(slow_call_path, memory(adr_type));
4931 }
4932
4933 // Remove unused edges.
4934 for (uint i = 1; i < result_region->req(); i++) {
4935 if (result_region->in(i) == NULL)
4936 result_region->init_req(i, top());
4937 }
4938
4939 // Finished; return the combined state.
4940 set_control( _gvn.transform(result_region) );
4941 set_i_o( _gvn.transform(result_i_o) );
4942 set_memory( _gvn.transform(result_memory), adr_type );
4943
4944 // The memory edges above are precise in order to model effects around
4945 // array copies accurately to allow value numbering of field loads around
4946 // arraycopy. Such field loads, both before and after, are common in Java
4947 // collections and similar classes involving header/array data structures.
4948 //
4949 // But with low number of register or when some registers are used or killed
4950 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4951 // The next memory barrier is added to avoid it. If the arraycopy can be
4952 // optimized away (which it can, sometimes) then we can manually remove
4953 // the membar also.
4954 //
4955 // Do not let reads from the cloned object float above the arraycopy.
4956 if (alloc != NULL) {
4957 // Do not let stores that initialize this object be reordered with
4958 // a subsequent store that would make this object accessible by
4959 // other threads.
4960 // Record what AllocateNode this StoreStore protects so that
4961 // escape analysis can go from the MemBarStoreStoreNode to the
4962 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4963 // based on the escape status of the AllocateNode.
4964 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4965 } else if (InsertMemBarAfterArraycopy)
4966 insert_mem_bar(Op_MemBarCPUOrder);
4967 }
4968
4969
4970 // Helper function which determines if an arraycopy immediately follows
4971 // an allocation, with no intervening tests or other escapes for the object.
4972 AllocateArrayNode*
4973 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4974 RegionNode* slow_region) {
4975 if (stopped()) return NULL; // no fast path
4976 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4977
4978 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4979 if (alloc == NULL) return NULL;
4980
4981 Node* rawmem = memory(Compile::AliasIdxRaw);
4982 // Is the allocation's memory state untouched?
4983 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4984 // Bail out if there have been raw-memory effects since the allocation.
4985 // (Example: There might have been a call or safepoint.)
4986 return NULL;
4987 }
4988 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4989 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4990 return NULL;
4991 }
4992
4993 // There must be no unexpected observers of this allocation.
4994 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4995 Node* obs = ptr->fast_out(i);
4996 if (obs != this->map()) {
4997 return NULL;
4998 }
4999 }
5000
5001 // This arraycopy must unconditionally follow the allocation of the ptr.
5002 Node* alloc_ctl = ptr->in(0);
5003 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5004
5005 Node* ctl = control();
5006 while (ctl != alloc_ctl) {
5007 // There may be guards which feed into the slow_region.
5008 // Any other control flow means that we might not get a chance
5009 // to finish initializing the allocated object.
5010 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5011 IfNode* iff = ctl->in(0)->as_If();
5012 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5013 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5014 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5015 ctl = iff->in(0); // This test feeds the known slow_region.
5016 continue;
5017 }
5018 // One more try: Various low-level checks bottom out in
5019 // uncommon traps. If the debug-info of the trap omits
5020 // any reference to the allocation, as we've already
5021 // observed, then there can be no objection to the trap.
5022 bool found_trap = false;
5023 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5024 Node* obs = not_ctl->fast_out(j);
5025 if (obs->in(0) == not_ctl && obs->is_Call() &&
5026 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5027 found_trap = true; break;
5028 }
5029 }
5030 if (found_trap) {
5031 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5032 continue;
5033 }
5034 }
5035 return NULL;
5036 }
5037
5038 // If we get this far, we have an allocation which immediately
5039 // precedes the arraycopy, and we can take over zeroing the new object.
5040 // The arraycopy will finish the initialization, and provide
5041 // a new control state to which we will anchor the destination pointer.
5042
5043 return alloc;
5044 }
5045
5046 // Helper for initialization of arrays, creating a ClearArray.
5047 // It writes zero bits in [start..end), within the body of an array object.
5048 // The memory effects are all chained onto the 'adr_type' alias category.
5049 //
5050 // Since the object is otherwise uninitialized, we are free
5051 // to put a little "slop" around the edges of the cleared area,
5052 // as long as it does not go back into the array's header,
5053 // or beyond the array end within the heap.
5054 //
5055 // The lower edge can be rounded down to the nearest jint and the
5056 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5057 //
5058 // Arguments:
5059 // adr_type memory slice where writes are generated
5060 // dest oop of the destination array
5061 // basic_elem_type element type of the destination
5062 // slice_idx array index of first element to store
5063 // slice_len number of elements to store (or NULL)
5064 // dest_size total size in bytes of the array object
5065 //
5066 // Exactly one of slice_len or dest_size must be non-NULL.
5067 // If dest_size is non-NULL, zeroing extends to the end of the object.
5068 // If slice_len is non-NULL, the slice_idx value must be a constant.
5069 void
5070 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5071 Node* dest,
5072 BasicType basic_elem_type,
5073 Node* slice_idx,
5074 Node* slice_len,
5075 Node* dest_size) {
5076 // one or the other but not both of slice_len and dest_size:
5077 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5078 if (slice_len == NULL) slice_len = top();
5079 if (dest_size == NULL) dest_size = top();
5080
5081 // operate on this memory slice:
5082 Node* mem = memory(adr_type); // memory slice to operate on
5083
5084 // scaling and rounding of indexes:
5085 int scale = exact_log2(type2aelembytes(basic_elem_type));
5086 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5087 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5088 int bump_bit = (-1 << scale) & BytesPerInt;
5089
5090 // determine constant starts and ends
5091 const intptr_t BIG_NEG = -128;
5092 assert(BIG_NEG + 2*abase < 0, "neg enough");
5093 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5094 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5095 if (slice_len_con == 0) {
5096 return; // nothing to do here
5097 }
5098 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5099 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5100 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5101 assert(end_con < 0, "not two cons");
5102 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5103 BytesPerLong);
5104 }
5105
5106 if (start_con >= 0 && end_con >= 0) {
5107 // Constant start and end. Simple.
5108 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5109 start_con, end_con, &_gvn);
5110 } else if (start_con >= 0 && dest_size != top()) {
5111 // Constant start, pre-rounded end after the tail of the array.
5112 Node* end = dest_size;
5113 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5114 start_con, end, &_gvn);
5115 } else if (start_con >= 0 && slice_len != top()) {
5116 // Constant start, non-constant end. End needs rounding up.
5117 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5118 intptr_t end_base = abase + (slice_idx_con << scale);
5119 int end_round = (-1 << scale) & (BytesPerLong - 1);
5120 Node* end = ConvI2X(slice_len);
5121 if (scale != 0)
5122 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5123 end_base += end_round;
5124 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5125 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5126 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5127 start_con, end, &_gvn);
5128 } else if (start_con < 0 && dest_size != top()) {
5129 // Non-constant start, pre-rounded end after the tail of the array.
5130 // This is almost certainly a "round-to-end" operation.
5131 Node* start = slice_idx;
5132 start = ConvI2X(start);
5133 if (scale != 0)
5134 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5135 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5136 if ((bump_bit | clear_low) != 0) {
5137 int to_clear = (bump_bit | clear_low);
5138 // Align up mod 8, then store a jint zero unconditionally
5139 // just before the mod-8 boundary.
5140 if (((abase + bump_bit) & ~to_clear) - bump_bit
5141 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5142 bump_bit = 0;
5143 assert((abase & to_clear) == 0, "array base must be long-aligned");
5144 } else {
5145 // Bump 'start' up to (or past) the next jint boundary:
5146 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5147 assert((abase & clear_low) == 0, "array base must be int-aligned");
5148 }
5149 // Round bumped 'start' down to jlong boundary in body of array.
5150 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5151 if (bump_bit != 0) {
5152 // Store a zero to the immediately preceding jint:
5153 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5154 Node* p1 = basic_plus_adr(dest, x1);
5155 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5156 mem = _gvn.transform(mem);
5157 }
5158 }
5159 Node* end = dest_size; // pre-rounded
5160 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5161 start, end, &_gvn);
5162 } else {
5163 // Non-constant start, unrounded non-constant end.
5164 // (Nobody zeroes a random midsection of an array using this routine.)
5165 ShouldNotReachHere(); // fix caller
5166 }
5167
5168 // Done.
5169 set_memory(mem, adr_type);
5170 }
5171
5172
5173 bool
5174 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5175 BasicType basic_elem_type,
5176 AllocateNode* alloc,
5177 Node* src, Node* src_offset,
5178 Node* dest, Node* dest_offset,
5179 Node* dest_size, bool dest_uninitialized) {
5180 // See if there is an advantage from block transfer.
5181 int scale = exact_log2(type2aelembytes(basic_elem_type));
5182 if (scale >= LogBytesPerLong)
5183 return false; // it is already a block transfer
5184
5185 // Look at the alignment of the starting offsets.
5186 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5187
5188 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5189 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5190 if (src_off_con < 0 || dest_off_con < 0)
5191 // At present, we can only understand constants.
5192 return false;
5193
5194 intptr_t src_off = abase + (src_off_con << scale);
5195 intptr_t dest_off = abase + (dest_off_con << scale);
5196
5197 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5198 // Non-aligned; too bad.
5199 // One more chance: Pick off an initial 32-bit word.
5200 // This is a common case, since abase can be odd mod 8.
5201 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5202 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5203 Node* sptr = basic_plus_adr(src, src_off);
5204 Node* dptr = basic_plus_adr(dest, dest_off);
5205 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5206 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5207 src_off += BytesPerInt;
5208 dest_off += BytesPerInt;
5209 } else {
5210 return false;
5211 }
5212 }
5213 assert(src_off % BytesPerLong == 0, "");
5214 assert(dest_off % BytesPerLong == 0, "");
5215
5216 // Do this copy by giant steps.
5217 Node* sptr = basic_plus_adr(src, src_off);
5218 Node* dptr = basic_plus_adr(dest, dest_off);
5219 Node* countx = dest_size;
5220 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5221 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5222
5223 bool disjoint_bases = true; // since alloc != NULL
5224 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5225 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5226
5227 return true;
5228 }
5229
5230
5231 // Helper function; generates code for the slow case.
5232 // We make a call to a runtime method which emulates the native method,
5233 // but without the native wrapper overhead.
5234 void
5235 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5236 Node* src, Node* src_offset,
5237 Node* dest, Node* dest_offset,
5238 Node* copy_length, bool dest_uninitialized) {
5239 assert(!dest_uninitialized, "Invariant");
5240 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5241 OptoRuntime::slow_arraycopy_Type(),
5242 OptoRuntime::slow_arraycopy_Java(),
5243 "slow_arraycopy", adr_type,
5244 src, src_offset, dest, dest_offset,
5245 copy_length);
5246
5247 // Handle exceptions thrown by this fellow:
5248 make_slow_call_ex(call, env()->Throwable_klass(), false);
5249 }
5250
5251 // Helper function; generates code for cases requiring runtime checks.
5252 Node*
5253 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5254 Node* dest_elem_klass,
5255 Node* src, Node* src_offset,
5256 Node* dest, Node* dest_offset,
5257 Node* copy_length, bool dest_uninitialized) {
5258 if (stopped()) return NULL;
5259
5260 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5261 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5262 return NULL;
5263 }
5264
5265 // Pick out the parameters required to perform a store-check
5266 // for the target array. This is an optimistic check. It will
5267 // look in each non-null element's class, at the desired klass's
5268 // super_check_offset, for the desired klass.
5269 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5270 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5271 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5272 Node* check_offset = ConvI2X(_gvn.transform(n3));
5273 Node* check_value = dest_elem_klass;
5274
5275 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5276 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5277
5278 // (We know the arrays are never conjoint, because their types differ.)
5279 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5280 OptoRuntime::checkcast_arraycopy_Type(),
5281 copyfunc_addr, "checkcast_arraycopy", adr_type,
5282 // five arguments, of which two are
5283 // intptr_t (jlong in LP64)
5284 src_start, dest_start,
5285 copy_length XTOP,
5286 check_offset XTOP,
5287 check_value);
5288
5289 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5290 }
5291
5292
5293 // Helper function; generates code for cases requiring runtime checks.
5294 Node*
5295 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5296 Node* src, Node* src_offset,
5297 Node* dest, Node* dest_offset,
5298 Node* copy_length, bool dest_uninitialized) {
5299 assert(!dest_uninitialized, "Invariant");
5300 if (stopped()) return NULL;
5301 address copyfunc_addr = StubRoutines::generic_arraycopy();
5302 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5303 return NULL;
5304 }
5305
5306 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5307 OptoRuntime::generic_arraycopy_Type(),
5308 copyfunc_addr, "generic_arraycopy", adr_type,
5309 src, src_offset, dest, dest_offset, copy_length);
5310
5311 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5312 }
5313
5314 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5315 void
5316 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5317 BasicType basic_elem_type,
5318 bool disjoint_bases,
5319 Node* src, Node* src_offset,
5320 Node* dest, Node* dest_offset,
5321 Node* copy_length, bool dest_uninitialized) {
5322 if (stopped()) return; // nothing to do
5323
5324 Node* src_start = src;
5325 Node* dest_start = dest;
5326 if (src_offset != NULL || dest_offset != NULL) {
5327 assert(src_offset != NULL && dest_offset != NULL, "");
5328 src_start = array_element_address(src, src_offset, basic_elem_type);
5329 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5330 }
5331
5332 // Figure out which arraycopy runtime method to call.
5333 const char* copyfunc_name = "arraycopy";
5334 address copyfunc_addr =
5335 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5336 disjoint_bases, copyfunc_name, dest_uninitialized);
5337
5338 // Call it. Note that the count_ix value is not scaled to a byte-size.
5339 make_runtime_call(RC_LEAF|RC_NO_FP,
5340 OptoRuntime::fast_arraycopy_Type(),
5341 copyfunc_addr, copyfunc_name, adr_type,
5342 src_start, dest_start, copy_length XTOP);
5343 }
5344
5345 //----------------------------inline_reference_get----------------------------
5346 // public T java.lang.ref.Reference.get();
5347 bool LibraryCallKit::inline_reference_get() {
5348 const int referent_offset = java_lang_ref_Reference::referent_offset;
5349 guarantee(referent_offset > 0, "should have already been set");
5350
5351 // Get the argument:
5352 Node* reference_obj = null_check_receiver();
5353 if (stopped()) return true;
5354
5355 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5356
5357 ciInstanceKlass* klass = env()->Object_klass();
5358 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5359
5360 Node* no_ctrl = NULL;
5361 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5362
5363 // Use the pre-barrier to record the value in the referent field
5364 pre_barrier(false /* do_load */,
5365 control(),
5366 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5367 result /* pre_val */,
5368 T_OBJECT);
5369
5370 // Add memory barrier to prevent commoning reads from this field
5371 // across safepoint since GC can change its value.
5372 insert_mem_bar(Op_MemBarCPUOrder);
5373
5374 set_result(result);
5375 return true;
5376 }
5377
5378
5379 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5380 bool is_exact=true, bool is_static=false) {
5381
5382 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5383 assert(tinst != NULL, "obj is null");
5384 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5385 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5386
5387 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5388 ciSymbol::make(fieldTypeString),
5389 is_static);
5390 if (field == NULL) return (Node *) NULL;
5391 assert (field != NULL, "undefined field");
5392
5393 // Next code copied from Parse::do_get_xxx():
5394
5395 // Compute address and memory type.
5396 int offset = field->offset_in_bytes();
5397 bool is_vol = field->is_volatile();
5398 ciType* field_klass = field->type();
5399 assert(field_klass->is_loaded(), "should be loaded");
5400 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5401 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5402 BasicType bt = field->layout_type();
5403
5404 // Build the resultant type of the load
5405 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5406
5407 // Build the load.
5408 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5409 return loadedField;
5410 }
5411
5412
5413 //------------------------------inline_aescrypt_Block-----------------------
5414 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5415 address stubAddr;
5416 const char *stubName;
5417 assert(UseAES, "need AES instruction support");
5418
5419 switch(id) {
5420 case vmIntrinsics::_aescrypt_encryptBlock:
5421 stubAddr = StubRoutines::aescrypt_encryptBlock();
5422 stubName = "aescrypt_encryptBlock";
5423 break;
5424 case vmIntrinsics::_aescrypt_decryptBlock:
5425 stubAddr = StubRoutines::aescrypt_decryptBlock();
5426 stubName = "aescrypt_decryptBlock";
5427 break;
5428 }
5429 if (stubAddr == NULL) return false;
5430
5431 Node* aescrypt_object = argument(0);
5432 Node* src = argument(1);
5433 Node* src_offset = argument(2);
5434 Node* dest = argument(3);
5435 Node* dest_offset = argument(4);
5436
5437 // (1) src and dest are arrays.
5438 const Type* src_type = src->Value(&_gvn);
5439 const Type* dest_type = dest->Value(&_gvn);
5440 const TypeAryPtr* top_src = src_type->isa_aryptr();
5441 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5442 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5443
5444 // for the quick and dirty code we will skip all the checks.
5445 // we are just trying to get the call to be generated.
5446 Node* src_start = src;
5447 Node* dest_start = dest;
5448 if (src_offset != NULL || dest_offset != NULL) {
5449 assert(src_offset != NULL && dest_offset != NULL, "");
5450 src_start = array_element_address(src, src_offset, T_BYTE);
5451 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5452 }
5453
5454 // now need to get the start of its expanded key array
5455 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5456 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5457 if (k_start == NULL) return false;
5458
5459 // Call the stub.
5460 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5461 stubAddr, stubName, TypePtr::BOTTOM,
5462 src_start, dest_start, k_start);
5463
5464 return true;
5465 }
5466
5467 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5468 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5469 address stubAddr;
5470 const char *stubName;
5471
5472 assert(UseAES, "need AES instruction support");
5473
5474 switch(id) {
5475 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5476 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5477 stubName = "cipherBlockChaining_encryptAESCrypt";
5478 break;
5479 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5480 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5481 stubName = "cipherBlockChaining_decryptAESCrypt";
5482 break;
5483 }
5484 if (stubAddr == NULL) return false;
5485
5486 Node* cipherBlockChaining_object = argument(0);
5487 Node* src = argument(1);
5488 Node* src_offset = argument(2);
5489 Node* len = argument(3);
5490 Node* dest = argument(4);
5491 Node* dest_offset = argument(5);
5492
5493 // (1) src and dest are arrays.
5494 const Type* src_type = src->Value(&_gvn);
5495 const Type* dest_type = dest->Value(&_gvn);
5496 const TypeAryPtr* top_src = src_type->isa_aryptr();
5497 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5498 assert (top_src != NULL && top_src->klass() != NULL
5499 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5500
5501 // checks are the responsibility of the caller
5502 Node* src_start = src;
5503 Node* dest_start = dest;
5504 if (src_offset != NULL || dest_offset != NULL) {
5505 assert(src_offset != NULL && dest_offset != NULL, "");
5506 src_start = array_element_address(src, src_offset, T_BYTE);
5507 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5508 }
5509
5510 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5511 // (because of the predicated logic executed earlier).
5512 // so we cast it here safely.
5513 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5514
5515 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5516 if (embeddedCipherObj == NULL) return false;
5517
5518 // cast it to what we know it will be at runtime
5519 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5520 assert(tinst != NULL, "CBC obj is null");
5521 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5522 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5523 if (!klass_AESCrypt->is_loaded()) return false;
5524
5525 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5526 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5527 const TypeOopPtr* xtype = aklass->as_instance_type();
5528 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5529 aescrypt_object = _gvn.transform(aescrypt_object);
5530
5531 // we need to get the start of the aescrypt_object's expanded key array
5532 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5533 if (k_start == NULL) return false;
5534
5535 // similarly, get the start address of the r vector
5536 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5537 if (objRvec == NULL) return false;
5538 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5539
5540 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5541 make_runtime_call(RC_LEAF|RC_NO_FP,
5542 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5543 stubAddr, stubName, TypePtr::BOTTOM,
5544 src_start, dest_start, k_start, r_start, len);
5545
5546 // return is void so no result needs to be pushed
5547
5548 return true;
5549 }
5550
5551 //------------------------------get_key_start_from_aescrypt_object-----------------------
5552 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5553 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5554 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5555 if (objAESCryptKey == NULL) return (Node *) NULL;
5556
5557 // now have the array, need to get the start address of the K array
5558 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5559 return k_start;
5560 }
5561
5562 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5563 // Return node representing slow path of predicate check.
5564 // the pseudo code we want to emulate with this predicate is:
5565 // for encryption:
5566 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5567 // for decryption:
5568 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5569 // note cipher==plain is more conservative than the original java code but that's OK
5570 //
5571 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5572 // First, check receiver for NULL since it is virtual method.
5573 Node* objCBC = argument(0);
5574 objCBC = null_check(objCBC);
5575
5576 if (stopped()) return NULL; // Always NULL
5577
5578 // Load embeddedCipher field of CipherBlockChaining object.
5579 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5580
5581 // get AESCrypt klass for instanceOf check
5582 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5583 // will have same classloader as CipherBlockChaining object
5584 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5585 assert(tinst != NULL, "CBCobj is null");
5586 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5587
5588 // we want to do an instanceof comparison against the AESCrypt class
5589 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5590 if (!klass_AESCrypt->is_loaded()) {
5591 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5592 Node* ctrl = control();
5593 set_control(top()); // no regular fast path
5594 return ctrl;
5595 }
5596 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5597
5598 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5599 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5600 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5601
5602 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5603
5604 // for encryption, we are done
5605 if (!decrypting)
5606 return instof_false; // even if it is NULL
5607
5608 // for decryption, we need to add a further check to avoid
5609 // taking the intrinsic path when cipher and plain are the same
5610 // see the original java code for why.
5611 RegionNode* region = new(C) RegionNode(3);
5612 region->init_req(1, instof_false);
5613 Node* src = argument(1);
5614 Node* dest = argument(4);
5615 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5616 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5617 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5618 region->init_req(2, src_dest_conjoint);
5619
5620 record_for_igvn(region);
5621 return _gvn.transform(region);
5622 }