Fri, 21 Dec 2012 01:39:34 -0800
Merge
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 */
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"
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;
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 };
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
74 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
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 }
97 virtual LibraryCallKit* is_LibraryCallKit() const { return (LibraryCallKit*)this; }
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(); }
105 bool try_to_inline();
106 Node* try_to_predicate();
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 }
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 }
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; }
125 virtual int reexecute_sp() { return _reexecute_sp; }
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);
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();
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();
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_unsafe_fence(vmIntrinsics::ID id);
286 bool inline_fp_conversions(vmIntrinsics::ID id);
287 bool inline_number_methods(vmIntrinsics::ID id);
288 bool inline_reference_get();
289 bool inline_aescrypt_Block(vmIntrinsics::ID id);
290 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
291 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
292 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
293 };
296 //---------------------------make_vm_intrinsic----------------------------
297 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
298 vmIntrinsics::ID id = m->intrinsic_id();
299 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
301 if (DisableIntrinsic[0] != '\0'
302 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
303 // disabled by a user request on the command line:
304 // example: -XX:DisableIntrinsic=_hashCode,_getClass
305 return NULL;
306 }
308 if (!m->is_loaded()) {
309 // do not attempt to inline unloaded methods
310 return NULL;
311 }
313 // Only a few intrinsics implement a virtual dispatch.
314 // They are expensive calls which are also frequently overridden.
315 if (is_virtual) {
316 switch (id) {
317 case vmIntrinsics::_hashCode:
318 case vmIntrinsics::_clone:
319 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
320 break;
321 default:
322 return NULL;
323 }
324 }
326 // -XX:-InlineNatives disables nearly all intrinsics:
327 if (!InlineNatives) {
328 switch (id) {
329 case vmIntrinsics::_indexOf:
330 case vmIntrinsics::_compareTo:
331 case vmIntrinsics::_equals:
332 case vmIntrinsics::_equalsC:
333 case vmIntrinsics::_getAndAddInt:
334 case vmIntrinsics::_getAndAddLong:
335 case vmIntrinsics::_getAndSetInt:
336 case vmIntrinsics::_getAndSetLong:
337 case vmIntrinsics::_getAndSetObject:
338 case vmIntrinsics::_loadFence:
339 case vmIntrinsics::_storeFence:
340 case vmIntrinsics::_fullFence:
341 break; // InlineNatives does not control String.compareTo
342 case vmIntrinsics::_Reference_get:
343 break; // InlineNatives does not control Reference.get
344 default:
345 return NULL;
346 }
347 }
349 bool is_predicted = false;
351 switch (id) {
352 case vmIntrinsics::_compareTo:
353 if (!SpecialStringCompareTo) return NULL;
354 if (!Matcher::match_rule_supported(Op_StrComp)) return NULL;
355 break;
356 case vmIntrinsics::_indexOf:
357 if (!SpecialStringIndexOf) return NULL;
358 break;
359 case vmIntrinsics::_equals:
360 if (!SpecialStringEquals) return NULL;
361 if (!Matcher::match_rule_supported(Op_StrEquals)) return NULL;
362 break;
363 case vmIntrinsics::_equalsC:
364 if (!SpecialArraysEquals) return NULL;
365 if (!Matcher::match_rule_supported(Op_AryEq)) return NULL;
366 break;
367 case vmIntrinsics::_arraycopy:
368 if (!InlineArrayCopy) return NULL;
369 break;
370 case vmIntrinsics::_copyMemory:
371 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
372 if (!InlineArrayCopy) return NULL;
373 break;
374 case vmIntrinsics::_hashCode:
375 if (!InlineObjectHash) return NULL;
376 break;
377 case vmIntrinsics::_clone:
378 case vmIntrinsics::_copyOf:
379 case vmIntrinsics::_copyOfRange:
380 if (!InlineObjectCopy) return NULL;
381 // These also use the arraycopy intrinsic mechanism:
382 if (!InlineArrayCopy) return NULL;
383 break;
384 case vmIntrinsics::_checkIndex:
385 // We do not intrinsify this. The optimizer does fine with it.
386 return NULL;
388 case vmIntrinsics::_getCallerClass:
389 if (!UseNewReflection) return NULL;
390 if (!InlineReflectionGetCallerClass) return NULL;
391 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
392 break;
394 case vmIntrinsics::_bitCount_i:
395 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
396 break;
398 case vmIntrinsics::_bitCount_l:
399 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
400 break;
402 case vmIntrinsics::_numberOfLeadingZeros_i:
403 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
404 break;
406 case vmIntrinsics::_numberOfLeadingZeros_l:
407 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
408 break;
410 case vmIntrinsics::_numberOfTrailingZeros_i:
411 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
412 break;
414 case vmIntrinsics::_numberOfTrailingZeros_l:
415 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
416 break;
418 case vmIntrinsics::_reverseBytes_c:
419 if (!Matcher::match_rule_supported(Op_ReverseBytesUS)) return NULL;
420 break;
421 case vmIntrinsics::_reverseBytes_s:
422 if (!Matcher::match_rule_supported(Op_ReverseBytesS)) return NULL;
423 break;
424 case vmIntrinsics::_reverseBytes_i:
425 if (!Matcher::match_rule_supported(Op_ReverseBytesI)) return NULL;
426 break;
427 case vmIntrinsics::_reverseBytes_l:
428 if (!Matcher::match_rule_supported(Op_ReverseBytesL)) return NULL;
429 break;
431 case vmIntrinsics::_Reference_get:
432 // Use the intrinsic version of Reference.get() so that the value in
433 // the referent field can be registered by the G1 pre-barrier code.
434 // Also add memory barrier to prevent commoning reads from this field
435 // across safepoint since GC can change it value.
436 break;
438 case vmIntrinsics::_compareAndSwapObject:
439 #ifdef _LP64
440 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
441 #endif
442 break;
444 case vmIntrinsics::_compareAndSwapLong:
445 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
446 break;
448 case vmIntrinsics::_getAndAddInt:
449 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
450 break;
452 case vmIntrinsics::_getAndAddLong:
453 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
454 break;
456 case vmIntrinsics::_getAndSetInt:
457 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
458 break;
460 case vmIntrinsics::_getAndSetLong:
461 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
462 break;
464 case vmIntrinsics::_getAndSetObject:
465 #ifdef _LP64
466 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
467 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
468 break;
469 #else
470 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
471 break;
472 #endif
474 case vmIntrinsics::_aescrypt_encryptBlock:
475 case vmIntrinsics::_aescrypt_decryptBlock:
476 if (!UseAESIntrinsics) return NULL;
477 break;
479 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
480 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
481 if (!UseAESIntrinsics) return NULL;
482 // these two require the predicated logic
483 is_predicted = true;
484 break;
486 default:
487 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
488 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
489 break;
490 }
492 // -XX:-InlineClassNatives disables natives from the Class class.
493 // The flag applies to all reflective calls, notably Array.newArray
494 // (visible to Java programmers as Array.newInstance).
495 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
496 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
497 if (!InlineClassNatives) return NULL;
498 }
500 // -XX:-InlineThreadNatives disables natives from the Thread class.
501 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
502 if (!InlineThreadNatives) return NULL;
503 }
505 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
506 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
507 m->holder()->name() == ciSymbol::java_lang_Float() ||
508 m->holder()->name() == ciSymbol::java_lang_Double()) {
509 if (!InlineMathNatives) return NULL;
510 }
512 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
513 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
514 if (!InlineUnsafeOps) return NULL;
515 }
517 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
518 }
520 //----------------------register_library_intrinsics-----------------------
521 // Initialize this file's data structures, for each Compile instance.
522 void Compile::register_library_intrinsics() {
523 // Nothing to do here.
524 }
526 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
527 LibraryCallKit kit(jvms, this);
528 Compile* C = kit.C;
529 int nodes = C->unique();
530 #ifndef PRODUCT
531 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
532 char buf[1000];
533 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
534 tty->print_cr("Intrinsic %s", str);
535 }
536 #endif
537 ciMethod* callee = kit.callee();
538 const int bci = kit.bci();
540 // Try to inline the intrinsic.
541 if (kit.try_to_inline()) {
542 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
543 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
544 }
545 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
546 if (C->log()) {
547 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
548 vmIntrinsics::name_at(intrinsic_id()),
549 (is_virtual() ? " virtual='1'" : ""),
550 C->unique() - nodes);
551 }
552 // Push the result from the inlined method onto the stack.
553 kit.push_result();
554 return kit.transfer_exceptions_into_jvms();
555 }
557 // The intrinsic bailed out
558 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
559 if (jvms->has_method()) {
560 // Not a root compile.
561 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
562 C->print_inlining(callee, jvms->depth() - 1, bci, msg);
563 } else {
564 // Root compile
565 tty->print("Did not generate intrinsic %s%s at bci:%d in",
566 vmIntrinsics::name_at(intrinsic_id()),
567 (is_virtual() ? " (virtual)" : ""), bci);
568 }
569 }
570 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
571 return NULL;
572 }
574 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
575 LibraryCallKit kit(jvms, this);
576 Compile* C = kit.C;
577 int nodes = C->unique();
578 #ifndef PRODUCT
579 assert(is_predicted(), "sanity");
580 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
581 char buf[1000];
582 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
583 tty->print_cr("Predicate for intrinsic %s", str);
584 }
585 #endif
586 ciMethod* callee = kit.callee();
587 const int bci = kit.bci();
589 Node* slow_ctl = kit.try_to_predicate();
590 if (!kit.failing()) {
591 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
592 C->print_inlining(callee, jvms->depth() - 1, bci, is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
593 }
594 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
595 if (C->log()) {
596 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
597 vmIntrinsics::name_at(intrinsic_id()),
598 (is_virtual() ? " virtual='1'" : ""),
599 C->unique() - nodes);
600 }
601 return slow_ctl; // Could be NULL if the check folds.
602 }
604 // The intrinsic bailed out
605 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
606 if (jvms->has_method()) {
607 // Not a root compile.
608 const char* msg = "failed to generate predicate for intrinsic";
609 C->print_inlining(kit.callee(), jvms->depth() - 1, bci, msg);
610 } else {
611 // Root compile
612 C->print_inlining_stream()->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
613 vmIntrinsics::name_at(intrinsic_id()),
614 (is_virtual() ? " (virtual)" : ""), bci);
615 }
616 }
617 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
618 return NULL;
619 }
621 bool LibraryCallKit::try_to_inline() {
622 // Handle symbolic names for otherwise undistinguished boolean switches:
623 const bool is_store = true;
624 const bool is_native_ptr = true;
625 const bool is_static = true;
626 const bool is_volatile = true;
628 if (!jvms()->has_method()) {
629 // Root JVMState has a null method.
630 assert(map()->memory()->Opcode() == Op_Parm, "");
631 // Insert the memory aliasing node
632 set_all_memory(reset_memory());
633 }
634 assert(merged_memory(), "");
637 switch (intrinsic_id()) {
638 case vmIntrinsics::_hashCode: return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
639 case vmIntrinsics::_identityHashCode: return inline_native_hashcode(/*!virtual*/ false, is_static);
640 case vmIntrinsics::_getClass: return inline_native_getClass();
642 case vmIntrinsics::_dsin:
643 case vmIntrinsics::_dcos:
644 case vmIntrinsics::_dtan:
645 case vmIntrinsics::_dabs:
646 case vmIntrinsics::_datan2:
647 case vmIntrinsics::_dsqrt:
648 case vmIntrinsics::_dexp:
649 case vmIntrinsics::_dlog:
650 case vmIntrinsics::_dlog10:
651 case vmIntrinsics::_dpow: return inline_math_native(intrinsic_id());
653 case vmIntrinsics::_min:
654 case vmIntrinsics::_max: return inline_min_max(intrinsic_id());
656 case vmIntrinsics::_arraycopy: return inline_arraycopy();
658 case vmIntrinsics::_compareTo: return inline_string_compareTo();
659 case vmIntrinsics::_indexOf: return inline_string_indexOf();
660 case vmIntrinsics::_equals: return inline_string_equals();
662 case vmIntrinsics::_getObject: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, !is_volatile);
663 case vmIntrinsics::_getBoolean: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, !is_volatile);
664 case vmIntrinsics::_getByte: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, !is_volatile);
665 case vmIntrinsics::_getShort: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, !is_volatile);
666 case vmIntrinsics::_getChar: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, !is_volatile);
667 case vmIntrinsics::_getInt: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, !is_volatile);
668 case vmIntrinsics::_getLong: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, !is_volatile);
669 case vmIntrinsics::_getFloat: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, !is_volatile);
670 case vmIntrinsics::_getDouble: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
672 case vmIntrinsics::_putObject: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, !is_volatile);
673 case vmIntrinsics::_putBoolean: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, !is_volatile);
674 case vmIntrinsics::_putByte: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, !is_volatile);
675 case vmIntrinsics::_putShort: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, !is_volatile);
676 case vmIntrinsics::_putChar: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, !is_volatile);
677 case vmIntrinsics::_putInt: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, !is_volatile);
678 case vmIntrinsics::_putLong: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, !is_volatile);
679 case vmIntrinsics::_putFloat: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, !is_volatile);
680 case vmIntrinsics::_putDouble: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, !is_volatile);
682 case vmIntrinsics::_getByte_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_BYTE, !is_volatile);
683 case vmIntrinsics::_getShort_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_SHORT, !is_volatile);
684 case vmIntrinsics::_getChar_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_CHAR, !is_volatile);
685 case vmIntrinsics::_getInt_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_INT, !is_volatile);
686 case vmIntrinsics::_getLong_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_LONG, !is_volatile);
687 case vmIntrinsics::_getFloat_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_FLOAT, !is_volatile);
688 case vmIntrinsics::_getDouble_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_DOUBLE, !is_volatile);
689 case vmIntrinsics::_getAddress_raw: return inline_unsafe_access( is_native_ptr, !is_store, T_ADDRESS, !is_volatile);
691 case vmIntrinsics::_putByte_raw: return inline_unsafe_access( is_native_ptr, is_store, T_BYTE, !is_volatile);
692 case vmIntrinsics::_putShort_raw: return inline_unsafe_access( is_native_ptr, is_store, T_SHORT, !is_volatile);
693 case vmIntrinsics::_putChar_raw: return inline_unsafe_access( is_native_ptr, is_store, T_CHAR, !is_volatile);
694 case vmIntrinsics::_putInt_raw: return inline_unsafe_access( is_native_ptr, is_store, T_INT, !is_volatile);
695 case vmIntrinsics::_putLong_raw: return inline_unsafe_access( is_native_ptr, is_store, T_LONG, !is_volatile);
696 case vmIntrinsics::_putFloat_raw: return inline_unsafe_access( is_native_ptr, is_store, T_FLOAT, !is_volatile);
697 case vmIntrinsics::_putDouble_raw: return inline_unsafe_access( is_native_ptr, is_store, T_DOUBLE, !is_volatile);
698 case vmIntrinsics::_putAddress_raw: return inline_unsafe_access( is_native_ptr, is_store, T_ADDRESS, !is_volatile);
700 case vmIntrinsics::_getObjectVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, is_volatile);
701 case vmIntrinsics::_getBooleanVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, is_volatile);
702 case vmIntrinsics::_getByteVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, is_volatile);
703 case vmIntrinsics::_getShortVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, is_volatile);
704 case vmIntrinsics::_getCharVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, is_volatile);
705 case vmIntrinsics::_getIntVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, is_volatile);
706 case vmIntrinsics::_getLongVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, is_volatile);
707 case vmIntrinsics::_getFloatVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, is_volatile);
708 case vmIntrinsics::_getDoubleVolatile: return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, is_volatile);
710 case vmIntrinsics::_putObjectVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, is_volatile);
711 case vmIntrinsics::_putBooleanVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, is_volatile);
712 case vmIntrinsics::_putByteVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, is_volatile);
713 case vmIntrinsics::_putShortVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, is_volatile);
714 case vmIntrinsics::_putCharVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, is_volatile);
715 case vmIntrinsics::_putIntVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_INT, is_volatile);
716 case vmIntrinsics::_putLongVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, is_volatile);
717 case vmIntrinsics::_putFloatVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, is_volatile);
718 case vmIntrinsics::_putDoubleVolatile: return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, is_volatile);
720 case vmIntrinsics::_prefetchRead: return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
721 case vmIntrinsics::_prefetchWrite: return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
722 case vmIntrinsics::_prefetchReadStatic: return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
723 case vmIntrinsics::_prefetchWriteStatic: return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
725 case vmIntrinsics::_compareAndSwapObject: return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
726 case vmIntrinsics::_compareAndSwapInt: return inline_unsafe_load_store(T_INT, LS_cmpxchg);
727 case vmIntrinsics::_compareAndSwapLong: return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
729 case vmIntrinsics::_putOrderedObject: return inline_unsafe_ordered_store(T_OBJECT);
730 case vmIntrinsics::_putOrderedInt: return inline_unsafe_ordered_store(T_INT);
731 case vmIntrinsics::_putOrderedLong: return inline_unsafe_ordered_store(T_LONG);
733 case vmIntrinsics::_getAndAddInt: return inline_unsafe_load_store(T_INT, LS_xadd);
734 case vmIntrinsics::_getAndAddLong: return inline_unsafe_load_store(T_LONG, LS_xadd);
735 case vmIntrinsics::_getAndSetInt: return inline_unsafe_load_store(T_INT, LS_xchg);
736 case vmIntrinsics::_getAndSetLong: return inline_unsafe_load_store(T_LONG, LS_xchg);
737 case vmIntrinsics::_getAndSetObject: return inline_unsafe_load_store(T_OBJECT, LS_xchg);
739 case vmIntrinsics::_loadFence:
740 case vmIntrinsics::_storeFence:
741 case vmIntrinsics::_fullFence: return inline_unsafe_fence(intrinsic_id());
743 case vmIntrinsics::_currentThread: return inline_native_currentThread();
744 case vmIntrinsics::_isInterrupted: return inline_native_isInterrupted();
746 #ifdef TRACE_HAVE_INTRINSICS
747 case vmIntrinsics::_classID: return inline_native_classID();
748 case vmIntrinsics::_threadID: return inline_native_threadID();
749 case vmIntrinsics::_counterTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
750 #endif
751 case vmIntrinsics::_currentTimeMillis: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
752 case vmIntrinsics::_nanoTime: return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
753 case vmIntrinsics::_allocateInstance: return inline_unsafe_allocate();
754 case vmIntrinsics::_copyMemory: return inline_unsafe_copyMemory();
755 case vmIntrinsics::_newArray: return inline_native_newArray();
756 case vmIntrinsics::_getLength: return inline_native_getLength();
757 case vmIntrinsics::_copyOf: return inline_array_copyOf(false);
758 case vmIntrinsics::_copyOfRange: return inline_array_copyOf(true);
759 case vmIntrinsics::_equalsC: return inline_array_equals();
760 case vmIntrinsics::_clone: return inline_native_clone(intrinsic()->is_virtual());
762 case vmIntrinsics::_isAssignableFrom: return inline_native_subtype_check();
764 case vmIntrinsics::_isInstance:
765 case vmIntrinsics::_getModifiers:
766 case vmIntrinsics::_isInterface:
767 case vmIntrinsics::_isArray:
768 case vmIntrinsics::_isPrimitive:
769 case vmIntrinsics::_getSuperclass:
770 case vmIntrinsics::_getComponentType:
771 case vmIntrinsics::_getClassAccessFlags: return inline_native_Class_query(intrinsic_id());
773 case vmIntrinsics::_floatToRawIntBits:
774 case vmIntrinsics::_floatToIntBits:
775 case vmIntrinsics::_intBitsToFloat:
776 case vmIntrinsics::_doubleToRawLongBits:
777 case vmIntrinsics::_doubleToLongBits:
778 case vmIntrinsics::_longBitsToDouble: return inline_fp_conversions(intrinsic_id());
780 case vmIntrinsics::_numberOfLeadingZeros_i:
781 case vmIntrinsics::_numberOfLeadingZeros_l:
782 case vmIntrinsics::_numberOfTrailingZeros_i:
783 case vmIntrinsics::_numberOfTrailingZeros_l:
784 case vmIntrinsics::_bitCount_i:
785 case vmIntrinsics::_bitCount_l:
786 case vmIntrinsics::_reverseBytes_i:
787 case vmIntrinsics::_reverseBytes_l:
788 case vmIntrinsics::_reverseBytes_s:
789 case vmIntrinsics::_reverseBytes_c: return inline_number_methods(intrinsic_id());
791 case vmIntrinsics::_getCallerClass: return inline_native_Reflection_getCallerClass();
793 case vmIntrinsics::_Reference_get: return inline_reference_get();
795 case vmIntrinsics::_aescrypt_encryptBlock:
796 case vmIntrinsics::_aescrypt_decryptBlock: return inline_aescrypt_Block(intrinsic_id());
798 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
799 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
800 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
802 default:
803 // If you get here, it may be that someone has added a new intrinsic
804 // to the list in vmSymbols.hpp without implementing it here.
805 #ifndef PRODUCT
806 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
807 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
808 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
809 }
810 #endif
811 return false;
812 }
813 }
815 Node* LibraryCallKit::try_to_predicate() {
816 if (!jvms()->has_method()) {
817 // Root JVMState has a null method.
818 assert(map()->memory()->Opcode() == Op_Parm, "");
819 // Insert the memory aliasing node
820 set_all_memory(reset_memory());
821 }
822 assert(merged_memory(), "");
824 switch (intrinsic_id()) {
825 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
826 return inline_cipherBlockChaining_AESCrypt_predicate(false);
827 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
828 return inline_cipherBlockChaining_AESCrypt_predicate(true);
830 default:
831 // If you get here, it may be that someone has added a new intrinsic
832 // to the list in vmSymbols.hpp without implementing it here.
833 #ifndef PRODUCT
834 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
835 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
836 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
837 }
838 #endif
839 Node* slow_ctl = control();
840 set_control(top()); // No fast path instrinsic
841 return slow_ctl;
842 }
843 }
845 //------------------------------set_result-------------------------------
846 // Helper function for finishing intrinsics.
847 void LibraryCallKit::set_result(RegionNode* region, PhiNode* value) {
848 record_for_igvn(region);
849 set_control(_gvn.transform(region));
850 set_result( _gvn.transform(value));
851 assert(value->type()->basic_type() == result()->bottom_type()->basic_type(), "sanity");
852 }
854 //------------------------------generate_guard---------------------------
855 // Helper function for generating guarded fast-slow graph structures.
856 // The given 'test', if true, guards a slow path. If the test fails
857 // then a fast path can be taken. (We generally hope it fails.)
858 // In all cases, GraphKit::control() is updated to the fast path.
859 // The returned value represents the control for the slow path.
860 // The return value is never 'top'; it is either a valid control
861 // or NULL if it is obvious that the slow path can never be taken.
862 // Also, if region and the slow control are not NULL, the slow edge
863 // is appended to the region.
864 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
865 if (stopped()) {
866 // Already short circuited.
867 return NULL;
868 }
870 // Build an if node and its projections.
871 // If test is true we take the slow path, which we assume is uncommon.
872 if (_gvn.type(test) == TypeInt::ZERO) {
873 // The slow branch is never taken. No need to build this guard.
874 return NULL;
875 }
877 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
879 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
880 if (if_slow == top()) {
881 // The slow branch is never taken. No need to build this guard.
882 return NULL;
883 }
885 if (region != NULL)
886 region->add_req(if_slow);
888 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
889 set_control(if_fast);
891 return if_slow;
892 }
894 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
895 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
896 }
897 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
898 return generate_guard(test, region, PROB_FAIR);
899 }
901 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
902 Node* *pos_index) {
903 if (stopped())
904 return NULL; // already stopped
905 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
906 return NULL; // index is already adequately typed
907 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
908 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
909 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
910 if (is_neg != NULL && pos_index != NULL) {
911 // Emulate effect of Parse::adjust_map_after_if.
912 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
913 ccast->set_req(0, control());
914 (*pos_index) = _gvn.transform(ccast);
915 }
916 return is_neg;
917 }
919 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
920 Node* *pos_index) {
921 if (stopped())
922 return NULL; // already stopped
923 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
924 return NULL; // index is already adequately typed
925 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
926 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
927 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
928 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
929 if (is_notp != NULL && pos_index != NULL) {
930 // Emulate effect of Parse::adjust_map_after_if.
931 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
932 ccast->set_req(0, control());
933 (*pos_index) = _gvn.transform(ccast);
934 }
935 return is_notp;
936 }
938 // Make sure that 'position' is a valid limit index, in [0..length].
939 // There are two equivalent plans for checking this:
940 // A. (offset + copyLength) unsigned<= arrayLength
941 // B. offset <= (arrayLength - copyLength)
942 // We require that all of the values above, except for the sum and
943 // difference, are already known to be non-negative.
944 // Plan A is robust in the face of overflow, if offset and copyLength
945 // are both hugely positive.
946 //
947 // Plan B is less direct and intuitive, but it does not overflow at
948 // all, since the difference of two non-negatives is always
949 // representable. Whenever Java methods must perform the equivalent
950 // check they generally use Plan B instead of Plan A.
951 // For the moment we use Plan A.
952 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
953 Node* subseq_length,
954 Node* array_length,
955 RegionNode* region) {
956 if (stopped())
957 return NULL; // already stopped
958 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
959 if (zero_offset && subseq_length->eqv_uncast(array_length))
960 return NULL; // common case of whole-array copy
961 Node* last = subseq_length;
962 if (!zero_offset) // last += offset
963 last = _gvn.transform( new (C) AddINode(last, offset));
964 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
965 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
966 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
967 return is_over;
968 }
971 //--------------------------generate_current_thread--------------------
972 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
973 ciKlass* thread_klass = env()->Thread_klass();
974 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
975 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
976 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
977 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
978 tls_output = thread;
979 return threadObj;
980 }
983 //------------------------------make_string_method_node------------------------
984 // Helper method for String intrinsic functions. This version is called
985 // with str1 and str2 pointing to String object nodes.
986 //
987 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
988 Node* no_ctrl = NULL;
990 // Get start addr of string
991 Node* str1_value = load_String_value(no_ctrl, str1);
992 Node* str1_offset = load_String_offset(no_ctrl, str1);
993 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
995 // Get length of string 1
996 Node* str1_len = load_String_length(no_ctrl, str1);
998 Node* str2_value = load_String_value(no_ctrl, str2);
999 Node* str2_offset = load_String_offset(no_ctrl, str2);
1000 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1002 Node* str2_len = NULL;
1003 Node* result = NULL;
1005 switch (opcode) {
1006 case Op_StrIndexOf:
1007 // Get length of string 2
1008 str2_len = load_String_length(no_ctrl, str2);
1010 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1011 str1_start, str1_len, str2_start, str2_len);
1012 break;
1013 case Op_StrComp:
1014 // Get length of string 2
1015 str2_len = load_String_length(no_ctrl, str2);
1017 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1018 str1_start, str1_len, str2_start, str2_len);
1019 break;
1020 case Op_StrEquals:
1021 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1022 str1_start, str2_start, str1_len);
1023 break;
1024 default:
1025 ShouldNotReachHere();
1026 return NULL;
1027 }
1029 // All these intrinsics have checks.
1030 C->set_has_split_ifs(true); // Has chance for split-if optimization
1032 return _gvn.transform(result);
1033 }
1035 // Helper method for String intrinsic functions. This version is called
1036 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1037 // to Int nodes containing the lenghts of str1 and str2.
1038 //
1039 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1040 Node* result = NULL;
1041 switch (opcode) {
1042 case Op_StrIndexOf:
1043 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1044 str1_start, cnt1, str2_start, cnt2);
1045 break;
1046 case Op_StrComp:
1047 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1048 str1_start, cnt1, str2_start, cnt2);
1049 break;
1050 case Op_StrEquals:
1051 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1052 str1_start, str2_start, cnt1);
1053 break;
1054 default:
1055 ShouldNotReachHere();
1056 return NULL;
1057 }
1059 // All these intrinsics have checks.
1060 C->set_has_split_ifs(true); // Has chance for split-if optimization
1062 return _gvn.transform(result);
1063 }
1065 //------------------------------inline_string_compareTo------------------------
1066 // public int java.lang.String.compareTo(String anotherString);
1067 bool LibraryCallKit::inline_string_compareTo() {
1068 Node* receiver = null_check(argument(0));
1069 Node* arg = null_check(argument(1));
1070 if (stopped()) {
1071 return true;
1072 }
1073 set_result(make_string_method_node(Op_StrComp, receiver, arg));
1074 return true;
1075 }
1077 //------------------------------inline_string_equals------------------------
1078 bool LibraryCallKit::inline_string_equals() {
1079 Node* receiver = null_check_receiver();
1080 // NOTE: Do not null check argument for String.equals() because spec
1081 // allows to specify NULL as argument.
1082 Node* argument = this->argument(1);
1083 if (stopped()) {
1084 return true;
1085 }
1087 // paths (plus control) merge
1088 RegionNode* region = new (C) RegionNode(5);
1089 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1091 // does source == target string?
1092 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1093 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1095 Node* if_eq = generate_slow_guard(bol, NULL);
1096 if (if_eq != NULL) {
1097 // receiver == argument
1098 phi->init_req(2, intcon(1));
1099 region->init_req(2, if_eq);
1100 }
1102 // get String klass for instanceOf
1103 ciInstanceKlass* klass = env()->String_klass();
1105 if (!stopped()) {
1106 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1107 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1108 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1110 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1111 //instanceOf == true, fallthrough
1113 if (inst_false != NULL) {
1114 phi->init_req(3, intcon(0));
1115 region->init_req(3, inst_false);
1116 }
1117 }
1119 if (!stopped()) {
1120 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1122 // Properly cast the argument to String
1123 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1124 // This path is taken only when argument's type is String:NotNull.
1125 argument = cast_not_null(argument, false);
1127 Node* no_ctrl = NULL;
1129 // Get start addr of receiver
1130 Node* receiver_val = load_String_value(no_ctrl, receiver);
1131 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1132 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1134 // Get length of receiver
1135 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1137 // Get start addr of argument
1138 Node* argument_val = load_String_value(no_ctrl, argument);
1139 Node* argument_offset = load_String_offset(no_ctrl, argument);
1140 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1142 // Get length of argument
1143 Node* argument_cnt = load_String_length(no_ctrl, argument);
1145 // Check for receiver count != argument count
1146 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1147 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1148 Node* if_ne = generate_slow_guard(bol, NULL);
1149 if (if_ne != NULL) {
1150 phi->init_req(4, intcon(0));
1151 region->init_req(4, if_ne);
1152 }
1154 // Check for count == 0 is done by assembler code for StrEquals.
1156 if (!stopped()) {
1157 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1158 phi->init_req(1, equals);
1159 region->init_req(1, control());
1160 }
1161 }
1163 // post merge
1164 set_control(_gvn.transform(region));
1165 record_for_igvn(region);
1167 set_result(_gvn.transform(phi));
1168 return true;
1169 }
1171 //------------------------------inline_array_equals----------------------------
1172 bool LibraryCallKit::inline_array_equals() {
1173 Node* arg1 = argument(0);
1174 Node* arg2 = argument(1);
1175 set_result(_gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS), arg1, arg2)));
1176 return true;
1177 }
1179 // Java version of String.indexOf(constant string)
1180 // class StringDecl {
1181 // StringDecl(char[] ca) {
1182 // offset = 0;
1183 // count = ca.length;
1184 // value = ca;
1185 // }
1186 // int offset;
1187 // int count;
1188 // char[] value;
1189 // }
1190 //
1191 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1192 // int targetOffset, int cache_i, int md2) {
1193 // int cache = cache_i;
1194 // int sourceOffset = string_object.offset;
1195 // int sourceCount = string_object.count;
1196 // int targetCount = target_object.length;
1197 //
1198 // int targetCountLess1 = targetCount - 1;
1199 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1200 //
1201 // char[] source = string_object.value;
1202 // char[] target = target_object;
1203 // int lastChar = target[targetCountLess1];
1204 //
1205 // outer_loop:
1206 // for (int i = sourceOffset; i < sourceEnd; ) {
1207 // int src = source[i + targetCountLess1];
1208 // if (src == lastChar) {
1209 // // With random strings and a 4-character alphabet,
1210 // // reverse matching at this point sets up 0.8% fewer
1211 // // frames, but (paradoxically) makes 0.3% more probes.
1212 // // Since those probes are nearer the lastChar probe,
1213 // // there is may be a net D$ win with reverse matching.
1214 // // But, reversing loop inhibits unroll of inner loop
1215 // // for unknown reason. So, does running outer loop from
1216 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1217 // for (int j = 0; j < targetCountLess1; j++) {
1218 // if (target[targetOffset + j] != source[i+j]) {
1219 // if ((cache & (1 << source[i+j])) == 0) {
1220 // if (md2 < j+1) {
1221 // i += j+1;
1222 // continue outer_loop;
1223 // }
1224 // }
1225 // i += md2;
1226 // continue outer_loop;
1227 // }
1228 // }
1229 // return i - sourceOffset;
1230 // }
1231 // if ((cache & (1 << src)) == 0) {
1232 // i += targetCountLess1;
1233 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1234 // i++;
1235 // }
1236 // return -1;
1237 // }
1239 //------------------------------string_indexOf------------------------
1240 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1241 jint cache_i, jint md2_i) {
1243 Node* no_ctrl = NULL;
1244 float likely = PROB_LIKELY(0.9);
1245 float unlikely = PROB_UNLIKELY(0.9);
1247 const int nargs = 0; // no arguments to push back for uncommon trap in predicate
1249 Node* source = load_String_value(no_ctrl, string_object);
1250 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1251 Node* sourceCount = load_String_length(no_ctrl, string_object);
1253 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1254 jint target_length = target_array->length();
1255 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1256 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1258 IdealKit kit(this, false, true);
1259 #define __ kit.
1260 Node* zero = __ ConI(0);
1261 Node* one = __ ConI(1);
1262 Node* cache = __ ConI(cache_i);
1263 Node* md2 = __ ConI(md2_i);
1264 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1265 Node* targetCount = __ ConI(target_length);
1266 Node* targetCountLess1 = __ ConI(target_length - 1);
1267 Node* targetOffset = __ ConI(targetOffset_i);
1268 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1270 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1271 Node* outer_loop = __ make_label(2 /* goto */);
1272 Node* return_ = __ make_label(1);
1274 __ set(rtn,__ ConI(-1));
1275 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1276 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1277 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1278 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1279 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1280 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1281 Node* tpj = __ AddI(targetOffset, __ value(j));
1282 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1283 Node* ipj = __ AddI(__ value(i), __ value(j));
1284 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1285 __ if_then(targ, BoolTest::ne, src2); {
1286 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1287 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1288 __ increment(i, __ AddI(__ value(j), one));
1289 __ goto_(outer_loop);
1290 } __ end_if(); __ dead(j);
1291 }__ end_if(); __ dead(j);
1292 __ increment(i, md2);
1293 __ goto_(outer_loop);
1294 }__ end_if();
1295 __ increment(j, one);
1296 }__ end_loop(); __ dead(j);
1297 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1298 __ goto_(return_);
1299 }__ end_if();
1300 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1301 __ increment(i, targetCountLess1);
1302 }__ end_if();
1303 __ increment(i, one);
1304 __ bind(outer_loop);
1305 }__ end_loop(); __ dead(i);
1306 __ bind(return_);
1308 // Final sync IdealKit and GraphKit.
1309 final_sync(kit);
1310 Node* result = __ value(rtn);
1311 #undef __
1312 C->set_has_loops(true);
1313 return result;
1314 }
1316 //------------------------------inline_string_indexOf------------------------
1317 bool LibraryCallKit::inline_string_indexOf() {
1318 Node* receiver = argument(0);
1319 Node* arg = argument(1);
1321 Node* result;
1322 // Disable the use of pcmpestri until it can be guaranteed that
1323 // the load doesn't cross into the uncommited space.
1324 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1325 UseSSE42Intrinsics) {
1326 // Generate SSE4.2 version of indexOf
1327 // We currently only have match rules that use SSE4.2
1329 receiver = null_check(receiver);
1330 arg = null_check(arg);
1331 if (stopped()) {
1332 return true;
1333 }
1335 ciInstanceKlass* str_klass = env()->String_klass();
1336 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1338 // Make the merge point
1339 RegionNode* result_rgn = new (C) RegionNode(4);
1340 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1341 Node* no_ctrl = NULL;
1343 // Get start addr of source string
1344 Node* source = load_String_value(no_ctrl, receiver);
1345 Node* source_offset = load_String_offset(no_ctrl, receiver);
1346 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1348 // Get length of source string
1349 Node* source_cnt = load_String_length(no_ctrl, receiver);
1351 // Get start addr of substring
1352 Node* substr = load_String_value(no_ctrl, arg);
1353 Node* substr_offset = load_String_offset(no_ctrl, arg);
1354 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1356 // Get length of source string
1357 Node* substr_cnt = load_String_length(no_ctrl, arg);
1359 // Check for substr count > string count
1360 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1361 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1362 Node* if_gt = generate_slow_guard(bol, NULL);
1363 if (if_gt != NULL) {
1364 result_phi->init_req(2, intcon(-1));
1365 result_rgn->init_req(2, if_gt);
1366 }
1368 if (!stopped()) {
1369 // Check for substr count == 0
1370 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1371 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1372 Node* if_zero = generate_slow_guard(bol, NULL);
1373 if (if_zero != NULL) {
1374 result_phi->init_req(3, intcon(0));
1375 result_rgn->init_req(3, if_zero);
1376 }
1377 }
1379 if (!stopped()) {
1380 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1381 result_phi->init_req(1, result);
1382 result_rgn->init_req(1, control());
1383 }
1384 set_control(_gvn.transform(result_rgn));
1385 record_for_igvn(result_rgn);
1386 result = _gvn.transform(result_phi);
1388 } else { // Use LibraryCallKit::string_indexOf
1389 // don't intrinsify if argument isn't a constant string.
1390 if (!arg->is_Con()) {
1391 return false;
1392 }
1393 const TypeOopPtr* str_type = _gvn.type(arg)->isa_oopptr();
1394 if (str_type == NULL) {
1395 return false;
1396 }
1397 ciInstanceKlass* klass = env()->String_klass();
1398 ciObject* str_const = str_type->const_oop();
1399 if (str_const == NULL || str_const->klass() != klass) {
1400 return false;
1401 }
1402 ciInstance* str = str_const->as_instance();
1403 assert(str != NULL, "must be instance");
1405 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1406 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1408 int o;
1409 int c;
1410 if (java_lang_String::has_offset_field()) {
1411 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1412 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1413 } else {
1414 o = 0;
1415 c = pat->length();
1416 }
1418 // constant strings have no offset and count == length which
1419 // simplifies the resulting code somewhat so lets optimize for that.
1420 if (o != 0 || c != pat->length()) {
1421 return false;
1422 }
1424 receiver = null_check(receiver, T_OBJECT);
1425 // NOTE: No null check on the argument is needed since it's a constant String oop.
1426 if (stopped()) {
1427 return true;
1428 }
1430 // The null string as a pattern always returns 0 (match at beginning of string)
1431 if (c == 0) {
1432 set_result(intcon(0));
1433 return true;
1434 }
1436 // Generate default indexOf
1437 jchar lastChar = pat->char_at(o + (c - 1));
1438 int cache = 0;
1439 int i;
1440 for (i = 0; i < c - 1; i++) {
1441 assert(i < pat->length(), "out of range");
1442 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1443 }
1445 int md2 = c;
1446 for (i = 0; i < c - 1; i++) {
1447 assert(i < pat->length(), "out of range");
1448 if (pat->char_at(o + i) == lastChar) {
1449 md2 = (c - 1) - i;
1450 }
1451 }
1453 result = string_indexOf(receiver, pat, o, cache, md2);
1454 }
1455 set_result(result);
1456 return true;
1457 }
1459 //--------------------------round_double_node--------------------------------
1460 // Round a double node if necessary.
1461 Node* LibraryCallKit::round_double_node(Node* n) {
1462 if (Matcher::strict_fp_requires_explicit_rounding && UseSSE <= 1)
1463 n = _gvn.transform(new (C) RoundDoubleNode(0, n));
1464 return n;
1465 }
1467 //------------------------------inline_math-----------------------------------
1468 // public static double Math.abs(double)
1469 // public static double Math.sqrt(double)
1470 // public static double Math.log(double)
1471 // public static double Math.log10(double)
1472 bool LibraryCallKit::inline_math(vmIntrinsics::ID id) {
1473 Node* arg = round_double_node(argument(0));
1474 Node* n;
1475 switch (id) {
1476 case vmIntrinsics::_dabs: n = new (C) AbsDNode( arg); break;
1477 case vmIntrinsics::_dsqrt: n = new (C) SqrtDNode(0, arg); break;
1478 case vmIntrinsics::_dlog: n = new (C) LogDNode( arg); break;
1479 case vmIntrinsics::_dlog10: n = new (C) Log10DNode( arg); break;
1480 default: fatal_unexpected_iid(id); break;
1481 }
1482 set_result(_gvn.transform(n));
1483 return true;
1484 }
1486 //------------------------------inline_trig----------------------------------
1487 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1488 // argument reduction which will turn into a fast/slow diamond.
1489 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1490 Node* arg = round_double_node(argument(0));
1491 Node* n = NULL;
1493 switch (id) {
1494 case vmIntrinsics::_dsin: n = new (C) SinDNode(arg); break;
1495 case vmIntrinsics::_dcos: n = new (C) CosDNode(arg); break;
1496 case vmIntrinsics::_dtan: n = new (C) TanDNode(arg); break;
1497 default: fatal_unexpected_iid(id); break;
1498 }
1499 n = _gvn.transform(n);
1501 // Rounding required? Check for argument reduction!
1502 if (Matcher::strict_fp_requires_explicit_rounding) {
1503 static const double pi_4 = 0.7853981633974483;
1504 static const double neg_pi_4 = -0.7853981633974483;
1505 // pi/2 in 80-bit extended precision
1506 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1507 // -pi/2 in 80-bit extended precision
1508 // 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};
1509 // Cutoff value for using this argument reduction technique
1510 //static const double pi_2_minus_epsilon = 1.564660403643354;
1511 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1513 // Pseudocode for sin:
1514 // if (x <= Math.PI / 4.0) {
1515 // if (x >= -Math.PI / 4.0) return fsin(x);
1516 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1517 // } else {
1518 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1519 // }
1520 // return StrictMath.sin(x);
1522 // Pseudocode for cos:
1523 // if (x <= Math.PI / 4.0) {
1524 // if (x >= -Math.PI / 4.0) return fcos(x);
1525 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1526 // } else {
1527 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1528 // }
1529 // return StrictMath.cos(x);
1531 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1532 // requires a special machine instruction to load it. Instead we'll try
1533 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1534 // probably do the math inside the SIN encoding.
1536 // Make the merge point
1537 RegionNode* r = new (C) RegionNode(3);
1538 Node* phi = new (C) PhiNode(r, Type::DOUBLE);
1540 // Flatten arg so we need only 1 test
1541 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1542 // Node for PI/4 constant
1543 Node *pi4 = makecon(TypeD::make(pi_4));
1544 // Check PI/4 : abs(arg)
1545 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1546 // Check: If PI/4 < abs(arg) then go slow
1547 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1548 // Branch either way
1549 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1550 set_control(opt_iff(r,iff));
1552 // Set fast path result
1553 phi->init_req(2, n);
1555 // Slow path - non-blocking leaf call
1556 Node* call = NULL;
1557 switch (id) {
1558 case vmIntrinsics::_dsin:
1559 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1560 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1561 "Sin", NULL, arg, top());
1562 break;
1563 case vmIntrinsics::_dcos:
1564 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1565 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1566 "Cos", NULL, arg, top());
1567 break;
1568 case vmIntrinsics::_dtan:
1569 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1570 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1571 "Tan", NULL, arg, top());
1572 break;
1573 }
1574 assert(control()->in(0) == call, "");
1575 Node* slow_result = _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
1576 r->init_req(1, control());
1577 phi->init_req(1, slow_result);
1579 // Post-merge
1580 set_control(_gvn.transform(r));
1581 record_for_igvn(r);
1582 n = _gvn.transform(phi);
1584 C->set_has_split_ifs(true); // Has chance for split-if optimization
1585 }
1586 set_result(n);
1587 return true;
1588 }
1590 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1591 //-------------------
1592 //result=(result.isNaN())? funcAddr():result;
1593 // Check: If isNaN() by checking result!=result? then either trap
1594 // or go to runtime
1595 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result, result));
1596 // Build the boolean node
1597 Node* bolisnum = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::eq));
1599 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1600 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1601 // The pow or exp intrinsic returned a NaN, which requires a call
1602 // to the runtime. Recompile with the runtime call.
1603 uncommon_trap(Deoptimization::Reason_intrinsic,
1604 Deoptimization::Action_make_not_entrant);
1605 }
1606 set_result(result);
1607 } else {
1608 // If this inlining ever returned NaN in the past, we compile a call
1609 // to the runtime to properly handle corner cases
1611 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1612 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1613 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1615 if (!if_slow->is_top()) {
1616 RegionNode* result_region = new (C) RegionNode(3);
1617 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1619 result_region->init_req(1, if_fast);
1620 result_val->init_req(1, result);
1622 set_control(if_slow);
1624 const TypePtr* no_memory_effects = NULL;
1625 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1626 no_memory_effects,
1627 x, top(), y, y ? top() : NULL);
1628 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1629 #ifdef ASSERT
1630 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1631 assert(value_top == top(), "second value must be top");
1632 #endif
1634 result_region->init_req(2, control());
1635 result_val->init_req(2, value);
1636 set_result(result_region, result_val);
1637 } else {
1638 set_result(result);
1639 }
1640 }
1641 }
1643 //------------------------------inline_exp-------------------------------------
1644 // Inline exp instructions, if possible. The Intel hardware only misses
1645 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1646 bool LibraryCallKit::inline_exp() {
1647 Node* arg = round_double_node(argument(0));
1648 Node* n = _gvn.transform(new (C) ExpDNode(0, arg));
1650 finish_pow_exp(n, arg, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1652 C->set_has_split_ifs(true); // Has chance for split-if optimization
1653 return true;
1654 }
1656 //------------------------------inline_pow-------------------------------------
1657 // Inline power instructions, if possible.
1658 bool LibraryCallKit::inline_pow() {
1659 // Pseudocode for pow
1660 // if (x <= 0.0) {
1661 // long longy = (long)y;
1662 // if ((double)longy == y) { // if y is long
1663 // if (y + 1 == y) longy = 0; // huge number: even
1664 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1665 // } else {
1666 // result = NaN;
1667 // }
1668 // } else {
1669 // result = DPow(x,y);
1670 // }
1671 // if (result != result)? {
1672 // result = uncommon_trap() or runtime_call();
1673 // }
1674 // return result;
1676 Node* x = round_double_node(argument(0));
1677 Node* y = round_double_node(argument(2));
1679 Node* result = NULL;
1681 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1682 // Short form: skip the fancy tests and just check for NaN result.
1683 result = _gvn.transform(new (C) PowDNode(0, x, y));
1684 } else {
1685 // If this inlining ever returned NaN in the past, include all
1686 // checks + call to the runtime.
1688 // Set the merge point for If node with condition of (x <= 0.0)
1689 // There are four possible paths to region node and phi node
1690 RegionNode *r = new (C) RegionNode(4);
1691 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1693 // Build the first if node: if (x <= 0.0)
1694 // Node for 0 constant
1695 Node *zeronode = makecon(TypeD::ZERO);
1696 // Check x:0
1697 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1698 // Check: If (x<=0) then go complex path
1699 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1700 // Branch either way
1701 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1702 // Fast path taken; set region slot 3
1703 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1704 r->init_req(3,fast_taken); // Capture fast-control
1706 // Fast path not-taken, i.e. slow path
1707 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1709 // Set fast path result
1710 Node *fast_result = _gvn.transform( new (C) PowDNode(0, x, y) );
1711 phi->init_req(3, fast_result);
1713 // Complex path
1714 // Build the second if node (if y is long)
1715 // Node for (long)y
1716 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1717 // Node for (double)((long) y)
1718 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1719 // Check (double)((long) y) : y
1720 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1721 // Check if (y isn't long) then go to slow path
1723 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1724 // Branch either way
1725 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1726 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1728 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1730 // Calculate DPow(abs(x), y)*(1 & (long)y)
1731 // Node for constant 1
1732 Node *conone = longcon(1);
1733 // 1& (long)y
1734 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1736 // A huge number is always even. Detect a huge number by checking
1737 // if y + 1 == y and set integer to be tested for parity to 0.
1738 // Required for corner case:
1739 // (long)9.223372036854776E18 = max_jlong
1740 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1741 // max_jlong is odd but 9.223372036854776E18 is even
1742 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1743 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1744 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1745 Node* correctedsign = NULL;
1746 if (ConditionalMoveLimit != 0) {
1747 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1748 } else {
1749 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1750 RegionNode *r = new (C) RegionNode(3);
1751 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1752 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1753 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1754 phi->init_req(1, signnode);
1755 phi->init_req(2, longcon(0));
1756 correctedsign = _gvn.transform(phi);
1757 ylong_path = _gvn.transform(r);
1758 record_for_igvn(r);
1759 }
1761 // zero node
1762 Node *conzero = longcon(0);
1763 // Check (1&(long)y)==0?
1764 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1765 // Check if (1&(long)y)!=0?, if so the result is negative
1766 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1767 // abs(x)
1768 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1769 // abs(x)^y
1770 Node *absxpowy = _gvn.transform( new (C) PowDNode(0, absx, y) );
1771 // -abs(x)^y
1772 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1773 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1774 Node *signresult = NULL;
1775 if (ConditionalMoveLimit != 0) {
1776 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1777 } else {
1778 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1779 RegionNode *r = new (C) RegionNode(3);
1780 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1781 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1782 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1783 phi->init_req(1, absxpowy);
1784 phi->init_req(2, negabsxpowy);
1785 signresult = _gvn.transform(phi);
1786 ylong_path = _gvn.transform(r);
1787 record_for_igvn(r);
1788 }
1789 // Set complex path fast result
1790 r->init_req(2, ylong_path);
1791 phi->init_req(2, signresult);
1793 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1794 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1795 r->init_req(1,slow_path);
1796 phi->init_req(1,slow_result);
1798 // Post merge
1799 set_control(_gvn.transform(r));
1800 record_for_igvn(r);
1801 result = _gvn.transform(phi);
1802 }
1804 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1806 C->set_has_split_ifs(true); // Has chance for split-if optimization
1807 return true;
1808 }
1810 //------------------------------runtime_math-----------------------------
1811 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1812 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1813 "must be (DD)D or (D)D type");
1815 // Inputs
1816 Node* a = round_double_node(argument(0));
1817 Node* b = (call_type == OptoRuntime::Math_DD_D_Type()) ? round_double_node(argument(2)) : NULL;
1819 const TypePtr* no_memory_effects = NULL;
1820 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1821 no_memory_effects,
1822 a, top(), b, b ? top() : NULL);
1823 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1824 #ifdef ASSERT
1825 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1826 assert(value_top == top(), "second value must be top");
1827 #endif
1829 set_result(value);
1830 return true;
1831 }
1833 //------------------------------inline_math_native-----------------------------
1834 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1835 #define FN_PTR(f) CAST_FROM_FN_PTR(address, f)
1836 switch (id) {
1837 // These intrinsics are not properly supported on all hardware
1838 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1839 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dcos), "COS");
1840 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1841 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dsin), "SIN");
1842 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1843 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dtan), "TAN");
1845 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_math(id) :
1846 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog), "LOG");
1847 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_math(id) :
1848 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dlog10), "LOG10");
1850 // These intrinsics are supported on all hardware
1851 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_math(id) : false;
1852 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_math(id) : false;
1854 case vmIntrinsics::_dexp: return Matcher::has_match_rule(Op_ExpD) ? inline_exp() :
1855 runtime_math(OptoRuntime::Math_D_D_Type(), FN_PTR(SharedRuntime::dexp), "EXP");
1856 case vmIntrinsics::_dpow: return Matcher::has_match_rule(Op_PowD) ? inline_pow() :
1857 runtime_math(OptoRuntime::Math_DD_D_Type(), FN_PTR(SharedRuntime::dpow), "POW");
1858 #undef FN_PTR
1860 // These intrinsics are not yet correctly implemented
1861 case vmIntrinsics::_datan2:
1862 return false;
1864 default:
1865 fatal_unexpected_iid(id);
1866 return false;
1867 }
1868 }
1870 static bool is_simple_name(Node* n) {
1871 return (n->req() == 1 // constant
1872 || (n->is_Type() && n->as_Type()->type()->singleton())
1873 || n->is_Proj() // parameter or return value
1874 || n->is_Phi() // local of some sort
1875 );
1876 }
1878 //----------------------------inline_min_max-----------------------------------
1879 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1880 set_result(generate_min_max(id, argument(0), argument(1)));
1881 return true;
1882 }
1884 Node*
1885 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1886 // These are the candidate return value:
1887 Node* xvalue = x0;
1888 Node* yvalue = y0;
1890 if (xvalue == yvalue) {
1891 return xvalue;
1892 }
1894 bool want_max = (id == vmIntrinsics::_max);
1896 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1897 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1898 if (txvalue == NULL || tyvalue == NULL) return top();
1899 // This is not really necessary, but it is consistent with a
1900 // hypothetical MaxINode::Value method:
1901 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1903 // %%% This folding logic should (ideally) be in a different place.
1904 // Some should be inside IfNode, and there to be a more reliable
1905 // transformation of ?: style patterns into cmoves. We also want
1906 // more powerful optimizations around cmove and min/max.
1908 // Try to find a dominating comparison of these guys.
1909 // It can simplify the index computation for Arrays.copyOf
1910 // and similar uses of System.arraycopy.
1911 // First, compute the normalized version of CmpI(x, y).
1912 int cmp_op = Op_CmpI;
1913 Node* xkey = xvalue;
1914 Node* ykey = yvalue;
1915 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
1916 if (ideal_cmpxy->is_Cmp()) {
1917 // E.g., if we have CmpI(length - offset, count),
1918 // it might idealize to CmpI(length, count + offset)
1919 cmp_op = ideal_cmpxy->Opcode();
1920 xkey = ideal_cmpxy->in(1);
1921 ykey = ideal_cmpxy->in(2);
1922 }
1924 // Start by locating any relevant comparisons.
1925 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1926 Node* cmpxy = NULL;
1927 Node* cmpyx = NULL;
1928 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1929 Node* cmp = start_from->fast_out(k);
1930 if (cmp->outcnt() > 0 && // must have prior uses
1931 cmp->in(0) == NULL && // must be context-independent
1932 cmp->Opcode() == cmp_op) { // right kind of compare
1933 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1934 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1935 }
1936 }
1938 const int NCMPS = 2;
1939 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1940 int cmpn;
1941 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1942 if (cmps[cmpn] != NULL) break; // find a result
1943 }
1944 if (cmpn < NCMPS) {
1945 // Look for a dominating test that tells us the min and max.
1946 int depth = 0; // Limit search depth for speed
1947 Node* dom = control();
1948 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1949 if (++depth >= 100) break;
1950 Node* ifproj = dom;
1951 if (!ifproj->is_Proj()) continue;
1952 Node* iff = ifproj->in(0);
1953 if (!iff->is_If()) continue;
1954 Node* bol = iff->in(1);
1955 if (!bol->is_Bool()) continue;
1956 Node* cmp = bol->in(1);
1957 if (cmp == NULL) continue;
1958 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1959 if (cmps[cmpn] == cmp) break;
1960 if (cmpn == NCMPS) continue;
1961 BoolTest::mask btest = bol->as_Bool()->_test._test;
1962 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1963 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1964 // At this point, we know that 'x btest y' is true.
1965 switch (btest) {
1966 case BoolTest::eq:
1967 // They are proven equal, so we can collapse the min/max.
1968 // Either value is the answer. Choose the simpler.
1969 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1970 return yvalue;
1971 return xvalue;
1972 case BoolTest::lt: // x < y
1973 case BoolTest::le: // x <= y
1974 return (want_max ? yvalue : xvalue);
1975 case BoolTest::gt: // x > y
1976 case BoolTest::ge: // x >= y
1977 return (want_max ? xvalue : yvalue);
1978 }
1979 }
1980 }
1982 // We failed to find a dominating test.
1983 // Let's pick a test that might GVN with prior tests.
1984 Node* best_bol = NULL;
1985 BoolTest::mask best_btest = BoolTest::illegal;
1986 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1987 Node* cmp = cmps[cmpn];
1988 if (cmp == NULL) continue;
1989 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1990 Node* bol = cmp->fast_out(j);
1991 if (!bol->is_Bool()) continue;
1992 BoolTest::mask btest = bol->as_Bool()->_test._test;
1993 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1994 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1995 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1996 best_bol = bol->as_Bool();
1997 best_btest = btest;
1998 }
1999 }
2000 }
2002 Node* answer_if_true = NULL;
2003 Node* answer_if_false = NULL;
2004 switch (best_btest) {
2005 default:
2006 if (cmpxy == NULL)
2007 cmpxy = ideal_cmpxy;
2008 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2009 // and fall through:
2010 case BoolTest::lt: // x < y
2011 case BoolTest::le: // x <= y
2012 answer_if_true = (want_max ? yvalue : xvalue);
2013 answer_if_false = (want_max ? xvalue : yvalue);
2014 break;
2015 case BoolTest::gt: // x > y
2016 case BoolTest::ge: // x >= y
2017 answer_if_true = (want_max ? xvalue : yvalue);
2018 answer_if_false = (want_max ? yvalue : xvalue);
2019 break;
2020 }
2022 jint hi, lo;
2023 if (want_max) {
2024 // We can sharpen the minimum.
2025 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2026 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2027 } else {
2028 // We can sharpen the maximum.
2029 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2030 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2031 }
2033 // Use a flow-free graph structure, to avoid creating excess control edges
2034 // which could hinder other optimizations.
2035 // Since Math.min/max is often used with arraycopy, we want
2036 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2037 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2038 answer_if_false, answer_if_true,
2039 TypeInt::make(lo, hi, widen));
2041 return _gvn.transform(cmov);
2043 /*
2044 // This is not as desirable as it may seem, since Min and Max
2045 // nodes do not have a full set of optimizations.
2046 // And they would interfere, anyway, with 'if' optimizations
2047 // and with CMoveI canonical forms.
2048 switch (id) {
2049 case vmIntrinsics::_min:
2050 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2051 case vmIntrinsics::_max:
2052 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2053 default:
2054 ShouldNotReachHere();
2055 }
2056 */
2057 }
2059 inline int
2060 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2061 const TypePtr* base_type = TypePtr::NULL_PTR;
2062 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2063 if (base_type == NULL) {
2064 // Unknown type.
2065 return Type::AnyPtr;
2066 } else if (base_type == TypePtr::NULL_PTR) {
2067 // Since this is a NULL+long form, we have to switch to a rawptr.
2068 base = _gvn.transform( new (C) CastX2PNode(offset) );
2069 offset = MakeConX(0);
2070 return Type::RawPtr;
2071 } else if (base_type->base() == Type::RawPtr) {
2072 return Type::RawPtr;
2073 } else if (base_type->isa_oopptr()) {
2074 // Base is never null => always a heap address.
2075 if (base_type->ptr() == TypePtr::NotNull) {
2076 return Type::OopPtr;
2077 }
2078 // Offset is small => always a heap address.
2079 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2080 if (offset_type != NULL &&
2081 base_type->offset() == 0 && // (should always be?)
2082 offset_type->_lo >= 0 &&
2083 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2084 return Type::OopPtr;
2085 }
2086 // Otherwise, it might either be oop+off or NULL+addr.
2087 return Type::AnyPtr;
2088 } else {
2089 // No information:
2090 return Type::AnyPtr;
2091 }
2092 }
2094 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2095 int kind = classify_unsafe_addr(base, offset);
2096 if (kind == Type::RawPtr) {
2097 return basic_plus_adr(top(), base, offset);
2098 } else {
2099 return basic_plus_adr(base, offset);
2100 }
2101 }
2103 //--------------------------inline_number_methods-----------------------------
2104 // inline int Integer.numberOfLeadingZeros(int)
2105 // inline int Long.numberOfLeadingZeros(long)
2106 //
2107 // inline int Integer.numberOfTrailingZeros(int)
2108 // inline int Long.numberOfTrailingZeros(long)
2109 //
2110 // inline int Integer.bitCount(int)
2111 // inline int Long.bitCount(long)
2112 //
2113 // inline char Character.reverseBytes(char)
2114 // inline short Short.reverseBytes(short)
2115 // inline int Integer.reverseBytes(int)
2116 // inline long Long.reverseBytes(long)
2117 bool LibraryCallKit::inline_number_methods(vmIntrinsics::ID id) {
2118 Node* arg = argument(0);
2119 Node* n;
2120 switch (id) {
2121 case vmIntrinsics::_numberOfLeadingZeros_i: n = new (C) CountLeadingZerosINode( arg); break;
2122 case vmIntrinsics::_numberOfLeadingZeros_l: n = new (C) CountLeadingZerosLNode( arg); break;
2123 case vmIntrinsics::_numberOfTrailingZeros_i: n = new (C) CountTrailingZerosINode(arg); break;
2124 case vmIntrinsics::_numberOfTrailingZeros_l: n = new (C) CountTrailingZerosLNode(arg); break;
2125 case vmIntrinsics::_bitCount_i: n = new (C) PopCountINode( arg); break;
2126 case vmIntrinsics::_bitCount_l: n = new (C) PopCountLNode( arg); break;
2127 case vmIntrinsics::_reverseBytes_c: n = new (C) ReverseBytesUSNode(0, arg); break;
2128 case vmIntrinsics::_reverseBytes_s: n = new (C) ReverseBytesSNode( 0, arg); break;
2129 case vmIntrinsics::_reverseBytes_i: n = new (C) ReverseBytesINode( 0, arg); break;
2130 case vmIntrinsics::_reverseBytes_l: n = new (C) ReverseBytesLNode( 0, arg); break;
2131 default: fatal_unexpected_iid(id); break;
2132 }
2133 set_result(_gvn.transform(n));
2134 return true;
2135 }
2137 //----------------------------inline_unsafe_access----------------------------
2139 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2141 // Helper that guards and inserts a pre-barrier.
2142 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2143 Node* pre_val, bool need_mem_bar) {
2144 // We could be accessing the referent field of a reference object. If so, when G1
2145 // is enabled, we need to log the value in the referent field in an SATB buffer.
2146 // This routine performs some compile time filters and generates suitable
2147 // runtime filters that guard the pre-barrier code.
2148 // Also add memory barrier for non volatile load from the referent field
2149 // to prevent commoning of loads across safepoint.
2150 if (!UseG1GC && !need_mem_bar)
2151 return;
2153 // Some compile time checks.
2155 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2156 const TypeX* otype = offset->find_intptr_t_type();
2157 if (otype != NULL && otype->is_con() &&
2158 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2159 // Constant offset but not the reference_offset so just return
2160 return;
2161 }
2163 // We only need to generate the runtime guards for instances.
2164 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2165 if (btype != NULL) {
2166 if (btype->isa_aryptr()) {
2167 // Array type so nothing to do
2168 return;
2169 }
2171 const TypeInstPtr* itype = btype->isa_instptr();
2172 if (itype != NULL) {
2173 // Can the klass of base_oop be statically determined to be
2174 // _not_ a sub-class of Reference and _not_ Object?
2175 ciKlass* klass = itype->klass();
2176 if ( klass->is_loaded() &&
2177 !klass->is_subtype_of(env()->Reference_klass()) &&
2178 !env()->Object_klass()->is_subtype_of(klass)) {
2179 return;
2180 }
2181 }
2182 }
2184 // The compile time filters did not reject base_oop/offset so
2185 // we need to generate the following runtime filters
2186 //
2187 // if (offset == java_lang_ref_Reference::_reference_offset) {
2188 // if (instance_of(base, java.lang.ref.Reference)) {
2189 // pre_barrier(_, pre_val, ...);
2190 // }
2191 // }
2193 float likely = PROB_LIKELY( 0.999);
2194 float unlikely = PROB_UNLIKELY(0.999);
2196 IdealKit ideal(this);
2197 #define __ ideal.
2199 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2201 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2202 // Update graphKit memory and control from IdealKit.
2203 sync_kit(ideal);
2205 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2206 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2208 // Update IdealKit memory and control from graphKit.
2209 __ sync_kit(this);
2211 Node* one = __ ConI(1);
2212 // is_instof == 0 if base_oop == NULL
2213 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2215 // Update graphKit from IdeakKit.
2216 sync_kit(ideal);
2218 // Use the pre-barrier to record the value in the referent field
2219 pre_barrier(false /* do_load */,
2220 __ ctrl(),
2221 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2222 pre_val /* pre_val */,
2223 T_OBJECT);
2224 if (need_mem_bar) {
2225 // Add memory barrier to prevent commoning reads from this field
2226 // across safepoint since GC can change its value.
2227 insert_mem_bar(Op_MemBarCPUOrder);
2228 }
2229 // Update IdealKit from graphKit.
2230 __ sync_kit(this);
2232 } __ end_if(); // _ref_type != ref_none
2233 } __ end_if(); // offset == referent_offset
2235 // Final sync IdealKit and GraphKit.
2236 final_sync(ideal);
2237 #undef __
2238 }
2241 // Interpret Unsafe.fieldOffset cookies correctly:
2242 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2244 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2245 // Attempt to infer a sharper value type from the offset and base type.
2246 ciKlass* sharpened_klass = NULL;
2248 // See if it is an instance field, with an object type.
2249 if (alias_type->field() != NULL) {
2250 assert(!is_native_ptr, "native pointer op cannot use a java address");
2251 if (alias_type->field()->type()->is_klass()) {
2252 sharpened_klass = alias_type->field()->type()->as_klass();
2253 }
2254 }
2256 // See if it is a narrow oop array.
2257 if (adr_type->isa_aryptr()) {
2258 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2259 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2260 if (elem_type != NULL) {
2261 sharpened_klass = elem_type->klass();
2262 }
2263 }
2264 }
2266 // The sharpened class might be unloaded if there is no class loader
2267 // contraint in place.
2268 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2269 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2271 #ifndef PRODUCT
2272 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2273 tty->print(" from base type: "); adr_type->dump();
2274 tty->print(" sharpened value: "); tjp->dump();
2275 }
2276 #endif
2277 // Sharpen the value type.
2278 return tjp;
2279 }
2280 return NULL;
2281 }
2283 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2284 if (callee()->is_static()) return false; // caller must have the capability!
2286 #ifndef PRODUCT
2287 {
2288 ResourceMark rm;
2289 // Check the signatures.
2290 ciSignature* sig = callee()->signature();
2291 #ifdef ASSERT
2292 if (!is_store) {
2293 // Object getObject(Object base, int/long offset), etc.
2294 BasicType rtype = sig->return_type()->basic_type();
2295 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2296 rtype = T_ADDRESS; // it is really a C void*
2297 assert(rtype == type, "getter must return the expected value");
2298 if (!is_native_ptr) {
2299 assert(sig->count() == 2, "oop getter has 2 arguments");
2300 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2301 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2302 } else {
2303 assert(sig->count() == 1, "native getter has 1 argument");
2304 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2305 }
2306 } else {
2307 // void putObject(Object base, int/long offset, Object x), etc.
2308 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2309 if (!is_native_ptr) {
2310 assert(sig->count() == 3, "oop putter has 3 arguments");
2311 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2312 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2313 } else {
2314 assert(sig->count() == 2, "native putter has 2 arguments");
2315 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2316 }
2317 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2318 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2319 vtype = T_ADDRESS; // it is really a C void*
2320 assert(vtype == type, "putter must accept the expected value");
2321 }
2322 #endif // ASSERT
2323 }
2324 #endif //PRODUCT
2326 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2328 Node* receiver = argument(0); // type: oop
2330 // Build address expression. See the code in inline_unsafe_prefetch.
2331 Node* adr;
2332 Node* heap_base_oop = top();
2333 Node* offset = top();
2334 Node* val;
2336 if (!is_native_ptr) {
2337 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2338 Node* base = argument(1); // type: oop
2339 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2340 offset = argument(2); // type: long
2341 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2342 // to be plain byte offsets, which are also the same as those accepted
2343 // by oopDesc::field_base.
2344 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2345 "fieldOffset must be byte-scaled");
2346 // 32-bit machines ignore the high half!
2347 offset = ConvL2X(offset);
2348 adr = make_unsafe_address(base, offset);
2349 heap_base_oop = base;
2350 val = is_store ? argument(4) : NULL;
2351 } else {
2352 Node* ptr = argument(1); // type: long
2353 ptr = ConvL2X(ptr); // adjust Java long to machine word
2354 adr = make_unsafe_address(NULL, ptr);
2355 val = is_store ? argument(3) : NULL;
2356 }
2358 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2360 // First guess at the value type.
2361 const Type *value_type = Type::get_const_basic_type(type);
2363 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2364 // there was not enough information to nail it down.
2365 Compile::AliasType* alias_type = C->alias_type(adr_type);
2366 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2368 // We will need memory barriers unless we can determine a unique
2369 // alias category for this reference. (Note: If for some reason
2370 // the barriers get omitted and the unsafe reference begins to "pollute"
2371 // the alias analysis of the rest of the graph, either Compile::can_alias
2372 // or Compile::must_alias will throw a diagnostic assert.)
2373 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2375 // If we are reading the value of the referent field of a Reference
2376 // object (either by using Unsafe directly or through reflection)
2377 // then, if G1 is enabled, we need to record the referent in an
2378 // SATB log buffer using the pre-barrier mechanism.
2379 // Also we need to add memory barrier to prevent commoning reads
2380 // from this field across safepoint since GC can change its value.
2381 bool need_read_barrier = !is_native_ptr && !is_store &&
2382 offset != top() && heap_base_oop != top();
2384 if (!is_store && type == T_OBJECT) {
2385 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2386 if (tjp != NULL) {
2387 value_type = tjp;
2388 }
2389 }
2391 receiver = null_check(receiver);
2392 if (stopped()) {
2393 return true;
2394 }
2395 // Heap pointers get a null-check from the interpreter,
2396 // as a courtesy. However, this is not guaranteed by Unsafe,
2397 // and it is not possible to fully distinguish unintended nulls
2398 // from intended ones in this API.
2400 if (is_volatile) {
2401 // We need to emit leading and trailing CPU membars (see below) in
2402 // addition to memory membars when is_volatile. This is a little
2403 // too strong, but avoids the need to insert per-alias-type
2404 // volatile membars (for stores; compare Parse::do_put_xxx), which
2405 // we cannot do effectively here because we probably only have a
2406 // rough approximation of type.
2407 need_mem_bar = true;
2408 // For Stores, place a memory ordering barrier now.
2409 if (is_store)
2410 insert_mem_bar(Op_MemBarRelease);
2411 }
2413 // Memory barrier to prevent normal and 'unsafe' accesses from
2414 // bypassing each other. Happens after null checks, so the
2415 // exception paths do not take memory state from the memory barrier,
2416 // so there's no problems making a strong assert about mixing users
2417 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2418 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2419 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2421 if (!is_store) {
2422 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2423 // load value
2424 switch (type) {
2425 case T_BOOLEAN:
2426 case T_CHAR:
2427 case T_BYTE:
2428 case T_SHORT:
2429 case T_INT:
2430 case T_LONG:
2431 case T_FLOAT:
2432 case T_DOUBLE:
2433 break;
2434 case T_OBJECT:
2435 if (need_read_barrier) {
2436 insert_pre_barrier(heap_base_oop, offset, p, !(is_volatile || need_mem_bar));
2437 }
2438 break;
2439 case T_ADDRESS:
2440 // Cast to an int type.
2441 p = _gvn.transform(new (C) CastP2XNode(NULL, p));
2442 p = ConvX2L(p);
2443 break;
2444 default:
2445 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2446 break;
2447 }
2448 // The load node has the control of the preceding MemBarCPUOrder. All
2449 // following nodes will have the control of the MemBarCPUOrder inserted at
2450 // the end of this method. So, pushing the load onto the stack at a later
2451 // point is fine.
2452 set_result(p);
2453 } else {
2454 // place effect of store into memory
2455 switch (type) {
2456 case T_DOUBLE:
2457 val = dstore_rounding(val);
2458 break;
2459 case T_ADDRESS:
2460 // Repackage the long as a pointer.
2461 val = ConvL2X(val);
2462 val = _gvn.transform( new (C) CastX2PNode(val) );
2463 break;
2464 }
2466 if (type != T_OBJECT ) {
2467 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2468 } else {
2469 // Possibly an oop being stored to Java heap or native memory
2470 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2471 // oop to Java heap.
2472 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2473 } else {
2474 // We can't tell at compile time if we are storing in the Java heap or outside
2475 // of it. So we need to emit code to conditionally do the proper type of
2476 // store.
2478 IdealKit ideal(this);
2479 #define __ ideal.
2480 // QQQ who knows what probability is here??
2481 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2482 // Sync IdealKit and graphKit.
2483 sync_kit(ideal);
2484 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2485 // Update IdealKit memory.
2486 __ sync_kit(this);
2487 } __ else_(); {
2488 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2489 } __ end_if();
2490 // Final sync IdealKit and GraphKit.
2491 final_sync(ideal);
2492 #undef __
2493 }
2494 }
2495 }
2497 if (is_volatile) {
2498 if (!is_store)
2499 insert_mem_bar(Op_MemBarAcquire);
2500 else
2501 insert_mem_bar(Op_MemBarVolatile);
2502 }
2504 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2506 return true;
2507 }
2509 //----------------------------inline_unsafe_prefetch----------------------------
2511 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2512 #ifndef PRODUCT
2513 {
2514 ResourceMark rm;
2515 // Check the signatures.
2516 ciSignature* sig = callee()->signature();
2517 #ifdef ASSERT
2518 // Object getObject(Object base, int/long offset), etc.
2519 BasicType rtype = sig->return_type()->basic_type();
2520 if (!is_native_ptr) {
2521 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2522 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2523 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2524 } else {
2525 assert(sig->count() == 1, "native prefetch has 1 argument");
2526 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2527 }
2528 #endif // ASSERT
2529 }
2530 #endif // !PRODUCT
2532 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2534 const int idx = is_static ? 0 : 1;
2535 if (!is_static) {
2536 null_check_receiver();
2537 if (stopped()) {
2538 return true;
2539 }
2540 }
2542 // Build address expression. See the code in inline_unsafe_access.
2543 Node *adr;
2544 if (!is_native_ptr) {
2545 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2546 Node* base = argument(idx + 0); // type: oop
2547 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2548 Node* offset = argument(idx + 1); // type: long
2549 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2550 // to be plain byte offsets, which are also the same as those accepted
2551 // by oopDesc::field_base.
2552 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2553 "fieldOffset must be byte-scaled");
2554 // 32-bit machines ignore the high half!
2555 offset = ConvL2X(offset);
2556 adr = make_unsafe_address(base, offset);
2557 } else {
2558 Node* ptr = argument(idx + 0); // type: long
2559 ptr = ConvL2X(ptr); // adjust Java long to machine word
2560 adr = make_unsafe_address(NULL, ptr);
2561 }
2563 // Generate the read or write prefetch
2564 Node *prefetch;
2565 if (is_store) {
2566 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2567 } else {
2568 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2569 }
2570 prefetch->init_req(0, control());
2571 set_i_o(_gvn.transform(prefetch));
2573 return true;
2574 }
2576 //----------------------------inline_unsafe_load_store----------------------------
2577 // This method serves a couple of different customers (depending on LoadStoreKind):
2578 //
2579 // LS_cmpxchg:
2580 // public final native boolean compareAndSwapObject(Object o, long offset, Object expected, Object x);
2581 // public final native boolean compareAndSwapInt( Object o, long offset, int expected, int x);
2582 // public final native boolean compareAndSwapLong( Object o, long offset, long expected, long x);
2583 //
2584 // LS_xadd:
2585 // public int getAndAddInt( Object o, long offset, int delta)
2586 // public long getAndAddLong(Object o, long offset, long delta)
2587 //
2588 // LS_xchg:
2589 // int getAndSet(Object o, long offset, int newValue)
2590 // long getAndSet(Object o, long offset, long newValue)
2591 // Object getAndSet(Object o, long offset, Object newValue)
2592 //
2593 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2594 // This basic scheme here is the same as inline_unsafe_access, but
2595 // differs in enough details that combining them would make the code
2596 // overly confusing. (This is a true fact! I originally combined
2597 // them, but even I was confused by it!) As much code/comments as
2598 // possible are retained from inline_unsafe_access though to make
2599 // the correspondences clearer. - dl
2601 if (callee()->is_static()) return false; // caller must have the capability!
2603 #ifndef PRODUCT
2604 BasicType rtype;
2605 {
2606 ResourceMark rm;
2607 // Check the signatures.
2608 ciSignature* sig = callee()->signature();
2609 rtype = sig->return_type()->basic_type();
2610 if (kind == LS_xadd || kind == LS_xchg) {
2611 // Check the signatures.
2612 #ifdef ASSERT
2613 assert(rtype == type, "get and set must return the expected type");
2614 assert(sig->count() == 3, "get and set has 3 arguments");
2615 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2616 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2617 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2618 #endif // ASSERT
2619 } else if (kind == LS_cmpxchg) {
2620 // Check the signatures.
2621 #ifdef ASSERT
2622 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2623 assert(sig->count() == 4, "CAS has 4 arguments");
2624 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2625 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2626 #endif // ASSERT
2627 } else {
2628 ShouldNotReachHere();
2629 }
2630 }
2631 #endif //PRODUCT
2633 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2635 // Get arguments:
2636 Node* receiver = NULL;
2637 Node* base = NULL;
2638 Node* offset = NULL;
2639 Node* oldval = NULL;
2640 Node* newval = NULL;
2641 if (kind == LS_cmpxchg) {
2642 const bool two_slot_type = type2size[type] == 2;
2643 receiver = argument(0); // type: oop
2644 base = argument(1); // type: oop
2645 offset = argument(2); // type: long
2646 oldval = argument(4); // type: oop, int, or long
2647 newval = argument(two_slot_type ? 6 : 5); // type: oop, int, or long
2648 } else if (kind == LS_xadd || kind == LS_xchg){
2649 receiver = argument(0); // type: oop
2650 base = argument(1); // type: oop
2651 offset = argument(2); // type: long
2652 oldval = NULL;
2653 newval = argument(4); // type: oop, int, or long
2654 }
2656 // Null check receiver.
2657 receiver = null_check(receiver);
2658 if (stopped()) {
2659 return true;
2660 }
2662 // Build field offset expression.
2663 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2664 // to be plain byte offsets, which are also the same as those accepted
2665 // by oopDesc::field_base.
2666 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2667 // 32-bit machines ignore the high half of long offsets
2668 offset = ConvL2X(offset);
2669 Node* adr = make_unsafe_address(base, offset);
2670 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2672 // For CAS, unlike inline_unsafe_access, there seems no point in
2673 // trying to refine types. Just use the coarse types here.
2674 const Type *value_type = Type::get_const_basic_type(type);
2675 Compile::AliasType* alias_type = C->alias_type(adr_type);
2676 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2678 if (kind == LS_xchg && type == T_OBJECT) {
2679 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2680 if (tjp != NULL) {
2681 value_type = tjp;
2682 }
2683 }
2685 int alias_idx = C->get_alias_index(adr_type);
2687 // Memory-model-wise, a LoadStore acts like a little synchronized
2688 // block, so needs barriers on each side. These don't translate
2689 // into actual barriers on most machines, but we still need rest of
2690 // compiler to respect ordering.
2692 insert_mem_bar(Op_MemBarRelease);
2693 insert_mem_bar(Op_MemBarCPUOrder);
2695 // 4984716: MemBars must be inserted before this
2696 // memory node in order to avoid a false
2697 // dependency which will confuse the scheduler.
2698 Node *mem = memory(alias_idx);
2700 // For now, we handle only those cases that actually exist: ints,
2701 // longs, and Object. Adding others should be straightforward.
2702 Node* load_store;
2703 switch(type) {
2704 case T_INT:
2705 if (kind == LS_xadd) {
2706 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2707 } else if (kind == LS_xchg) {
2708 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2709 } else if (kind == LS_cmpxchg) {
2710 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2711 } else {
2712 ShouldNotReachHere();
2713 }
2714 break;
2715 case T_LONG:
2716 if (kind == LS_xadd) {
2717 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2718 } else if (kind == LS_xchg) {
2719 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2720 } else if (kind == LS_cmpxchg) {
2721 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2722 } else {
2723 ShouldNotReachHere();
2724 }
2725 break;
2726 case T_OBJECT:
2727 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2728 // could be delayed during Parse (for example, in adjust_map_after_if()).
2729 // Execute transformation here to avoid barrier generation in such case.
2730 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2731 newval = _gvn.makecon(TypePtr::NULL_PTR);
2733 // Reference stores need a store barrier.
2734 pre_barrier(true /* do_load*/,
2735 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2736 NULL /* pre_val*/,
2737 T_OBJECT);
2738 #ifdef _LP64
2739 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2740 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2741 if (kind == LS_xchg) {
2742 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2743 newval_enc, adr_type, value_type->make_narrowoop()));
2744 } else {
2745 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2746 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2747 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2748 newval_enc, oldval_enc));
2749 }
2750 } else
2751 #endif
2752 {
2753 if (kind == LS_xchg) {
2754 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
2755 } else {
2756 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2757 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2758 }
2759 }
2760 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
2761 break;
2762 default:
2763 fatal(err_msg_res("unexpected type %d: %s", type, type2name(type)));
2764 break;
2765 }
2767 // SCMemProjNodes represent the memory state of a LoadStore. Their
2768 // main role is to prevent LoadStore nodes from being optimized away
2769 // when their results aren't used.
2770 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
2771 set_memory(proj, alias_idx);
2773 // Add the trailing membar surrounding the access
2774 insert_mem_bar(Op_MemBarCPUOrder);
2775 insert_mem_bar(Op_MemBarAcquire);
2777 #ifdef _LP64
2778 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
2779 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
2780 }
2781 #endif
2783 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
2784 set_result(load_store);
2785 return true;
2786 }
2788 //----------------------------inline_unsafe_ordered_store----------------------
2789 // public native void sun.misc.Unsafe.putOrderedObject(Object o, long offset, Object x);
2790 // public native void sun.misc.Unsafe.putOrderedInt(Object o, long offset, int x);
2791 // public native void sun.misc.Unsafe.putOrderedLong(Object o, long offset, long x);
2792 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2793 // This is another variant of inline_unsafe_access, differing in
2794 // that it always issues store-store ("release") barrier and ensures
2795 // store-atomicity (which only matters for "long").
2797 if (callee()->is_static()) return false; // caller must have the capability!
2799 #ifndef PRODUCT
2800 {
2801 ResourceMark rm;
2802 // Check the signatures.
2803 ciSignature* sig = callee()->signature();
2804 #ifdef ASSERT
2805 BasicType rtype = sig->return_type()->basic_type();
2806 assert(rtype == T_VOID, "must return void");
2807 assert(sig->count() == 3, "has 3 arguments");
2808 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2809 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2810 #endif // ASSERT
2811 }
2812 #endif //PRODUCT
2814 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2816 // Get arguments:
2817 Node* receiver = argument(0); // type: oop
2818 Node* base = argument(1); // type: oop
2819 Node* offset = argument(2); // type: long
2820 Node* val = argument(4); // type: oop, int, or long
2822 // Null check receiver.
2823 receiver = null_check(receiver);
2824 if (stopped()) {
2825 return true;
2826 }
2828 // Build field offset expression.
2829 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2830 // 32-bit machines ignore the high half of long offsets
2831 offset = ConvL2X(offset);
2832 Node* adr = make_unsafe_address(base, offset);
2833 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2834 const Type *value_type = Type::get_const_basic_type(type);
2835 Compile::AliasType* alias_type = C->alias_type(adr_type);
2837 insert_mem_bar(Op_MemBarRelease);
2838 insert_mem_bar(Op_MemBarCPUOrder);
2839 // Ensure that the store is atomic for longs:
2840 const bool require_atomic_access = true;
2841 Node* store;
2842 if (type == T_OBJECT) // reference stores need a store barrier.
2843 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2844 else {
2845 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2846 }
2847 insert_mem_bar(Op_MemBarCPUOrder);
2848 return true;
2849 }
2851 bool LibraryCallKit::inline_unsafe_fence(vmIntrinsics::ID id) {
2852 // Regardless of form, don't allow previous ld/st to move down,
2853 // then issue acquire, release, or volatile mem_bar.
2854 insert_mem_bar(Op_MemBarCPUOrder);
2855 switch(id) {
2856 case vmIntrinsics::_loadFence:
2857 insert_mem_bar(Op_MemBarAcquire);
2858 return true;
2859 case vmIntrinsics::_storeFence:
2860 insert_mem_bar(Op_MemBarRelease);
2861 return true;
2862 case vmIntrinsics::_fullFence:
2863 insert_mem_bar(Op_MemBarVolatile);
2864 return true;
2865 default:
2866 fatal_unexpected_iid(id);
2867 return false;
2868 }
2869 }
2871 //----------------------------inline_unsafe_allocate---------------------------
2872 // public native Object sun.mics.Unsafe.allocateInstance(Class<?> cls);
2873 bool LibraryCallKit::inline_unsafe_allocate() {
2874 if (callee()->is_static()) return false; // caller must have the capability!
2876 null_check_receiver(); // null-check, then ignore
2877 Node* cls = null_check(argument(1));
2878 if (stopped()) return true;
2880 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2881 kls = null_check(kls);
2882 if (stopped()) return true; // argument was like int.class
2884 // Note: The argument might still be an illegal value like
2885 // Serializable.class or Object[].class. The runtime will handle it.
2886 // But we must make an explicit check for initialization.
2887 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
2888 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
2889 // can generate code to load it as unsigned byte.
2890 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2891 Node* bits = intcon(InstanceKlass::fully_initialized);
2892 Node* test = _gvn.transform(new (C) SubINode(inst, bits));
2893 // The 'test' is non-zero if we need to take a slow path.
2895 Node* obj = new_instance(kls, test);
2896 set_result(obj);
2897 return true;
2898 }
2900 #ifdef TRACE_HAVE_INTRINSICS
2901 /*
2902 * oop -> myklass
2903 * myklass->trace_id |= USED
2904 * return myklass->trace_id & ~0x3
2905 */
2906 bool LibraryCallKit::inline_native_classID() {
2907 null_check_receiver(); // null-check, then ignore
2908 Node* cls = null_check(argument(1), T_OBJECT);
2909 Node* kls = load_klass_from_mirror(cls, false, NULL, 0);
2910 kls = null_check(kls, T_OBJECT);
2911 ByteSize offset = TRACE_ID_OFFSET;
2912 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2913 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2914 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2915 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
2916 Node* clsused = longcon(0x01l); // set the class bit
2917 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
2919 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2920 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2921 set_result(andl);
2922 return true;
2923 }
2925 bool LibraryCallKit::inline_native_threadID() {
2926 Node* tls_ptr = NULL;
2927 Node* cur_thr = generate_current_thread(tls_ptr);
2928 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2929 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2930 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2932 Node* threadid = NULL;
2933 size_t thread_id_size = OSThread::thread_id_size();
2934 if (thread_id_size == (size_t) BytesPerLong) {
2935 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2936 } else if (thread_id_size == (size_t) BytesPerInt) {
2937 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2938 } else {
2939 ShouldNotReachHere();
2940 }
2941 set_result(threadid);
2942 return true;
2943 }
2944 #endif
2946 //------------------------inline_native_time_funcs--------------
2947 // inline code for System.currentTimeMillis() and System.nanoTime()
2948 // these have the same type and signature
2949 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2950 const TypeFunc* tf = OptoRuntime::void_long_Type();
2951 const TypePtr* no_memory_effects = NULL;
2952 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2953 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
2954 #ifdef ASSERT
2955 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+1));
2956 assert(value_top == top(), "second value must be top");
2957 #endif
2958 set_result(value);
2959 return true;
2960 }
2962 //------------------------inline_native_currentThread------------------
2963 bool LibraryCallKit::inline_native_currentThread() {
2964 Node* junk = NULL;
2965 set_result(generate_current_thread(junk));
2966 return true;
2967 }
2969 //------------------------inline_native_isInterrupted------------------
2970 // private native boolean java.lang.Thread.isInterrupted(boolean ClearInterrupted);
2971 bool LibraryCallKit::inline_native_isInterrupted() {
2972 // Add a fast path to t.isInterrupted(clear_int):
2973 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2974 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2975 // So, in the common case that the interrupt bit is false,
2976 // we avoid making a call into the VM. Even if the interrupt bit
2977 // is true, if the clear_int argument is false, we avoid the VM call.
2978 // However, if the receiver is not currentThread, we must call the VM,
2979 // because there must be some locking done around the operation.
2981 // We only go to the fast case code if we pass two guards.
2982 // Paths which do not pass are accumulated in the slow_region.
2984 enum {
2985 no_int_result_path = 1, // t == Thread.current() && !TLS._osthread._interrupted
2986 no_clear_result_path = 2, // t == Thread.current() && TLS._osthread._interrupted && !clear_int
2987 slow_result_path = 3, // slow path: t.isInterrupted(clear_int)
2988 PATH_LIMIT
2989 };
2991 // Ensure that it's not possible to move the load of TLS._osthread._interrupted flag
2992 // out of the function.
2993 insert_mem_bar(Op_MemBarCPUOrder);
2995 RegionNode* result_rgn = new (C) RegionNode(PATH_LIMIT);
2996 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
2998 RegionNode* slow_region = new (C) RegionNode(1);
2999 record_for_igvn(slow_region);
3001 // (a) Receiving thread must be the current thread.
3002 Node* rec_thr = argument(0);
3003 Node* tls_ptr = NULL;
3004 Node* cur_thr = generate_current_thread(tls_ptr);
3005 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3006 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3008 generate_slow_guard(bol_thr, slow_region);
3010 // (b) Interrupt bit on TLS must be false.
3011 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3012 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3013 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3015 // Set the control input on the field _interrupted read to prevent it floating up.
3016 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3017 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3018 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3020 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3022 // First fast path: if (!TLS._interrupted) return false;
3023 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3024 result_rgn->init_req(no_int_result_path, false_bit);
3025 result_val->init_req(no_int_result_path, intcon(0));
3027 // drop through to next case
3028 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3030 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3031 Node* clr_arg = argument(1);
3032 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3033 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3034 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3036 // Second fast path: ... else if (!clear_int) return true;
3037 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3038 result_rgn->init_req(no_clear_result_path, false_arg);
3039 result_val->init_req(no_clear_result_path, intcon(1));
3041 // drop through to next case
3042 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3044 // (d) Otherwise, go to the slow path.
3045 slow_region->add_req(control());
3046 set_control( _gvn.transform(slow_region) );
3048 if (stopped()) {
3049 // There is no slow path.
3050 result_rgn->init_req(slow_result_path, top());
3051 result_val->init_req(slow_result_path, top());
3052 } else {
3053 // non-virtual because it is a private non-static
3054 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3056 Node* slow_val = set_results_for_java_call(slow_call);
3057 // this->control() comes from set_results_for_java_call
3059 Node* fast_io = slow_call->in(TypeFunc::I_O);
3060 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3062 // These two phis are pre-filled with copies of of the fast IO and Memory
3063 PhiNode* result_mem = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3064 PhiNode* result_io = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3066 result_rgn->init_req(slow_result_path, control());
3067 result_io ->init_req(slow_result_path, i_o());
3068 result_mem->init_req(slow_result_path, reset_memory());
3069 result_val->init_req(slow_result_path, slow_val);
3071 set_all_memory(_gvn.transform(result_mem));
3072 set_i_o( _gvn.transform(result_io));
3073 }
3075 C->set_has_split_ifs(true); // Has chance for split-if optimization
3076 set_result(result_rgn, result_val);
3077 return true;
3078 }
3080 //---------------------------load_mirror_from_klass----------------------------
3081 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3082 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3083 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3084 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3085 }
3087 //-----------------------load_klass_from_mirror_common-------------------------
3088 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3089 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3090 // and branch to the given path on the region.
3091 // If never_see_null, take an uncommon trap on null, so we can optimistically
3092 // compile for the non-null case.
3093 // If the region is NULL, force never_see_null = true.
3094 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3095 bool never_see_null,
3096 RegionNode* region,
3097 int null_path,
3098 int offset) {
3099 if (region == NULL) never_see_null = true;
3100 Node* p = basic_plus_adr(mirror, offset);
3101 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3102 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3103 Node* null_ctl = top();
3104 kls = null_check_oop(kls, &null_ctl, never_see_null);
3105 if (region != NULL) {
3106 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3107 region->init_req(null_path, null_ctl);
3108 } else {
3109 assert(null_ctl == top(), "no loose ends");
3110 }
3111 return kls;
3112 }
3114 //--------------------(inline_native_Class_query helpers)---------------------
3115 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3116 // Fall through if (mods & mask) == bits, take the guard otherwise.
3117 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3118 // Branch around if the given klass has the given modifier bit set.
3119 // Like generate_guard, adds a new path onto the region.
3120 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3121 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3122 Node* mask = intcon(modifier_mask);
3123 Node* bits = intcon(modifier_bits);
3124 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3125 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3126 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3127 return generate_fair_guard(bol, region);
3128 }
3129 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3130 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3131 }
3133 //-------------------------inline_native_Class_query-------------------
3134 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3135 const Type* return_type = TypeInt::BOOL;
3136 Node* prim_return_value = top(); // what happens if it's a primitive class?
3137 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3138 bool expect_prim = false; // most of these guys expect to work on refs
3140 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3142 Node* mirror = argument(0);
3143 Node* obj = top();
3145 switch (id) {
3146 case vmIntrinsics::_isInstance:
3147 // nothing is an instance of a primitive type
3148 prim_return_value = intcon(0);
3149 obj = argument(1);
3150 break;
3151 case vmIntrinsics::_getModifiers:
3152 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3153 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3154 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3155 break;
3156 case vmIntrinsics::_isInterface:
3157 prim_return_value = intcon(0);
3158 break;
3159 case vmIntrinsics::_isArray:
3160 prim_return_value = intcon(0);
3161 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3162 break;
3163 case vmIntrinsics::_isPrimitive:
3164 prim_return_value = intcon(1);
3165 expect_prim = true; // obviously
3166 break;
3167 case vmIntrinsics::_getSuperclass:
3168 prim_return_value = null();
3169 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3170 break;
3171 case vmIntrinsics::_getComponentType:
3172 prim_return_value = null();
3173 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3174 break;
3175 case vmIntrinsics::_getClassAccessFlags:
3176 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3177 return_type = TypeInt::INT; // not bool! 6297094
3178 break;
3179 default:
3180 fatal_unexpected_iid(id);
3181 break;
3182 }
3184 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3185 if (mirror_con == NULL) return false; // cannot happen?
3187 #ifndef PRODUCT
3188 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3189 ciType* k = mirror_con->java_mirror_type();
3190 if (k) {
3191 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3192 k->print_name();
3193 tty->cr();
3194 }
3195 }
3196 #endif
3198 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3199 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3200 record_for_igvn(region);
3201 PhiNode* phi = new (C) PhiNode(region, return_type);
3203 // The mirror will never be null of Reflection.getClassAccessFlags, however
3204 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3205 // if it is. See bug 4774291.
3207 // For Reflection.getClassAccessFlags(), the null check occurs in
3208 // the wrong place; see inline_unsafe_access(), above, for a similar
3209 // situation.
3210 mirror = null_check(mirror);
3211 // If mirror or obj is dead, only null-path is taken.
3212 if (stopped()) return true;
3214 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3216 // Now load the mirror's klass metaobject, and null-check it.
3217 // Side-effects region with the control path if the klass is null.
3218 Node* kls = load_klass_from_mirror(mirror, never_see_null, region, _prim_path);
3219 // If kls is null, we have a primitive mirror.
3220 phi->init_req(_prim_path, prim_return_value);
3221 if (stopped()) { set_result(region, phi); return true; }
3223 Node* p; // handy temp
3224 Node* null_ctl;
3226 // Now that we have the non-null klass, we can perform the real query.
3227 // For constant classes, the query will constant-fold in LoadNode::Value.
3228 Node* query_value = top();
3229 switch (id) {
3230 case vmIntrinsics::_isInstance:
3231 // nothing is an instance of a primitive type
3232 query_value = gen_instanceof(obj, kls);
3233 break;
3235 case vmIntrinsics::_getModifiers:
3236 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3237 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3238 break;
3240 case vmIntrinsics::_isInterface:
3241 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3242 if (generate_interface_guard(kls, region) != NULL)
3243 // A guard was added. If the guard is taken, it was an interface.
3244 phi->add_req(intcon(1));
3245 // If we fall through, it's a plain class.
3246 query_value = intcon(0);
3247 break;
3249 case vmIntrinsics::_isArray:
3250 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3251 if (generate_array_guard(kls, region) != NULL)
3252 // A guard was added. If the guard is taken, it was an array.
3253 phi->add_req(intcon(1));
3254 // If we fall through, it's a plain class.
3255 query_value = intcon(0);
3256 break;
3258 case vmIntrinsics::_isPrimitive:
3259 query_value = intcon(0); // "normal" path produces false
3260 break;
3262 case vmIntrinsics::_getSuperclass:
3263 // The rules here are somewhat unfortunate, but we can still do better
3264 // with random logic than with a JNI call.
3265 // Interfaces store null or Object as _super, but must report null.
3266 // Arrays store an intermediate super as _super, but must report Object.
3267 // Other types can report the actual _super.
3268 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3269 if (generate_interface_guard(kls, region) != NULL)
3270 // A guard was added. If the guard is taken, it was an interface.
3271 phi->add_req(null());
3272 if (generate_array_guard(kls, region) != NULL)
3273 // A guard was added. If the guard is taken, it was an array.
3274 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3275 // If we fall through, it's a plain class. Get its _super.
3276 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3277 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3278 null_ctl = top();
3279 kls = null_check_oop(kls, &null_ctl);
3280 if (null_ctl != top()) {
3281 // If the guard is taken, Object.superClass is null (both klass and mirror).
3282 region->add_req(null_ctl);
3283 phi ->add_req(null());
3284 }
3285 if (!stopped()) {
3286 query_value = load_mirror_from_klass(kls);
3287 }
3288 break;
3290 case vmIntrinsics::_getComponentType:
3291 if (generate_array_guard(kls, region) != NULL) {
3292 // Be sure to pin the oop load to the guard edge just created:
3293 Node* is_array_ctrl = region->in(region->req()-1);
3294 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3295 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3296 phi->add_req(cmo);
3297 }
3298 query_value = null(); // non-array case is null
3299 break;
3301 case vmIntrinsics::_getClassAccessFlags:
3302 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3303 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3304 break;
3306 default:
3307 fatal_unexpected_iid(id);
3308 break;
3309 }
3311 // Fall-through is the normal case of a query to a real class.
3312 phi->init_req(1, query_value);
3313 region->init_req(1, control());
3315 C->set_has_split_ifs(true); // Has chance for split-if optimization
3316 set_result(region, phi);
3317 return true;
3318 }
3320 //--------------------------inline_native_subtype_check------------------------
3321 // This intrinsic takes the JNI calls out of the heart of
3322 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3323 bool LibraryCallKit::inline_native_subtype_check() {
3324 // Pull both arguments off the stack.
3325 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3326 args[0] = argument(0);
3327 args[1] = argument(1);
3328 Node* klasses[2]; // corresponding Klasses: superk, subk
3329 klasses[0] = klasses[1] = top();
3331 enum {
3332 // A full decision tree on {superc is prim, subc is prim}:
3333 _prim_0_path = 1, // {P,N} => false
3334 // {P,P} & superc!=subc => false
3335 _prim_same_path, // {P,P} & superc==subc => true
3336 _prim_1_path, // {N,P} => false
3337 _ref_subtype_path, // {N,N} & subtype check wins => true
3338 _both_ref_path, // {N,N} & subtype check loses => false
3339 PATH_LIMIT
3340 };
3342 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3343 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3344 record_for_igvn(region);
3346 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3347 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3348 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3350 // First null-check both mirrors and load each mirror's klass metaobject.
3351 int which_arg;
3352 for (which_arg = 0; which_arg <= 1; which_arg++) {
3353 Node* arg = args[which_arg];
3354 arg = null_check(arg);
3355 if (stopped()) break;
3356 args[which_arg] = arg;
3358 Node* p = basic_plus_adr(arg, class_klass_offset);
3359 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3360 klasses[which_arg] = _gvn.transform(kls);
3361 }
3363 // Having loaded both klasses, test each for null.
3364 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3365 for (which_arg = 0; which_arg <= 1; which_arg++) {
3366 Node* kls = klasses[which_arg];
3367 Node* null_ctl = top();
3368 kls = null_check_oop(kls, &null_ctl, never_see_null);
3369 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3370 region->init_req(prim_path, null_ctl);
3371 if (stopped()) break;
3372 klasses[which_arg] = kls;
3373 }
3375 if (!stopped()) {
3376 // now we have two reference types, in klasses[0..1]
3377 Node* subk = klasses[1]; // the argument to isAssignableFrom
3378 Node* superk = klasses[0]; // the receiver
3379 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3380 // now we have a successful reference subtype check
3381 region->set_req(_ref_subtype_path, control());
3382 }
3384 // If both operands are primitive (both klasses null), then
3385 // we must return true when they are identical primitives.
3386 // It is convenient to test this after the first null klass check.
3387 set_control(region->in(_prim_0_path)); // go back to first null check
3388 if (!stopped()) {
3389 // Since superc is primitive, make a guard for the superc==subc case.
3390 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3391 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3392 generate_guard(bol_eq, region, PROB_FAIR);
3393 if (region->req() == PATH_LIMIT+1) {
3394 // A guard was added. If the added guard is taken, superc==subc.
3395 region->swap_edges(PATH_LIMIT, _prim_same_path);
3396 region->del_req(PATH_LIMIT);
3397 }
3398 region->set_req(_prim_0_path, control()); // Not equal after all.
3399 }
3401 // these are the only paths that produce 'true':
3402 phi->set_req(_prim_same_path, intcon(1));
3403 phi->set_req(_ref_subtype_path, intcon(1));
3405 // pull together the cases:
3406 assert(region->req() == PATH_LIMIT, "sane region");
3407 for (uint i = 1; i < region->req(); i++) {
3408 Node* ctl = region->in(i);
3409 if (ctl == NULL || ctl == top()) {
3410 region->set_req(i, top());
3411 phi ->set_req(i, top());
3412 } else if (phi->in(i) == NULL) {
3413 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3414 }
3415 }
3417 set_control(_gvn.transform(region));
3418 set_result(_gvn.transform(phi));
3419 return true;
3420 }
3422 //---------------------generate_array_guard_common------------------------
3423 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3424 bool obj_array, bool not_array) {
3425 // If obj_array/non_array==false/false:
3426 // Branch around if the given klass is in fact an array (either obj or prim).
3427 // If obj_array/non_array==false/true:
3428 // Branch around if the given klass is not an array klass of any kind.
3429 // If obj_array/non_array==true/true:
3430 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3431 // If obj_array/non_array==true/false:
3432 // Branch around if the kls is an oop array (Object[] or subtype)
3433 //
3434 // Like generate_guard, adds a new path onto the region.
3435 jint layout_con = 0;
3436 Node* layout_val = get_layout_helper(kls, layout_con);
3437 if (layout_val == NULL) {
3438 bool query = (obj_array
3439 ? Klass::layout_helper_is_objArray(layout_con)
3440 : Klass::layout_helper_is_array(layout_con));
3441 if (query == not_array) {
3442 return NULL; // never a branch
3443 } else { // always a branch
3444 Node* always_branch = control();
3445 if (region != NULL)
3446 region->add_req(always_branch);
3447 set_control(top());
3448 return always_branch;
3449 }
3450 }
3451 // Now test the correct condition.
3452 jint nval = (obj_array
3453 ? ((jint)Klass::_lh_array_tag_type_value
3454 << Klass::_lh_array_tag_shift)
3455 : Klass::_lh_neutral_value);
3456 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3457 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3458 // invert the test if we are looking for a non-array
3459 if (not_array) btest = BoolTest(btest).negate();
3460 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3461 return generate_fair_guard(bol, region);
3462 }
3465 //-----------------------inline_native_newArray--------------------------
3466 // private static native Object java.lang.reflect.newArray(Class<?> componentType, int length);
3467 bool LibraryCallKit::inline_native_newArray() {
3468 Node* mirror = argument(0);
3469 Node* count_val = argument(1);
3471 mirror = null_check(mirror);
3472 // If mirror or obj is dead, only null-path is taken.
3473 if (stopped()) return true;
3475 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3476 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3477 PhiNode* result_val = new(C) PhiNode(result_reg,
3478 TypeInstPtr::NOTNULL);
3479 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3480 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3481 TypePtr::BOTTOM);
3483 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3484 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3485 result_reg, _slow_path);
3486 Node* normal_ctl = control();
3487 Node* no_array_ctl = result_reg->in(_slow_path);
3489 // Generate code for the slow case. We make a call to newArray().
3490 set_control(no_array_ctl);
3491 if (!stopped()) {
3492 // Either the input type is void.class, or else the
3493 // array klass has not yet been cached. Either the
3494 // ensuing call will throw an exception, or else it
3495 // will cache the array klass for next time.
3496 PreserveJVMState pjvms(this);
3497 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3498 Node* slow_result = set_results_for_java_call(slow_call);
3499 // this->control() comes from set_results_for_java_call
3500 result_reg->set_req(_slow_path, control());
3501 result_val->set_req(_slow_path, slow_result);
3502 result_io ->set_req(_slow_path, i_o());
3503 result_mem->set_req(_slow_path, reset_memory());
3504 }
3506 set_control(normal_ctl);
3507 if (!stopped()) {
3508 // Normal case: The array type has been cached in the java.lang.Class.
3509 // The following call works fine even if the array type is polymorphic.
3510 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3511 Node* obj = new_array(klass_node, count_val, 0); // no arguments to push
3512 result_reg->init_req(_normal_path, control());
3513 result_val->init_req(_normal_path, obj);
3514 result_io ->init_req(_normal_path, i_o());
3515 result_mem->init_req(_normal_path, reset_memory());
3516 }
3518 // Return the combined state.
3519 set_i_o( _gvn.transform(result_io) );
3520 set_all_memory( _gvn.transform(result_mem) );
3522 C->set_has_split_ifs(true); // Has chance for split-if optimization
3523 set_result(result_reg, result_val);
3524 return true;
3525 }
3527 //----------------------inline_native_getLength--------------------------
3528 // public static native int java.lang.reflect.Array.getLength(Object array);
3529 bool LibraryCallKit::inline_native_getLength() {
3530 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3532 Node* array = null_check(argument(0));
3533 // If array is dead, only null-path is taken.
3534 if (stopped()) return true;
3536 // Deoptimize if it is a non-array.
3537 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3539 if (non_array != NULL) {
3540 PreserveJVMState pjvms(this);
3541 set_control(non_array);
3542 uncommon_trap(Deoptimization::Reason_intrinsic,
3543 Deoptimization::Action_maybe_recompile);
3544 }
3546 // If control is dead, only non-array-path is taken.
3547 if (stopped()) return true;
3549 // The works fine even if the array type is polymorphic.
3550 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3551 Node* result = load_array_length(array);
3553 C->set_has_split_ifs(true); // Has chance for split-if optimization
3554 set_result(result);
3555 return true;
3556 }
3558 //------------------------inline_array_copyOf----------------------------
3559 // public static <T,U> T[] java.util.Arrays.copyOf( U[] original, int newLength, Class<? extends T[]> newType);
3560 // public static <T,U> T[] java.util.Arrays.copyOfRange(U[] original, int from, int to, Class<? extends T[]> newType);
3561 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3562 return false;
3563 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3565 // Get the arguments.
3566 Node* original = argument(0);
3567 Node* start = is_copyOfRange? argument(1): intcon(0);
3568 Node* end = is_copyOfRange? argument(2): argument(1);
3569 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3571 Node* newcopy;
3573 // Set the original stack and the reexecute bit for the interpreter to reexecute
3574 // the bytecode that invokes Arrays.copyOf if deoptimization happens.
3575 { PreserveReexecuteState preexecs(this);
3576 jvms()->set_should_reexecute(true);
3578 array_type_mirror = null_check(array_type_mirror);
3579 original = null_check(original);
3581 // Check if a null path was taken unconditionally.
3582 if (stopped()) return true;
3584 Node* orig_length = load_array_length(original);
3586 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, NULL, 0);
3587 klass_node = null_check(klass_node);
3589 RegionNode* bailout = new (C) RegionNode(1);
3590 record_for_igvn(bailout);
3592 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3593 // Bail out if that is so.
3594 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3595 if (not_objArray != NULL) {
3596 // Improve the klass node's type from the new optimistic assumption:
3597 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3598 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3599 Node* cast = new (C) CastPPNode(klass_node, akls);
3600 cast->init_req(0, control());
3601 klass_node = _gvn.transform(cast);
3602 }
3604 // Bail out if either start or end is negative.
3605 generate_negative_guard(start, bailout, &start);
3606 generate_negative_guard(end, bailout, &end);
3608 Node* length = end;
3609 if (_gvn.type(start) != TypeInt::ZERO) {
3610 length = _gvn.transform(new (C) SubINode(end, start));
3611 }
3613 // Bail out if length is negative.
3614 // Without this the new_array would throw
3615 // NegativeArraySizeException but IllegalArgumentException is what
3616 // should be thrown
3617 generate_negative_guard(length, bailout, &length);
3619 if (bailout->req() > 1) {
3620 PreserveJVMState pjvms(this);
3621 set_control(_gvn.transform(bailout));
3622 uncommon_trap(Deoptimization::Reason_intrinsic,
3623 Deoptimization::Action_maybe_recompile);
3624 }
3626 if (!stopped()) {
3627 // How many elements will we copy from the original?
3628 // The answer is MinI(orig_length - start, length).
3629 Node* orig_tail = _gvn.transform(new (C) SubINode(orig_length, start));
3630 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3632 newcopy = new_array(klass_node, length, 0); // no argments to push
3634 // Generate a direct call to the right arraycopy function(s).
3635 // We know the copy is disjoint but we might not know if the
3636 // oop stores need checking.
3637 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3638 // This will fail a store-check if x contains any non-nulls.
3639 bool disjoint_bases = true;
3640 // if start > orig_length then the length of the copy may be
3641 // negative.
3642 bool length_never_negative = !is_copyOfRange;
3643 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3644 original, start, newcopy, intcon(0), moved,
3645 disjoint_bases, length_never_negative);
3646 }
3647 } // original reexecute is set back here
3649 C->set_has_split_ifs(true); // Has chance for split-if optimization
3650 if (!stopped()) {
3651 set_result(newcopy);
3652 }
3653 return true;
3654 }
3657 //----------------------generate_virtual_guard---------------------------
3658 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3659 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3660 RegionNode* slow_region) {
3661 ciMethod* method = callee();
3662 int vtable_index = method->vtable_index();
3663 // Get the Method* out of the appropriate vtable entry.
3664 int entry_offset = (InstanceKlass::vtable_start_offset() +
3665 vtable_index*vtableEntry::size()) * wordSize +
3666 vtableEntry::method_offset_in_bytes();
3667 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3668 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3670 // Compare the target method with the expected method (e.g., Object.hashCode).
3671 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3673 Node* native_call = makecon(native_call_addr);
3674 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3675 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3677 return generate_slow_guard(test_native, slow_region);
3678 }
3680 //-----------------------generate_method_call----------------------------
3681 // Use generate_method_call to make a slow-call to the real
3682 // method if the fast path fails. An alternative would be to
3683 // use a stub like OptoRuntime::slow_arraycopy_Java.
3684 // This only works for expanding the current library call,
3685 // not another intrinsic. (E.g., don't use this for making an
3686 // arraycopy call inside of the copyOf intrinsic.)
3687 CallJavaNode*
3688 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3689 // When compiling the intrinsic method itself, do not use this technique.
3690 guarantee(callee() != C->method(), "cannot make slow-call to self");
3692 ciMethod* method = callee();
3693 // ensure the JVMS we have will be correct for this call
3694 guarantee(method_id == method->intrinsic_id(), "must match");
3696 const TypeFunc* tf = TypeFunc::make(method);
3697 CallJavaNode* slow_call;
3698 if (is_static) {
3699 assert(!is_virtual, "");
3700 slow_call = new(C) CallStaticJavaNode(tf,
3701 SharedRuntime::get_resolve_static_call_stub(),
3702 method, bci());
3703 } else if (is_virtual) {
3704 null_check_receiver();
3705 int vtable_index = Method::invalid_vtable_index;
3706 if (UseInlineCaches) {
3707 // Suppress the vtable call
3708 } else {
3709 // hashCode and clone are not a miranda methods,
3710 // so the vtable index is fixed.
3711 // No need to use the linkResolver to get it.
3712 vtable_index = method->vtable_index();
3713 }
3714 slow_call = new(C) CallDynamicJavaNode(tf,
3715 SharedRuntime::get_resolve_virtual_call_stub(),
3716 method, vtable_index, bci());
3717 } else { // neither virtual nor static: opt_virtual
3718 null_check_receiver();
3719 slow_call = new(C) CallStaticJavaNode(tf,
3720 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3721 method, bci());
3722 slow_call->set_optimized_virtual(true);
3723 }
3724 set_arguments_for_java_call(slow_call);
3725 set_edges_for_java_call(slow_call);
3726 return slow_call;
3727 }
3730 //------------------------------inline_native_hashcode--------------------
3731 // Build special case code for calls to hashCode on an object.
3732 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3733 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3734 assert(!(is_virtual && is_static), "either virtual, special, or static");
3736 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3738 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3739 PhiNode* result_val = new(C) PhiNode(result_reg,
3740 TypeInt::INT);
3741 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3742 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3743 TypePtr::BOTTOM);
3744 Node* obj = NULL;
3745 if (!is_static) {
3746 // Check for hashing null object
3747 obj = null_check_receiver();
3748 if (stopped()) return true; // unconditionally null
3749 result_reg->init_req(_null_path, top());
3750 result_val->init_req(_null_path, top());
3751 } else {
3752 // Do a null check, and return zero if null.
3753 // System.identityHashCode(null) == 0
3754 obj = argument(0);
3755 Node* null_ctl = top();
3756 obj = null_check_oop(obj, &null_ctl);
3757 result_reg->init_req(_null_path, null_ctl);
3758 result_val->init_req(_null_path, _gvn.intcon(0));
3759 }
3761 // Unconditionally null? Then return right away.
3762 if (stopped()) {
3763 set_control( result_reg->in(_null_path));
3764 if (!stopped())
3765 set_result(result_val->in(_null_path));
3766 return true;
3767 }
3769 // After null check, get the object's klass.
3770 Node* obj_klass = load_object_klass(obj);
3772 // This call may be virtual (invokevirtual) or bound (invokespecial).
3773 // For each case we generate slightly different code.
3775 // We only go to the fast case code if we pass a number of guards. The
3776 // paths which do not pass are accumulated in the slow_region.
3777 RegionNode* slow_region = new (C) RegionNode(1);
3778 record_for_igvn(slow_region);
3780 // If this is a virtual call, we generate a funny guard. We pull out
3781 // the vtable entry corresponding to hashCode() from the target object.
3782 // If the target method which we are calling happens to be the native
3783 // Object hashCode() method, we pass the guard. We do not need this
3784 // guard for non-virtual calls -- the caller is known to be the native
3785 // Object hashCode().
3786 if (is_virtual) {
3787 generate_virtual_guard(obj_klass, slow_region);
3788 }
3790 // Get the header out of the object, use LoadMarkNode when available
3791 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3792 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3794 // Test the header to see if it is unlocked.
3795 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3796 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
3797 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3798 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
3799 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
3801 generate_slow_guard(test_unlocked, slow_region);
3803 // Get the hash value and check to see that it has been properly assigned.
3804 // We depend on hash_mask being at most 32 bits and avoid the use of
3805 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3806 // vm: see markOop.hpp.
3807 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3808 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3809 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
3810 // This hack lets the hash bits live anywhere in the mark object now, as long
3811 // as the shift drops the relevant bits into the low 32 bits. Note that
3812 // Java spec says that HashCode is an int so there's no point in capturing
3813 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3814 hshifted_header = ConvX2I(hshifted_header);
3815 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
3817 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3818 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
3819 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
3821 generate_slow_guard(test_assigned, slow_region);
3823 Node* init_mem = reset_memory();
3824 // fill in the rest of the null path:
3825 result_io ->init_req(_null_path, i_o());
3826 result_mem->init_req(_null_path, init_mem);
3828 result_val->init_req(_fast_path, hash_val);
3829 result_reg->init_req(_fast_path, control());
3830 result_io ->init_req(_fast_path, i_o());
3831 result_mem->init_req(_fast_path, init_mem);
3833 // Generate code for the slow case. We make a call to hashCode().
3834 set_control(_gvn.transform(slow_region));
3835 if (!stopped()) {
3836 // No need for PreserveJVMState, because we're using up the present state.
3837 set_all_memory(init_mem);
3838 vmIntrinsics::ID hashCode_id = is_static ? vmIntrinsics::_identityHashCode : vmIntrinsics::_hashCode;
3839 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3840 Node* slow_result = set_results_for_java_call(slow_call);
3841 // this->control() comes from set_results_for_java_call
3842 result_reg->init_req(_slow_path, control());
3843 result_val->init_req(_slow_path, slow_result);
3844 result_io ->set_req(_slow_path, i_o());
3845 result_mem ->set_req(_slow_path, reset_memory());
3846 }
3848 // Return the combined state.
3849 set_i_o( _gvn.transform(result_io) );
3850 set_all_memory( _gvn.transform(result_mem) );
3852 set_result(result_reg, result_val);
3853 return true;
3854 }
3856 //---------------------------inline_native_getClass----------------------------
3857 // public final native Class<?> java.lang.Object.getClass();
3858 //
3859 // Build special case code for calls to getClass on an object.
3860 bool LibraryCallKit::inline_native_getClass() {
3861 Node* obj = null_check_receiver();
3862 if (stopped()) return true;
3863 set_result(load_mirror_from_klass(load_object_klass(obj)));
3864 return true;
3865 }
3867 //-----------------inline_native_Reflection_getCallerClass---------------------
3868 // public static native Class<?> sun.reflect.Reflection.getCallerClass(int realFramesToSkip);
3869 //
3870 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3871 //
3872 // NOTE that this code must perform the same logic as
3873 // vframeStream::security_get_caller_frame in that it must skip
3874 // Method.invoke() and auxiliary frames.
3875 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3876 #ifndef PRODUCT
3877 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3878 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3879 }
3880 #endif
3882 Node* caller_depth_node = argument(0);
3884 // The depth value must be a constant in order for the runtime call
3885 // to be eliminated.
3886 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3887 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3888 #ifndef PRODUCT
3889 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3890 tty->print_cr(" Bailing out because caller depth was not a constant");
3891 }
3892 #endif
3893 return false;
3894 }
3895 // Note that the JVM state at this point does not include the
3896 // getCallerClass() frame which we are trying to inline. The
3897 // semantics of getCallerClass(), however, are that the "first"
3898 // frame is the getCallerClass() frame, so we subtract one from the
3899 // requested depth before continuing. We don't inline requests of
3900 // getCallerClass(0).
3901 int caller_depth = caller_depth_type->get_con() - 1;
3902 if (caller_depth < 0) {
3903 #ifndef PRODUCT
3904 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3905 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3906 }
3907 #endif
3908 return false;
3909 }
3911 if (!jvms()->has_method()) {
3912 #ifndef PRODUCT
3913 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3914 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3915 }
3916 #endif
3917 return false;
3918 }
3919 int _depth = jvms()->depth(); // cache call chain depth
3921 // Walk back up the JVM state to find the caller at the required
3922 // depth. NOTE that this code must perform the same logic as
3923 // vframeStream::security_get_caller_frame in that it must skip
3924 // Method.invoke() and auxiliary frames. Note also that depth is
3925 // 1-based (1 is the bottom of the inlining).
3926 int inlining_depth = _depth;
3927 JVMState* caller_jvms = NULL;
3929 if (inlining_depth > 0) {
3930 caller_jvms = jvms();
3931 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3932 do {
3933 // The following if-tests should be performed in this order
3934 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3935 // Skip a Method.invoke() or auxiliary frame
3936 } else if (caller_depth > 0) {
3937 // Skip real frame
3938 --caller_depth;
3939 } else {
3940 // We're done: reached desired caller after skipping.
3941 break;
3942 }
3943 caller_jvms = caller_jvms->caller();
3944 --inlining_depth;
3945 } while (inlining_depth > 0);
3946 }
3948 if (inlining_depth == 0) {
3949 #ifndef PRODUCT
3950 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3951 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3952 tty->print_cr(" JVM state at this point:");
3953 for (int i = _depth; i >= 1; i--) {
3954 ciMethod* m = jvms()->of_depth(i)->method();
3955 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3956 }
3957 }
3958 #endif
3959 return false; // Reached end of inlining
3960 }
3962 // Acquire method holder as java.lang.Class
3963 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3964 ciInstance* caller_mirror = caller_klass->java_mirror();
3966 // Push this as a constant
3967 set_result(makecon(TypeInstPtr::make(caller_mirror)));
3969 #ifndef PRODUCT
3970 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3971 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);
3972 tty->print_cr(" JVM state at this point:");
3973 for (int i = _depth; i >= 1; i--) {
3974 ciMethod* m = jvms()->of_depth(i)->method();
3975 tty->print_cr(" %d) %s.%s", i, m->holder()->name()->as_utf8(), m->name()->as_utf8());
3976 }
3977 }
3978 #endif
3979 return true;
3980 }
3982 // Helper routine for above
3983 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3984 ciMethod* method = jvms->method();
3986 // Is this the Method.invoke method itself?
3987 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3988 return true;
3990 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3991 ciKlass* k = method->holder();
3992 if (k->is_instance_klass()) {
3993 ciInstanceKlass* ik = k->as_instance_klass();
3994 for (; ik != NULL; ik = ik->super()) {
3995 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3996 ik == env()->find_system_klass(ik->name())) {
3997 return true;
3998 }
3999 }
4000 }
4001 else if (method->is_method_handle_intrinsic() ||
4002 method->is_compiled_lambda_form()) {
4003 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4004 return true;
4005 }
4007 return false;
4008 }
4010 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4011 Node* arg = argument(0);
4012 Node* result;
4014 switch (id) {
4015 case vmIntrinsics::_floatToRawIntBits: result = new (C) MoveF2INode(arg); break;
4016 case vmIntrinsics::_intBitsToFloat: result = new (C) MoveI2FNode(arg); break;
4017 case vmIntrinsics::_doubleToRawLongBits: result = new (C) MoveD2LNode(arg); break;
4018 case vmIntrinsics::_longBitsToDouble: result = new (C) MoveL2DNode(arg); break;
4020 case vmIntrinsics::_doubleToLongBits: {
4021 // two paths (plus control) merge in a wood
4022 RegionNode *r = new (C) RegionNode(3);
4023 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4025 Node *cmpisnan = _gvn.transform(new (C) CmpDNode(arg, arg));
4026 // Build the boolean node
4027 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4029 // Branch either way.
4030 // NaN case is less traveled, which makes all the difference.
4031 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4032 Node *opt_isnan = _gvn.transform(ifisnan);
4033 assert( opt_isnan->is_If(), "Expect an IfNode");
4034 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4035 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4037 set_control(iftrue);
4039 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4040 Node *slow_result = longcon(nan_bits); // return NaN
4041 phi->init_req(1, _gvn.transform( slow_result ));
4042 r->init_req(1, iftrue);
4044 // Else fall through
4045 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4046 set_control(iffalse);
4048 phi->init_req(2, _gvn.transform(new (C) MoveD2LNode(arg)));
4049 r->init_req(2, iffalse);
4051 // Post merge
4052 set_control(_gvn.transform(r));
4053 record_for_igvn(r);
4055 C->set_has_split_ifs(true); // Has chance for split-if optimization
4056 result = phi;
4057 assert(result->bottom_type()->isa_long(), "must be");
4058 break;
4059 }
4061 case vmIntrinsics::_floatToIntBits: {
4062 // two paths (plus control) merge in a wood
4063 RegionNode *r = new (C) RegionNode(3);
4064 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4066 Node *cmpisnan = _gvn.transform(new (C) CmpFNode(arg, arg));
4067 // Build the boolean node
4068 Node *bolisnan = _gvn.transform(new (C) BoolNode(cmpisnan, BoolTest::ne));
4070 // Branch either way.
4071 // NaN case is less traveled, which makes all the difference.
4072 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4073 Node *opt_isnan = _gvn.transform(ifisnan);
4074 assert( opt_isnan->is_If(), "Expect an IfNode");
4075 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4076 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4078 set_control(iftrue);
4080 static const jint nan_bits = 0x7fc00000;
4081 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4082 phi->init_req(1, _gvn.transform( slow_result ));
4083 r->init_req(1, iftrue);
4085 // Else fall through
4086 Node *iffalse = _gvn.transform(new (C) IfFalseNode(opt_ifisnan));
4087 set_control(iffalse);
4089 phi->init_req(2, _gvn.transform(new (C) MoveF2INode(arg)));
4090 r->init_req(2, iffalse);
4092 // Post merge
4093 set_control(_gvn.transform(r));
4094 record_for_igvn(r);
4096 C->set_has_split_ifs(true); // Has chance for split-if optimization
4097 result = phi;
4098 assert(result->bottom_type()->isa_int(), "must be");
4099 break;
4100 }
4102 default:
4103 fatal_unexpected_iid(id);
4104 break;
4105 }
4106 set_result(_gvn.transform(result));
4107 return true;
4108 }
4110 #ifdef _LP64
4111 #define XTOP ,top() /*additional argument*/
4112 #else //_LP64
4113 #define XTOP /*no additional argument*/
4114 #endif //_LP64
4116 //----------------------inline_unsafe_copyMemory-------------------------
4117 // public native void sun.misc.Unsafe.copyMemory(Object srcBase, long srcOffset, Object destBase, long destOffset, long bytes);
4118 bool LibraryCallKit::inline_unsafe_copyMemory() {
4119 if (callee()->is_static()) return false; // caller must have the capability!
4120 null_check_receiver(); // null-check receiver
4121 if (stopped()) return true;
4123 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4125 Node* src_ptr = argument(1); // type: oop
4126 Node* src_off = ConvL2X(argument(2)); // type: long
4127 Node* dst_ptr = argument(4); // type: oop
4128 Node* dst_off = ConvL2X(argument(5)); // type: long
4129 Node* size = ConvL2X(argument(7)); // type: long
4131 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4132 "fieldOffset must be byte-scaled");
4134 Node* src = make_unsafe_address(src_ptr, src_off);
4135 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4137 // Conservatively insert a memory barrier on all memory slices.
4138 // Do not let writes of the copy source or destination float below the copy.
4139 insert_mem_bar(Op_MemBarCPUOrder);
4141 // Call it. Note that the length argument is not scaled.
4142 make_runtime_call(RC_LEAF|RC_NO_FP,
4143 OptoRuntime::fast_arraycopy_Type(),
4144 StubRoutines::unsafe_arraycopy(),
4145 "unsafe_arraycopy",
4146 TypeRawPtr::BOTTOM,
4147 src, dst, size XTOP);
4149 // Do not let reads of the copy destination float above the copy.
4150 insert_mem_bar(Op_MemBarCPUOrder);
4152 return true;
4153 }
4155 //------------------------clone_coping-----------------------------------
4156 // Helper function for inline_native_clone.
4157 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4158 assert(obj_size != NULL, "");
4159 Node* raw_obj = alloc_obj->in(1);
4160 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4162 AllocateNode* alloc = NULL;
4163 if (ReduceBulkZeroing) {
4164 // We will be completely responsible for initializing this object -
4165 // mark Initialize node as complete.
4166 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4167 // The object was just allocated - there should be no any stores!
4168 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4169 // Mark as complete_with_arraycopy so that on AllocateNode
4170 // expansion, we know this AllocateNode is initialized by an array
4171 // copy and a StoreStore barrier exists after the array copy.
4172 alloc->initialization()->set_complete_with_arraycopy();
4173 }
4175 // Copy the fastest available way.
4176 // TODO: generate fields copies for small objects instead.
4177 Node* src = obj;
4178 Node* dest = alloc_obj;
4179 Node* size = _gvn.transform(obj_size);
4181 // Exclude the header but include array length to copy by 8 bytes words.
4182 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4183 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4184 instanceOopDesc::base_offset_in_bytes();
4185 // base_off:
4186 // 8 - 32-bit VM
4187 // 12 - 64-bit VM, compressed klass
4188 // 16 - 64-bit VM, normal klass
4189 if (base_off % BytesPerLong != 0) {
4190 assert(UseCompressedKlassPointers, "");
4191 if (is_array) {
4192 // Exclude length to copy by 8 bytes words.
4193 base_off += sizeof(int);
4194 } else {
4195 // Include klass to copy by 8 bytes words.
4196 base_off = instanceOopDesc::klass_offset_in_bytes();
4197 }
4198 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4199 }
4200 src = basic_plus_adr(src, base_off);
4201 dest = basic_plus_adr(dest, base_off);
4203 // Compute the length also, if needed:
4204 Node* countx = size;
4205 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4206 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4208 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4209 bool disjoint_bases = true;
4210 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4211 src, NULL, dest, NULL, countx,
4212 /*dest_uninitialized*/true);
4214 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4215 if (card_mark) {
4216 assert(!is_array, "");
4217 // Put in store barrier for any and all oops we are sticking
4218 // into this object. (We could avoid this if we could prove
4219 // that the object type contains no oop fields at all.)
4220 Node* no_particular_value = NULL;
4221 Node* no_particular_field = NULL;
4222 int raw_adr_idx = Compile::AliasIdxRaw;
4223 post_barrier(control(),
4224 memory(raw_adr_type),
4225 alloc_obj,
4226 no_particular_field,
4227 raw_adr_idx,
4228 no_particular_value,
4229 T_OBJECT,
4230 false);
4231 }
4233 // Do not let reads from the cloned object float above the arraycopy.
4234 if (alloc != NULL) {
4235 // Do not let stores that initialize this object be reordered with
4236 // a subsequent store that would make this object accessible by
4237 // other threads.
4238 // Record what AllocateNode this StoreStore protects so that
4239 // escape analysis can go from the MemBarStoreStoreNode to the
4240 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4241 // based on the escape status of the AllocateNode.
4242 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4243 } else {
4244 insert_mem_bar(Op_MemBarCPUOrder);
4245 }
4246 }
4248 //------------------------inline_native_clone----------------------------
4249 // protected native Object java.lang.Object.clone();
4250 //
4251 // Here are the simple edge cases:
4252 // null receiver => normal trap
4253 // virtual and clone was overridden => slow path to out-of-line clone
4254 // not cloneable or finalizer => slow path to out-of-line Object.clone
4255 //
4256 // The general case has two steps, allocation and copying.
4257 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4258 //
4259 // Copying also has two cases, oop arrays and everything else.
4260 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4261 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4262 //
4263 // These steps fold up nicely if and when the cloned object's klass
4264 // can be sharply typed as an object array, a type array, or an instance.
4265 //
4266 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4267 PhiNode* result_val;
4269 // Set the reexecute bit for the interpreter to reexecute
4270 // the bytecode that invokes Object.clone if deoptimization happens.
4271 { PreserveReexecuteState preexecs(this);
4272 jvms()->set_should_reexecute(true);
4274 Node* obj = null_check_receiver();
4275 if (stopped()) return true;
4277 Node* obj_klass = load_object_klass(obj);
4278 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4279 const TypeOopPtr* toop = ((tklass != NULL)
4280 ? tklass->as_instance_type()
4281 : TypeInstPtr::NOTNULL);
4283 // Conservatively insert a memory barrier on all memory slices.
4284 // Do not let writes into the original float below the clone.
4285 insert_mem_bar(Op_MemBarCPUOrder);
4287 // paths into result_reg:
4288 enum {
4289 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4290 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4291 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4292 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4293 PATH_LIMIT
4294 };
4295 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4296 result_val = new(C) PhiNode(result_reg,
4297 TypeInstPtr::NOTNULL);
4298 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4299 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4300 TypePtr::BOTTOM);
4301 record_for_igvn(result_reg);
4303 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4304 int raw_adr_idx = Compile::AliasIdxRaw;
4306 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4307 if (array_ctl != NULL) {
4308 // It's an array.
4309 PreserveJVMState pjvms(this);
4310 set_control(array_ctl);
4311 Node* obj_length = load_array_length(obj);
4312 Node* obj_size = NULL;
4313 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size); // no arguments to push
4315 if (!use_ReduceInitialCardMarks()) {
4316 // If it is an oop array, it requires very special treatment,
4317 // because card marking is required on each card of the array.
4318 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4319 if (is_obja != NULL) {
4320 PreserveJVMState pjvms2(this);
4321 set_control(is_obja);
4322 // Generate a direct call to the right arraycopy function(s).
4323 bool disjoint_bases = true;
4324 bool length_never_negative = true;
4325 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4326 obj, intcon(0), alloc_obj, intcon(0),
4327 obj_length,
4328 disjoint_bases, length_never_negative);
4329 result_reg->init_req(_objArray_path, control());
4330 result_val->init_req(_objArray_path, alloc_obj);
4331 result_i_o ->set_req(_objArray_path, i_o());
4332 result_mem ->set_req(_objArray_path, reset_memory());
4333 }
4334 }
4335 // Otherwise, there are no card marks to worry about.
4336 // (We can dispense with card marks if we know the allocation
4337 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4338 // causes the non-eden paths to take compensating steps to
4339 // simulate a fresh allocation, so that no further
4340 // card marks are required in compiled code to initialize
4341 // the object.)
4343 if (!stopped()) {
4344 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4346 // Present the results of the copy.
4347 result_reg->init_req(_array_path, control());
4348 result_val->init_req(_array_path, alloc_obj);
4349 result_i_o ->set_req(_array_path, i_o());
4350 result_mem ->set_req(_array_path, reset_memory());
4351 }
4352 }
4354 // We only go to the instance fast case code if we pass a number of guards.
4355 // The paths which do not pass are accumulated in the slow_region.
4356 RegionNode* slow_region = new (C) RegionNode(1);
4357 record_for_igvn(slow_region);
4358 if (!stopped()) {
4359 // It's an instance (we did array above). Make the slow-path tests.
4360 // If this is a virtual call, we generate a funny guard. We grab
4361 // the vtable entry corresponding to clone() from the target object.
4362 // If the target method which we are calling happens to be the
4363 // Object clone() method, we pass the guard. We do not need this
4364 // guard for non-virtual calls; the caller is known to be the native
4365 // Object clone().
4366 if (is_virtual) {
4367 generate_virtual_guard(obj_klass, slow_region);
4368 }
4370 // The object must be cloneable and must not have a finalizer.
4371 // Both of these conditions may be checked in a single test.
4372 // We could optimize the cloneable test further, but we don't care.
4373 generate_access_flags_guard(obj_klass,
4374 // Test both conditions:
4375 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4376 // Must be cloneable but not finalizer:
4377 JVM_ACC_IS_CLONEABLE,
4378 slow_region);
4379 }
4381 if (!stopped()) {
4382 // It's an instance, and it passed the slow-path tests.
4383 PreserveJVMState pjvms(this);
4384 Node* obj_size = NULL;
4385 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4387 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4389 // Present the results of the slow call.
4390 result_reg->init_req(_instance_path, control());
4391 result_val->init_req(_instance_path, alloc_obj);
4392 result_i_o ->set_req(_instance_path, i_o());
4393 result_mem ->set_req(_instance_path, reset_memory());
4394 }
4396 // Generate code for the slow case. We make a call to clone().
4397 set_control(_gvn.transform(slow_region));
4398 if (!stopped()) {
4399 PreserveJVMState pjvms(this);
4400 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4401 Node* slow_result = set_results_for_java_call(slow_call);
4402 // this->control() comes from set_results_for_java_call
4403 result_reg->init_req(_slow_path, control());
4404 result_val->init_req(_slow_path, slow_result);
4405 result_i_o ->set_req(_slow_path, i_o());
4406 result_mem ->set_req(_slow_path, reset_memory());
4407 }
4409 // Return the combined state.
4410 set_control( _gvn.transform(result_reg) );
4411 set_i_o( _gvn.transform(result_i_o) );
4412 set_all_memory( _gvn.transform(result_mem) );
4413 } // original reexecute is set back here
4415 set_result(_gvn.transform(result_val));
4416 return true;
4417 }
4419 //------------------------------basictype2arraycopy----------------------------
4420 address LibraryCallKit::basictype2arraycopy(BasicType t,
4421 Node* src_offset,
4422 Node* dest_offset,
4423 bool disjoint_bases,
4424 const char* &name,
4425 bool dest_uninitialized) {
4426 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4427 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4429 bool aligned = false;
4430 bool disjoint = disjoint_bases;
4432 // if the offsets are the same, we can treat the memory regions as
4433 // disjoint, because either the memory regions are in different arrays,
4434 // or they are identical (which we can treat as disjoint.) We can also
4435 // treat a copy with a destination index less that the source index
4436 // as disjoint since a low->high copy will work correctly in this case.
4437 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4438 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4439 // both indices are constants
4440 int s_offs = src_offset_inttype->get_con();
4441 int d_offs = dest_offset_inttype->get_con();
4442 int element_size = type2aelembytes(t);
4443 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4444 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4445 if (s_offs >= d_offs) disjoint = true;
4446 } else if (src_offset == dest_offset && src_offset != NULL) {
4447 // This can occur if the offsets are identical non-constants.
4448 disjoint = true;
4449 }
4451 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4452 }
4455 //------------------------------inline_arraycopy-----------------------
4456 // public static native void java.lang.System.arraycopy(Object src, int srcPos,
4457 // Object dest, int destPos,
4458 // int length);
4459 bool LibraryCallKit::inline_arraycopy() {
4460 // Get the arguments.
4461 Node* src = argument(0); // type: oop
4462 Node* src_offset = argument(1); // type: int
4463 Node* dest = argument(2); // type: oop
4464 Node* dest_offset = argument(3); // type: int
4465 Node* length = argument(4); // type: int
4467 // Compile time checks. If any of these checks cannot be verified at compile time,
4468 // we do not make a fast path for this call. Instead, we let the call remain as it
4469 // is. The checks we choose to mandate at compile time are:
4470 //
4471 // (1) src and dest are arrays.
4472 const Type* src_type = src->Value(&_gvn);
4473 const Type* dest_type = dest->Value(&_gvn);
4474 const TypeAryPtr* top_src = src_type->isa_aryptr();
4475 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4476 if (top_src == NULL || top_src->klass() == NULL ||
4477 top_dest == NULL || top_dest->klass() == NULL) {
4478 // Conservatively insert a memory barrier on all memory slices.
4479 // Do not let writes into the source float below the arraycopy.
4480 insert_mem_bar(Op_MemBarCPUOrder);
4482 // Call StubRoutines::generic_arraycopy stub.
4483 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4484 src, src_offset, dest, dest_offset, length);
4486 // Do not let reads from the destination float above the arraycopy.
4487 // Since we cannot type the arrays, we don't know which slices
4488 // might be affected. We could restrict this barrier only to those
4489 // memory slices which pertain to array elements--but don't bother.
4490 if (!InsertMemBarAfterArraycopy)
4491 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4492 insert_mem_bar(Op_MemBarCPUOrder);
4493 return true;
4494 }
4496 // (2) src and dest arrays must have elements of the same BasicType
4497 // Figure out the size and type of the elements we will be copying.
4498 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4499 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4500 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4501 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4503 if (src_elem != dest_elem || dest_elem == T_VOID) {
4504 // The component types are not the same or are not recognized. Punt.
4505 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4506 generate_slow_arraycopy(TypePtr::BOTTOM,
4507 src, src_offset, dest, dest_offset, length,
4508 /*dest_uninitialized*/false);
4509 return true;
4510 }
4512 //---------------------------------------------------------------------------
4513 // We will make a fast path for this call to arraycopy.
4515 // We have the following tests left to perform:
4516 //
4517 // (3) src and dest must not be null.
4518 // (4) src_offset must not be negative.
4519 // (5) dest_offset must not be negative.
4520 // (6) length must not be negative.
4521 // (7) src_offset + length must not exceed length of src.
4522 // (8) dest_offset + length must not exceed length of dest.
4523 // (9) each element of an oop array must be assignable
4525 RegionNode* slow_region = new (C) RegionNode(1);
4526 record_for_igvn(slow_region);
4528 // (3) operands must not be null
4529 // We currently perform our null checks with the null_check routine.
4530 // This means that the null exceptions will be reported in the caller
4531 // rather than (correctly) reported inside of the native arraycopy call.
4532 // This should be corrected, given time. We do our null check with the
4533 // stack pointer restored.
4534 src = null_check(src, T_ARRAY);
4535 dest = null_check(dest, T_ARRAY);
4537 // (4) src_offset must not be negative.
4538 generate_negative_guard(src_offset, slow_region);
4540 // (5) dest_offset must not be negative.
4541 generate_negative_guard(dest_offset, slow_region);
4543 // (6) length must not be negative (moved to generate_arraycopy()).
4544 // generate_negative_guard(length, slow_region);
4546 // (7) src_offset + length must not exceed length of src.
4547 generate_limit_guard(src_offset, length,
4548 load_array_length(src),
4549 slow_region);
4551 // (8) dest_offset + length must not exceed length of dest.
4552 generate_limit_guard(dest_offset, length,
4553 load_array_length(dest),
4554 slow_region);
4556 // (9) each element of an oop array must be assignable
4557 // The generate_arraycopy subroutine checks this.
4559 // This is where the memory effects are placed:
4560 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4561 generate_arraycopy(adr_type, dest_elem,
4562 src, src_offset, dest, dest_offset, length,
4563 false, false, slow_region);
4565 return true;
4566 }
4568 //-----------------------------generate_arraycopy----------------------
4569 // Generate an optimized call to arraycopy.
4570 // Caller must guard against non-arrays.
4571 // Caller must determine a common array basic-type for both arrays.
4572 // Caller must validate offsets against array bounds.
4573 // The slow_region has already collected guard failure paths
4574 // (such as out of bounds length or non-conformable array types).
4575 // The generated code has this shape, in general:
4576 //
4577 // if (length == 0) return // via zero_path
4578 // slowval = -1
4579 // if (types unknown) {
4580 // slowval = call generic copy loop
4581 // if (slowval == 0) return // via checked_path
4582 // } else if (indexes in bounds) {
4583 // if ((is object array) && !(array type check)) {
4584 // slowval = call checked copy loop
4585 // if (slowval == 0) return // via checked_path
4586 // } else {
4587 // call bulk copy loop
4588 // return // via fast_path
4589 // }
4590 // }
4591 // // adjust params for remaining work:
4592 // if (slowval != -1) {
4593 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4594 // }
4595 // slow_region:
4596 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4597 // return // via slow_call_path
4598 //
4599 // This routine is used from several intrinsics: System.arraycopy,
4600 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4601 //
4602 void
4603 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4604 BasicType basic_elem_type,
4605 Node* src, Node* src_offset,
4606 Node* dest, Node* dest_offset,
4607 Node* copy_length,
4608 bool disjoint_bases,
4609 bool length_never_negative,
4610 RegionNode* slow_region) {
4612 if (slow_region == NULL) {
4613 slow_region = new(C) RegionNode(1);
4614 record_for_igvn(slow_region);
4615 }
4617 Node* original_dest = dest;
4618 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4619 bool dest_uninitialized = false;
4621 // See if this is the initialization of a newly-allocated array.
4622 // If so, we will take responsibility here for initializing it to zero.
4623 // (Note: Because tightly_coupled_allocation performs checks on the
4624 // out-edges of the dest, we need to avoid making derived pointers
4625 // from it until we have checked its uses.)
4626 if (ReduceBulkZeroing
4627 && !ZeroTLAB // pointless if already zeroed
4628 && basic_elem_type != T_CONFLICT // avoid corner case
4629 && !src->eqv_uncast(dest)
4630 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4631 != NULL)
4632 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4633 && alloc->maybe_set_complete(&_gvn)) {
4634 // "You break it, you buy it."
4635 InitializeNode* init = alloc->initialization();
4636 assert(init->is_complete(), "we just did this");
4637 init->set_complete_with_arraycopy();
4638 assert(dest->is_CheckCastPP(), "sanity");
4639 assert(dest->in(0)->in(0) == init, "dest pinned");
4640 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4641 // From this point on, every exit path is responsible for
4642 // initializing any non-copied parts of the object to zero.
4643 // Also, if this flag is set we make sure that arraycopy interacts properly
4644 // with G1, eliding pre-barriers. See CR 6627983.
4645 dest_uninitialized = true;
4646 } else {
4647 // No zeroing elimination here.
4648 alloc = NULL;
4649 //original_dest = dest;
4650 //dest_uninitialized = false;
4651 }
4653 // Results are placed here:
4654 enum { fast_path = 1, // normal void-returning assembly stub
4655 checked_path = 2, // special assembly stub with cleanup
4656 slow_call_path = 3, // something went wrong; call the VM
4657 zero_path = 4, // bypass when length of copy is zero
4658 bcopy_path = 5, // copy primitive array by 64-bit blocks
4659 PATH_LIMIT = 6
4660 };
4661 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4662 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4663 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4664 record_for_igvn(result_region);
4665 _gvn.set_type_bottom(result_i_o);
4666 _gvn.set_type_bottom(result_memory);
4667 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4669 // The slow_control path:
4670 Node* slow_control;
4671 Node* slow_i_o = i_o();
4672 Node* slow_mem = memory(adr_type);
4673 debug_only(slow_control = (Node*) badAddress);
4675 // Checked control path:
4676 Node* checked_control = top();
4677 Node* checked_mem = NULL;
4678 Node* checked_i_o = NULL;
4679 Node* checked_value = NULL;
4681 if (basic_elem_type == T_CONFLICT) {
4682 assert(!dest_uninitialized, "");
4683 Node* cv = generate_generic_arraycopy(adr_type,
4684 src, src_offset, dest, dest_offset,
4685 copy_length, dest_uninitialized);
4686 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4687 checked_control = control();
4688 checked_i_o = i_o();
4689 checked_mem = memory(adr_type);
4690 checked_value = cv;
4691 set_control(top()); // no fast path
4692 }
4694 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4695 if (not_pos != NULL) {
4696 PreserveJVMState pjvms(this);
4697 set_control(not_pos);
4699 // (6) length must not be negative.
4700 if (!length_never_negative) {
4701 generate_negative_guard(copy_length, slow_region);
4702 }
4704 // copy_length is 0.
4705 if (!stopped() && dest_uninitialized) {
4706 Node* dest_length = alloc->in(AllocateNode::ALength);
4707 if (copy_length->eqv_uncast(dest_length)
4708 || _gvn.find_int_con(dest_length, 1) <= 0) {
4709 // There is no zeroing to do. No need for a secondary raw memory barrier.
4710 } else {
4711 // Clear the whole thing since there are no source elements to copy.
4712 generate_clear_array(adr_type, dest, basic_elem_type,
4713 intcon(0), NULL,
4714 alloc->in(AllocateNode::AllocSize));
4715 // Use a secondary InitializeNode as raw memory barrier.
4716 // Currently it is needed only on this path since other
4717 // paths have stub or runtime calls as raw memory barriers.
4718 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4719 Compile::AliasIdxRaw,
4720 top())->as_Initialize();
4721 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4722 }
4723 }
4725 // Present the results of the fast call.
4726 result_region->init_req(zero_path, control());
4727 result_i_o ->init_req(zero_path, i_o());
4728 result_memory->init_req(zero_path, memory(adr_type));
4729 }
4731 if (!stopped() && dest_uninitialized) {
4732 // We have to initialize the *uncopied* part of the array to zero.
4733 // The copy destination is the slice dest[off..off+len]. The other slices
4734 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4735 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4736 Node* dest_length = alloc->in(AllocateNode::ALength);
4737 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
4738 copy_length) );
4740 // If there is a head section that needs zeroing, do it now.
4741 if (find_int_con(dest_offset, -1) != 0) {
4742 generate_clear_array(adr_type, dest, basic_elem_type,
4743 intcon(0), dest_offset,
4744 NULL);
4745 }
4747 // Next, perform a dynamic check on the tail length.
4748 // It is often zero, and we can win big if we prove this.
4749 // There are two wins: Avoid generating the ClearArray
4750 // with its attendant messy index arithmetic, and upgrade
4751 // the copy to a more hardware-friendly word size of 64 bits.
4752 Node* tail_ctl = NULL;
4753 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4754 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
4755 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
4756 tail_ctl = generate_slow_guard(bol_lt, NULL);
4757 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4758 }
4760 // At this point, let's assume there is no tail.
4761 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4762 // There is no tail. Try an upgrade to a 64-bit copy.
4763 bool didit = false;
4764 { PreserveJVMState pjvms(this);
4765 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4766 src, src_offset, dest, dest_offset,
4767 dest_size, dest_uninitialized);
4768 if (didit) {
4769 // Present the results of the block-copying fast call.
4770 result_region->init_req(bcopy_path, control());
4771 result_i_o ->init_req(bcopy_path, i_o());
4772 result_memory->init_req(bcopy_path, memory(adr_type));
4773 }
4774 }
4775 if (didit)
4776 set_control(top()); // no regular fast path
4777 }
4779 // Clear the tail, if any.
4780 if (tail_ctl != NULL) {
4781 Node* notail_ctl = stopped() ? NULL : control();
4782 set_control(tail_ctl);
4783 if (notail_ctl == NULL) {
4784 generate_clear_array(adr_type, dest, basic_elem_type,
4785 dest_tail, NULL,
4786 dest_size);
4787 } else {
4788 // Make a local merge.
4789 Node* done_ctl = new(C) RegionNode(3);
4790 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
4791 done_ctl->init_req(1, notail_ctl);
4792 done_mem->init_req(1, memory(adr_type));
4793 generate_clear_array(adr_type, dest, basic_elem_type,
4794 dest_tail, NULL,
4795 dest_size);
4796 done_ctl->init_req(2, control());
4797 done_mem->init_req(2, memory(adr_type));
4798 set_control( _gvn.transform(done_ctl) );
4799 set_memory( _gvn.transform(done_mem), adr_type );
4800 }
4801 }
4802 }
4804 BasicType copy_type = basic_elem_type;
4805 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4806 if (!stopped() && copy_type == T_OBJECT) {
4807 // If src and dest have compatible element types, we can copy bits.
4808 // Types S[] and D[] are compatible if D is a supertype of S.
4809 //
4810 // If they are not, we will use checked_oop_disjoint_arraycopy,
4811 // which performs a fast optimistic per-oop check, and backs off
4812 // further to JVM_ArrayCopy on the first per-oop check that fails.
4813 // (Actually, we don't move raw bits only; the GC requires card marks.)
4815 // Get the Klass* for both src and dest
4816 Node* src_klass = load_object_klass(src);
4817 Node* dest_klass = load_object_klass(dest);
4819 // Generate the subtype check.
4820 // This might fold up statically, or then again it might not.
4821 //
4822 // Non-static example: Copying List<String>.elements to a new String[].
4823 // The backing store for a List<String> is always an Object[],
4824 // but its elements are always type String, if the generic types
4825 // are correct at the source level.
4826 //
4827 // Test S[] against D[], not S against D, because (probably)
4828 // the secondary supertype cache is less busy for S[] than S.
4829 // This usually only matters when D is an interface.
4830 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4831 // Plug failing path into checked_oop_disjoint_arraycopy
4832 if (not_subtype_ctrl != top()) {
4833 PreserveJVMState pjvms(this);
4834 set_control(not_subtype_ctrl);
4835 // (At this point we can assume disjoint_bases, since types differ.)
4836 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
4837 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4838 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4839 Node* dest_elem_klass = _gvn.transform(n1);
4840 Node* cv = generate_checkcast_arraycopy(adr_type,
4841 dest_elem_klass,
4842 src, src_offset, dest, dest_offset,
4843 ConvI2X(copy_length), dest_uninitialized);
4844 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4845 checked_control = control();
4846 checked_i_o = i_o();
4847 checked_mem = memory(adr_type);
4848 checked_value = cv;
4849 }
4850 // At this point we know we do not need type checks on oop stores.
4852 // Let's see if we need card marks:
4853 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4854 // If we do not need card marks, copy using the jint or jlong stub.
4855 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4856 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4857 "sizes agree");
4858 }
4859 }
4861 if (!stopped()) {
4862 // Generate the fast path, if possible.
4863 PreserveJVMState pjvms(this);
4864 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4865 src, src_offset, dest, dest_offset,
4866 ConvI2X(copy_length), dest_uninitialized);
4868 // Present the results of the fast call.
4869 result_region->init_req(fast_path, control());
4870 result_i_o ->init_req(fast_path, i_o());
4871 result_memory->init_req(fast_path, memory(adr_type));
4872 }
4874 // Here are all the slow paths up to this point, in one bundle:
4875 slow_control = top();
4876 if (slow_region != NULL)
4877 slow_control = _gvn.transform(slow_region);
4878 DEBUG_ONLY(slow_region = (RegionNode*)badAddress);
4880 set_control(checked_control);
4881 if (!stopped()) {
4882 // Clean up after the checked call.
4883 // The returned value is either 0 or -1^K,
4884 // where K = number of partially transferred array elements.
4885 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
4886 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
4887 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4889 // If it is 0, we are done, so transfer to the end.
4890 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
4891 result_region->init_req(checked_path, checks_done);
4892 result_i_o ->init_req(checked_path, checked_i_o);
4893 result_memory->init_req(checked_path, checked_mem);
4895 // If it is not zero, merge into the slow call.
4896 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
4897 RegionNode* slow_reg2 = new(C) RegionNode(3);
4898 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
4899 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4900 record_for_igvn(slow_reg2);
4901 slow_reg2 ->init_req(1, slow_control);
4902 slow_i_o2 ->init_req(1, slow_i_o);
4903 slow_mem2 ->init_req(1, slow_mem);
4904 slow_reg2 ->init_req(2, control());
4905 slow_i_o2 ->init_req(2, checked_i_o);
4906 slow_mem2 ->init_req(2, checked_mem);
4908 slow_control = _gvn.transform(slow_reg2);
4909 slow_i_o = _gvn.transform(slow_i_o2);
4910 slow_mem = _gvn.transform(slow_mem2);
4912 if (alloc != NULL) {
4913 // We'll restart from the very beginning, after zeroing the whole thing.
4914 // This can cause double writes, but that's OK since dest is brand new.
4915 // So we ignore the low 31 bits of the value returned from the stub.
4916 } else {
4917 // We must continue the copy exactly where it failed, or else
4918 // another thread might see the wrong number of writes to dest.
4919 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
4920 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
4921 slow_offset->init_req(1, intcon(0));
4922 slow_offset->init_req(2, checked_offset);
4923 slow_offset = _gvn.transform(slow_offset);
4925 // Adjust the arguments by the conditionally incoming offset.
4926 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
4927 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
4928 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
4930 // Tweak the node variables to adjust the code produced below:
4931 src_offset = src_off_plus;
4932 dest_offset = dest_off_plus;
4933 copy_length = length_minus;
4934 }
4935 }
4937 set_control(slow_control);
4938 if (!stopped()) {
4939 // Generate the slow path, if needed.
4940 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4942 set_memory(slow_mem, adr_type);
4943 set_i_o(slow_i_o);
4945 if (dest_uninitialized) {
4946 generate_clear_array(adr_type, dest, basic_elem_type,
4947 intcon(0), NULL,
4948 alloc->in(AllocateNode::AllocSize));
4949 }
4951 generate_slow_arraycopy(adr_type,
4952 src, src_offset, dest, dest_offset,
4953 copy_length, /*dest_uninitialized*/false);
4955 result_region->init_req(slow_call_path, control());
4956 result_i_o ->init_req(slow_call_path, i_o());
4957 result_memory->init_req(slow_call_path, memory(adr_type));
4958 }
4960 // Remove unused edges.
4961 for (uint i = 1; i < result_region->req(); i++) {
4962 if (result_region->in(i) == NULL)
4963 result_region->init_req(i, top());
4964 }
4966 // Finished; return the combined state.
4967 set_control( _gvn.transform(result_region) );
4968 set_i_o( _gvn.transform(result_i_o) );
4969 set_memory( _gvn.transform(result_memory), adr_type );
4971 // The memory edges above are precise in order to model effects around
4972 // array copies accurately to allow value numbering of field loads around
4973 // arraycopy. Such field loads, both before and after, are common in Java
4974 // collections and similar classes involving header/array data structures.
4975 //
4976 // But with low number of register or when some registers are used or killed
4977 // by arraycopy calls it causes registers spilling on stack. See 6544710.
4978 // The next memory barrier is added to avoid it. If the arraycopy can be
4979 // optimized away (which it can, sometimes) then we can manually remove
4980 // the membar also.
4981 //
4982 // Do not let reads from the cloned object float above the arraycopy.
4983 if (alloc != NULL) {
4984 // Do not let stores that initialize this object be reordered with
4985 // a subsequent store that would make this object accessible by
4986 // other threads.
4987 // Record what AllocateNode this StoreStore protects so that
4988 // escape analysis can go from the MemBarStoreStoreNode to the
4989 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4990 // based on the escape status of the AllocateNode.
4991 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4992 } else if (InsertMemBarAfterArraycopy)
4993 insert_mem_bar(Op_MemBarCPUOrder);
4994 }
4997 // Helper function which determines if an arraycopy immediately follows
4998 // an allocation, with no intervening tests or other escapes for the object.
4999 AllocateArrayNode*
5000 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5001 RegionNode* slow_region) {
5002 if (stopped()) return NULL; // no fast path
5003 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5005 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5006 if (alloc == NULL) return NULL;
5008 Node* rawmem = memory(Compile::AliasIdxRaw);
5009 // Is the allocation's memory state untouched?
5010 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5011 // Bail out if there have been raw-memory effects since the allocation.
5012 // (Example: There might have been a call or safepoint.)
5013 return NULL;
5014 }
5015 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5016 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5017 return NULL;
5018 }
5020 // There must be no unexpected observers of this allocation.
5021 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5022 Node* obs = ptr->fast_out(i);
5023 if (obs != this->map()) {
5024 return NULL;
5025 }
5026 }
5028 // This arraycopy must unconditionally follow the allocation of the ptr.
5029 Node* alloc_ctl = ptr->in(0);
5030 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5032 Node* ctl = control();
5033 while (ctl != alloc_ctl) {
5034 // There may be guards which feed into the slow_region.
5035 // Any other control flow means that we might not get a chance
5036 // to finish initializing the allocated object.
5037 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5038 IfNode* iff = ctl->in(0)->as_If();
5039 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5040 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5041 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5042 ctl = iff->in(0); // This test feeds the known slow_region.
5043 continue;
5044 }
5045 // One more try: Various low-level checks bottom out in
5046 // uncommon traps. If the debug-info of the trap omits
5047 // any reference to the allocation, as we've already
5048 // observed, then there can be no objection to the trap.
5049 bool found_trap = false;
5050 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5051 Node* obs = not_ctl->fast_out(j);
5052 if (obs->in(0) == not_ctl && obs->is_Call() &&
5053 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5054 found_trap = true; break;
5055 }
5056 }
5057 if (found_trap) {
5058 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5059 continue;
5060 }
5061 }
5062 return NULL;
5063 }
5065 // If we get this far, we have an allocation which immediately
5066 // precedes the arraycopy, and we can take over zeroing the new object.
5067 // The arraycopy will finish the initialization, and provide
5068 // a new control state to which we will anchor the destination pointer.
5070 return alloc;
5071 }
5073 // Helper for initialization of arrays, creating a ClearArray.
5074 // It writes zero bits in [start..end), within the body of an array object.
5075 // The memory effects are all chained onto the 'adr_type' alias category.
5076 //
5077 // Since the object is otherwise uninitialized, we are free
5078 // to put a little "slop" around the edges of the cleared area,
5079 // as long as it does not go back into the array's header,
5080 // or beyond the array end within the heap.
5081 //
5082 // The lower edge can be rounded down to the nearest jint and the
5083 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5084 //
5085 // Arguments:
5086 // adr_type memory slice where writes are generated
5087 // dest oop of the destination array
5088 // basic_elem_type element type of the destination
5089 // slice_idx array index of first element to store
5090 // slice_len number of elements to store (or NULL)
5091 // dest_size total size in bytes of the array object
5092 //
5093 // Exactly one of slice_len or dest_size must be non-NULL.
5094 // If dest_size is non-NULL, zeroing extends to the end of the object.
5095 // If slice_len is non-NULL, the slice_idx value must be a constant.
5096 void
5097 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5098 Node* dest,
5099 BasicType basic_elem_type,
5100 Node* slice_idx,
5101 Node* slice_len,
5102 Node* dest_size) {
5103 // one or the other but not both of slice_len and dest_size:
5104 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5105 if (slice_len == NULL) slice_len = top();
5106 if (dest_size == NULL) dest_size = top();
5108 // operate on this memory slice:
5109 Node* mem = memory(adr_type); // memory slice to operate on
5111 // scaling and rounding of indexes:
5112 int scale = exact_log2(type2aelembytes(basic_elem_type));
5113 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5114 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5115 int bump_bit = (-1 << scale) & BytesPerInt;
5117 // determine constant starts and ends
5118 const intptr_t BIG_NEG = -128;
5119 assert(BIG_NEG + 2*abase < 0, "neg enough");
5120 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5121 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5122 if (slice_len_con == 0) {
5123 return; // nothing to do here
5124 }
5125 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5126 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5127 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5128 assert(end_con < 0, "not two cons");
5129 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5130 BytesPerLong);
5131 }
5133 if (start_con >= 0 && end_con >= 0) {
5134 // Constant start and end. Simple.
5135 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5136 start_con, end_con, &_gvn);
5137 } else if (start_con >= 0 && dest_size != top()) {
5138 // Constant start, pre-rounded end after the tail of the array.
5139 Node* end = dest_size;
5140 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5141 start_con, end, &_gvn);
5142 } else if (start_con >= 0 && slice_len != top()) {
5143 // Constant start, non-constant end. End needs rounding up.
5144 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5145 intptr_t end_base = abase + (slice_idx_con << scale);
5146 int end_round = (-1 << scale) & (BytesPerLong - 1);
5147 Node* end = ConvI2X(slice_len);
5148 if (scale != 0)
5149 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5150 end_base += end_round;
5151 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5152 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5153 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5154 start_con, end, &_gvn);
5155 } else if (start_con < 0 && dest_size != top()) {
5156 // Non-constant start, pre-rounded end after the tail of the array.
5157 // This is almost certainly a "round-to-end" operation.
5158 Node* start = slice_idx;
5159 start = ConvI2X(start);
5160 if (scale != 0)
5161 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5162 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5163 if ((bump_bit | clear_low) != 0) {
5164 int to_clear = (bump_bit | clear_low);
5165 // Align up mod 8, then store a jint zero unconditionally
5166 // just before the mod-8 boundary.
5167 if (((abase + bump_bit) & ~to_clear) - bump_bit
5168 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5169 bump_bit = 0;
5170 assert((abase & to_clear) == 0, "array base must be long-aligned");
5171 } else {
5172 // Bump 'start' up to (or past) the next jint boundary:
5173 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5174 assert((abase & clear_low) == 0, "array base must be int-aligned");
5175 }
5176 // Round bumped 'start' down to jlong boundary in body of array.
5177 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5178 if (bump_bit != 0) {
5179 // Store a zero to the immediately preceding jint:
5180 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5181 Node* p1 = basic_plus_adr(dest, x1);
5182 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5183 mem = _gvn.transform(mem);
5184 }
5185 }
5186 Node* end = dest_size; // pre-rounded
5187 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5188 start, end, &_gvn);
5189 } else {
5190 // Non-constant start, unrounded non-constant end.
5191 // (Nobody zeroes a random midsection of an array using this routine.)
5192 ShouldNotReachHere(); // fix caller
5193 }
5195 // Done.
5196 set_memory(mem, adr_type);
5197 }
5200 bool
5201 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5202 BasicType basic_elem_type,
5203 AllocateNode* alloc,
5204 Node* src, Node* src_offset,
5205 Node* dest, Node* dest_offset,
5206 Node* dest_size, bool dest_uninitialized) {
5207 // See if there is an advantage from block transfer.
5208 int scale = exact_log2(type2aelembytes(basic_elem_type));
5209 if (scale >= LogBytesPerLong)
5210 return false; // it is already a block transfer
5212 // Look at the alignment of the starting offsets.
5213 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5215 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5216 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5217 if (src_off_con < 0 || dest_off_con < 0)
5218 // At present, we can only understand constants.
5219 return false;
5221 intptr_t src_off = abase + (src_off_con << scale);
5222 intptr_t dest_off = abase + (dest_off_con << scale);
5224 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5225 // Non-aligned; too bad.
5226 // One more chance: Pick off an initial 32-bit word.
5227 // This is a common case, since abase can be odd mod 8.
5228 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5229 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5230 Node* sptr = basic_plus_adr(src, src_off);
5231 Node* dptr = basic_plus_adr(dest, dest_off);
5232 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5233 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5234 src_off += BytesPerInt;
5235 dest_off += BytesPerInt;
5236 } else {
5237 return false;
5238 }
5239 }
5240 assert(src_off % BytesPerLong == 0, "");
5241 assert(dest_off % BytesPerLong == 0, "");
5243 // Do this copy by giant steps.
5244 Node* sptr = basic_plus_adr(src, src_off);
5245 Node* dptr = basic_plus_adr(dest, dest_off);
5246 Node* countx = dest_size;
5247 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5248 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5250 bool disjoint_bases = true; // since alloc != NULL
5251 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5252 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5254 return true;
5255 }
5258 // Helper function; generates code for the slow case.
5259 // We make a call to a runtime method which emulates the native method,
5260 // but without the native wrapper overhead.
5261 void
5262 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5263 Node* src, Node* src_offset,
5264 Node* dest, Node* dest_offset,
5265 Node* copy_length, bool dest_uninitialized) {
5266 assert(!dest_uninitialized, "Invariant");
5267 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5268 OptoRuntime::slow_arraycopy_Type(),
5269 OptoRuntime::slow_arraycopy_Java(),
5270 "slow_arraycopy", adr_type,
5271 src, src_offset, dest, dest_offset,
5272 copy_length);
5274 // Handle exceptions thrown by this fellow:
5275 make_slow_call_ex(call, env()->Throwable_klass(), false);
5276 }
5278 // Helper function; generates code for cases requiring runtime checks.
5279 Node*
5280 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5281 Node* dest_elem_klass,
5282 Node* src, Node* src_offset,
5283 Node* dest, Node* dest_offset,
5284 Node* copy_length, bool dest_uninitialized) {
5285 if (stopped()) return NULL;
5287 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5288 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5289 return NULL;
5290 }
5292 // Pick out the parameters required to perform a store-check
5293 // for the target array. This is an optimistic check. It will
5294 // look in each non-null element's class, at the desired klass's
5295 // super_check_offset, for the desired klass.
5296 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5297 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5298 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5299 Node* check_offset = ConvI2X(_gvn.transform(n3));
5300 Node* check_value = dest_elem_klass;
5302 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5303 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5305 // (We know the arrays are never conjoint, because their types differ.)
5306 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5307 OptoRuntime::checkcast_arraycopy_Type(),
5308 copyfunc_addr, "checkcast_arraycopy", adr_type,
5309 // five arguments, of which two are
5310 // intptr_t (jlong in LP64)
5311 src_start, dest_start,
5312 copy_length XTOP,
5313 check_offset XTOP,
5314 check_value);
5316 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5317 }
5320 // Helper function; generates code for cases requiring runtime checks.
5321 Node*
5322 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5323 Node* src, Node* src_offset,
5324 Node* dest, Node* dest_offset,
5325 Node* copy_length, bool dest_uninitialized) {
5326 assert(!dest_uninitialized, "Invariant");
5327 if (stopped()) return NULL;
5328 address copyfunc_addr = StubRoutines::generic_arraycopy();
5329 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5330 return NULL;
5331 }
5333 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5334 OptoRuntime::generic_arraycopy_Type(),
5335 copyfunc_addr, "generic_arraycopy", adr_type,
5336 src, src_offset, dest, dest_offset, copy_length);
5338 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5339 }
5341 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5342 void
5343 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5344 BasicType basic_elem_type,
5345 bool disjoint_bases,
5346 Node* src, Node* src_offset,
5347 Node* dest, Node* dest_offset,
5348 Node* copy_length, bool dest_uninitialized) {
5349 if (stopped()) return; // nothing to do
5351 Node* src_start = src;
5352 Node* dest_start = dest;
5353 if (src_offset != NULL || dest_offset != NULL) {
5354 assert(src_offset != NULL && dest_offset != NULL, "");
5355 src_start = array_element_address(src, src_offset, basic_elem_type);
5356 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5357 }
5359 // Figure out which arraycopy runtime method to call.
5360 const char* copyfunc_name = "arraycopy";
5361 address copyfunc_addr =
5362 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5363 disjoint_bases, copyfunc_name, dest_uninitialized);
5365 // Call it. Note that the count_ix value is not scaled to a byte-size.
5366 make_runtime_call(RC_LEAF|RC_NO_FP,
5367 OptoRuntime::fast_arraycopy_Type(),
5368 copyfunc_addr, copyfunc_name, adr_type,
5369 src_start, dest_start, copy_length XTOP);
5370 }
5372 //----------------------------inline_reference_get----------------------------
5373 // public T java.lang.ref.Reference.get();
5374 bool LibraryCallKit::inline_reference_get() {
5375 const int referent_offset = java_lang_ref_Reference::referent_offset;
5376 guarantee(referent_offset > 0, "should have already been set");
5378 // Get the argument:
5379 Node* reference_obj = null_check_receiver();
5380 if (stopped()) return true;
5382 Node* adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5384 ciInstanceKlass* klass = env()->Object_klass();
5385 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5387 Node* no_ctrl = NULL;
5388 Node* result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5390 // Use the pre-barrier to record the value in the referent field
5391 pre_barrier(false /* do_load */,
5392 control(),
5393 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5394 result /* pre_val */,
5395 T_OBJECT);
5397 // Add memory barrier to prevent commoning reads from this field
5398 // across safepoint since GC can change its value.
5399 insert_mem_bar(Op_MemBarCPUOrder);
5401 set_result(result);
5402 return true;
5403 }
5406 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5407 bool is_exact=true, bool is_static=false) {
5409 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5410 assert(tinst != NULL, "obj is null");
5411 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5412 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5414 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5415 ciSymbol::make(fieldTypeString),
5416 is_static);
5417 if (field == NULL) return (Node *) NULL;
5418 assert (field != NULL, "undefined field");
5420 // Next code copied from Parse::do_get_xxx():
5422 // Compute address and memory type.
5423 int offset = field->offset_in_bytes();
5424 bool is_vol = field->is_volatile();
5425 ciType* field_klass = field->type();
5426 assert(field_klass->is_loaded(), "should be loaded");
5427 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5428 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5429 BasicType bt = field->layout_type();
5431 // Build the resultant type of the load
5432 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5434 // Build the load.
5435 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5436 return loadedField;
5437 }
5440 //------------------------------inline_aescrypt_Block-----------------------
5441 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5442 address stubAddr;
5443 const char *stubName;
5444 assert(UseAES, "need AES instruction support");
5446 switch(id) {
5447 case vmIntrinsics::_aescrypt_encryptBlock:
5448 stubAddr = StubRoutines::aescrypt_encryptBlock();
5449 stubName = "aescrypt_encryptBlock";
5450 break;
5451 case vmIntrinsics::_aescrypt_decryptBlock:
5452 stubAddr = StubRoutines::aescrypt_decryptBlock();
5453 stubName = "aescrypt_decryptBlock";
5454 break;
5455 }
5456 if (stubAddr == NULL) return false;
5458 Node* aescrypt_object = argument(0);
5459 Node* src = argument(1);
5460 Node* src_offset = argument(2);
5461 Node* dest = argument(3);
5462 Node* dest_offset = argument(4);
5464 // (1) src and dest are arrays.
5465 const Type* src_type = src->Value(&_gvn);
5466 const Type* dest_type = dest->Value(&_gvn);
5467 const TypeAryPtr* top_src = src_type->isa_aryptr();
5468 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5469 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5471 // for the quick and dirty code we will skip all the checks.
5472 // we are just trying to get the call to be generated.
5473 Node* src_start = src;
5474 Node* dest_start = dest;
5475 if (src_offset != NULL || dest_offset != NULL) {
5476 assert(src_offset != NULL && dest_offset != NULL, "");
5477 src_start = array_element_address(src, src_offset, T_BYTE);
5478 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5479 }
5481 // now need to get the start of its expanded key array
5482 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5483 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5484 if (k_start == NULL) return false;
5486 // Call the stub.
5487 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5488 stubAddr, stubName, TypePtr::BOTTOM,
5489 src_start, dest_start, k_start);
5491 return true;
5492 }
5494 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5495 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5496 address stubAddr;
5497 const char *stubName;
5499 assert(UseAES, "need AES instruction support");
5501 switch(id) {
5502 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5503 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5504 stubName = "cipherBlockChaining_encryptAESCrypt";
5505 break;
5506 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5507 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5508 stubName = "cipherBlockChaining_decryptAESCrypt";
5509 break;
5510 }
5511 if (stubAddr == NULL) return false;
5513 Node* cipherBlockChaining_object = argument(0);
5514 Node* src = argument(1);
5515 Node* src_offset = argument(2);
5516 Node* len = argument(3);
5517 Node* dest = argument(4);
5518 Node* dest_offset = argument(5);
5520 // (1) src and dest are arrays.
5521 const Type* src_type = src->Value(&_gvn);
5522 const Type* dest_type = dest->Value(&_gvn);
5523 const TypeAryPtr* top_src = src_type->isa_aryptr();
5524 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5525 assert (top_src != NULL && top_src->klass() != NULL
5526 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5528 // checks are the responsibility of the caller
5529 Node* src_start = src;
5530 Node* dest_start = dest;
5531 if (src_offset != NULL || dest_offset != NULL) {
5532 assert(src_offset != NULL && dest_offset != NULL, "");
5533 src_start = array_element_address(src, src_offset, T_BYTE);
5534 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5535 }
5537 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5538 // (because of the predicated logic executed earlier).
5539 // so we cast it here safely.
5540 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5542 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5543 if (embeddedCipherObj == NULL) return false;
5545 // cast it to what we know it will be at runtime
5546 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5547 assert(tinst != NULL, "CBC obj is null");
5548 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5549 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5550 if (!klass_AESCrypt->is_loaded()) return false;
5552 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5553 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5554 const TypeOopPtr* xtype = aklass->as_instance_type();
5555 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5556 aescrypt_object = _gvn.transform(aescrypt_object);
5558 // we need to get the start of the aescrypt_object's expanded key array
5559 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5560 if (k_start == NULL) return false;
5562 // similarly, get the start address of the r vector
5563 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5564 if (objRvec == NULL) return false;
5565 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5567 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5568 make_runtime_call(RC_LEAF|RC_NO_FP,
5569 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5570 stubAddr, stubName, TypePtr::BOTTOM,
5571 src_start, dest_start, k_start, r_start, len);
5573 // return is void so no result needs to be pushed
5575 return true;
5576 }
5578 //------------------------------get_key_start_from_aescrypt_object-----------------------
5579 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5580 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5581 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5582 if (objAESCryptKey == NULL) return (Node *) NULL;
5584 // now have the array, need to get the start address of the K array
5585 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5586 return k_start;
5587 }
5589 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5590 // Return node representing slow path of predicate check.
5591 // the pseudo code we want to emulate with this predicate is:
5592 // for encryption:
5593 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5594 // for decryption:
5595 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5596 // note cipher==plain is more conservative than the original java code but that's OK
5597 //
5598 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5599 // First, check receiver for NULL since it is virtual method.
5600 Node* objCBC = argument(0);
5601 objCBC = null_check(objCBC);
5603 if (stopped()) return NULL; // Always NULL
5605 // Load embeddedCipher field of CipherBlockChaining object.
5606 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5608 // get AESCrypt klass for instanceOf check
5609 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5610 // will have same classloader as CipherBlockChaining object
5611 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5612 assert(tinst != NULL, "CBCobj is null");
5613 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5615 // we want to do an instanceof comparison against the AESCrypt class
5616 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5617 if (!klass_AESCrypt->is_loaded()) {
5618 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5619 Node* ctrl = control();
5620 set_control(top()); // no regular fast path
5621 return ctrl;
5622 }
5623 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5625 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5626 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5627 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5629 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5631 // for encryption, we are done
5632 if (!decrypting)
5633 return instof_false; // even if it is NULL
5635 // for decryption, we need to add a further check to avoid
5636 // taking the intrinsic path when cipher and plain are the same
5637 // see the original java code for why.
5638 RegionNode* region = new(C) RegionNode(3);
5639 region->init_req(1, instof_false);
5640 Node* src = argument(1);
5641 Node* dest = argument(4);
5642 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5643 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5644 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5645 region->init_req(2, src_dest_conjoint);
5647 record_for_igvn(region);
5648 return _gvn.transform(region);
5649 }