1 /*
2 * Copyright (c) 1999, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
41
42 class LibraryIntrinsic : public InlineCallGenerator {
43 // Extend the set of intrinsics known to the runtime:
44 public:
45 private:
46 bool _is_virtual;
47 bool _is_predicted;
48 vmIntrinsics::ID _intrinsic_id;
49
50 public:
51 LibraryIntrinsic(ciMethod* m, bool is_virtual, bool is_predicted, vmIntrinsics::ID id)
52 : InlineCallGenerator(m),
53 _is_virtual(is_virtual),
54 _is_predicted(is_predicted),
55 _intrinsic_id(id)
56 {
57 }
58 virtual bool is_intrinsic() const { return true; }
59 virtual bool is_virtual() const { return _is_virtual; }
60 virtual bool is_predicted() const { return _is_predicted; }
61 virtual JVMState* generate(JVMState* jvms);
62 virtual Node* generate_predicate(JVMState* jvms);
63 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
64 };
65
66
67 // Local helper class for LibraryIntrinsic:
68 class LibraryCallKit : public GraphKit {
69 private:
70 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
71 Node* _result; // the result node, if any
72 int _reexecute_sp; // the stack pointer when bytecode needs to be reexecuted
73
74 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
75
76 public:
77 LibraryCallKit(JVMState* jvms, LibraryIntrinsic* intrinsic)
78 : GraphKit(jvms),
79 _intrinsic(intrinsic),
80 _result(NULL)
81 {
82 // Check if this is a root compile. In that case we don't have a caller.
83 if (!jvms->has_method()) {
84 _reexecute_sp = sp();
85 } else {
86 // Find out how many arguments the interpreter needs when deoptimizing
87 // and save the stack pointer value so it can used by uncommon_trap.
88 // We find the argument count by looking at the declared signature.
89 bool ignored_will_link;
90 ciSignature* declared_signature = NULL;
91 ciMethod* ignored_callee = caller()->get_method_at_bci(bci(), ignored_will_link, &declared_signature);
92 const int nargs = declared_signature->arg_size_for_bc(caller()->java_code_at_bci(bci()));
93 _reexecute_sp = sp() + nargs; // "push" arguments back on stack
94 }
95 }
96
97 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
98
99 ciMethod* caller() const { return jvms()->method(); }
100 int bci() const { return jvms()->bci(); }
101 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
102 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
103 ciMethod* callee() const { return _intrinsic->method(); }
104
105 bool try_to_inline();
106 Node* try_to_predicate();
107
108 void push_result() {
109 // Push the result onto the stack.
110 if (!stopped() && result() != NULL) {
111 BasicType bt = result()->bottom_type()->basic_type();
112 push_node(bt, result());
113 }
114 }
115
116 private:
117 void fatal_unexpected_iid(vmIntrinsics::ID iid) {
118 fatal(err_msg_res("unexpected intrinsic %d: %s", iid, vmIntrinsics::name_at(iid)));
119 }
120
121 void set_result(Node* n) { assert(_result == NULL, "only set once"); _result = n; }
122 void set_result(RegionNode* region, PhiNode* value);
123 Node* result() { return _result; }
124
125 virtual int reexecute_sp() { return _reexecute_sp; }
126
127 // Helper functions to inline natives
128 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
129 Node* generate_slow_guard(Node* test, RegionNode* region);
130 Node* generate_fair_guard(Node* test, RegionNode* region);
131 Node* generate_negative_guard(Node* index, RegionNode* region,
132 // resulting CastII of index:
133 Node* *pos_index = NULL);
134 Node* generate_nonpositive_guard(Node* index, bool never_negative,
135 // resulting CastII of index:
136 Node* *pos_index = NULL);
137 Node* generate_limit_guard(Node* offset, Node* subseq_length,
138 Node* array_length,
139 RegionNode* region);
140 Node* generate_current_thread(Node* &tls_output);
141 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
142 bool disjoint_bases, const char* &name, bool dest_uninitialized);
143 Node* load_mirror_from_klass(Node* klass);
144 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
145 RegionNode* region, int null_path,
146 int offset);
147 Node* load_klass_from_mirror(Node* mirror, bool never_see_null,
148 RegionNode* region, int null_path) {
149 int offset = java_lang_Class::klass_offset_in_bytes();
150 return load_klass_from_mirror_common(mirror, never_see_null,
151 region, null_path,
152 offset);
153 }
154 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
155 RegionNode* region, int null_path) {
156 int offset = java_lang_Class::array_klass_offset_in_bytes();
157 return load_klass_from_mirror_common(mirror, never_see_null,
158 region, null_path,
159 offset);
160 }
161 Node* generate_access_flags_guard(Node* kls,
162 int modifier_mask, int modifier_bits,
163 RegionNode* region);
164 Node* generate_interface_guard(Node* kls, RegionNode* region);
165 Node* generate_array_guard(Node* kls, RegionNode* region) {
166 return generate_array_guard_common(kls, region, false, false);
167 }
168 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
169 return generate_array_guard_common(kls, region, false, true);
170 }
171 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
172 return generate_array_guard_common(kls, region, true, false);
173 }
174 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
175 return generate_array_guard_common(kls, region, true, true);
176 }
177 Node* generate_array_guard_common(Node* kls, RegionNode* region,
178 bool obj_array, bool not_array);
179 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
180 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
181 bool is_virtual = false, bool is_static = false);
182 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
183 return generate_method_call(method_id, false, true);
184 }
185 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
186 return generate_method_call(method_id, true, false);
187 }
188 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
189
190 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
191 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
192 bool inline_string_compareTo();
193 bool inline_string_indexOf();
194 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
195 bool inline_string_equals();
196 Node* round_double_node(Node* n);
197 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
198 bool inline_math_native(vmIntrinsics::ID id);
199 bool inline_trig(vmIntrinsics::ID id);
200 bool inline_math(vmIntrinsics::ID id);
201 bool inline_exp();
202 bool inline_pow();
203 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
204 bool inline_min_max(vmIntrinsics::ID id);
205 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
206 // This returns Type::AnyPtr, RawPtr, or OopPtr.
207 int classify_unsafe_addr(Node* &base, Node* &offset);
208 Node* make_unsafe_address(Node* base, Node* offset);
209 // Helper for inline_unsafe_access.
210 // Generates the guards that check whether the result of
211 // Unsafe.getObject should be recorded in an SATB log buffer.
212 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, bool need_mem_bar);
213 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
214 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
215 bool inline_unsafe_allocate();
216 bool inline_unsafe_copyMemory();
217 bool inline_native_currentThread();
218 #ifdef TRACE_HAVE_INTRINSICS
219 bool inline_native_classID();
220 bool inline_native_threadID();
221 #endif
222 bool inline_native_time_funcs(address method, const char* funcName);
223 bool inline_native_isInterrupted();
224 bool inline_native_Class_query(vmIntrinsics::ID id);
225 bool inline_native_subtype_check();
226
227 bool inline_native_newArray();
228 bool inline_native_getLength();
229 bool inline_array_copyOf(bool is_copyOfRange);
230 bool inline_array_equals();
231 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
232 bool inline_native_clone(bool is_virtual);
233 bool inline_native_Reflection_getCallerClass();
234 bool is_method_invoke_or_aux_frame(JVMState* jvms);
235 // Helper function for inlining native object hash method
236 bool inline_native_hashcode(bool is_virtual, bool is_static);
237 bool inline_native_getClass();
238
239 // Helper functions for inlining arraycopy
240 bool inline_arraycopy();
241 void generate_arraycopy(const TypePtr* adr_type,
242 BasicType basic_elem_type,
243 Node* src, Node* src_offset,
244 Node* dest, Node* dest_offset,
245 Node* copy_length,
246 bool disjoint_bases = false,
247 bool length_never_negative = false,
248 RegionNode* slow_region = NULL);
249 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
250 RegionNode* slow_region);
251 void generate_clear_array(const TypePtr* adr_type,
252 Node* dest,
253 BasicType basic_elem_type,
254 Node* slice_off,
255 Node* slice_len,
256 Node* slice_end);
257 bool generate_block_arraycopy(const TypePtr* adr_type,
258 BasicType basic_elem_type,
259 AllocateNode* alloc,
260 Node* src, Node* src_offset,
261 Node* dest, Node* dest_offset,
262 Node* dest_size, bool dest_uninitialized);
263 void generate_slow_arraycopy(const TypePtr* adr_type,
264 Node* src, Node* src_offset,
265 Node* dest, Node* dest_offset,
266 Node* copy_length, bool dest_uninitialized);
267 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
268 Node* dest_elem_klass,
269 Node* src, Node* src_offset,
270 Node* dest, Node* dest_offset,
271 Node* copy_length, bool dest_uninitialized);
272 Node* generate_generic_arraycopy(const TypePtr* adr_type,
273 Node* src, Node* src_offset,
274 Node* dest, Node* dest_offset,
275 Node* copy_length, bool dest_uninitialized);
276 void generate_unchecked_arraycopy(const TypePtr* adr_type,
277 BasicType basic_elem_type,
278 bool disjoint_bases,
279 Node* src, Node* src_offset,
280 Node* dest, Node* dest_offset,
281 Node* copy_length, bool dest_uninitialized);
282 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
283 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
284 bool inline_unsafe_ordered_store(BasicType type);
285 bool inline_fp_conversions(vmIntrinsics::ID id);
286 bool inline_number_methods(vmIntrinsics::ID id);
287 bool inline_reference_get();
288 bool inline_aescrypt_Block(vmIntrinsics::ID id);
289 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
290 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
291 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
292 };
293
294
295 //---------------------------make_vm_intrinsic----------------------------
296 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
297 vmIntrinsics::ID id = m->intrinsic_id();
298 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
299
300 if (DisableIntrinsic[0] != '\0'
301 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
302 // disabled by a user request on the command line:
303 // example: -XX:DisableIntrinsic=_hashCode,_getClass
304 return NULL;
305 }
306
307 if (!m->is_loaded()) {
308 // do not attempt to inline unloaded methods
309 return NULL;
310 }
311
312 // Only a few intrinsics implement a virtual dispatch.
313 // They are expensive calls which are also frequently overridden.
314 if (is_virtual) {
315 switch (id) {
316 case vmIntrinsics::_hashCode:
317 case vmIntrinsics::_clone:
318 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
319 break;
320 default:
321 return NULL;
322 }
323 }
324
325 // -XX:-InlineNatives disables nearly all intrinsics:
326 if (!InlineNatives) {
327 switch (id) {
328 case vmIntrinsics::_indexOf:
329 case vmIntrinsics::_compareTo:
330 case vmIntrinsics::_equals:
331 case vmIntrinsics::_equalsC:
332 case vmIntrinsics::_getAndAddInt:
333 case vmIntrinsics::_getAndAddLong:
334 case vmIntrinsics::_getAndSetInt:
335 case vmIntrinsics::_getAndSetLong:
336 case vmIntrinsics::_getAndSetObject:
337 break; // InlineNatives does not control String.compareTo
338 case vmIntrinsics::_Reference_get:
339 break; // InlineNatives does not control Reference.get
340 default:
341 return NULL;
342 }
343 }
344
345 bool is_predicted = false;
346
347 switch (id) {
348 case vmIntrinsics::_compareTo:
349 if (!SpecialStringCompareTo) return NULL;
350 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
351 break;
352 case vmIntrinsics::_indexOf:
353 if (!SpecialStringIndexOf) return NULL;
354 break;
355 case vmIntrinsics::_equals:
356 if (!SpecialStringEquals) return NULL;
357 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
358 break;
359 case vmIntrinsics::_equalsC:
360 if (!SpecialArraysEquals) return NULL;
361 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
362 break;
363 case vmIntrinsics::_arraycopy:
364 if (!InlineArrayCopy) return NULL;
365 break;
366 case vmIntrinsics::_copyMemory:
367 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
368 if (!InlineArrayCopy) return NULL;
369 break;
370 case vmIntrinsics::_hashCode:
371 if (!InlineObjectHash) return NULL;
372 break;
373 case vmIntrinsics::_clone:
374 case vmIntrinsics::_copyOf:
375 case vmIntrinsics::_copyOfRange:
376 if (!InlineObjectCopy) return NULL;
377 // These also use the arraycopy intrinsic mechanism:
378 if (!InlineArrayCopy) return NULL;
379 break;
380 case vmIntrinsics::_checkIndex:
381 // We do not intrinsify this. The optimizer does fine with it.
382 return NULL;
383
384 case vmIntrinsics::_getCallerClass:
385 if (!UseNewReflection) return NULL;
386 if (!InlineReflectionGetCallerClass) return NULL;
387 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
388 break;
389
390 case vmIntrinsics::_bitCount_i:
391 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
392 break;
393
394 case vmIntrinsics::_bitCount_l:
395 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
396 break;
397
398 case vmIntrinsics::_numberOfLeadingZeros_i:
399 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
400 break;
401
402 case vmIntrinsics::_numberOfLeadingZeros_l:
403 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
404 break;
405
406 case vmIntrinsics::_numberOfTrailingZeros_i:
407 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
408 break;
409
410 case vmIntrinsics::_numberOfTrailingZeros_l:
411 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
412 break;
413
414 case vmIntrinsics::_reverseBytes_c:
415 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return false;
416 break;
417 case vmIntrinsics::_reverseBytes_s:
418 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return false;
419 break;
420 case vmIntrinsics::_reverseBytes_i:
421 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return false;
422 break;
423 case vmIntrinsics::_reverseBytes_l:
424 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return false;
425 break;
426
427 case vmIntrinsics::_Reference_get:
428 // Use the intrinsic version of Reference.get() so that the value in
429 // the referent field can be registered by the G1 pre-barrier code.
430 // Also add memory barrier to prevent commoning reads from this field
431 // across safepoint since GC can change it value.
432 break;
433
434 case vmIntrinsics::_compareAndSwapObject:
435 #ifdef _LP64
436 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
437 #endif
438 break;
439
440 case vmIntrinsics::_compareAndSwapLong:
441 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
442 break;
443
444 case vmIntrinsics::_getAndAddInt:
445 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
446 break;
447
448 case vmIntrinsics::_getAndAddLong:
449 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
450 break;
451
452 case vmIntrinsics::_getAndSetInt:
453 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
454 break;
455
456 case vmIntrinsics::_getAndSetLong:
457 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
458 break;
459
460 case vmIntrinsics::_getAndSetObject:
461 #ifdef _LP64
462 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
463 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
464 break;
465 #else
466 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
467 break;
468 #endif
469
470 case vmIntrinsics::_aescrypt_encryptBlock:
471 case vmIntrinsics::_aescrypt_decryptBlock:
472 if (!UseAESIntrinsics) return NULL;
473 break;
474
475 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
476 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
477 if (!UseAESIntrinsics) return NULL;
478 // these two require the predicated logic
479 is_predicted = true;
480 break;
481
482 default:
483 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
484 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
485 break;
486 }
487
488 // -XX:-InlineClassNatives disables natives from the Class class.
489 // The flag applies to all reflective calls, notably Array.newArray
490 // (visible to Java programmers as Array.newInstance).
491 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
492 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
493 if (!InlineClassNatives) return NULL;
494 }
495
496 // -XX:-InlineThreadNatives disables natives from the Thread class.
497 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
498 if (!InlineThreadNatives) return NULL;
499 }
500
501 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
502 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
503 m->holder()->name() == ciSymbol::java_lang_Float() ||
504 m->holder()->name() == ciSymbol::java_lang_Double()) {
505 if (!InlineMathNatives) return NULL;
506 }
507
508 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
509 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
510 if (!InlineUnsafeOps) return NULL;
511 }
512
513 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
514 }
515
516 //----------------------register_library_intrinsics-----------------------
517 // Initialize this file's data structures, for each Compile instance.
518 void Compile::register_library_intrinsics() {
519 // Nothing to do here.
520 }
521
522 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
523 LibraryCallKit kit(jvms, this);
524 Compile* C = kit.C;
525 int nodes = C->unique();
526 #ifndef PRODUCT
527 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
528 char buf[1000];
529 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
530 tty->print_cr("Intrinsic %s", str);
531 }
532 #endif
533 ciMethod* callee = kit.callee();
534 const int bci = kit.bci();
535
536 // Try to inline the intrinsic.
537 if (kit.try_to_inline()) {
538 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
539 CompileTask::print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
540 }
541 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
542 if (C->log()) {
543 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
544 vmIntrinsics::name_at(intrinsic_id()),
545 (is_virtual() ? " virtual='1'" : ""),
546 C->unique() - nodes);
547 }
548 // Push the result from the inlined method onto the stack.
549 kit.push_result();
550 return kit.transfer_exceptions_into_jvms();
551 }
552
553 // The intrinsic bailed out
554 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
555 if (jvms->has_method()) {
556 // Not a root compile.
557 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
558 CompileTask::print_inlining(callee, jvms->depth() - 1, bci, msg);
559 } else {
560 // Root compile
561 tty->print("Did not generate intrinsic %s%s at bci:%d in",
562 vmIntrinsics::name_at(intrinsic_id()),
563 (is_virtual() ? " (virtual)" : ""), bci);
564 }
565 }
566 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
567 return NULL;
568 }
569
570 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
571 LibraryCallKit kit(jvms, this);
572 Compile* C = kit.C;
573 int nodes = C->unique();
574 #ifndef PRODUCT
575 assert(is_predicted(), "sanity");
576 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
577 char buf[1000];
578 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
579 tty->print_cr("Predicate for intrinsic %s", str);
580 }
581 #endif
582 ciMethod* callee = kit.callee();
583 const int bci = kit.bci();
584
585 Node* slow_ctl = kit.try_to_predicate();
586 if (!kit.failing()) {
587 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
588 CompileTask::print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
589 }
590 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
591 if (C->log()) {
592 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
593 vmIntrinsics::name_at(intrinsic_id()),
594 (is_virtual() ? " virtual='1'" : ""),
595 C->unique() - nodes);
596 }
597 return slow_ctl; // Could be NULL if the check folds.
598 }
599
600 // The intrinsic bailed out
601 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
602 if (jvms->has_method()) {
603 // Not a root compile.
604 const char* msg = "failed to generate predicate for intrinsic";
605 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
606 } else {
607 // Root compile
608 tty->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
609 vmIntrinsics::name_at(intrinsic_id()),
610 (is_virtual() ? " (virtual)" : ""), bci);
611 }
612 }
613 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
614 return NULL;
615 }
616
617 bool LibraryCallKit::try_to_inline() {
618 // Handle symbolic names for otherwise undistinguished boolean switches:
619 const bool is_store = true;
620 const bool is_native_ptr = true;
621 const bool is_static = true;
622 const bool is_volatile = true;
623
624 if (!jvms()->has_method()) {
625 // Root JVMState has a null method.
626 assert(map()->memory()->Opcode() == Op_Parm, "");
627 // Insert the memory aliasing node
628 set_all_memory(reset_memory());
629 }
630 assert(merged_memory(), "");
631
632
633 switch (intrinsic_id()) {
634 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
635 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
636 case vmIntrinsics::_getClass: return inline_native_getClass();
637
638 case vmIntrinsics::_dsin:
639 case vmIntrinsics::_dcos:
640 case vmIntrinsics::_dtan:
641 case vmIntrinsics::_dabs:
642 case vmIntrinsics::_datan2:
643 case vmIntrinsics::_dsqrt:
644 case vmIntrinsics::_dexp:
645 case vmIntrinsics::_dlog:
646 case vmIntrinsics::_dlog10:
647 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
648
649 case vmIntrinsics::_min:
650 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
651
652 case vmIntrinsics::_arraycopy: return inline_arraycopy();
653
654 case vmIntrinsics::_compareTo: return inline_string_compareTo();
655 case vmIntrinsics::_indexOf: return inline_string_indexOf();
656 case vmIntrinsics::_equals: return inline_string_equals();
657
658 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
659 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
660 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
661 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
662 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
663 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
664 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
665 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
666 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
667
668 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
669 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
670 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
671 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
672 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
673 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
674 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
675 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
676 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
677
678 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
679 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
680 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
681 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
682 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
683 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
684 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
685 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
686
687 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
688 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
689 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
690 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
691 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
692 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
693 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
694 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
695
696 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
697 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
698 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
699 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
700 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
701 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
702 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
703 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
704 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
705
706 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
707 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
708 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
709 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
710 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
711 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
712 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
713 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
714 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
715
716 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
717 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
718 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
719 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
720
721 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
722 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
723 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
724
725 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
726 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
727 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
728
729 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
730 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
731 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
732 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
733 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
734
735 case vmIntrinsics::_currentThread: return inline_native_currentThread();
736 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
737
738 #ifdef TRACE_HAVE_INTRINSICS
739 case vmIntrinsics::_classID: return inline_native_classID();
740 case vmIntrinsics::_threadID: return inline_native_threadID();
741 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
742 #endif
743 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
744 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
745 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
746 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
747 case vmIntrinsics::_newArray: return inline_native_newArray();
748 case vmIntrinsics::_getLength: return inline_native_getLength();
749 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
750 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
751 case vmIntrinsics::_equalsC: return inline_array_equals();
752 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
753
754 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
755
756 case vmIntrinsics::_isInstance:
757 case vmIntrinsics::_getModifiers:
758 case vmIntrinsics::_isInterface:
759 case vmIntrinsics::_isArray:
760 case vmIntrinsics::_isPrimitive:
761 case vmIntrinsics::_getSuperclass:
762 case vmIntrinsics::_getComponentType:
763 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
764
765 case vmIntrinsics::_floatToRawIntBits:
766 case vmIntrinsics::_floatToIntBits:
767 case vmIntrinsics::_intBitsToFloat:
768 case vmIntrinsics::_doubleToRawLongBits:
769 case vmIntrinsics::_doubleToLongBits:
770 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
771
772 case vmIntrinsics::_numberOfLeadingZeros_i:
773 case vmIntrinsics::_numberOfLeadingZeros_l:
774 case vmIntrinsics::_numberOfTrailingZeros_i:
775 case vmIntrinsics::_numberOfTrailingZeros_l:
776 case vmIntrinsics::_bitCount_i:
777 case vmIntrinsics::_bitCount_l:
778 case vmIntrinsics::_reverseBytes_i:
779 case vmIntrinsics::_reverseBytes_l:
780 case vmIntrinsics::_reverseBytes_s:
781 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
782
783 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
784
785 case vmIntrinsics::_Reference_get: return inline_reference_get();
786
787 case vmIntrinsics::_aescrypt_encryptBlock:
788 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
789
790 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
791 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
792 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
793
794 default:
795 // If you get here, it may be that someone has added a new intrinsic
796 // to the list in vmSymbols.hpp without implementing it here.
797 #ifndef PRODUCT
798 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
799 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
800 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
801 }
802 #endif
803 return false;
804 }
805 }
806
807 Node* LibraryCallKit::try_to_predicate() {
808 if (!jvms()->has_method()) {
809 // Root JVMState has a null method.
810 assert(map()->memory()->Opcode() == Op_Parm, "");
811 // Insert the memory aliasing node
812 set_all_memory(reset_memory());
813 }
814 assert(merged_memory(), "");
815
816 switch (intrinsic_id()) {
817 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
818 return inline_cipherBlockChaining_AESCrypt_predicate(false);
819 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
820 return inline_cipherBlockChaining_AESCrypt_predicate(true);
821
822 default:
823 // If you get here, it may be that someone has added a new intrinsic
824 // to the list in vmSymbols.hpp without implementing it here.
825 #ifndef PRODUCT
826 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
827 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
828 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
829 }
830 #endif
831 Node* slow_ctl = control();
832 set_control(top()); // No fast path instrinsic
833 return slow_ctl;
834 }
835 }
836
837 //------------------------------set_result-------------------------------
838 // Helper function for finishing intrinsics.
839 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
840 record_for_igvn(region);
841 set_control(_gvn.transform(region));
842 set_result( _gvn.transform(value));
843 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
844 }
845
846 //------------------------------generate_guard---------------------------
847 // Helper function for generating guarded fast-slow graph structures.
848 // The given 'test', if true, guards a slow path. If the test fails
849 // then a fast path can be taken. (We generally hope it fails.)
850 // In all cases, GraphKit::control() is updated to the fast path.
851 // The returned value represents the control for the slow path.
852 // The return value is never 'top'; it is either a valid control
853 // or NULL if it is obvious that the slow path can never be taken.
854 // Also, if region and the slow control are not NULL, the slow edge
855 // is appended to the region.
856 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
857 if (stopped()) {
858 // Already short circuited.
859 return NULL;
860 }
861
862 // Build an if node and its projections.
863 // If test is true we take the slow path, which we assume is uncommon.
864 if (_gvn.type(test) == TypeInt::ZERO) {
865 // The slow branch is never taken. No need to build this guard.
866 return NULL;
867 }
868
869 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
870
871 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
872 if (if_slow == top()) {
873 // The slow branch is never taken. No need to build this guard.
874 return NULL;
875 }
876
877 if (region != NULL)
878 region->add_req(if_slow);
879
880 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
881 set_control(if_fast);
882
883 return if_slow;
884 }
885
886 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
887 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
888 }
889 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
890 return generate_guard(test, region, PROB_FAIR);
891 }
892
893 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
894 Node* *pos_index) {
895 if (stopped())
896 return NULL; // already stopped
897 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
898 return NULL; // index is already adequately typed
899 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
900 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
901 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
902 if (is_neg != NULL && pos_index != NULL) {
903 // Emulate effect of Parse::adjust_map_after_if.
904 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
905 ccast->set_req(0, control());
906 (*pos_index) = _gvn.transform(ccast);
907 }
908 return is_neg;
909 }
910
911 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
912 Node* *pos_index) {
913 if (stopped())
914 return NULL; // already stopped
915 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
916 return NULL; // index is already adequately typed
917 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
918 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
919 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
920 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
921 if (is_notp != NULL && pos_index != NULL) {
922 // Emulate effect of Parse::adjust_map_after_if.
923 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
924 ccast->set_req(0, control());
925 (*pos_index) = _gvn.transform(ccast);
926 }
927 return is_notp;
928 }
929
930 // Make sure that 'position' is a valid limit index, in [0..length].
931 // There are two equivalent plans for checking this:
932 // A. (offset + copyLength) unsigned<= arrayLength
933 // B. offset <= (arrayLength - copyLength)
934 // We require that all of the values above, except for the sum and
935 // difference, are already known to be non-negative.
936 // Plan A is robust in the face of overflow, if offset and copyLength
937 // are both hugely positive.
938 //
939 // Plan B is less direct and intuitive, but it does not overflow at
940 // all, since the difference of two non-negatives is always
941 // representable. Whenever Java methods must perform the equivalent
942 // check they generally use Plan B instead of Plan A.
943 // For the moment we use Plan A.
944 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
945 Node* subseq_length,
946 Node* array_length,
947 RegionNode* region) {
948 if (stopped())
949 return NULL; // already stopped
950 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
951 if (zero_offset && subseq_length->eqv_uncast(array_length))
952 return NULL; // common case of whole-array copy
953 Node* last = subseq_length;
954 if (!zero_offset) // last += offset
955 last = _gvn.transform( new (C) AddINode(last, offset));
956 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
957 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
958 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
959 return is_over;
960 }
961
962
963 //--------------------------generate_current_thread--------------------
964 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
965 ciKlass* thread_klass = env()->Thread_klass();
966 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
967 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
968 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
969 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
970 tls_output = thread;
971 return threadObj;
972 }
973
974
975 //------------------------------make_string_method_node------------------------
976 // Helper method for String intrinsic functions. This version is called
977 // with str1 and str2 pointing to String object nodes.
978 //
979 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
980 Node* no_ctrl = NULL;
981
982 // Get start addr of string
983 Node* str1_value = load_String_value(no_ctrl, str1);
984 Node* str1_offset = load_String_offset(no_ctrl, str1);
985 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
986
987 // Get length of string 1
988 Node* str1_len = load_String_length(no_ctrl, str1);
989
990 Node* str2_value = load_String_value(no_ctrl, str2);
991 Node* str2_offset = load_String_offset(no_ctrl, str2);
992 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
993
994 Node* str2_len = NULL;
995 Node* result = NULL;
996
997 switch (opcode) {
998 case Op_StrIndexOf:
999 // Get length of string 2
1000 str2_len = load_String_length(no_ctrl, str2);
1001
1002 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1003 str1_start, str1_len, str2_start, str2_len);
1004 break;
1005 case Op_StrComp:
1006 // Get length of string 2
1007 str2_len = load_String_length(no_ctrl, str2);
1008
1009 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1010 str1_start, str1_len, str2_start, str2_len);
1011 break;
1012 case Op_StrEquals:
1013 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1014 str1_start, str2_start, str1_len);
1015 break;
1016 default:
1017 ShouldNotReachHere();
1018 return NULL;
1019 }
1020
1021 // All these intrinsics have checks.
1022 C->set_has_split_ifs(true); // Has chance for split-if optimization
1023
1024 return _gvn.transform(result);
1025 }
1026
1027 // Helper method for String intrinsic functions. This version is called
1028 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1029 // to Int nodes containing the lenghts of str1 and str2.
1030 //
1031 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1032 Node* result = NULL;
1033 switch (opcode) {
1034 case Op_StrIndexOf:
1035 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1036 str1_start, cnt1, str2_start, cnt2);
1037 break;
1038 case Op_StrComp:
1039 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1040 str1_start, cnt1, str2_start, cnt2);
1041 break;
1042 case Op_StrEquals:
1043 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1044 str1_start, str2_start, cnt1);
1045 break;
1046 default:
1047 ShouldNotReachHere();
1048 return NULL;
1049 }
1050
1051 // All these intrinsics have checks.
1052 C->set_has_split_ifs(true); // Has chance for split-if optimization
1053
1054 return _gvn.transform(result);
1055 }
1056
1057 //------------------------------inline_string_compareTo------------------------
1058 // public int java.lang.String.compareTo(String anotherString);
1059 bool LibraryCallKit::inline_string_compareTo() {
1060 Node* receiver = null_check(argument(0));
1061 Node* arg = null_check(argument(1));
1062 if (stopped()) {
1063 return true;
1064 }
1065 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1066 return true;
1067 }
1068
1069 //------------------------------inline_string_equals------------------------
1070 bool LibraryCallKit::inline_string_equals() {
1071 Node* receiver = null_check_receiver();
1072 // NOTE: Do not null check argument for String.equals() because spec
1073 // allows to specify NULL as argument.
1074 Node* argument = this->argument(1);
1075 if (stopped()) {
1076 return true;
1077 }
1078
1079 // paths (plus control) merge
1080 RegionNode* region = new (C) RegionNode(5);
1081 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1082
1083 // does source == target string?
1084 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1085 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1086
1087 Node* if_eq = generate_slow_guard(bol, NULL);
1088 if (if_eq != NULL) {
1089 // receiver == argument
1090 phi->init_req(2, intcon(1));
1091 region->init_req(2, if_eq);
1092 }
1093
1094 // get String klass for instanceOf
1095 ciInstanceKlass* klass = env()->String_klass();
1096
1097 if (!stopped()) {
1098 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1099 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1100 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1101
1102 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1103 //instanceOf == true, fallthrough
1104
1105 if (inst_false != NULL) {
1106 phi->init_req(3, intcon(0));
1107 region->init_req(3, inst_false);
1108 }
1109 }
1110
1111 if (!stopped()) {
1112 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1113
1114 // Properly cast the argument to String
1115 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1116 // This path is taken only when argument's type is String:NotNull.
1117 argument = cast_not_null(argument, false);
1118
1119 Node* no_ctrl = NULL;
1120
1121 // Get start addr of receiver
1122 Node* receiver_val = load_String_value(no_ctrl, receiver);
1123 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1124 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1125
1126 // Get length of receiver
1127 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1128
1129 // Get start addr of argument
1130 Node* argument_val = load_String_value(no_ctrl, argument);
1131 Node* argument_offset = load_String_offset(no_ctrl, argument);
1132 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1133
1134 // Get length of argument
1135 Node* argument_cnt = load_String_length(no_ctrl, argument);
1136
1137 // Check for receiver count != argument count
1138 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1139 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1140 Node* if_ne = generate_slow_guard(bol, NULL);
1141 if (if_ne != NULL) {
1142 phi->init_req(4, intcon(0));
1143 region->init_req(4, if_ne);
1144 }
1145
1146 // Check for count == 0 is done by assembler code for StrEquals.
1147
1148 if (!stopped()) {
1149 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1150 phi->init_req(1, equals);
1151 region->init_req(1, control());
1152 }
1153 }
1154
1155 // post merge
1156 set_control(_gvn.transform(region));
1157 record_for_igvn(region);
1158
1159 set_result(_gvn.transform(phi));
1160 return true;
1161 }
1162
1163 //------------------------------inline_array_equals----------------------------
1164 bool LibraryCallKit::inline_array_equals() {
1165 Node* arg1 = argument(0);
1166 Node* arg2 = argument(1);
1167 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1168 return true;
1169 }
1170
1171 // Java version of String.indexOf(constant string)
1172 // class StringDecl {
1173 // StringDecl(char[] ca) {
1174 // offset = 0;
1175 // count = ca.length;
1176 // value = ca;
1177 // }
1178 // int offset;
1179 // int count;
1180 // char[] value;
1181 // }
1182 //
1183 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1184 // int targetOffset, int cache_i, int md2) {
1185 // int cache = cache_i;
1186 // int sourceOffset = string_object.offset;
1187 // int sourceCount = string_object.count;
1188 // int targetCount = target_object.length;
1189 //
1190 // int targetCountLess1 = targetCount - 1;
1191 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1192 //
1193 // char[] source = string_object.value;
1194 // char[] target = target_object;
1195 // int lastChar = target[targetCountLess1];
1196 //
1197 // outer_loop:
1198 // for (int i = sourceOffset; i < sourceEnd; ) {
1199 // int src = source[i + targetCountLess1];
1200 // if (src == lastChar) {
1201 // // With random strings and a 4-character alphabet,
1202 // // reverse matching at this point sets up 0.8% fewer
1203 // // frames, but (paradoxically) makes 0.3% more probes.
1204 // // Since those probes are nearer the lastChar probe,
1205 // // there is may be a net D$ win with reverse matching.
1206 // // But, reversing loop inhibits unroll of inner loop
1207 // // for unknown reason. So, does running outer loop from
1208 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1209 // for (int j = 0; j < targetCountLess1; j++) {
1210 // if (target[targetOffset + j] != source[i+j]) {
1211 // if ((cache & (1 << source[i+j])) == 0) {
1212 // if (md2 < j+1) {
1213 // i += j+1;
1214 // continue outer_loop;
1215 // }
1216 // }
1217 // i += md2;
1218 // continue outer_loop;
1219 // }
1220 // }
1221 // return i - sourceOffset;
1222 // }
1223 // if ((cache & (1 << src)) == 0) {
1224 // i += targetCountLess1;
1225 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1226 // i++;
1227 // }
1228 // return -1;
1229 // }
1230
1231 //------------------------------string_indexOf------------------------
1232 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1233 jint cache_i, jint md2_i) {
1234
1235 Node* no_ctrl = NULL;
1236 float likely = PROB_LIKELY(0.9);
1237 float unlikely = PROB_UNLIKELY(0.9);
1238
1239 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1240
1241 Node* source = load_String_value(no_ctrl, string_object);
1242 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1243 Node* sourceCount = load_String_length(no_ctrl, string_object);
1244
1245 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1246 jint target_length = target_array->length();
1247 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1248 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1249
1250 IdealKit kit(this, false, true);
1251 #define __ kit.
1252 Node* zero = __ ConI(0);
1253 Node* one = __ ConI(1);
1254 Node* cache = __ ConI(cache_i);
1255 Node* md2 = __ ConI(md2_i);
1256 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1257 Node* targetCount = __ ConI(target_length);
1258 Node* targetCountLess1 = __ ConI(target_length - 1);
1259 Node* targetOffset = __ ConI(targetOffset_i);
1260 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1261
1262 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1263 Node* outer_loop = __ make_label(2 /* goto */);
1264 Node* return_ = __ make_label(1);
1265
1266 __ set(rtn,__ ConI(-1));
1267 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1268 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1269 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1270 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1271 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1272 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1273 Node* tpj = __ AddI(targetOffset, __ value(j));
1274 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1275 Node* ipj = __ AddI(__ value(i), __ value(j));
1276 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1277 __ if_then(targ, BoolTest::ne, src2); {
1278 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1279 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1280 __ increment(i, __ AddI(__ value(j), one));
1281 __ goto_(outer_loop);
1282 } __ end_if(); __ dead(j);
1283 }__ end_if(); __ dead(j);
1284 __ increment(i, md2);
1285 __ goto_(outer_loop);
1286 }__ end_if();
1287 __ increment(j, one);
1288 }__ end_loop(); __ dead(j);
1289 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1290 __ goto_(return_);
1291 }__ end_if();
1292 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1293 __ increment(i, targetCountLess1);
1294 }__ end_if();
1295 __ increment(i, one);
1296 __ bind(outer_loop);
1297 }__ end_loop(); __ dead(i);
1298 __ bind(return_);
1299
1300 // Final sync IdealKit and GraphKit.
1301 final_sync(kit);
1302 Node* result = __ value(rtn);
1303 #undef __
1304 C->set_has_loops(true);
1305 return result;
1306 }
1307
1308 //------------------------------inline_string_indexOf------------------------
1309 bool LibraryCallKit::inline_string_indexOf() {
1310 Node* receiver = argument(0);
1311 Node* arg = argument(1);
1312
1313 Node* result;
1314 // Disable the use of pcmpestri until it can be guaranteed that
1315 // the load doesn't cross into the uncommited space.
1316 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1317 UseSSE42Intrinsics) {
1318 // Generate SSE4.2 version of indexOf
1319 // We currently only have match rules that use SSE4.2
1320
1321 receiver = null_check(receiver);
1322 arg = null_check(arg);
1323 if (stopped()) {
1324 return true;
1325 }
1326
1327 ciInstanceKlass* str_klass = env()->String_klass();
1328 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1329
1330 // Make the merge point
1331 RegionNode* result_rgn = new (C) RegionNode(4);
1332 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1333 Node* no_ctrl = NULL;
1334
1335 // Get start addr of source string
1336 Node* source = load_String_value(no_ctrl, receiver);
1337 Node* source_offset = load_String_offset(no_ctrl, receiver);
1338 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1339
1340 // Get length of source string
1341 Node* source_cnt = load_String_length(no_ctrl, receiver);
1342
1343 // Get start addr of substring
1344 Node* substr = load_String_value(no_ctrl, arg);
1345 Node* substr_offset = load_String_offset(no_ctrl, arg);
1346 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1347
1348 // Get length of source string
1349 Node* substr_cnt = load_String_length(no_ctrl, arg);
1350
1351 // Check for substr count > string count
1352 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1353 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1354 Node* if_gt = generate_slow_guard(bol, NULL);
1355 if (if_gt != NULL) {
1356 result_phi->init_req(2, intcon(-1));
1357 result_rgn->init_req(2, if_gt);
1358 }
1359
1360 if (!stopped()) {
1361 // Check for substr count == 0
1362 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1363 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1364 Node* if_zero = generate_slow_guard(bol, NULL);
1365 if (if_zero != NULL) {
1366 result_phi->init_req(3, intcon(0));
1367 result_rgn->init_req(3, if_zero);
1368 }
1369 }
1370
1371 if (!stopped()) {
1372 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1373 result_phi->init_req(1, result);
1374 result_rgn->init_req(1, control());
1375 }
1376 set_control(_gvn.transform(result_rgn));
1377 record_for_igvn(result_rgn);
1378 result = _gvn.transform(result_phi);
1379
1380 } else { // Use LibraryCallKit::string_indexOf
1381 // don't intrinsify if argument isn't a constant string.
1382 if (!arg->is_Con()) {
1383 return false;
1384 }
1385 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1386 if (str_type == NULL) {
1387 return false;
1388 }
1389 ciInstanceKlass* klass = env()->String_klass();
1390 ciObject* str_const = str_type->const_oop();
1391 if (str_const == NULL || str_const->klass() != klass) {
1392 return false;
1393 }
1394 ciInstance* str = str_const->as_instance();
1395 assert(str != NULL, "must be instance");
1396
1397 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1398 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1399
1400 int o;
1401 int c;
1402 if (java_lang_String::has_offset_field()) {
1403 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1404 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1405 } else {
1406 o = 0;
1407 c = pat->length();
1408 }
1409
1410 // constant strings have no offset and count == length which
1411 // simplifies the resulting code somewhat so lets optimize for that.
1412 if (o != 0 || c != pat->length()) {
1413 return false;
1414 }
1415
1416 receiver = null_check(receiver, T_OBJECT);
1417 // NOTE: No null check on the argument is needed since it's a constant String oop.
1418 if (stopped()) {
1419 return true;
1420 }
1421
1422 // The null string as a pattern always returns 0 (match at beginning of string)
1423 if (c == 0) {
1424 set_result(intcon(0));
1425 return true;
1426 }
1427
1428 // Generate default indexOf
1429 jchar lastChar = pat->char_at(o + (c - 1));
1430 int cache = 0;
1431 int i;
1432 for (i = 0; i < c - 1; i++) {
1433 assert(i < pat->length(), "out of range");
1434 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1435 }
1436
1437 int md2 = c;
1438 for (i = 0; i < c - 1; i++) {
1439 assert(i < pat->length(), "out of range");
1440 if (pat->char_at(o + i) == lastChar) {
1441 md2 = (c - 1) - i;
1442 }
1443 }
1444
1445 result = string_indexOf(receiver, pat, o, cache, md2);
1446 }
1447 set_result(result);
1448 return true;
1449 }
1450
1451 //--------------------------round_double_node--------------------------------
1452 // Round a double node if necessary.
1453 Node* LibraryCallKit::round_double_node(Node* n) {
1454 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1455 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1456 return n;
1457 }
1458
1459 //------------------------------inline_math-----------------------------------
1460 // public static double Math.abs(double)
1461 // public static double Math.sqrt(double)
1462 // public static double Math.log(double)
1463 // public static double Math.log10(double)
1464 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1465 Node* arg = round_double_node(argument(0));
1466 Node* n;
1467 switch (id) {
1468 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1469 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(0, arg); break;
1470 case vmIntrinsics::_dlog: n = new (C) LogDNode( arg); break;
1471 case vmIntrinsics::_dlog10: n = new (C) Log10DNode( arg); break;
1472 default: fatal_unexpected_iid(id); break;
1473 }
1474 set_result(_gvn.transform(n));
1475 return true;
1476 }
1477
1478 //------------------------------inline_trig----------------------------------
1479 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1480 // argument reduction which will turn into a fast/slow diamond.
1481 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1482 Node* arg = round_double_node(argument(0));
1483 Node* n = NULL;
1484
1485 switch (id) {
1486 case vmIntrinsics::_dsin: n = new (C) SinDNode(arg); break;
1487 case vmIntrinsics::_dcos: n = new (C) CosDNode(arg); break;
1488 case vmIntrinsics::_dtan: n = new (C) TanDNode(arg); break;
1489 default: fatal_unexpected_iid(id); break;
1490 }
1491 n = _gvn.transform(n);
1492
1493 // Rounding required? Check for argument reduction!
1494 if (Matcher::strict_fp_requires_explicit_rounding) {
1495 static const double pi_4 = 0.7853981633974483;
1496 static const double neg_pi_4 = -0.7853981633974483;
1497 // pi/2 in 80-bit extended precision
1498 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1499 // -pi/2 in 80-bit extended precision
1500 // 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};
1501 // Cutoff value for using this argument reduction technique
1502 //static const double pi_2_minus_epsilon = 1.564660403643354;
1503 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1504
1505 // Pseudocode for sin:
1506 // if (x <= Math.PI / 4.0) {
1507 // if (x >= -Math.PI / 4.0) return fsin(x);
1508 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1509 // } else {
1510 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1511 // }
1512 // return StrictMath.sin(x);
1513
1514 // Pseudocode for cos:
1515 // if (x <= Math.PI / 4.0) {
1516 // if (x >= -Math.PI / 4.0) return fcos(x);
1517 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1518 // } else {
1519 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1520 // }
1521 // return StrictMath.cos(x);
1522
1523 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1524 // requires a special machine instruction to load it. Instead we'll try
1525 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1526 // probably do the math inside the SIN encoding.
1527
1528 // Make the merge point
1529 RegionNode* r = new (C) RegionNode(3);
1530 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1531
1532 // Flatten arg so we need only 1 test
1533 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1534 // Node for PI/4 constant
1535 Node *pi4 = makecon(TypeD::make(pi_4));
1536 // Check PI/4 : abs(arg)
1537 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1538 // Check: If PI/4 < abs(arg) then go slow
1539 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1540 // Branch either way
1541 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1542 set_control(opt_iff(r,iff));
1543
1544 // Set fast path result
1545 phi->init_req(2, n);
1546
1547 // Slow path - non-blocking leaf call
1548 Node* call = NULL;
1549 switch (id) {
1550 case vmIntrinsics::_dsin:
1551 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1552 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1553 "Sin", NULL, arg, top());
1554 break;
1555 case vmIntrinsics::_dcos:
1556 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1557 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1558 "Cos", NULL, arg, top());
1559 break;
1560 case vmIntrinsics::_dtan:
1561 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1562 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1563 "Tan", NULL, arg, top());
1564 break;
1565 }
1566 assert(control()->in(0) == call, "");
1567 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1568 r->init_req(1, control());
1569 phi->init_req(1, slow_result);
1570
1571 // Post-merge
1572 set_control(_gvn.transform(r));
1573 record_for_igvn(r);
1574 n = _gvn.transform(phi);
1575
1576 C->set_has_split_ifs(true); // Has chance for split-if optimization
1577 }
1578 set_result(n);
1579 return true;
1580 }
1581
1582 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1583 //-------------------
1584 //result=(result.isNaN())? funcAddr():result;
1585 // Check: If isNaN() by checking result!=result? then either trap
1586 // or go to runtime
1587 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1588 // Build the boolean node
1589 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1590
1591 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1592 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1593 // The pow or exp intrinsic returned a NaN, which requires a call
1594 // to the runtime. Recompile with the runtime call.
1595 uncommon_trap(Deoptimization::Reason_intrinsic,
1596 Deoptimization::Action_make_not_entrant);
1597 }
1598 set_result(result);
1599 } else {
1600 // If this inlining ever returned NaN in the past, we compile a call
1601 // to the runtime to properly handle corner cases
1602
1603 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1604 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1605 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1606
1607 if (!if_slow->is_top()) {
1608 RegionNode* result_region = new (C) RegionNode(3);
1609 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1610
1611 result_region->init_req(1, if_fast);
1612 result_val->init_req(1, result);
1613
1614 set_control(if_slow);
1615
1616 const TypePtr* no_memory_effects = NULL;
1617 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1618 no_memory_effects,
1619 x, top(), y, y ? top() : NULL);
1620 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1621 #ifdef ASSERT
1622 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1623 assert(value_top == top(), "second value must be top");
1624 #endif
1625
1626 result_region->init_req(2, control());
1627 result_val->init_req(2, value);
1628 set_result(result_region, result_val);
1629 } else {
1630 set_result(result);
1631 }
1632 }
1633 }
1634
1635 //------------------------------inline_exp-------------------------------------
1636 // Inline exp instructions, if possible. The Intel hardware only misses
1637 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1638 bool LibraryCallKit::inline_exp() {
1639 Node* arg = round_double_node(argument(0));
1640 Node* n = _gvn.transform(new (C) ExpDNode(0, arg));
1641
1642 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1643
1644 C->set_has_split_ifs(true); // Has chance for split-if optimization
1645 return true;
1646 }
1647
1648 //------------------------------inline_pow-------------------------------------
1649 // Inline power instructions, if possible.
1650 bool LibraryCallKit::inline_pow() {
1651 // Pseudocode for pow
1652 // if (x <= 0.0) {
1653 // long longy = (long)y;
1654 // if ((double)longy == y) { // if y is long
1655 // if (y + 1 == y) longy = 0; // huge number: even
1656 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1657 // } else {
1658 // result = NaN;
1659 // }
1660 // } else {
1661 // result = DPow(x,y);
1662 // }
1663 // if (result != result)? {
1664 // result = uncommon_trap() or runtime_call();
1665 // }
1666 // return result;
1667
1668 Node* x = round_double_node(argument(0));
1669 Node* y = round_double_node(argument(2));
1670
1671 Node* result = NULL;
1672
1673 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1674 // Short form: skip the fancy tests and just check for NaN result.
1675 result = _gvn.transform(new (C) PowDNode(0, x, y));
1676 } else {
1677 // If this inlining ever returned NaN in the past, include all
1678 // checks + call to the runtime.
1679
1680 // Set the merge point for If node with condition of (x <= 0.0)
1681 // There are four possible paths to region node and phi node
1682 RegionNode *r = new (C) RegionNode(4);
1683 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1684
1685 // Build the first if node: if (x <= 0.0)
1686 // Node for 0 constant
1687 Node *zeronode = makecon(TypeD::ZERO);
1688 // Check x:0
1689 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1690 // Check: If (x<=0) then go complex path
1691 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1692 // Branch either way
1693 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1694 // Fast path taken; set region slot 3
1695 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1696 r->init_req(3,fast_taken); // Capture fast-control
1697
1698 // Fast path not-taken, i.e. slow path
1699 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1700
1701 // Set fast path result
1702 Node *fast_result = _gvn.transform( new (C) PowDNode(0, x, y) );
1703 phi->init_req(3, fast_result);
1704
1705 // Complex path
1706 // Build the second if node (if y is long)
1707 // Node for (long)y
1708 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1709 // Node for (double)((long) y)
1710 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1711 // Check (double)((long) y) : y
1712 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1713 // Check if (y isn't long) then go to slow path
1714
1715 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1716 // Branch either way
1717 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1718 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1719
1720 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1721
1722 // Calculate DPow(abs(x), y)*(1 & (long)y)
1723 // Node for constant 1
1724 Node *conone = longcon(1);
1725 // 1& (long)y
1726 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1727
1728 // A huge number is always even. Detect a huge number by checking
1729 // if y + 1 == y and set integer to be tested for parity to 0.
1730 // Required for corner case:
1731 // (long)9.223372036854776E18 = max_jlong
1732 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1733 // max_jlong is odd but 9.223372036854776E18 is even
1734 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1735 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1736 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1737 Node* correctedsign = NULL;
1738 if (ConditionalMoveLimit != 0) {
1739 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1740 } else {
1741 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1742 RegionNode *r = new (C) RegionNode(3);
1743 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1744 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1745 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1746 phi->init_req(1, signnode);
1747 phi->init_req(2, longcon(0));
1748 correctedsign = _gvn.transform(phi);
1749 ylong_path = _gvn.transform(r);
1750 record_for_igvn(r);
1751 }
1752
1753 // zero node
1754 Node *conzero = longcon(0);
1755 // Check (1&(long)y)==0?
1756 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1757 // Check if (1&(long)y)!=0?, if so the result is negative
1758 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1759 // abs(x)
1760 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1761 // abs(x)^y
1762 Node *absxpowy = _gvn.transform( new (C) PowDNode(0, absx, y) );
1763 // -abs(x)^y
1764 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1765 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1766 Node *signresult = NULL;
1767 if (ConditionalMoveLimit != 0) {
1768 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1769 } else {
1770 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1771 RegionNode *r = new (C) RegionNode(3);
1772 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1773 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1774 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1775 phi->init_req(1, absxpowy);
1776 phi->init_req(2, negabsxpowy);
1777 signresult = _gvn.transform(phi);
1778 ylong_path = _gvn.transform(r);
1779 record_for_igvn(r);
1780 }
1781 // Set complex path fast result
1782 r->init_req(2, ylong_path);
1783 phi->init_req(2, signresult);
1784
1785 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1786 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1787 r->init_req(1,slow_path);
1788 phi->init_req(1,slow_result);
1789
1790 // Post merge
1791 set_control(_gvn.transform(r));
1792 record_for_igvn(r);
1793 result = _gvn.transform(phi);
1794 }
1795
1796 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1797
1798 C->set_has_split_ifs(true); // Has chance for split-if optimization
1799 return true;
1800 }
1801
1802 //------------------------------runtime_math-----------------------------
1803 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1804 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1805 "must be (DD)D or (D)D type");
1806
1807 // Inputs
1808 Node* a = round_double_node(argument(0));
1809 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1810
1811 const TypePtr* no_memory_effects = NULL;
1812 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1813 no_memory_effects,
1814 a, top(), b, b ? top() : NULL);
1815 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1816 #ifdef ASSERT
1817 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1818 assert(value_top == top(), "second value must be top");
1819 #endif
1820
1821 set_result(value);
1822 return true;
1823 }
1824
1825 //------------------------------inline_math_native-----------------------------
1826 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1827 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1828 switch (id) {
1829 // These intrinsics are not properly supported on all hardware
1830 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1831 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1832 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1833 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1834 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1835 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1836
1837 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1838 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1839 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1840 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1841
1842 // These intrinsics are supported on all hardware
1843 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1844 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1845
1846 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1847 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1848 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1849 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1850 #undef FN_PTR
1851
1852 // These intrinsics are not yet correctly implemented
1853 case vmIntrinsics::_datan2:
1854 return false;
1855
1856 default:
1857 fatal_unexpected_iid(id);
1858 return false;
1859 }
1860 }
1861
1862 static bool is_simple_name(Node* n) {
1863 return (n->req() == 1 // constant
1864 || (n->is_Type() && n->as_Type()->type()->singleton())
1865 || n->is_Proj() // parameter or return value
1866 || n->is_Phi() // local of some sort
1867 );
1868 }
1869
1870 //----------------------------inline_min_max-----------------------------------
1871 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1872 set_result(generate_min_max(id, argument(0), argument(1)));
1873 return true;
1874 }
1875
1876 Node*
1877 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1878 // These are the candidate return value:
1879 Node* xvalue = x0;
1880 Node* yvalue = y0;
1881
1882 if (xvalue == yvalue) {
1883 return xvalue;
1884 }
1885
1886 bool want_max = (id == vmIntrinsics::_max);
1887
1888 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1889 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1890 if (txvalue == NULL || tyvalue == NULL) return top();
1891 // This is not really necessary, but it is consistent with a
1892 // hypothetical MaxINode::Value method:
1893 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1894
1895 // %%% This folding logic should (ideally) be in a different place.
1896 // Some should be inside IfNode, and there to be a more reliable
1897 // transformation of ?: style patterns into cmoves. We also want
1898 // more powerful optimizations around cmove and min/max.
1899
1900 // Try to find a dominating comparison of these guys.
1901 // It can simplify the index computation for Arrays.copyOf
1902 // and similar uses of System.arraycopy.
1903 // First, compute the normalized version of CmpI(x, y).
1904 int cmp_op = Op_CmpI;
1905 Node* xkey = xvalue;
1906 Node* ykey = yvalue;
1907 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1908 if (ideal_cmpxy->is_Cmp()) {
1909 // E.g., if we have CmpI(length - offset, count),
1910 // it might idealize to CmpI(length, count + offset)
1911 cmp_op = ideal_cmpxy->Opcode();
1912 xkey = ideal_cmpxy->in(1);
1913 ykey = ideal_cmpxy->in(2);
1914 }
1915
1916 // Start by locating any relevant comparisons.
1917 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1918 Node* cmpxy = NULL;
1919 Node* cmpyx = NULL;
1920 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1921 Node* cmp = start_from->fast_out(k);
1922 if (cmp->outcnt() > 0 && // must have prior uses
1923 cmp->in(0) == NULL && // must be context-independent
1924 cmp->Opcode() == cmp_op) { // right kind of compare
1925 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1926 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1927 }
1928 }
1929
1930 const int NCMPS = 2;
1931 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1932 int cmpn;
1933 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1934 if (cmps[cmpn] != NULL) break; // find a result
1935 }
1936 if (cmpn < NCMPS) {
1937 // Look for a dominating test that tells us the min and max.
1938 int depth = 0; // Limit search depth for speed
1939 Node* dom = control();
1940 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1941 if (++depth >= 100) break;
1942 Node* ifproj = dom;
1943 if (!ifproj->is_Proj()) continue;
1944 Node* iff = ifproj->in(0);
1945 if (!iff->is_If()) continue;
1946 Node* bol = iff->in(1);
1947 if (!bol->is_Bool()) continue;
1948 Node* cmp = bol->in(1);
1949 if (cmp == NULL) continue;
1950 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1951 if (cmps[cmpn] == cmp) break;
1952 if (cmpn == NCMPS) continue;
1953 BoolTest::mask btest = bol->as_Bool()->_test._test;
1954 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1955 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1956 // At this point, we know that 'x btest y' is true.
1957 switch (btest) {
1958 case BoolTest::eq:
1959 // They are proven equal, so we can collapse the min/max.
1960 // Either value is the answer. Choose the simpler.
1961 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1962 return yvalue;
1963 return xvalue;
1964 case BoolTest::lt: // x < y
1965 case BoolTest::le: // x <= y
1966 return (want_max ? yvalue : xvalue);
1967 case BoolTest::gt: // x > y
1968 case BoolTest::ge: // x >= y
1969 return (want_max ? xvalue : yvalue);
1970 }
1971 }
1972 }
1973
1974 // We failed to find a dominating test.
1975 // Let's pick a test that might GVN with prior tests.
1976 Node* best_bol = NULL;
1977 BoolTest::mask best_btest = BoolTest::illegal;
1978 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1979 Node* cmp = cmps[cmpn];
1980 if (cmp == NULL) continue;
1981 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1982 Node* bol = cmp->fast_out(j);
1983 if (!bol->is_Bool()) continue;
1984 BoolTest::mask btest = bol->as_Bool()->_test._test;
1985 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1986 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1987 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1988 best_bol = bol->as_Bool();
1989 best_btest = btest;
1990 }
1991 }
1992 }
1993
1994 Node* answer_if_true = NULL;
1995 Node* answer_if_false = NULL;
1996 switch (best_btest) {
1997 default:
1998 if (cmpxy == NULL)
1999 cmpxy = ideal_cmpxy;
2000 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2001 // and fall through:
2002 case BoolTest::lt: // x < y
2003 case BoolTest::le: // x <= y
2004 answer_if_true = (want_max ? yvalue : xvalue);
2005 answer_if_false = (want_max ? xvalue : yvalue);
2006 break;
2007 case BoolTest::gt: // x > y
2008 case BoolTest::ge: // x >= y
2009 answer_if_true = (want_max ? xvalue : yvalue);
2010 answer_if_false = (want_max ? yvalue : xvalue);
2011 break;
2012 }
2013
2014 jint hi, lo;
2015 if (want_max) {
2016 // We can sharpen the minimum.
2017 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2018 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2019 } else {
2020 // We can sharpen the maximum.
2021 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2022 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2023 }
2024
2025 // Use a flow-free graph structure, to avoid creating excess control edges
2026 // which could hinder other optimizations.
2027 // Since Math.min/max is often used with arraycopy, we want
2028 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2029 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2030 answer_if_false, answer_if_true,
2031 TypeInt::make(lo, hi, widen));
2032
2033 return _gvn.transform(cmov);
2034
2035 /*
2036 // This is not as desirable as it may seem, since Min and Max
2037 // nodes do not have a full set of optimizations.
2038 // And they would interfere, anyway, with 'if' optimizations
2039 // and with CMoveI canonical forms.
2040 switch (id) {
2041 case vmIntrinsics::_min:
2042 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2043 case vmIntrinsics::_max:
2044 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2045 default:
2046 ShouldNotReachHere();
2047 }
2048 */
2049 }
2050
2051 inline int
2052 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2053 const TypePtr* base_type = TypePtr::NULL_PTR;
2054 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2055 if (base_type == NULL) {
2056 // Unknown type.
2057 return Type::AnyPtr;
2058 } else if (base_type == TypePtr::NULL_PTR) {
2059 // Since this is a NULL+long form, we have to switch to a rawptr.
2060 base = _gvn.transform( new (C) CastX2PNode(offset) );
2061 offset = MakeConX(0);
2062 return Type::RawPtr;
2063 } else if (base_type->base() == Type::RawPtr) {
2064 return Type::RawPtr;
2065 } else if (base_type->isa_oopptr()) {
2066 // Base is never null => always a heap address.
2067 if (base_type->ptr() == TypePtr::NotNull) {
2068 return Type::OopPtr;
2069 }
2070 // Offset is small => always a heap address.
2071 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2072 if (offset_type != NULL &&
2073 base_type->offset() == 0 && // (should always be?)
2074 offset_type->_lo >= 0 &&
2075 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2076 return Type::OopPtr;
2077 }
2078 // Otherwise, it might either be oop+off or NULL+addr.
2079 return Type::AnyPtr;
2080 } else {
2081 // No information:
2082 return Type::AnyPtr;
2083 }
2084 }
2085
2086 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2087 int kind = classify_unsafe_addr(base, offset);
2088 if (kind == Type::RawPtr) {
2089 return basic_plus_adr(top(), base, offset);
2090 } else {
2091 return basic_plus_adr(base, offset);
2092 }
2093 }
2094
2095 //--------------------------inline_number_methods-----------------------------
2096 // inline int Integer.numberOfLeadingZeros(int)
2097 // inline int Long.numberOfLeadingZeros(long)
2098 //
2099 // inline int Integer.numberOfTrailingZeros(int)
2100 // inline int Long.numberOfTrailingZeros(long)
2101 //
2102 // inline int Integer.bitCount(int)
2103 // inline int Long.bitCount(long)
2104 //
2105 // inline char Character.reverseBytes(char)
2106 // inline short Short.reverseBytes(short)
2107 // inline int Integer.reverseBytes(int)
2108 // inline long Long.reverseBytes(long)
2109 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2110 Node* arg = argument(0);
2111 Node* n;
2112 switch (id) {
2113 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2114 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2115 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2116 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2117 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2118 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2119 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2120 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2121 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2122 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2123 default: fatal_unexpected_iid(id); break;
2124 }
2125 set_result(_gvn.transform(n));
2126 return true;
2127 }
2128
2129 //----------------------------inline_unsafe_access----------------------------
2130
2131 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2132
2133 // Helper that guards and inserts a pre-barrier.
2134 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2135 Node* pre_val, bool need_mem_bar) {
2136 // We could be accessing the referent field of a reference object. If so, when G1
2137 // is enabled, we need to log the value in the referent field in an SATB buffer.
2138 // This routine performs some compile time filters and generates suitable
2139 // runtime filters that guard the pre-barrier code.
2140 // Also add memory barrier for non volatile load from the referent field
2141 // to prevent commoning of loads across safepoint.
2142 if (!UseG1GC && !need_mem_bar)
2143 return;
2144
2145 // Some compile time checks.
2146
2147 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2148 const TypeX* otype = offset->find_intptr_t_type();
2149 if (otype != NULL && otype->is_con() &&
2150 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2151 // Constant offset but not the reference_offset so just return
2152 return;
2153 }
2154
2155 // We only need to generate the runtime guards for instances.
2156 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2157 if (btype != NULL) {
2158 if (btype->isa_aryptr()) {
2159 // Array type so nothing to do
2160 return;
2161 }
2162
2163 const TypeInstPtr* itype = btype->isa_instptr();
2164 if (itype != NULL) {
2165 // Can the klass of base_oop be statically determined to be
2166 // _not_ a sub-class of Reference and _not_ Object?
2167 ciKlass* klass = itype->klass();
2168 if ( klass->is_loaded() &&
2169 !klass->is_subtype_of(env()->Reference_klass()) &&
2170 !env()->Object_klass()->is_subtype_of(klass)) {
2171 return;
2172 }
2173 }
2174 }
2175
2176 // The compile time filters did not reject base_oop/offset so
2177 // we need to generate the following runtime filters
2178 //
2179 // if (offset == java_lang_ref_Reference::_reference_offset) {
2180 // if (instance_of(base, java.lang.ref.Reference)) {
2181 // pre_barrier(_, pre_val, ...);
2182 // }
2183 // }
2184
2185 float likely = PROB_LIKELY( 0.999);
2186 float unlikely = PROB_UNLIKELY(0.999);
2187
2188 IdealKit ideal(this);
2189 #define __ ideal.
2190
2191 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2192
2193 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2194 // Update graphKit memory and control from IdealKit.
2195 sync_kit(ideal);
2196
2197 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2198 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2199
2200 // Update IdealKit memory and control from graphKit.
2201 __ sync_kit(this);
2202
2203 Node* one = __ ConI(1);
2204 // is_instof == 0 if base_oop == NULL
2205 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2206
2207 // Update graphKit from IdeakKit.
2208 sync_kit(ideal);
2209
2210 // Use the pre-barrier to record the value in the referent field
2211 pre_barrier(false /* do_load */,
2212 __ ctrl(),
2213 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2214 pre_val /* pre_val */,
2215 T_OBJECT);
2216 if (need_mem_bar) {
2217 // Add memory barrier to prevent commoning reads from this field
2218 // across safepoint since GC can change its value.
2219 insert_mem_bar(Op_MemBarCPUOrder);
2220 }
2221 // Update IdealKit from graphKit.
2222 __ sync_kit(this);
2223
2224 } __ end_if(); // _ref_type != ref_none
2225 } __ end_if(); // offset == referent_offset
2226
2227 // Final sync IdealKit and GraphKit.
2228 final_sync(ideal);
2229 #undef __
2230 }
2231
2232
2233 // Interpret Unsafe.fieldOffset cookies correctly:
2234 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2235
2236 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2237 // Attempt to infer a sharper value type from the offset and base type.
2238 ciKlass* sharpened_klass = NULL;
2239
2240 // See if it is an instance field, with an object type.
2241 if (alias_type->field() != NULL) {
2242 assert(!is_native_ptr, "native pointer op cannot use a java address");
2243 if (alias_type->field()->type()->is_klass()) {
2244 sharpened_klass = alias_type->field()->type()->as_klass();
2245 }
2246 }
2247
2248 // See if it is a narrow oop array.
2249 if (adr_type->isa_aryptr()) {
2250 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2251 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2252 if (elem_type != NULL) {
2253 sharpened_klass = elem_type->klass();
2254 }
2255 }
2256 }
2257
2258 // The sharpened class might be unloaded if there is no class loader
2259 // contraint in place.
2260 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2261 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2262
2263 #ifndef PRODUCT
2264 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2265 tty->print(" from base type: "); adr_type->dump();
2266 tty->print(" sharpened value: "); tjp->dump();
2267 }
2268 #endif
2269 // Sharpen the value type.
2270 return tjp;
2271 }
2272 return NULL;
2273 }
2274
2275 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2276 if (callee()->is_static()) return false; // caller must have the capability!
2277
2278 #ifndef PRODUCT
2279 {
2280 ResourceMark rm;
2281 // Check the signatures.
2282 ciSignature* sig = callee()->signature();
2283 #ifdef ASSERT
2284 if (!is_store) {
2285 // Object getObject(Object base, int/long offset), etc.
2286 BasicType rtype = sig->return_type()->basic_type();
2287 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2288 rtype = T_ADDRESS; // it is really a C void*
2289 assert(rtype == type, "getter must return the expected value");
2290 if (!is_native_ptr) {
2291 assert(sig->count() == 2, "oop getter has 2 arguments");
2292 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2293 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2294 } else {
2295 assert(sig->count() == 1, "native getter has 1 argument");
2296 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2297 }
2298 } else {
2299 // void putObject(Object base, int/long offset, Object x), etc.
2300 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2301 if (!is_native_ptr) {
2302 assert(sig->count() == 3, "oop putter has 3 arguments");
2303 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2304 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2305 } else {
2306 assert(sig->count() == 2, "native putter has 2 arguments");
2307 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2308 }
2309 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2310 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2311 vtype = T_ADDRESS; // it is really a C void*
2312 assert(vtype == type, "putter must accept the expected value");
2313 }
2314 #endif // ASSERT
2315 }
2316 #endif //PRODUCT
2317
2318 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2319
2320 Node* receiver = argument(0); // type: oop
2321
2322 // Build address expression. See the code in inline_unsafe_prefetch.
2323 Node* adr;
2324 Node* heap_base_oop = top();
2325 Node* offset = top();
2326 Node* val;
2327
2328 if (!is_native_ptr) {
2329 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2330 Node* base = argument(1); // type: oop
2331 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2332 offset = argument(2); // type: long
2333 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2334 // to be plain byte offsets, which are also the same as those accepted
2335 // by oopDesc::field_base.
2336 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2337 "fieldOffset must be byte-scaled");
2338 // 32-bit machines ignore the high half!
2339 offset = ConvL2X(offset);
2340 adr = make_unsafe_address(base, offset);
2341 heap_base_oop = base;
2342 val = is_store ? argument(4) : NULL;
2343 } else {
2344 Node* ptr = argument(1); // type: long
2345 ptr = ConvL2X(ptr); // adjust Java long to machine word
2346 adr = make_unsafe_address(NULL, ptr);
2347 val = is_store ? argument(3) : NULL;
2348 }
2349
2350 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2351
2352 // First guess at the value type.
2353 const Type *value_type = Type::get_const_basic_type(type);
2354
2355 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2356 // there was not enough information to nail it down.
2357 Compile::AliasType* alias_type = C->alias_type(adr_type);
2358 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2359
2360 // We will need memory barriers unless we can determine a unique
2361 // alias category for this reference. (Note: If for some reason
2362 // the barriers get omitted and the unsafe reference begins to "pollute"
2363 // the alias analysis of the rest of the graph, either Compile::can_alias
2364 // or Compile::must_alias will throw a diagnostic assert.)
2365 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2366
2367 // If we are reading the value of the referent field of a Reference
2368 // object (either by using Unsafe directly or through reflection)
2369 // then, if G1 is enabled, we need to record the referent in an
2370 // SATB log buffer using the pre-barrier mechanism.
2371 // Also we need to add memory barrier to prevent commoning reads
2372 // from this field across safepoint since GC can change its value.
2373 bool need_read_barrier = !is_native_ptr && !is_store &&
2374 offset != top() && heap_base_oop != top();
2375
2376 if (!is_store && type == T_OBJECT) {
2377 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2378 if (tjp != NULL) {
2379 value_type = tjp;
2380 }
2381 }
2382
2383 receiver = null_check(receiver);
2384 if (stopped()) {
2385 return true;
2386 }
2387 // Heap pointers get a null-check from the interpreter,
2388 // as a courtesy. However, this is not guaranteed by Unsafe,
2389 // and it is not possible to fully distinguish unintended nulls
2390 // from intended ones in this API.
2391
2392 if (is_volatile) {
2393 // We need to emit leading and trailing CPU membars (see below) in
2394 // addition to memory membars when is_volatile. This is a little
2395 // too strong, but avoids the need to insert per-alias-type
2396 // volatile membars (for stores; compare Parse::do_put_xxx), which
2397 // we cannot do effectively here because we probably only have a
2398 // rough approximation of type.
2399 need_mem_bar = true;
2400 // For Stores, place a memory ordering barrier now.
2401 if (is_store)
2402 insert_mem_bar(Op_MemBarRelease);
2403 }
2404
2405 // Memory barrier to prevent normal and 'unsafe' accesses from
2406 // bypassing each other. Happens after null checks, so the
2407 // exception paths do not take memory state from the memory barrier,
2408 // so there's no problems making a strong assert about mixing users
2409 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2410 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2411 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2412
2413 if (!is_store) {
2414 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2415 // load value
2416 switch (type) {
2417 case T_BOOLEAN:
2418 case T_CHAR:
2419 case T_BYTE:
2420 case T_SHORT:
2421 case T_INT:
2422 case T_LONG:
2423 case T_FLOAT:
2424 case T_DOUBLE:
2425 break;
2426 case T_OBJECT:
2427 if (need_read_barrier) {
2428 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2429 }
2430 break;
2431 case T_ADDRESS:
2432 // Cast to an int type.
2433 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2434 p = ConvX2L(p);
2435 break;
2436 default:
2437 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2438 break;
2439 }
2440 // The load node has the control of the preceding MemBarCPUOrder. All
2441 // following nodes will have the control of the MemBarCPUOrder inserted at
2442 // the end of this method. So, pushing the load onto the stack at a later
2443 // point is fine.
2444 set_result(p);
2445 } else {
2446 // place effect of store into memory
2447 switch (type) {
2448 case T_DOUBLE:
2449 val = dstore_rounding(val);
2450 break;
2451 case T_ADDRESS:
2452 // Repackage the long as a pointer.
2453 val = ConvL2X(val);
2454 val = _gvn.transform( new (C) CastX2PNode(val) );
2455 break;
2456 }
2457
2458 if (type != T_OBJECT ) {
2459 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2460 } else {
2461 // Possibly an oop being stored to Java heap or native memory
2462 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2463 // oop to Java heap.
2464 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2465 } else {
2466 // We can't tell at compile time if we are storing in the Java heap or outside
2467 // of it. So we need to emit code to conditionally do the proper type of
2468 // store.
2469
2470 IdealKit ideal(this);
2471 #define __ ideal.
2472 // QQQ who knows what probability is here??
2473 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2474 // Sync IdealKit and graphKit.
2475 sync_kit(ideal);
2476 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2477 // Update IdealKit memory.
2478 __ sync_kit(this);
2479 } __ else_(); {
2480 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2481 } __ end_if();
2482 // Final sync IdealKit and GraphKit.
2483 final_sync(ideal);
2484 #undef __
2485 }
2486 }
2487 }
2488
2489 if (is_volatile) {
2490 if (!is_store)
2491 insert_mem_bar(Op_MemBarAcquire);
2492 else
2493 insert_mem_bar(Op_MemBarVolatile);
2494 }
2495
2496 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2497
2498 return true;
2499 }
2500
2501 //----------------------------inline_unsafe_prefetch----------------------------
2502
2503 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2504 #ifndef PRODUCT
2505 {
2506 ResourceMark rm;
2507 // Check the signatures.
2508 ciSignature* sig = callee()->signature();
2509 #ifdef ASSERT
2510 // Object getObject(Object base, int/long offset), etc.
2511 BasicType rtype = sig->return_type()->basic_type();
2512 if (!is_native_ptr) {
2513 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2514 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2515 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2516 } else {
2517 assert(sig->count() == 1, "native prefetch has 1 argument");
2518 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2519 }
2520 #endif // ASSERT
2521 }
2522 #endif // !PRODUCT
2523
2524 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2525
2526 const int idx = is_static ? 0 : 1;
2527 if (!is_static) {
2528 null_check_receiver();
2529 if (stopped()) {
2530 return true;
2531 }
2532 }
2533
2534 // Build address expression. See the code in inline_unsafe_access.
2535 Node *adr;
2536 if (!is_native_ptr) {
2537 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2538 Node* base = argument(idx + 0); // type: oop
2539 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2540 Node* offset = argument(idx + 1); // type: long
2541 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2542 // to be plain byte offsets, which are also the same as those accepted
2543 // by oopDesc::field_base.
2544 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2545 "fieldOffset must be byte-scaled");
2546 // 32-bit machines ignore the high half!
2547 offset = ConvL2X(offset);
2548 adr = make_unsafe_address(base, offset);
2549 } else {
2550 Node* ptr = argument(idx + 0); // type: long
2551 ptr = ConvL2X(ptr); // adjust Java long to machine word
2552 adr = make_unsafe_address(NULL, ptr);
2553 }
2554
2555 // Generate the read or write prefetch
2556 Node *prefetch;
2557 if (is_store) {
2558 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2559 } else {
2560 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2561 }
2562 prefetch->init_req(0, control());
2563 set_i_o(_gvn.transform(prefetch));
2564
2565 return true;
2566 }
2567
2568 //----------------------------inline_unsafe_load_store----------------------------
2569 // This method serves a couple of different customers (depending on LoadStoreKind):
2570 //
2571 // LS_cmpxchg:
2572 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2573 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2574 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2575 //
2576 // LS_xadd:
2577 // public int getAndAddInt( Object o, long offset, int delta)
2578 // public long getAndAddLong(Object o, long offset, long delta)
2579 //
2580 // LS_xchg:
2581 // int getAndSet(Object o, long offset, int newValue)
2582 // long getAndSet(Object o, long offset, long newValue)
2583 // Object getAndSet(Object o, long offset, Object newValue)
2584 //
2585 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2586 // This basic scheme here is the same as inline_unsafe_access, but
2587 // differs in enough details that combining them would make the code
2588 // overly confusing. (This is a true fact! I originally combined
2589 // them, but even I was confused by it!) As much code/comments as
2590 // possible are retained from inline_unsafe_access though to make
2591 // the correspondences clearer. - dl
2592
2593 if (callee()->is_static()) return false; // caller must have the capability!
2594
2595 #ifndef PRODUCT
2596 BasicType rtype;
2597 {
2598 ResourceMark rm;
2599 // Check the signatures.
2600 ciSignature* sig = callee()->signature();
2601 rtype = sig->return_type()->basic_type();
2602 if (kind == LS_xadd || kind == LS_xchg) {
2603 // Check the signatures.
2604 #ifdef ASSERT
2605 assert(rtype == type, "get and set must return the expected type");
2606 assert(sig->count() == 3, "get and set has 3 arguments");
2607 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2608 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2609 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2610 #endif // ASSERT
2611 } else if (kind == LS_cmpxchg) {
2612 // Check the signatures.
2613 #ifdef ASSERT
2614 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2615 assert(sig->count() == 4, "CAS has 4 arguments");
2616 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2617 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2618 #endif // ASSERT
2619 } else {
2620 ShouldNotReachHere();
2621 }
2622 }
2623 #endif //PRODUCT
2624
2625 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2626
2627 // Get arguments:
2628 Node* receiver = NULL;
2629 Node* base = NULL;
2630 Node* offset = NULL;
2631 Node* oldval = NULL;
2632 Node* newval = NULL;
2633 if (kind == LS_cmpxchg) {
2634 const bool two_slot_type = type2size[type] == 2;
2635 receiver = argument(0); // type: oop
2636 base = argument(1); // type: oop
2637 offset = argument(2); // type: long
2638 oldval = argument(4); // type: oop, int, or long
2639 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2640 } else if (kind == LS_xadd || kind == LS_xchg){
2641 receiver = argument(0); // type: oop
2642 base = argument(1); // type: oop
2643 offset = argument(2); // type: long
2644 oldval = NULL;
2645 newval = argument(4); // type: oop, int, or long
2646 }
2647
2648 // Null check receiver.
2649 receiver = null_check(receiver);
2650 if (stopped()) {
2651 return true;
2652 }
2653
2654 // Build field offset expression.
2655 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2656 // to be plain byte offsets, which are also the same as those accepted
2657 // by oopDesc::field_base.
2658 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2659 // 32-bit machines ignore the high half of long offsets
2660 offset = ConvL2X(offset);
2661 Node* adr = make_unsafe_address(base, offset);
2662 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2663
2664 // For CAS, unlike inline_unsafe_access, there seems no point in
2665 // trying to refine types. Just use the coarse types here.
2666 const Type *value_type = Type::get_const_basic_type(type);
2667 Compile::AliasType* alias_type = C->alias_type(adr_type);
2668 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2669
2670 if (kind == LS_xchg && type == T_OBJECT) {
2671 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2672 if (tjp != NULL) {
2673 value_type = tjp;
2674 }
2675 }
2676
2677 int alias_idx = C->get_alias_index(adr_type);
2678
2679 // Memory-model-wise, a LoadStore acts like a little synchronized
2680 // block, so needs barriers on each side. These don't translate
2681 // into actual barriers on most machines, but we still need rest of
2682 // compiler to respect ordering.
2683
2684 insert_mem_bar(Op_MemBarRelease);
2685 insert_mem_bar(Op_MemBarCPUOrder);
2686
2687 // 4984716: MemBars must be inserted before this
2688 // memory node in order to avoid a false
2689 // dependency which will confuse the scheduler.
2690 Node *mem = memory(alias_idx);
2691
2692 // For now, we handle only those cases that actually exist: ints,
2693 // longs, and Object. Adding others should be straightforward.
2694 Node* load_store;
2695 switch(type) {
2696 case T_INT:
2697 if (kind == LS_xadd) {
2698 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2699 } else if (kind == LS_xchg) {
2700 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2701 } else if (kind == LS_cmpxchg) {
2702 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2703 } else {
2704 ShouldNotReachHere();
2705 }
2706 break;
2707 case T_LONG:
2708 if (kind == LS_xadd) {
2709 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2710 } else if (kind == LS_xchg) {
2711 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2712 } else if (kind == LS_cmpxchg) {
2713 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2714 } else {
2715 ShouldNotReachHere();
2716 }
2717 break;
2718 case T_OBJECT:
2719 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2720 // could be delayed during Parse (for example, in adjust_map_after_if()).
2721 // Execute transformation here to avoid barrier generation in such case.
2722 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2723 newval = _gvn.makecon(TypePtr::NULL_PTR);
2724
2725 // Reference stores need a store barrier.
2726 pre_barrier(true /* do_load*/,
2727 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2728 NULL /* pre_val*/,
2729 T_OBJECT);
2730 #ifdef _LP64
2731 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2732 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2733 if (kind == LS_xchg) {
2734 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2735 newval_enc, adr_type, value_type->make_narrowoop()));
2736 } else {
2737 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2738 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2739 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2740 newval_enc, oldval_enc));
2741 }
2742 } else
2743 #endif
2744 {
2745 if (kind == LS_xchg) {
2746 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2747 } else {
2748 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2749 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2750 }
2751 }
2752 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2753 break;
2754 default:
2755 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2756 break;
2757 }
2758
2759 // SCMemProjNodes represent the memory state of a LoadStore. Their
2760 // main role is to prevent LoadStore nodes from being optimized away
2761 // when their results aren't used.
2762 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2763 set_memory(proj, alias_idx);
2764
2765 // Add the trailing membar surrounding the access
2766 insert_mem_bar(Op_MemBarCPUOrder);
2767 insert_mem_bar(Op_MemBarAcquire);
2768
2769 #ifdef _LP64
2770 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2771 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
2772 }
2773 #endif
2774
2775 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2776 set_result(load_store);
2777 return true;
2778 }
2779
2780 //----------------------------inline_unsafe_ordered_store----------------------
2781 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2782 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2783 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2784 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2785 // This is another variant of inline_unsafe_access, differing in
2786 // that it always issues store-store ("release") barrier and ensures
2787 // store-atomicity (which only matters for "long").
2788
2789 if (callee()->is_static()) return false; // caller must have the capability!
2790
2791 #ifndef PRODUCT
2792 {
2793 ResourceMark rm;
2794 // Check the signatures.
2795 ciSignature* sig = callee()->signature();
2796 #ifdef ASSERT
2797 BasicType rtype = sig->return_type()->basic_type();
2798 assert(rtype == T_VOID, "must return void");
2799 assert(sig->count() == 3, "has 3 arguments");
2800 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2801 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2802 #endif // ASSERT
2803 }
2804 #endif //PRODUCT
2805
2806 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2807
2808 // Get arguments:
2809 Node* receiver = argument(0); // type: oop
2810 Node* base = argument(1); // type: oop
2811 Node* offset = argument(2); // type: long
2812 Node* val = argument(4); // type: oop, int, or long
2813
2814 // Null check receiver.
2815 receiver = null_check(receiver);
2816 if (stopped()) {
2817 return true;
2818 }
2819
2820 // Build field offset expression.
2821 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2822 // 32-bit machines ignore the high half of long offsets
2823 offset = ConvL2X(offset);
2824 Node* adr = make_unsafe_address(base, offset);
2825 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2826 const Type *value_type = Type::get_const_basic_type(type);
2827 Compile::AliasType* alias_type = C->alias_type(adr_type);
2828
2829 insert_mem_bar(Op_MemBarRelease);
2830 insert_mem_bar(Op_MemBarCPUOrder);
2831 // Ensure that the store is atomic for longs:
2832 const bool require_atomic_access = true;
2833 Node* store;
2834 if (type == T_OBJECT) // reference stores need a store barrier.
2835 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2836 else {
2837 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2838 }
2839 insert_mem_bar(Op_MemBarCPUOrder);
2840 return true;
2841 }
2842
2843 //----------------------------inline_unsafe_allocate---------------------------
2844 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2845 bool LibraryCallKit::inline_unsafe_allocate() {
2846 if (callee()->is_static()) return false; // caller must have the capability!
2847
2848 null_check_receiver(); // null-check, then ignore
2849 Node* cls = null_check(argument(1));
2850 if (stopped()) return true;
2851
2852 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2853 kls = null_check(kls);
2854 if (stopped()) return true; // argument was like int.class
2855
2856 // Note: The argument might still be an illegal value like
2857 // Serializable.class or Object[].class. The runtime will handle it.
2858 // But we must make an explicit check for initialization.
2859 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2860 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2861 // can generate code to load it as unsigned byte.
2862 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2863 Node* bits = intcon(InstanceKlass::fully_initialized);
2864 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2865 // The 'test' is non-zero if we need to take a slow path.
2866
2867 Node* obj = new_instance(kls, test);
2868 set_result(obj);
2869 return true;
2870 }
2871
2872 #ifdef TRACE_HAVE_INTRINSICS
2873 /*
2874 * oop -> myklass
2875 * myklass->trace_id |= USED
2876 * return myklass->trace_id & ~0x3
2877 */
2878 bool LibraryCallKit::inline_native_classID() {
2879 null_check_receiver(); // null-check, then ignore
2880 Node* cls = null_check(argument(1), T_OBJECT);
2881 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2882 kls = null_check(kls, T_OBJECT);
2883 ByteSize offset = TRACE_ID_OFFSET;
2884 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2885 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2886 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2887 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2888 Node* clsused = longcon(0x01l); // set the class bit
2889 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2890
2891 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2892 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2893 set_result(andl);
2894 return true;
2895 }
2896
2897 bool LibraryCallKit::inline_native_threadID() {
2898 Node* tls_ptr = NULL;
2899 Node* cur_thr = generate_current_thread(tls_ptr);
2900 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2901 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2902 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2903
2904 Node* threadid = NULL;
2905 size_t thread_id_size = OSThread::thread_id_size();
2906 if (thread_id_size == (size_t) BytesPerLong) {
2907 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2908 } else if (thread_id_size == (size_t) BytesPerInt) {
2909 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2910 } else {
2911 ShouldNotReachHere();
2912 }
2913 set_result(threadid);
2914 return true;
2915 }
2916 #endif
2917
2918 //------------------------inline_native_time_funcs--------------
2919 // inline code for System.currentTimeMillis() and System.nanoTime()
2920 // these have the same type and signature
2921 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2922 const TypeFunc* tf = OptoRuntime::void_long_Type();
2923 const TypePtr* no_memory_effects = NULL;
2924 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2925 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2926 #ifdef ASSERT
2927 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2928 assert(value_top == top(), "second value must be top");
2929 #endif
2930 set_result(value);
2931 return true;
2932 }
2933
2934 //------------------------inline_native_currentThread------------------
2935 bool LibraryCallKit::inline_native_currentThread() {
2936 Node* junk = NULL;
2937 set_result(generate_current_thread(junk));
2938 return true;
2939 }
2940
2941 //------------------------inline_native_isInterrupted------------------
2942 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2943 bool LibraryCallKit::inline_native_isInterrupted() {
2944 // Add a fast path to t.isInterrupted(clear_int):
2945 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2946 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2947 // So, in the common case that the interrupt bit is false,
2948 // we avoid making a call into the VM. Even if the interrupt bit
2949 // is true, if the clear_int argument is false, we avoid the VM call.
2950 // However, if the receiver is not currentThread, we must call the VM,
2951 // because there must be some locking done around the operation.
2952
2953 // We only go to the fast case code if we pass two guards.
2954 // Paths which do not pass are accumulated in the slow_region.
2955 RegionNode* slow_region = new (C) RegionNode(1);
2956 record_for_igvn(slow_region);
2957 RegionNode* result_rgn = new (C) RegionNode(1+3); // fast1, fast2, slow
2958 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
2959 enum { no_int_result_path = 1,
2960 no_clear_result_path = 2,
2961 slow_result_path = 3
2962 };
2963
2964 // (a) Receiving thread must be the current thread.
2965 Node* rec_thr = argument(0);
2966 Node* tls_ptr = NULL;
2967 Node* cur_thr = generate_current_thread(tls_ptr);
2968 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
2969 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
2970
2971 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2972 if (!known_current_thread)
2973 generate_slow_guard(bol_thr, slow_region);
2974
2975 // (b) Interrupt bit on TLS must be false.
2976 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2977 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2978 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2979 // Set the control input on the field _interrupted read to prevent it floating up.
2980 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2981 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
2982 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
2983
2984 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2985
2986 // First fast path: if (!TLS._interrupted) return false;
2987 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
2988 result_rgn->init_req(no_int_result_path, false_bit);
2989 result_val->init_req(no_int_result_path, intcon(0));
2990
2991 // drop through to next case
2992 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
2993
2994 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2995 Node* clr_arg = argument(1);
2996 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
2997 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
2998 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2999
3000 // Second fast path: ... else if (!clear_int) return true;
3001 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3002 result_rgn->init_req(no_clear_result_path, false_arg);
3003 result_val->init_req(no_clear_result_path, intcon(1));
3004
3005 // drop through to next case
3006 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3007
3008 // (d) Otherwise, go to the slow path.
3009 slow_region->add_req(control());
3010 set_control( _gvn.transform(slow_region) );
3011
3012 if (stopped()) {
3013 // There is no slow path.
3014 result_rgn->init_req(slow_result_path, top());
3015 result_val->init_req(slow_result_path, top());
3016 } else {
3017 // non-virtual because it is a private non-static
3018 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3019
3020 Node* slow_val = set_results_for_java_call(slow_call);
3021 // this->control() comes from set_results_for_java_call
3022
3023 // If we know that the result of the slow call will be true, tell the optimizer!
3024 if (known_current_thread) slow_val = intcon(1);
3025
3026 Node* fast_io = slow_call->in(TypeFunc::I_O);
3027 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3028 // These two phis are pre-filled with copies of of the fast IO and Memory
3029 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3030 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3031
3032 result_rgn->init_req(slow_result_path, control());
3033 io_phi ->init_req(slow_result_path, i_o());
3034 mem_phi ->init_req(slow_result_path, reset_memory());
3035 result_val->init_req(slow_result_path, slow_val);
3036
3037 set_all_memory( _gvn.transform(mem_phi) );
3038 set_i_o( _gvn.transform(io_phi) );
3039 }
3040
3041 C->set_has_split_ifs(true); // Has chance for split-if optimization
3042 set_result(result_rgn, result_val);
3043 return true;
3044 }
3045
3046 //---------------------------load_mirror_from_klass----------------------------
3047 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3048 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3049 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3050 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3051 }
3052
3053 //-----------------------load_klass_from_mirror_common-------------------------
3054 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3055 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3056 // and branch to the given path on the region.
3057 // If never_see_null, take an uncommon trap on null, so we can optimistically
3058 // compile for the non-null case.
3059 // If the region is NULL, force never_see_null = true.
3060 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3061 bool never_see_null,
3062 RegionNode* region,
3063 int null_path,
3064 int offset) {
3065 if (region == NULL) never_see_null = true;
3066 Node* p = basic_plus_adr(mirror, offset);
3067 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3068 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3069 Node* null_ctl = top();
3070 kls = null_check_oop(kls, &null_ctl, never_see_null);
3071 if (region != NULL) {
3072 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3073 region->init_req(null_path, null_ctl);
3074 } else {
3075 assert(null_ctl == top(), "no loose ends");
3076 }
3077 return kls;
3078 }
3079
3080 //--------------------(inline_native_Class_query helpers)---------------------
3081 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3082 // Fall through if (mods & mask) == bits, take the guard otherwise.
3083 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3084 // Branch around if the given klass has the given modifier bit set.
3085 // Like generate_guard, adds a new path onto the region.
3086 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3087 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3088 Node* mask = intcon(modifier_mask);
3089 Node* bits = intcon(modifier_bits);
3090 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3091 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3092 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3093 return generate_fair_guard(bol, region);
3094 }
3095 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3096 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3097 }
3098
3099 //-------------------------inline_native_Class_query-------------------
3100 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3101 const Type* return_type = TypeInt::BOOL;
3102 Node* prim_return_value = top(); // what happens if it's a primitive class?
3103 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3104 bool expect_prim = false; // most of these guys expect to work on refs
3105
3106 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3107
3108 Node* mirror = argument(0);
3109 Node* obj = top();
3110
3111 switch (id) {
3112 case vmIntrinsics::_isInstance:
3113 // nothing is an instance of a primitive type
3114 prim_return_value = intcon(0);
3115 obj = argument(1);
3116 break;
3117 case vmIntrinsics::_getModifiers:
3118 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3119 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3120 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3121 break;
3122 case vmIntrinsics::_isInterface:
3123 prim_return_value = intcon(0);
3124 break;
3125 case vmIntrinsics::_isArray:
3126 prim_return_value = intcon(0);
3127 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3128 break;
3129 case vmIntrinsics::_isPrimitive:
3130 prim_return_value = intcon(1);
3131 expect_prim = true; // obviously
3132 break;
3133 case vmIntrinsics::_getSuperclass:
3134 prim_return_value = null();
3135 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3136 break;
3137 case vmIntrinsics::_getComponentType:
3138 prim_return_value = null();
3139 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3140 break;
3141 case vmIntrinsics::_getClassAccessFlags:
3142 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3143 return_type = TypeInt::INT; // not bool! 6297094
3144 break;
3145 default:
3146 fatal_unexpected_iid(id);
3147 break;
3148 }
3149
3150 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3151 if (mirror_con == NULL) return false; // cannot happen?
3152
3153 #ifndef PRODUCT
3154 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3155 ciType* k = mirror_con->java_mirror_type();
3156 if (k) {
3157 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3158 k->print_name();
3159 tty->cr();
3160 }
3161 }
3162 #endif
3163
3164 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3165 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3166 record_for_igvn(region);
3167 PhiNode* phi = new (C) PhiNode(region, return_type);
3168
3169 // The mirror will never be null of Reflection.getClassAccessFlags, however
3170 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3171 // if it is. See bug 4774291.
3172
3173 // For Reflection.getClassAccessFlags(), the null check occurs in
3174 // the wrong place; see inline_unsafe_access(), above, for a similar
3175 // situation.
3176 mirror = null_check(mirror);
3177 // If mirror or obj is dead, only null-path is taken.
3178 if (stopped()) return true;
3179
3180 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3181
3182 // Now load the mirror's klass metaobject, and null-check it.
3183 // Side-effects region with the control path if the klass is null.
3184 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3185 // If kls is null, we have a primitive mirror.
3186 phi->init_req(_prim_path, prim_return_value);
3187 if (stopped()) { set_result(region, phi); return true; }
3188
3189 Node* p; // handy temp
3190 Node* null_ctl;
3191
3192 // Now that we have the non-null klass, we can perform the real query.
3193 // For constant classes, the query will constant-fold in LoadNode::Value.
3194 Node* query_value = top();
3195 switch (id) {
3196 case vmIntrinsics::_isInstance:
3197 // nothing is an instance of a primitive type
3198 query_value = gen_instanceof(obj, kls);
3199 break;
3200
3201 case vmIntrinsics::_getModifiers:
3202 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3203 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3204 break;
3205
3206 case vmIntrinsics::_isInterface:
3207 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3208 if (generate_interface_guard(kls, region) != NULL)
3209 // A guard was added. If the guard is taken, it was an interface.
3210 phi->add_req(intcon(1));
3211 // If we fall through, it's a plain class.
3212 query_value = intcon(0);
3213 break;
3214
3215 case vmIntrinsics::_isArray:
3216 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3217 if (generate_array_guard(kls, region) != NULL)
3218 // A guard was added. If the guard is taken, it was an array.
3219 phi->add_req(intcon(1));
3220 // If we fall through, it's a plain class.
3221 query_value = intcon(0);
3222 break;
3223
3224 case vmIntrinsics::_isPrimitive:
3225 query_value = intcon(0); // "normal" path produces false
3226 break;
3227
3228 case vmIntrinsics::_getSuperclass:
3229 // The rules here are somewhat unfortunate, but we can still do better
3230 // with random logic than with a JNI call.
3231 // Interfaces store null or Object as _super, but must report null.
3232 // Arrays store an intermediate super as _super, but must report Object.
3233 // Other types can report the actual _super.
3234 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3235 if (generate_interface_guard(kls, region) != NULL)
3236 // A guard was added. If the guard is taken, it was an interface.
3237 phi->add_req(null());
3238 if (generate_array_guard(kls, region) != NULL)
3239 // A guard was added. If the guard is taken, it was an array.
3240 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3241 // If we fall through, it's a plain class. Get its _super.
3242 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3243 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3244 null_ctl = top();
3245 kls = null_check_oop(kls, &null_ctl);
3246 if (null_ctl != top()) {
3247 // If the guard is taken, Object.superClass is null (both klass and mirror).
3248 region->add_req(null_ctl);
3249 phi ->add_req(null());
3250 }
3251 if (!stopped()) {
3252 query_value = load_mirror_from_klass(kls);
3253 }
3254 break;
3255
3256 case vmIntrinsics::_getComponentType:
3257 if (generate_array_guard(kls, region) != NULL) {
3258 // Be sure to pin the oop load to the guard edge just created:
3259 Node* is_array_ctrl = region->in(region->req()-1);
3260 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3261 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3262 phi->add_req(cmo);
3263 }
3264 query_value = null(); // non-array case is null
3265 break;
3266
3267 case vmIntrinsics::_getClassAccessFlags:
3268 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3269 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3270 break;
3271
3272 default:
3273 fatal_unexpected_iid(id);
3274 break;
3275 }
3276
3277 // Fall-through is the normal case of a query to a real class.
3278 phi->init_req(1, query_value);
3279 region->init_req(1, control());
3280
3281 C->set_has_split_ifs(true); // Has chance for split-if optimization
3282 set_result(region, phi);
3283 return true;
3284 }
3285
3286 //--------------------------inline_native_subtype_check------------------------
3287 // This intrinsic takes the JNI calls out of the heart of
3288 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3289 bool LibraryCallKit::inline_native_subtype_check() {
3290 // Pull both arguments off the stack.
3291 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3292 args[0] = argument(0);
3293 args[1] = argument(1);
3294 Node* klasses[2]; // corresponding Klasses: superk, subk
3295 klasses[0] = klasses[1] = top();
3296
3297 enum {
3298 // A full decision tree on {superc is prim, subc is prim}:
3299 _prim_0_path = 1, // {P,N} => false
3300 // {P,P} & superc!=subc => false
3301 _prim_same_path, // {P,P} & superc==subc => true
3302 _prim_1_path, // {N,P} => false
3303 _ref_subtype_path, // {N,N} & subtype check wins => true
3304 _both_ref_path, // {N,N} & subtype check loses => false
3305 PATH_LIMIT
3306 };
3307
3308 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3309 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3310 record_for_igvn(region);
3311
3312 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3313 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3314 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3315
3316 // First null-check both mirrors and load each mirror's klass metaobject.
3317 int which_arg;
3318 for (which_arg = 0; which_arg <= 1; which_arg++) {
3319 Node* arg = args[which_arg];
3320 arg = null_check(arg);
3321 if (stopped()) break;
3322 args[which_arg] = _gvn.transform(arg);
3323
3324 Node* p = basic_plus_adr(arg, class_klass_offset);
3325 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3326 klasses[which_arg] = _gvn.transform(kls);
3327 }
3328
3329 // Having loaded both klasses, test each for null.
3330 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3331 for (which_arg = 0; which_arg <= 1; which_arg++) {
3332 Node* kls = klasses[which_arg];
3333 Node* null_ctl = top();
3334 kls = null_check_oop(kls, &null_ctl, never_see_null);
3335 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3336 region->init_req(prim_path, null_ctl);
3337 if (stopped()) break;
3338 klasses[which_arg] = kls;
3339 }
3340
3341 if (!stopped()) {
3342 // now we have two reference types, in klasses[0..1]
3343 Node* subk = klasses[1]; // the argument to isAssignableFrom
3344 Node* superk = klasses[0]; // the receiver
3345 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3346 // now we have a successful reference subtype check
3347 region->set_req(_ref_subtype_path, control());
3348 }
3349
3350 // If both operands are primitive (both klasses null), then
3351 // we must return true when they are identical primitives.
3352 // It is convenient to test this after the first null klass check.
3353 set_control(region->in(_prim_0_path)); // go back to first null check
3354 if (!stopped()) {
3355 // Since superc is primitive, make a guard for the superc==subc case.
3356 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3357 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3358 generate_guard(bol_eq, region, PROB_FAIR);
3359 if (region->req() == PATH_LIMIT+1) {
3360 // A guard was added. If the added guard is taken, superc==subc.
3361 region->swap_edges(PATH_LIMIT, _prim_same_path);
3362 region->del_req(PATH_LIMIT);
3363 }
3364 region->set_req(_prim_0_path, control()); // Not equal after all.
3365 }
3366
3367 // these are the only paths that produce 'true':
3368 phi->set_req(_prim_same_path, intcon(1));
3369 phi->set_req(_ref_subtype_path, intcon(1));
3370
3371 // pull together the cases:
3372 assert(region->req() == PATH_LIMIT, "sane region");
3373 for (uint i = 1; i < region->req(); i++) {
3374 Node* ctl = region->in(i);
3375 if (ctl == NULL || ctl == top()) {
3376 region->set_req(i, top());
3377 phi ->set_req(i, top());
3378 } else if (phi->in(i) == NULL) {
3379 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3380 }
3381 }
3382
3383 set_control(_gvn.transform(region));
3384 set_result(_gvn.transform(phi));
3385 return true;
3386 }
3387
3388 //---------------------generate_array_guard_common------------------------
3389 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3390 bool obj_array, bool not_array) {
3391 // If obj_array/non_array==false/false:
3392 // Branch around if the given klass is in fact an array (either obj or prim).
3393 // If obj_array/non_array==false/true:
3394 // Branch around if the given klass is not an array klass of any kind.
3395 // If obj_array/non_array==true/true:
3396 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3397 // If obj_array/non_array==true/false:
3398 // Branch around if the kls is an oop array (Object[] or subtype)
3399 //
3400 // Like generate_guard, adds a new path onto the region.
3401 jint layout_con = 0;
3402 Node* layout_val = get_layout_helper(kls, layout_con);
3403 if (layout_val == NULL) {
3404 bool query = (obj_array
3405 ? Klass::layout_helper_is_objArray(layout_con)
3406 : Klass::layout_helper_is_array(layout_con));
3407 if (query == not_array) {
3408 return NULL; // never a branch
3409 } else { // always a branch
3410 Node* always_branch = control();
3411 if (region != NULL)
3412 region->add_req(always_branch);
3413 set_control(top());
3414 return always_branch;
3415 }
3416 }
3417 // Now test the correct condition.
3418 jint nval = (obj_array
3419 ? ((jint)Klass::_lh_array_tag_type_value
3420 << Klass::_lh_array_tag_shift)
3421 : Klass::_lh_neutral_value);
3422 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3423 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3424 // invert the test if we are looking for a non-array
3425 if (not_array) btest = BoolTest(btest).negate();
3426 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3427 return generate_fair_guard(bol, region);
3428 }
3429
3430
3431 //-----------------------inline_native_newArray--------------------------
3432 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3433 bool LibraryCallKit::inline_native_newArray() {
3434 Node* mirror = argument(0);
3435 Node* count_val = argument(1);
3436
3437 mirror = null_check(mirror);
3438 // If mirror or obj is dead, only null-path is taken.
3439 if (stopped()) return true;
3440
3441 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3442 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3443 PhiNode* result_val = new(C) PhiNode(result_reg,
3444 TypeInstPtr::NOTNULL);
3445 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3446 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3447 TypePtr::BOTTOM);
3448
3449 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3450 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3451 result_reg, _slow_path);
3452 Node* normal_ctl = control();
3453 Node* no_array_ctl = result_reg->in(_slow_path);
3454
3455 // Generate code for the slow case. We make a call to newArray().
3456 set_control(no_array_ctl);
3457 if (!stopped()) {
3458 // Either the input type is void.class, or else the
3459 // array klass has not yet been cached. Either the
3460 // ensuing call will throw an exception, or else it
3461 // will cache the array klass for next time.
3462 PreserveJVMState pjvms(this);
3463 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3464 Node* slow_result = set_results_for_java_call(slow_call);
3465 // this->control() comes from set_results_for_java_call
3466 result_reg->set_req(_slow_path, control());
3467 result_val->set_req(_slow_path, slow_result);
3468 result_io ->set_req(_slow_path, i_o());
3469 result_mem->set_req(_slow_path, reset_memory());
3470 }
3471
3472 set_control(normal_ctl);
3473 if (!stopped()) {
3474 // Normal case: The array type has been cached in the java.lang.Class.
3475 // The following call works fine even if the array type is polymorphic.
3476 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3477 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3478 result_reg->init_req(_normal_path, control());
3479 result_val->init_req(_normal_path, obj);
3480 result_io ->init_req(_normal_path, i_o());
3481 result_mem->init_req(_normal_path, reset_memory());
3482 }
3483
3484 // Return the combined state.
3485 set_i_o( _gvn.transform(result_io) );
3486 set_all_memory( _gvn.transform(result_mem) );
3487
3488 C->set_has_split_ifs(true); // Has chance for split-if optimization
3489 set_result(result_reg, result_val);
3490 return true;
3491 }
3492
3493 //----------------------inline_native_getLength--------------------------
3494 // public static native int java.lang.reflect.Array.getLength(Object array);
3495 bool LibraryCallKit::inline_native_getLength() {
3496 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3497
3498 Node* array = null_check(argument(0));
3499 // If array is dead, only null-path is taken.
3500 if (stopped()) return true;
3501
3502 // Deoptimize if it is a non-array.
3503 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3504
3505 if (non_array != NULL) {
3506 PreserveJVMState pjvms(this);
3507 set_control(non_array);
3508 uncommon_trap(Deoptimization::Reason_intrinsic,
3509 Deoptimization::Action_maybe_recompile);
3510 }
3511
3512 // If control is dead, only non-array-path is taken.
3513 if (stopped()) return true;
3514
3515 // The works fine even if the array type is polymorphic.
3516 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3517 Node* result = load_array_length(array);
3518
3519 C->set_has_split_ifs(true); // Has chance for split-if optimization
3520 set_result(result);
3521 return true;
3522 }
3523
3524 //------------------------inline_array_copyOf----------------------------
3525 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3526 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3527 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3528 return false;
3529 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3530
3531 // Get the arguments.
3532 Node* original = argument(0);
3533 Node* start = is_copyOfRange? argument(1): intcon(0);
3534 Node* end = is_copyOfRange? argument(2): argument(1);
3535 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3536
3537 Node* newcopy;
3538
3539 // Set the original stack and the reexecute bit for the interpreter to reexecute
3540 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3541 { PreserveReexecuteState preexecs(this);
3542 jvms()->set_should_reexecute(true);
3543
3544 array_type_mirror = null_check(array_type_mirror);
3545 original = null_check(original);
3546
3547 // Check if a null path was taken unconditionally.
3548 if (stopped()) return true;
3549
3550 Node* orig_length = load_array_length(original);
3551
3552 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3553 klass_node = null_check(klass_node);
3554
3555 RegionNode* bailout = new (C) RegionNode(1);
3556 record_for_igvn(bailout);
3557
3558 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3559 // Bail out if that is so.
3560 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3561 if (not_objArray != NULL) {
3562 // Improve the klass node's type from the new optimistic assumption:
3563 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3564 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3565 Node* cast = new (C) CastPPNode(klass_node, akls);
3566 cast->init_req(0, control());
3567 klass_node = _gvn.transform(cast);
3568 }
3569
3570 // Bail out if either start or end is negative.
3571 generate_negative_guard(start, bailout, &start);
3572 generate_negative_guard(end, bailout, &end);
3573
3574 Node* length = end;
3575 if (_gvn.type(start) != TypeInt::ZERO) {
3576 length = _gvn.transform(new (C) SubINode(end, start));
3577 }
3578
3579 // Bail out if length is negative.
3580 // Without this the new_array would throw
3581 // NegativeArraySizeException but IllegalArgumentException is what
3582 // should be thrown
3583 generate_negative_guard(length, bailout, &length);
3584
3585 if (bailout->req() > 1) {
3586 PreserveJVMState pjvms(this);
3587 set_control(_gvn.transform(bailout));
3588 uncommon_trap(Deoptimization::Reason_intrinsic,
3589 Deoptimization::Action_maybe_recompile);
3590 }
3591
3592 if (!stopped()) {
3593 // How many elements will we copy from the original?
3594 // The answer is MinI(orig_length - start, length).
3595 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3596 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3597
3598 newcopy = new_array(klass_node, length, 0); // no argments to push
3599
3600 // Generate a direct call to the right arraycopy function(s).
3601 // We know the copy is disjoint but we might not know if the
3602 // oop stores need checking.
3603 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3604 // This will fail a store-check if x contains any non-nulls.
3605 bool disjoint_bases = true;
3606 // if start > orig_length then the length of the copy may be
3607 // negative.
3608 bool length_never_negative = !is_copyOfRange;
3609 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3610 original, start, newcopy, intcon(0), moved,
3611 disjoint_bases, length_never_negative);
3612 }
3613 } // original reexecute is set back here
3614
3615 C->set_has_split_ifs(true); // Has chance for split-if optimization
3616 if (!stopped()) {
3617 set_result(newcopy);
3618 }
3619 return true;
3620 }
3621
3622
3623 //----------------------generate_virtual_guard---------------------------
3624 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3625 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3626 RegionNode* slow_region) {
3627 ciMethod* method = callee();
3628 int vtable_index = method->vtable_index();
3629 // Get the Method* out of the appropriate vtable entry.
3630 int entry_offset = (InstanceKlass::vtable_start_offset() +
3631 vtable_index*vtableEntry::size()) * wordSize +
3632 vtableEntry::method_offset_in_bytes();
3633 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3634 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3635
3636 // Compare the target method with the expected method (e.g., Object.hashCode).
3637 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3638
3639 Node* native_call = makecon(native_call_addr);
3640 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3641 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3642
3643 return generate_slow_guard(test_native, slow_region);
3644 }
3645
3646 //-----------------------generate_method_call----------------------------
3647 // Use generate_method_call to make a slow-call to the real
3648 // method if the fast path fails. An alternative would be to
3649 // use a stub like OptoRuntime::slow_arraycopy_Java.
3650 // This only works for expanding the current library call,
3651 // not another intrinsic. (E.g., don't use this for making an
3652 // arraycopy call inside of the copyOf intrinsic.)
3653 CallJavaNode*
3654 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3655 // When compiling the intrinsic method itself, do not use this technique.
3656 guarantee(callee() != C->method(), "cannot make slow-call to self");
3657
3658 ciMethod* method = callee();
3659 // ensure the JVMS we have will be correct for this call
3660 guarantee(method_id == method->intrinsic_id(), "must match");
3661
3662 const TypeFunc* tf = TypeFunc::make(method);
3663 CallJavaNode* slow_call;
3664 if (is_static) {
3665 assert(!is_virtual, "");
3666 slow_call = new(C) CallStaticJavaNode(tf,
3667 SharedRuntime::get_resolve_static_call_stub(),
3668 method, bci());
3669 } else if (is_virtual) {
3670 null_check_receiver();
3671 int vtable_index = Method::invalid_vtable_index;
3672 if (UseInlineCaches) {
3673 // Suppress the vtable call
3674 } else {
3675 // hashCode and clone are not a miranda methods,
3676 // so the vtable index is fixed.
3677 // No need to use the linkResolver to get it.
3678 vtable_index = method->vtable_index();
3679 }
3680 slow_call = new(C) CallDynamicJavaNode(tf,
3681 SharedRuntime::get_resolve_virtual_call_stub(),
3682 method, vtable_index, bci());
3683 } else { // neither virtual nor static: opt_virtual
3684 null_check_receiver();
3685 slow_call = new(C) CallStaticJavaNode(tf,
3686 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3687 method, bci());
3688 slow_call->set_optimized_virtual(true);
3689 }
3690 set_arguments_for_java_call(slow_call);
3691 set_edges_for_java_call(slow_call);
3692 return slow_call;
3693 }
3694
3695
3696 //------------------------------inline_native_hashcode--------------------
3697 // Build special case code for calls to hashCode on an object.
3698 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3699 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3700 assert(!(is_virtual && is_static), "either virtual, special, or static");
3701
3702 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3703
3704 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3705 PhiNode* result_val = new(C) PhiNode(result_reg,
3706 TypeInt::INT);
3707 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3708 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3709 TypePtr::BOTTOM);
3710 Node* obj = NULL;
3711 if (!is_static) {
3712 // Check for hashing null object
3713 obj = null_check_receiver();
3714 if (stopped()) return true; // unconditionally null
3715 result_reg->init_req(_null_path, top());
3716 result_val->init_req(_null_path, top());
3717 } else {
3718 // Do a null check, and return zero if null.
3719 // System.identityHashCode(null) == 0
3720 obj = argument(0);
3721 Node* null_ctl = top();
3722 obj = null_check_oop(obj, &null_ctl);
3723 result_reg->init_req(_null_path, null_ctl);
3724 result_val->init_req(_null_path, _gvn.intcon(0));
3725 }
3726
3727 // Unconditionally null? Then return right away.
3728 if (stopped()) {
3729 set_control( result_reg->in(_null_path));
3730 if (!stopped())
3731 set_result(result_val->in(_null_path));
3732 return true;
3733 }
3734
3735 // After null check, get the object's klass.
3736 Node* obj_klass = load_object_klass(obj);
3737
3738 // This call may be virtual (invokevirtual) or bound (invokespecial).
3739 // For each case we generate slightly different code.
3740
3741 // We only go to the fast case code if we pass a number of guards. The
3742 // paths which do not pass are accumulated in the slow_region.
3743 RegionNode* slow_region = new (C) RegionNode(1);
3744 record_for_igvn(slow_region);
3745
3746 // If this is a virtual call, we generate a funny guard. We pull out
3747 // the vtable entry corresponding to hashCode() from the target object.
3748 // If the target method which we are calling happens to be the native
3749 // Object hashCode() method, we pass the guard. We do not need this
3750 // guard for non-virtual calls -- the caller is known to be the native
3751 // Object hashCode().
3752 if (is_virtual) {
3753 generate_virtual_guard(obj_klass, slow_region);
3754 }
3755
3756 // Get the header out of the object, use LoadMarkNode when available
3757 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3758 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3759
3760 // Test the header to see if it is unlocked.
3761 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3762 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3763 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3764 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3765 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3766
3767 generate_slow_guard(test_unlocked, slow_region);
3768
3769 // Get the hash value and check to see that it has been properly assigned.
3770 // We depend on hash_mask being at most 32 bits and avoid the use of
3771 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3772 // vm: see markOop.hpp.
3773 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3774 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3775 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3776 // This hack lets the hash bits live anywhere in the mark object now, as long
3777 // as the shift drops the relevant bits into the low 32 bits. Note that
3778 // Java spec says that HashCode is an int so there's no point in capturing
3779 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3780 hshifted_header = ConvX2I(hshifted_header);
3781 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3782
3783 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3784 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3785 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3786
3787 generate_slow_guard(test_assigned, slow_region);
3788
3789 Node* init_mem = reset_memory();
3790 // fill in the rest of the null path:
3791 result_io ->init_req(_null_path, i_o());
3792 result_mem->init_req(_null_path, init_mem);
3793
3794 result_val->init_req(_fast_path, hash_val);
3795 result_reg->init_req(_fast_path, control());
3796 result_io ->init_req(_fast_path, i_o());
3797 result_mem->init_req(_fast_path, init_mem);
3798
3799 // Generate code for the slow case. We make a call to hashCode().
3800 set_control(_gvn.transform(slow_region));
3801 if (!stopped()) {
3802 // No need for PreserveJVMState, because we're using up the present state.
3803 set_all_memory(init_mem);
3804 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3805 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3806 Node* slow_result = set_results_for_java_call(slow_call);
3807 // this->control() comes from set_results_for_java_call
3808 result_reg->init_req(_slow_path, control());
3809 result_val->init_req(_slow_path, slow_result);
3810 result_io ->set_req(_slow_path, i_o());
3811 result_mem ->set_req(_slow_path, reset_memory());
3812 }
3813
3814 // Return the combined state.
3815 set_i_o( _gvn.transform(result_io) );
3816 set_all_memory( _gvn.transform(result_mem) );
3817
3818 set_result(result_reg, result_val);
3819 return true;
3820 }
3821
3822 //---------------------------inline_native_getClass----------------------------
3823 // public final native Class<?> java.lang.Object.getClass();
3824 //
3825 // Build special case code for calls to getClass on an object.
3826 bool LibraryCallKit::inline_native_getClass() {
3827 Node* obj = null_check_receiver();
3828 if (stopped()) return true;
3829 set_result(load_mirror_from_klass(load_object_klass(obj)));
3830 return true;
3831 }
3832
3833 //-----------------inline_native_Reflection_getCallerClass---------------------
3834 // public static native Class<?> sun.reflect.Reflection.getCallerClass(int realFramesToSkip);
3835 //
3836 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3837 //
3838 // NOTE that this code must perform the same logic as
3839 // vframeStream::security_get_caller_frame in that it must skip
3840 // Method.invoke() and auxiliary frames.
3841 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3842 #ifndef PRODUCT
3843 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3844 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3845 }
3846 #endif
3847
3848 Node* caller_depth_node = argument(0);
3849
3850 // The depth value must be a constant in order for the runtime call
3851 // to be eliminated.
3852 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3853 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3854 #ifndef PRODUCT
3855 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3856 tty->print_cr(" Bailing out because caller depth was not a constant");
3857 }
3858 #endif
3859 return false;
3860 }
3861 // Note that the JVM state at this point does not include the
3862 // getCallerClass() frame which we are trying to inline. The
3863 // semantics of getCallerClass(), however, are that the "first"
3864 // frame is the getCallerClass() frame, so we subtract one from the
3865 // requested depth before continuing. We don't inline requests of
3866 // getCallerClass(0).
3867 int caller_depth = caller_depth_type->get_con() - 1;
3868 if (caller_depth < 0) {
3869 #ifndef PRODUCT
3870 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3871 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3872 }
3873 #endif
3874 return false;
3875 }
3876
3877 if (!jvms()->has_method()) {
3878 #ifndef PRODUCT
3879 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3880 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3881 }
3882 #endif
3883 return false;
3884 }
3885 int _depth = jvms()->depth(); // cache call chain depth
3886
3887 // Walk back up the JVM state to find the caller at the required
3888 // depth. NOTE that this code must perform the same logic as
3889 // vframeStream::security_get_caller_frame in that it must skip
3890 // Method.invoke() and auxiliary frames. Note also that depth is
3891 // 1-based (1 is the bottom of the inlining).
3892 int inlining_depth = _depth;
3893 JVMState* caller_jvms = NULL;
3894
3895 if (inlining_depth > 0) {
3896 caller_jvms = jvms();
3897 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3898 do {
3899 // The following if-tests should be performed in this order
3900 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3901 // Skip a Method.invoke() or auxiliary frame
3902 } else if (caller_depth > 0) {
3903 // Skip real frame
3904 --caller_depth;
3905 } else {
3906 // We're done: reached desired caller after skipping.
3907 break;
3908 }
3909 caller_jvms = caller_jvms->caller();
3910 --inlining_depth;
3911 } while (inlining_depth > 0);
3912 }
3913
3914 if (inlining_depth == 0) {
3915 #ifndef PRODUCT
3916 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3917 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3918 tty->print_cr(" JVM state at this point:");
3919 for (int i = _depth; i >= 1; i--) {
3920 ciMethod* m = jvms()->of_depth(i)->method();
3921 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3922 }
3923 }
3924 #endif
3925 return false; // Reached end of inlining
3926 }
3927
3928 // Acquire method holder as java.lang.Class
3929 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3930 ciInstance* caller_mirror = caller_klass->java_mirror();
3931
3932 // Push this as a constant
3933 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3934
3935 #ifndef PRODUCT
3936 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3937 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);
3938 tty->print_cr(" JVM state at this point:");
3939 for (int i = _depth; i >= 1; i--) {
3940 ciMethod* m = jvms()->of_depth(i)->method();
3941 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3942 }
3943 }
3944 #endif
3945 return true;
3946 }
3947
3948 // Helper routine for above
3949 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3950 ciMethod* method = jvms->method();
3951
3952 // Is this the Method.invoke method itself?
3953 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3954 return true;
3955
3956 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3957 ciKlass* k = method->holder();
3958 if (k->is_instance_klass()) {
3959 ciInstanceKlass* ik = k->as_instance_klass();
3960 for (; ik != NULL; ik = ik->super()) {
3961 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3962 ik == env()->find_system_klass(ik->name())) {
3963 return true;
3964 }
3965 }
3966 }
3967 else if (method->is_method_handle_intrinsic() ||
3968 method->is_compiled_lambda_form()) {
3969 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3970 return true;
3971 }
3972
3973 return false;
3974 }
3975
3976 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
3977 Node* arg = argument(0);
3978 Node* result;
3979
3980 switch (id) {
3981 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
3982 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
3983 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
3984 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
3985
3986 case vmIntrinsics::_doubleToLongBits: {
3987 // two paths (plus control) merge in a wood
3988 RegionNode *r = new (C) RegionNode(3);
3989 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
3990
3991 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
3992 // Build the boolean node
3993 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
3994
3995 // Branch either way.
3996 // NaN case is less traveled, which makes all the difference.
3997 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3998 Node *opt_isnan = _gvn.transform(ifisnan);
3999 assert( opt_isnan->is_If(), "Expect an IfNode");
4000 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4001 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4002
4003 set_control(iftrue);
4004
4005 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4006 Node *slow_result = longcon(nan_bits); // return NaN
4007 phi->init_req(1, _gvn.transform( slow_result ));
4008 r->init_req(1, iftrue);
4009
4010 // Else fall through
4011 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4012 set_control(iffalse);
4013
4014 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4015 r->init_req(2, iffalse);
4016
4017 // Post merge
4018 set_control(_gvn.transform(r));
4019 record_for_igvn(r);
4020
4021 C->set_has_split_ifs(true); // Has chance for split-if optimization
4022 result = phi;
4023 assert(result->bottom_type()->isa_long(), "must be");
4024 break;
4025 }
4026
4027 case vmIntrinsics::_floatToIntBits: {
4028 // two paths (plus control) merge in a wood
4029 RegionNode *r = new (C) RegionNode(3);
4030 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4031
4032 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4033 // Build the boolean node
4034 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4035
4036 // Branch either way.
4037 // NaN case is less traveled, which makes all the difference.
4038 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4039 Node *opt_isnan = _gvn.transform(ifisnan);
4040 assert( opt_isnan->is_If(), "Expect an IfNode");
4041 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4042 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4043
4044 set_control(iftrue);
4045
4046 static const jint nan_bits = 0x7fc00000;
4047 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4048 phi->init_req(1, _gvn.transform( slow_result ));
4049 r->init_req(1, iftrue);
4050
4051 // Else fall through
4052 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4053 set_control(iffalse);
4054
4055 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4056 r->init_req(2, iffalse);
4057
4058 // Post merge
4059 set_control(_gvn.transform(r));
4060 record_for_igvn(r);
4061
4062 C->set_has_split_ifs(true); // Has chance for split-if optimization
4063 result = phi;
4064 assert(result->bottom_type()->isa_int(), "must be");
4065 break;
4066 }
4067
4068 default:
4069 fatal_unexpected_iid(id);
4070 break;
4071 }
4072 set_result(_gvn.transform(result));
4073 return true;
4074 }
4075
4076 #ifdef _LP64
4077 #define XTOP ,top() /*additional argument*/
4078 #else //_LP64
4079 #define XTOP /*no additional argument*/
4080 #endif //_LP64
4081
4082 //----------------------inline_unsafe_copyMemory-------------------------
4083 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4084 bool LibraryCallKit::inline_unsafe_copyMemory() {
4085 if (callee()->is_static()) return false; // caller must have the capability!
4086 null_check_receiver(); // null-check receiver
4087 if (stopped()) return true;
4088
4089 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4090
4091 Node* src_ptr = argument(1); // type: oop
4092 Node* src_off = ConvL2X(argument(2)); // type: long
4093 Node* dst_ptr = argument(4); // type: oop
4094 Node* dst_off = ConvL2X(argument(5)); // type: long
4095 Node* size = ConvL2X(argument(7)); // type: long
4096
4097 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4098 "fieldOffset must be byte-scaled");
4099
4100 Node* src = make_unsafe_address(src_ptr, src_off);
4101 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4102
4103 // Conservatively insert a memory barrier on all memory slices.
4104 // Do not let writes of the copy source or destination float below the copy.
4105 insert_mem_bar(Op_MemBarCPUOrder);
4106
4107 // Call it. Note that the length argument is not scaled.
4108 make_runtime_call(RC_LEAF|RC_NO_FP,
4109 OptoRuntime::fast_arraycopy_Type(),
4110 StubRoutines::unsafe_arraycopy(),
4111 "unsafe_arraycopy",
4112 TypeRawPtr::BOTTOM,
4113 src, dst, size XTOP);
4114
4115 // Do not let reads of the copy destination float above the copy.
4116 insert_mem_bar(Op_MemBarCPUOrder);
4117
4118 return true;
4119 }
4120
4121 //------------------------clone_coping-----------------------------------
4122 // Helper function for inline_native_clone.
4123 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4124 assert(obj_size != NULL, "");
4125 Node* raw_obj = alloc_obj->in(1);
4126 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4127
4128 AllocateNode* alloc = NULL;
4129 if (ReduceBulkZeroing) {
4130 // We will be completely responsible for initializing this object -
4131 // mark Initialize node as complete.
4132 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4133 // The object was just allocated - there should be no any stores!
4134 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4135 // Mark as complete_with_arraycopy so that on AllocateNode
4136 // expansion, we know this AllocateNode is initialized by an array
4137 // copy and a StoreStore barrier exists after the array copy.
4138 alloc->initialization()->set_complete_with_arraycopy();
4139 }
4140
4141 // Copy the fastest available way.
4142 // TODO: generate fields copies for small objects instead.
4143 Node* src = obj;
4144 Node* dest = alloc_obj;
4145 Node* size = _gvn.transform(obj_size);
4146
4147 // Exclude the header but include array length to copy by 8 bytes words.
4148 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4149 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4150 instanceOopDesc::base_offset_in_bytes();
4151 // base_off:
4152 // 8 - 32-bit VM
4153 // 12 - 64-bit VM, compressed klass
4154 // 16 - 64-bit VM, normal klass
4155 if (base_off % BytesPerLong != 0) {
4156 assert(UseCompressedKlassPointers, "");
4157 if (is_array) {
4158 // Exclude length to copy by 8 bytes words.
4159 base_off += sizeof(int);
4160 } else {
4161 // Include klass to copy by 8 bytes words.
4162 base_off = instanceOopDesc::klass_offset_in_bytes();
4163 }
4164 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4165 }
4166 src = basic_plus_adr(src, base_off);
4167 dest = basic_plus_adr(dest, base_off);
4168
4169 // Compute the length also, if needed:
4170 Node* countx = size;
4171 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4172 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4173
4174 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4175 bool disjoint_bases = true;
4176 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4177 src, NULL, dest, NULL, countx,
4178 /*dest_uninitialized*/true);
4179
4180 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4181 if (card_mark) {
4182 assert(!is_array, "");
4183 // Put in store barrier for any and all oops we are sticking
4184 // into this object. (We could avoid this if we could prove
4185 // that the object type contains no oop fields at all.)
4186 Node* no_particular_value = NULL;
4187 Node* no_particular_field = NULL;
4188 int raw_adr_idx = Compile::AliasIdxRaw;
4189 post_barrier(control(),
4190 memory(raw_adr_type),
4191 alloc_obj,
4192 no_particular_field,
4193 raw_adr_idx,
4194 no_particular_value,
4195 T_OBJECT,
4196 false);
4197 }
4198
4199 // Do not let reads from the cloned object float above the arraycopy.
4200 if (alloc != NULL) {
4201 // Do not let stores that initialize this object be reordered with
4202 // a subsequent store that would make this object accessible by
4203 // other threads.
4204 // Record what AllocateNode this StoreStore protects so that
4205 // escape analysis can go from the MemBarStoreStoreNode to the
4206 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4207 // based on the escape status of the AllocateNode.
4208 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4209 } else {
4210 insert_mem_bar(Op_MemBarCPUOrder);
4211 }
4212 }
4213
4214 //------------------------inline_native_clone----------------------------
4215 // protected native Object java.lang.Object.clone();
4216 //
4217 // Here are the simple edge cases:
4218 // null receiver => normal trap
4219 // virtual and clone was overridden => slow path to out-of-line clone
4220 // not cloneable or finalizer => slow path to out-of-line Object.clone
4221 //
4222 // The general case has two steps, allocation and copying.
4223 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4224 //
4225 // Copying also has two cases, oop arrays and everything else.
4226 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4227 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4228 //
4229 // These steps fold up nicely if and when the cloned object's klass
4230 // can be sharply typed as an object array, a type array, or an instance.
4231 //
4232 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4233 PhiNode* result_val;
4234
4235 // Set the reexecute bit for the interpreter to reexecute
4236 // the bytecode that invokes Object.clone if deoptimization happens.
4237 { PreserveReexecuteState preexecs(this);
4238 jvms()->set_should_reexecute(true);
4239
4240 Node* obj = null_check_receiver();
4241 if (stopped()) return true;
4242
4243 Node* obj_klass = load_object_klass(obj);
4244 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4245 const TypeOopPtr* toop = ((tklass != NULL)
4246 ? tklass->as_instance_type()
4247 : TypeInstPtr::NOTNULL);
4248
4249 // Conservatively insert a memory barrier on all memory slices.
4250 // Do not let writes into the original float below the clone.
4251 insert_mem_bar(Op_MemBarCPUOrder);
4252
4253 // paths into result_reg:
4254 enum {
4255 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4256 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4257 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4258 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4259 PATH_LIMIT
4260 };
4261 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4262 result_val = new(C) PhiNode(result_reg,
4263 TypeInstPtr::NOTNULL);
4264 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4265 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4266 TypePtr::BOTTOM);
4267 record_for_igvn(result_reg);
4268
4269 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4270 int raw_adr_idx = Compile::AliasIdxRaw;
4271
4272 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4273 if (array_ctl != NULL) {
4274 // It's an array.
4275 PreserveJVMState pjvms(this);
4276 set_control(array_ctl);
4277 Node* obj_length = load_array_length(obj);
4278 Node* obj_size = NULL;
4279 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4280
4281 if (!use_ReduceInitialCardMarks()) {
4282 // If it is an oop array, it requires very special treatment,
4283 // because card marking is required on each card of the array.
4284 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4285 if (is_obja != NULL) {
4286 PreserveJVMState pjvms2(this);
4287 set_control(is_obja);
4288 // Generate a direct call to the right arraycopy function(s).
4289 bool disjoint_bases = true;
4290 bool length_never_negative = true;
4291 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4292 obj, intcon(0), alloc_obj, intcon(0),
4293 obj_length,
4294 disjoint_bases, length_never_negative);
4295 result_reg->init_req(_objArray_path, control());
4296 result_val->init_req(_objArray_path, alloc_obj);
4297 result_i_o ->set_req(_objArray_path, i_o());
4298 result_mem ->set_req(_objArray_path, reset_memory());
4299 }
4300 }
4301 // Otherwise, there are no card marks to worry about.
4302 // (We can dispense with card marks if we know the allocation
4303 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4304 // causes the non-eden paths to take compensating steps to
4305 // simulate a fresh allocation, so that no further
4306 // card marks are required in compiled code to initialize
4307 // the object.)
4308
4309 if (!stopped()) {
4310 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4311
4312 // Present the results of the copy.
4313 result_reg->init_req(_array_path, control());
4314 result_val->init_req(_array_path, alloc_obj);
4315 result_i_o ->set_req(_array_path, i_o());
4316 result_mem ->set_req(_array_path, reset_memory());
4317 }
4318 }
4319
4320 // We only go to the instance fast case code if we pass a number of guards.
4321 // The paths which do not pass are accumulated in the slow_region.
4322 RegionNode* slow_region = new (C) RegionNode(1);
4323 record_for_igvn(slow_region);
4324 if (!stopped()) {
4325 // It's an instance (we did array above). Make the slow-path tests.
4326 // If this is a virtual call, we generate a funny guard. We grab
4327 // the vtable entry corresponding to clone() from the target object.
4328 // If the target method which we are calling happens to be the
4329 // Object clone() method, we pass the guard. We do not need this
4330 // guard for non-virtual calls; the caller is known to be the native
4331 // Object clone().
4332 if (is_virtual) {
4333 generate_virtual_guard(obj_klass, slow_region);
4334 }
4335
4336 // The object must be cloneable and must not have a finalizer.
4337 // Both of these conditions may be checked in a single test.
4338 // We could optimize the cloneable test further, but we don't care.
4339 generate_access_flags_guard(obj_klass,
4340 // Test both conditions:
4341 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4342 // Must be cloneable but not finalizer:
4343 JVM_ACC_IS_CLONEABLE,
4344 slow_region);
4345 }
4346
4347 if (!stopped()) {
4348 // It's an instance, and it passed the slow-path tests.
4349 PreserveJVMState pjvms(this);
4350 Node* obj_size = NULL;
4351 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4352
4353 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4354
4355 // Present the results of the slow call.
4356 result_reg->init_req(_instance_path, control());
4357 result_val->init_req(_instance_path, alloc_obj);
4358 result_i_o ->set_req(_instance_path, i_o());
4359 result_mem ->set_req(_instance_path, reset_memory());
4360 }
4361
4362 // Generate code for the slow case. We make a call to clone().
4363 set_control(_gvn.transform(slow_region));
4364 if (!stopped()) {
4365 PreserveJVMState pjvms(this);
4366 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4367 Node* slow_result = set_results_for_java_call(slow_call);
4368 // this->control() comes from set_results_for_java_call
4369 result_reg->init_req(_slow_path, control());
4370 result_val->init_req(_slow_path, slow_result);
4371 result_i_o ->set_req(_slow_path, i_o());
4372 result_mem ->set_req(_slow_path, reset_memory());
4373 }
4374
4375 // Return the combined state.
4376 set_control( _gvn.transform(result_reg) );
4377 set_i_o( _gvn.transform(result_i_o) );
4378 set_all_memory( _gvn.transform(result_mem) );
4379 } // original reexecute is set back here
4380
4381 set_result(_gvn.transform(result_val));
4382 return true;
4383 }
4384
4385 //------------------------------basictype2arraycopy----------------------------
4386 address LibraryCallKit::basictype2arraycopy(BasicType t,
4387 Node* src_offset,
4388 Node* dest_offset,
4389 bool disjoint_bases,
4390 const char* &name,
4391 bool dest_uninitialized) {
4392 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4393 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4394
4395 bool aligned = false;
4396 bool disjoint = disjoint_bases;
4397
4398 // if the offsets are the same, we can treat the memory regions as
4399 // disjoint, because either the memory regions are in different arrays,
4400 // or they are identical (which we can treat as disjoint.) We can also
4401 // treat a copy with a destination index less that the source index
4402 // as disjoint since a low->high copy will work correctly in this case.
4403 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4404 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4405 // both indices are constants
4406 int s_offs = src_offset_inttype->get_con();
4407 int d_offs = dest_offset_inttype->get_con();
4408 int element_size = type2aelembytes(t);
4409 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4410 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4411 if (s_offs >= d_offs) disjoint = true;
4412 } else if (src_offset == dest_offset && src_offset != NULL) {
4413 // This can occur if the offsets are identical non-constants.
4414 disjoint = true;
4415 }
4416
4417 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4418 }
4419
4420
4421 //------------------------------inline_arraycopy-----------------------
4422 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4423 // Object dest, int destPos,
4424 // int length);
4425 bool LibraryCallKit::inline_arraycopy() {
4426 // Get the arguments.
4427 Node* src = argument(0); // type: oop
4428 Node* src_offset = argument(1); // type: int
4429 Node* dest = argument(2); // type: oop
4430 Node* dest_offset = argument(3); // type: int
4431 Node* length = argument(4); // type: int
4432
4433 // Compile time checks. If any of these checks cannot be verified at compile time,
4434 // we do not make a fast path for this call. Instead, we let the call remain as it
4435 // is. The checks we choose to mandate at compile time are:
4436 //
4437 // (1) src and dest are arrays.
4438 const Type* src_type = src->Value(&_gvn);
4439 const Type* dest_type = dest->Value(&_gvn);
4440 const TypeAryPtr* top_src = src_type->isa_aryptr();
4441 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4442 if (top_src == NULL || top_src->klass() == NULL ||
4443 top_dest == NULL || top_dest->klass() == NULL) {
4444 // Conservatively insert a memory barrier on all memory slices.
4445 // Do not let writes into the source float below the arraycopy.
4446 insert_mem_bar(Op_MemBarCPUOrder);
4447
4448 // Call StubRoutines::generic_arraycopy stub.
4449 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4450 src, src_offset, dest, dest_offset, length);
4451
4452 // Do not let reads from the destination float above the arraycopy.
4453 // Since we cannot type the arrays, we don't know which slices
4454 // might be affected. We could restrict this barrier only to those
4455 // memory slices which pertain to array elements--but don't bother.
4456 if (!InsertMemBarAfterArraycopy)
4457 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4458 insert_mem_bar(Op_MemBarCPUOrder);
4459 return true;
4460 }
4461
4462 // (2) src and dest arrays must have elements of the same BasicType
4463 // Figure out the size and type of the elements we will be copying.
4464 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4465 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4466 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4467 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4468
4469 if (src_elem != dest_elem || dest_elem == T_VOID) {
4470 // The component types are not the same or are not recognized. Punt.
4471 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4472 generate_slow_arraycopy(TypePtr::BOTTOM,
4473 src, src_offset, dest, dest_offset, length,
4474 /*dest_uninitialized*/false);
4475 return true;
4476 }
4477
4478 //---------------------------------------------------------------------------
4479 // We will make a fast path for this call to arraycopy.
4480
4481 // We have the following tests left to perform:
4482 //
4483 // (3) src and dest must not be null.
4484 // (4) src_offset must not be negative.
4485 // (5) dest_offset must not be negative.
4486 // (6) length must not be negative.
4487 // (7) src_offset + length must not exceed length of src.
4488 // (8) dest_offset + length must not exceed length of dest.
4489 // (9) each element of an oop array must be assignable
4490
4491 RegionNode* slow_region = new (C) RegionNode(1);
4492 record_for_igvn(slow_region);
4493
4494 // (3) operands must not be null
4495 // We currently perform our null checks with the null_check routine.
4496 // This means that the null exceptions will be reported in the caller
4497 // rather than (correctly) reported inside of the native arraycopy call.
4498 // This should be corrected, given time. We do our null check with the
4499 // stack pointer restored.
4500 src = null_check(src, T_ARRAY);
4501 dest = null_check(dest, T_ARRAY);
4502
4503 // (4) src_offset must not be negative.
4504 generate_negative_guard(src_offset, slow_region);
4505
4506 // (5) dest_offset must not be negative.
4507 generate_negative_guard(dest_offset, slow_region);
4508
4509 // (6) length must not be negative (moved to generate_arraycopy()).
4510 // generate_negative_guard(length, slow_region);
4511
4512 // (7) src_offset + length must not exceed length of src.
4513 generate_limit_guard(src_offset, length,
4514 load_array_length(src),
4515 slow_region);
4516
4517 // (8) dest_offset + length must not exceed length of dest.
4518 generate_limit_guard(dest_offset, length,
4519 load_array_length(dest),
4520 slow_region);
4521
4522 // (9) each element of an oop array must be assignable
4523 // The generate_arraycopy subroutine checks this.
4524
4525 // This is where the memory effects are placed:
4526 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4527 generate_arraycopy(adr_type, dest_elem,
4528 src, src_offset, dest, dest_offset, length,
4529 false, false, slow_region);
4530
4531 return true;
4532 }
4533
4534 //-----------------------------generate_arraycopy----------------------
4535 // Generate an optimized call to arraycopy.
4536 // Caller must guard against non-arrays.
4537 // Caller must determine a common array basic-type for both arrays.
4538 // Caller must validate offsets against array bounds.
4539 // The slow_region has already collected guard failure paths
4540 // (such as out of bounds length or non-conformable array types).
4541 // The generated code has this shape, in general:
4542 //
4543 // if (length == 0) return // via zero_path
4544 // slowval = -1
4545 // if (types unknown) {
4546 // slowval = call generic copy loop
4547 // if (slowval == 0) return // via checked_path
4548 // } else if (indexes in bounds) {
4549 // if ((is object array) && !(array type check)) {
4550 // slowval = call checked copy loop
4551 // if (slowval == 0) return // via checked_path
4552 // } else {
4553 // call bulk copy loop
4554 // return // via fast_path
4555 // }
4556 // }
4557 // // adjust params for remaining work:
4558 // if (slowval != -1) {
4559 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4560 // }
4561 // slow_region:
4562 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4563 // return // via slow_call_path
4564 //
4565 // This routine is used from several intrinsics: System.arraycopy,
4566 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4567 //
4568 void
4569 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4570 BasicType basic_elem_type,
4571 Node* src, Node* src_offset,
4572 Node* dest, Node* dest_offset,
4573 Node* copy_length,
4574 bool disjoint_bases,
4575 bool length_never_negative,
4576 RegionNode* slow_region) {
4577
4578 if (slow_region == NULL) {
4579 slow_region = new(C) RegionNode(1);
4580 record_for_igvn(slow_region);
4581 }
4582
4583 Node* original_dest = dest;
4584 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4585 bool dest_uninitialized = false;
4586
4587 // See if this is the initialization of a newly-allocated array.
4588 // If so, we will take responsibility here for initializing it to zero.
4589 // (Note: Because tightly_coupled_allocation performs checks on the
4590 // out-edges of the dest, we need to avoid making derived pointers
4591 // from it until we have checked its uses.)
4592 if (ReduceBulkZeroing
4593 && !ZeroTLAB // pointless if already zeroed
4594 && basic_elem_type != T_CONFLICT // avoid corner case
4595 && !src->eqv_uncast(dest)
4596 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4597 != NULL)
4598 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4599 && alloc->maybe_set_complete(&_gvn)) {
4600 // "You break it, you buy it."
4601 InitializeNode* init = alloc->initialization();
4602 assert(init->is_complete(), "we just did this");
4603 init->set_complete_with_arraycopy();
4604 assert(dest->is_CheckCastPP(), "sanity");
4605 assert(dest->in(0)->in(0) == init, "dest pinned");
4606 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4607 // From this point on, every exit path is responsible for
4608 // initializing any non-copied parts of the object to zero.
4609 // Also, if this flag is set we make sure that arraycopy interacts properly
4610 // with G1, eliding pre-barriers. See CR 6627983.
4611 dest_uninitialized = true;
4612 } else {
4613 // No zeroing elimination here.
4614 alloc = NULL;
4615 //original_dest = dest;
4616 //dest_uninitialized = false;
4617 }
4618
4619 // Results are placed here:
4620 enum { fast_path = 1, // normal void-returning assembly stub
4621 checked_path = 2, // special assembly stub with cleanup
4622 slow_call_path = 3, // something went wrong; call the VM
4623 zero_path = 4, // bypass when length of copy is zero
4624 bcopy_path = 5, // copy primitive array by 64-bit blocks
4625 PATH_LIMIT = 6
4626 };
4627 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4628 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4629 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4630 record_for_igvn(result_region);
4631 _gvn.set_type_bottom(result_i_o);
4632 _gvn.set_type_bottom(result_memory);
4633 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4634
4635 // The slow_control path:
4636 Node* slow_control;
4637 Node* slow_i_o = i_o();
4638 Node* slow_mem = memory(adr_type);
4639 debug_only(slow_control = (Node*) badAddress);
4640
4641 // Checked control path:
4642 Node* checked_control = top();
4643 Node* checked_mem = NULL;
4644 Node* checked_i_o = NULL;
4645 Node* checked_value = NULL;
4646
4647 if (basic_elem_type == T_CONFLICT) {
4648 assert(!dest_uninitialized, "");
4649 Node* cv = generate_generic_arraycopy(adr_type,
4650 src, src_offset, dest, dest_offset,
4651 copy_length, dest_uninitialized);
4652 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4653 checked_control = control();
4654 checked_i_o = i_o();
4655 checked_mem = memory(adr_type);
4656 checked_value = cv;
4657 set_control(top()); // no fast path
4658 }
4659
4660 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4661 if (not_pos != NULL) {
4662 PreserveJVMState pjvms(this);
4663 set_control(not_pos);
4664
4665 // (6) length must not be negative.
4666 if (!length_never_negative) {
4667 generate_negative_guard(copy_length, slow_region);
4668 }
4669
4670 // copy_length is 0.
4671 if (!stopped() && dest_uninitialized) {
4672 Node* dest_length = alloc->in(AllocateNode::ALength);
4673 if (copy_length->eqv_uncast(dest_length)
4674 || _gvn.find_int_con(dest_length, 1) <= 0) {
4675 // There is no zeroing to do. No need for a secondary raw memory barrier.
4676 } else {
4677 // Clear the whole thing since there are no source elements to copy.
4678 generate_clear_array(adr_type, dest, basic_elem_type,
4679 intcon(0), NULL,
4680 alloc->in(AllocateNode::AllocSize));
4681 // Use a secondary InitializeNode as raw memory barrier.
4682 // Currently it is needed only on this path since other
4683 // paths have stub or runtime calls as raw memory barriers.
4684 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4685 Compile::AliasIdxRaw,
4686 top())->as_Initialize();
4687 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4688 }
4689 }
4690
4691 // Present the results of the fast call.
4692 result_region->init_req(zero_path, control());
4693 result_i_o ->init_req(zero_path, i_o());
4694 result_memory->init_req(zero_path, memory(adr_type));
4695 }
4696
4697 if (!stopped() && dest_uninitialized) {
4698 // We have to initialize the *uncopied* part of the array to zero.
4699 // The copy destination is the slice dest[off..off+len]. The other slices
4700 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4701 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4702 Node* dest_length = alloc->in(AllocateNode::ALength);
4703 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4704 copy_length) );
4705
4706 // If there is a head section that needs zeroing, do it now.
4707 if (find_int_con(dest_offset, -1) != 0) {
4708 generate_clear_array(adr_type, dest, basic_elem_type,
4709 intcon(0), dest_offset,
4710 NULL);
4711 }
4712
4713 // Next, perform a dynamic check on the tail length.
4714 // It is often zero, and we can win big if we prove this.
4715 // There are two wins: Avoid generating the ClearArray
4716 // with its attendant messy index arithmetic, and upgrade
4717 // the copy to a more hardware-friendly word size of 64 bits.
4718 Node* tail_ctl = NULL;
4719 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4720 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4721 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4722 tail_ctl = generate_slow_guard(bol_lt, NULL);
4723 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4724 }
4725
4726 // At this point, let's assume there is no tail.
4727 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4728 // There is no tail. Try an upgrade to a 64-bit copy.
4729 bool didit = false;
4730 { PreserveJVMState pjvms(this);
4731 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4732 src, src_offset, dest, dest_offset,
4733 dest_size, dest_uninitialized);
4734 if (didit) {
4735 // Present the results of the block-copying fast call.
4736 result_region->init_req(bcopy_path, control());
4737 result_i_o ->init_req(bcopy_path, i_o());
4738 result_memory->init_req(bcopy_path, memory(adr_type));
4739 }
4740 }
4741 if (didit)
4742 set_control(top()); // no regular fast path
4743 }
4744
4745 // Clear the tail, if any.
4746 if (tail_ctl != NULL) {
4747 Node* notail_ctl = stopped() ? NULL : control();
4748 set_control(tail_ctl);
4749 if (notail_ctl == NULL) {
4750 generate_clear_array(adr_type, dest, basic_elem_type,
4751 dest_tail, NULL,
4752 dest_size);
4753 } else {
4754 // Make a local merge.
4755 Node* done_ctl = new(C) RegionNode(3);
4756 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4757 done_ctl->init_req(1, notail_ctl);
4758 done_mem->init_req(1, memory(adr_type));
4759 generate_clear_array(adr_type, dest, basic_elem_type,
4760 dest_tail, NULL,
4761 dest_size);
4762 done_ctl->init_req(2, control());
4763 done_mem->init_req(2, memory(adr_type));
4764 set_control( _gvn.transform(done_ctl) );
4765 set_memory( _gvn.transform(done_mem), adr_type );
4766 }
4767 }
4768 }
4769
4770 BasicType copy_type = basic_elem_type;
4771 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4772 if (!stopped() && copy_type == T_OBJECT) {
4773 // If src and dest have compatible element types, we can copy bits.
4774 // Types S[] and D[] are compatible if D is a supertype of S.
4775 //
4776 // If they are not, we will use checked_oop_disjoint_arraycopy,
4777 // which performs a fast optimistic per-oop check, and backs off
4778 // further to JVM_ArrayCopy on the first per-oop check that fails.
4779 // (Actually, we don't move raw bits only; the GC requires card marks.)
4780
4781 // Get the Klass* for both src and dest
4782 Node* src_klass = load_object_klass(src);
4783 Node* dest_klass = load_object_klass(dest);
4784
4785 // Generate the subtype check.
4786 // This might fold up statically, or then again it might not.
4787 //
4788 // Non-static example: Copying List<String>.elements to a new String[].
4789 // The backing store for a List<String> is always an Object[],
4790 // but its elements are always type String, if the generic types
4791 // are correct at the source level.
4792 //
4793 // Test S[] against D[], not S against D, because (probably)
4794 // the secondary supertype cache is less busy for S[] than S.
4795 // This usually only matters when D is an interface.
4796 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4797 // Plug failing path into checked_oop_disjoint_arraycopy
4798 if (not_subtype_ctrl != top()) {
4799 PreserveJVMState pjvms(this);
4800 set_control(not_subtype_ctrl);
4801 // (At this point we can assume disjoint_bases, since types differ.)
4802 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
4803 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4804 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4805 Node* dest_elem_klass = _gvn.transform(n1);
4806 Node* cv = generate_checkcast_arraycopy(adr_type,
4807 dest_elem_klass,
4808 src, src_offset, dest, dest_offset,
4809 ConvI2X(copy_length), dest_uninitialized);
4810 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4811 checked_control = control();
4812 checked_i_o = i_o();
4813 checked_mem = memory(adr_type);
4814 checked_value = cv;
4815 }
4816 // At this point we know we do not need type checks on oop stores.
4817
4818 // Let's see if we need card marks:
4819 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4820 // If we do not need card marks, copy using the jint or jlong stub.
4821 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4822 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4823 "sizes agree");
4824 }
4825 }
4826
4827 if (!stopped()) {
4828 // Generate the fast path, if possible.
4829 PreserveJVMState pjvms(this);
4830 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4831 src, src_offset, dest, dest_offset,
4832 ConvI2X(copy_length), dest_uninitialized);
4833
4834 // Present the results of the fast call.
4835 result_region->init_req(fast_path, control());
4836 result_i_o ->init_req(fast_path, i_o());
4837 result_memory->init_req(fast_path, memory(adr_type));
4838 }
4839
4840 // Here are all the slow paths up to this point, in one bundle:
4841 slow_control = top();
4842 if (slow_region != NULL)
4843 slow_control = _gvn.transform(slow_region);
4844 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4845
4846 set_control(checked_control);
4847 if (!stopped()) {
4848 // Clean up after the checked call.
4849 // The returned value is either 0 or -1^K,
4850 // where K = number of partially transferred array elements.
4851 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4852 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4853 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4854
4855 // If it is 0, we are done, so transfer to the end.
4856 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4857 result_region->init_req(checked_path, checks_done);
4858 result_i_o ->init_req(checked_path, checked_i_o);
4859 result_memory->init_req(checked_path, checked_mem);
4860
4861 // If it is not zero, merge into the slow call.
4862 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4863 RegionNode* slow_reg2 = new(C) RegionNode(3);
4864 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4865 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4866 record_for_igvn(slow_reg2);
4867 slow_reg2 ->init_req(1, slow_control);
4868 slow_i_o2 ->init_req(1, slow_i_o);
4869 slow_mem2 ->init_req(1, slow_mem);
4870 slow_reg2 ->init_req(2, control());
4871 slow_i_o2 ->init_req(2, checked_i_o);
4872 slow_mem2 ->init_req(2, checked_mem);
4873
4874 slow_control = _gvn.transform(slow_reg2);
4875 slow_i_o = _gvn.transform(slow_i_o2);
4876 slow_mem = _gvn.transform(slow_mem2);
4877
4878 if (alloc != NULL) {
4879 // We'll restart from the very beginning, after zeroing the whole thing.
4880 // This can cause double writes, but that's OK since dest is brand new.
4881 // So we ignore the low 31 bits of the value returned from the stub.
4882 } else {
4883 // We must continue the copy exactly where it failed, or else
4884 // another thread might see the wrong number of writes to dest.
4885 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4886 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4887 slow_offset->init_req(1, intcon(0));
4888 slow_offset->init_req(2, checked_offset);
4889 slow_offset = _gvn.transform(slow_offset);
4890
4891 // Adjust the arguments by the conditionally incoming offset.
4892 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4893 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4894 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4895
4896 // Tweak the node variables to adjust the code produced below:
4897 src_offset = src_off_plus;
4898 dest_offset = dest_off_plus;
4899 copy_length = length_minus;
4900 }
4901 }
4902
4903 set_control(slow_control);
4904 if (!stopped()) {
4905 // Generate the slow path, if needed.
4906 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4907
4908 set_memory(slow_mem, adr_type);
4909 set_i_o(slow_i_o);
4910
4911 if (dest_uninitialized) {
4912 generate_clear_array(adr_type, dest, basic_elem_type,
4913 intcon(0), NULL,
4914 alloc->in(AllocateNode::AllocSize));
4915 }
4916
4917 generate_slow_arraycopy(adr_type,
4918 src, src_offset, dest, dest_offset,
4919 copy_length, /*dest_uninitialized*/false);
4920
4921 result_region->init_req(slow_call_path, control());
4922 result_i_o ->init_req(slow_call_path, i_o());
4923 result_memory->init_req(slow_call_path, memory(adr_type));
4924 }
4925
4926 // Remove unused edges.
4927 for (uint i = 1; i < result_region->req(); i++) {
4928 if (result_region->in(i) == NULL)
4929 result_region->init_req(i, top());
4930 }
4931
4932 // Finished; return the combined state.
4933 set_control( _gvn.transform(result_region) );
4934 set_i_o( _gvn.transform(result_i_o) );
4935 set_memory( _gvn.transform(result_memory), adr_type );
4936
4937 // The memory edges above are precise in order to model effects around
4938 // array copies accurately to allow value numbering of field loads around
4939 // arraycopy. Such field loads, both before and after, are common in Java
4940 // collections and similar classes involving header/array data structures.
4941 //
4942 // But with low number of register or when some registers are used or killed
4943 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4944 // The next memory barrier is added to avoid it. If the arraycopy can be
4945 // optimized away (which it can, sometimes) then we can manually remove
4946 // the membar also.
4947 //
4948 // Do not let reads from the cloned object float above the arraycopy.
4949 if (alloc != NULL) {
4950 // Do not let stores that initialize this object be reordered with
4951 // a subsequent store that would make this object accessible by
4952 // other threads.
4953 // Record what AllocateNode this StoreStore protects so that
4954 // escape analysis can go from the MemBarStoreStoreNode to the
4955 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4956 // based on the escape status of the AllocateNode.
4957 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4958 } else if (InsertMemBarAfterArraycopy)
4959 insert_mem_bar(Op_MemBarCPUOrder);
4960 }
4961
4962
4963 // Helper function which determines if an arraycopy immediately follows
4964 // an allocation, with no intervening tests or other escapes for the object.
4965 AllocateArrayNode*
4966 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
4967 RegionNode* slow_region) {
4968 if (stopped()) return NULL; // no fast path
4969 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
4970
4971 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
4972 if (alloc == NULL) return NULL;
4973
4974 Node* rawmem = memory(Compile::AliasIdxRaw);
4975 // Is the allocation's memory state untouched?
4976 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
4977 // Bail out if there have been raw-memory effects since the allocation.
4978 // (Example: There might have been a call or safepoint.)
4979 return NULL;
4980 }
4981 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
4982 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
4983 return NULL;
4984 }
4985
4986 // There must be no unexpected observers of this allocation.
4987 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
4988 Node* obs = ptr->fast_out(i);
4989 if (obs != this->map()) {
4990 return NULL;
4991 }
4992 }
4993
4994 // This arraycopy must unconditionally follow the allocation of the ptr.
4995 Node* alloc_ctl = ptr->in(0);
4996 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
4997
4998 Node* ctl = control();
4999 while (ctl != alloc_ctl) {
5000 // There may be guards which feed into the slow_region.
5001 // Any other control flow means that we might not get a chance
5002 // to finish initializing the allocated object.
5003 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5004 IfNode* iff = ctl->in(0)->as_If();
5005 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5006 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5007 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5008 ctl = iff->in(0); // This test feeds the known slow_region.
5009 continue;
5010 }
5011 // One more try: Various low-level checks bottom out in
5012 // uncommon traps. If the debug-info of the trap omits
5013 // any reference to the allocation, as we've already
5014 // observed, then there can be no objection to the trap.
5015 bool found_trap = false;
5016 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5017 Node* obs = not_ctl->fast_out(j);
5018 if (obs->in(0) == not_ctl && obs->is_Call() &&
5019 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5020 found_trap = true; break;
5021 }
5022 }
5023 if (found_trap) {
5024 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5025 continue;
5026 }
5027 }
5028 return NULL;
5029 }
5030
5031 // If we get this far, we have an allocation which immediately
5032 // precedes the arraycopy, and we can take over zeroing the new object.
5033 // The arraycopy will finish the initialization, and provide
5034 // a new control state to which we will anchor the destination pointer.
5035
5036 return alloc;
5037 }
5038
5039 // Helper for initialization of arrays, creating a ClearArray.
5040 // It writes zero bits in [start..end), within the body of an array object.
5041 // The memory effects are all chained onto the 'adr_type' alias category.
5042 //
5043 // Since the object is otherwise uninitialized, we are free
5044 // to put a little "slop" around the edges of the cleared area,
5045 // as long as it does not go back into the array's header,
5046 // or beyond the array end within the heap.
5047 //
5048 // The lower edge can be rounded down to the nearest jint and the
5049 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5050 //
5051 // Arguments:
5052 // adr_type memory slice where writes are generated
5053 // dest oop of the destination array
5054 // basic_elem_type element type of the destination
5055 // slice_idx array index of first element to store
5056 // slice_len number of elements to store (or NULL)
5057 // dest_size total size in bytes of the array object
5058 //
5059 // Exactly one of slice_len or dest_size must be non-NULL.
5060 // If dest_size is non-NULL, zeroing extends to the end of the object.
5061 // If slice_len is non-NULL, the slice_idx value must be a constant.
5062 void
5063 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5064 Node* dest,
5065 BasicType basic_elem_type,
5066 Node* slice_idx,
5067 Node* slice_len,
5068 Node* dest_size) {
5069 // one or the other but not both of slice_len and dest_size:
5070 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5071 if (slice_len == NULL) slice_len = top();
5072 if (dest_size == NULL) dest_size = top();
5073
5074 // operate on this memory slice:
5075 Node* mem = memory(adr_type); // memory slice to operate on
5076
5077 // scaling and rounding of indexes:
5078 int scale = exact_log2(type2aelembytes(basic_elem_type));
5079 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5080 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5081 int bump_bit = (-1 << scale) & BytesPerInt;
5082
5083 // determine constant starts and ends
5084 const intptr_t BIG_NEG = -128;
5085 assert(BIG_NEG + 2*abase < 0, "neg enough");
5086 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5087 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5088 if (slice_len_con == 0) {
5089 return; // nothing to do here
5090 }
5091 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5092 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5093 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5094 assert(end_con < 0, "not two cons");
5095 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5096 BytesPerLong);
5097 }
5098
5099 if (start_con >= 0 && end_con >= 0) {
5100 // Constant start and end. Simple.
5101 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5102 start_con, end_con, &_gvn);
5103 } else if (start_con >= 0 && dest_size != top()) {
5104 // Constant start, pre-rounded end after the tail of the array.
5105 Node* end = dest_size;
5106 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5107 start_con, end, &_gvn);
5108 } else if (start_con >= 0 && slice_len != top()) {
5109 // Constant start, non-constant end. End needs rounding up.
5110 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5111 intptr_t end_base = abase + (slice_idx_con << scale);
5112 int end_round = (-1 << scale) & (BytesPerLong - 1);
5113 Node* end = ConvI2X(slice_len);
5114 if (scale != 0)
5115 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5116 end_base += end_round;
5117 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5118 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5119 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5120 start_con, end, &_gvn);
5121 } else if (start_con < 0 && dest_size != top()) {
5122 // Non-constant start, pre-rounded end after the tail of the array.
5123 // This is almost certainly a "round-to-end" operation.
5124 Node* start = slice_idx;
5125 start = ConvI2X(start);
5126 if (scale != 0)
5127 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5128 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5129 if ((bump_bit | clear_low) != 0) {
5130 int to_clear = (bump_bit | clear_low);
5131 // Align up mod 8, then store a jint zero unconditionally
5132 // just before the mod-8 boundary.
5133 if (((abase + bump_bit) & ~to_clear) - bump_bit
5134 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5135 bump_bit = 0;
5136 assert((abase & to_clear) == 0, "array base must be long-aligned");
5137 } else {
5138 // Bump 'start' up to (or past) the next jint boundary:
5139 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5140 assert((abase & clear_low) == 0, "array base must be int-aligned");
5141 }
5142 // Round bumped 'start' down to jlong boundary in body of array.
5143 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5144 if (bump_bit != 0) {
5145 // Store a zero to the immediately preceding jint:
5146 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5147 Node* p1 = basic_plus_adr(dest, x1);
5148 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5149 mem = _gvn.transform(mem);
5150 }
5151 }
5152 Node* end = dest_size; // pre-rounded
5153 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5154 start, end, &_gvn);
5155 } else {
5156 // Non-constant start, unrounded non-constant end.
5157 // (Nobody zeroes a random midsection of an array using this routine.)
5158 ShouldNotReachHere(); // fix caller
5159 }
5160
5161 // Done.
5162 set_memory(mem, adr_type);
5163 }
5164
5165
5166 bool
5167 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5168 BasicType basic_elem_type,
5169 AllocateNode* alloc,
5170 Node* src, Node* src_offset,
5171 Node* dest, Node* dest_offset,
5172 Node* dest_size, bool dest_uninitialized) {
5173 // See if there is an advantage from block transfer.
5174 int scale = exact_log2(type2aelembytes(basic_elem_type));
5175 if (scale >= LogBytesPerLong)
5176 return false; // it is already a block transfer
5177
5178 // Look at the alignment of the starting offsets.
5179 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5180
5181 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5182 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5183 if (src_off_con < 0 || dest_off_con < 0)
5184 // At present, we can only understand constants.
5185 return false;
5186
5187 intptr_t src_off = abase + (src_off_con << scale);
5188 intptr_t dest_off = abase + (dest_off_con << scale);
5189
5190 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5191 // Non-aligned; too bad.
5192 // One more chance: Pick off an initial 32-bit word.
5193 // This is a common case, since abase can be odd mod 8.
5194 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5195 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5196 Node* sptr = basic_plus_adr(src, src_off);
5197 Node* dptr = basic_plus_adr(dest, dest_off);
5198 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5199 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5200 src_off += BytesPerInt;
5201 dest_off += BytesPerInt;
5202 } else {
5203 return false;
5204 }
5205 }
5206 assert(src_off % BytesPerLong == 0, "");
5207 assert(dest_off % BytesPerLong == 0, "");
5208
5209 // Do this copy by giant steps.
5210 Node* sptr = basic_plus_adr(src, src_off);
5211 Node* dptr = basic_plus_adr(dest, dest_off);
5212 Node* countx = dest_size;
5213 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5214 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5215
5216 bool disjoint_bases = true; // since alloc != NULL
5217 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5218 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5219
5220 return true;
5221 }
5222
5223
5224 // Helper function; generates code for the slow case.
5225 // We make a call to a runtime method which emulates the native method,
5226 // but without the native wrapper overhead.
5227 void
5228 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5229 Node* src, Node* src_offset,
5230 Node* dest, Node* dest_offset,
5231 Node* copy_length, bool dest_uninitialized) {
5232 assert(!dest_uninitialized, "Invariant");
5233 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5234 OptoRuntime::slow_arraycopy_Type(),
5235 OptoRuntime::slow_arraycopy_Java(),
5236 "slow_arraycopy", adr_type,
5237 src, src_offset, dest, dest_offset,
5238 copy_length);
5239
5240 // Handle exceptions thrown by this fellow:
5241 make_slow_call_ex(call, env()->Throwable_klass(), false);
5242 }
5243
5244 // Helper function; generates code for cases requiring runtime checks.
5245 Node*
5246 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5247 Node* dest_elem_klass,
5248 Node* src, Node* src_offset,
5249 Node* dest, Node* dest_offset,
5250 Node* copy_length, bool dest_uninitialized) {
5251 if (stopped()) return NULL;
5252
5253 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5254 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5255 return NULL;
5256 }
5257
5258 // Pick out the parameters required to perform a store-check
5259 // for the target array. This is an optimistic check. It will
5260 // look in each non-null element's class, at the desired klass's
5261 // super_check_offset, for the desired klass.
5262 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5263 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5264 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5265 Node* check_offset = ConvI2X(_gvn.transform(n3));
5266 Node* check_value = dest_elem_klass;
5267
5268 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5269 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5270
5271 // (We know the arrays are never conjoint, because their types differ.)
5272 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5273 OptoRuntime::checkcast_arraycopy_Type(),
5274 copyfunc_addr, "checkcast_arraycopy", adr_type,
5275 // five arguments, of which two are
5276 // intptr_t (jlong in LP64)
5277 src_start, dest_start,
5278 copy_length XTOP,
5279 check_offset XTOP,
5280 check_value);
5281
5282 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5283 }
5284
5285
5286 // Helper function; generates code for cases requiring runtime checks.
5287 Node*
5288 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5289 Node* src, Node* src_offset,
5290 Node* dest, Node* dest_offset,
5291 Node* copy_length, bool dest_uninitialized) {
5292 assert(!dest_uninitialized, "Invariant");
5293 if (stopped()) return NULL;
5294 address copyfunc_addr = StubRoutines::generic_arraycopy();
5295 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5296 return NULL;
5297 }
5298
5299 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5300 OptoRuntime::generic_arraycopy_Type(),
5301 copyfunc_addr, "generic_arraycopy", adr_type,
5302 src, src_offset, dest, dest_offset, copy_length);
5303
5304 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5305 }
5306
5307 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5308 void
5309 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5310 BasicType basic_elem_type,
5311 bool disjoint_bases,
5312 Node* src, Node* src_offset,
5313 Node* dest, Node* dest_offset,
5314 Node* copy_length, bool dest_uninitialized) {
5315 if (stopped()) return; // nothing to do
5316
5317 Node* src_start = src;
5318 Node* dest_start = dest;
5319 if (src_offset != NULL || dest_offset != NULL) {
5320 assert(src_offset != NULL && dest_offset != NULL, "");
5321 src_start = array_element_address(src, src_offset, basic_elem_type);
5322 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5323 }
5324
5325 // Figure out which arraycopy runtime method to call.
5326 const char* copyfunc_name = "arraycopy";
5327 address copyfunc_addr =
5328 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5329 disjoint_bases, copyfunc_name, dest_uninitialized);
5330
5331 // Call it. Note that the count_ix value is not scaled to a byte-size.
5332 make_runtime_call(RC_LEAF|RC_NO_FP,
5333 OptoRuntime::fast_arraycopy_Type(),
5334 copyfunc_addr, copyfunc_name, adr_type,
5335 src_start, dest_start, copy_length XTOP);
5336 }
5337
5338 //----------------------------inline_reference_get----------------------------
5339 // public T java.lang.ref.Reference.get();
5340 bool LibraryCallKit::inline_reference_get() {
5341 const int referent_offset = java_lang_ref_Reference::referent_offset;
5342 guarantee(referent_offset > 0, "should have already been set");
5343
5344 // Get the argument:
5345 Node* reference_obj = null_check_receiver();
5346 if (stopped()) return true;
5347
5348 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5349
5350 ciInstanceKlass* klass = env()->Object_klass();
5351 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5352
5353 Node* no_ctrl = NULL;
5354 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5355
5356 // Use the pre-barrier to record the value in the referent field
5357 pre_barrier(false /* do_load */,
5358 control(),
5359 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5360 result /* pre_val */,
5361 T_OBJECT);
5362
5363 // Add memory barrier to prevent commoning reads from this field
5364 // across safepoint since GC can change its value.
5365 insert_mem_bar(Op_MemBarCPUOrder);
5366
5367 set_result(result);
5368 return true;
5369 }
5370
5371
5372 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5373 bool is_exact=true, bool is_static=false) {
5374
5375 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5376 assert(tinst != NULL, "obj is null");
5377 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5378 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5379
5380 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5381 ciSymbol::make(fieldTypeString),
5382 is_static);
5383 if (field == NULL) return (Node *) NULL;
5384 assert (field != NULL, "undefined field");
5385
5386 // Next code copied from Parse::do_get_xxx():
5387
5388 // Compute address and memory type.
5389 int offset = field->offset_in_bytes();
5390 bool is_vol = field->is_volatile();
5391 ciType* field_klass = field->type();
5392 assert(field_klass->is_loaded(), "should be loaded");
5393 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5394 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5395 BasicType bt = field->layout_type();
5396
5397 // Build the resultant type of the load
5398 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5399
5400 // Build the load.
5401 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5402 return loadedField;
5403 }
5404
5405
5406 //------------------------------inline_aescrypt_Block-----------------------
5407 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5408 address stubAddr;
5409 const char *stubName;
5410 assert(UseAES, "need AES instruction support");
5411
5412 switch(id) {
5413 case vmIntrinsics::_aescrypt_encryptBlock:
5414 stubAddr = StubRoutines::aescrypt_encryptBlock();
5415 stubName = "aescrypt_encryptBlock";
5416 break;
5417 case vmIntrinsics::_aescrypt_decryptBlock:
5418 stubAddr = StubRoutines::aescrypt_decryptBlock();
5419 stubName = "aescrypt_decryptBlock";
5420 break;
5421 }
5422 if (stubAddr == NULL) return false;
5423
5424 Node* aescrypt_object = argument(0);
5425 Node* src = argument(1);
5426 Node* src_offset = argument(2);
5427 Node* dest = argument(3);
5428 Node* dest_offset = argument(4);
5429
5430 // (1) src and dest are arrays.
5431 const Type* src_type = src->Value(&_gvn);
5432 const Type* dest_type = dest->Value(&_gvn);
5433 const TypeAryPtr* top_src = src_type->isa_aryptr();
5434 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5435 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5436
5437 // for the quick and dirty code we will skip all the checks.
5438 // we are just trying to get the call to be generated.
5439 Node* src_start = src;
5440 Node* dest_start = dest;
5441 if (src_offset != NULL || dest_offset != NULL) {
5442 assert(src_offset != NULL && dest_offset != NULL, "");
5443 src_start = array_element_address(src, src_offset, T_BYTE);
5444 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5445 }
5446
5447 // now need to get the start of its expanded key array
5448 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5449 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5450 if (k_start == NULL) return false;
5451
5452 // Call the stub.
5453 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5454 stubAddr, stubName, TypePtr::BOTTOM,
5455 src_start, dest_start, k_start);
5456
5457 return true;
5458 }
5459
5460 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5461 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5462 address stubAddr;
5463 const char *stubName;
5464
5465 assert(UseAES, "need AES instruction support");
5466
5467 switch(id) {
5468 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5469 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5470 stubName = "cipherBlockChaining_encryptAESCrypt";
5471 break;
5472 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5473 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5474 stubName = "cipherBlockChaining_decryptAESCrypt";
5475 break;
5476 }
5477 if (stubAddr == NULL) return false;
5478
5479 Node* cipherBlockChaining_object = argument(0);
5480 Node* src = argument(1);
5481 Node* src_offset = argument(2);
5482 Node* len = argument(3);
5483 Node* dest = argument(4);
5484 Node* dest_offset = argument(5);
5485
5486 // (1) src and dest are arrays.
5487 const Type* src_type = src->Value(&_gvn);
5488 const Type* dest_type = dest->Value(&_gvn);
5489 const TypeAryPtr* top_src = src_type->isa_aryptr();
5490 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5491 assert (top_src != NULL && top_src->klass() != NULL
5492 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5493
5494 // checks are the responsibility of the caller
5495 Node* src_start = src;
5496 Node* dest_start = dest;
5497 if (src_offset != NULL || dest_offset != NULL) {
5498 assert(src_offset != NULL && dest_offset != NULL, "");
5499 src_start = array_element_address(src, src_offset, T_BYTE);
5500 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5501 }
5502
5503 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5504 // (because of the predicated logic executed earlier).
5505 // so we cast it here safely.
5506 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5507
5508 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5509 if (embeddedCipherObj == NULL) return false;
5510
5511 // cast it to what we know it will be at runtime
5512 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5513 assert(tinst != NULL, "CBC obj is null");
5514 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5515 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5516 if (!klass_AESCrypt->is_loaded()) return false;
5517
5518 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5519 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5520 const TypeOopPtr* xtype = aklass->as_instance_type();
5521 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5522 aescrypt_object = _gvn.transform(aescrypt_object);
5523
5524 // we need to get the start of the aescrypt_object's expanded key array
5525 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5526 if (k_start == NULL) return false;
5527
5528 // similarly, get the start address of the r vector
5529 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5530 if (objRvec == NULL) return false;
5531 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5532
5533 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5534 make_runtime_call(RC_LEAF|RC_NO_FP,
5535 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5536 stubAddr, stubName, TypePtr::BOTTOM,
5537 src_start, dest_start, k_start, r_start, len);
5538
5539 // return is void so no result needs to be pushed
5540
5541 return true;
5542 }
5543
5544 //------------------------------get_key_start_from_aescrypt_object-----------------------
5545 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5546 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5547 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5548 if (objAESCryptKey == NULL) return (Node *) NULL;
5549
5550 // now have the array, need to get the start address of the K array
5551 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5552 return k_start;
5553 }
5554
5555 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5556 // Return node representing slow path of predicate check.
5557 // the pseudo code we want to emulate with this predicate is:
5558 // for encryption:
5559 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5560 // for decryption:
5561 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5562 // note cipher==plain is more conservative than the original java code but that's OK
5563 //
5564 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5565 // First, check receiver for NULL since it is virtual method.
5566 Node* objCBC = argument(0);
5567 objCBC = null_check(objCBC);
5568
5569 if (stopped()) return NULL; // Always NULL
5570
5571 // Load embeddedCipher field of CipherBlockChaining object.
5572 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5573
5574 // get AESCrypt klass for instanceOf check
5575 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5576 // will have same classloader as CipherBlockChaining object
5577 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5578 assert(tinst != NULL, "CBCobj is null");
5579 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5580
5581 // we want to do an instanceof comparison against the AESCrypt class
5582 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5583 if (!klass_AESCrypt->is_loaded()) {
5584 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5585 Node* ctrl = control();
5586 set_control(top()); // no regular fast path
5587 return ctrl;
5588 }
5589 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5590
5591 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5592 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5593 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5594
5595 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5596
5597 // for encryption, we are done
5598 if (!decrypting)
5599 return instof_false; // even if it is NULL
5600
5601 // for decryption, we need to add a further check to avoid
5602 // taking the intrinsic path when cipher and plain are the same
5603 // see the original java code for why.
5604 RegionNode* region = new(C) RegionNode(3);
5605 region->init_req(1, instof_false);
5606 Node* src = argument(1);
5607 Node* dest = argument(4);
5608 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5609 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5610 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5611 region->init_req(2, src_dest_conjoint);
5612
5613 record_for_igvn(region);
5614 return _gvn.transform(region);
5615 }