Wed, 24 Oct 2012 14:33:22 -0700
7184394: add intrinsics to use AES instructions
Summary: Use new x86 AES instructions for AESCrypt.
Reviewed-by: twisti, kvn, roland
Contributed-by: tom.deneau@amd.com
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
72 const TypeOopPtr* sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr = false);
74 public:
75 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
76 : GraphKit(caller),
77 _intrinsic(intrinsic)
78 {
79 }
81 ciMethod* caller() const { return jvms()->method(); }
82 int bci() const { return jvms()->bci(); }
83 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
84 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
85 ciMethod* callee() const { return _intrinsic->method(); }
86 ciSignature* signature() const { return callee()->signature(); }
87 int arg_size() const { return callee()->arg_size(); }
89 bool try_to_inline();
90 Node* try_to_predicate();
92 // Helper functions to inline natives
93 void push_result(RegionNode* region, PhiNode* value);
94 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
95 Node* generate_slow_guard(Node* test, RegionNode* region);
96 Node* generate_fair_guard(Node* test, RegionNode* region);
97 Node* generate_negative_guard(Node* index, RegionNode* region,
98 // resulting CastII of index:
99 Node* *pos_index = NULL);
100 Node* generate_nonpositive_guard(Node* index, bool never_negative,
101 // resulting CastII of index:
102 Node* *pos_index = NULL);
103 Node* generate_limit_guard(Node* offset, Node* subseq_length,
104 Node* array_length,
105 RegionNode* region);
106 Node* generate_current_thread(Node* &tls_output);
107 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
108 bool disjoint_bases, const char* &name, bool dest_uninitialized);
109 Node* load_mirror_from_klass(Node* klass);
110 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
111 int nargs,
112 RegionNode* region, int null_path,
113 int offset);
114 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
115 RegionNode* region, int null_path) {
116 int offset = java_lang_Class::klass_offset_in_bytes();
117 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
118 region, null_path,
119 offset);
120 }
121 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
122 int nargs,
123 RegionNode* region, int null_path) {
124 int offset = java_lang_Class::array_klass_offset_in_bytes();
125 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
126 region, null_path,
127 offset);
128 }
129 Node* generate_access_flags_guard(Node* kls,
130 int modifier_mask, int modifier_bits,
131 RegionNode* region);
132 Node* generate_interface_guard(Node* kls, RegionNode* region);
133 Node* generate_array_guard(Node* kls, RegionNode* region) {
134 return generate_array_guard_common(kls, region, false, false);
135 }
136 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
137 return generate_array_guard_common(kls, region, false, true);
138 }
139 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
140 return generate_array_guard_common(kls, region, true, false);
141 }
142 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
143 return generate_array_guard_common(kls, region, true, true);
144 }
145 Node* generate_array_guard_common(Node* kls, RegionNode* region,
146 bool obj_array, bool not_array);
147 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
148 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
149 bool is_virtual = false, bool is_static = false);
150 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
151 return generate_method_call(method_id, false, true);
152 }
153 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
154 return generate_method_call(method_id, true, false);
155 }
156 Node * load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString, bool is_exact, bool is_static);
158 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
159 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
160 bool inline_string_compareTo();
161 bool inline_string_indexOf();
162 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
163 bool inline_string_equals();
164 Node* pop_math_arg();
165 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
166 bool inline_math_native(vmIntrinsics::ID id);
167 bool inline_trig(vmIntrinsics::ID id);
168 bool inline_trans(vmIntrinsics::ID id);
169 bool inline_abs(vmIntrinsics::ID id);
170 bool inline_sqrt(vmIntrinsics::ID id);
171 void finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName);
172 bool inline_pow(vmIntrinsics::ID id);
173 bool inline_exp(vmIntrinsics::ID id);
174 bool inline_min_max(vmIntrinsics::ID id);
175 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
176 // This returns Type::AnyPtr, RawPtr, or OopPtr.
177 int classify_unsafe_addr(Node* &base, Node* &offset);
178 Node* make_unsafe_address(Node* base, Node* offset);
179 // Helper for inline_unsafe_access.
180 // Generates the guards that check whether the result of
181 // Unsafe.getObject should be recorded in an SATB log buffer.
182 void insert_pre_barrier(Node* base_oop, Node* offset, Node* pre_val, int nargs, bool need_mem_bar);
183 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
184 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
185 bool inline_unsafe_allocate();
186 bool inline_unsafe_copyMemory();
187 bool inline_native_currentThread();
188 #ifdef TRACE_HAVE_INTRINSICS
189 bool inline_native_classID();
190 bool inline_native_threadID();
191 #endif
192 bool inline_native_time_funcs(address method, const char* funcName);
193 bool inline_native_isInterrupted();
194 bool inline_native_Class_query(vmIntrinsics::ID id);
195 bool inline_native_subtype_check();
197 bool inline_native_newArray();
198 bool inline_native_getLength();
199 bool inline_array_copyOf(bool is_copyOfRange);
200 bool inline_array_equals();
201 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
202 bool inline_native_clone(bool is_virtual);
203 bool inline_native_Reflection_getCallerClass();
204 bool is_method_invoke_or_aux_frame(JVMState* jvms);
205 // Helper function for inlining native object hash method
206 bool inline_native_hashcode(bool is_virtual, bool is_static);
207 bool inline_native_getClass();
209 // Helper functions for inlining arraycopy
210 bool inline_arraycopy();
211 void generate_arraycopy(const TypePtr* adr_type,
212 BasicType basic_elem_type,
213 Node* src, Node* src_offset,
214 Node* dest, Node* dest_offset,
215 Node* copy_length,
216 bool disjoint_bases = false,
217 bool length_never_negative = false,
218 RegionNode* slow_region = NULL);
219 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
220 RegionNode* slow_region);
221 void generate_clear_array(const TypePtr* adr_type,
222 Node* dest,
223 BasicType basic_elem_type,
224 Node* slice_off,
225 Node* slice_len,
226 Node* slice_end);
227 bool generate_block_arraycopy(const TypePtr* adr_type,
228 BasicType basic_elem_type,
229 AllocateNode* alloc,
230 Node* src, Node* src_offset,
231 Node* dest, Node* dest_offset,
232 Node* dest_size, bool dest_uninitialized);
233 void generate_slow_arraycopy(const TypePtr* adr_type,
234 Node* src, Node* src_offset,
235 Node* dest, Node* dest_offset,
236 Node* copy_length, bool dest_uninitialized);
237 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
238 Node* dest_elem_klass,
239 Node* src, Node* src_offset,
240 Node* dest, Node* dest_offset,
241 Node* copy_length, bool dest_uninitialized);
242 Node* generate_generic_arraycopy(const TypePtr* adr_type,
243 Node* src, Node* src_offset,
244 Node* dest, Node* dest_offset,
245 Node* copy_length, bool dest_uninitialized);
246 void generate_unchecked_arraycopy(const TypePtr* adr_type,
247 BasicType basic_elem_type,
248 bool disjoint_bases,
249 Node* src, Node* src_offset,
250 Node* dest, Node* dest_offset,
251 Node* copy_length, bool dest_uninitialized);
252 typedef enum { LS_xadd, LS_xchg, LS_cmpxchg } LoadStoreKind;
253 bool inline_unsafe_load_store(BasicType type, LoadStoreKind kind);
254 bool inline_unsafe_ordered_store(BasicType type);
255 bool inline_fp_conversions(vmIntrinsics::ID id);
256 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
257 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
258 bool inline_bitCount(vmIntrinsics::ID id);
259 bool inline_reverseBytes(vmIntrinsics::ID id);
261 bool inline_reference_get();
262 bool inline_aescrypt_Block(vmIntrinsics::ID id);
263 bool inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id);
264 Node* inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting);
265 Node* get_key_start_from_aescrypt_object(Node* aescrypt_object);
266 };
269 //---------------------------make_vm_intrinsic----------------------------
270 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
271 vmIntrinsics::ID id = m->intrinsic_id();
272 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
274 if (DisableIntrinsic[0] != '\0'
275 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
276 // disabled by a user request on the command line:
277 // example: -XX:DisableIntrinsic=_hashCode,_getClass
278 return NULL;
279 }
281 if (!m->is_loaded()) {
282 // do not attempt to inline unloaded methods
283 return NULL;
284 }
286 // Only a few intrinsics implement a virtual dispatch.
287 // They are expensive calls which are also frequently overridden.
288 if (is_virtual) {
289 switch (id) {
290 case vmIntrinsics::_hashCode:
291 case vmIntrinsics::_clone:
292 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
293 break;
294 default:
295 return NULL;
296 }
297 }
299 // -XX:-InlineNatives disables nearly all intrinsics:
300 if (!InlineNatives) {
301 switch (id) {
302 case vmIntrinsics::_indexOf:
303 case vmIntrinsics::_compareTo:
304 case vmIntrinsics::_equals:
305 case vmIntrinsics::_equalsC:
306 case vmIntrinsics::_getAndAddInt:
307 case vmIntrinsics::_getAndAddLong:
308 case vmIntrinsics::_getAndSetInt:
309 case vmIntrinsics::_getAndSetLong:
310 case vmIntrinsics::_getAndSetObject:
311 break; // InlineNatives does not control String.compareTo
312 case vmIntrinsics::_Reference_get:
313 break; // InlineNatives does not control Reference.get
314 default:
315 return NULL;
316 }
317 }
319 bool is_predicted = false;
321 switch (id) {
322 case vmIntrinsics::_compareTo:
323 if (!SpecialStringCompareTo) return NULL;
324 break;
325 case vmIntrinsics::_indexOf:
326 if (!SpecialStringIndexOf) return NULL;
327 break;
328 case vmIntrinsics::_equals:
329 if (!SpecialStringEquals) return NULL;
330 break;
331 case vmIntrinsics::_equalsC:
332 if (!SpecialArraysEquals) return NULL;
333 break;
334 case vmIntrinsics::_arraycopy:
335 if (!InlineArrayCopy) return NULL;
336 break;
337 case vmIntrinsics::_copyMemory:
338 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
339 if (!InlineArrayCopy) return NULL;
340 break;
341 case vmIntrinsics::_hashCode:
342 if (!InlineObjectHash) return NULL;
343 break;
344 case vmIntrinsics::_clone:
345 case vmIntrinsics::_copyOf:
346 case vmIntrinsics::_copyOfRange:
347 if (!InlineObjectCopy) return NULL;
348 // These also use the arraycopy intrinsic mechanism:
349 if (!InlineArrayCopy) return NULL;
350 break;
351 case vmIntrinsics::_checkIndex:
352 // We do not intrinsify this. The optimizer does fine with it.
353 return NULL;
355 case vmIntrinsics::_getCallerClass:
356 if (!UseNewReflection) return NULL;
357 if (!InlineReflectionGetCallerClass) return NULL;
358 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
359 break;
361 case vmIntrinsics::_bitCount_i:
362 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
363 break;
365 case vmIntrinsics::_bitCount_l:
366 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
367 break;
369 case vmIntrinsics::_numberOfLeadingZeros_i:
370 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
371 break;
373 case vmIntrinsics::_numberOfLeadingZeros_l:
374 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
375 break;
377 case vmIntrinsics::_numberOfTrailingZeros_i:
378 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
379 break;
381 case vmIntrinsics::_numberOfTrailingZeros_l:
382 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
383 break;
385 case vmIntrinsics::_Reference_get:
386 // Use the intrinsic version of Reference.get() so that the value in
387 // the referent field can be registered by the G1 pre-barrier code.
388 // Also add memory barrier to prevent commoning reads from this field
389 // across safepoint since GC can change it value.
390 break;
392 case vmIntrinsics::_compareAndSwapObject:
393 #ifdef _LP64
394 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_CompareAndSwapP)) return NULL;
395 #endif
396 break;
398 case vmIntrinsics::_compareAndSwapLong:
399 if (!Matcher::match_rule_supported(Op_CompareAndSwapL)) return NULL;
400 break;
402 case vmIntrinsics::_getAndAddInt:
403 if (!Matcher::match_rule_supported(Op_GetAndAddI)) return NULL;
404 break;
406 case vmIntrinsics::_getAndAddLong:
407 if (!Matcher::match_rule_supported(Op_GetAndAddL)) return NULL;
408 break;
410 case vmIntrinsics::_getAndSetInt:
411 if (!Matcher::match_rule_supported(Op_GetAndSetI)) return NULL;
412 break;
414 case vmIntrinsics::_getAndSetLong:
415 if (!Matcher::match_rule_supported(Op_GetAndSetL)) return NULL;
416 break;
418 case vmIntrinsics::_getAndSetObject:
419 #ifdef _LP64
420 if (!UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
421 if (UseCompressedOops && !Matcher::match_rule_supported(Op_GetAndSetN)) return NULL;
422 break;
423 #else
424 if (!Matcher::match_rule_supported(Op_GetAndSetP)) return NULL;
425 break;
426 #endif
428 case vmIntrinsics::_aescrypt_encryptBlock:
429 case vmIntrinsics::_aescrypt_decryptBlock:
430 if (!UseAESIntrinsics) return NULL;
431 break;
433 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
434 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
435 if (!UseAESIntrinsics) return NULL;
436 // these two require the predicated logic
437 is_predicted = true;
438 break;
440 default:
441 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
442 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
443 break;
444 }
446 // -XX:-InlineClassNatives disables natives from the Class class.
447 // The flag applies to all reflective calls, notably Array.newArray
448 // (visible to Java programmers as Array.newInstance).
449 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
450 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
451 if (!InlineClassNatives) return NULL;
452 }
454 // -XX:-InlineThreadNatives disables natives from the Thread class.
455 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
456 if (!InlineThreadNatives) return NULL;
457 }
459 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
460 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
461 m->holder()->name() == ciSymbol::java_lang_Float() ||
462 m->holder()->name() == ciSymbol::java_lang_Double()) {
463 if (!InlineMathNatives) return NULL;
464 }
466 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
467 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
468 if (!InlineUnsafeOps) return NULL;
469 }
471 return new LibraryIntrinsic(m, is_virtual, is_predicted, (vmIntrinsics::ID) id);
472 }
474 //----------------------register_library_intrinsics-----------------------
475 // Initialize this file's data structures, for each Compile instance.
476 void Compile::register_library_intrinsics() {
477 // Nothing to do here.
478 }
480 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
481 LibraryCallKit kit(jvms, this);
482 Compile* C = kit.C;
483 int nodes = C->unique();
484 #ifndef PRODUCT
485 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
486 char buf[1000];
487 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
488 tty->print_cr("Intrinsic %s", str);
489 }
490 #endif
492 if (kit.try_to_inline()) {
493 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
494 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
495 }
496 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
497 if (C->log()) {
498 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
499 vmIntrinsics::name_at(intrinsic_id()),
500 (is_virtual() ? " virtual='1'" : ""),
501 C->unique() - nodes);
502 }
503 return kit.transfer_exceptions_into_jvms();
504 }
506 // The intrinsic bailed out
507 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
508 if (jvms->has_method()) {
509 // Not a root compile.
510 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
511 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
512 } else {
513 // Root compile
514 tty->print("Did not generate intrinsic %s%s at bci:%d in",
515 vmIntrinsics::name_at(intrinsic_id()),
516 (is_virtual() ? " (virtual)" : ""), kit.bci());
517 }
518 }
519 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
520 return NULL;
521 }
523 Node* LibraryIntrinsic::generate_predicate(JVMState* jvms) {
524 LibraryCallKit kit(jvms, this);
525 Compile* C = kit.C;
526 int nodes = C->unique();
527 #ifndef PRODUCT
528 assert(is_predicted(), "sanity");
529 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
530 char buf[1000];
531 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
532 tty->print_cr("Predicate for intrinsic %s", str);
533 }
534 #endif
536 Node* slow_ctl = kit.try_to_predicate();
537 if (!kit.failing()) {
538 if (C->log()) {
539 C->log()->elem("predicate_intrinsic id='%s'%s nodes='%d'",
540 vmIntrinsics::name_at(intrinsic_id()),
541 (is_virtual() ? " virtual='1'" : ""),
542 C->unique() - nodes);
543 }
544 return slow_ctl; // Could be NULL if the check folds.
545 }
547 // The intrinsic bailed out
548 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
549 if (jvms->has_method()) {
550 // Not a root compile.
551 const char* msg = "failed to generate predicate for intrinsic";
552 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
553 } else {
554 // Root compile
555 tty->print("Did not generate predicate for intrinsic %s%s at bci:%d in",
556 vmIntrinsics::name_at(intrinsic_id()),
557 (is_virtual() ? " (virtual)" : ""), kit.bci());
558 }
559 }
560 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
561 return NULL;
562 }
564 bool LibraryCallKit::try_to_inline() {
565 // Handle symbolic names for otherwise undistinguished boolean switches:
566 const bool is_store = true;
567 const bool is_native_ptr = true;
568 const bool is_static = true;
570 if (!jvms()->has_method()) {
571 // Root JVMState has a null method.
572 assert(map()->memory()->Opcode() == Op_Parm, "");
573 // Insert the memory aliasing node
574 set_all_memory(reset_memory());
575 }
576 assert(merged_memory(), "");
578 switch (intrinsic_id()) {
579 case vmIntrinsics::_hashCode:
580 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
581 case vmIntrinsics::_identityHashCode:
582 return inline_native_hashcode(/*!virtual*/ false, is_static);
583 case vmIntrinsics::_getClass:
584 return inline_native_getClass();
586 case vmIntrinsics::_dsin:
587 case vmIntrinsics::_dcos:
588 case vmIntrinsics::_dtan:
589 case vmIntrinsics::_dabs:
590 case vmIntrinsics::_datan2:
591 case vmIntrinsics::_dsqrt:
592 case vmIntrinsics::_dexp:
593 case vmIntrinsics::_dlog:
594 case vmIntrinsics::_dlog10:
595 case vmIntrinsics::_dpow:
596 return inline_math_native(intrinsic_id());
598 case vmIntrinsics::_min:
599 case vmIntrinsics::_max:
600 return inline_min_max(intrinsic_id());
602 case vmIntrinsics::_arraycopy:
603 return inline_arraycopy();
605 case vmIntrinsics::_compareTo:
606 return inline_string_compareTo();
607 case vmIntrinsics::_indexOf:
608 return inline_string_indexOf();
609 case vmIntrinsics::_equals:
610 return inline_string_equals();
612 case vmIntrinsics::_getObject:
613 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
614 case vmIntrinsics::_getBoolean:
615 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
616 case vmIntrinsics::_getByte:
617 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
618 case vmIntrinsics::_getShort:
619 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
620 case vmIntrinsics::_getChar:
621 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
622 case vmIntrinsics::_getInt:
623 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
624 case vmIntrinsics::_getLong:
625 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
626 case vmIntrinsics::_getFloat:
627 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
628 case vmIntrinsics::_getDouble:
629 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
631 case vmIntrinsics::_putObject:
632 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
633 case vmIntrinsics::_putBoolean:
634 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
635 case vmIntrinsics::_putByte:
636 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
637 case vmIntrinsics::_putShort:
638 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
639 case vmIntrinsics::_putChar:
640 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
641 case vmIntrinsics::_putInt:
642 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
643 case vmIntrinsics::_putLong:
644 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
645 case vmIntrinsics::_putFloat:
646 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
647 case vmIntrinsics::_putDouble:
648 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
650 case vmIntrinsics::_getByte_raw:
651 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
652 case vmIntrinsics::_getShort_raw:
653 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
654 case vmIntrinsics::_getChar_raw:
655 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
656 case vmIntrinsics::_getInt_raw:
657 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
658 case vmIntrinsics::_getLong_raw:
659 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
660 case vmIntrinsics::_getFloat_raw:
661 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
662 case vmIntrinsics::_getDouble_raw:
663 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
664 case vmIntrinsics::_getAddress_raw:
665 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
667 case vmIntrinsics::_putByte_raw:
668 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
669 case vmIntrinsics::_putShort_raw:
670 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
671 case vmIntrinsics::_putChar_raw:
672 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
673 case vmIntrinsics::_putInt_raw:
674 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
675 case vmIntrinsics::_putLong_raw:
676 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
677 case vmIntrinsics::_putFloat_raw:
678 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
679 case vmIntrinsics::_putDouble_raw:
680 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
681 case vmIntrinsics::_putAddress_raw:
682 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
684 case vmIntrinsics::_getObjectVolatile:
685 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
686 case vmIntrinsics::_getBooleanVolatile:
687 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
688 case vmIntrinsics::_getByteVolatile:
689 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
690 case vmIntrinsics::_getShortVolatile:
691 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
692 case vmIntrinsics::_getCharVolatile:
693 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
694 case vmIntrinsics::_getIntVolatile:
695 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
696 case vmIntrinsics::_getLongVolatile:
697 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
698 case vmIntrinsics::_getFloatVolatile:
699 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
700 case vmIntrinsics::_getDoubleVolatile:
701 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
703 case vmIntrinsics::_putObjectVolatile:
704 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
705 case vmIntrinsics::_putBooleanVolatile:
706 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
707 case vmIntrinsics::_putByteVolatile:
708 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
709 case vmIntrinsics::_putShortVolatile:
710 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
711 case vmIntrinsics::_putCharVolatile:
712 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
713 case vmIntrinsics::_putIntVolatile:
714 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
715 case vmIntrinsics::_putLongVolatile:
716 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
717 case vmIntrinsics::_putFloatVolatile:
718 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
719 case vmIntrinsics::_putDoubleVolatile:
720 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
722 case vmIntrinsics::_prefetchRead:
723 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
724 case vmIntrinsics::_prefetchWrite:
725 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
726 case vmIntrinsics::_prefetchReadStatic:
727 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
728 case vmIntrinsics::_prefetchWriteStatic:
729 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
731 case vmIntrinsics::_compareAndSwapObject:
732 return inline_unsafe_load_store(T_OBJECT, LS_cmpxchg);
733 case vmIntrinsics::_compareAndSwapInt:
734 return inline_unsafe_load_store(T_INT, LS_cmpxchg);
735 case vmIntrinsics::_compareAndSwapLong:
736 return inline_unsafe_load_store(T_LONG, LS_cmpxchg);
738 case vmIntrinsics::_putOrderedObject:
739 return inline_unsafe_ordered_store(T_OBJECT);
740 case vmIntrinsics::_putOrderedInt:
741 return inline_unsafe_ordered_store(T_INT);
742 case vmIntrinsics::_putOrderedLong:
743 return inline_unsafe_ordered_store(T_LONG);
745 case vmIntrinsics::_getAndAddInt:
746 return inline_unsafe_load_store(T_INT, LS_xadd);
747 case vmIntrinsics::_getAndAddLong:
748 return inline_unsafe_load_store(T_LONG, LS_xadd);
749 case vmIntrinsics::_getAndSetInt:
750 return inline_unsafe_load_store(T_INT, LS_xchg);
751 case vmIntrinsics::_getAndSetLong:
752 return inline_unsafe_load_store(T_LONG, LS_xchg);
753 case vmIntrinsics::_getAndSetObject:
754 return inline_unsafe_load_store(T_OBJECT, LS_xchg);
756 case vmIntrinsics::_currentThread:
757 return inline_native_currentThread();
758 case vmIntrinsics::_isInterrupted:
759 return inline_native_isInterrupted();
761 #ifdef TRACE_HAVE_INTRINSICS
762 case vmIntrinsics::_classID:
763 return inline_native_classID();
764 case vmIntrinsics::_threadID:
765 return inline_native_threadID();
766 case vmIntrinsics::_counterTime:
767 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
768 #endif
769 case vmIntrinsics::_currentTimeMillis:
770 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
771 case vmIntrinsics::_nanoTime:
772 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
773 case vmIntrinsics::_allocateInstance:
774 return inline_unsafe_allocate();
775 case vmIntrinsics::_copyMemory:
776 return inline_unsafe_copyMemory();
777 case vmIntrinsics::_newArray:
778 return inline_native_newArray();
779 case vmIntrinsics::_getLength:
780 return inline_native_getLength();
781 case vmIntrinsics::_copyOf:
782 return inline_array_copyOf(false);
783 case vmIntrinsics::_copyOfRange:
784 return inline_array_copyOf(true);
785 case vmIntrinsics::_equalsC:
786 return inline_array_equals();
787 case vmIntrinsics::_clone:
788 return inline_native_clone(intrinsic()->is_virtual());
790 case vmIntrinsics::_isAssignableFrom:
791 return inline_native_subtype_check();
793 case vmIntrinsics::_isInstance:
794 case vmIntrinsics::_getModifiers:
795 case vmIntrinsics::_isInterface:
796 case vmIntrinsics::_isArray:
797 case vmIntrinsics::_isPrimitive:
798 case vmIntrinsics::_getSuperclass:
799 case vmIntrinsics::_getComponentType:
800 case vmIntrinsics::_getClassAccessFlags:
801 return inline_native_Class_query(intrinsic_id());
803 case vmIntrinsics::_floatToRawIntBits:
804 case vmIntrinsics::_floatToIntBits:
805 case vmIntrinsics::_intBitsToFloat:
806 case vmIntrinsics::_doubleToRawLongBits:
807 case vmIntrinsics::_doubleToLongBits:
808 case vmIntrinsics::_longBitsToDouble:
809 return inline_fp_conversions(intrinsic_id());
811 case vmIntrinsics::_numberOfLeadingZeros_i:
812 case vmIntrinsics::_numberOfLeadingZeros_l:
813 return inline_numberOfLeadingZeros(intrinsic_id());
815 case vmIntrinsics::_numberOfTrailingZeros_i:
816 case vmIntrinsics::_numberOfTrailingZeros_l:
817 return inline_numberOfTrailingZeros(intrinsic_id());
819 case vmIntrinsics::_bitCount_i:
820 case vmIntrinsics::_bitCount_l:
821 return inline_bitCount(intrinsic_id());
823 case vmIntrinsics::_reverseBytes_i:
824 case vmIntrinsics::_reverseBytes_l:
825 case vmIntrinsics::_reverseBytes_s:
826 case vmIntrinsics::_reverseBytes_c:
827 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
829 case vmIntrinsics::_getCallerClass:
830 return inline_native_Reflection_getCallerClass();
832 case vmIntrinsics::_Reference_get:
833 return inline_reference_get();
835 case vmIntrinsics::_aescrypt_encryptBlock:
836 case vmIntrinsics::_aescrypt_decryptBlock:
837 return inline_aescrypt_Block(intrinsic_id());
839 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
840 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
841 return inline_cipherBlockChaining_AESCrypt(intrinsic_id());
843 default:
844 // If you get here, it may be that someone has added a new intrinsic
845 // to the list in vmSymbols.hpp without implementing it here.
846 #ifndef PRODUCT
847 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
848 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
849 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
850 }
851 #endif
852 return false;
853 }
854 }
856 Node* LibraryCallKit::try_to_predicate() {
857 if (!jvms()->has_method()) {
858 // Root JVMState has a null method.
859 assert(map()->memory()->Opcode() == Op_Parm, "");
860 // Insert the memory aliasing node
861 set_all_memory(reset_memory());
862 }
863 assert(merged_memory(), "");
865 switch (intrinsic_id()) {
866 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
867 return inline_cipherBlockChaining_AESCrypt_predicate(false);
868 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
869 return inline_cipherBlockChaining_AESCrypt_predicate(true);
871 default:
872 // If you get here, it may be that someone has added a new intrinsic
873 // to the list in vmSymbols.hpp without implementing it here.
874 #ifndef PRODUCT
875 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
876 tty->print_cr("*** Warning: Unimplemented predicate for intrinsic %s(%d)",
877 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
878 }
879 #endif
880 Node* slow_ctl = control();
881 set_control(top()); // No fast path instrinsic
882 return slow_ctl;
883 }
884 }
886 //------------------------------push_result------------------------------
887 // Helper function for finishing intrinsics.
888 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
889 record_for_igvn(region);
890 set_control(_gvn.transform(region));
891 BasicType value_type = value->type()->basic_type();
892 push_node(value_type, _gvn.transform(value));
893 }
895 //------------------------------generate_guard---------------------------
896 // Helper function for generating guarded fast-slow graph structures.
897 // The given 'test', if true, guards a slow path. If the test fails
898 // then a fast path can be taken. (We generally hope it fails.)
899 // In all cases, GraphKit::control() is updated to the fast path.
900 // The returned value represents the control for the slow path.
901 // The return value is never 'top'; it is either a valid control
902 // or NULL if it is obvious that the slow path can never be taken.
903 // Also, if region and the slow control are not NULL, the slow edge
904 // is appended to the region.
905 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
906 if (stopped()) {
907 // Already short circuited.
908 return NULL;
909 }
911 // Build an if node and its projections.
912 // If test is true we take the slow path, which we assume is uncommon.
913 if (_gvn.type(test) == TypeInt::ZERO) {
914 // The slow branch is never taken. No need to build this guard.
915 return NULL;
916 }
918 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
920 Node* if_slow = _gvn.transform( new (C) IfTrueNode(iff) );
921 if (if_slow == top()) {
922 // The slow branch is never taken. No need to build this guard.
923 return NULL;
924 }
926 if (region != NULL)
927 region->add_req(if_slow);
929 Node* if_fast = _gvn.transform( new (C) IfFalseNode(iff) );
930 set_control(if_fast);
932 return if_slow;
933 }
935 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
936 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
937 }
938 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
939 return generate_guard(test, region, PROB_FAIR);
940 }
942 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
943 Node* *pos_index) {
944 if (stopped())
945 return NULL; // already stopped
946 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
947 return NULL; // index is already adequately typed
948 Node* cmp_lt = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
949 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
950 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
951 if (is_neg != NULL && pos_index != NULL) {
952 // Emulate effect of Parse::adjust_map_after_if.
953 Node* ccast = new (C) CastIINode(index, TypeInt::POS);
954 ccast->set_req(0, control());
955 (*pos_index) = _gvn.transform(ccast);
956 }
957 return is_neg;
958 }
960 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
961 Node* *pos_index) {
962 if (stopped())
963 return NULL; // already stopped
964 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
965 return NULL; // index is already adequately typed
966 Node* cmp_le = _gvn.transform( new (C) CmpINode(index, intcon(0)) );
967 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
968 Node* bol_le = _gvn.transform( new (C) BoolNode(cmp_le, le_or_eq) );
969 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
970 if (is_notp != NULL && pos_index != NULL) {
971 // Emulate effect of Parse::adjust_map_after_if.
972 Node* ccast = new (C) CastIINode(index, TypeInt::POS1);
973 ccast->set_req(0, control());
974 (*pos_index) = _gvn.transform(ccast);
975 }
976 return is_notp;
977 }
979 // Make sure that 'position' is a valid limit index, in [0..length].
980 // There are two equivalent plans for checking this:
981 // A. (offset + copyLength) unsigned<= arrayLength
982 // B. offset <= (arrayLength - copyLength)
983 // We require that all of the values above, except for the sum and
984 // difference, are already known to be non-negative.
985 // Plan A is robust in the face of overflow, if offset and copyLength
986 // are both hugely positive.
987 //
988 // Plan B is less direct and intuitive, but it does not overflow at
989 // all, since the difference of two non-negatives is always
990 // representable. Whenever Java methods must perform the equivalent
991 // check they generally use Plan B instead of Plan A.
992 // For the moment we use Plan A.
993 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
994 Node* subseq_length,
995 Node* array_length,
996 RegionNode* region) {
997 if (stopped())
998 return NULL; // already stopped
999 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
1000 if (zero_offset && subseq_length->eqv_uncast(array_length))
1001 return NULL; // common case of whole-array copy
1002 Node* last = subseq_length;
1003 if (!zero_offset) // last += offset
1004 last = _gvn.transform( new (C) AddINode(last, offset));
1005 Node* cmp_lt = _gvn.transform( new (C) CmpUNode(array_length, last) );
1006 Node* bol_lt = _gvn.transform( new (C) BoolNode(cmp_lt, BoolTest::lt) );
1007 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
1008 return is_over;
1009 }
1012 //--------------------------generate_current_thread--------------------
1013 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
1014 ciKlass* thread_klass = env()->Thread_klass();
1015 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
1016 Node* thread = _gvn.transform(new (C) ThreadLocalNode());
1017 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
1018 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
1019 tls_output = thread;
1020 return threadObj;
1021 }
1024 //------------------------------make_string_method_node------------------------
1025 // Helper method for String intrinsic functions. This version is called
1026 // with str1 and str2 pointing to String object nodes.
1027 //
1028 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
1029 Node* no_ctrl = NULL;
1031 // Get start addr of string
1032 Node* str1_value = load_String_value(no_ctrl, str1);
1033 Node* str1_offset = load_String_offset(no_ctrl, str1);
1034 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
1036 // Get length of string 1
1037 Node* str1_len = load_String_length(no_ctrl, str1);
1039 Node* str2_value = load_String_value(no_ctrl, str2);
1040 Node* str2_offset = load_String_offset(no_ctrl, str2);
1041 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
1043 Node* str2_len = NULL;
1044 Node* result = NULL;
1046 switch (opcode) {
1047 case Op_StrIndexOf:
1048 // Get length of string 2
1049 str2_len = load_String_length(no_ctrl, str2);
1051 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1052 str1_start, str1_len, str2_start, str2_len);
1053 break;
1054 case Op_StrComp:
1055 // Get length of string 2
1056 str2_len = load_String_length(no_ctrl, str2);
1058 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1059 str1_start, str1_len, str2_start, str2_len);
1060 break;
1061 case Op_StrEquals:
1062 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1063 str1_start, str2_start, str1_len);
1064 break;
1065 default:
1066 ShouldNotReachHere();
1067 return NULL;
1068 }
1070 // All these intrinsics have checks.
1071 C->set_has_split_ifs(true); // Has chance for split-if optimization
1073 return _gvn.transform(result);
1074 }
1076 // Helper method for String intrinsic functions. This version is called
1077 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
1078 // to Int nodes containing the lenghts of str1 and str2.
1079 //
1080 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
1082 Node* result = NULL;
1083 switch (opcode) {
1084 case Op_StrIndexOf:
1085 result = new (C) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
1086 str1_start, cnt1, str2_start, cnt2);
1087 break;
1088 case Op_StrComp:
1089 result = new (C) StrCompNode(control(), memory(TypeAryPtr::CHARS),
1090 str1_start, cnt1, str2_start, cnt2);
1091 break;
1092 case Op_StrEquals:
1093 result = new (C) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
1094 str1_start, str2_start, cnt1);
1095 break;
1096 default:
1097 ShouldNotReachHere();
1098 return NULL;
1099 }
1101 // All these intrinsics have checks.
1102 C->set_has_split_ifs(true); // Has chance for split-if optimization
1104 return _gvn.transform(result);
1105 }
1107 //------------------------------inline_string_compareTo------------------------
1108 bool LibraryCallKit::inline_string_compareTo() {
1110 if (!Matcher::has_match_rule(Op_StrComp)) return false;
1112 _sp += 2;
1113 Node *argument = pop(); // pop non-receiver first: it was pushed second
1114 Node *receiver = pop();
1116 // Null check on self without removing any arguments. The argument
1117 // null check technically happens in the wrong place, which can lead to
1118 // invalid stack traces when string compare is inlined into a method
1119 // which handles NullPointerExceptions.
1120 _sp += 2;
1121 receiver = do_null_check(receiver, T_OBJECT);
1122 argument = do_null_check(argument, T_OBJECT);
1123 _sp -= 2;
1124 if (stopped()) {
1125 return true;
1126 }
1128 Node* compare = make_string_method_node(Op_StrComp, receiver, argument);
1129 push(compare);
1130 return true;
1131 }
1133 //------------------------------inline_string_equals------------------------
1134 bool LibraryCallKit::inline_string_equals() {
1136 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
1138 int nargs = 2;
1139 _sp += nargs;
1140 Node* argument = pop(); // pop non-receiver first: it was pushed second
1141 Node* receiver = pop();
1143 // Null check on self without removing any arguments. The argument
1144 // null check technically happens in the wrong place, which can lead to
1145 // invalid stack traces when string compare is inlined into a method
1146 // which handles NullPointerExceptions.
1147 _sp += nargs;
1148 receiver = do_null_check(receiver, T_OBJECT);
1149 //should not do null check for argument for String.equals(), because spec
1150 //allows to specify NULL as argument.
1151 _sp -= nargs;
1153 if (stopped()) {
1154 return true;
1155 }
1157 // paths (plus control) merge
1158 RegionNode* region = new (C) RegionNode(5);
1159 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
1161 // does source == target string?
1162 Node* cmp = _gvn.transform(new (C) CmpPNode(receiver, argument));
1163 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::eq));
1165 Node* if_eq = generate_slow_guard(bol, NULL);
1166 if (if_eq != NULL) {
1167 // receiver == argument
1168 phi->init_req(2, intcon(1));
1169 region->init_req(2, if_eq);
1170 }
1172 // get String klass for instanceOf
1173 ciInstanceKlass* klass = env()->String_klass();
1175 if (!stopped()) {
1176 _sp += nargs; // gen_instanceof might do an uncommon trap
1177 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1178 _sp -= nargs;
1179 Node* cmp = _gvn.transform(new (C) CmpINode(inst, intcon(1)));
1180 Node* bol = _gvn.transform(new (C) BoolNode(cmp, BoolTest::ne));
1182 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1183 //instanceOf == true, fallthrough
1185 if (inst_false != NULL) {
1186 phi->init_req(3, intcon(0));
1187 region->init_req(3, inst_false);
1188 }
1189 }
1191 if (!stopped()) {
1192 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1194 // Properly cast the argument to String
1195 argument = _gvn.transform(new (C) CheckCastPPNode(control(), argument, string_type));
1196 // This path is taken only when argument's type is String:NotNull.
1197 argument = cast_not_null(argument, false);
1199 Node* no_ctrl = NULL;
1201 // Get start addr of receiver
1202 Node* receiver_val = load_String_value(no_ctrl, receiver);
1203 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1204 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1206 // Get length of receiver
1207 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1209 // Get start addr of argument
1210 Node* argument_val = load_String_value(no_ctrl, argument);
1211 Node* argument_offset = load_String_offset(no_ctrl, argument);
1212 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1214 // Get length of argument
1215 Node* argument_cnt = load_String_length(no_ctrl, argument);
1217 // Check for receiver count != argument count
1218 Node* cmp = _gvn.transform( new(C) CmpINode(receiver_cnt, argument_cnt) );
1219 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::ne) );
1220 Node* if_ne = generate_slow_guard(bol, NULL);
1221 if (if_ne != NULL) {
1222 phi->init_req(4, intcon(0));
1223 region->init_req(4, if_ne);
1224 }
1226 // Check for count == 0 is done by assembler code for StrEquals.
1228 if (!stopped()) {
1229 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1230 phi->init_req(1, equals);
1231 region->init_req(1, control());
1232 }
1233 }
1235 // post merge
1236 set_control(_gvn.transform(region));
1237 record_for_igvn(region);
1239 push(_gvn.transform(phi));
1241 return true;
1242 }
1244 //------------------------------inline_array_equals----------------------------
1245 bool LibraryCallKit::inline_array_equals() {
1247 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1249 _sp += 2;
1250 Node *argument2 = pop();
1251 Node *argument1 = pop();
1253 Node* equals =
1254 _gvn.transform(new (C) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1255 argument1, argument2) );
1256 push(equals);
1257 return true;
1258 }
1260 // Java version of String.indexOf(constant string)
1261 // class StringDecl {
1262 // StringDecl(char[] ca) {
1263 // offset = 0;
1264 // count = ca.length;
1265 // value = ca;
1266 // }
1267 // int offset;
1268 // int count;
1269 // char[] value;
1270 // }
1271 //
1272 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1273 // int targetOffset, int cache_i, int md2) {
1274 // int cache = cache_i;
1275 // int sourceOffset = string_object.offset;
1276 // int sourceCount = string_object.count;
1277 // int targetCount = target_object.length;
1278 //
1279 // int targetCountLess1 = targetCount - 1;
1280 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1281 //
1282 // char[] source = string_object.value;
1283 // char[] target = target_object;
1284 // int lastChar = target[targetCountLess1];
1285 //
1286 // outer_loop:
1287 // for (int i = sourceOffset; i < sourceEnd; ) {
1288 // int src = source[i + targetCountLess1];
1289 // if (src == lastChar) {
1290 // // With random strings and a 4-character alphabet,
1291 // // reverse matching at this point sets up 0.8% fewer
1292 // // frames, but (paradoxically) makes 0.3% more probes.
1293 // // Since those probes are nearer the lastChar probe,
1294 // // there is may be a net D$ win with reverse matching.
1295 // // But, reversing loop inhibits unroll of inner loop
1296 // // for unknown reason. So, does running outer loop from
1297 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1298 // for (int j = 0; j < targetCountLess1; j++) {
1299 // if (target[targetOffset + j] != source[i+j]) {
1300 // if ((cache & (1 << source[i+j])) == 0) {
1301 // if (md2 < j+1) {
1302 // i += j+1;
1303 // continue outer_loop;
1304 // }
1305 // }
1306 // i += md2;
1307 // continue outer_loop;
1308 // }
1309 // }
1310 // return i - sourceOffset;
1311 // }
1312 // if ((cache & (1 << src)) == 0) {
1313 // i += targetCountLess1;
1314 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1315 // i++;
1316 // }
1317 // return -1;
1318 // }
1320 //------------------------------string_indexOf------------------------
1321 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1322 jint cache_i, jint md2_i) {
1324 Node* no_ctrl = NULL;
1325 float likely = PROB_LIKELY(0.9);
1326 float unlikely = PROB_UNLIKELY(0.9);
1328 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1330 Node* source = load_String_value(no_ctrl, string_object);
1331 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1332 Node* sourceCount = load_String_length(no_ctrl, string_object);
1334 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1335 jint target_length = target_array->length();
1336 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1337 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1339 IdealKit kit(this, false, true);
1340 #define __ kit.
1341 Node* zero = __ ConI(0);
1342 Node* one = __ ConI(1);
1343 Node* cache = __ ConI(cache_i);
1344 Node* md2 = __ ConI(md2_i);
1345 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1346 Node* targetCount = __ ConI(target_length);
1347 Node* targetCountLess1 = __ ConI(target_length - 1);
1348 Node* targetOffset = __ ConI(targetOffset_i);
1349 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1351 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1352 Node* outer_loop = __ make_label(2 /* goto */);
1353 Node* return_ = __ make_label(1);
1355 __ set(rtn,__ ConI(-1));
1356 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1357 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1358 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1359 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1360 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1361 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1362 Node* tpj = __ AddI(targetOffset, __ value(j));
1363 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1364 Node* ipj = __ AddI(__ value(i), __ value(j));
1365 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1366 __ if_then(targ, BoolTest::ne, src2); {
1367 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1368 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1369 __ increment(i, __ AddI(__ value(j), one));
1370 __ goto_(outer_loop);
1371 } __ end_if(); __ dead(j);
1372 }__ end_if(); __ dead(j);
1373 __ increment(i, md2);
1374 __ goto_(outer_loop);
1375 }__ end_if();
1376 __ increment(j, one);
1377 }__ end_loop(); __ dead(j);
1378 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1379 __ goto_(return_);
1380 }__ end_if();
1381 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1382 __ increment(i, targetCountLess1);
1383 }__ end_if();
1384 __ increment(i, one);
1385 __ bind(outer_loop);
1386 }__ end_loop(); __ dead(i);
1387 __ bind(return_);
1389 // Final sync IdealKit and GraphKit.
1390 final_sync(kit);
1391 Node* result = __ value(rtn);
1392 #undef __
1393 C->set_has_loops(true);
1394 return result;
1395 }
1397 //------------------------------inline_string_indexOf------------------------
1398 bool LibraryCallKit::inline_string_indexOf() {
1400 _sp += 2;
1401 Node *argument = pop(); // pop non-receiver first: it was pushed second
1402 Node *receiver = pop();
1404 Node* result;
1405 // Disable the use of pcmpestri until it can be guaranteed that
1406 // the load doesn't cross into the uncommited space.
1407 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1408 UseSSE42Intrinsics) {
1409 // Generate SSE4.2 version of indexOf
1410 // We currently only have match rules that use SSE4.2
1412 // Null check on self without removing any arguments. The argument
1413 // null check technically happens in the wrong place, which can lead to
1414 // invalid stack traces when string compare is inlined into a method
1415 // which handles NullPointerExceptions.
1416 _sp += 2;
1417 receiver = do_null_check(receiver, T_OBJECT);
1418 argument = do_null_check(argument, T_OBJECT);
1419 _sp -= 2;
1421 if (stopped()) {
1422 return true;
1423 }
1425 ciInstanceKlass* str_klass = env()->String_klass();
1426 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1428 // Make the merge point
1429 RegionNode* result_rgn = new (C) RegionNode(4);
1430 Node* result_phi = new (C) PhiNode(result_rgn, TypeInt::INT);
1431 Node* no_ctrl = NULL;
1433 // Get start addr of source string
1434 Node* source = load_String_value(no_ctrl, receiver);
1435 Node* source_offset = load_String_offset(no_ctrl, receiver);
1436 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1438 // Get length of source string
1439 Node* source_cnt = load_String_length(no_ctrl, receiver);
1441 // Get start addr of substring
1442 Node* substr = load_String_value(no_ctrl, argument);
1443 Node* substr_offset = load_String_offset(no_ctrl, argument);
1444 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1446 // Get length of source string
1447 Node* substr_cnt = load_String_length(no_ctrl, argument);
1449 // Check for substr count > string count
1450 Node* cmp = _gvn.transform( new(C) CmpINode(substr_cnt, source_cnt) );
1451 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::gt) );
1452 Node* if_gt = generate_slow_guard(bol, NULL);
1453 if (if_gt != NULL) {
1454 result_phi->init_req(2, intcon(-1));
1455 result_rgn->init_req(2, if_gt);
1456 }
1458 if (!stopped()) {
1459 // Check for substr count == 0
1460 cmp = _gvn.transform( new(C) CmpINode(substr_cnt, intcon(0)) );
1461 bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
1462 Node* if_zero = generate_slow_guard(bol, NULL);
1463 if (if_zero != NULL) {
1464 result_phi->init_req(3, intcon(0));
1465 result_rgn->init_req(3, if_zero);
1466 }
1467 }
1469 if (!stopped()) {
1470 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1471 result_phi->init_req(1, result);
1472 result_rgn->init_req(1, control());
1473 }
1474 set_control(_gvn.transform(result_rgn));
1475 record_for_igvn(result_rgn);
1476 result = _gvn.transform(result_phi);
1478 } else { // Use LibraryCallKit::string_indexOf
1479 // don't intrinsify if argument isn't a constant string.
1480 if (!argument->is_Con()) {
1481 return false;
1482 }
1483 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1484 if (str_type == NULL) {
1485 return false;
1486 }
1487 ciInstanceKlass* klass = env()->String_klass();
1488 ciObject* str_const = str_type->const_oop();
1489 if (str_const == NULL || str_const->klass() != klass) {
1490 return false;
1491 }
1492 ciInstance* str = str_const->as_instance();
1493 assert(str != NULL, "must be instance");
1495 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1496 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1498 int o;
1499 int c;
1500 if (java_lang_String::has_offset_field()) {
1501 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1502 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1503 } else {
1504 o = 0;
1505 c = pat->length();
1506 }
1508 // constant strings have no offset and count == length which
1509 // simplifies the resulting code somewhat so lets optimize for that.
1510 if (o != 0 || c != pat->length()) {
1511 return false;
1512 }
1514 // Null check on self without removing any arguments. The argument
1515 // null check technically happens in the wrong place, which can lead to
1516 // invalid stack traces when string compare is inlined into a method
1517 // which handles NullPointerExceptions.
1518 _sp += 2;
1519 receiver = do_null_check(receiver, T_OBJECT);
1520 // No null check on the argument is needed since it's a constant String oop.
1521 _sp -= 2;
1522 if (stopped()) {
1523 return true;
1524 }
1526 // The null string as a pattern always returns 0 (match at beginning of string)
1527 if (c == 0) {
1528 push(intcon(0));
1529 return true;
1530 }
1532 // Generate default indexOf
1533 jchar lastChar = pat->char_at(o + (c - 1));
1534 int cache = 0;
1535 int i;
1536 for (i = 0; i < c - 1; i++) {
1537 assert(i < pat->length(), "out of range");
1538 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1539 }
1541 int md2 = c;
1542 for (i = 0; i < c - 1; i++) {
1543 assert(i < pat->length(), "out of range");
1544 if (pat->char_at(o + i) == lastChar) {
1545 md2 = (c - 1) - i;
1546 }
1547 }
1549 result = string_indexOf(receiver, pat, o, cache, md2);
1550 }
1552 push(result);
1553 return true;
1554 }
1556 //--------------------------pop_math_arg--------------------------------
1557 // Pop a double argument to a math function from the stack
1558 // rounding it if necessary.
1559 Node * LibraryCallKit::pop_math_arg() {
1560 Node *arg = pop_pair();
1561 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1562 arg = _gvn.transform( new (C) RoundDoubleNode(0, arg) );
1563 return arg;
1564 }
1566 //------------------------------inline_trig----------------------------------
1567 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1568 // argument reduction which will turn into a fast/slow diamond.
1569 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1570 _sp += arg_size(); // restore stack pointer
1571 Node* arg = pop_math_arg();
1572 Node* trig = NULL;
1574 switch (id) {
1575 case vmIntrinsics::_dsin:
1576 trig = _gvn.transform((Node*)new (C) SinDNode(arg));
1577 break;
1578 case vmIntrinsics::_dcos:
1579 trig = _gvn.transform((Node*)new (C) CosDNode(arg));
1580 break;
1581 case vmIntrinsics::_dtan:
1582 trig = _gvn.transform((Node*)new (C) TanDNode(arg));
1583 break;
1584 default:
1585 assert(false, "bad intrinsic was passed in");
1586 return false;
1587 }
1589 // Rounding required? Check for argument reduction!
1590 if( Matcher::strict_fp_requires_explicit_rounding ) {
1592 static const double pi_4 = 0.7853981633974483;
1593 static const double neg_pi_4 = -0.7853981633974483;
1594 // pi/2 in 80-bit extended precision
1595 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1596 // -pi/2 in 80-bit extended precision
1597 // 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};
1598 // Cutoff value for using this argument reduction technique
1599 //static const double pi_2_minus_epsilon = 1.564660403643354;
1600 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1602 // Pseudocode for sin:
1603 // if (x <= Math.PI / 4.0) {
1604 // if (x >= -Math.PI / 4.0) return fsin(x);
1605 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1606 // } else {
1607 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1608 // }
1609 // return StrictMath.sin(x);
1611 // Pseudocode for cos:
1612 // if (x <= Math.PI / 4.0) {
1613 // if (x >= -Math.PI / 4.0) return fcos(x);
1614 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1615 // } else {
1616 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1617 // }
1618 // return StrictMath.cos(x);
1620 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1621 // requires a special machine instruction to load it. Instead we'll try
1622 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1623 // probably do the math inside the SIN encoding.
1625 // Make the merge point
1626 RegionNode *r = new (C) RegionNode(3);
1627 Node *phi = new (C) PhiNode(r,Type::DOUBLE);
1629 // Flatten arg so we need only 1 test
1630 Node *abs = _gvn.transform(new (C) AbsDNode(arg));
1631 // Node for PI/4 constant
1632 Node *pi4 = makecon(TypeD::make(pi_4));
1633 // Check PI/4 : abs(arg)
1634 Node *cmp = _gvn.transform(new (C) CmpDNode(pi4,abs));
1635 // Check: If PI/4 < abs(arg) then go slow
1636 Node *bol = _gvn.transform( new (C) BoolNode( cmp, BoolTest::lt ) );
1637 // Branch either way
1638 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1639 set_control(opt_iff(r,iff));
1641 // Set fast path result
1642 phi->init_req(2,trig);
1644 // Slow path - non-blocking leaf call
1645 Node* call = NULL;
1646 switch (id) {
1647 case vmIntrinsics::_dsin:
1648 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1649 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1650 "Sin", NULL, arg, top());
1651 break;
1652 case vmIntrinsics::_dcos:
1653 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1654 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1655 "Cos", NULL, arg, top());
1656 break;
1657 case vmIntrinsics::_dtan:
1658 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1659 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1660 "Tan", NULL, arg, top());
1661 break;
1662 }
1663 assert(control()->in(0) == call, "");
1664 Node* slow_result = _gvn.transform(new (C) ProjNode(call,TypeFunc::Parms));
1665 r->init_req(1,control());
1666 phi->init_req(1,slow_result);
1668 // Post-merge
1669 set_control(_gvn.transform(r));
1670 record_for_igvn(r);
1671 trig = _gvn.transform(phi);
1673 C->set_has_split_ifs(true); // Has chance for split-if optimization
1674 }
1675 // Push result back on JVM stack
1676 push_pair(trig);
1677 return true;
1678 }
1680 //------------------------------inline_sqrt-------------------------------------
1681 // Inline square root instruction, if possible.
1682 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1683 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1684 _sp += arg_size(); // restore stack pointer
1685 push_pair(_gvn.transform(new (C) SqrtDNode(0, pop_math_arg())));
1686 return true;
1687 }
1689 //------------------------------inline_abs-------------------------------------
1690 // Inline absolute value instruction, if possible.
1691 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1692 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1693 _sp += arg_size(); // restore stack pointer
1694 push_pair(_gvn.transform(new (C) AbsDNode(pop_math_arg())));
1695 return true;
1696 }
1698 void LibraryCallKit::finish_pow_exp(Node* result, Node* x, Node* y, const TypeFunc* call_type, address funcAddr, const char* funcName) {
1699 //-------------------
1700 //result=(result.isNaN())? funcAddr():result;
1701 // Check: If isNaN() by checking result!=result? then either trap
1702 // or go to runtime
1703 Node* cmpisnan = _gvn.transform(new (C) CmpDNode(result,result));
1704 // Build the boolean node
1705 Node* bolisnum = _gvn.transform( new (C) BoolNode(cmpisnan, BoolTest::eq) );
1707 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1708 {
1709 BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1710 // End the current control-flow path
1711 push_pair(x);
1712 if (y != NULL) {
1713 push_pair(y);
1714 }
1715 // The pow or exp intrinsic returned a NaN, which requires a call
1716 // to the runtime. Recompile with the runtime call.
1717 uncommon_trap(Deoptimization::Reason_intrinsic,
1718 Deoptimization::Action_make_not_entrant);
1719 }
1720 push_pair(result);
1721 } else {
1722 // If this inlining ever returned NaN in the past, we compile a call
1723 // to the runtime to properly handle corner cases
1725 IfNode* iff = create_and_xform_if(control(), bolisnum, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1726 Node* if_slow = _gvn.transform( new (C) IfFalseNode(iff) );
1727 Node* if_fast = _gvn.transform( new (C) IfTrueNode(iff) );
1729 if (!if_slow->is_top()) {
1730 RegionNode* result_region = new(C) RegionNode(3);
1731 PhiNode* result_val = new (C) PhiNode(result_region, Type::DOUBLE);
1733 result_region->init_req(1, if_fast);
1734 result_val->init_req(1, result);
1736 set_control(if_slow);
1738 const TypePtr* no_memory_effects = NULL;
1739 Node* rt = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1740 no_memory_effects,
1741 x, top(), y, y ? top() : NULL);
1742 Node* value = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+0));
1743 #ifdef ASSERT
1744 Node* value_top = _gvn.transform(new (C) ProjNode(rt, TypeFunc::Parms+1));
1745 assert(value_top == top(), "second value must be top");
1746 #endif
1748 result_region->init_req(2, control());
1749 result_val->init_req(2, value);
1750 push_result(result_region, result_val);
1751 } else {
1752 push_pair(result);
1753 }
1754 }
1755 }
1757 //------------------------------inline_exp-------------------------------------
1758 // Inline exp instructions, if possible. The Intel hardware only misses
1759 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1760 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1761 assert(id == vmIntrinsics::_dexp, "Not exp");
1763 _sp += arg_size(); // restore stack pointer
1764 Node *x = pop_math_arg();
1765 Node *result = _gvn.transform(new (C) ExpDNode(0,x));
1767 finish_pow_exp(result, x, NULL, OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1769 C->set_has_split_ifs(true); // Has chance for split-if optimization
1771 return true;
1772 }
1774 //------------------------------inline_pow-------------------------------------
1775 // Inline power instructions, if possible.
1776 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1777 assert(id == vmIntrinsics::_dpow, "Not pow");
1779 // Pseudocode for pow
1780 // if (x <= 0.0) {
1781 // long longy = (long)y;
1782 // if ((double)longy == y) { // if y is long
1783 // if (y + 1 == y) longy = 0; // huge number: even
1784 // result = ((1&longy) == 0)?-DPow(abs(x), y):DPow(abs(x), y);
1785 // } else {
1786 // result = NaN;
1787 // }
1788 // } else {
1789 // result = DPow(x,y);
1790 // }
1791 // if (result != result)? {
1792 // result = uncommon_trap() or runtime_call();
1793 // }
1794 // return result;
1796 _sp += arg_size(); // restore stack pointer
1797 Node* y = pop_math_arg();
1798 Node* x = pop_math_arg();
1800 Node* result = NULL;
1802 if (!too_many_traps(Deoptimization::Reason_intrinsic)) {
1803 // Short form: skip the fancy tests and just check for NaN result.
1804 result = _gvn.transform( new (C) PowDNode(0, x, y) );
1805 } else {
1806 // If this inlining ever returned NaN in the past, include all
1807 // checks + call to the runtime.
1809 // Set the merge point for If node with condition of (x <= 0.0)
1810 // There are four possible paths to region node and phi node
1811 RegionNode *r = new (C) RegionNode(4);
1812 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1814 // Build the first if node: if (x <= 0.0)
1815 // Node for 0 constant
1816 Node *zeronode = makecon(TypeD::ZERO);
1817 // Check x:0
1818 Node *cmp = _gvn.transform(new (C) CmpDNode(x, zeronode));
1819 // Check: If (x<=0) then go complex path
1820 Node *bol1 = _gvn.transform( new (C) BoolNode( cmp, BoolTest::le ) );
1821 // Branch either way
1822 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1823 // Fast path taken; set region slot 3
1824 Node *fast_taken = _gvn.transform( new (C) IfFalseNode(if1) );
1825 r->init_req(3,fast_taken); // Capture fast-control
1827 // Fast path not-taken, i.e. slow path
1828 Node *complex_path = _gvn.transform( new (C) IfTrueNode(if1) );
1830 // Set fast path result
1831 Node *fast_result = _gvn.transform( new (C) PowDNode(0, x, y) );
1832 phi->init_req(3, fast_result);
1834 // Complex path
1835 // Build the second if node (if y is long)
1836 // Node for (long)y
1837 Node *longy = _gvn.transform( new (C) ConvD2LNode(y));
1838 // Node for (double)((long) y)
1839 Node *doublelongy= _gvn.transform( new (C) ConvL2DNode(longy));
1840 // Check (double)((long) y) : y
1841 Node *cmplongy= _gvn.transform(new (C) CmpDNode(doublelongy, y));
1842 // Check if (y isn't long) then go to slow path
1844 Node *bol2 = _gvn.transform( new (C) BoolNode( cmplongy, BoolTest::ne ) );
1845 // Branch either way
1846 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1847 Node* ylong_path = _gvn.transform( new (C) IfFalseNode(if2));
1849 Node *slow_path = _gvn.transform( new (C) IfTrueNode(if2) );
1851 // Calculate DPow(abs(x), y)*(1 & (long)y)
1852 // Node for constant 1
1853 Node *conone = longcon(1);
1854 // 1& (long)y
1855 Node *signnode= _gvn.transform( new (C) AndLNode(conone, longy) );
1857 // A huge number is always even. Detect a huge number by checking
1858 // if y + 1 == y and set integer to be tested for parity to 0.
1859 // Required for corner case:
1860 // (long)9.223372036854776E18 = max_jlong
1861 // (double)(long)9.223372036854776E18 = 9.223372036854776E18
1862 // max_jlong is odd but 9.223372036854776E18 is even
1863 Node* yplus1 = _gvn.transform( new (C) AddDNode(y, makecon(TypeD::make(1))));
1864 Node *cmpyplus1= _gvn.transform(new (C) CmpDNode(yplus1, y));
1865 Node *bolyplus1 = _gvn.transform( new (C) BoolNode( cmpyplus1, BoolTest::eq ) );
1866 Node* correctedsign = NULL;
1867 if (ConditionalMoveLimit != 0) {
1868 correctedsign = _gvn.transform( CMoveNode::make(C, NULL, bolyplus1, signnode, longcon(0), TypeLong::LONG));
1869 } else {
1870 IfNode *ifyplus1 = create_and_xform_if(ylong_path,bolyplus1, PROB_FAIR, COUNT_UNKNOWN);
1871 RegionNode *r = new (C) RegionNode(3);
1872 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
1873 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyplus1)));
1874 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyplus1)));
1875 phi->init_req(1, signnode);
1876 phi->init_req(2, longcon(0));
1877 correctedsign = _gvn.transform(phi);
1878 ylong_path = _gvn.transform(r);
1879 record_for_igvn(r);
1880 }
1882 // zero node
1883 Node *conzero = longcon(0);
1884 // Check (1&(long)y)==0?
1885 Node *cmpeq1 = _gvn.transform(new (C) CmpLNode(correctedsign, conzero));
1886 // Check if (1&(long)y)!=0?, if so the result is negative
1887 Node *bol3 = _gvn.transform( new (C) BoolNode( cmpeq1, BoolTest::ne ) );
1888 // abs(x)
1889 Node *absx=_gvn.transform( new (C) AbsDNode(x));
1890 // abs(x)^y
1891 Node *absxpowy = _gvn.transform( new (C) PowDNode(0, absx, y) );
1892 // -abs(x)^y
1893 Node *negabsxpowy = _gvn.transform(new (C) NegDNode (absxpowy));
1894 // (1&(long)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1895 Node *signresult = NULL;
1896 if (ConditionalMoveLimit != 0) {
1897 signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1898 } else {
1899 IfNode *ifyeven = create_and_xform_if(ylong_path,bol3, PROB_FAIR, COUNT_UNKNOWN);
1900 RegionNode *r = new (C) RegionNode(3);
1901 Node *phi = new (C) PhiNode(r, Type::DOUBLE);
1902 r->init_req(1, _gvn.transform( new (C) IfFalseNode(ifyeven)));
1903 r->init_req(2, _gvn.transform( new (C) IfTrueNode(ifyeven)));
1904 phi->init_req(1, absxpowy);
1905 phi->init_req(2, negabsxpowy);
1906 signresult = _gvn.transform(phi);
1907 ylong_path = _gvn.transform(r);
1908 record_for_igvn(r);
1909 }
1910 // Set complex path fast result
1911 r->init_req(2, ylong_path);
1912 phi->init_req(2, signresult);
1914 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1915 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1916 r->init_req(1,slow_path);
1917 phi->init_req(1,slow_result);
1919 // Post merge
1920 set_control(_gvn.transform(r));
1921 record_for_igvn(r);
1922 result=_gvn.transform(phi);
1923 }
1925 finish_pow_exp(result, x, y, OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1927 C->set_has_split_ifs(true); // Has chance for split-if optimization
1929 return true;
1930 }
1932 //------------------------------inline_trans-------------------------------------
1933 // Inline transcendental instructions, if possible. The Intel hardware gets
1934 // these right, no funny corner cases missed.
1935 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1936 _sp += arg_size(); // restore stack pointer
1937 Node* arg = pop_math_arg();
1938 Node* trans = NULL;
1940 switch (id) {
1941 case vmIntrinsics::_dlog:
1942 trans = _gvn.transform((Node*)new (C) LogDNode(arg));
1943 break;
1944 case vmIntrinsics::_dlog10:
1945 trans = _gvn.transform((Node*)new (C) Log10DNode(arg));
1946 break;
1947 default:
1948 assert(false, "bad intrinsic was passed in");
1949 return false;
1950 }
1952 // Push result back on JVM stack
1953 push_pair(trans);
1954 return true;
1955 }
1957 //------------------------------runtime_math-----------------------------
1958 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1959 Node* a = NULL;
1960 Node* b = NULL;
1962 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1963 "must be (DD)D or (D)D type");
1965 // Inputs
1966 _sp += arg_size(); // restore stack pointer
1967 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1968 b = pop_math_arg();
1969 }
1970 a = pop_math_arg();
1972 const TypePtr* no_memory_effects = NULL;
1973 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1974 no_memory_effects,
1975 a, top(), b, b ? top() : NULL);
1976 Node* value = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+0));
1977 #ifdef ASSERT
1978 Node* value_top = _gvn.transform(new (C) ProjNode(trig, TypeFunc::Parms+1));
1979 assert(value_top == top(), "second value must be top");
1980 #endif
1982 push_pair(value);
1983 return true;
1984 }
1986 //------------------------------inline_math_native-----------------------------
1987 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1988 switch (id) {
1989 // These intrinsics are not properly supported on all hardware
1990 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1991 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1992 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1993 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1994 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1995 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1997 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1998 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1999 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
2000 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
2002 // These intrinsics are supported on all hardware
2003 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
2004 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
2006 case vmIntrinsics::_dexp: return
2007 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) :
2008 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
2009 case vmIntrinsics::_dpow: return
2010 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) :
2011 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
2013 // These intrinsics are not yet correctly implemented
2014 case vmIntrinsics::_datan2:
2015 return false;
2017 default:
2018 ShouldNotReachHere();
2019 return false;
2020 }
2021 }
2023 static bool is_simple_name(Node* n) {
2024 return (n->req() == 1 // constant
2025 || (n->is_Type() && n->as_Type()->type()->singleton())
2026 || n->is_Proj() // parameter or return value
2027 || n->is_Phi() // local of some sort
2028 );
2029 }
2031 //----------------------------inline_min_max-----------------------------------
2032 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
2033 push(generate_min_max(id, argument(0), argument(1)));
2035 return true;
2036 }
2038 Node*
2039 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
2040 // These are the candidate return value:
2041 Node* xvalue = x0;
2042 Node* yvalue = y0;
2044 if (xvalue == yvalue) {
2045 return xvalue;
2046 }
2048 bool want_max = (id == vmIntrinsics::_max);
2050 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
2051 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
2052 if (txvalue == NULL || tyvalue == NULL) return top();
2053 // This is not really necessary, but it is consistent with a
2054 // hypothetical MaxINode::Value method:
2055 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
2057 // %%% This folding logic should (ideally) be in a different place.
2058 // Some should be inside IfNode, and there to be a more reliable
2059 // transformation of ?: style patterns into cmoves. We also want
2060 // more powerful optimizations around cmove and min/max.
2062 // Try to find a dominating comparison of these guys.
2063 // It can simplify the index computation for Arrays.copyOf
2064 // and similar uses of System.arraycopy.
2065 // First, compute the normalized version of CmpI(x, y).
2066 int cmp_op = Op_CmpI;
2067 Node* xkey = xvalue;
2068 Node* ykey = yvalue;
2069 Node* ideal_cmpxy = _gvn.transform( new(C) CmpINode(xkey, ykey) );
2070 if (ideal_cmpxy->is_Cmp()) {
2071 // E.g., if we have CmpI(length - offset, count),
2072 // it might idealize to CmpI(length, count + offset)
2073 cmp_op = ideal_cmpxy->Opcode();
2074 xkey = ideal_cmpxy->in(1);
2075 ykey = ideal_cmpxy->in(2);
2076 }
2078 // Start by locating any relevant comparisons.
2079 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
2080 Node* cmpxy = NULL;
2081 Node* cmpyx = NULL;
2082 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
2083 Node* cmp = start_from->fast_out(k);
2084 if (cmp->outcnt() > 0 && // must have prior uses
2085 cmp->in(0) == NULL && // must be context-independent
2086 cmp->Opcode() == cmp_op) { // right kind of compare
2087 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
2088 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
2089 }
2090 }
2092 const int NCMPS = 2;
2093 Node* cmps[NCMPS] = { cmpxy, cmpyx };
2094 int cmpn;
2095 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2096 if (cmps[cmpn] != NULL) break; // find a result
2097 }
2098 if (cmpn < NCMPS) {
2099 // Look for a dominating test that tells us the min and max.
2100 int depth = 0; // Limit search depth for speed
2101 Node* dom = control();
2102 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
2103 if (++depth >= 100) break;
2104 Node* ifproj = dom;
2105 if (!ifproj->is_Proj()) continue;
2106 Node* iff = ifproj->in(0);
2107 if (!iff->is_If()) continue;
2108 Node* bol = iff->in(1);
2109 if (!bol->is_Bool()) continue;
2110 Node* cmp = bol->in(1);
2111 if (cmp == NULL) continue;
2112 for (cmpn = 0; cmpn < NCMPS; cmpn++)
2113 if (cmps[cmpn] == cmp) break;
2114 if (cmpn == NCMPS) continue;
2115 BoolTest::mask btest = bol->as_Bool()->_test._test;
2116 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
2117 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2118 // At this point, we know that 'x btest y' is true.
2119 switch (btest) {
2120 case BoolTest::eq:
2121 // They are proven equal, so we can collapse the min/max.
2122 // Either value is the answer. Choose the simpler.
2123 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
2124 return yvalue;
2125 return xvalue;
2126 case BoolTest::lt: // x < y
2127 case BoolTest::le: // x <= y
2128 return (want_max ? yvalue : xvalue);
2129 case BoolTest::gt: // x > y
2130 case BoolTest::ge: // x >= y
2131 return (want_max ? xvalue : yvalue);
2132 }
2133 }
2134 }
2136 // We failed to find a dominating test.
2137 // Let's pick a test that might GVN with prior tests.
2138 Node* best_bol = NULL;
2139 BoolTest::mask best_btest = BoolTest::illegal;
2140 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
2141 Node* cmp = cmps[cmpn];
2142 if (cmp == NULL) continue;
2143 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
2144 Node* bol = cmp->fast_out(j);
2145 if (!bol->is_Bool()) continue;
2146 BoolTest::mask btest = bol->as_Bool()->_test._test;
2147 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
2148 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
2149 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
2150 best_bol = bol->as_Bool();
2151 best_btest = btest;
2152 }
2153 }
2154 }
2156 Node* answer_if_true = NULL;
2157 Node* answer_if_false = NULL;
2158 switch (best_btest) {
2159 default:
2160 if (cmpxy == NULL)
2161 cmpxy = ideal_cmpxy;
2162 best_bol = _gvn.transform( new(C) BoolNode(cmpxy, BoolTest::lt) );
2163 // and fall through:
2164 case BoolTest::lt: // x < y
2165 case BoolTest::le: // x <= y
2166 answer_if_true = (want_max ? yvalue : xvalue);
2167 answer_if_false = (want_max ? xvalue : yvalue);
2168 break;
2169 case BoolTest::gt: // x > y
2170 case BoolTest::ge: // x >= y
2171 answer_if_true = (want_max ? xvalue : yvalue);
2172 answer_if_false = (want_max ? yvalue : xvalue);
2173 break;
2174 }
2176 jint hi, lo;
2177 if (want_max) {
2178 // We can sharpen the minimum.
2179 hi = MAX2(txvalue->_hi, tyvalue->_hi);
2180 lo = MAX2(txvalue->_lo, tyvalue->_lo);
2181 } else {
2182 // We can sharpen the maximum.
2183 hi = MIN2(txvalue->_hi, tyvalue->_hi);
2184 lo = MIN2(txvalue->_lo, tyvalue->_lo);
2185 }
2187 // Use a flow-free graph structure, to avoid creating excess control edges
2188 // which could hinder other optimizations.
2189 // Since Math.min/max is often used with arraycopy, we want
2190 // tightly_coupled_allocation to be able to see beyond min/max expressions.
2191 Node* cmov = CMoveNode::make(C, NULL, best_bol,
2192 answer_if_false, answer_if_true,
2193 TypeInt::make(lo, hi, widen));
2195 return _gvn.transform(cmov);
2197 /*
2198 // This is not as desirable as it may seem, since Min and Max
2199 // nodes do not have a full set of optimizations.
2200 // And they would interfere, anyway, with 'if' optimizations
2201 // and with CMoveI canonical forms.
2202 switch (id) {
2203 case vmIntrinsics::_min:
2204 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2205 case vmIntrinsics::_max:
2206 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2207 default:
2208 ShouldNotReachHere();
2209 }
2210 */
2211 }
2213 inline int
2214 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2215 const TypePtr* base_type = TypePtr::NULL_PTR;
2216 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2217 if (base_type == NULL) {
2218 // Unknown type.
2219 return Type::AnyPtr;
2220 } else if (base_type == TypePtr::NULL_PTR) {
2221 // Since this is a NULL+long form, we have to switch to a rawptr.
2222 base = _gvn.transform( new (C) CastX2PNode(offset) );
2223 offset = MakeConX(0);
2224 return Type::RawPtr;
2225 } else if (base_type->base() == Type::RawPtr) {
2226 return Type::RawPtr;
2227 } else if (base_type->isa_oopptr()) {
2228 // Base is never null => always a heap address.
2229 if (base_type->ptr() == TypePtr::NotNull) {
2230 return Type::OopPtr;
2231 }
2232 // Offset is small => always a heap address.
2233 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2234 if (offset_type != NULL &&
2235 base_type->offset() == 0 && // (should always be?)
2236 offset_type->_lo >= 0 &&
2237 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2238 return Type::OopPtr;
2239 }
2240 // Otherwise, it might either be oop+off or NULL+addr.
2241 return Type::AnyPtr;
2242 } else {
2243 // No information:
2244 return Type::AnyPtr;
2245 }
2246 }
2248 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2249 int kind = classify_unsafe_addr(base, offset);
2250 if (kind == Type::RawPtr) {
2251 return basic_plus_adr(top(), base, offset);
2252 } else {
2253 return basic_plus_adr(base, offset);
2254 }
2255 }
2257 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2258 // inline int Integer.numberOfLeadingZeros(int)
2259 // inline int Long.numberOfLeadingZeros(long)
2260 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2261 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2262 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2263 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2264 _sp += arg_size(); // restore stack pointer
2265 switch (id) {
2266 case vmIntrinsics::_numberOfLeadingZeros_i:
2267 push(_gvn.transform(new (C) CountLeadingZerosINode(pop())));
2268 break;
2269 case vmIntrinsics::_numberOfLeadingZeros_l:
2270 push(_gvn.transform(new (C) CountLeadingZerosLNode(pop_pair())));
2271 break;
2272 default:
2273 ShouldNotReachHere();
2274 }
2275 return true;
2276 }
2278 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2279 // inline int Integer.numberOfTrailingZeros(int)
2280 // inline int Long.numberOfTrailingZeros(long)
2281 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2282 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2283 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2284 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2285 _sp += arg_size(); // restore stack pointer
2286 switch (id) {
2287 case vmIntrinsics::_numberOfTrailingZeros_i:
2288 push(_gvn.transform(new (C) CountTrailingZerosINode(pop())));
2289 break;
2290 case vmIntrinsics::_numberOfTrailingZeros_l:
2291 push(_gvn.transform(new (C) CountTrailingZerosLNode(pop_pair())));
2292 break;
2293 default:
2294 ShouldNotReachHere();
2295 }
2296 return true;
2297 }
2299 //----------------------------inline_bitCount_int/long-----------------------
2300 // inline int Integer.bitCount(int)
2301 // inline int Long.bitCount(long)
2302 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2303 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2304 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2305 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2306 _sp += arg_size(); // restore stack pointer
2307 switch (id) {
2308 case vmIntrinsics::_bitCount_i:
2309 push(_gvn.transform(new (C) PopCountINode(pop())));
2310 break;
2311 case vmIntrinsics::_bitCount_l:
2312 push(_gvn.transform(new (C) PopCountLNode(pop_pair())));
2313 break;
2314 default:
2315 ShouldNotReachHere();
2316 }
2317 return true;
2318 }
2320 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2321 // inline Integer.reverseBytes(int)
2322 // inline Long.reverseBytes(long)
2323 // inline Character.reverseBytes(char)
2324 // inline Short.reverseBytes(short)
2325 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2326 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2327 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2328 "not reverse Bytes");
2329 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2330 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2331 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2332 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2333 _sp += arg_size(); // restore stack pointer
2334 switch (id) {
2335 case vmIntrinsics::_reverseBytes_i:
2336 push(_gvn.transform(new (C) ReverseBytesINode(0, pop())));
2337 break;
2338 case vmIntrinsics::_reverseBytes_l:
2339 push_pair(_gvn.transform(new (C) ReverseBytesLNode(0, pop_pair())));
2340 break;
2341 case vmIntrinsics::_reverseBytes_c:
2342 push(_gvn.transform(new (C) ReverseBytesUSNode(0, pop())));
2343 break;
2344 case vmIntrinsics::_reverseBytes_s:
2345 push(_gvn.transform(new (C) ReverseBytesSNode(0, pop())));
2346 break;
2347 default:
2348 ;
2349 }
2350 return true;
2351 }
2353 //----------------------------inline_unsafe_access----------------------------
2355 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2357 // Helper that guards and inserts a pre-barrier.
2358 void LibraryCallKit::insert_pre_barrier(Node* base_oop, Node* offset,
2359 Node* pre_val, int nargs, bool need_mem_bar) {
2360 // We could be accessing the referent field of a reference object. If so, when G1
2361 // is enabled, we need to log the value in the referent field in an SATB buffer.
2362 // This routine performs some compile time filters and generates suitable
2363 // runtime filters that guard the pre-barrier code.
2364 // Also add memory barrier for non volatile load from the referent field
2365 // to prevent commoning of loads across safepoint.
2366 if (!UseG1GC && !need_mem_bar)
2367 return;
2369 // Some compile time checks.
2371 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2372 const TypeX* otype = offset->find_intptr_t_type();
2373 if (otype != NULL && otype->is_con() &&
2374 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2375 // Constant offset but not the reference_offset so just return
2376 return;
2377 }
2379 // We only need to generate the runtime guards for instances.
2380 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2381 if (btype != NULL) {
2382 if (btype->isa_aryptr()) {
2383 // Array type so nothing to do
2384 return;
2385 }
2387 const TypeInstPtr* itype = btype->isa_instptr();
2388 if (itype != NULL) {
2389 // Can the klass of base_oop be statically determined to be
2390 // _not_ a sub-class of Reference and _not_ Object?
2391 ciKlass* klass = itype->klass();
2392 if ( klass->is_loaded() &&
2393 !klass->is_subtype_of(env()->Reference_klass()) &&
2394 !env()->Object_klass()->is_subtype_of(klass)) {
2395 return;
2396 }
2397 }
2398 }
2400 // The compile time filters did not reject base_oop/offset so
2401 // we need to generate the following runtime filters
2402 //
2403 // if (offset == java_lang_ref_Reference::_reference_offset) {
2404 // if (instance_of(base, java.lang.ref.Reference)) {
2405 // pre_barrier(_, pre_val, ...);
2406 // }
2407 // }
2409 float likely = PROB_LIKELY(0.999);
2410 float unlikely = PROB_UNLIKELY(0.999);
2412 IdealKit ideal(this);
2413 #define __ ideal.
2415 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2417 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2418 // Update graphKit memory and control from IdealKit.
2419 sync_kit(ideal);
2421 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2422 _sp += nargs; // gen_instanceof might do an uncommon trap
2423 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2424 _sp -= nargs;
2426 // Update IdealKit memory and control from graphKit.
2427 __ sync_kit(this);
2429 Node* one = __ ConI(1);
2430 // is_instof == 0 if base_oop == NULL
2431 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2433 // Update graphKit from IdeakKit.
2434 sync_kit(ideal);
2436 // Use the pre-barrier to record the value in the referent field
2437 pre_barrier(false /* do_load */,
2438 __ ctrl(),
2439 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2440 pre_val /* pre_val */,
2441 T_OBJECT);
2442 if (need_mem_bar) {
2443 // Add memory barrier to prevent commoning reads from this field
2444 // across safepoint since GC can change its value.
2445 insert_mem_bar(Op_MemBarCPUOrder);
2446 }
2447 // Update IdealKit from graphKit.
2448 __ sync_kit(this);
2450 } __ end_if(); // _ref_type != ref_none
2451 } __ end_if(); // offset == referent_offset
2453 // Final sync IdealKit and GraphKit.
2454 final_sync(ideal);
2455 #undef __
2456 }
2459 // Interpret Unsafe.fieldOffset cookies correctly:
2460 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2462 const TypeOopPtr* LibraryCallKit::sharpen_unsafe_type(Compile::AliasType* alias_type, const TypePtr *adr_type, bool is_native_ptr) {
2463 // Attempt to infer a sharper value type from the offset and base type.
2464 ciKlass* sharpened_klass = NULL;
2466 // See if it is an instance field, with an object type.
2467 if (alias_type->field() != NULL) {
2468 assert(!is_native_ptr, "native pointer op cannot use a java address");
2469 if (alias_type->field()->type()->is_klass()) {
2470 sharpened_klass = alias_type->field()->type()->as_klass();
2471 }
2472 }
2474 // See if it is a narrow oop array.
2475 if (adr_type->isa_aryptr()) {
2476 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2477 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2478 if (elem_type != NULL) {
2479 sharpened_klass = elem_type->klass();
2480 }
2481 }
2482 }
2484 // The sharpened class might be unloaded if there is no class loader
2485 // contraint in place.
2486 if (sharpened_klass != NULL && sharpened_klass->is_loaded()) {
2487 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2489 #ifndef PRODUCT
2490 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2491 tty->print(" from base type: "); adr_type->dump();
2492 tty->print(" sharpened value: "); tjp->dump();
2493 }
2494 #endif
2495 // Sharpen the value type.
2496 return tjp;
2497 }
2498 return NULL;
2499 }
2501 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2502 if (callee()->is_static()) return false; // caller must have the capability!
2504 #ifndef PRODUCT
2505 {
2506 ResourceMark rm;
2507 // Check the signatures.
2508 ciSignature* sig = signature();
2509 #ifdef ASSERT
2510 if (!is_store) {
2511 // Object getObject(Object base, int/long offset), etc.
2512 BasicType rtype = sig->return_type()->basic_type();
2513 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2514 rtype = T_ADDRESS; // it is really a C void*
2515 assert(rtype == type, "getter must return the expected value");
2516 if (!is_native_ptr) {
2517 assert(sig->count() == 2, "oop getter has 2 arguments");
2518 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2519 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2520 } else {
2521 assert(sig->count() == 1, "native getter has 1 argument");
2522 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2523 }
2524 } else {
2525 // void putObject(Object base, int/long offset, Object x), etc.
2526 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2527 if (!is_native_ptr) {
2528 assert(sig->count() == 3, "oop putter has 3 arguments");
2529 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2530 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2531 } else {
2532 assert(sig->count() == 2, "native putter has 2 arguments");
2533 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2534 }
2535 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2536 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2537 vtype = T_ADDRESS; // it is really a C void*
2538 assert(vtype == type, "putter must accept the expected value");
2539 }
2540 #endif // ASSERT
2541 }
2542 #endif //PRODUCT
2544 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2546 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2548 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2549 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2550 assert(callee()->arg_size() == nargs, "must be");
2552 debug_only(int saved_sp = _sp);
2553 _sp += nargs;
2555 Node* val;
2556 debug_only(val = (Node*)(uintptr_t)-1);
2559 if (is_store) {
2560 // Get the value being stored. (Pop it first; it was pushed last.)
2561 switch (type) {
2562 case T_DOUBLE:
2563 case T_LONG:
2564 case T_ADDRESS:
2565 val = pop_pair();
2566 break;
2567 default:
2568 val = pop();
2569 }
2570 }
2572 // Build address expression. See the code in inline_unsafe_prefetch.
2573 Node *adr;
2574 Node *heap_base_oop = top();
2575 Node* offset = top();
2577 if (!is_native_ptr) {
2578 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2579 offset = pop_pair();
2580 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2581 Node* base = pop();
2582 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2583 // to be plain byte offsets, which are also the same as those accepted
2584 // by oopDesc::field_base.
2585 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2586 "fieldOffset must be byte-scaled");
2587 // 32-bit machines ignore the high half!
2588 offset = ConvL2X(offset);
2589 adr = make_unsafe_address(base, offset);
2590 heap_base_oop = base;
2591 } else {
2592 Node* ptr = pop_pair();
2593 // Adjust Java long to machine word:
2594 ptr = ConvL2X(ptr);
2595 adr = make_unsafe_address(NULL, ptr);
2596 }
2598 // Pop receiver last: it was pushed first.
2599 Node *receiver = pop();
2601 assert(saved_sp == _sp, "must have correct argument count");
2603 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2605 // First guess at the value type.
2606 const Type *value_type = Type::get_const_basic_type(type);
2608 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2609 // there was not enough information to nail it down.
2610 Compile::AliasType* alias_type = C->alias_type(adr_type);
2611 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2613 // We will need memory barriers unless we can determine a unique
2614 // alias category for this reference. (Note: If for some reason
2615 // the barriers get omitted and the unsafe reference begins to "pollute"
2616 // the alias analysis of the rest of the graph, either Compile::can_alias
2617 // or Compile::must_alias will throw a diagnostic assert.)
2618 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2620 // If we are reading the value of the referent field of a Reference
2621 // object (either by using Unsafe directly or through reflection)
2622 // then, if G1 is enabled, we need to record the referent in an
2623 // SATB log buffer using the pre-barrier mechanism.
2624 // Also we need to add memory barrier to prevent commoning reads
2625 // from this field across safepoint since GC can change its value.
2626 bool need_read_barrier = !is_native_ptr && !is_store &&
2627 offset != top() && heap_base_oop != top();
2629 if (!is_store && type == T_OBJECT) {
2630 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type, is_native_ptr);
2631 if (tjp != NULL) {
2632 value_type = tjp;
2633 }
2634 }
2636 // Null check on self without removing any arguments. The argument
2637 // null check technically happens in the wrong place, which can lead to
2638 // invalid stack traces when the primitive is inlined into a method
2639 // which handles NullPointerExceptions.
2640 _sp += nargs;
2641 do_null_check(receiver, T_OBJECT);
2642 _sp -= nargs;
2643 if (stopped()) {
2644 return true;
2645 }
2646 // Heap pointers get a null-check from the interpreter,
2647 // as a courtesy. However, this is not guaranteed by Unsafe,
2648 // and it is not possible to fully distinguish unintended nulls
2649 // from intended ones in this API.
2651 if (is_volatile) {
2652 // We need to emit leading and trailing CPU membars (see below) in
2653 // addition to memory membars when is_volatile. This is a little
2654 // too strong, but avoids the need to insert per-alias-type
2655 // volatile membars (for stores; compare Parse::do_put_xxx), which
2656 // we cannot do effectively here because we probably only have a
2657 // rough approximation of type.
2658 need_mem_bar = true;
2659 // For Stores, place a memory ordering barrier now.
2660 if (is_store)
2661 insert_mem_bar(Op_MemBarRelease);
2662 }
2664 // Memory barrier to prevent normal and 'unsafe' accesses from
2665 // bypassing each other. Happens after null checks, so the
2666 // exception paths do not take memory state from the memory barrier,
2667 // so there's no problems making a strong assert about mixing users
2668 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2669 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2670 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2672 if (!is_store) {
2673 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2674 // load value and push onto stack
2675 switch (type) {
2676 case T_BOOLEAN:
2677 case T_CHAR:
2678 case T_BYTE:
2679 case T_SHORT:
2680 case T_INT:
2681 case T_FLOAT:
2682 push(p);
2683 break;
2684 case T_OBJECT:
2685 if (need_read_barrier) {
2686 insert_pre_barrier(heap_base_oop, offset, p, nargs, !(is_volatile || need_mem_bar));
2687 }
2688 push(p);
2689 break;
2690 case T_ADDRESS:
2691 // Cast to an int type.
2692 p = _gvn.transform( new (C) CastP2XNode(NULL,p) );
2693 p = ConvX2L(p);
2694 push_pair(p);
2695 break;
2696 case T_DOUBLE:
2697 case T_LONG:
2698 push_pair( p );
2699 break;
2700 default: ShouldNotReachHere();
2701 }
2702 } else {
2703 // place effect of store into memory
2704 switch (type) {
2705 case T_DOUBLE:
2706 val = dstore_rounding(val);
2707 break;
2708 case T_ADDRESS:
2709 // Repackage the long as a pointer.
2710 val = ConvL2X(val);
2711 val = _gvn.transform( new (C) CastX2PNode(val) );
2712 break;
2713 }
2715 if (type != T_OBJECT ) {
2716 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2717 } else {
2718 // Possibly an oop being stored to Java heap or native memory
2719 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2720 // oop to Java heap.
2721 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2722 } else {
2723 // We can't tell at compile time if we are storing in the Java heap or outside
2724 // of it. So we need to emit code to conditionally do the proper type of
2725 // store.
2727 IdealKit ideal(this);
2728 #define __ ideal.
2729 // QQQ who knows what probability is here??
2730 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2731 // Sync IdealKit and graphKit.
2732 sync_kit(ideal);
2733 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2734 // Update IdealKit memory.
2735 __ sync_kit(this);
2736 } __ else_(); {
2737 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2738 } __ end_if();
2739 // Final sync IdealKit and GraphKit.
2740 final_sync(ideal);
2741 #undef __
2742 }
2743 }
2744 }
2746 if (is_volatile) {
2747 if (!is_store)
2748 insert_mem_bar(Op_MemBarAcquire);
2749 else
2750 insert_mem_bar(Op_MemBarVolatile);
2751 }
2753 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2755 return true;
2756 }
2758 //----------------------------inline_unsafe_prefetch----------------------------
2760 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2761 #ifndef PRODUCT
2762 {
2763 ResourceMark rm;
2764 // Check the signatures.
2765 ciSignature* sig = signature();
2766 #ifdef ASSERT
2767 // Object getObject(Object base, int/long offset), etc.
2768 BasicType rtype = sig->return_type()->basic_type();
2769 if (!is_native_ptr) {
2770 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2771 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2772 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2773 } else {
2774 assert(sig->count() == 1, "native prefetch has 1 argument");
2775 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2776 }
2777 #endif // ASSERT
2778 }
2779 #endif // !PRODUCT
2781 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2783 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2784 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2786 debug_only(int saved_sp = _sp);
2787 _sp += nargs;
2789 // Build address expression. See the code in inline_unsafe_access.
2790 Node *adr;
2791 if (!is_native_ptr) {
2792 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2793 Node* offset = pop_pair();
2794 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2795 Node* base = pop();
2796 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2797 // to be plain byte offsets, which are also the same as those accepted
2798 // by oopDesc::field_base.
2799 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2800 "fieldOffset must be byte-scaled");
2801 // 32-bit machines ignore the high half!
2802 offset = ConvL2X(offset);
2803 adr = make_unsafe_address(base, offset);
2804 } else {
2805 Node* ptr = pop_pair();
2806 // Adjust Java long to machine word:
2807 ptr = ConvL2X(ptr);
2808 adr = make_unsafe_address(NULL, ptr);
2809 }
2811 if (is_static) {
2812 assert(saved_sp == _sp, "must have correct argument count");
2813 } else {
2814 // Pop receiver last: it was pushed first.
2815 Node *receiver = pop();
2816 assert(saved_sp == _sp, "must have correct argument count");
2818 // Null check on self without removing any arguments. The argument
2819 // null check technically happens in the wrong place, which can lead to
2820 // invalid stack traces when the primitive is inlined into a method
2821 // which handles NullPointerExceptions.
2822 _sp += nargs;
2823 do_null_check(receiver, T_OBJECT);
2824 _sp -= nargs;
2825 if (stopped()) {
2826 return true;
2827 }
2828 }
2830 // Generate the read or write prefetch
2831 Node *prefetch;
2832 if (is_store) {
2833 prefetch = new (C) PrefetchWriteNode(i_o(), adr);
2834 } else {
2835 prefetch = new (C) PrefetchReadNode(i_o(), adr);
2836 }
2837 prefetch->init_req(0, control());
2838 set_i_o(_gvn.transform(prefetch));
2840 return true;
2841 }
2843 //----------------------------inline_unsafe_load_store----------------------------
2845 bool LibraryCallKit::inline_unsafe_load_store(BasicType type, LoadStoreKind kind) {
2846 // This basic scheme here is the same as inline_unsafe_access, but
2847 // differs in enough details that combining them would make the code
2848 // overly confusing. (This is a true fact! I originally combined
2849 // them, but even I was confused by it!) As much code/comments as
2850 // possible are retained from inline_unsafe_access though to make
2851 // the correspondences clearer. - dl
2853 if (callee()->is_static()) return false; // caller must have the capability!
2855 #ifndef PRODUCT
2856 BasicType rtype;
2857 {
2858 ResourceMark rm;
2859 ciSignature* sig = signature();
2860 rtype = sig->return_type()->basic_type();
2861 if (kind == LS_xadd || kind == LS_xchg) {
2862 // Check the signatures.
2863 #ifdef ASSERT
2864 assert(rtype == type, "get and set must return the expected type");
2865 assert(sig->count() == 3, "get and set has 3 arguments");
2866 assert(sig->type_at(0)->basic_type() == T_OBJECT, "get and set base is object");
2867 assert(sig->type_at(1)->basic_type() == T_LONG, "get and set offset is long");
2868 assert(sig->type_at(2)->basic_type() == type, "get and set must take expected type as new value/delta");
2869 #endif // ASSERT
2870 } else if (kind == LS_cmpxchg) {
2871 // Check the signatures.
2872 #ifdef ASSERT
2873 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2874 assert(sig->count() == 4, "CAS has 4 arguments");
2875 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2876 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2877 #endif // ASSERT
2878 } else {
2879 ShouldNotReachHere();
2880 }
2881 }
2882 #endif //PRODUCT
2884 // number of stack slots per value argument (1 or 2)
2885 int type_words = type2size[type];
2887 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2889 // Argument words: "this" plus oop plus offset (plus oldvalue) plus newvalue/delta;
2890 int nargs = 1 + 1 + 2 + ((kind == LS_cmpxchg) ? type_words : 0) + type_words;
2892 // pop arguments: newval, offset, base, and receiver
2893 debug_only(int saved_sp = _sp);
2894 _sp += nargs;
2895 Node* newval = (type_words == 1) ? pop() : pop_pair();
2896 Node* oldval = (kind == LS_cmpxchg) ? ((type_words == 1) ? pop() : pop_pair()) : NULL;
2897 Node *offset = pop_pair();
2898 Node *base = pop();
2899 Node *receiver = pop();
2900 assert(saved_sp == _sp, "must have correct argument count");
2902 // Null check receiver.
2903 _sp += nargs;
2904 do_null_check(receiver, T_OBJECT);
2905 _sp -= nargs;
2906 if (stopped()) {
2907 return true;
2908 }
2910 // Build field offset expression.
2911 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2912 // to be plain byte offsets, which are also the same as those accepted
2913 // by oopDesc::field_base.
2914 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2915 // 32-bit machines ignore the high half of long offsets
2916 offset = ConvL2X(offset);
2917 Node* adr = make_unsafe_address(base, offset);
2918 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2920 // For CAS, unlike inline_unsafe_access, there seems no point in
2921 // trying to refine types. Just use the coarse types here.
2922 const Type *value_type = Type::get_const_basic_type(type);
2923 Compile::AliasType* alias_type = C->alias_type(adr_type);
2924 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2926 if (kind == LS_xchg && type == T_OBJECT) {
2927 const TypeOopPtr* tjp = sharpen_unsafe_type(alias_type, adr_type);
2928 if (tjp != NULL) {
2929 value_type = tjp;
2930 }
2931 }
2933 int alias_idx = C->get_alias_index(adr_type);
2935 // Memory-model-wise, a LoadStore acts like a little synchronized
2936 // block, so needs barriers on each side. These don't translate
2937 // into actual barriers on most machines, but we still need rest of
2938 // compiler to respect ordering.
2940 insert_mem_bar(Op_MemBarRelease);
2941 insert_mem_bar(Op_MemBarCPUOrder);
2943 // 4984716: MemBars must be inserted before this
2944 // memory node in order to avoid a false
2945 // dependency which will confuse the scheduler.
2946 Node *mem = memory(alias_idx);
2948 // For now, we handle only those cases that actually exist: ints,
2949 // longs, and Object. Adding others should be straightforward.
2950 Node* load_store;
2951 switch(type) {
2952 case T_INT:
2953 if (kind == LS_xadd) {
2954 load_store = _gvn.transform(new (C) GetAndAddINode(control(), mem, adr, newval, adr_type));
2955 } else if (kind == LS_xchg) {
2956 load_store = _gvn.transform(new (C) GetAndSetINode(control(), mem, adr, newval, adr_type));
2957 } else if (kind == LS_cmpxchg) {
2958 load_store = _gvn.transform(new (C) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2959 } else {
2960 ShouldNotReachHere();
2961 }
2962 break;
2963 case T_LONG:
2964 if (kind == LS_xadd) {
2965 load_store = _gvn.transform(new (C) GetAndAddLNode(control(), mem, adr, newval, adr_type));
2966 } else if (kind == LS_xchg) {
2967 load_store = _gvn.transform(new (C) GetAndSetLNode(control(), mem, adr, newval, adr_type));
2968 } else if (kind == LS_cmpxchg) {
2969 load_store = _gvn.transform(new (C) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2970 } else {
2971 ShouldNotReachHere();
2972 }
2973 break;
2974 case T_OBJECT:
2975 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2976 // could be delayed during Parse (for example, in adjust_map_after_if()).
2977 // Execute transformation here to avoid barrier generation in such case.
2978 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2979 newval = _gvn.makecon(TypePtr::NULL_PTR);
2981 // Reference stores need a store barrier.
2982 pre_barrier(true /* do_load*/,
2983 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2984 NULL /* pre_val*/,
2985 T_OBJECT);
2986 #ifdef _LP64
2987 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2988 Node *newval_enc = _gvn.transform(new (C) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2989 if (kind == LS_xchg) {
2990 load_store = _gvn.transform(new (C) GetAndSetNNode(control(), mem, adr,
2991 newval_enc, adr_type, value_type->make_narrowoop()));
2992 } else {
2993 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
2994 Node *oldval_enc = _gvn.transform(new (C) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2995 load_store = _gvn.transform(new (C) CompareAndSwapNNode(control(), mem, adr,
2996 newval_enc, oldval_enc));
2997 }
2998 } else
2999 #endif
3000 {
3001 if (kind == LS_xchg) {
3002 load_store = _gvn.transform(new (C) GetAndSetPNode(control(), mem, adr, newval, adr_type, value_type->is_oopptr()));
3003 } else {
3004 assert(kind == LS_cmpxchg, "wrong LoadStore operation");
3005 load_store = _gvn.transform(new (C) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
3006 }
3007 }
3008 post_barrier(control(), load_store, base, adr, alias_idx, newval, T_OBJECT, true);
3009 break;
3010 default:
3011 ShouldNotReachHere();
3012 break;
3013 }
3015 // SCMemProjNodes represent the memory state of a LoadStore. Their
3016 // main role is to prevent LoadStore nodes from being optimized away
3017 // when their results aren't used.
3018 Node* proj = _gvn.transform( new (C) SCMemProjNode(load_store));
3019 set_memory(proj, alias_idx);
3021 // Add the trailing membar surrounding the access
3022 insert_mem_bar(Op_MemBarCPUOrder);
3023 insert_mem_bar(Op_MemBarAcquire);
3025 #ifdef _LP64
3026 if (type == T_OBJECT && adr->bottom_type()->is_ptr_to_narrowoop() && kind == LS_xchg) {
3027 load_store = _gvn.transform(new (C) DecodeNNode(load_store, load_store->bottom_type()->make_ptr()));
3028 }
3029 #endif
3031 assert(type2size[load_store->bottom_type()->basic_type()] == type2size[rtype], "result type should match");
3032 push_node(load_store->bottom_type()->basic_type(), load_store);
3033 return true;
3034 }
3036 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
3037 // This is another variant of inline_unsafe_access, differing in
3038 // that it always issues store-store ("release") barrier and ensures
3039 // store-atomicity (which only matters for "long").
3041 if (callee()->is_static()) return false; // caller must have the capability!
3043 #ifndef PRODUCT
3044 {
3045 ResourceMark rm;
3046 // Check the signatures.
3047 ciSignature* sig = signature();
3048 #ifdef ASSERT
3049 BasicType rtype = sig->return_type()->basic_type();
3050 assert(rtype == T_VOID, "must return void");
3051 assert(sig->count() == 3, "has 3 arguments");
3052 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
3053 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
3054 #endif // ASSERT
3055 }
3056 #endif //PRODUCT
3058 // number of stack slots per value argument (1 or 2)
3059 int type_words = type2size[type];
3061 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
3063 // Argument words: "this" plus oop plus offset plus value;
3064 int nargs = 1 + 1 + 2 + type_words;
3066 // pop arguments: val, offset, base, and receiver
3067 debug_only(int saved_sp = _sp);
3068 _sp += nargs;
3069 Node* val = (type_words == 1) ? pop() : pop_pair();
3070 Node *offset = pop_pair();
3071 Node *base = pop();
3072 Node *receiver = pop();
3073 assert(saved_sp == _sp, "must have correct argument count");
3075 // Null check receiver.
3076 _sp += nargs;
3077 do_null_check(receiver, T_OBJECT);
3078 _sp -= nargs;
3079 if (stopped()) {
3080 return true;
3081 }
3083 // Build field offset expression.
3084 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
3085 // 32-bit machines ignore the high half of long offsets
3086 offset = ConvL2X(offset);
3087 Node* adr = make_unsafe_address(base, offset);
3088 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
3089 const Type *value_type = Type::get_const_basic_type(type);
3090 Compile::AliasType* alias_type = C->alias_type(adr_type);
3092 insert_mem_bar(Op_MemBarRelease);
3093 insert_mem_bar(Op_MemBarCPUOrder);
3094 // Ensure that the store is atomic for longs:
3095 bool require_atomic_access = true;
3096 Node* store;
3097 if (type == T_OBJECT) // reference stores need a store barrier.
3098 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
3099 else {
3100 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
3101 }
3102 insert_mem_bar(Op_MemBarCPUOrder);
3103 return true;
3104 }
3106 bool LibraryCallKit::inline_unsafe_allocate() {
3107 if (callee()->is_static()) return false; // caller must have the capability!
3108 int nargs = 1 + 1;
3109 assert(signature()->size() == nargs-1, "alloc has 1 argument");
3110 null_check_receiver(callee()); // check then ignore argument(0)
3111 _sp += nargs; // set original stack for use by uncommon_trap
3112 Node* cls = do_null_check(argument(1), T_OBJECT);
3113 _sp -= nargs;
3114 if (stopped()) return true;
3116 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
3117 _sp += nargs; // set original stack for use by uncommon_trap
3118 kls = do_null_check(kls, T_OBJECT);
3119 _sp -= nargs;
3120 if (stopped()) return true; // argument was like int.class
3122 // Note: The argument might still be an illegal value like
3123 // Serializable.class or Object[].class. The runtime will handle it.
3124 // But we must make an explicit check for initialization.
3125 Node* insp = basic_plus_adr(kls, in_bytes(InstanceKlass::init_state_offset()));
3126 // Use T_BOOLEAN for InstanceKlass::_init_state so the compiler
3127 // can generate code to load it as unsigned byte.
3128 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
3129 Node* bits = intcon(InstanceKlass::fully_initialized);
3130 Node* test = _gvn.transform( new (C) SubINode(inst, bits) );
3131 // The 'test' is non-zero if we need to take a slow path.
3133 Node* obj = new_instance(kls, test);
3134 push(obj);
3136 return true;
3137 }
3139 #ifdef TRACE_HAVE_INTRINSICS
3140 /*
3141 * oop -> myklass
3142 * myklass->trace_id |= USED
3143 * return myklass->trace_id & ~0x3
3144 */
3145 bool LibraryCallKit::inline_native_classID() {
3146 int nargs = 1 + 1;
3147 null_check_receiver(callee()); // check then ignore argument(0)
3148 _sp += nargs;
3149 Node* cls = do_null_check(argument(1), T_OBJECT);
3150 _sp -= nargs;
3151 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
3152 _sp += nargs;
3153 kls = do_null_check(kls, T_OBJECT);
3154 _sp -= nargs;
3155 ByteSize offset = TRACE_ID_OFFSET;
3156 Node* insp = basic_plus_adr(kls, in_bytes(offset));
3157 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
3158 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
3159 Node* andl = _gvn.transform(new (C) AndLNode(tvalue, bits));
3160 Node* clsused = longcon(0x01l); // set the class bit
3161 Node* orl = _gvn.transform(new (C) OrLNode(tvalue, clsused));
3163 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
3164 store_to_memory(control(), insp, orl, T_LONG, adr_type);
3165 push_pair(andl);
3166 return true;
3167 }
3169 bool LibraryCallKit::inline_native_threadID() {
3170 Node* tls_ptr = NULL;
3171 Node* cur_thr = generate_current_thread(tls_ptr);
3172 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3173 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3174 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
3176 Node* threadid = NULL;
3177 size_t thread_id_size = OSThread::thread_id_size();
3178 if (thread_id_size == (size_t) BytesPerLong) {
3179 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
3180 push(threadid);
3181 } else if (thread_id_size == (size_t) BytesPerInt) {
3182 threadid = make_load(control(), p, TypeInt::INT, T_INT);
3183 push(threadid);
3184 } else {
3185 ShouldNotReachHere();
3186 }
3187 return true;
3188 }
3189 #endif
3191 //------------------------inline_native_time_funcs--------------
3192 // inline code for System.currentTimeMillis() and System.nanoTime()
3193 // these have the same type and signature
3194 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
3195 const TypeFunc *tf = OptoRuntime::void_long_Type();
3196 const TypePtr* no_memory_effects = NULL;
3197 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
3198 Node* value = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms+0));
3199 #ifdef ASSERT
3200 Node* value_top = _gvn.transform(new (C) ProjNode(time, TypeFunc::Parms + 1));
3201 assert(value_top == top(), "second value must be top");
3202 #endif
3203 push_pair(value);
3204 return true;
3205 }
3207 //------------------------inline_native_currentThread------------------
3208 bool LibraryCallKit::inline_native_currentThread() {
3209 Node* junk = NULL;
3210 push(generate_current_thread(junk));
3211 return true;
3212 }
3214 //------------------------inline_native_isInterrupted------------------
3215 bool LibraryCallKit::inline_native_isInterrupted() {
3216 const int nargs = 1+1; // receiver + boolean
3217 assert(nargs == arg_size(), "sanity");
3218 // Add a fast path to t.isInterrupted(clear_int):
3219 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
3220 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
3221 // So, in the common case that the interrupt bit is false,
3222 // we avoid making a call into the VM. Even if the interrupt bit
3223 // is true, if the clear_int argument is false, we avoid the VM call.
3224 // However, if the receiver is not currentThread, we must call the VM,
3225 // because there must be some locking done around the operation.
3227 // We only go to the fast case code if we pass two guards.
3228 // Paths which do not pass are accumulated in the slow_region.
3229 RegionNode* slow_region = new (C) RegionNode(1);
3230 record_for_igvn(slow_region);
3231 RegionNode* result_rgn = new (C) RegionNode(1+3); // fast1, fast2, slow
3232 PhiNode* result_val = new (C) PhiNode(result_rgn, TypeInt::BOOL);
3233 enum { no_int_result_path = 1,
3234 no_clear_result_path = 2,
3235 slow_result_path = 3
3236 };
3238 // (a) Receiving thread must be the current thread.
3239 Node* rec_thr = argument(0);
3240 Node* tls_ptr = NULL;
3241 Node* cur_thr = generate_current_thread(tls_ptr);
3242 Node* cmp_thr = _gvn.transform( new (C) CmpPNode(cur_thr, rec_thr) );
3243 Node* bol_thr = _gvn.transform( new (C) BoolNode(cmp_thr, BoolTest::ne) );
3245 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
3246 if (!known_current_thread)
3247 generate_slow_guard(bol_thr, slow_region);
3249 // (b) Interrupt bit on TLS must be false.
3250 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
3251 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
3252 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
3253 // Set the control input on the field _interrupted read to prevent it floating up.
3254 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
3255 Node* cmp_bit = _gvn.transform( new (C) CmpINode(int_bit, intcon(0)) );
3256 Node* bol_bit = _gvn.transform( new (C) BoolNode(cmp_bit, BoolTest::ne) );
3258 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
3260 // First fast path: if (!TLS._interrupted) return false;
3261 Node* false_bit = _gvn.transform( new (C) IfFalseNode(iff_bit) );
3262 result_rgn->init_req(no_int_result_path, false_bit);
3263 result_val->init_req(no_int_result_path, intcon(0));
3265 // drop through to next case
3266 set_control( _gvn.transform(new (C) IfTrueNode(iff_bit)) );
3268 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
3269 Node* clr_arg = argument(1);
3270 Node* cmp_arg = _gvn.transform( new (C) CmpINode(clr_arg, intcon(0)) );
3271 Node* bol_arg = _gvn.transform( new (C) BoolNode(cmp_arg, BoolTest::ne) );
3272 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3274 // Second fast path: ... else if (!clear_int) return true;
3275 Node* false_arg = _gvn.transform( new (C) IfFalseNode(iff_arg) );
3276 result_rgn->init_req(no_clear_result_path, false_arg);
3277 result_val->init_req(no_clear_result_path, intcon(1));
3279 // drop through to next case
3280 set_control( _gvn.transform(new (C) IfTrueNode(iff_arg)) );
3282 // (d) Otherwise, go to the slow path.
3283 slow_region->add_req(control());
3284 set_control( _gvn.transform(slow_region) );
3286 if (stopped()) {
3287 // There is no slow path.
3288 result_rgn->init_req(slow_result_path, top());
3289 result_val->init_req(slow_result_path, top());
3290 } else {
3291 // non-virtual because it is a private non-static
3292 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3294 Node* slow_val = set_results_for_java_call(slow_call);
3295 // this->control() comes from set_results_for_java_call
3297 // If we know that the result of the slow call will be true, tell the optimizer!
3298 if (known_current_thread) slow_val = intcon(1);
3300 Node* fast_io = slow_call->in(TypeFunc::I_O);
3301 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3302 // These two phis are pre-filled with copies of of the fast IO and Memory
3303 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3304 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3306 result_rgn->init_req(slow_result_path, control());
3307 io_phi ->init_req(slow_result_path, i_o());
3308 mem_phi ->init_req(slow_result_path, reset_memory());
3309 result_val->init_req(slow_result_path, slow_val);
3311 set_all_memory( _gvn.transform(mem_phi) );
3312 set_i_o( _gvn.transform(io_phi) );
3313 }
3315 push_result(result_rgn, result_val);
3316 C->set_has_split_ifs(true); // Has chance for split-if optimization
3318 return true;
3319 }
3321 //---------------------------load_mirror_from_klass----------------------------
3322 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3323 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3324 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3325 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3326 }
3328 //-----------------------load_klass_from_mirror_common-------------------------
3329 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3330 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3331 // and branch to the given path on the region.
3332 // If never_see_null, take an uncommon trap on null, so we can optimistically
3333 // compile for the non-null case.
3334 // If the region is NULL, force never_see_null = true.
3335 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3336 bool never_see_null,
3337 int nargs,
3338 RegionNode* region,
3339 int null_path,
3340 int offset) {
3341 if (region == NULL) never_see_null = true;
3342 Node* p = basic_plus_adr(mirror, offset);
3343 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3344 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3345 _sp += nargs; // any deopt will start just before call to enclosing method
3346 Node* null_ctl = top();
3347 kls = null_check_oop(kls, &null_ctl, never_see_null);
3348 if (region != NULL) {
3349 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3350 region->init_req(null_path, null_ctl);
3351 } else {
3352 assert(null_ctl == top(), "no loose ends");
3353 }
3354 _sp -= nargs;
3355 return kls;
3356 }
3358 //--------------------(inline_native_Class_query helpers)---------------------
3359 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3360 // Fall through if (mods & mask) == bits, take the guard otherwise.
3361 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3362 // Branch around if the given klass has the given modifier bit set.
3363 // Like generate_guard, adds a new path onto the region.
3364 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3365 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3366 Node* mask = intcon(modifier_mask);
3367 Node* bits = intcon(modifier_bits);
3368 Node* mbit = _gvn.transform( new (C) AndINode(mods, mask) );
3369 Node* cmp = _gvn.transform( new (C) CmpINode(mbit, bits) );
3370 Node* bol = _gvn.transform( new (C) BoolNode(cmp, BoolTest::ne) );
3371 return generate_fair_guard(bol, region);
3372 }
3373 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3374 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3375 }
3377 //-------------------------inline_native_Class_query-------------------
3378 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3379 int nargs = 1+0; // just the Class mirror, in most cases
3380 const Type* return_type = TypeInt::BOOL;
3381 Node* prim_return_value = top(); // what happens if it's a primitive class?
3382 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3383 bool expect_prim = false; // most of these guys expect to work on refs
3385 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3387 switch (id) {
3388 case vmIntrinsics::_isInstance:
3389 nargs = 1+1; // the Class mirror, plus the object getting queried about
3390 // nothing is an instance of a primitive type
3391 prim_return_value = intcon(0);
3392 break;
3393 case vmIntrinsics::_getModifiers:
3394 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3395 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3396 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3397 break;
3398 case vmIntrinsics::_isInterface:
3399 prim_return_value = intcon(0);
3400 break;
3401 case vmIntrinsics::_isArray:
3402 prim_return_value = intcon(0);
3403 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3404 break;
3405 case vmIntrinsics::_isPrimitive:
3406 prim_return_value = intcon(1);
3407 expect_prim = true; // obviously
3408 break;
3409 case vmIntrinsics::_getSuperclass:
3410 prim_return_value = null();
3411 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3412 break;
3413 case vmIntrinsics::_getComponentType:
3414 prim_return_value = null();
3415 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3416 break;
3417 case vmIntrinsics::_getClassAccessFlags:
3418 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3419 return_type = TypeInt::INT; // not bool! 6297094
3420 break;
3421 default:
3422 ShouldNotReachHere();
3423 }
3425 Node* mirror = argument(0);
3426 Node* obj = (nargs <= 1)? top(): argument(1);
3428 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3429 if (mirror_con == NULL) return false; // cannot happen?
3431 #ifndef PRODUCT
3432 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3433 ciType* k = mirror_con->java_mirror_type();
3434 if (k) {
3435 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3436 k->print_name();
3437 tty->cr();
3438 }
3439 }
3440 #endif
3442 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3443 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3444 record_for_igvn(region);
3445 PhiNode* phi = new (C) PhiNode(region, return_type);
3447 // The mirror will never be null of Reflection.getClassAccessFlags, however
3448 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3449 // if it is. See bug 4774291.
3451 // For Reflection.getClassAccessFlags(), the null check occurs in
3452 // the wrong place; see inline_unsafe_access(), above, for a similar
3453 // situation.
3454 _sp += nargs; // set original stack for use by uncommon_trap
3455 mirror = do_null_check(mirror, T_OBJECT);
3456 _sp -= nargs;
3457 // If mirror or obj is dead, only null-path is taken.
3458 if (stopped()) return true;
3460 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3462 // Now load the mirror's klass metaobject, and null-check it.
3463 // Side-effects region with the control path if the klass is null.
3464 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3465 region, _prim_path);
3466 // If kls is null, we have a primitive mirror.
3467 phi->init_req(_prim_path, prim_return_value);
3468 if (stopped()) { push_result(region, phi); return true; }
3470 Node* p; // handy temp
3471 Node* null_ctl;
3473 // Now that we have the non-null klass, we can perform the real query.
3474 // For constant classes, the query will constant-fold in LoadNode::Value.
3475 Node* query_value = top();
3476 switch (id) {
3477 case vmIntrinsics::_isInstance:
3478 // nothing is an instance of a primitive type
3479 _sp += nargs; // gen_instanceof might do an uncommon trap
3480 query_value = gen_instanceof(obj, kls);
3481 _sp -= nargs;
3482 break;
3484 case vmIntrinsics::_getModifiers:
3485 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3486 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3487 break;
3489 case vmIntrinsics::_isInterface:
3490 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3491 if (generate_interface_guard(kls, region) != NULL)
3492 // A guard was added. If the guard is taken, it was an interface.
3493 phi->add_req(intcon(1));
3494 // If we fall through, it's a plain class.
3495 query_value = intcon(0);
3496 break;
3498 case vmIntrinsics::_isArray:
3499 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3500 if (generate_array_guard(kls, region) != NULL)
3501 // A guard was added. If the guard is taken, it was an array.
3502 phi->add_req(intcon(1));
3503 // If we fall through, it's a plain class.
3504 query_value = intcon(0);
3505 break;
3507 case vmIntrinsics::_isPrimitive:
3508 query_value = intcon(0); // "normal" path produces false
3509 break;
3511 case vmIntrinsics::_getSuperclass:
3512 // The rules here are somewhat unfortunate, but we can still do better
3513 // with random logic than with a JNI call.
3514 // Interfaces store null or Object as _super, but must report null.
3515 // Arrays store an intermediate super as _super, but must report Object.
3516 // Other types can report the actual _super.
3517 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3518 if (generate_interface_guard(kls, region) != NULL)
3519 // A guard was added. If the guard is taken, it was an interface.
3520 phi->add_req(null());
3521 if (generate_array_guard(kls, region) != NULL)
3522 // A guard was added. If the guard is taken, it was an array.
3523 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3524 // If we fall through, it's a plain class. Get its _super.
3525 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3526 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3527 null_ctl = top();
3528 kls = null_check_oop(kls, &null_ctl);
3529 if (null_ctl != top()) {
3530 // If the guard is taken, Object.superClass is null (both klass and mirror).
3531 region->add_req(null_ctl);
3532 phi ->add_req(null());
3533 }
3534 if (!stopped()) {
3535 query_value = load_mirror_from_klass(kls);
3536 }
3537 break;
3539 case vmIntrinsics::_getComponentType:
3540 if (generate_array_guard(kls, region) != NULL) {
3541 // Be sure to pin the oop load to the guard edge just created:
3542 Node* is_array_ctrl = region->in(region->req()-1);
3543 Node* cma = basic_plus_adr(kls, in_bytes(ArrayKlass::component_mirror_offset()));
3544 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3545 phi->add_req(cmo);
3546 }
3547 query_value = null(); // non-array case is null
3548 break;
3550 case vmIntrinsics::_getClassAccessFlags:
3551 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3552 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3553 break;
3555 default:
3556 ShouldNotReachHere();
3557 }
3559 // Fall-through is the normal case of a query to a real class.
3560 phi->init_req(1, query_value);
3561 region->init_req(1, control());
3563 push_result(region, phi);
3564 C->set_has_split_ifs(true); // Has chance for split-if optimization
3566 return true;
3567 }
3569 //--------------------------inline_native_subtype_check------------------------
3570 // This intrinsic takes the JNI calls out of the heart of
3571 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3572 bool LibraryCallKit::inline_native_subtype_check() {
3573 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3575 // Pull both arguments off the stack.
3576 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3577 args[0] = argument(0);
3578 args[1] = argument(1);
3579 Node* klasses[2]; // corresponding Klasses: superk, subk
3580 klasses[0] = klasses[1] = top();
3582 enum {
3583 // A full decision tree on {superc is prim, subc is prim}:
3584 _prim_0_path = 1, // {P,N} => false
3585 // {P,P} & superc!=subc => false
3586 _prim_same_path, // {P,P} & superc==subc => true
3587 _prim_1_path, // {N,P} => false
3588 _ref_subtype_path, // {N,N} & subtype check wins => true
3589 _both_ref_path, // {N,N} & subtype check loses => false
3590 PATH_LIMIT
3591 };
3593 RegionNode* region = new (C) RegionNode(PATH_LIMIT);
3594 Node* phi = new (C) PhiNode(region, TypeInt::BOOL);
3595 record_for_igvn(region);
3597 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3598 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3599 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3601 // First null-check both mirrors and load each mirror's klass metaobject.
3602 int which_arg;
3603 for (which_arg = 0; which_arg <= 1; which_arg++) {
3604 Node* arg = args[which_arg];
3605 _sp += nargs; // set original stack for use by uncommon_trap
3606 arg = do_null_check(arg, T_OBJECT);
3607 _sp -= nargs;
3608 if (stopped()) break;
3609 args[which_arg] = _gvn.transform(arg);
3611 Node* p = basic_plus_adr(arg, class_klass_offset);
3612 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3613 klasses[which_arg] = _gvn.transform(kls);
3614 }
3616 // Having loaded both klasses, test each for null.
3617 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3618 for (which_arg = 0; which_arg <= 1; which_arg++) {
3619 Node* kls = klasses[which_arg];
3620 Node* null_ctl = top();
3621 _sp += nargs; // set original stack for use by uncommon_trap
3622 kls = null_check_oop(kls, &null_ctl, never_see_null);
3623 _sp -= nargs;
3624 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3625 region->init_req(prim_path, null_ctl);
3626 if (stopped()) break;
3627 klasses[which_arg] = kls;
3628 }
3630 if (!stopped()) {
3631 // now we have two reference types, in klasses[0..1]
3632 Node* subk = klasses[1]; // the argument to isAssignableFrom
3633 Node* superk = klasses[0]; // the receiver
3634 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3635 // now we have a successful reference subtype check
3636 region->set_req(_ref_subtype_path, control());
3637 }
3639 // If both operands are primitive (both klasses null), then
3640 // we must return true when they are identical primitives.
3641 // It is convenient to test this after the first null klass check.
3642 set_control(region->in(_prim_0_path)); // go back to first null check
3643 if (!stopped()) {
3644 // Since superc is primitive, make a guard for the superc==subc case.
3645 Node* cmp_eq = _gvn.transform( new (C) CmpPNode(args[0], args[1]) );
3646 Node* bol_eq = _gvn.transform( new (C) BoolNode(cmp_eq, BoolTest::eq) );
3647 generate_guard(bol_eq, region, PROB_FAIR);
3648 if (region->req() == PATH_LIMIT+1) {
3649 // A guard was added. If the added guard is taken, superc==subc.
3650 region->swap_edges(PATH_LIMIT, _prim_same_path);
3651 region->del_req(PATH_LIMIT);
3652 }
3653 region->set_req(_prim_0_path, control()); // Not equal after all.
3654 }
3656 // these are the only paths that produce 'true':
3657 phi->set_req(_prim_same_path, intcon(1));
3658 phi->set_req(_ref_subtype_path, intcon(1));
3660 // pull together the cases:
3661 assert(region->req() == PATH_LIMIT, "sane region");
3662 for (uint i = 1; i < region->req(); i++) {
3663 Node* ctl = region->in(i);
3664 if (ctl == NULL || ctl == top()) {
3665 region->set_req(i, top());
3666 phi ->set_req(i, top());
3667 } else if (phi->in(i) == NULL) {
3668 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3669 }
3670 }
3672 set_control(_gvn.transform(region));
3673 push(_gvn.transform(phi));
3675 return true;
3676 }
3678 //---------------------generate_array_guard_common------------------------
3679 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3680 bool obj_array, bool not_array) {
3681 // If obj_array/non_array==false/false:
3682 // Branch around if the given klass is in fact an array (either obj or prim).
3683 // If obj_array/non_array==false/true:
3684 // Branch around if the given klass is not an array klass of any kind.
3685 // If obj_array/non_array==true/true:
3686 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3687 // If obj_array/non_array==true/false:
3688 // Branch around if the kls is an oop array (Object[] or subtype)
3689 //
3690 // Like generate_guard, adds a new path onto the region.
3691 jint layout_con = 0;
3692 Node* layout_val = get_layout_helper(kls, layout_con);
3693 if (layout_val == NULL) {
3694 bool query = (obj_array
3695 ? Klass::layout_helper_is_objArray(layout_con)
3696 : Klass::layout_helper_is_array(layout_con));
3697 if (query == not_array) {
3698 return NULL; // never a branch
3699 } else { // always a branch
3700 Node* always_branch = control();
3701 if (region != NULL)
3702 region->add_req(always_branch);
3703 set_control(top());
3704 return always_branch;
3705 }
3706 }
3707 // Now test the correct condition.
3708 jint nval = (obj_array
3709 ? ((jint)Klass::_lh_array_tag_type_value
3710 << Klass::_lh_array_tag_shift)
3711 : Klass::_lh_neutral_value);
3712 Node* cmp = _gvn.transform( new(C) CmpINode(layout_val, intcon(nval)) );
3713 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3714 // invert the test if we are looking for a non-array
3715 if (not_array) btest = BoolTest(btest).negate();
3716 Node* bol = _gvn.transform( new(C) BoolNode(cmp, btest) );
3717 return generate_fair_guard(bol, region);
3718 }
3721 //-----------------------inline_native_newArray--------------------------
3722 bool LibraryCallKit::inline_native_newArray() {
3723 int nargs = 2;
3724 Node* mirror = argument(0);
3725 Node* count_val = argument(1);
3727 _sp += nargs; // set original stack for use by uncommon_trap
3728 mirror = do_null_check(mirror, T_OBJECT);
3729 _sp -= nargs;
3730 // If mirror or obj is dead, only null-path is taken.
3731 if (stopped()) return true;
3733 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3734 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
3735 PhiNode* result_val = new(C) PhiNode(result_reg,
3736 TypeInstPtr::NOTNULL);
3737 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
3738 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
3739 TypePtr::BOTTOM);
3741 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3742 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3743 nargs,
3744 result_reg, _slow_path);
3745 Node* normal_ctl = control();
3746 Node* no_array_ctl = result_reg->in(_slow_path);
3748 // Generate code for the slow case. We make a call to newArray().
3749 set_control(no_array_ctl);
3750 if (!stopped()) {
3751 // Either the input type is void.class, or else the
3752 // array klass has not yet been cached. Either the
3753 // ensuing call will throw an exception, or else it
3754 // will cache the array klass for next time.
3755 PreserveJVMState pjvms(this);
3756 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3757 Node* slow_result = set_results_for_java_call(slow_call);
3758 // this->control() comes from set_results_for_java_call
3759 result_reg->set_req(_slow_path, control());
3760 result_val->set_req(_slow_path, slow_result);
3761 result_io ->set_req(_slow_path, i_o());
3762 result_mem->set_req(_slow_path, reset_memory());
3763 }
3765 set_control(normal_ctl);
3766 if (!stopped()) {
3767 // Normal case: The array type has been cached in the java.lang.Class.
3768 // The following call works fine even if the array type is polymorphic.
3769 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3770 Node* obj = new_array(klass_node, count_val, nargs);
3771 result_reg->init_req(_normal_path, control());
3772 result_val->init_req(_normal_path, obj);
3773 result_io ->init_req(_normal_path, i_o());
3774 result_mem->init_req(_normal_path, reset_memory());
3775 }
3777 // Return the combined state.
3778 set_i_o( _gvn.transform(result_io) );
3779 set_all_memory( _gvn.transform(result_mem) );
3780 push_result(result_reg, result_val);
3781 C->set_has_split_ifs(true); // Has chance for split-if optimization
3783 return true;
3784 }
3786 //----------------------inline_native_getLength--------------------------
3787 bool LibraryCallKit::inline_native_getLength() {
3788 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3790 int nargs = 1;
3791 Node* array = argument(0);
3793 _sp += nargs; // set original stack for use by uncommon_trap
3794 array = do_null_check(array, T_OBJECT);
3795 _sp -= nargs;
3797 // If array is dead, only null-path is taken.
3798 if (stopped()) return true;
3800 // Deoptimize if it is a non-array.
3801 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3803 if (non_array != NULL) {
3804 PreserveJVMState pjvms(this);
3805 set_control(non_array);
3806 _sp += nargs; // push the arguments back on the stack
3807 uncommon_trap(Deoptimization::Reason_intrinsic,
3808 Deoptimization::Action_maybe_recompile);
3809 }
3811 // If control is dead, only non-array-path is taken.
3812 if (stopped()) return true;
3814 // The works fine even if the array type is polymorphic.
3815 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3816 push( load_array_length(array) );
3818 C->set_has_split_ifs(true); // Has chance for split-if optimization
3820 return true;
3821 }
3823 //------------------------inline_array_copyOf----------------------------
3824 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3825 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3827 // Restore the stack and pop off the arguments.
3828 int nargs = 3 + (is_copyOfRange? 1: 0);
3829 Node* original = argument(0);
3830 Node* start = is_copyOfRange? argument(1): intcon(0);
3831 Node* end = is_copyOfRange? argument(2): argument(1);
3832 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3834 Node* newcopy;
3836 //set the original stack and the reexecute bit for the interpreter to reexecute
3837 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3838 { PreserveReexecuteState preexecs(this);
3839 _sp += nargs;
3840 jvms()->set_should_reexecute(true);
3842 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3843 original = do_null_check(original, T_OBJECT);
3845 // Check if a null path was taken unconditionally.
3846 if (stopped()) return true;
3848 Node* orig_length = load_array_length(original);
3850 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3851 NULL, 0);
3852 klass_node = do_null_check(klass_node, T_OBJECT);
3854 RegionNode* bailout = new (C) RegionNode(1);
3855 record_for_igvn(bailout);
3857 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3858 // Bail out if that is so.
3859 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3860 if (not_objArray != NULL) {
3861 // Improve the klass node's type from the new optimistic assumption:
3862 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3863 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3864 Node* cast = new (C) CastPPNode(klass_node, akls);
3865 cast->init_req(0, control());
3866 klass_node = _gvn.transform(cast);
3867 }
3869 // Bail out if either start or end is negative.
3870 generate_negative_guard(start, bailout, &start);
3871 generate_negative_guard(end, bailout, &end);
3873 Node* length = end;
3874 if (_gvn.type(start) != TypeInt::ZERO) {
3875 length = _gvn.transform( new (C) SubINode(end, start) );
3876 }
3878 // Bail out if length is negative.
3879 // Without this the new_array would throw
3880 // NegativeArraySizeException but IllegalArgumentException is what
3881 // should be thrown
3882 generate_negative_guard(length, bailout, &length);
3884 if (bailout->req() > 1) {
3885 PreserveJVMState pjvms(this);
3886 set_control( _gvn.transform(bailout) );
3887 uncommon_trap(Deoptimization::Reason_intrinsic,
3888 Deoptimization::Action_maybe_recompile);
3889 }
3891 if (!stopped()) {
3893 // How many elements will we copy from the original?
3894 // The answer is MinI(orig_length - start, length).
3895 Node* orig_tail = _gvn.transform( new(C) SubINode(orig_length, start) );
3896 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3898 newcopy = new_array(klass_node, length, 0);
3900 // Generate a direct call to the right arraycopy function(s).
3901 // We know the copy is disjoint but we might not know if the
3902 // oop stores need checking.
3903 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3904 // This will fail a store-check if x contains any non-nulls.
3905 bool disjoint_bases = true;
3906 // if start > orig_length then the length of the copy may be
3907 // negative.
3908 bool length_never_negative = !is_copyOfRange;
3909 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3910 original, start, newcopy, intcon(0), moved,
3911 disjoint_bases, length_never_negative);
3912 }
3913 } //original reexecute and sp are set back here
3915 if(!stopped()) {
3916 push(newcopy);
3917 }
3919 C->set_has_split_ifs(true); // Has chance for split-if optimization
3921 return true;
3922 }
3925 //----------------------generate_virtual_guard---------------------------
3926 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3927 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3928 RegionNode* slow_region) {
3929 ciMethod* method = callee();
3930 int vtable_index = method->vtable_index();
3931 // Get the Method* out of the appropriate vtable entry.
3932 int entry_offset = (InstanceKlass::vtable_start_offset() +
3933 vtable_index*vtableEntry::size()) * wordSize +
3934 vtableEntry::method_offset_in_bytes();
3935 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3936 Node* target_call = make_load(NULL, entry_addr, TypePtr::NOTNULL, T_ADDRESS);
3938 // Compare the target method with the expected method (e.g., Object.hashCode).
3939 const TypePtr* native_call_addr = TypeMetadataPtr::make(method);
3941 Node* native_call = makecon(native_call_addr);
3942 Node* chk_native = _gvn.transform( new(C) CmpPNode(target_call, native_call) );
3943 Node* test_native = _gvn.transform( new(C) BoolNode(chk_native, BoolTest::ne) );
3945 return generate_slow_guard(test_native, slow_region);
3946 }
3948 //-----------------------generate_method_call----------------------------
3949 // Use generate_method_call to make a slow-call to the real
3950 // method if the fast path fails. An alternative would be to
3951 // use a stub like OptoRuntime::slow_arraycopy_Java.
3952 // This only works for expanding the current library call,
3953 // not another intrinsic. (E.g., don't use this for making an
3954 // arraycopy call inside of the copyOf intrinsic.)
3955 CallJavaNode*
3956 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3957 // When compiling the intrinsic method itself, do not use this technique.
3958 guarantee(callee() != C->method(), "cannot make slow-call to self");
3960 ciMethod* method = callee();
3961 // ensure the JVMS we have will be correct for this call
3962 guarantee(method_id == method->intrinsic_id(), "must match");
3964 const TypeFunc* tf = TypeFunc::make(method);
3965 CallJavaNode* slow_call;
3966 if (is_static) {
3967 assert(!is_virtual, "");
3968 slow_call = new(C) CallStaticJavaNode(tf,
3969 SharedRuntime::get_resolve_static_call_stub(),
3970 method, bci());
3971 } else if (is_virtual) {
3972 null_check_receiver(method);
3973 int vtable_index = Method::invalid_vtable_index;
3974 if (UseInlineCaches) {
3975 // Suppress the vtable call
3976 } else {
3977 // hashCode and clone are not a miranda methods,
3978 // so the vtable index is fixed.
3979 // No need to use the linkResolver to get it.
3980 vtable_index = method->vtable_index();
3981 }
3982 slow_call = new(C) CallDynamicJavaNode(tf,
3983 SharedRuntime::get_resolve_virtual_call_stub(),
3984 method, vtable_index, bci());
3985 } else { // neither virtual nor static: opt_virtual
3986 null_check_receiver(method);
3987 slow_call = new(C) CallStaticJavaNode(tf,
3988 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3989 method, bci());
3990 slow_call->set_optimized_virtual(true);
3991 }
3992 set_arguments_for_java_call(slow_call);
3993 set_edges_for_java_call(slow_call);
3994 return slow_call;
3995 }
3998 //------------------------------inline_native_hashcode--------------------
3999 // Build special case code for calls to hashCode on an object.
4000 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
4001 assert(is_static == callee()->is_static(), "correct intrinsic selection");
4002 assert(!(is_virtual && is_static), "either virtual, special, or static");
4004 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
4006 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4007 PhiNode* result_val = new(C) PhiNode(result_reg,
4008 TypeInt::INT);
4009 PhiNode* result_io = new(C) PhiNode(result_reg, Type::ABIO);
4010 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4011 TypePtr::BOTTOM);
4012 Node* obj = NULL;
4013 if (!is_static) {
4014 // Check for hashing null object
4015 obj = null_check_receiver(callee());
4016 if (stopped()) return true; // unconditionally null
4017 result_reg->init_req(_null_path, top());
4018 result_val->init_req(_null_path, top());
4019 } else {
4020 // Do a null check, and return zero if null.
4021 // System.identityHashCode(null) == 0
4022 obj = argument(0);
4023 Node* null_ctl = top();
4024 obj = null_check_oop(obj, &null_ctl);
4025 result_reg->init_req(_null_path, null_ctl);
4026 result_val->init_req(_null_path, _gvn.intcon(0));
4027 }
4029 // Unconditionally null? Then return right away.
4030 if (stopped()) {
4031 set_control( result_reg->in(_null_path) );
4032 if (!stopped())
4033 push( result_val ->in(_null_path) );
4034 return true;
4035 }
4037 // After null check, get the object's klass.
4038 Node* obj_klass = load_object_klass(obj);
4040 // This call may be virtual (invokevirtual) or bound (invokespecial).
4041 // For each case we generate slightly different code.
4043 // We only go to the fast case code if we pass a number of guards. The
4044 // paths which do not pass are accumulated in the slow_region.
4045 RegionNode* slow_region = new (C) RegionNode(1);
4046 record_for_igvn(slow_region);
4048 // If this is a virtual call, we generate a funny guard. We pull out
4049 // the vtable entry corresponding to hashCode() from the target object.
4050 // If the target method which we are calling happens to be the native
4051 // Object hashCode() method, we pass the guard. We do not need this
4052 // guard for non-virtual calls -- the caller is known to be the native
4053 // Object hashCode().
4054 if (is_virtual) {
4055 generate_virtual_guard(obj_klass, slow_region);
4056 }
4058 // Get the header out of the object, use LoadMarkNode when available
4059 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
4060 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
4062 // Test the header to see if it is unlocked.
4063 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
4064 Node *lmasked_header = _gvn.transform( new (C) AndXNode(header, lock_mask) );
4065 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
4066 Node *chk_unlocked = _gvn.transform( new (C) CmpXNode( lmasked_header, unlocked_val));
4067 Node *test_unlocked = _gvn.transform( new (C) BoolNode( chk_unlocked, BoolTest::ne) );
4069 generate_slow_guard(test_unlocked, slow_region);
4071 // Get the hash value and check to see that it has been properly assigned.
4072 // We depend on hash_mask being at most 32 bits and avoid the use of
4073 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
4074 // vm: see markOop.hpp.
4075 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
4076 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
4077 Node *hshifted_header= _gvn.transform( new (C) URShiftXNode(header, hash_shift) );
4078 // This hack lets the hash bits live anywhere in the mark object now, as long
4079 // as the shift drops the relevant bits into the low 32 bits. Note that
4080 // Java spec says that HashCode is an int so there's no point in capturing
4081 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
4082 hshifted_header = ConvX2I(hshifted_header);
4083 Node *hash_val = _gvn.transform( new (C) AndINode(hshifted_header, hash_mask) );
4085 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
4086 Node *chk_assigned = _gvn.transform( new (C) CmpINode( hash_val, no_hash_val));
4087 Node *test_assigned = _gvn.transform( new (C) BoolNode( chk_assigned, BoolTest::eq) );
4089 generate_slow_guard(test_assigned, slow_region);
4091 Node* init_mem = reset_memory();
4092 // fill in the rest of the null path:
4093 result_io ->init_req(_null_path, i_o());
4094 result_mem->init_req(_null_path, init_mem);
4096 result_val->init_req(_fast_path, hash_val);
4097 result_reg->init_req(_fast_path, control());
4098 result_io ->init_req(_fast_path, i_o());
4099 result_mem->init_req(_fast_path, init_mem);
4101 // Generate code for the slow case. We make a call to hashCode().
4102 set_control(_gvn.transform(slow_region));
4103 if (!stopped()) {
4104 // No need for PreserveJVMState, because we're using up the present state.
4105 set_all_memory(init_mem);
4106 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
4107 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
4108 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
4109 Node* slow_result = set_results_for_java_call(slow_call);
4110 // this->control() comes from set_results_for_java_call
4111 result_reg->init_req(_slow_path, control());
4112 result_val->init_req(_slow_path, slow_result);
4113 result_io ->set_req(_slow_path, i_o());
4114 result_mem ->set_req(_slow_path, reset_memory());
4115 }
4117 // Return the combined state.
4118 set_i_o( _gvn.transform(result_io) );
4119 set_all_memory( _gvn.transform(result_mem) );
4120 push_result(result_reg, result_val);
4122 return true;
4123 }
4125 //---------------------------inline_native_getClass----------------------------
4126 // Build special case code for calls to getClass on an object.
4127 bool LibraryCallKit::inline_native_getClass() {
4128 Node* obj = null_check_receiver(callee());
4129 if (stopped()) return true;
4130 push( load_mirror_from_klass(load_object_klass(obj)) );
4131 return true;
4132 }
4134 //-----------------inline_native_Reflection_getCallerClass---------------------
4135 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
4136 //
4137 // NOTE that this code must perform the same logic as
4138 // vframeStream::security_get_caller_frame in that it must skip
4139 // Method.invoke() and auxiliary frames.
4144 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
4145 ciMethod* method = callee();
4147 #ifndef PRODUCT
4148 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4149 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
4150 }
4151 #endif
4153 debug_only(int saved_sp = _sp);
4155 // Argument words: (int depth)
4156 int nargs = 1;
4158 _sp += nargs;
4159 Node* caller_depth_node = pop();
4161 assert(saved_sp == _sp, "must have correct argument count");
4163 // The depth value must be a constant in order for the runtime call
4164 // to be eliminated.
4165 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
4166 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
4167 #ifndef PRODUCT
4168 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4169 tty->print_cr(" Bailing out because caller depth was not a constant");
4170 }
4171 #endif
4172 return false;
4173 }
4174 // Note that the JVM state at this point does not include the
4175 // getCallerClass() frame which we are trying to inline. The
4176 // semantics of getCallerClass(), however, are that the "first"
4177 // frame is the getCallerClass() frame, so we subtract one from the
4178 // requested depth before continuing. We don't inline requests of
4179 // getCallerClass(0).
4180 int caller_depth = caller_depth_type->get_con() - 1;
4181 if (caller_depth < 0) {
4182 #ifndef PRODUCT
4183 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4184 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
4185 }
4186 #endif
4187 return false;
4188 }
4190 if (!jvms()->has_method()) {
4191 #ifndef PRODUCT
4192 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4193 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
4194 }
4195 #endif
4196 return false;
4197 }
4198 int _depth = jvms()->depth(); // cache call chain depth
4200 // Walk back up the JVM state to find the caller at the required
4201 // depth. NOTE that this code must perform the same logic as
4202 // vframeStream::security_get_caller_frame in that it must skip
4203 // Method.invoke() and auxiliary frames. Note also that depth is
4204 // 1-based (1 is the bottom of the inlining).
4205 int inlining_depth = _depth;
4206 JVMState* caller_jvms = NULL;
4208 if (inlining_depth > 0) {
4209 caller_jvms = jvms();
4210 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
4211 do {
4212 // The following if-tests should be performed in this order
4213 if (is_method_invoke_or_aux_frame(caller_jvms)) {
4214 // Skip a Method.invoke() or auxiliary frame
4215 } else if (caller_depth > 0) {
4216 // Skip real frame
4217 --caller_depth;
4218 } else {
4219 // We're done: reached desired caller after skipping.
4220 break;
4221 }
4222 caller_jvms = caller_jvms->caller();
4223 --inlining_depth;
4224 } while (inlining_depth > 0);
4225 }
4227 if (inlining_depth == 0) {
4228 #ifndef PRODUCT
4229 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4230 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
4231 tty->print_cr(" JVM state at this point:");
4232 for (int i = _depth; i >= 1; i--) {
4233 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
4234 }
4235 }
4236 #endif
4237 return false; // Reached end of inlining
4238 }
4240 // Acquire method holder as java.lang.Class
4241 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
4242 ciInstance* caller_mirror = caller_klass->java_mirror();
4243 // Push this as a constant
4244 push(makecon(TypeInstPtr::make(caller_mirror)));
4245 #ifndef PRODUCT
4246 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
4247 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);
4248 tty->print_cr(" JVM state at this point:");
4249 for (int i = _depth; i >= 1; i--) {
4250 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
4251 }
4252 }
4253 #endif
4254 return true;
4255 }
4257 // Helper routine for above
4258 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
4259 ciMethod* method = jvms->method();
4261 // Is this the Method.invoke method itself?
4262 if (method->intrinsic_id() == vmIntrinsics::_invoke)
4263 return true;
4265 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
4266 ciKlass* k = method->holder();
4267 if (k->is_instance_klass()) {
4268 ciInstanceKlass* ik = k->as_instance_klass();
4269 for (; ik != NULL; ik = ik->super()) {
4270 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
4271 ik == env()->find_system_klass(ik->name())) {
4272 return true;
4273 }
4274 }
4275 }
4276 else if (method->is_method_handle_intrinsic() ||
4277 method->is_compiled_lambda_form()) {
4278 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4279 return true;
4280 }
4282 return false;
4283 }
4285 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4286 // restore the arguments
4287 _sp += arg_size();
4289 switch (id) {
4290 case vmIntrinsics::_floatToRawIntBits:
4291 push(_gvn.transform( new (C) MoveF2INode(pop())));
4292 break;
4294 case vmIntrinsics::_intBitsToFloat:
4295 push(_gvn.transform( new (C) MoveI2FNode(pop())));
4296 break;
4298 case vmIntrinsics::_doubleToRawLongBits:
4299 push_pair(_gvn.transform( new (C) MoveD2LNode(pop_pair())));
4300 break;
4302 case vmIntrinsics::_longBitsToDouble:
4303 push_pair(_gvn.transform( new (C) MoveL2DNode(pop_pair())));
4304 break;
4306 case vmIntrinsics::_doubleToLongBits: {
4307 Node* value = pop_pair();
4309 // two paths (plus control) merge in a wood
4310 RegionNode *r = new (C) RegionNode(3);
4311 Node *phi = new (C) PhiNode(r, TypeLong::LONG);
4313 Node *cmpisnan = _gvn.transform( new (C) CmpDNode(value, value));
4314 // Build the boolean node
4315 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) );
4317 // Branch either way.
4318 // NaN case is less traveled, which makes all the difference.
4319 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4320 Node *opt_isnan = _gvn.transform(ifisnan);
4321 assert( opt_isnan->is_If(), "Expect an IfNode");
4322 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4323 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4325 set_control(iftrue);
4327 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4328 Node *slow_result = longcon(nan_bits); // return NaN
4329 phi->init_req(1, _gvn.transform( slow_result ));
4330 r->init_req(1, iftrue);
4332 // Else fall through
4333 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) );
4334 set_control(iffalse);
4336 phi->init_req(2, _gvn.transform( new (C) MoveD2LNode(value)));
4337 r->init_req(2, iffalse);
4339 // Post merge
4340 set_control(_gvn.transform(r));
4341 record_for_igvn(r);
4343 Node* result = _gvn.transform(phi);
4344 assert(result->bottom_type()->isa_long(), "must be");
4345 push_pair(result);
4347 C->set_has_split_ifs(true); // Has chance for split-if optimization
4349 break;
4350 }
4352 case vmIntrinsics::_floatToIntBits: {
4353 Node* value = pop();
4355 // two paths (plus control) merge in a wood
4356 RegionNode *r = new (C) RegionNode(3);
4357 Node *phi = new (C) PhiNode(r, TypeInt::INT);
4359 Node *cmpisnan = _gvn.transform( new (C) CmpFNode(value, value));
4360 // Build the boolean node
4361 Node *bolisnan = _gvn.transform( new (C) BoolNode( cmpisnan, BoolTest::ne ) );
4363 // Branch either way.
4364 // NaN case is less traveled, which makes all the difference.
4365 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4366 Node *opt_isnan = _gvn.transform(ifisnan);
4367 assert( opt_isnan->is_If(), "Expect an IfNode");
4368 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4369 Node *iftrue = _gvn.transform( new (C) IfTrueNode(opt_ifisnan) );
4371 set_control(iftrue);
4373 static const jint nan_bits = 0x7fc00000;
4374 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4375 phi->init_req(1, _gvn.transform( slow_result ));
4376 r->init_req(1, iftrue);
4378 // Else fall through
4379 Node *iffalse = _gvn.transform( new (C) IfFalseNode(opt_ifisnan) );
4380 set_control(iffalse);
4382 phi->init_req(2, _gvn.transform( new (C) MoveF2INode(value)));
4383 r->init_req(2, iffalse);
4385 // Post merge
4386 set_control(_gvn.transform(r));
4387 record_for_igvn(r);
4389 Node* result = _gvn.transform(phi);
4390 assert(result->bottom_type()->isa_int(), "must be");
4391 push(result);
4393 C->set_has_split_ifs(true); // Has chance for split-if optimization
4395 break;
4396 }
4398 default:
4399 ShouldNotReachHere();
4400 }
4402 return true;
4403 }
4405 #ifdef _LP64
4406 #define XTOP ,top() /*additional argument*/
4407 #else //_LP64
4408 #define XTOP /*no additional argument*/
4409 #endif //_LP64
4411 //----------------------inline_unsafe_copyMemory-------------------------
4412 bool LibraryCallKit::inline_unsafe_copyMemory() {
4413 if (callee()->is_static()) return false; // caller must have the capability!
4414 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4415 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4416 null_check_receiver(callee()); // check then ignore argument(0)
4417 if (stopped()) return true;
4419 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4421 Node* src_ptr = argument(1);
4422 Node* src_off = ConvL2X(argument(2));
4423 assert(argument(3)->is_top(), "2nd half of long");
4424 Node* dst_ptr = argument(4);
4425 Node* dst_off = ConvL2X(argument(5));
4426 assert(argument(6)->is_top(), "2nd half of long");
4427 Node* size = ConvL2X(argument(7));
4428 assert(argument(8)->is_top(), "2nd half of long");
4430 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4431 "fieldOffset must be byte-scaled");
4433 Node* src = make_unsafe_address(src_ptr, src_off);
4434 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4436 // Conservatively insert a memory barrier on all memory slices.
4437 // Do not let writes of the copy source or destination float below the copy.
4438 insert_mem_bar(Op_MemBarCPUOrder);
4440 // Call it. Note that the length argument is not scaled.
4441 make_runtime_call(RC_LEAF|RC_NO_FP,
4442 OptoRuntime::fast_arraycopy_Type(),
4443 StubRoutines::unsafe_arraycopy(),
4444 "unsafe_arraycopy",
4445 TypeRawPtr::BOTTOM,
4446 src, dst, size XTOP);
4448 // Do not let reads of the copy destination float above the copy.
4449 insert_mem_bar(Op_MemBarCPUOrder);
4451 return true;
4452 }
4454 //------------------------clone_coping-----------------------------------
4455 // Helper function for inline_native_clone.
4456 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4457 assert(obj_size != NULL, "");
4458 Node* raw_obj = alloc_obj->in(1);
4459 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4461 AllocateNode* alloc = NULL;
4462 if (ReduceBulkZeroing) {
4463 // We will be completely responsible for initializing this object -
4464 // mark Initialize node as complete.
4465 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4466 // The object was just allocated - there should be no any stores!
4467 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4468 // Mark as complete_with_arraycopy so that on AllocateNode
4469 // expansion, we know this AllocateNode is initialized by an array
4470 // copy and a StoreStore barrier exists after the array copy.
4471 alloc->initialization()->set_complete_with_arraycopy();
4472 }
4474 // Copy the fastest available way.
4475 // TODO: generate fields copies for small objects instead.
4476 Node* src = obj;
4477 Node* dest = alloc_obj;
4478 Node* size = _gvn.transform(obj_size);
4480 // Exclude the header but include array length to copy by 8 bytes words.
4481 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4482 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4483 instanceOopDesc::base_offset_in_bytes();
4484 // base_off:
4485 // 8 - 32-bit VM
4486 // 12 - 64-bit VM, compressed klass
4487 // 16 - 64-bit VM, normal klass
4488 if (base_off % BytesPerLong != 0) {
4489 assert(UseCompressedKlassPointers, "");
4490 if (is_array) {
4491 // Exclude length to copy by 8 bytes words.
4492 base_off += sizeof(int);
4493 } else {
4494 // Include klass to copy by 8 bytes words.
4495 base_off = instanceOopDesc::klass_offset_in_bytes();
4496 }
4497 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4498 }
4499 src = basic_plus_adr(src, base_off);
4500 dest = basic_plus_adr(dest, base_off);
4502 // Compute the length also, if needed:
4503 Node* countx = size;
4504 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(base_off)) );
4505 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4507 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4508 bool disjoint_bases = true;
4509 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4510 src, NULL, dest, NULL, countx,
4511 /*dest_uninitialized*/true);
4513 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4514 if (card_mark) {
4515 assert(!is_array, "");
4516 // Put in store barrier for any and all oops we are sticking
4517 // into this object. (We could avoid this if we could prove
4518 // that the object type contains no oop fields at all.)
4519 Node* no_particular_value = NULL;
4520 Node* no_particular_field = NULL;
4521 int raw_adr_idx = Compile::AliasIdxRaw;
4522 post_barrier(control(),
4523 memory(raw_adr_type),
4524 alloc_obj,
4525 no_particular_field,
4526 raw_adr_idx,
4527 no_particular_value,
4528 T_OBJECT,
4529 false);
4530 }
4532 // Do not let reads from the cloned object float above the arraycopy.
4533 if (alloc != NULL) {
4534 // Do not let stores that initialize this object be reordered with
4535 // a subsequent store that would make this object accessible by
4536 // other threads.
4537 // Record what AllocateNode this StoreStore protects so that
4538 // escape analysis can go from the MemBarStoreStoreNode to the
4539 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4540 // based on the escape status of the AllocateNode.
4541 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4542 } else {
4543 insert_mem_bar(Op_MemBarCPUOrder);
4544 }
4545 }
4547 //------------------------inline_native_clone----------------------------
4548 // Here are the simple edge cases:
4549 // null receiver => normal trap
4550 // virtual and clone was overridden => slow path to out-of-line clone
4551 // not cloneable or finalizer => slow path to out-of-line Object.clone
4552 //
4553 // The general case has two steps, allocation and copying.
4554 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4555 //
4556 // Copying also has two cases, oop arrays and everything else.
4557 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4558 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4559 //
4560 // These steps fold up nicely if and when the cloned object's klass
4561 // can be sharply typed as an object array, a type array, or an instance.
4562 //
4563 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4564 int nargs = 1;
4565 PhiNode* result_val;
4567 //set the original stack and the reexecute bit for the interpreter to reexecute
4568 //the bytecode that invokes Object.clone if deoptimization happens
4569 { PreserveReexecuteState preexecs(this);
4570 jvms()->set_should_reexecute(true);
4572 //null_check_receiver will adjust _sp (push and pop)
4573 Node* obj = null_check_receiver(callee());
4574 if (stopped()) return true;
4576 _sp += nargs;
4578 Node* obj_klass = load_object_klass(obj);
4579 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4580 const TypeOopPtr* toop = ((tklass != NULL)
4581 ? tklass->as_instance_type()
4582 : TypeInstPtr::NOTNULL);
4584 // Conservatively insert a memory barrier on all memory slices.
4585 // Do not let writes into the original float below the clone.
4586 insert_mem_bar(Op_MemBarCPUOrder);
4588 // paths into result_reg:
4589 enum {
4590 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4591 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4592 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4593 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4594 PATH_LIMIT
4595 };
4596 RegionNode* result_reg = new(C) RegionNode(PATH_LIMIT);
4597 result_val = new(C) PhiNode(result_reg,
4598 TypeInstPtr::NOTNULL);
4599 PhiNode* result_i_o = new(C) PhiNode(result_reg, Type::ABIO);
4600 PhiNode* result_mem = new(C) PhiNode(result_reg, Type::MEMORY,
4601 TypePtr::BOTTOM);
4602 record_for_igvn(result_reg);
4604 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4605 int raw_adr_idx = Compile::AliasIdxRaw;
4607 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4608 if (array_ctl != NULL) {
4609 // It's an array.
4610 PreserveJVMState pjvms(this);
4611 set_control(array_ctl);
4612 Node* obj_length = load_array_length(obj);
4613 Node* obj_size = NULL;
4614 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4616 if (!use_ReduceInitialCardMarks()) {
4617 // If it is an oop array, it requires very special treatment,
4618 // because card marking is required on each card of the array.
4619 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4620 if (is_obja != NULL) {
4621 PreserveJVMState pjvms2(this);
4622 set_control(is_obja);
4623 // Generate a direct call to the right arraycopy function(s).
4624 bool disjoint_bases = true;
4625 bool length_never_negative = true;
4626 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4627 obj, intcon(0), alloc_obj, intcon(0),
4628 obj_length,
4629 disjoint_bases, length_never_negative);
4630 result_reg->init_req(_objArray_path, control());
4631 result_val->init_req(_objArray_path, alloc_obj);
4632 result_i_o ->set_req(_objArray_path, i_o());
4633 result_mem ->set_req(_objArray_path, reset_memory());
4634 }
4635 }
4636 // Otherwise, there are no card marks to worry about.
4637 // (We can dispense with card marks if we know the allocation
4638 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4639 // causes the non-eden paths to take compensating steps to
4640 // simulate a fresh allocation, so that no further
4641 // card marks are required in compiled code to initialize
4642 // the object.)
4644 if (!stopped()) {
4645 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4647 // Present the results of the copy.
4648 result_reg->init_req(_array_path, control());
4649 result_val->init_req(_array_path, alloc_obj);
4650 result_i_o ->set_req(_array_path, i_o());
4651 result_mem ->set_req(_array_path, reset_memory());
4652 }
4653 }
4655 // We only go to the instance fast case code if we pass a number of guards.
4656 // The paths which do not pass are accumulated in the slow_region.
4657 RegionNode* slow_region = new (C) RegionNode(1);
4658 record_for_igvn(slow_region);
4659 if (!stopped()) {
4660 // It's an instance (we did array above). Make the slow-path tests.
4661 // If this is a virtual call, we generate a funny guard. We grab
4662 // the vtable entry corresponding to clone() from the target object.
4663 // If the target method which we are calling happens to be the
4664 // Object clone() method, we pass the guard. We do not need this
4665 // guard for non-virtual calls; the caller is known to be the native
4666 // Object clone().
4667 if (is_virtual) {
4668 generate_virtual_guard(obj_klass, slow_region);
4669 }
4671 // The object must be cloneable and must not have a finalizer.
4672 // Both of these conditions may be checked in a single test.
4673 // We could optimize the cloneable test further, but we don't care.
4674 generate_access_flags_guard(obj_klass,
4675 // Test both conditions:
4676 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4677 // Must be cloneable but not finalizer:
4678 JVM_ACC_IS_CLONEABLE,
4679 slow_region);
4680 }
4682 if (!stopped()) {
4683 // It's an instance, and it passed the slow-path tests.
4684 PreserveJVMState pjvms(this);
4685 Node* obj_size = NULL;
4686 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4688 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4690 // Present the results of the slow call.
4691 result_reg->init_req(_instance_path, control());
4692 result_val->init_req(_instance_path, alloc_obj);
4693 result_i_o ->set_req(_instance_path, i_o());
4694 result_mem ->set_req(_instance_path, reset_memory());
4695 }
4697 // Generate code for the slow case. We make a call to clone().
4698 set_control(_gvn.transform(slow_region));
4699 if (!stopped()) {
4700 PreserveJVMState pjvms(this);
4701 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4702 Node* slow_result = set_results_for_java_call(slow_call);
4703 // this->control() comes from set_results_for_java_call
4704 result_reg->init_req(_slow_path, control());
4705 result_val->init_req(_slow_path, slow_result);
4706 result_i_o ->set_req(_slow_path, i_o());
4707 result_mem ->set_req(_slow_path, reset_memory());
4708 }
4710 // Return the combined state.
4711 set_control( _gvn.transform(result_reg) );
4712 set_i_o( _gvn.transform(result_i_o) );
4713 set_all_memory( _gvn.transform(result_mem) );
4714 } //original reexecute and sp are set back here
4716 push(_gvn.transform(result_val));
4718 return true;
4719 }
4721 //------------------------------basictype2arraycopy----------------------------
4722 address LibraryCallKit::basictype2arraycopy(BasicType t,
4723 Node* src_offset,
4724 Node* dest_offset,
4725 bool disjoint_bases,
4726 const char* &name,
4727 bool dest_uninitialized) {
4728 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4729 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4731 bool aligned = false;
4732 bool disjoint = disjoint_bases;
4734 // if the offsets are the same, we can treat the memory regions as
4735 // disjoint, because either the memory regions are in different arrays,
4736 // or they are identical (which we can treat as disjoint.) We can also
4737 // treat a copy with a destination index less that the source index
4738 // as disjoint since a low->high copy will work correctly in this case.
4739 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4740 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4741 // both indices are constants
4742 int s_offs = src_offset_inttype->get_con();
4743 int d_offs = dest_offset_inttype->get_con();
4744 int element_size = type2aelembytes(t);
4745 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4746 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4747 if (s_offs >= d_offs) disjoint = true;
4748 } else if (src_offset == dest_offset && src_offset != NULL) {
4749 // This can occur if the offsets are identical non-constants.
4750 disjoint = true;
4751 }
4753 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4754 }
4757 //------------------------------inline_arraycopy-----------------------
4758 bool LibraryCallKit::inline_arraycopy() {
4759 // Restore the stack and pop off the arguments.
4760 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4761 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4763 Node *src = argument(0);
4764 Node *src_offset = argument(1);
4765 Node *dest = argument(2);
4766 Node *dest_offset = argument(3);
4767 Node *length = argument(4);
4769 // Compile time checks. If any of these checks cannot be verified at compile time,
4770 // we do not make a fast path for this call. Instead, we let the call remain as it
4771 // is. The checks we choose to mandate at compile time are:
4772 //
4773 // (1) src and dest are arrays.
4774 const Type* src_type = src->Value(&_gvn);
4775 const Type* dest_type = dest->Value(&_gvn);
4776 const TypeAryPtr* top_src = src_type->isa_aryptr();
4777 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4778 if (top_src == NULL || top_src->klass() == NULL ||
4779 top_dest == NULL || top_dest->klass() == NULL) {
4780 // Conservatively insert a memory barrier on all memory slices.
4781 // Do not let writes into the source float below the arraycopy.
4782 insert_mem_bar(Op_MemBarCPUOrder);
4784 // Call StubRoutines::generic_arraycopy stub.
4785 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4786 src, src_offset, dest, dest_offset, length);
4788 // Do not let reads from the destination float above the arraycopy.
4789 // Since we cannot type the arrays, we don't know which slices
4790 // might be affected. We could restrict this barrier only to those
4791 // memory slices which pertain to array elements--but don't bother.
4792 if (!InsertMemBarAfterArraycopy)
4793 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4794 insert_mem_bar(Op_MemBarCPUOrder);
4795 return true;
4796 }
4798 // (2) src and dest arrays must have elements of the same BasicType
4799 // Figure out the size and type of the elements we will be copying.
4800 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4801 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4802 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4803 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4805 if (src_elem != dest_elem || dest_elem == T_VOID) {
4806 // The component types are not the same or are not recognized. Punt.
4807 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4808 generate_slow_arraycopy(TypePtr::BOTTOM,
4809 src, src_offset, dest, dest_offset, length,
4810 /*dest_uninitialized*/false);
4811 return true;
4812 }
4814 //---------------------------------------------------------------------------
4815 // We will make a fast path for this call to arraycopy.
4817 // We have the following tests left to perform:
4818 //
4819 // (3) src and dest must not be null.
4820 // (4) src_offset must not be negative.
4821 // (5) dest_offset must not be negative.
4822 // (6) length must not be negative.
4823 // (7) src_offset + length must not exceed length of src.
4824 // (8) dest_offset + length must not exceed length of dest.
4825 // (9) each element of an oop array must be assignable
4827 RegionNode* slow_region = new (C) RegionNode(1);
4828 record_for_igvn(slow_region);
4830 // (3) operands must not be null
4831 // We currently perform our null checks with the do_null_check routine.
4832 // This means that the null exceptions will be reported in the caller
4833 // rather than (correctly) reported inside of the native arraycopy call.
4834 // This should be corrected, given time. We do our null check with the
4835 // stack pointer restored.
4836 _sp += nargs;
4837 src = do_null_check(src, T_ARRAY);
4838 dest = do_null_check(dest, T_ARRAY);
4839 _sp -= nargs;
4841 // (4) src_offset must not be negative.
4842 generate_negative_guard(src_offset, slow_region);
4844 // (5) dest_offset must not be negative.
4845 generate_negative_guard(dest_offset, slow_region);
4847 // (6) length must not be negative (moved to generate_arraycopy()).
4848 // generate_negative_guard(length, slow_region);
4850 // (7) src_offset + length must not exceed length of src.
4851 generate_limit_guard(src_offset, length,
4852 load_array_length(src),
4853 slow_region);
4855 // (8) dest_offset + length must not exceed length of dest.
4856 generate_limit_guard(dest_offset, length,
4857 load_array_length(dest),
4858 slow_region);
4860 // (9) each element of an oop array must be assignable
4861 // The generate_arraycopy subroutine checks this.
4863 // This is where the memory effects are placed:
4864 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4865 generate_arraycopy(adr_type, dest_elem,
4866 src, src_offset, dest, dest_offset, length,
4867 false, false, slow_region);
4869 return true;
4870 }
4872 //-----------------------------generate_arraycopy----------------------
4873 // Generate an optimized call to arraycopy.
4874 // Caller must guard against non-arrays.
4875 // Caller must determine a common array basic-type for both arrays.
4876 // Caller must validate offsets against array bounds.
4877 // The slow_region has already collected guard failure paths
4878 // (such as out of bounds length or non-conformable array types).
4879 // The generated code has this shape, in general:
4880 //
4881 // if (length == 0) return // via zero_path
4882 // slowval = -1
4883 // if (types unknown) {
4884 // slowval = call generic copy loop
4885 // if (slowval == 0) return // via checked_path
4886 // } else if (indexes in bounds) {
4887 // if ((is object array) && !(array type check)) {
4888 // slowval = call checked copy loop
4889 // if (slowval == 0) return // via checked_path
4890 // } else {
4891 // call bulk copy loop
4892 // return // via fast_path
4893 // }
4894 // }
4895 // // adjust params for remaining work:
4896 // if (slowval != -1) {
4897 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4898 // }
4899 // slow_region:
4900 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4901 // return // via slow_call_path
4902 //
4903 // This routine is used from several intrinsics: System.arraycopy,
4904 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4905 //
4906 void
4907 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4908 BasicType basic_elem_type,
4909 Node* src, Node* src_offset,
4910 Node* dest, Node* dest_offset,
4911 Node* copy_length,
4912 bool disjoint_bases,
4913 bool length_never_negative,
4914 RegionNode* slow_region) {
4916 if (slow_region == NULL) {
4917 slow_region = new(C) RegionNode(1);
4918 record_for_igvn(slow_region);
4919 }
4921 Node* original_dest = dest;
4922 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4923 bool dest_uninitialized = false;
4925 // See if this is the initialization of a newly-allocated array.
4926 // If so, we will take responsibility here for initializing it to zero.
4927 // (Note: Because tightly_coupled_allocation performs checks on the
4928 // out-edges of the dest, we need to avoid making derived pointers
4929 // from it until we have checked its uses.)
4930 if (ReduceBulkZeroing
4931 && !ZeroTLAB // pointless if already zeroed
4932 && basic_elem_type != T_CONFLICT // avoid corner case
4933 && !src->eqv_uncast(dest)
4934 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4935 != NULL)
4936 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4937 && alloc->maybe_set_complete(&_gvn)) {
4938 // "You break it, you buy it."
4939 InitializeNode* init = alloc->initialization();
4940 assert(init->is_complete(), "we just did this");
4941 init->set_complete_with_arraycopy();
4942 assert(dest->is_CheckCastPP(), "sanity");
4943 assert(dest->in(0)->in(0) == init, "dest pinned");
4944 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4945 // From this point on, every exit path is responsible for
4946 // initializing any non-copied parts of the object to zero.
4947 // Also, if this flag is set we make sure that arraycopy interacts properly
4948 // with G1, eliding pre-barriers. See CR 6627983.
4949 dest_uninitialized = true;
4950 } else {
4951 // No zeroing elimination here.
4952 alloc = NULL;
4953 //original_dest = dest;
4954 //dest_uninitialized = false;
4955 }
4957 // Results are placed here:
4958 enum { fast_path = 1, // normal void-returning assembly stub
4959 checked_path = 2, // special assembly stub with cleanup
4960 slow_call_path = 3, // something went wrong; call the VM
4961 zero_path = 4, // bypass when length of copy is zero
4962 bcopy_path = 5, // copy primitive array by 64-bit blocks
4963 PATH_LIMIT = 6
4964 };
4965 RegionNode* result_region = new(C) RegionNode(PATH_LIMIT);
4966 PhiNode* result_i_o = new(C) PhiNode(result_region, Type::ABIO);
4967 PhiNode* result_memory = new(C) PhiNode(result_region, Type::MEMORY, adr_type);
4968 record_for_igvn(result_region);
4969 _gvn.set_type_bottom(result_i_o);
4970 _gvn.set_type_bottom(result_memory);
4971 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4973 // The slow_control path:
4974 Node* slow_control;
4975 Node* slow_i_o = i_o();
4976 Node* slow_mem = memory(adr_type);
4977 debug_only(slow_control = (Node*) badAddress);
4979 // Checked control path:
4980 Node* checked_control = top();
4981 Node* checked_mem = NULL;
4982 Node* checked_i_o = NULL;
4983 Node* checked_value = NULL;
4985 if (basic_elem_type == T_CONFLICT) {
4986 assert(!dest_uninitialized, "");
4987 Node* cv = generate_generic_arraycopy(adr_type,
4988 src, src_offset, dest, dest_offset,
4989 copy_length, dest_uninitialized);
4990 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4991 checked_control = control();
4992 checked_i_o = i_o();
4993 checked_mem = memory(adr_type);
4994 checked_value = cv;
4995 set_control(top()); // no fast path
4996 }
4998 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4999 if (not_pos != NULL) {
5000 PreserveJVMState pjvms(this);
5001 set_control(not_pos);
5003 // (6) length must not be negative.
5004 if (!length_never_negative) {
5005 generate_negative_guard(copy_length, slow_region);
5006 }
5008 // copy_length is 0.
5009 if (!stopped() && dest_uninitialized) {
5010 Node* dest_length = alloc->in(AllocateNode::ALength);
5011 if (copy_length->eqv_uncast(dest_length)
5012 || _gvn.find_int_con(dest_length, 1) <= 0) {
5013 // There is no zeroing to do. No need for a secondary raw memory barrier.
5014 } else {
5015 // Clear the whole thing since there are no source elements to copy.
5016 generate_clear_array(adr_type, dest, basic_elem_type,
5017 intcon(0), NULL,
5018 alloc->in(AllocateNode::AllocSize));
5019 // Use a secondary InitializeNode as raw memory barrier.
5020 // Currently it is needed only on this path since other
5021 // paths have stub or runtime calls as raw memory barriers.
5022 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
5023 Compile::AliasIdxRaw,
5024 top())->as_Initialize();
5025 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
5026 }
5027 }
5029 // Present the results of the fast call.
5030 result_region->init_req(zero_path, control());
5031 result_i_o ->init_req(zero_path, i_o());
5032 result_memory->init_req(zero_path, memory(adr_type));
5033 }
5035 if (!stopped() && dest_uninitialized) {
5036 // We have to initialize the *uncopied* part of the array to zero.
5037 // The copy destination is the slice dest[off..off+len]. The other slices
5038 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
5039 Node* dest_size = alloc->in(AllocateNode::AllocSize);
5040 Node* dest_length = alloc->in(AllocateNode::ALength);
5041 Node* dest_tail = _gvn.transform( new(C) AddINode(dest_offset,
5042 copy_length) );
5044 // If there is a head section that needs zeroing, do it now.
5045 if (find_int_con(dest_offset, -1) != 0) {
5046 generate_clear_array(adr_type, dest, basic_elem_type,
5047 intcon(0), dest_offset,
5048 NULL);
5049 }
5051 // Next, perform a dynamic check on the tail length.
5052 // It is often zero, and we can win big if we prove this.
5053 // There are two wins: Avoid generating the ClearArray
5054 // with its attendant messy index arithmetic, and upgrade
5055 // the copy to a more hardware-friendly word size of 64 bits.
5056 Node* tail_ctl = NULL;
5057 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
5058 Node* cmp_lt = _gvn.transform( new(C) CmpINode(dest_tail, dest_length) );
5059 Node* bol_lt = _gvn.transform( new(C) BoolNode(cmp_lt, BoolTest::lt) );
5060 tail_ctl = generate_slow_guard(bol_lt, NULL);
5061 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
5062 }
5064 // At this point, let's assume there is no tail.
5065 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
5066 // There is no tail. Try an upgrade to a 64-bit copy.
5067 bool didit = false;
5068 { PreserveJVMState pjvms(this);
5069 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
5070 src, src_offset, dest, dest_offset,
5071 dest_size, dest_uninitialized);
5072 if (didit) {
5073 // Present the results of the block-copying fast call.
5074 result_region->init_req(bcopy_path, control());
5075 result_i_o ->init_req(bcopy_path, i_o());
5076 result_memory->init_req(bcopy_path, memory(adr_type));
5077 }
5078 }
5079 if (didit)
5080 set_control(top()); // no regular fast path
5081 }
5083 // Clear the tail, if any.
5084 if (tail_ctl != NULL) {
5085 Node* notail_ctl = stopped() ? NULL : control();
5086 set_control(tail_ctl);
5087 if (notail_ctl == NULL) {
5088 generate_clear_array(adr_type, dest, basic_elem_type,
5089 dest_tail, NULL,
5090 dest_size);
5091 } else {
5092 // Make a local merge.
5093 Node* done_ctl = new(C) RegionNode(3);
5094 Node* done_mem = new(C) PhiNode(done_ctl, Type::MEMORY, adr_type);
5095 done_ctl->init_req(1, notail_ctl);
5096 done_mem->init_req(1, memory(adr_type));
5097 generate_clear_array(adr_type, dest, basic_elem_type,
5098 dest_tail, NULL,
5099 dest_size);
5100 done_ctl->init_req(2, control());
5101 done_mem->init_req(2, memory(adr_type));
5102 set_control( _gvn.transform(done_ctl) );
5103 set_memory( _gvn.transform(done_mem), adr_type );
5104 }
5105 }
5106 }
5108 BasicType copy_type = basic_elem_type;
5109 assert(basic_elem_type != T_ARRAY, "caller must fix this");
5110 if (!stopped() && copy_type == T_OBJECT) {
5111 // If src and dest have compatible element types, we can copy bits.
5112 // Types S[] and D[] are compatible if D is a supertype of S.
5113 //
5114 // If they are not, we will use checked_oop_disjoint_arraycopy,
5115 // which performs a fast optimistic per-oop check, and backs off
5116 // further to JVM_ArrayCopy on the first per-oop check that fails.
5117 // (Actually, we don't move raw bits only; the GC requires card marks.)
5119 // Get the Klass* for both src and dest
5120 Node* src_klass = load_object_klass(src);
5121 Node* dest_klass = load_object_klass(dest);
5123 // Generate the subtype check.
5124 // This might fold up statically, or then again it might not.
5125 //
5126 // Non-static example: Copying List<String>.elements to a new String[].
5127 // The backing store for a List<String> is always an Object[],
5128 // but its elements are always type String, if the generic types
5129 // are correct at the source level.
5130 //
5131 // Test S[] against D[], not S against D, because (probably)
5132 // the secondary supertype cache is less busy for S[] than S.
5133 // This usually only matters when D is an interface.
5134 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
5135 // Plug failing path into checked_oop_disjoint_arraycopy
5136 if (not_subtype_ctrl != top()) {
5137 PreserveJVMState pjvms(this);
5138 set_control(not_subtype_ctrl);
5139 // (At this point we can assume disjoint_bases, since types differ.)
5140 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
5141 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
5142 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
5143 Node* dest_elem_klass = _gvn.transform(n1);
5144 Node* cv = generate_checkcast_arraycopy(adr_type,
5145 dest_elem_klass,
5146 src, src_offset, dest, dest_offset,
5147 ConvI2X(copy_length), dest_uninitialized);
5148 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
5149 checked_control = control();
5150 checked_i_o = i_o();
5151 checked_mem = memory(adr_type);
5152 checked_value = cv;
5153 }
5154 // At this point we know we do not need type checks on oop stores.
5156 // Let's see if we need card marks:
5157 if (alloc != NULL && use_ReduceInitialCardMarks()) {
5158 // If we do not need card marks, copy using the jint or jlong stub.
5159 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
5160 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
5161 "sizes agree");
5162 }
5163 }
5165 if (!stopped()) {
5166 // Generate the fast path, if possible.
5167 PreserveJVMState pjvms(this);
5168 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5169 src, src_offset, dest, dest_offset,
5170 ConvI2X(copy_length), dest_uninitialized);
5172 // Present the results of the fast call.
5173 result_region->init_req(fast_path, control());
5174 result_i_o ->init_req(fast_path, i_o());
5175 result_memory->init_req(fast_path, memory(adr_type));
5176 }
5178 // Here are all the slow paths up to this point, in one bundle:
5179 slow_control = top();
5180 if (slow_region != NULL)
5181 slow_control = _gvn.transform(slow_region);
5182 debug_only(slow_region = (RegionNode*)badAddress);
5184 set_control(checked_control);
5185 if (!stopped()) {
5186 // Clean up after the checked call.
5187 // The returned value is either 0 or -1^K,
5188 // where K = number of partially transferred array elements.
5189 Node* cmp = _gvn.transform( new(C) CmpINode(checked_value, intcon(0)) );
5190 Node* bol = _gvn.transform( new(C) BoolNode(cmp, BoolTest::eq) );
5191 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5193 // If it is 0, we are done, so transfer to the end.
5194 Node* checks_done = _gvn.transform( new(C) IfTrueNode(iff) );
5195 result_region->init_req(checked_path, checks_done);
5196 result_i_o ->init_req(checked_path, checked_i_o);
5197 result_memory->init_req(checked_path, checked_mem);
5199 // If it is not zero, merge into the slow call.
5200 set_control( _gvn.transform( new(C) IfFalseNode(iff) ));
5201 RegionNode* slow_reg2 = new(C) RegionNode(3);
5202 PhiNode* slow_i_o2 = new(C) PhiNode(slow_reg2, Type::ABIO);
5203 PhiNode* slow_mem2 = new(C) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5204 record_for_igvn(slow_reg2);
5205 slow_reg2 ->init_req(1, slow_control);
5206 slow_i_o2 ->init_req(1, slow_i_o);
5207 slow_mem2 ->init_req(1, slow_mem);
5208 slow_reg2 ->init_req(2, control());
5209 slow_i_o2 ->init_req(2, checked_i_o);
5210 slow_mem2 ->init_req(2, checked_mem);
5212 slow_control = _gvn.transform(slow_reg2);
5213 slow_i_o = _gvn.transform(slow_i_o2);
5214 slow_mem = _gvn.transform(slow_mem2);
5216 if (alloc != NULL) {
5217 // We'll restart from the very beginning, after zeroing the whole thing.
5218 // This can cause double writes, but that's OK since dest is brand new.
5219 // So we ignore the low 31 bits of the value returned from the stub.
5220 } else {
5221 // We must continue the copy exactly where it failed, or else
5222 // another thread might see the wrong number of writes to dest.
5223 Node* checked_offset = _gvn.transform( new(C) XorINode(checked_value, intcon(-1)) );
5224 Node* slow_offset = new(C) PhiNode(slow_reg2, TypeInt::INT);
5225 slow_offset->init_req(1, intcon(0));
5226 slow_offset->init_req(2, checked_offset);
5227 slow_offset = _gvn.transform(slow_offset);
5229 // Adjust the arguments by the conditionally incoming offset.
5230 Node* src_off_plus = _gvn.transform( new(C) AddINode(src_offset, slow_offset) );
5231 Node* dest_off_plus = _gvn.transform( new(C) AddINode(dest_offset, slow_offset) );
5232 Node* length_minus = _gvn.transform( new(C) SubINode(copy_length, slow_offset) );
5234 // Tweak the node variables to adjust the code produced below:
5235 src_offset = src_off_plus;
5236 dest_offset = dest_off_plus;
5237 copy_length = length_minus;
5238 }
5239 }
5241 set_control(slow_control);
5242 if (!stopped()) {
5243 // Generate the slow path, if needed.
5244 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5246 set_memory(slow_mem, adr_type);
5247 set_i_o(slow_i_o);
5249 if (dest_uninitialized) {
5250 generate_clear_array(adr_type, dest, basic_elem_type,
5251 intcon(0), NULL,
5252 alloc->in(AllocateNode::AllocSize));
5253 }
5255 generate_slow_arraycopy(adr_type,
5256 src, src_offset, dest, dest_offset,
5257 copy_length, /*dest_uninitialized*/false);
5259 result_region->init_req(slow_call_path, control());
5260 result_i_o ->init_req(slow_call_path, i_o());
5261 result_memory->init_req(slow_call_path, memory(adr_type));
5262 }
5264 // Remove unused edges.
5265 for (uint i = 1; i < result_region->req(); i++) {
5266 if (result_region->in(i) == NULL)
5267 result_region->init_req(i, top());
5268 }
5270 // Finished; return the combined state.
5271 set_control( _gvn.transform(result_region) );
5272 set_i_o( _gvn.transform(result_i_o) );
5273 set_memory( _gvn.transform(result_memory), adr_type );
5275 // The memory edges above are precise in order to model effects around
5276 // array copies accurately to allow value numbering of field loads around
5277 // arraycopy. Such field loads, both before and after, are common in Java
5278 // collections and similar classes involving header/array data structures.
5279 //
5280 // But with low number of register or when some registers are used or killed
5281 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5282 // The next memory barrier is added to avoid it. If the arraycopy can be
5283 // optimized away (which it can, sometimes) then we can manually remove
5284 // the membar also.
5285 //
5286 // Do not let reads from the cloned object float above the arraycopy.
5287 if (alloc != NULL) {
5288 // Do not let stores that initialize this object be reordered with
5289 // a subsequent store that would make this object accessible by
5290 // other threads.
5291 // Record what AllocateNode this StoreStore protects so that
5292 // escape analysis can go from the MemBarStoreStoreNode to the
5293 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5294 // based on the escape status of the AllocateNode.
5295 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5296 } else if (InsertMemBarAfterArraycopy)
5297 insert_mem_bar(Op_MemBarCPUOrder);
5298 }
5301 // Helper function which determines if an arraycopy immediately follows
5302 // an allocation, with no intervening tests or other escapes for the object.
5303 AllocateArrayNode*
5304 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5305 RegionNode* slow_region) {
5306 if (stopped()) return NULL; // no fast path
5307 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5309 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5310 if (alloc == NULL) return NULL;
5312 Node* rawmem = memory(Compile::AliasIdxRaw);
5313 // Is the allocation's memory state untouched?
5314 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5315 // Bail out if there have been raw-memory effects since the allocation.
5316 // (Example: There might have been a call or safepoint.)
5317 return NULL;
5318 }
5319 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5320 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5321 return NULL;
5322 }
5324 // There must be no unexpected observers of this allocation.
5325 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5326 Node* obs = ptr->fast_out(i);
5327 if (obs != this->map()) {
5328 return NULL;
5329 }
5330 }
5332 // This arraycopy must unconditionally follow the allocation of the ptr.
5333 Node* alloc_ctl = ptr->in(0);
5334 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5336 Node* ctl = control();
5337 while (ctl != alloc_ctl) {
5338 // There may be guards which feed into the slow_region.
5339 // Any other control flow means that we might not get a chance
5340 // to finish initializing the allocated object.
5341 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5342 IfNode* iff = ctl->in(0)->as_If();
5343 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5344 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5345 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5346 ctl = iff->in(0); // This test feeds the known slow_region.
5347 continue;
5348 }
5349 // One more try: Various low-level checks bottom out in
5350 // uncommon traps. If the debug-info of the trap omits
5351 // any reference to the allocation, as we've already
5352 // observed, then there can be no objection to the trap.
5353 bool found_trap = false;
5354 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5355 Node* obs = not_ctl->fast_out(j);
5356 if (obs->in(0) == not_ctl && obs->is_Call() &&
5357 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5358 found_trap = true; break;
5359 }
5360 }
5361 if (found_trap) {
5362 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5363 continue;
5364 }
5365 }
5366 return NULL;
5367 }
5369 // If we get this far, we have an allocation which immediately
5370 // precedes the arraycopy, and we can take over zeroing the new object.
5371 // The arraycopy will finish the initialization, and provide
5372 // a new control state to which we will anchor the destination pointer.
5374 return alloc;
5375 }
5377 // Helper for initialization of arrays, creating a ClearArray.
5378 // It writes zero bits in [start..end), within the body of an array object.
5379 // The memory effects are all chained onto the 'adr_type' alias category.
5380 //
5381 // Since the object is otherwise uninitialized, we are free
5382 // to put a little "slop" around the edges of the cleared area,
5383 // as long as it does not go back into the array's header,
5384 // or beyond the array end within the heap.
5385 //
5386 // The lower edge can be rounded down to the nearest jint and the
5387 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5388 //
5389 // Arguments:
5390 // adr_type memory slice where writes are generated
5391 // dest oop of the destination array
5392 // basic_elem_type element type of the destination
5393 // slice_idx array index of first element to store
5394 // slice_len number of elements to store (or NULL)
5395 // dest_size total size in bytes of the array object
5396 //
5397 // Exactly one of slice_len or dest_size must be non-NULL.
5398 // If dest_size is non-NULL, zeroing extends to the end of the object.
5399 // If slice_len is non-NULL, the slice_idx value must be a constant.
5400 void
5401 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5402 Node* dest,
5403 BasicType basic_elem_type,
5404 Node* slice_idx,
5405 Node* slice_len,
5406 Node* dest_size) {
5407 // one or the other but not both of slice_len and dest_size:
5408 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5409 if (slice_len == NULL) slice_len = top();
5410 if (dest_size == NULL) dest_size = top();
5412 // operate on this memory slice:
5413 Node* mem = memory(adr_type); // memory slice to operate on
5415 // scaling and rounding of indexes:
5416 int scale = exact_log2(type2aelembytes(basic_elem_type));
5417 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5418 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5419 int bump_bit = (-1 << scale) & BytesPerInt;
5421 // determine constant starts and ends
5422 const intptr_t BIG_NEG = -128;
5423 assert(BIG_NEG + 2*abase < 0, "neg enough");
5424 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5425 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5426 if (slice_len_con == 0) {
5427 return; // nothing to do here
5428 }
5429 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5430 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5431 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5432 assert(end_con < 0, "not two cons");
5433 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5434 BytesPerLong);
5435 }
5437 if (start_con >= 0 && end_con >= 0) {
5438 // Constant start and end. Simple.
5439 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5440 start_con, end_con, &_gvn);
5441 } else if (start_con >= 0 && dest_size != top()) {
5442 // Constant start, pre-rounded end after the tail of the array.
5443 Node* end = dest_size;
5444 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5445 start_con, end, &_gvn);
5446 } else if (start_con >= 0 && slice_len != top()) {
5447 // Constant start, non-constant end. End needs rounding up.
5448 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5449 intptr_t end_base = abase + (slice_idx_con << scale);
5450 int end_round = (-1 << scale) & (BytesPerLong - 1);
5451 Node* end = ConvI2X(slice_len);
5452 if (scale != 0)
5453 end = _gvn.transform( new(C) LShiftXNode(end, intcon(scale) ));
5454 end_base += end_round;
5455 end = _gvn.transform( new(C) AddXNode(end, MakeConX(end_base)) );
5456 end = _gvn.transform( new(C) AndXNode(end, MakeConX(~end_round)) );
5457 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5458 start_con, end, &_gvn);
5459 } else if (start_con < 0 && dest_size != top()) {
5460 // Non-constant start, pre-rounded end after the tail of the array.
5461 // This is almost certainly a "round-to-end" operation.
5462 Node* start = slice_idx;
5463 start = ConvI2X(start);
5464 if (scale != 0)
5465 start = _gvn.transform( new(C) LShiftXNode( start, intcon(scale) ));
5466 start = _gvn.transform( new(C) AddXNode(start, MakeConX(abase)) );
5467 if ((bump_bit | clear_low) != 0) {
5468 int to_clear = (bump_bit | clear_low);
5469 // Align up mod 8, then store a jint zero unconditionally
5470 // just before the mod-8 boundary.
5471 if (((abase + bump_bit) & ~to_clear) - bump_bit
5472 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5473 bump_bit = 0;
5474 assert((abase & to_clear) == 0, "array base must be long-aligned");
5475 } else {
5476 // Bump 'start' up to (or past) the next jint boundary:
5477 start = _gvn.transform( new(C) AddXNode(start, MakeConX(bump_bit)) );
5478 assert((abase & clear_low) == 0, "array base must be int-aligned");
5479 }
5480 // Round bumped 'start' down to jlong boundary in body of array.
5481 start = _gvn.transform( new(C) AndXNode(start, MakeConX(~to_clear)) );
5482 if (bump_bit != 0) {
5483 // Store a zero to the immediately preceding jint:
5484 Node* x1 = _gvn.transform( new(C) AddXNode(start, MakeConX(-bump_bit)) );
5485 Node* p1 = basic_plus_adr(dest, x1);
5486 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5487 mem = _gvn.transform(mem);
5488 }
5489 }
5490 Node* end = dest_size; // pre-rounded
5491 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5492 start, end, &_gvn);
5493 } else {
5494 // Non-constant start, unrounded non-constant end.
5495 // (Nobody zeroes a random midsection of an array using this routine.)
5496 ShouldNotReachHere(); // fix caller
5497 }
5499 // Done.
5500 set_memory(mem, adr_type);
5501 }
5504 bool
5505 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5506 BasicType basic_elem_type,
5507 AllocateNode* alloc,
5508 Node* src, Node* src_offset,
5509 Node* dest, Node* dest_offset,
5510 Node* dest_size, bool dest_uninitialized) {
5511 // See if there is an advantage from block transfer.
5512 int scale = exact_log2(type2aelembytes(basic_elem_type));
5513 if (scale >= LogBytesPerLong)
5514 return false; // it is already a block transfer
5516 // Look at the alignment of the starting offsets.
5517 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5519 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5520 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5521 if (src_off_con < 0 || dest_off_con < 0)
5522 // At present, we can only understand constants.
5523 return false;
5525 intptr_t src_off = abase + (src_off_con << scale);
5526 intptr_t dest_off = abase + (dest_off_con << scale);
5528 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5529 // Non-aligned; too bad.
5530 // One more chance: Pick off an initial 32-bit word.
5531 // This is a common case, since abase can be odd mod 8.
5532 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5533 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5534 Node* sptr = basic_plus_adr(src, src_off);
5535 Node* dptr = basic_plus_adr(dest, dest_off);
5536 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5537 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5538 src_off += BytesPerInt;
5539 dest_off += BytesPerInt;
5540 } else {
5541 return false;
5542 }
5543 }
5544 assert(src_off % BytesPerLong == 0, "");
5545 assert(dest_off % BytesPerLong == 0, "");
5547 // Do this copy by giant steps.
5548 Node* sptr = basic_plus_adr(src, src_off);
5549 Node* dptr = basic_plus_adr(dest, dest_off);
5550 Node* countx = dest_size;
5551 countx = _gvn.transform( new (C) SubXNode(countx, MakeConX(dest_off)) );
5552 countx = _gvn.transform( new (C) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5554 bool disjoint_bases = true; // since alloc != NULL
5555 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5556 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5558 return true;
5559 }
5562 // Helper function; generates code for the slow case.
5563 // We make a call to a runtime method which emulates the native method,
5564 // but without the native wrapper overhead.
5565 void
5566 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5567 Node* src, Node* src_offset,
5568 Node* dest, Node* dest_offset,
5569 Node* copy_length, bool dest_uninitialized) {
5570 assert(!dest_uninitialized, "Invariant");
5571 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5572 OptoRuntime::slow_arraycopy_Type(),
5573 OptoRuntime::slow_arraycopy_Java(),
5574 "slow_arraycopy", adr_type,
5575 src, src_offset, dest, dest_offset,
5576 copy_length);
5578 // Handle exceptions thrown by this fellow:
5579 make_slow_call_ex(call, env()->Throwable_klass(), false);
5580 }
5582 // Helper function; generates code for cases requiring runtime checks.
5583 Node*
5584 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5585 Node* dest_elem_klass,
5586 Node* src, Node* src_offset,
5587 Node* dest, Node* dest_offset,
5588 Node* copy_length, bool dest_uninitialized) {
5589 if (stopped()) return NULL;
5591 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5592 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5593 return NULL;
5594 }
5596 // Pick out the parameters required to perform a store-check
5597 // for the target array. This is an optimistic check. It will
5598 // look in each non-null element's class, at the desired klass's
5599 // super_check_offset, for the desired klass.
5600 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5601 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5602 Node* n3 = new(C) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5603 Node* check_offset = ConvI2X(_gvn.transform(n3));
5604 Node* check_value = dest_elem_klass;
5606 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5607 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5609 // (We know the arrays are never conjoint, because their types differ.)
5610 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5611 OptoRuntime::checkcast_arraycopy_Type(),
5612 copyfunc_addr, "checkcast_arraycopy", adr_type,
5613 // five arguments, of which two are
5614 // intptr_t (jlong in LP64)
5615 src_start, dest_start,
5616 copy_length XTOP,
5617 check_offset XTOP,
5618 check_value);
5620 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5621 }
5624 // Helper function; generates code for cases requiring runtime checks.
5625 Node*
5626 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5627 Node* src, Node* src_offset,
5628 Node* dest, Node* dest_offset,
5629 Node* copy_length, bool dest_uninitialized) {
5630 assert(!dest_uninitialized, "Invariant");
5631 if (stopped()) return NULL;
5632 address copyfunc_addr = StubRoutines::generic_arraycopy();
5633 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5634 return NULL;
5635 }
5637 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5638 OptoRuntime::generic_arraycopy_Type(),
5639 copyfunc_addr, "generic_arraycopy", adr_type,
5640 src, src_offset, dest, dest_offset, copy_length);
5642 return _gvn.transform(new (C) ProjNode(call, TypeFunc::Parms));
5643 }
5645 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5646 void
5647 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5648 BasicType basic_elem_type,
5649 bool disjoint_bases,
5650 Node* src, Node* src_offset,
5651 Node* dest, Node* dest_offset,
5652 Node* copy_length, bool dest_uninitialized) {
5653 if (stopped()) return; // nothing to do
5655 Node* src_start = src;
5656 Node* dest_start = dest;
5657 if (src_offset != NULL || dest_offset != NULL) {
5658 assert(src_offset != NULL && dest_offset != NULL, "");
5659 src_start = array_element_address(src, src_offset, basic_elem_type);
5660 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5661 }
5663 // Figure out which arraycopy runtime method to call.
5664 const char* copyfunc_name = "arraycopy";
5665 address copyfunc_addr =
5666 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5667 disjoint_bases, copyfunc_name, dest_uninitialized);
5669 // Call it. Note that the count_ix value is not scaled to a byte-size.
5670 make_runtime_call(RC_LEAF|RC_NO_FP,
5671 OptoRuntime::fast_arraycopy_Type(),
5672 copyfunc_addr, copyfunc_name, adr_type,
5673 src_start, dest_start, copy_length XTOP);
5674 }
5676 //----------------------------inline_reference_get----------------------------
5678 bool LibraryCallKit::inline_reference_get() {
5679 const int nargs = 1; // self
5681 guarantee(java_lang_ref_Reference::referent_offset > 0,
5682 "should have already been set");
5684 int referent_offset = java_lang_ref_Reference::referent_offset;
5686 // Restore the stack and pop off the argument
5687 _sp += nargs;
5688 Node *reference_obj = pop();
5690 // Null check on self without removing any arguments.
5691 _sp += nargs;
5692 reference_obj = do_null_check(reference_obj, T_OBJECT);
5693 _sp -= nargs;;
5695 if (stopped()) return true;
5697 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5699 ciInstanceKlass* klass = env()->Object_klass();
5700 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5702 Node* no_ctrl = NULL;
5703 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5705 // Use the pre-barrier to record the value in the referent field
5706 pre_barrier(false /* do_load */,
5707 control(),
5708 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5709 result /* pre_val */,
5710 T_OBJECT);
5712 // Add memory barrier to prevent commoning reads from this field
5713 // across safepoint since GC can change its value.
5714 insert_mem_bar(Op_MemBarCPUOrder);
5716 push(result);
5717 return true;
5718 }
5721 Node * LibraryCallKit::load_field_from_object(Node * fromObj, const char * fieldName, const char * fieldTypeString,
5722 bool is_exact=true, bool is_static=false) {
5724 const TypeInstPtr* tinst = _gvn.type(fromObj)->isa_instptr();
5725 assert(tinst != NULL, "obj is null");
5726 assert(tinst->klass()->is_loaded(), "obj is not loaded");
5727 assert(!is_exact || tinst->klass_is_exact(), "klass not exact");
5729 ciField* field = tinst->klass()->as_instance_klass()->get_field_by_name(ciSymbol::make(fieldName),
5730 ciSymbol::make(fieldTypeString),
5731 is_static);
5732 if (field == NULL) return (Node *) NULL;
5733 assert (field != NULL, "undefined field");
5735 // Next code copied from Parse::do_get_xxx():
5737 // Compute address and memory type.
5738 int offset = field->offset_in_bytes();
5739 bool is_vol = field->is_volatile();
5740 ciType* field_klass = field->type();
5741 assert(field_klass->is_loaded(), "should be loaded");
5742 const TypePtr* adr_type = C->alias_type(field)->adr_type();
5743 Node *adr = basic_plus_adr(fromObj, fromObj, offset);
5744 BasicType bt = field->layout_type();
5746 // Build the resultant type of the load
5747 const Type *type = TypeOopPtr::make_from_klass(field_klass->as_klass());
5749 // Build the load.
5750 Node* loadedField = make_load(NULL, adr, type, bt, adr_type, is_vol);
5751 return loadedField;
5752 }
5755 //------------------------------inline_aescrypt_Block-----------------------
5756 bool LibraryCallKit::inline_aescrypt_Block(vmIntrinsics::ID id) {
5757 address stubAddr;
5758 const char *stubName;
5759 assert(UseAES, "need AES instruction support");
5761 switch(id) {
5762 case vmIntrinsics::_aescrypt_encryptBlock:
5763 stubAddr = StubRoutines::aescrypt_encryptBlock();
5764 stubName = "aescrypt_encryptBlock";
5765 break;
5766 case vmIntrinsics::_aescrypt_decryptBlock:
5767 stubAddr = StubRoutines::aescrypt_decryptBlock();
5768 stubName = "aescrypt_decryptBlock";
5769 break;
5770 }
5771 if (stubAddr == NULL) return false;
5773 // Restore the stack and pop off the arguments.
5774 int nargs = 5; // this + 2 oop/offset combos
5775 assert(callee()->signature()->size() == nargs-1, "encryptBlock has 4 arguments");
5777 Node *aescrypt_object = argument(0);
5778 Node *src = argument(1);
5779 Node *src_offset = argument(2);
5780 Node *dest = argument(3);
5781 Node *dest_offset = argument(4);
5783 // (1) src and dest are arrays.
5784 const Type* src_type = src->Value(&_gvn);
5785 const Type* dest_type = dest->Value(&_gvn);
5786 const TypeAryPtr* top_src = src_type->isa_aryptr();
5787 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5788 assert (top_src != NULL && top_src->klass() != NULL && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5790 // for the quick and dirty code we will skip all the checks.
5791 // we are just trying to get the call to be generated.
5792 Node* src_start = src;
5793 Node* dest_start = dest;
5794 if (src_offset != NULL || dest_offset != NULL) {
5795 assert(src_offset != NULL && dest_offset != NULL, "");
5796 src_start = array_element_address(src, src_offset, T_BYTE);
5797 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5798 }
5800 // now need to get the start of its expanded key array
5801 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5802 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5803 if (k_start == NULL) return false;
5805 // Call the stub.
5806 make_runtime_call(RC_LEAF|RC_NO_FP, OptoRuntime::aescrypt_block_Type(),
5807 stubAddr, stubName, TypePtr::BOTTOM,
5808 src_start, dest_start, k_start);
5810 return true;
5811 }
5813 //------------------------------inline_cipherBlockChaining_AESCrypt-----------------------
5814 bool LibraryCallKit::inline_cipherBlockChaining_AESCrypt(vmIntrinsics::ID id) {
5815 address stubAddr;
5816 const char *stubName;
5818 assert(UseAES, "need AES instruction support");
5820 switch(id) {
5821 case vmIntrinsics::_cipherBlockChaining_encryptAESCrypt:
5822 stubAddr = StubRoutines::cipherBlockChaining_encryptAESCrypt();
5823 stubName = "cipherBlockChaining_encryptAESCrypt";
5824 break;
5825 case vmIntrinsics::_cipherBlockChaining_decryptAESCrypt:
5826 stubAddr = StubRoutines::cipherBlockChaining_decryptAESCrypt();
5827 stubName = "cipherBlockChaining_decryptAESCrypt";
5828 break;
5829 }
5830 if (stubAddr == NULL) return false;
5833 // Restore the stack and pop off the arguments.
5834 int nargs = 6; // this + oop/offset + len + oop/offset
5835 assert(callee()->signature()->size() == nargs-1, "wrong number of arguments");
5836 Node *cipherBlockChaining_object = argument(0);
5837 Node *src = argument(1);
5838 Node *src_offset = argument(2);
5839 Node *len = argument(3);
5840 Node *dest = argument(4);
5841 Node *dest_offset = argument(5);
5843 // (1) src and dest are arrays.
5844 const Type* src_type = src->Value(&_gvn);
5845 const Type* dest_type = dest->Value(&_gvn);
5846 const TypeAryPtr* top_src = src_type->isa_aryptr();
5847 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
5848 assert (top_src != NULL && top_src->klass() != NULL
5849 && top_dest != NULL && top_dest->klass() != NULL, "args are strange");
5851 // checks are the responsibility of the caller
5852 Node* src_start = src;
5853 Node* dest_start = dest;
5854 if (src_offset != NULL || dest_offset != NULL) {
5855 assert(src_offset != NULL && dest_offset != NULL, "");
5856 src_start = array_element_address(src, src_offset, T_BYTE);
5857 dest_start = array_element_address(dest, dest_offset, T_BYTE);
5858 }
5860 // if we are in this set of code, we "know" the embeddedCipher is an AESCrypt object
5861 // (because of the predicated logic executed earlier).
5862 // so we cast it here safely.
5863 // this requires a newer class file that has this array as littleEndian ints, otherwise we revert to java
5865 Node* embeddedCipherObj = load_field_from_object(cipherBlockChaining_object, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5866 if (embeddedCipherObj == NULL) return false;
5868 // cast it to what we know it will be at runtime
5869 const TypeInstPtr* tinst = _gvn.type(cipherBlockChaining_object)->isa_instptr();
5870 assert(tinst != NULL, "CBC obj is null");
5871 assert(tinst->klass()->is_loaded(), "CBC obj is not loaded");
5872 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5873 if (!klass_AESCrypt->is_loaded()) return false;
5875 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5876 const TypeKlassPtr* aklass = TypeKlassPtr::make(instklass_AESCrypt);
5877 const TypeOopPtr* xtype = aklass->as_instance_type();
5878 Node* aescrypt_object = new(C) CheckCastPPNode(control(), embeddedCipherObj, xtype);
5879 aescrypt_object = _gvn.transform(aescrypt_object);
5881 // we need to get the start of the aescrypt_object's expanded key array
5882 Node* k_start = get_key_start_from_aescrypt_object(aescrypt_object);
5883 if (k_start == NULL) return false;
5885 // similarly, get the start address of the r vector
5886 Node* objRvec = load_field_from_object(cipherBlockChaining_object, "r", "[B", /*is_exact*/ false);
5887 if (objRvec == NULL) return false;
5888 Node* r_start = array_element_address(objRvec, intcon(0), T_BYTE);
5890 // Call the stub, passing src_start, dest_start, k_start, r_start and src_len
5891 make_runtime_call(RC_LEAF|RC_NO_FP,
5892 OptoRuntime::cipherBlockChaining_aescrypt_Type(),
5893 stubAddr, stubName, TypePtr::BOTTOM,
5894 src_start, dest_start, k_start, r_start, len);
5896 // return is void so no result needs to be pushed
5898 return true;
5899 }
5901 //------------------------------get_key_start_from_aescrypt_object-----------------------
5902 Node * LibraryCallKit::get_key_start_from_aescrypt_object(Node *aescrypt_object) {
5903 Node* objAESCryptKey = load_field_from_object(aescrypt_object, "K", "[I", /*is_exact*/ false);
5904 assert (objAESCryptKey != NULL, "wrong version of com.sun.crypto.provider.AESCrypt");
5905 if (objAESCryptKey == NULL) return (Node *) NULL;
5907 // now have the array, need to get the start address of the K array
5908 Node* k_start = array_element_address(objAESCryptKey, intcon(0), T_INT);
5909 return k_start;
5910 }
5912 //----------------------------inline_cipherBlockChaining_AESCrypt_predicate----------------------------
5913 // Return node representing slow path of predicate check.
5914 // the pseudo code we want to emulate with this predicate is:
5915 // for encryption:
5916 // if (embeddedCipherObj instanceof AESCrypt) do_intrinsic, else do_javapath
5917 // for decryption:
5918 // if ((embeddedCipherObj instanceof AESCrypt) && (cipher!=plain)) do_intrinsic, else do_javapath
5919 // note cipher==plain is more conservative than the original java code but that's OK
5920 //
5921 Node* LibraryCallKit::inline_cipherBlockChaining_AESCrypt_predicate(bool decrypting) {
5922 // First, check receiver for NULL since it is virtual method.
5923 int nargs = arg_size();
5924 Node* objCBC = argument(0);
5925 _sp += nargs;
5926 objCBC = do_null_check(objCBC, T_OBJECT);
5927 _sp -= nargs;
5929 if (stopped()) return NULL; // Always NULL
5931 // Load embeddedCipher field of CipherBlockChaining object.
5932 Node* embeddedCipherObj = load_field_from_object(objCBC, "embeddedCipher", "Lcom/sun/crypto/provider/SymmetricCipher;", /*is_exact*/ false);
5934 // get AESCrypt klass for instanceOf check
5935 // AESCrypt might not be loaded yet if some other SymmetricCipher got us to this compile point
5936 // will have same classloader as CipherBlockChaining object
5937 const TypeInstPtr* tinst = _gvn.type(objCBC)->isa_instptr();
5938 assert(tinst != NULL, "CBCobj is null");
5939 assert(tinst->klass()->is_loaded(), "CBCobj is not loaded");
5941 // we want to do an instanceof comparison against the AESCrypt class
5942 ciKlass* klass_AESCrypt = tinst->klass()->as_instance_klass()->find_klass(ciSymbol::make("com/sun/crypto/provider/AESCrypt"));
5943 if (!klass_AESCrypt->is_loaded()) {
5944 // if AESCrypt is not even loaded, we never take the intrinsic fast path
5945 Node* ctrl = control();
5946 set_control(top()); // no regular fast path
5947 return ctrl;
5948 }
5949 ciInstanceKlass* instklass_AESCrypt = klass_AESCrypt->as_instance_klass();
5951 _sp += nargs; // gen_instanceof might do an uncommon trap
5952 Node* instof = gen_instanceof(embeddedCipherObj, makecon(TypeKlassPtr::make(instklass_AESCrypt)));
5953 _sp -= nargs;
5954 Node* cmp_instof = _gvn.transform(new (C) CmpINode(instof, intcon(1)));
5955 Node* bool_instof = _gvn.transform(new (C) BoolNode(cmp_instof, BoolTest::ne));
5957 Node* instof_false = generate_guard(bool_instof, NULL, PROB_MIN);
5959 // for encryption, we are done
5960 if (!decrypting)
5961 return instof_false; // even if it is NULL
5963 // for decryption, we need to add a further check to avoid
5964 // taking the intrinsic path when cipher and plain are the same
5965 // see the original java code for why.
5966 RegionNode* region = new(C) RegionNode(3);
5967 region->init_req(1, instof_false);
5968 Node* src = argument(1);
5969 Node *dest = argument(4);
5970 Node* cmp_src_dest = _gvn.transform(new (C) CmpPNode(src, dest));
5971 Node* bool_src_dest = _gvn.transform(new (C) BoolNode(cmp_src_dest, BoolTest::eq));
5972 Node* src_dest_conjoint = generate_guard(bool_src_dest, NULL, PROB_MIN);
5973 region->init_req(2, src_dest_conjoint);
5975 record_for_igvn(region);
5976 return _gvn.transform(region);
5978 }