Fri, 25 May 2012 07:53:11 -0700
7170463: C2 should recognize "obj.getClass() == A.class" code pattern
Summary: optimize this code pattern obj.getClass() == A.class.
Reviewed-by: jrose, kvn
Contributed-by: Krystal Mok <sajia@taobao.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 vmIntrinsics::ID _intrinsic_id;
49 public:
50 LibraryIntrinsic(ciMethod* m, bool is_virtual, vmIntrinsics::ID id)
51 : InlineCallGenerator(m),
52 _is_virtual(is_virtual),
53 _intrinsic_id(id)
54 {
55 }
56 virtual bool is_intrinsic() const { return true; }
57 virtual bool is_virtual() const { return _is_virtual; }
58 virtual JVMState* generate(JVMState* jvms);
59 vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
60 };
63 // Local helper class for LibraryIntrinsic:
64 class LibraryCallKit : public GraphKit {
65 private:
66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
68 public:
69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
70 : GraphKit(caller),
71 _intrinsic(intrinsic)
72 {
73 }
75 ciMethod* caller() const { return jvms()->method(); }
76 int bci() const { return jvms()->bci(); }
77 LibraryIntrinsic* intrinsic() const { return _intrinsic; }
78 vmIntrinsics::ID intrinsic_id() const { return _intrinsic->intrinsic_id(); }
79 ciMethod* callee() const { return _intrinsic->method(); }
80 ciSignature* signature() const { return callee()->signature(); }
81 int arg_size() const { return callee()->arg_size(); }
83 bool try_to_inline();
85 // Helper functions to inline natives
86 void push_result(RegionNode* region, PhiNode* value);
87 Node* generate_guard(Node* test, RegionNode* region, float true_prob);
88 Node* generate_slow_guard(Node* test, RegionNode* region);
89 Node* generate_fair_guard(Node* test, RegionNode* region);
90 Node* generate_negative_guard(Node* index, RegionNode* region,
91 // resulting CastII of index:
92 Node* *pos_index = NULL);
93 Node* generate_nonpositive_guard(Node* index, bool never_negative,
94 // resulting CastII of index:
95 Node* *pos_index = NULL);
96 Node* generate_limit_guard(Node* offset, Node* subseq_length,
97 Node* array_length,
98 RegionNode* region);
99 Node* generate_current_thread(Node* &tls_output);
100 address basictype2arraycopy(BasicType t, Node *src_offset, Node *dest_offset,
101 bool disjoint_bases, const char* &name, bool dest_uninitialized);
102 Node* load_mirror_from_klass(Node* klass);
103 Node* load_klass_from_mirror_common(Node* mirror, bool never_see_null,
104 int nargs,
105 RegionNode* region, int null_path,
106 int offset);
107 Node* load_klass_from_mirror(Node* mirror, bool never_see_null, int nargs,
108 RegionNode* region, int null_path) {
109 int offset = java_lang_Class::klass_offset_in_bytes();
110 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
111 region, null_path,
112 offset);
113 }
114 Node* load_array_klass_from_mirror(Node* mirror, bool never_see_null,
115 int nargs,
116 RegionNode* region, int null_path) {
117 int offset = java_lang_Class::array_klass_offset_in_bytes();
118 return load_klass_from_mirror_common(mirror, never_see_null, nargs,
119 region, null_path,
120 offset);
121 }
122 Node* generate_access_flags_guard(Node* kls,
123 int modifier_mask, int modifier_bits,
124 RegionNode* region);
125 Node* generate_interface_guard(Node* kls, RegionNode* region);
126 Node* generate_array_guard(Node* kls, RegionNode* region) {
127 return generate_array_guard_common(kls, region, false, false);
128 }
129 Node* generate_non_array_guard(Node* kls, RegionNode* region) {
130 return generate_array_guard_common(kls, region, false, true);
131 }
132 Node* generate_objArray_guard(Node* kls, RegionNode* region) {
133 return generate_array_guard_common(kls, region, true, false);
134 }
135 Node* generate_non_objArray_guard(Node* kls, RegionNode* region) {
136 return generate_array_guard_common(kls, region, true, true);
137 }
138 Node* generate_array_guard_common(Node* kls, RegionNode* region,
139 bool obj_array, bool not_array);
140 Node* generate_virtual_guard(Node* obj_klass, RegionNode* slow_region);
141 CallJavaNode* generate_method_call(vmIntrinsics::ID method_id,
142 bool is_virtual = false, bool is_static = false);
143 CallJavaNode* generate_method_call_static(vmIntrinsics::ID method_id) {
144 return generate_method_call(method_id, false, true);
145 }
146 CallJavaNode* generate_method_call_virtual(vmIntrinsics::ID method_id) {
147 return generate_method_call(method_id, true, false);
148 }
150 Node* make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2);
151 Node* make_string_method_node(int opcode, Node* str1, Node* str2);
152 bool inline_string_compareTo();
153 bool inline_string_indexOf();
154 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
155 bool inline_string_equals();
156 Node* pop_math_arg();
157 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
158 bool inline_math_native(vmIntrinsics::ID id);
159 bool inline_trig(vmIntrinsics::ID id);
160 bool inline_trans(vmIntrinsics::ID id);
161 bool inline_abs(vmIntrinsics::ID id);
162 bool inline_sqrt(vmIntrinsics::ID id);
163 bool inline_pow(vmIntrinsics::ID id);
164 bool inline_exp(vmIntrinsics::ID id);
165 bool inline_min_max(vmIntrinsics::ID id);
166 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
167 // This returns Type::AnyPtr, RawPtr, or OopPtr.
168 int classify_unsafe_addr(Node* &base, Node* &offset);
169 Node* make_unsafe_address(Node* base, Node* offset);
170 // Helper for inline_unsafe_access.
171 // Generates the guards that check whether the result of
172 // Unsafe.getObject should be recorded in an SATB log buffer.
173 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val);
174 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
175 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
176 bool inline_unsafe_allocate();
177 bool inline_unsafe_copyMemory();
178 bool inline_native_currentThread();
179 #ifdef TRACE_HAVE_INTRINSICS
180 bool inline_native_classID();
181 bool inline_native_threadID();
182 #endif
183 bool inline_native_time_funcs(address method, const char* funcName);
184 bool inline_native_isInterrupted();
185 bool inline_native_Class_query(vmIntrinsics::ID id);
186 bool inline_native_subtype_check();
188 bool inline_native_newArray();
189 bool inline_native_getLength();
190 bool inline_array_copyOf(bool is_copyOfRange);
191 bool inline_array_equals();
192 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
193 bool inline_native_clone(bool is_virtual);
194 bool inline_native_Reflection_getCallerClass();
195 bool inline_native_AtomicLong_get();
196 bool inline_native_AtomicLong_attemptUpdate();
197 bool is_method_invoke_or_aux_frame(JVMState* jvms);
198 // Helper function for inlining native object hash method
199 bool inline_native_hashcode(bool is_virtual, bool is_static);
200 bool inline_native_getClass();
202 // Helper functions for inlining arraycopy
203 bool inline_arraycopy();
204 void generate_arraycopy(const TypePtr* adr_type,
205 BasicType basic_elem_type,
206 Node* src, Node* src_offset,
207 Node* dest, Node* dest_offset,
208 Node* copy_length,
209 bool disjoint_bases = false,
210 bool length_never_negative = false,
211 RegionNode* slow_region = NULL);
212 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
213 RegionNode* slow_region);
214 void generate_clear_array(const TypePtr* adr_type,
215 Node* dest,
216 BasicType basic_elem_type,
217 Node* slice_off,
218 Node* slice_len,
219 Node* slice_end);
220 bool generate_block_arraycopy(const TypePtr* adr_type,
221 BasicType basic_elem_type,
222 AllocateNode* alloc,
223 Node* src, Node* src_offset,
224 Node* dest, Node* dest_offset,
225 Node* dest_size, bool dest_uninitialized);
226 void generate_slow_arraycopy(const TypePtr* adr_type,
227 Node* src, Node* src_offset,
228 Node* dest, Node* dest_offset,
229 Node* copy_length, bool dest_uninitialized);
230 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
231 Node* dest_elem_klass,
232 Node* src, Node* src_offset,
233 Node* dest, Node* dest_offset,
234 Node* copy_length, bool dest_uninitialized);
235 Node* generate_generic_arraycopy(const TypePtr* adr_type,
236 Node* src, Node* src_offset,
237 Node* dest, Node* dest_offset,
238 Node* copy_length, bool dest_uninitialized);
239 void generate_unchecked_arraycopy(const TypePtr* adr_type,
240 BasicType basic_elem_type,
241 bool disjoint_bases,
242 Node* src, Node* src_offset,
243 Node* dest, Node* dest_offset,
244 Node* copy_length, bool dest_uninitialized);
245 bool inline_unsafe_CAS(BasicType type);
246 bool inline_unsafe_ordered_store(BasicType type);
247 bool inline_fp_conversions(vmIntrinsics::ID id);
248 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
249 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
250 bool inline_bitCount(vmIntrinsics::ID id);
251 bool inline_reverseBytes(vmIntrinsics::ID id);
253 bool inline_reference_get();
254 };
257 //---------------------------make_vm_intrinsic----------------------------
258 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
259 vmIntrinsics::ID id = m->intrinsic_id();
260 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
262 if (DisableIntrinsic[0] != '\0'
263 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
264 // disabled by a user request on the command line:
265 // example: -XX:DisableIntrinsic=_hashCode,_getClass
266 return NULL;
267 }
269 if (!m->is_loaded()) {
270 // do not attempt to inline unloaded methods
271 return NULL;
272 }
274 // Only a few intrinsics implement a virtual dispatch.
275 // They are expensive calls which are also frequently overridden.
276 if (is_virtual) {
277 switch (id) {
278 case vmIntrinsics::_hashCode:
279 case vmIntrinsics::_clone:
280 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
281 break;
282 default:
283 return NULL;
284 }
285 }
287 // -XX:-InlineNatives disables nearly all intrinsics:
288 if (!InlineNatives) {
289 switch (id) {
290 case vmIntrinsics::_indexOf:
291 case vmIntrinsics::_compareTo:
292 case vmIntrinsics::_equals:
293 case vmIntrinsics::_equalsC:
294 break; // InlineNatives does not control String.compareTo
295 default:
296 return NULL;
297 }
298 }
300 switch (id) {
301 case vmIntrinsics::_compareTo:
302 if (!SpecialStringCompareTo) return NULL;
303 break;
304 case vmIntrinsics::_indexOf:
305 if (!SpecialStringIndexOf) return NULL;
306 break;
307 case vmIntrinsics::_equals:
308 if (!SpecialStringEquals) return NULL;
309 break;
310 case vmIntrinsics::_equalsC:
311 if (!SpecialArraysEquals) return NULL;
312 break;
313 case vmIntrinsics::_arraycopy:
314 if (!InlineArrayCopy) return NULL;
315 break;
316 case vmIntrinsics::_copyMemory:
317 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
318 if (!InlineArrayCopy) return NULL;
319 break;
320 case vmIntrinsics::_hashCode:
321 if (!InlineObjectHash) return NULL;
322 break;
323 case vmIntrinsics::_clone:
324 case vmIntrinsics::_copyOf:
325 case vmIntrinsics::_copyOfRange:
326 if (!InlineObjectCopy) return NULL;
327 // These also use the arraycopy intrinsic mechanism:
328 if (!InlineArrayCopy) return NULL;
329 break;
330 case vmIntrinsics::_checkIndex:
331 // We do not intrinsify this. The optimizer does fine with it.
332 return NULL;
334 case vmIntrinsics::_get_AtomicLong:
335 case vmIntrinsics::_attemptUpdate:
336 if (!InlineAtomicLong) return NULL;
337 break;
339 case vmIntrinsics::_getCallerClass:
340 if (!UseNewReflection) return NULL;
341 if (!InlineReflectionGetCallerClass) return NULL;
342 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
343 break;
345 case vmIntrinsics::_bitCount_i:
346 if (!Matcher::match_rule_supported(Op_PopCountI)) return NULL;
347 break;
349 case vmIntrinsics::_bitCount_l:
350 if (!Matcher::match_rule_supported(Op_PopCountL)) return NULL;
351 break;
353 case vmIntrinsics::_numberOfLeadingZeros_i:
354 if (!Matcher::match_rule_supported(Op_CountLeadingZerosI)) return NULL;
355 break;
357 case vmIntrinsics::_numberOfLeadingZeros_l:
358 if (!Matcher::match_rule_supported(Op_CountLeadingZerosL)) return NULL;
359 break;
361 case vmIntrinsics::_numberOfTrailingZeros_i:
362 if (!Matcher::match_rule_supported(Op_CountTrailingZerosI)) return NULL;
363 break;
365 case vmIntrinsics::_numberOfTrailingZeros_l:
366 if (!Matcher::match_rule_supported(Op_CountTrailingZerosL)) return NULL;
367 break;
369 case vmIntrinsics::_Reference_get:
370 // It is only when G1 is enabled that we absolutely
371 // need to use the intrinsic version of Reference.get()
372 // so that the value in the referent field, if necessary,
373 // can be registered by the pre-barrier code.
374 if (!UseG1GC) return NULL;
375 break;
377 default:
378 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
379 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
380 break;
381 }
383 // -XX:-InlineClassNatives disables natives from the Class class.
384 // The flag applies to all reflective calls, notably Array.newArray
385 // (visible to Java programmers as Array.newInstance).
386 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
387 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
388 if (!InlineClassNatives) return NULL;
389 }
391 // -XX:-InlineThreadNatives disables natives from the Thread class.
392 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
393 if (!InlineThreadNatives) return NULL;
394 }
396 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
397 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
398 m->holder()->name() == ciSymbol::java_lang_Float() ||
399 m->holder()->name() == ciSymbol::java_lang_Double()) {
400 if (!InlineMathNatives) return NULL;
401 }
403 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
404 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
405 if (!InlineUnsafeOps) return NULL;
406 }
408 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
409 }
411 //----------------------register_library_intrinsics-----------------------
412 // Initialize this file's data structures, for each Compile instance.
413 void Compile::register_library_intrinsics() {
414 // Nothing to do here.
415 }
417 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
418 LibraryCallKit kit(jvms, this);
419 Compile* C = kit.C;
420 int nodes = C->unique();
421 #ifndef PRODUCT
422 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
423 char buf[1000];
424 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
425 tty->print_cr("Intrinsic %s", str);
426 }
427 #endif
429 if (kit.try_to_inline()) {
430 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
431 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
432 }
433 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
434 if (C->log()) {
435 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
436 vmIntrinsics::name_at(intrinsic_id()),
437 (is_virtual() ? " virtual='1'" : ""),
438 C->unique() - nodes);
439 }
440 return kit.transfer_exceptions_into_jvms();
441 }
443 // The intrinsic bailed out
444 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
445 if (jvms->has_method()) {
446 // Not a root compile.
447 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
448 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
449 } else {
450 // Root compile
451 tty->print("Did not generate intrinsic %s%s at bci:%d in",
452 vmIntrinsics::name_at(intrinsic_id()),
453 (is_virtual() ? " (virtual)" : ""), kit.bci());
454 }
455 }
456 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
457 return NULL;
458 }
460 bool LibraryCallKit::try_to_inline() {
461 // Handle symbolic names for otherwise undistinguished boolean switches:
462 const bool is_store = true;
463 const bool is_native_ptr = true;
464 const bool is_static = true;
466 if (!jvms()->has_method()) {
467 // Root JVMState has a null method.
468 assert(map()->memory()->Opcode() == Op_Parm, "");
469 // Insert the memory aliasing node
470 set_all_memory(reset_memory());
471 }
472 assert(merged_memory(), "");
474 switch (intrinsic_id()) {
475 case vmIntrinsics::_hashCode:
476 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
477 case vmIntrinsics::_identityHashCode:
478 return inline_native_hashcode(/*!virtual*/ false, is_static);
479 case vmIntrinsics::_getClass:
480 return inline_native_getClass();
482 case vmIntrinsics::_dsin:
483 case vmIntrinsics::_dcos:
484 case vmIntrinsics::_dtan:
485 case vmIntrinsics::_dabs:
486 case vmIntrinsics::_datan2:
487 case vmIntrinsics::_dsqrt:
488 case vmIntrinsics::_dexp:
489 case vmIntrinsics::_dlog:
490 case vmIntrinsics::_dlog10:
491 case vmIntrinsics::_dpow:
492 return inline_math_native(intrinsic_id());
494 case vmIntrinsics::_min:
495 case vmIntrinsics::_max:
496 return inline_min_max(intrinsic_id());
498 case vmIntrinsics::_arraycopy:
499 return inline_arraycopy();
501 case vmIntrinsics::_compareTo:
502 return inline_string_compareTo();
503 case vmIntrinsics::_indexOf:
504 return inline_string_indexOf();
505 case vmIntrinsics::_equals:
506 return inline_string_equals();
508 case vmIntrinsics::_getObject:
509 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
510 case vmIntrinsics::_getBoolean:
511 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
512 case vmIntrinsics::_getByte:
513 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
514 case vmIntrinsics::_getShort:
515 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
516 case vmIntrinsics::_getChar:
517 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
518 case vmIntrinsics::_getInt:
519 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
520 case vmIntrinsics::_getLong:
521 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
522 case vmIntrinsics::_getFloat:
523 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
524 case vmIntrinsics::_getDouble:
525 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
527 case vmIntrinsics::_putObject:
528 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
529 case vmIntrinsics::_putBoolean:
530 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
531 case vmIntrinsics::_putByte:
532 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
533 case vmIntrinsics::_putShort:
534 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
535 case vmIntrinsics::_putChar:
536 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
537 case vmIntrinsics::_putInt:
538 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
539 case vmIntrinsics::_putLong:
540 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
541 case vmIntrinsics::_putFloat:
542 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
543 case vmIntrinsics::_putDouble:
544 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
546 case vmIntrinsics::_getByte_raw:
547 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
548 case vmIntrinsics::_getShort_raw:
549 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
550 case vmIntrinsics::_getChar_raw:
551 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
552 case vmIntrinsics::_getInt_raw:
553 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
554 case vmIntrinsics::_getLong_raw:
555 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
556 case vmIntrinsics::_getFloat_raw:
557 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
558 case vmIntrinsics::_getDouble_raw:
559 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
560 case vmIntrinsics::_getAddress_raw:
561 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
563 case vmIntrinsics::_putByte_raw:
564 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
565 case vmIntrinsics::_putShort_raw:
566 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
567 case vmIntrinsics::_putChar_raw:
568 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
569 case vmIntrinsics::_putInt_raw:
570 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
571 case vmIntrinsics::_putLong_raw:
572 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
573 case vmIntrinsics::_putFloat_raw:
574 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
575 case vmIntrinsics::_putDouble_raw:
576 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
577 case vmIntrinsics::_putAddress_raw:
578 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
580 case vmIntrinsics::_getObjectVolatile:
581 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
582 case vmIntrinsics::_getBooleanVolatile:
583 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
584 case vmIntrinsics::_getByteVolatile:
585 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
586 case vmIntrinsics::_getShortVolatile:
587 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
588 case vmIntrinsics::_getCharVolatile:
589 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
590 case vmIntrinsics::_getIntVolatile:
591 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
592 case vmIntrinsics::_getLongVolatile:
593 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
594 case vmIntrinsics::_getFloatVolatile:
595 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
596 case vmIntrinsics::_getDoubleVolatile:
597 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
599 case vmIntrinsics::_putObjectVolatile:
600 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
601 case vmIntrinsics::_putBooleanVolatile:
602 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
603 case vmIntrinsics::_putByteVolatile:
604 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
605 case vmIntrinsics::_putShortVolatile:
606 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
607 case vmIntrinsics::_putCharVolatile:
608 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
609 case vmIntrinsics::_putIntVolatile:
610 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
611 case vmIntrinsics::_putLongVolatile:
612 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
613 case vmIntrinsics::_putFloatVolatile:
614 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
615 case vmIntrinsics::_putDoubleVolatile:
616 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
618 case vmIntrinsics::_prefetchRead:
619 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
620 case vmIntrinsics::_prefetchWrite:
621 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
622 case vmIntrinsics::_prefetchReadStatic:
623 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
624 case vmIntrinsics::_prefetchWriteStatic:
625 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
627 case vmIntrinsics::_compareAndSwapObject:
628 return inline_unsafe_CAS(T_OBJECT);
629 case vmIntrinsics::_compareAndSwapInt:
630 return inline_unsafe_CAS(T_INT);
631 case vmIntrinsics::_compareAndSwapLong:
632 return inline_unsafe_CAS(T_LONG);
634 case vmIntrinsics::_putOrderedObject:
635 return inline_unsafe_ordered_store(T_OBJECT);
636 case vmIntrinsics::_putOrderedInt:
637 return inline_unsafe_ordered_store(T_INT);
638 case vmIntrinsics::_putOrderedLong:
639 return inline_unsafe_ordered_store(T_LONG);
641 case vmIntrinsics::_currentThread:
642 return inline_native_currentThread();
643 case vmIntrinsics::_isInterrupted:
644 return inline_native_isInterrupted();
646 #ifdef TRACE_HAVE_INTRINSICS
647 case vmIntrinsics::_classID:
648 return inline_native_classID();
649 case vmIntrinsics::_threadID:
650 return inline_native_threadID();
651 case vmIntrinsics::_counterTime:
652 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, TRACE_TIME_METHOD), "counterTime");
653 #endif
654 case vmIntrinsics::_currentTimeMillis:
655 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeMillis), "currentTimeMillis");
656 case vmIntrinsics::_nanoTime:
657 return inline_native_time_funcs(CAST_FROM_FN_PTR(address, os::javaTimeNanos), "nanoTime");
658 case vmIntrinsics::_allocateInstance:
659 return inline_unsafe_allocate();
660 case vmIntrinsics::_copyMemory:
661 return inline_unsafe_copyMemory();
662 case vmIntrinsics::_newArray:
663 return inline_native_newArray();
664 case vmIntrinsics::_getLength:
665 return inline_native_getLength();
666 case vmIntrinsics::_copyOf:
667 return inline_array_copyOf(false);
668 case vmIntrinsics::_copyOfRange:
669 return inline_array_copyOf(true);
670 case vmIntrinsics::_equalsC:
671 return inline_array_equals();
672 case vmIntrinsics::_clone:
673 return inline_native_clone(intrinsic()->is_virtual());
675 case vmIntrinsics::_isAssignableFrom:
676 return inline_native_subtype_check();
678 case vmIntrinsics::_isInstance:
679 case vmIntrinsics::_getModifiers:
680 case vmIntrinsics::_isInterface:
681 case vmIntrinsics::_isArray:
682 case vmIntrinsics::_isPrimitive:
683 case vmIntrinsics::_getSuperclass:
684 case vmIntrinsics::_getComponentType:
685 case vmIntrinsics::_getClassAccessFlags:
686 return inline_native_Class_query(intrinsic_id());
688 case vmIntrinsics::_floatToRawIntBits:
689 case vmIntrinsics::_floatToIntBits:
690 case vmIntrinsics::_intBitsToFloat:
691 case vmIntrinsics::_doubleToRawLongBits:
692 case vmIntrinsics::_doubleToLongBits:
693 case vmIntrinsics::_longBitsToDouble:
694 return inline_fp_conversions(intrinsic_id());
696 case vmIntrinsics::_numberOfLeadingZeros_i:
697 case vmIntrinsics::_numberOfLeadingZeros_l:
698 return inline_numberOfLeadingZeros(intrinsic_id());
700 case vmIntrinsics::_numberOfTrailingZeros_i:
701 case vmIntrinsics::_numberOfTrailingZeros_l:
702 return inline_numberOfTrailingZeros(intrinsic_id());
704 case vmIntrinsics::_bitCount_i:
705 case vmIntrinsics::_bitCount_l:
706 return inline_bitCount(intrinsic_id());
708 case vmIntrinsics::_reverseBytes_i:
709 case vmIntrinsics::_reverseBytes_l:
710 case vmIntrinsics::_reverseBytes_s:
711 case vmIntrinsics::_reverseBytes_c:
712 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
714 case vmIntrinsics::_get_AtomicLong:
715 return inline_native_AtomicLong_get();
716 case vmIntrinsics::_attemptUpdate:
717 return inline_native_AtomicLong_attemptUpdate();
719 case vmIntrinsics::_getCallerClass:
720 return inline_native_Reflection_getCallerClass();
722 case vmIntrinsics::_Reference_get:
723 return inline_reference_get();
725 default:
726 // If you get here, it may be that someone has added a new intrinsic
727 // to the list in vmSymbols.hpp without implementing it here.
728 #ifndef PRODUCT
729 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
730 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
731 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
732 }
733 #endif
734 return false;
735 }
736 }
738 //------------------------------push_result------------------------------
739 // Helper function for finishing intrinsics.
740 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
741 record_for_igvn(region);
742 set_control(_gvn.transform(region));
743 BasicType value_type = value->type()->basic_type();
744 push_node(value_type, _gvn.transform(value));
745 }
747 //------------------------------generate_guard---------------------------
748 // Helper function for generating guarded fast-slow graph structures.
749 // The given 'test', if true, guards a slow path. If the test fails
750 // then a fast path can be taken. (We generally hope it fails.)
751 // In all cases, GraphKit::control() is updated to the fast path.
752 // The returned value represents the control for the slow path.
753 // The return value is never 'top'; it is either a valid control
754 // or NULL if it is obvious that the slow path can never be taken.
755 // Also, if region and the slow control are not NULL, the slow edge
756 // is appended to the region.
757 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
758 if (stopped()) {
759 // Already short circuited.
760 return NULL;
761 }
763 // Build an if node and its projections.
764 // If test is true we take the slow path, which we assume is uncommon.
765 if (_gvn.type(test) == TypeInt::ZERO) {
766 // The slow branch is never taken. No need to build this guard.
767 return NULL;
768 }
770 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
772 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
773 if (if_slow == top()) {
774 // The slow branch is never taken. No need to build this guard.
775 return NULL;
776 }
778 if (region != NULL)
779 region->add_req(if_slow);
781 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
782 set_control(if_fast);
784 return if_slow;
785 }
787 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
788 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
789 }
790 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
791 return generate_guard(test, region, PROB_FAIR);
792 }
794 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
795 Node* *pos_index) {
796 if (stopped())
797 return NULL; // already stopped
798 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
799 return NULL; // index is already adequately typed
800 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
801 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
802 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
803 if (is_neg != NULL && pos_index != NULL) {
804 // Emulate effect of Parse::adjust_map_after_if.
805 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
806 ccast->set_req(0, control());
807 (*pos_index) = _gvn.transform(ccast);
808 }
809 return is_neg;
810 }
812 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
813 Node* *pos_index) {
814 if (stopped())
815 return NULL; // already stopped
816 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
817 return NULL; // index is already adequately typed
818 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
819 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
820 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
821 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
822 if (is_notp != NULL && pos_index != NULL) {
823 // Emulate effect of Parse::adjust_map_after_if.
824 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
825 ccast->set_req(0, control());
826 (*pos_index) = _gvn.transform(ccast);
827 }
828 return is_notp;
829 }
831 // Make sure that 'position' is a valid limit index, in [0..length].
832 // There are two equivalent plans for checking this:
833 // A. (offset + copyLength) unsigned<= arrayLength
834 // B. offset <= (arrayLength - copyLength)
835 // We require that all of the values above, except for the sum and
836 // difference, are already known to be non-negative.
837 // Plan A is robust in the face of overflow, if offset and copyLength
838 // are both hugely positive.
839 //
840 // Plan B is less direct and intuitive, but it does not overflow at
841 // all, since the difference of two non-negatives is always
842 // representable. Whenever Java methods must perform the equivalent
843 // check they generally use Plan B instead of Plan A.
844 // For the moment we use Plan A.
845 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
846 Node* subseq_length,
847 Node* array_length,
848 RegionNode* region) {
849 if (stopped())
850 return NULL; // already stopped
851 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
852 if (zero_offset && subseq_length->eqv_uncast(array_length))
853 return NULL; // common case of whole-array copy
854 Node* last = subseq_length;
855 if (!zero_offset) // last += offset
856 last = _gvn.transform( new (C, 3) AddINode(last, offset));
857 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
858 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
859 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
860 return is_over;
861 }
864 //--------------------------generate_current_thread--------------------
865 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
866 ciKlass* thread_klass = env()->Thread_klass();
867 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
868 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
869 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
870 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
871 tls_output = thread;
872 return threadObj;
873 }
876 //------------------------------make_string_method_node------------------------
877 // Helper method for String intrinsic functions. This version is called
878 // with str1 and str2 pointing to String object nodes.
879 //
880 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* str2) {
881 Node* no_ctrl = NULL;
883 // Get start addr of string
884 Node* str1_value = load_String_value(no_ctrl, str1);
885 Node* str1_offset = load_String_offset(no_ctrl, str1);
886 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
888 // Get length of string 1
889 Node* str1_len = load_String_length(no_ctrl, str1);
891 Node* str2_value = load_String_value(no_ctrl, str2);
892 Node* str2_offset = load_String_offset(no_ctrl, str2);
893 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
895 Node* str2_len = NULL;
896 Node* result = NULL;
898 switch (opcode) {
899 case Op_StrIndexOf:
900 // Get length of string 2
901 str2_len = load_String_length(no_ctrl, str2);
903 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
904 str1_start, str1_len, str2_start, str2_len);
905 break;
906 case Op_StrComp:
907 // Get length of string 2
908 str2_len = load_String_length(no_ctrl, str2);
910 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
911 str1_start, str1_len, str2_start, str2_len);
912 break;
913 case Op_StrEquals:
914 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
915 str1_start, str2_start, str1_len);
916 break;
917 default:
918 ShouldNotReachHere();
919 return NULL;
920 }
922 // All these intrinsics have checks.
923 C->set_has_split_ifs(true); // Has chance for split-if optimization
925 return _gvn.transform(result);
926 }
928 // Helper method for String intrinsic functions. This version is called
929 // with str1 and str2 pointing to char[] nodes, with cnt1 and cnt2 pointing
930 // to Int nodes containing the lenghts of str1 and str2.
931 //
932 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1_start, Node* cnt1, Node* str2_start, Node* cnt2) {
934 Node* result = NULL;
935 switch (opcode) {
936 case Op_StrIndexOf:
937 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
938 str1_start, cnt1, str2_start, cnt2);
939 break;
940 case Op_StrComp:
941 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
942 str1_start, cnt1, str2_start, cnt2);
943 break;
944 case Op_StrEquals:
945 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
946 str1_start, str2_start, cnt1);
947 break;
948 default:
949 ShouldNotReachHere();
950 return NULL;
951 }
953 // All these intrinsics have checks.
954 C->set_has_split_ifs(true); // Has chance for split-if optimization
956 return _gvn.transform(result);
957 }
959 //------------------------------inline_string_compareTo------------------------
960 bool LibraryCallKit::inline_string_compareTo() {
962 if (!Matcher::has_match_rule(Op_StrComp)) return false;
964 _sp += 2;
965 Node *argument = pop(); // pop non-receiver first: it was pushed second
966 Node *receiver = pop();
968 // Null check on self without removing any arguments. The argument
969 // null check technically happens in the wrong place, which can lead to
970 // invalid stack traces when string compare is inlined into a method
971 // which handles NullPointerExceptions.
972 _sp += 2;
973 receiver = do_null_check(receiver, T_OBJECT);
974 argument = do_null_check(argument, T_OBJECT);
975 _sp -= 2;
976 if (stopped()) {
977 return true;
978 }
980 Node* compare = make_string_method_node(Op_StrComp, receiver, argument);
981 push(compare);
982 return true;
983 }
985 //------------------------------inline_string_equals------------------------
986 bool LibraryCallKit::inline_string_equals() {
988 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
990 int nargs = 2;
991 _sp += nargs;
992 Node* argument = pop(); // pop non-receiver first: it was pushed second
993 Node* receiver = pop();
995 // Null check on self without removing any arguments. The argument
996 // null check technically happens in the wrong place, which can lead to
997 // invalid stack traces when string compare is inlined into a method
998 // which handles NullPointerExceptions.
999 _sp += nargs;
1000 receiver = do_null_check(receiver, T_OBJECT);
1001 //should not do null check for argument for String.equals(), because spec
1002 //allows to specify NULL as argument.
1003 _sp -= nargs;
1005 if (stopped()) {
1006 return true;
1007 }
1009 // paths (plus control) merge
1010 RegionNode* region = new (C, 5) RegionNode(5);
1011 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
1013 // does source == target string?
1014 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
1015 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
1017 Node* if_eq = generate_slow_guard(bol, NULL);
1018 if (if_eq != NULL) {
1019 // receiver == argument
1020 phi->init_req(2, intcon(1));
1021 region->init_req(2, if_eq);
1022 }
1024 // get String klass for instanceOf
1025 ciInstanceKlass* klass = env()->String_klass();
1027 if (!stopped()) {
1028 _sp += nargs; // gen_instanceof might do an uncommon trap
1029 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
1030 _sp -= nargs;
1031 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
1032 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
1034 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1035 //instanceOf == true, fallthrough
1037 if (inst_false != NULL) {
1038 phi->init_req(3, intcon(0));
1039 region->init_req(3, inst_false);
1040 }
1041 }
1043 if (!stopped()) {
1044 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1046 // Properly cast the argument to String
1047 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1048 // This path is taken only when argument's type is String:NotNull.
1049 argument = cast_not_null(argument, false);
1051 Node* no_ctrl = NULL;
1053 // Get start addr of receiver
1054 Node* receiver_val = load_String_value(no_ctrl, receiver);
1055 Node* receiver_offset = load_String_offset(no_ctrl, receiver);
1056 Node* receiver_start = array_element_address(receiver_val, receiver_offset, T_CHAR);
1058 // Get length of receiver
1059 Node* receiver_cnt = load_String_length(no_ctrl, receiver);
1061 // Get start addr of argument
1062 Node* argument_val = load_String_value(no_ctrl, argument);
1063 Node* argument_offset = load_String_offset(no_ctrl, argument);
1064 Node* argument_start = array_element_address(argument_val, argument_offset, T_CHAR);
1066 // Get length of argument
1067 Node* argument_cnt = load_String_length(no_ctrl, argument);
1069 // Check for receiver count != argument count
1070 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1071 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1072 Node* if_ne = generate_slow_guard(bol, NULL);
1073 if (if_ne != NULL) {
1074 phi->init_req(4, intcon(0));
1075 region->init_req(4, if_ne);
1076 }
1078 // Check for count == 0 is done by assembler code for StrEquals.
1080 if (!stopped()) {
1081 Node* equals = make_string_method_node(Op_StrEquals, receiver_start, receiver_cnt, argument_start, argument_cnt);
1082 phi->init_req(1, equals);
1083 region->init_req(1, control());
1084 }
1085 }
1087 // post merge
1088 set_control(_gvn.transform(region));
1089 record_for_igvn(region);
1091 push(_gvn.transform(phi));
1093 return true;
1094 }
1096 //------------------------------inline_array_equals----------------------------
1097 bool LibraryCallKit::inline_array_equals() {
1099 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1101 _sp += 2;
1102 Node *argument2 = pop();
1103 Node *argument1 = pop();
1105 Node* equals =
1106 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1107 argument1, argument2) );
1108 push(equals);
1109 return true;
1110 }
1112 // Java version of String.indexOf(constant string)
1113 // class StringDecl {
1114 // StringDecl(char[] ca) {
1115 // offset = 0;
1116 // count = ca.length;
1117 // value = ca;
1118 // }
1119 // int offset;
1120 // int count;
1121 // char[] value;
1122 // }
1123 //
1124 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1125 // int targetOffset, int cache_i, int md2) {
1126 // int cache = cache_i;
1127 // int sourceOffset = string_object.offset;
1128 // int sourceCount = string_object.count;
1129 // int targetCount = target_object.length;
1130 //
1131 // int targetCountLess1 = targetCount - 1;
1132 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1133 //
1134 // char[] source = string_object.value;
1135 // char[] target = target_object;
1136 // int lastChar = target[targetCountLess1];
1137 //
1138 // outer_loop:
1139 // for (int i = sourceOffset; i < sourceEnd; ) {
1140 // int src = source[i + targetCountLess1];
1141 // if (src == lastChar) {
1142 // // With random strings and a 4-character alphabet,
1143 // // reverse matching at this point sets up 0.8% fewer
1144 // // frames, but (paradoxically) makes 0.3% more probes.
1145 // // Since those probes are nearer the lastChar probe,
1146 // // there is may be a net D$ win with reverse matching.
1147 // // But, reversing loop inhibits unroll of inner loop
1148 // // for unknown reason. So, does running outer loop from
1149 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1150 // for (int j = 0; j < targetCountLess1; j++) {
1151 // if (target[targetOffset + j] != source[i+j]) {
1152 // if ((cache & (1 << source[i+j])) == 0) {
1153 // if (md2 < j+1) {
1154 // i += j+1;
1155 // continue outer_loop;
1156 // }
1157 // }
1158 // i += md2;
1159 // continue outer_loop;
1160 // }
1161 // }
1162 // return i - sourceOffset;
1163 // }
1164 // if ((cache & (1 << src)) == 0) {
1165 // i += targetCountLess1;
1166 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1167 // i++;
1168 // }
1169 // return -1;
1170 // }
1172 //------------------------------string_indexOf------------------------
1173 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1174 jint cache_i, jint md2_i) {
1176 Node* no_ctrl = NULL;
1177 float likely = PROB_LIKELY(0.9);
1178 float unlikely = PROB_UNLIKELY(0.9);
1180 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1182 Node* source = load_String_value(no_ctrl, string_object);
1183 Node* sourceOffset = load_String_offset(no_ctrl, string_object);
1184 Node* sourceCount = load_String_length(no_ctrl, string_object);
1186 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1187 jint target_length = target_array->length();
1188 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1189 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1191 IdealKit kit(this, false, true);
1192 #define __ kit.
1193 Node* zero = __ ConI(0);
1194 Node* one = __ ConI(1);
1195 Node* cache = __ ConI(cache_i);
1196 Node* md2 = __ ConI(md2_i);
1197 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1198 Node* targetCount = __ ConI(target_length);
1199 Node* targetCountLess1 = __ ConI(target_length - 1);
1200 Node* targetOffset = __ ConI(targetOffset_i);
1201 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1203 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1204 Node* outer_loop = __ make_label(2 /* goto */);
1205 Node* return_ = __ make_label(1);
1207 __ set(rtn,__ ConI(-1));
1208 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1209 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1210 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1211 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1212 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1213 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1214 Node* tpj = __ AddI(targetOffset, __ value(j));
1215 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1216 Node* ipj = __ AddI(__ value(i), __ value(j));
1217 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1218 __ if_then(targ, BoolTest::ne, src2); {
1219 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1220 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1221 __ increment(i, __ AddI(__ value(j), one));
1222 __ goto_(outer_loop);
1223 } __ end_if(); __ dead(j);
1224 }__ end_if(); __ dead(j);
1225 __ increment(i, md2);
1226 __ goto_(outer_loop);
1227 }__ end_if();
1228 __ increment(j, one);
1229 }__ end_loop(); __ dead(j);
1230 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1231 __ goto_(return_);
1232 }__ end_if();
1233 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1234 __ increment(i, targetCountLess1);
1235 }__ end_if();
1236 __ increment(i, one);
1237 __ bind(outer_loop);
1238 }__ end_loop(); __ dead(i);
1239 __ bind(return_);
1241 // Final sync IdealKit and GraphKit.
1242 final_sync(kit);
1243 Node* result = __ value(rtn);
1244 #undef __
1245 C->set_has_loops(true);
1246 return result;
1247 }
1249 //------------------------------inline_string_indexOf------------------------
1250 bool LibraryCallKit::inline_string_indexOf() {
1252 _sp += 2;
1253 Node *argument = pop(); // pop non-receiver first: it was pushed second
1254 Node *receiver = pop();
1256 Node* result;
1257 // Disable the use of pcmpestri until it can be guaranteed that
1258 // the load doesn't cross into the uncommited space.
1259 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1260 UseSSE42Intrinsics) {
1261 // Generate SSE4.2 version of indexOf
1262 // We currently only have match rules that use SSE4.2
1264 // Null check on self without removing any arguments. The argument
1265 // null check technically happens in the wrong place, which can lead to
1266 // invalid stack traces when string compare is inlined into a method
1267 // which handles NullPointerExceptions.
1268 _sp += 2;
1269 receiver = do_null_check(receiver, T_OBJECT);
1270 argument = do_null_check(argument, T_OBJECT);
1271 _sp -= 2;
1273 if (stopped()) {
1274 return true;
1275 }
1277 ciInstanceKlass* str_klass = env()->String_klass();
1278 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1280 // Make the merge point
1281 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1282 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1283 Node* no_ctrl = NULL;
1285 // Get start addr of source string
1286 Node* source = load_String_value(no_ctrl, receiver);
1287 Node* source_offset = load_String_offset(no_ctrl, receiver);
1288 Node* source_start = array_element_address(source, source_offset, T_CHAR);
1290 // Get length of source string
1291 Node* source_cnt = load_String_length(no_ctrl, receiver);
1293 // Get start addr of substring
1294 Node* substr = load_String_value(no_ctrl, argument);
1295 Node* substr_offset = load_String_offset(no_ctrl, argument);
1296 Node* substr_start = array_element_address(substr, substr_offset, T_CHAR);
1298 // Get length of source string
1299 Node* substr_cnt = load_String_length(no_ctrl, argument);
1301 // Check for substr count > string count
1302 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1303 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1304 Node* if_gt = generate_slow_guard(bol, NULL);
1305 if (if_gt != NULL) {
1306 result_phi->init_req(2, intcon(-1));
1307 result_rgn->init_req(2, if_gt);
1308 }
1310 if (!stopped()) {
1311 // Check for substr count == 0
1312 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1313 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1314 Node* if_zero = generate_slow_guard(bol, NULL);
1315 if (if_zero != NULL) {
1316 result_phi->init_req(3, intcon(0));
1317 result_rgn->init_req(3, if_zero);
1318 }
1319 }
1321 if (!stopped()) {
1322 result = make_string_method_node(Op_StrIndexOf, source_start, source_cnt, substr_start, substr_cnt);
1323 result_phi->init_req(1, result);
1324 result_rgn->init_req(1, control());
1325 }
1326 set_control(_gvn.transform(result_rgn));
1327 record_for_igvn(result_rgn);
1328 result = _gvn.transform(result_phi);
1330 } else { // Use LibraryCallKit::string_indexOf
1331 // don't intrinsify if argument isn't a constant string.
1332 if (!argument->is_Con()) {
1333 return false;
1334 }
1335 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1336 if (str_type == NULL) {
1337 return false;
1338 }
1339 ciInstanceKlass* klass = env()->String_klass();
1340 ciObject* str_const = str_type->const_oop();
1341 if (str_const == NULL || str_const->klass() != klass) {
1342 return false;
1343 }
1344 ciInstance* str = str_const->as_instance();
1345 assert(str != NULL, "must be instance");
1347 ciObject* v = str->field_value_by_offset(java_lang_String::value_offset_in_bytes()).as_object();
1348 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1350 int o;
1351 int c;
1352 if (java_lang_String::has_offset_field()) {
1353 o = str->field_value_by_offset(java_lang_String::offset_offset_in_bytes()).as_int();
1354 c = str->field_value_by_offset(java_lang_String::count_offset_in_bytes()).as_int();
1355 } else {
1356 o = 0;
1357 c = pat->length();
1358 }
1360 // constant strings have no offset and count == length which
1361 // simplifies the resulting code somewhat so lets optimize for that.
1362 if (o != 0 || c != pat->length()) {
1363 return false;
1364 }
1366 // Null check on self without removing any arguments. The argument
1367 // null check technically happens in the wrong place, which can lead to
1368 // invalid stack traces when string compare is inlined into a method
1369 // which handles NullPointerExceptions.
1370 _sp += 2;
1371 receiver = do_null_check(receiver, T_OBJECT);
1372 // No null check on the argument is needed since it's a constant String oop.
1373 _sp -= 2;
1374 if (stopped()) {
1375 return true;
1376 }
1378 // The null string as a pattern always returns 0 (match at beginning of string)
1379 if (c == 0) {
1380 push(intcon(0));
1381 return true;
1382 }
1384 // Generate default indexOf
1385 jchar lastChar = pat->char_at(o + (c - 1));
1386 int cache = 0;
1387 int i;
1388 for (i = 0; i < c - 1; i++) {
1389 assert(i < pat->length(), "out of range");
1390 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1391 }
1393 int md2 = c;
1394 for (i = 0; i < c - 1; i++) {
1395 assert(i < pat->length(), "out of range");
1396 if (pat->char_at(o + i) == lastChar) {
1397 md2 = (c - 1) - i;
1398 }
1399 }
1401 result = string_indexOf(receiver, pat, o, cache, md2);
1402 }
1404 push(result);
1405 return true;
1406 }
1408 //--------------------------pop_math_arg--------------------------------
1409 // Pop a double argument to a math function from the stack
1410 // rounding it if necessary.
1411 Node * LibraryCallKit::pop_math_arg() {
1412 Node *arg = pop_pair();
1413 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1414 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1415 return arg;
1416 }
1418 //------------------------------inline_trig----------------------------------
1419 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1420 // argument reduction which will turn into a fast/slow diamond.
1421 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1422 _sp += arg_size(); // restore stack pointer
1423 Node* arg = pop_math_arg();
1424 Node* trig = NULL;
1426 switch (id) {
1427 case vmIntrinsics::_dsin:
1428 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1429 break;
1430 case vmIntrinsics::_dcos:
1431 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1432 break;
1433 case vmIntrinsics::_dtan:
1434 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1435 break;
1436 default:
1437 assert(false, "bad intrinsic was passed in");
1438 return false;
1439 }
1441 // Rounding required? Check for argument reduction!
1442 if( Matcher::strict_fp_requires_explicit_rounding ) {
1444 static const double pi_4 = 0.7853981633974483;
1445 static const double neg_pi_4 = -0.7853981633974483;
1446 // pi/2 in 80-bit extended precision
1447 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1448 // -pi/2 in 80-bit extended precision
1449 // 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};
1450 // Cutoff value for using this argument reduction technique
1451 //static const double pi_2_minus_epsilon = 1.564660403643354;
1452 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1454 // Pseudocode for sin:
1455 // if (x <= Math.PI / 4.0) {
1456 // if (x >= -Math.PI / 4.0) return fsin(x);
1457 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1458 // } else {
1459 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1460 // }
1461 // return StrictMath.sin(x);
1463 // Pseudocode for cos:
1464 // if (x <= Math.PI / 4.0) {
1465 // if (x >= -Math.PI / 4.0) return fcos(x);
1466 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1467 // } else {
1468 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1469 // }
1470 // return StrictMath.cos(x);
1472 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1473 // requires a special machine instruction to load it. Instead we'll try
1474 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1475 // probably do the math inside the SIN encoding.
1477 // Make the merge point
1478 RegionNode *r = new (C, 3) RegionNode(3);
1479 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1481 // Flatten arg so we need only 1 test
1482 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1483 // Node for PI/4 constant
1484 Node *pi4 = makecon(TypeD::make(pi_4));
1485 // Check PI/4 : abs(arg)
1486 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1487 // Check: If PI/4 < abs(arg) then go slow
1488 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1489 // Branch either way
1490 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1491 set_control(opt_iff(r,iff));
1493 // Set fast path result
1494 phi->init_req(2,trig);
1496 // Slow path - non-blocking leaf call
1497 Node* call = NULL;
1498 switch (id) {
1499 case vmIntrinsics::_dsin:
1500 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1501 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1502 "Sin", NULL, arg, top());
1503 break;
1504 case vmIntrinsics::_dcos:
1505 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1506 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1507 "Cos", NULL, arg, top());
1508 break;
1509 case vmIntrinsics::_dtan:
1510 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1511 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1512 "Tan", NULL, arg, top());
1513 break;
1514 }
1515 assert(control()->in(0) == call, "");
1516 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1517 r->init_req(1,control());
1518 phi->init_req(1,slow_result);
1520 // Post-merge
1521 set_control(_gvn.transform(r));
1522 record_for_igvn(r);
1523 trig = _gvn.transform(phi);
1525 C->set_has_split_ifs(true); // Has chance for split-if optimization
1526 }
1527 // Push result back on JVM stack
1528 push_pair(trig);
1529 return true;
1530 }
1532 //------------------------------inline_sqrt-------------------------------------
1533 // Inline square root instruction, if possible.
1534 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1535 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1536 _sp += arg_size(); // restore stack pointer
1537 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1538 return true;
1539 }
1541 //------------------------------inline_abs-------------------------------------
1542 // Inline absolute value instruction, if possible.
1543 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1544 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1545 _sp += arg_size(); // restore stack pointer
1546 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1547 return true;
1548 }
1550 //------------------------------inline_exp-------------------------------------
1551 // Inline exp instructions, if possible. The Intel hardware only misses
1552 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1553 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1554 assert(id == vmIntrinsics::_dexp, "Not exp");
1556 // If this inlining ever returned NaN in the past, we do not intrinsify it
1557 // every again. NaN results requires StrictMath.exp handling.
1558 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1560 _sp += arg_size(); // restore stack pointer
1561 Node *x = pop_math_arg();
1562 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1564 //-------------------
1565 //result=(result.isNaN())? StrictMath::exp():result;
1566 // Check: If isNaN() by checking result!=result? then go to Strict Math
1567 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1568 // Build the boolean node
1569 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1571 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1572 // End the current control-flow path
1573 push_pair(x);
1574 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1575 // to handle. Recompile without intrinsifying Math.exp
1576 uncommon_trap(Deoptimization::Reason_intrinsic,
1577 Deoptimization::Action_make_not_entrant);
1578 }
1580 C->set_has_split_ifs(true); // Has chance for split-if optimization
1582 push_pair(result);
1584 return true;
1585 }
1587 //------------------------------inline_pow-------------------------------------
1588 // Inline power instructions, if possible.
1589 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1590 assert(id == vmIntrinsics::_dpow, "Not pow");
1592 // If this inlining ever returned NaN in the past, we do not intrinsify it
1593 // every again. NaN results requires StrictMath.pow handling.
1594 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1596 // Do not intrinsify on older platforms which lack cmove.
1597 if (ConditionalMoveLimit == 0) return false;
1599 // Pseudocode for pow
1600 // if (x <= 0.0) {
1601 // if ((double)((int)y)==y) { // if y is int
1602 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1603 // } else {
1604 // result = NaN;
1605 // }
1606 // } else {
1607 // result = DPow(x,y);
1608 // }
1609 // if (result != result)? {
1610 // uncommon_trap();
1611 // }
1612 // return result;
1614 _sp += arg_size(); // restore stack pointer
1615 Node* y = pop_math_arg();
1616 Node* x = pop_math_arg();
1618 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1620 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1621 // inside of something) then skip the fancy tests and just check for
1622 // NaN result.
1623 Node *result = NULL;
1624 if( jvms()->depth() >= 1 ) {
1625 result = fast_result;
1626 } else {
1628 // Set the merge point for If node with condition of (x <= 0.0)
1629 // There are four possible paths to region node and phi node
1630 RegionNode *r = new (C, 4) RegionNode(4);
1631 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1633 // Build the first if node: if (x <= 0.0)
1634 // Node for 0 constant
1635 Node *zeronode = makecon(TypeD::ZERO);
1636 // Check x:0
1637 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1638 // Check: If (x<=0) then go complex path
1639 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1640 // Branch either way
1641 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1642 Node *opt_test = _gvn.transform(if1);
1643 //assert( opt_test->is_If(), "Expect an IfNode");
1644 IfNode *opt_if1 = (IfNode*)opt_test;
1645 // Fast path taken; set region slot 3
1646 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1647 r->init_req(3,fast_taken); // Capture fast-control
1649 // Fast path not-taken, i.e. slow path
1650 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1652 // Set fast path result
1653 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1654 phi->init_req(3, fast_result);
1656 // Complex path
1657 // Build the second if node (if y is int)
1658 // Node for (int)y
1659 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1660 // Node for (double)((int) y)
1661 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1662 // Check (double)((int) y) : y
1663 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1664 // Check if (y isn't int) then go to slow path
1666 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1667 // Branch either way
1668 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1669 Node *slow_path = opt_iff(r,if2); // Set region path 2
1671 // Calculate DPow(abs(x), y)*(1 & (int)y)
1672 // Node for constant 1
1673 Node *conone = intcon(1);
1674 // 1& (int)y
1675 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1676 // zero node
1677 Node *conzero = intcon(0);
1678 // Check (1&(int)y)==0?
1679 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1680 // Check if (1&(int)y)!=0?, if so the result is negative
1681 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1682 // abs(x)
1683 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1684 // abs(x)^y
1685 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1686 // -abs(x)^y
1687 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1688 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1689 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1690 // Set complex path fast result
1691 phi->init_req(2, signresult);
1693 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1694 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1695 r->init_req(1,slow_path);
1696 phi->init_req(1,slow_result);
1698 // Post merge
1699 set_control(_gvn.transform(r));
1700 record_for_igvn(r);
1701 result=_gvn.transform(phi);
1702 }
1704 //-------------------
1705 //result=(result.isNaN())? uncommon_trap():result;
1706 // Check: If isNaN() by checking result!=result? then go to Strict Math
1707 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1708 // Build the boolean node
1709 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1711 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1712 // End the current control-flow path
1713 push_pair(x);
1714 push_pair(y);
1715 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1716 // to handle. Recompile without intrinsifying Math.pow.
1717 uncommon_trap(Deoptimization::Reason_intrinsic,
1718 Deoptimization::Action_make_not_entrant);
1719 }
1721 C->set_has_split_ifs(true); // Has chance for split-if optimization
1723 push_pair(result);
1725 return true;
1726 }
1728 //------------------------------inline_trans-------------------------------------
1729 // Inline transcendental instructions, if possible. The Intel hardware gets
1730 // these right, no funny corner cases missed.
1731 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1732 _sp += arg_size(); // restore stack pointer
1733 Node* arg = pop_math_arg();
1734 Node* trans = NULL;
1736 switch (id) {
1737 case vmIntrinsics::_dlog:
1738 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1739 break;
1740 case vmIntrinsics::_dlog10:
1741 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1742 break;
1743 default:
1744 assert(false, "bad intrinsic was passed in");
1745 return false;
1746 }
1748 // Push result back on JVM stack
1749 push_pair(trans);
1750 return true;
1751 }
1753 //------------------------------runtime_math-----------------------------
1754 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1755 Node* a = NULL;
1756 Node* b = NULL;
1758 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1759 "must be (DD)D or (D)D type");
1761 // Inputs
1762 _sp += arg_size(); // restore stack pointer
1763 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1764 b = pop_math_arg();
1765 }
1766 a = pop_math_arg();
1768 const TypePtr* no_memory_effects = NULL;
1769 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1770 no_memory_effects,
1771 a, top(), b, b ? top() : NULL);
1772 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1773 #ifdef ASSERT
1774 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1775 assert(value_top == top(), "second value must be top");
1776 #endif
1778 push_pair(value);
1779 return true;
1780 }
1782 //------------------------------inline_math_native-----------------------------
1783 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1784 switch (id) {
1785 // These intrinsics are not properly supported on all hardware
1786 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1787 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1788 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1789 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1790 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1791 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1793 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1794 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1795 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1796 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1798 // These intrinsics are supported on all hardware
1799 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1800 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1802 case vmIntrinsics::_dexp: return
1803 Matcher::has_match_rule(Op_ExpD) ? inline_exp(id) :
1804 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1805 case vmIntrinsics::_dpow: return
1806 Matcher::has_match_rule(Op_PowD) ? inline_pow(id) :
1807 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1809 // These intrinsics are not yet correctly implemented
1810 case vmIntrinsics::_datan2:
1811 return false;
1813 default:
1814 ShouldNotReachHere();
1815 return false;
1816 }
1817 }
1819 static bool is_simple_name(Node* n) {
1820 return (n->req() == 1 // constant
1821 || (n->is_Type() && n->as_Type()->type()->singleton())
1822 || n->is_Proj() // parameter or return value
1823 || n->is_Phi() // local of some sort
1824 );
1825 }
1827 //----------------------------inline_min_max-----------------------------------
1828 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1829 push(generate_min_max(id, argument(0), argument(1)));
1831 return true;
1832 }
1834 Node*
1835 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1836 // These are the candidate return value:
1837 Node* xvalue = x0;
1838 Node* yvalue = y0;
1840 if (xvalue == yvalue) {
1841 return xvalue;
1842 }
1844 bool want_max = (id == vmIntrinsics::_max);
1846 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1847 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1848 if (txvalue == NULL || tyvalue == NULL) return top();
1849 // This is not really necessary, but it is consistent with a
1850 // hypothetical MaxINode::Value method:
1851 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1853 // %%% This folding logic should (ideally) be in a different place.
1854 // Some should be inside IfNode, and there to be a more reliable
1855 // transformation of ?: style patterns into cmoves. We also want
1856 // more powerful optimizations around cmove and min/max.
1858 // Try to find a dominating comparison of these guys.
1859 // It can simplify the index computation for Arrays.copyOf
1860 // and similar uses of System.arraycopy.
1861 // First, compute the normalized version of CmpI(x, y).
1862 int cmp_op = Op_CmpI;
1863 Node* xkey = xvalue;
1864 Node* ykey = yvalue;
1865 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1866 if (ideal_cmpxy->is_Cmp()) {
1867 // E.g., if we have CmpI(length - offset, count),
1868 // it might idealize to CmpI(length, count + offset)
1869 cmp_op = ideal_cmpxy->Opcode();
1870 xkey = ideal_cmpxy->in(1);
1871 ykey = ideal_cmpxy->in(2);
1872 }
1874 // Start by locating any relevant comparisons.
1875 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1876 Node* cmpxy = NULL;
1877 Node* cmpyx = NULL;
1878 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1879 Node* cmp = start_from->fast_out(k);
1880 if (cmp->outcnt() > 0 && // must have prior uses
1881 cmp->in(0) == NULL && // must be context-independent
1882 cmp->Opcode() == cmp_op) { // right kind of compare
1883 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1884 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1885 }
1886 }
1888 const int NCMPS = 2;
1889 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1890 int cmpn;
1891 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1892 if (cmps[cmpn] != NULL) break; // find a result
1893 }
1894 if (cmpn < NCMPS) {
1895 // Look for a dominating test that tells us the min and max.
1896 int depth = 0; // Limit search depth for speed
1897 Node* dom = control();
1898 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1899 if (++depth >= 100) break;
1900 Node* ifproj = dom;
1901 if (!ifproj->is_Proj()) continue;
1902 Node* iff = ifproj->in(0);
1903 if (!iff->is_If()) continue;
1904 Node* bol = iff->in(1);
1905 if (!bol->is_Bool()) continue;
1906 Node* cmp = bol->in(1);
1907 if (cmp == NULL) continue;
1908 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1909 if (cmps[cmpn] == cmp) break;
1910 if (cmpn == NCMPS) continue;
1911 BoolTest::mask btest = bol->as_Bool()->_test._test;
1912 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1913 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1914 // At this point, we know that 'x btest y' is true.
1915 switch (btest) {
1916 case BoolTest::eq:
1917 // They are proven equal, so we can collapse the min/max.
1918 // Either value is the answer. Choose the simpler.
1919 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1920 return yvalue;
1921 return xvalue;
1922 case BoolTest::lt: // x < y
1923 case BoolTest::le: // x <= y
1924 return (want_max ? yvalue : xvalue);
1925 case BoolTest::gt: // x > y
1926 case BoolTest::ge: // x >= y
1927 return (want_max ? xvalue : yvalue);
1928 }
1929 }
1930 }
1932 // We failed to find a dominating test.
1933 // Let's pick a test that might GVN with prior tests.
1934 Node* best_bol = NULL;
1935 BoolTest::mask best_btest = BoolTest::illegal;
1936 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1937 Node* cmp = cmps[cmpn];
1938 if (cmp == NULL) continue;
1939 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1940 Node* bol = cmp->fast_out(j);
1941 if (!bol->is_Bool()) continue;
1942 BoolTest::mask btest = bol->as_Bool()->_test._test;
1943 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1944 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1945 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1946 best_bol = bol->as_Bool();
1947 best_btest = btest;
1948 }
1949 }
1950 }
1952 Node* answer_if_true = NULL;
1953 Node* answer_if_false = NULL;
1954 switch (best_btest) {
1955 default:
1956 if (cmpxy == NULL)
1957 cmpxy = ideal_cmpxy;
1958 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1959 // and fall through:
1960 case BoolTest::lt: // x < y
1961 case BoolTest::le: // x <= y
1962 answer_if_true = (want_max ? yvalue : xvalue);
1963 answer_if_false = (want_max ? xvalue : yvalue);
1964 break;
1965 case BoolTest::gt: // x > y
1966 case BoolTest::ge: // x >= y
1967 answer_if_true = (want_max ? xvalue : yvalue);
1968 answer_if_false = (want_max ? yvalue : xvalue);
1969 break;
1970 }
1972 jint hi, lo;
1973 if (want_max) {
1974 // We can sharpen the minimum.
1975 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1976 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1977 } else {
1978 // We can sharpen the maximum.
1979 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1980 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1981 }
1983 // Use a flow-free graph structure, to avoid creating excess control edges
1984 // which could hinder other optimizations.
1985 // Since Math.min/max is often used with arraycopy, we want
1986 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1987 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1988 answer_if_false, answer_if_true,
1989 TypeInt::make(lo, hi, widen));
1991 return _gvn.transform(cmov);
1993 /*
1994 // This is not as desirable as it may seem, since Min and Max
1995 // nodes do not have a full set of optimizations.
1996 // And they would interfere, anyway, with 'if' optimizations
1997 // and with CMoveI canonical forms.
1998 switch (id) {
1999 case vmIntrinsics::_min:
2000 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
2001 case vmIntrinsics::_max:
2002 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
2003 default:
2004 ShouldNotReachHere();
2005 }
2006 */
2007 }
2009 inline int
2010 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
2011 const TypePtr* base_type = TypePtr::NULL_PTR;
2012 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
2013 if (base_type == NULL) {
2014 // Unknown type.
2015 return Type::AnyPtr;
2016 } else if (base_type == TypePtr::NULL_PTR) {
2017 // Since this is a NULL+long form, we have to switch to a rawptr.
2018 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
2019 offset = MakeConX(0);
2020 return Type::RawPtr;
2021 } else if (base_type->base() == Type::RawPtr) {
2022 return Type::RawPtr;
2023 } else if (base_type->isa_oopptr()) {
2024 // Base is never null => always a heap address.
2025 if (base_type->ptr() == TypePtr::NotNull) {
2026 return Type::OopPtr;
2027 }
2028 // Offset is small => always a heap address.
2029 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
2030 if (offset_type != NULL &&
2031 base_type->offset() == 0 && // (should always be?)
2032 offset_type->_lo >= 0 &&
2033 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
2034 return Type::OopPtr;
2035 }
2036 // Otherwise, it might either be oop+off or NULL+addr.
2037 return Type::AnyPtr;
2038 } else {
2039 // No information:
2040 return Type::AnyPtr;
2041 }
2042 }
2044 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2045 int kind = classify_unsafe_addr(base, offset);
2046 if (kind == Type::RawPtr) {
2047 return basic_plus_adr(top(), base, offset);
2048 } else {
2049 return basic_plus_adr(base, offset);
2050 }
2051 }
2053 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2054 // inline int Integer.numberOfLeadingZeros(int)
2055 // inline int Long.numberOfLeadingZeros(long)
2056 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2057 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2058 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2059 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2060 _sp += arg_size(); // restore stack pointer
2061 switch (id) {
2062 case vmIntrinsics::_numberOfLeadingZeros_i:
2063 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2064 break;
2065 case vmIntrinsics::_numberOfLeadingZeros_l:
2066 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2067 break;
2068 default:
2069 ShouldNotReachHere();
2070 }
2071 return true;
2072 }
2074 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2075 // inline int Integer.numberOfTrailingZeros(int)
2076 // inline int Long.numberOfTrailingZeros(long)
2077 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2078 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2079 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2080 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2081 _sp += arg_size(); // restore stack pointer
2082 switch (id) {
2083 case vmIntrinsics::_numberOfTrailingZeros_i:
2084 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2085 break;
2086 case vmIntrinsics::_numberOfTrailingZeros_l:
2087 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2088 break;
2089 default:
2090 ShouldNotReachHere();
2091 }
2092 return true;
2093 }
2095 //----------------------------inline_bitCount_int/long-----------------------
2096 // inline int Integer.bitCount(int)
2097 // inline int Long.bitCount(long)
2098 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2099 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2100 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2101 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2102 _sp += arg_size(); // restore stack pointer
2103 switch (id) {
2104 case vmIntrinsics::_bitCount_i:
2105 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2106 break;
2107 case vmIntrinsics::_bitCount_l:
2108 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2109 break;
2110 default:
2111 ShouldNotReachHere();
2112 }
2113 return true;
2114 }
2116 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2117 // inline Integer.reverseBytes(int)
2118 // inline Long.reverseBytes(long)
2119 // inline Character.reverseBytes(char)
2120 // inline Short.reverseBytes(short)
2121 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2122 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2123 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2124 "not reverse Bytes");
2125 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2126 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2127 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2128 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2129 _sp += arg_size(); // restore stack pointer
2130 switch (id) {
2131 case vmIntrinsics::_reverseBytes_i:
2132 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2133 break;
2134 case vmIntrinsics::_reverseBytes_l:
2135 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2136 break;
2137 case vmIntrinsics::_reverseBytes_c:
2138 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2139 break;
2140 case vmIntrinsics::_reverseBytes_s:
2141 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2142 break;
2143 default:
2144 ;
2145 }
2146 return true;
2147 }
2149 //----------------------------inline_unsafe_access----------------------------
2151 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2153 // Helper that guards and inserts a G1 pre-barrier.
2154 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2155 assert(UseG1GC, "should not call this otherwise");
2157 // We could be accessing the referent field of a reference object. If so, when G1
2158 // is enabled, we need to log the value in the referent field in an SATB buffer.
2159 // This routine performs some compile time filters and generates suitable
2160 // runtime filters that guard the pre-barrier code.
2162 // Some compile time checks.
2164 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2165 const TypeX* otype = offset->find_intptr_t_type();
2166 if (otype != NULL && otype->is_con() &&
2167 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2168 // Constant offset but not the reference_offset so just return
2169 return;
2170 }
2172 // We only need to generate the runtime guards for instances.
2173 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2174 if (btype != NULL) {
2175 if (btype->isa_aryptr()) {
2176 // Array type so nothing to do
2177 return;
2178 }
2180 const TypeInstPtr* itype = btype->isa_instptr();
2181 if (itype != NULL) {
2182 // Can the klass of base_oop be statically determined
2183 // to be _not_ a sub-class of Reference?
2184 ciKlass* klass = itype->klass();
2185 if (klass->is_subtype_of(env()->Reference_klass()) &&
2186 !env()->Reference_klass()->is_subtype_of(klass)) {
2187 return;
2188 }
2189 }
2190 }
2192 // The compile time filters did not reject base_oop/offset so
2193 // we need to generate the following runtime filters
2194 //
2195 // if (offset == java_lang_ref_Reference::_reference_offset) {
2196 // if (base != null) {
2197 // if (instance_of(base, java.lang.ref.Reference)) {
2198 // pre_barrier(_, pre_val, ...);
2199 // }
2200 // }
2201 // }
2203 float likely = PROB_LIKELY(0.999);
2204 float unlikely = PROB_UNLIKELY(0.999);
2206 IdealKit ideal(this);
2207 #define __ ideal.
2209 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2211 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2212 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2214 // Update graphKit memory and control from IdealKit.
2215 sync_kit(ideal);
2217 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2218 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2220 // Update IdealKit memory and control from graphKit.
2221 __ sync_kit(this);
2223 Node* one = __ ConI(1);
2225 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2227 // Update graphKit from IdeakKit.
2228 sync_kit(ideal);
2230 // Use the pre-barrier to record the value in the referent field
2231 pre_barrier(false /* do_load */,
2232 __ ctrl(),
2233 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2234 pre_val /* pre_val */,
2235 T_OBJECT);
2237 // Update IdealKit from graphKit.
2238 __ sync_kit(this);
2240 } __ end_if(); // _ref_type != ref_none
2241 } __ end_if(); // base != NULL
2242 } __ end_if(); // offset == referent_offset
2244 // Final sync IdealKit and GraphKit.
2245 final_sync(ideal);
2246 #undef __
2247 }
2250 // Interpret Unsafe.fieldOffset cookies correctly:
2251 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2253 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2254 if (callee()->is_static()) return false; // caller must have the capability!
2256 #ifndef PRODUCT
2257 {
2258 ResourceMark rm;
2259 // Check the signatures.
2260 ciSignature* sig = signature();
2261 #ifdef ASSERT
2262 if (!is_store) {
2263 // Object getObject(Object base, int/long offset), etc.
2264 BasicType rtype = sig->return_type()->basic_type();
2265 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2266 rtype = T_ADDRESS; // it is really a C void*
2267 assert(rtype == type, "getter must return the expected value");
2268 if (!is_native_ptr) {
2269 assert(sig->count() == 2, "oop getter has 2 arguments");
2270 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2271 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2272 } else {
2273 assert(sig->count() == 1, "native getter has 1 argument");
2274 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2275 }
2276 } else {
2277 // void putObject(Object base, int/long offset, Object x), etc.
2278 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2279 if (!is_native_ptr) {
2280 assert(sig->count() == 3, "oop putter has 3 arguments");
2281 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2282 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2283 } else {
2284 assert(sig->count() == 2, "native putter has 2 arguments");
2285 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2286 }
2287 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2288 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2289 vtype = T_ADDRESS; // it is really a C void*
2290 assert(vtype == type, "putter must accept the expected value");
2291 }
2292 #endif // ASSERT
2293 }
2294 #endif //PRODUCT
2296 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2298 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2300 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2301 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2303 debug_only(int saved_sp = _sp);
2304 _sp += nargs;
2306 Node* val;
2307 debug_only(val = (Node*)(uintptr_t)-1);
2310 if (is_store) {
2311 // Get the value being stored. (Pop it first; it was pushed last.)
2312 switch (type) {
2313 case T_DOUBLE:
2314 case T_LONG:
2315 case T_ADDRESS:
2316 val = pop_pair();
2317 break;
2318 default:
2319 val = pop();
2320 }
2321 }
2323 // Build address expression. See the code in inline_unsafe_prefetch.
2324 Node *adr;
2325 Node *heap_base_oop = top();
2326 Node* offset = top();
2328 if (!is_native_ptr) {
2329 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2330 offset = pop_pair();
2331 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2332 Node* base = pop();
2333 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2334 // to be plain byte offsets, which are also the same as those accepted
2335 // by oopDesc::field_base.
2336 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2337 "fieldOffset must be byte-scaled");
2338 // 32-bit machines ignore the high half!
2339 offset = ConvL2X(offset);
2340 adr = make_unsafe_address(base, offset);
2341 heap_base_oop = base;
2342 } else {
2343 Node* ptr = pop_pair();
2344 // Adjust Java long to machine word:
2345 ptr = ConvL2X(ptr);
2346 adr = make_unsafe_address(NULL, ptr);
2347 }
2349 // Pop receiver last: it was pushed first.
2350 Node *receiver = pop();
2352 assert(saved_sp == _sp, "must have correct argument count");
2354 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2356 // First guess at the value type.
2357 const Type *value_type = Type::get_const_basic_type(type);
2359 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2360 // there was not enough information to nail it down.
2361 Compile::AliasType* alias_type = C->alias_type(adr_type);
2362 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2364 // We will need memory barriers unless we can determine a unique
2365 // alias category for this reference. (Note: If for some reason
2366 // the barriers get omitted and the unsafe reference begins to "pollute"
2367 // the alias analysis of the rest of the graph, either Compile::can_alias
2368 // or Compile::must_alias will throw a diagnostic assert.)
2369 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2371 // If we are reading the value of the referent field of a Reference
2372 // object (either by using Unsafe directly or through reflection)
2373 // then, if G1 is enabled, we need to record the referent in an
2374 // SATB log buffer using the pre-barrier mechanism.
2375 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2376 offset != top() && heap_base_oop != top();
2378 if (!is_store && type == T_OBJECT) {
2379 // Attempt to infer a sharper value type from the offset and base type.
2380 ciKlass* sharpened_klass = NULL;
2382 // See if it is an instance field, with an object type.
2383 if (alias_type->field() != NULL) {
2384 assert(!is_native_ptr, "native pointer op cannot use a java address");
2385 if (alias_type->field()->type()->is_klass()) {
2386 sharpened_klass = alias_type->field()->type()->as_klass();
2387 }
2388 }
2390 // See if it is a narrow oop array.
2391 if (adr_type->isa_aryptr()) {
2392 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2393 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2394 if (elem_type != NULL) {
2395 sharpened_klass = elem_type->klass();
2396 }
2397 }
2398 }
2400 if (sharpened_klass != NULL) {
2401 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2403 // Sharpen the value type.
2404 value_type = tjp;
2406 #ifndef PRODUCT
2407 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2408 tty->print(" from base type: "); adr_type->dump();
2409 tty->print(" sharpened value: "); value_type->dump();
2410 }
2411 #endif
2412 }
2413 }
2415 // Null check on self without removing any arguments. The argument
2416 // null check technically happens in the wrong place, which can lead to
2417 // invalid stack traces when the primitive is inlined into a method
2418 // which handles NullPointerExceptions.
2419 _sp += nargs;
2420 do_null_check(receiver, T_OBJECT);
2421 _sp -= nargs;
2422 if (stopped()) {
2423 return true;
2424 }
2425 // Heap pointers get a null-check from the interpreter,
2426 // as a courtesy. However, this is not guaranteed by Unsafe,
2427 // and it is not possible to fully distinguish unintended nulls
2428 // from intended ones in this API.
2430 if (is_volatile) {
2431 // We need to emit leading and trailing CPU membars (see below) in
2432 // addition to memory membars when is_volatile. This is a little
2433 // too strong, but avoids the need to insert per-alias-type
2434 // volatile membars (for stores; compare Parse::do_put_xxx), which
2435 // we cannot do effectively here because we probably only have a
2436 // rough approximation of type.
2437 need_mem_bar = true;
2438 // For Stores, place a memory ordering barrier now.
2439 if (is_store)
2440 insert_mem_bar(Op_MemBarRelease);
2441 }
2443 // Memory barrier to prevent normal and 'unsafe' accesses from
2444 // bypassing each other. Happens after null checks, so the
2445 // exception paths do not take memory state from the memory barrier,
2446 // so there's no problems making a strong assert about mixing users
2447 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2448 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2449 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2451 if (!is_store) {
2452 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2453 // load value and push onto stack
2454 switch (type) {
2455 case T_BOOLEAN:
2456 case T_CHAR:
2457 case T_BYTE:
2458 case T_SHORT:
2459 case T_INT:
2460 case T_FLOAT:
2461 push(p);
2462 break;
2463 case T_OBJECT:
2464 if (need_read_barrier) {
2465 insert_g1_pre_barrier(heap_base_oop, offset, p);
2466 }
2467 push(p);
2468 break;
2469 case T_ADDRESS:
2470 // Cast to an int type.
2471 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2472 p = ConvX2L(p);
2473 push_pair(p);
2474 break;
2475 case T_DOUBLE:
2476 case T_LONG:
2477 push_pair( p );
2478 break;
2479 default: ShouldNotReachHere();
2480 }
2481 } else {
2482 // place effect of store into memory
2483 switch (type) {
2484 case T_DOUBLE:
2485 val = dstore_rounding(val);
2486 break;
2487 case T_ADDRESS:
2488 // Repackage the long as a pointer.
2489 val = ConvL2X(val);
2490 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2491 break;
2492 }
2494 if (type != T_OBJECT ) {
2495 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2496 } else {
2497 // Possibly an oop being stored to Java heap or native memory
2498 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2499 // oop to Java heap.
2500 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2501 } else {
2502 // We can't tell at compile time if we are storing in the Java heap or outside
2503 // of it. So we need to emit code to conditionally do the proper type of
2504 // store.
2506 IdealKit ideal(this);
2507 #define __ ideal.
2508 // QQQ who knows what probability is here??
2509 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2510 // Sync IdealKit and graphKit.
2511 sync_kit(ideal);
2512 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2513 // Update IdealKit memory.
2514 __ sync_kit(this);
2515 } __ else_(); {
2516 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2517 } __ end_if();
2518 // Final sync IdealKit and GraphKit.
2519 final_sync(ideal);
2520 #undef __
2521 }
2522 }
2523 }
2525 if (is_volatile) {
2526 if (!is_store)
2527 insert_mem_bar(Op_MemBarAcquire);
2528 else
2529 insert_mem_bar(Op_MemBarVolatile);
2530 }
2532 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2534 return true;
2535 }
2537 //----------------------------inline_unsafe_prefetch----------------------------
2539 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2540 #ifndef PRODUCT
2541 {
2542 ResourceMark rm;
2543 // Check the signatures.
2544 ciSignature* sig = signature();
2545 #ifdef ASSERT
2546 // Object getObject(Object base, int/long offset), etc.
2547 BasicType rtype = sig->return_type()->basic_type();
2548 if (!is_native_ptr) {
2549 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2550 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2551 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2552 } else {
2553 assert(sig->count() == 1, "native prefetch has 1 argument");
2554 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2555 }
2556 #endif // ASSERT
2557 }
2558 #endif // !PRODUCT
2560 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2562 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2563 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2565 debug_only(int saved_sp = _sp);
2566 _sp += nargs;
2568 // Build address expression. See the code in inline_unsafe_access.
2569 Node *adr;
2570 if (!is_native_ptr) {
2571 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2572 Node* offset = pop_pair();
2573 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2574 Node* base = pop();
2575 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2576 // to be plain byte offsets, which are also the same as those accepted
2577 // by oopDesc::field_base.
2578 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2579 "fieldOffset must be byte-scaled");
2580 // 32-bit machines ignore the high half!
2581 offset = ConvL2X(offset);
2582 adr = make_unsafe_address(base, offset);
2583 } else {
2584 Node* ptr = pop_pair();
2585 // Adjust Java long to machine word:
2586 ptr = ConvL2X(ptr);
2587 adr = make_unsafe_address(NULL, ptr);
2588 }
2590 if (is_static) {
2591 assert(saved_sp == _sp, "must have correct argument count");
2592 } else {
2593 // Pop receiver last: it was pushed first.
2594 Node *receiver = pop();
2595 assert(saved_sp == _sp, "must have correct argument count");
2597 // Null check on self without removing any arguments. The argument
2598 // null check technically happens in the wrong place, which can lead to
2599 // invalid stack traces when the primitive is inlined into a method
2600 // which handles NullPointerExceptions.
2601 _sp += nargs;
2602 do_null_check(receiver, T_OBJECT);
2603 _sp -= nargs;
2604 if (stopped()) {
2605 return true;
2606 }
2607 }
2609 // Generate the read or write prefetch
2610 Node *prefetch;
2611 if (is_store) {
2612 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2613 } else {
2614 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2615 }
2616 prefetch->init_req(0, control());
2617 set_i_o(_gvn.transform(prefetch));
2619 return true;
2620 }
2622 //----------------------------inline_unsafe_CAS----------------------------
2624 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2625 // This basic scheme here is the same as inline_unsafe_access, but
2626 // differs in enough details that combining them would make the code
2627 // overly confusing. (This is a true fact! I originally combined
2628 // them, but even I was confused by it!) As much code/comments as
2629 // possible are retained from inline_unsafe_access though to make
2630 // the correspondences clearer. - dl
2632 if (callee()->is_static()) return false; // caller must have the capability!
2634 #ifndef PRODUCT
2635 {
2636 ResourceMark rm;
2637 // Check the signatures.
2638 ciSignature* sig = signature();
2639 #ifdef ASSERT
2640 BasicType rtype = sig->return_type()->basic_type();
2641 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2642 assert(sig->count() == 4, "CAS has 4 arguments");
2643 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2644 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2645 #endif // ASSERT
2646 }
2647 #endif //PRODUCT
2649 // number of stack slots per value argument (1 or 2)
2650 int type_words = type2size[type];
2652 // Cannot inline wide CAS on machines that don't support it natively
2653 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2654 return false;
2656 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2658 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2659 int nargs = 1 + 1 + 2 + type_words + type_words;
2661 // pop arguments: newval, oldval, offset, base, and receiver
2662 debug_only(int saved_sp = _sp);
2663 _sp += nargs;
2664 Node* newval = (type_words == 1) ? pop() : pop_pair();
2665 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2666 Node *offset = pop_pair();
2667 Node *base = pop();
2668 Node *receiver = pop();
2669 assert(saved_sp == _sp, "must have correct argument count");
2671 // Null check receiver.
2672 _sp += nargs;
2673 do_null_check(receiver, T_OBJECT);
2674 _sp -= nargs;
2675 if (stopped()) {
2676 return true;
2677 }
2679 // Build field offset expression.
2680 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2681 // to be plain byte offsets, which are also the same as those accepted
2682 // by oopDesc::field_base.
2683 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2684 // 32-bit machines ignore the high half of long offsets
2685 offset = ConvL2X(offset);
2686 Node* adr = make_unsafe_address(base, offset);
2687 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2689 // (Unlike inline_unsafe_access, there seems no point in trying
2690 // to refine types. Just use the coarse types here.
2691 const Type *value_type = Type::get_const_basic_type(type);
2692 Compile::AliasType* alias_type = C->alias_type(adr_type);
2693 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2694 int alias_idx = C->get_alias_index(adr_type);
2696 // Memory-model-wise, a CAS acts like a little synchronized block,
2697 // so needs barriers on each side. These don't translate into
2698 // actual barriers on most machines, but we still need rest of
2699 // compiler to respect ordering.
2701 insert_mem_bar(Op_MemBarRelease);
2702 insert_mem_bar(Op_MemBarCPUOrder);
2704 // 4984716: MemBars must be inserted before this
2705 // memory node in order to avoid a false
2706 // dependency which will confuse the scheduler.
2707 Node *mem = memory(alias_idx);
2709 // For now, we handle only those cases that actually exist: ints,
2710 // longs, and Object. Adding others should be straightforward.
2711 Node* cas;
2712 switch(type) {
2713 case T_INT:
2714 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2715 break;
2716 case T_LONG:
2717 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2718 break;
2719 case T_OBJECT:
2720 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2721 // could be delayed during Parse (for example, in adjust_map_after_if()).
2722 // Execute transformation here to avoid barrier generation in such case.
2723 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2724 newval = _gvn.makecon(TypePtr::NULL_PTR);
2726 // Reference stores need a store barrier.
2727 // (They don't if CAS fails, but it isn't worth checking.)
2728 pre_barrier(true /* do_load*/,
2729 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2730 NULL /* pre_val*/,
2731 T_OBJECT);
2732 #ifdef _LP64
2733 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2734 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2735 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2736 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2737 newval_enc, oldval_enc));
2738 } else
2739 #endif
2740 {
2741 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2742 }
2743 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2744 break;
2745 default:
2746 ShouldNotReachHere();
2747 break;
2748 }
2750 // SCMemProjNodes represent the memory state of CAS. Their main
2751 // role is to prevent CAS nodes from being optimized away when their
2752 // results aren't used.
2753 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2754 set_memory(proj, alias_idx);
2756 // Add the trailing membar surrounding the access
2757 insert_mem_bar(Op_MemBarCPUOrder);
2758 insert_mem_bar(Op_MemBarAcquire);
2760 push(cas);
2761 return true;
2762 }
2764 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2765 // This is another variant of inline_unsafe_access, differing in
2766 // that it always issues store-store ("release") barrier and ensures
2767 // store-atomicity (which only matters for "long").
2769 if (callee()->is_static()) return false; // caller must have the capability!
2771 #ifndef PRODUCT
2772 {
2773 ResourceMark rm;
2774 // Check the signatures.
2775 ciSignature* sig = signature();
2776 #ifdef ASSERT
2777 BasicType rtype = sig->return_type()->basic_type();
2778 assert(rtype == T_VOID, "must return void");
2779 assert(sig->count() == 3, "has 3 arguments");
2780 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2781 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2782 #endif // ASSERT
2783 }
2784 #endif //PRODUCT
2786 // number of stack slots per value argument (1 or 2)
2787 int type_words = type2size[type];
2789 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2791 // Argument words: "this" plus oop plus offset plus value;
2792 int nargs = 1 + 1 + 2 + type_words;
2794 // pop arguments: val, offset, base, and receiver
2795 debug_only(int saved_sp = _sp);
2796 _sp += nargs;
2797 Node* val = (type_words == 1) ? pop() : pop_pair();
2798 Node *offset = pop_pair();
2799 Node *base = pop();
2800 Node *receiver = pop();
2801 assert(saved_sp == _sp, "must have correct argument count");
2803 // Null check receiver.
2804 _sp += nargs;
2805 do_null_check(receiver, T_OBJECT);
2806 _sp -= nargs;
2807 if (stopped()) {
2808 return true;
2809 }
2811 // Build field offset expression.
2812 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2813 // 32-bit machines ignore the high half of long offsets
2814 offset = ConvL2X(offset);
2815 Node* adr = make_unsafe_address(base, offset);
2816 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2817 const Type *value_type = Type::get_const_basic_type(type);
2818 Compile::AliasType* alias_type = C->alias_type(adr_type);
2820 insert_mem_bar(Op_MemBarRelease);
2821 insert_mem_bar(Op_MemBarCPUOrder);
2822 // Ensure that the store is atomic for longs:
2823 bool require_atomic_access = true;
2824 Node* store;
2825 if (type == T_OBJECT) // reference stores need a store barrier.
2826 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2827 else {
2828 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2829 }
2830 insert_mem_bar(Op_MemBarCPUOrder);
2831 return true;
2832 }
2834 bool LibraryCallKit::inline_unsafe_allocate() {
2835 if (callee()->is_static()) return false; // caller must have the capability!
2836 int nargs = 1 + 1;
2837 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2838 null_check_receiver(callee()); // check then ignore argument(0)
2839 _sp += nargs; // set original stack for use by uncommon_trap
2840 Node* cls = do_null_check(argument(1), T_OBJECT);
2841 _sp -= nargs;
2842 if (stopped()) return true;
2844 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2845 _sp += nargs; // set original stack for use by uncommon_trap
2846 kls = do_null_check(kls, T_OBJECT);
2847 _sp -= nargs;
2848 if (stopped()) return true; // argument was like int.class
2850 // Note: The argument might still be an illegal value like
2851 // Serializable.class or Object[].class. The runtime will handle it.
2852 // But we must make an explicit check for initialization.
2853 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2854 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2855 // can generate code to load it as unsigned byte.
2856 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2857 Node* bits = intcon(instanceKlass::fully_initialized);
2858 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2859 // The 'test' is non-zero if we need to take a slow path.
2861 Node* obj = new_instance(kls, test);
2862 push(obj);
2864 return true;
2865 }
2867 #ifdef TRACE_HAVE_INTRINSICS
2868 /*
2869 * oop -> myklass
2870 * myklass->trace_id |= USED
2871 * return myklass->trace_id & ~0x3
2872 */
2873 bool LibraryCallKit::inline_native_classID() {
2874 int nargs = 1 + 1;
2875 null_check_receiver(callee()); // check then ignore argument(0)
2876 _sp += nargs;
2877 Node* cls = do_null_check(argument(1), T_OBJECT);
2878 _sp -= nargs;
2879 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2880 _sp += nargs;
2881 kls = do_null_check(kls, T_OBJECT);
2882 _sp -= nargs;
2883 ByteSize offset = TRACE_ID_OFFSET;
2884 Node* insp = basic_plus_adr(kls, in_bytes(offset));
2885 Node* tvalue = make_load(NULL, insp, TypeLong::LONG, T_LONG);
2886 Node* bits = longcon(~0x03l); // ignore bit 0 & 1
2887 Node* andl = _gvn.transform(new (C, 3) AndLNode(tvalue, bits));
2888 Node* clsused = longcon(0x01l); // set the class bit
2889 Node* orl = _gvn.transform(new (C, 3) OrLNode(tvalue, clsused));
2891 const TypePtr *adr_type = _gvn.type(insp)->isa_ptr();
2892 store_to_memory(control(), insp, orl, T_LONG, adr_type);
2893 push_pair(andl);
2894 return true;
2895 }
2897 bool LibraryCallKit::inline_native_threadID() {
2898 Node* tls_ptr = NULL;
2899 Node* cur_thr = generate_current_thread(tls_ptr);
2900 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2901 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2902 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::thread_id_offset()));
2904 Node* threadid = NULL;
2905 size_t thread_id_size = OSThread::thread_id_size();
2906 if (thread_id_size == (size_t) BytesPerLong) {
2907 threadid = ConvL2I(make_load(control(), p, TypeLong::LONG, T_LONG));
2908 push(threadid);
2909 } else if (thread_id_size == (size_t) BytesPerInt) {
2910 threadid = make_load(control(), p, TypeInt::INT, T_INT);
2911 push(threadid);
2912 } else {
2913 ShouldNotReachHere();
2914 }
2915 return true;
2916 }
2917 #endif
2919 //------------------------inline_native_time_funcs--------------
2920 // inline code for System.currentTimeMillis() and System.nanoTime()
2921 // these have the same type and signature
2922 bool LibraryCallKit::inline_native_time_funcs(address funcAddr, const char* funcName) {
2923 const TypeFunc *tf = OptoRuntime::void_long_Type();
2924 const TypePtr* no_memory_effects = NULL;
2925 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2926 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2927 #ifdef ASSERT
2928 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2929 assert(value_top == top(), "second value must be top");
2930 #endif
2931 push_pair(value);
2932 return true;
2933 }
2935 //------------------------inline_native_currentThread------------------
2936 bool LibraryCallKit::inline_native_currentThread() {
2937 Node* junk = NULL;
2938 push(generate_current_thread(junk));
2939 return true;
2940 }
2942 //------------------------inline_native_isInterrupted------------------
2943 bool LibraryCallKit::inline_native_isInterrupted() {
2944 const int nargs = 1+1; // receiver + boolean
2945 assert(nargs == arg_size(), "sanity");
2946 // Add a fast path to t.isInterrupted(clear_int):
2947 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2948 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2949 // So, in the common case that the interrupt bit is false,
2950 // we avoid making a call into the VM. Even if the interrupt bit
2951 // is true, if the clear_int argument is false, we avoid the VM call.
2952 // However, if the receiver is not currentThread, we must call the VM,
2953 // because there must be some locking done around the operation.
2955 // We only go to the fast case code if we pass two guards.
2956 // Paths which do not pass are accumulated in the slow_region.
2957 RegionNode* slow_region = new (C, 1) RegionNode(1);
2958 record_for_igvn(slow_region);
2959 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2960 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2961 enum { no_int_result_path = 1,
2962 no_clear_result_path = 2,
2963 slow_result_path = 3
2964 };
2966 // (a) Receiving thread must be the current thread.
2967 Node* rec_thr = argument(0);
2968 Node* tls_ptr = NULL;
2969 Node* cur_thr = generate_current_thread(tls_ptr);
2970 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2971 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2973 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2974 if (!known_current_thread)
2975 generate_slow_guard(bol_thr, slow_region);
2977 // (b) Interrupt bit on TLS must be false.
2978 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2979 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2980 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2981 // Set the control input on the field _interrupted read to prevent it floating up.
2982 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2983 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2984 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2986 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2988 // First fast path: if (!TLS._interrupted) return false;
2989 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2990 result_rgn->init_req(no_int_result_path, false_bit);
2991 result_val->init_req(no_int_result_path, intcon(0));
2993 // drop through to next case
2994 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2996 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2997 Node* clr_arg = argument(1);
2998 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2999 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
3000 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
3002 // Second fast path: ... else if (!clear_int) return true;
3003 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
3004 result_rgn->init_req(no_clear_result_path, false_arg);
3005 result_val->init_req(no_clear_result_path, intcon(1));
3007 // drop through to next case
3008 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
3010 // (d) Otherwise, go to the slow path.
3011 slow_region->add_req(control());
3012 set_control( _gvn.transform(slow_region) );
3014 if (stopped()) {
3015 // There is no slow path.
3016 result_rgn->init_req(slow_result_path, top());
3017 result_val->init_req(slow_result_path, top());
3018 } else {
3019 // non-virtual because it is a private non-static
3020 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
3022 Node* slow_val = set_results_for_java_call(slow_call);
3023 // this->control() comes from set_results_for_java_call
3025 // If we know that the result of the slow call will be true, tell the optimizer!
3026 if (known_current_thread) slow_val = intcon(1);
3028 Node* fast_io = slow_call->in(TypeFunc::I_O);
3029 Node* fast_mem = slow_call->in(TypeFunc::Memory);
3030 // These two phis are pre-filled with copies of of the fast IO and Memory
3031 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
3032 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
3034 result_rgn->init_req(slow_result_path, control());
3035 io_phi ->init_req(slow_result_path, i_o());
3036 mem_phi ->init_req(slow_result_path, reset_memory());
3037 result_val->init_req(slow_result_path, slow_val);
3039 set_all_memory( _gvn.transform(mem_phi) );
3040 set_i_o( _gvn.transform(io_phi) );
3041 }
3043 push_result(result_rgn, result_val);
3044 C->set_has_split_ifs(true); // Has chance for split-if optimization
3046 return true;
3047 }
3049 //---------------------------load_mirror_from_klass----------------------------
3050 // Given a klass oop, load its java mirror (a java.lang.Class oop).
3051 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
3052 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
3053 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
3054 }
3056 //-----------------------load_klass_from_mirror_common-------------------------
3057 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
3058 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
3059 // and branch to the given path on the region.
3060 // If never_see_null, take an uncommon trap on null, so we can optimistically
3061 // compile for the non-null case.
3062 // If the region is NULL, force never_see_null = true.
3063 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
3064 bool never_see_null,
3065 int nargs,
3066 RegionNode* region,
3067 int null_path,
3068 int offset) {
3069 if (region == NULL) never_see_null = true;
3070 Node* p = basic_plus_adr(mirror, offset);
3071 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3072 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
3073 _sp += nargs; // any deopt will start just before call to enclosing method
3074 Node* null_ctl = top();
3075 kls = null_check_oop(kls, &null_ctl, never_see_null);
3076 if (region != NULL) {
3077 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
3078 region->init_req(null_path, null_ctl);
3079 } else {
3080 assert(null_ctl == top(), "no loose ends");
3081 }
3082 _sp -= nargs;
3083 return kls;
3084 }
3086 //--------------------(inline_native_Class_query helpers)---------------------
3087 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3088 // Fall through if (mods & mask) == bits, take the guard otherwise.
3089 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3090 // Branch around if the given klass has the given modifier bit set.
3091 // Like generate_guard, adds a new path onto the region.
3092 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3093 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3094 Node* mask = intcon(modifier_mask);
3095 Node* bits = intcon(modifier_bits);
3096 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3097 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3098 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3099 return generate_fair_guard(bol, region);
3100 }
3101 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3102 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3103 }
3105 //-------------------------inline_native_Class_query-------------------
3106 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3107 int nargs = 1+0; // just the Class mirror, in most cases
3108 const Type* return_type = TypeInt::BOOL;
3109 Node* prim_return_value = top(); // what happens if it's a primitive class?
3110 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3111 bool expect_prim = false; // most of these guys expect to work on refs
3113 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3115 switch (id) {
3116 case vmIntrinsics::_isInstance:
3117 nargs = 1+1; // the Class mirror, plus the object getting queried about
3118 // nothing is an instance of a primitive type
3119 prim_return_value = intcon(0);
3120 break;
3121 case vmIntrinsics::_getModifiers:
3122 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3123 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3124 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3125 break;
3126 case vmIntrinsics::_isInterface:
3127 prim_return_value = intcon(0);
3128 break;
3129 case vmIntrinsics::_isArray:
3130 prim_return_value = intcon(0);
3131 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3132 break;
3133 case vmIntrinsics::_isPrimitive:
3134 prim_return_value = intcon(1);
3135 expect_prim = true; // obviously
3136 break;
3137 case vmIntrinsics::_getSuperclass:
3138 prim_return_value = null();
3139 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3140 break;
3141 case vmIntrinsics::_getComponentType:
3142 prim_return_value = null();
3143 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3144 break;
3145 case vmIntrinsics::_getClassAccessFlags:
3146 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3147 return_type = TypeInt::INT; // not bool! 6297094
3148 break;
3149 default:
3150 ShouldNotReachHere();
3151 }
3153 Node* mirror = argument(0);
3154 Node* obj = (nargs <= 1)? top(): argument(1);
3156 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3157 if (mirror_con == NULL) return false; // cannot happen?
3159 #ifndef PRODUCT
3160 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3161 ciType* k = mirror_con->java_mirror_type();
3162 if (k) {
3163 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3164 k->print_name();
3165 tty->cr();
3166 }
3167 }
3168 #endif
3170 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3171 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3172 record_for_igvn(region);
3173 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3175 // The mirror will never be null of Reflection.getClassAccessFlags, however
3176 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3177 // if it is. See bug 4774291.
3179 // For Reflection.getClassAccessFlags(), the null check occurs in
3180 // the wrong place; see inline_unsafe_access(), above, for a similar
3181 // situation.
3182 _sp += nargs; // set original stack for use by uncommon_trap
3183 mirror = do_null_check(mirror, T_OBJECT);
3184 _sp -= nargs;
3185 // If mirror or obj is dead, only null-path is taken.
3186 if (stopped()) return true;
3188 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3190 // Now load the mirror's klass metaobject, and null-check it.
3191 // Side-effects region with the control path if the klass is null.
3192 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3193 region, _prim_path);
3194 // If kls is null, we have a primitive mirror.
3195 phi->init_req(_prim_path, prim_return_value);
3196 if (stopped()) { push_result(region, phi); return true; }
3198 Node* p; // handy temp
3199 Node* null_ctl;
3201 // Now that we have the non-null klass, we can perform the real query.
3202 // For constant classes, the query will constant-fold in LoadNode::Value.
3203 Node* query_value = top();
3204 switch (id) {
3205 case vmIntrinsics::_isInstance:
3206 // nothing is an instance of a primitive type
3207 _sp += nargs; // gen_instanceof might do an uncommon trap
3208 query_value = gen_instanceof(obj, kls);
3209 _sp -= nargs;
3210 break;
3212 case vmIntrinsics::_getModifiers:
3213 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3214 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3215 break;
3217 case vmIntrinsics::_isInterface:
3218 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3219 if (generate_interface_guard(kls, region) != NULL)
3220 // A guard was added. If the guard is taken, it was an interface.
3221 phi->add_req(intcon(1));
3222 // If we fall through, it's a plain class.
3223 query_value = intcon(0);
3224 break;
3226 case vmIntrinsics::_isArray:
3227 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3228 if (generate_array_guard(kls, region) != NULL)
3229 // A guard was added. If the guard is taken, it was an array.
3230 phi->add_req(intcon(1));
3231 // If we fall through, it's a plain class.
3232 query_value = intcon(0);
3233 break;
3235 case vmIntrinsics::_isPrimitive:
3236 query_value = intcon(0); // "normal" path produces false
3237 break;
3239 case vmIntrinsics::_getSuperclass:
3240 // The rules here are somewhat unfortunate, but we can still do better
3241 // with random logic than with a JNI call.
3242 // Interfaces store null or Object as _super, but must report null.
3243 // Arrays store an intermediate super as _super, but must report Object.
3244 // Other types can report the actual _super.
3245 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3246 if (generate_interface_guard(kls, region) != NULL)
3247 // A guard was added. If the guard is taken, it was an interface.
3248 phi->add_req(null());
3249 if (generate_array_guard(kls, region) != NULL)
3250 // A guard was added. If the guard is taken, it was an array.
3251 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3252 // If we fall through, it's a plain class. Get its _super.
3253 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3254 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3255 null_ctl = top();
3256 kls = null_check_oop(kls, &null_ctl);
3257 if (null_ctl != top()) {
3258 // If the guard is taken, Object.superClass is null (both klass and mirror).
3259 region->add_req(null_ctl);
3260 phi ->add_req(null());
3261 }
3262 if (!stopped()) {
3263 query_value = load_mirror_from_klass(kls);
3264 }
3265 break;
3267 case vmIntrinsics::_getComponentType:
3268 if (generate_array_guard(kls, region) != NULL) {
3269 // Be sure to pin the oop load to the guard edge just created:
3270 Node* is_array_ctrl = region->in(region->req()-1);
3271 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3272 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3273 phi->add_req(cmo);
3274 }
3275 query_value = null(); // non-array case is null
3276 break;
3278 case vmIntrinsics::_getClassAccessFlags:
3279 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3280 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3281 break;
3283 default:
3284 ShouldNotReachHere();
3285 }
3287 // Fall-through is the normal case of a query to a real class.
3288 phi->init_req(1, query_value);
3289 region->init_req(1, control());
3291 push_result(region, phi);
3292 C->set_has_split_ifs(true); // Has chance for split-if optimization
3294 return true;
3295 }
3297 //--------------------------inline_native_subtype_check------------------------
3298 // This intrinsic takes the JNI calls out of the heart of
3299 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3300 bool LibraryCallKit::inline_native_subtype_check() {
3301 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3303 // Pull both arguments off the stack.
3304 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3305 args[0] = argument(0);
3306 args[1] = argument(1);
3307 Node* klasses[2]; // corresponding Klasses: superk, subk
3308 klasses[0] = klasses[1] = top();
3310 enum {
3311 // A full decision tree on {superc is prim, subc is prim}:
3312 _prim_0_path = 1, // {P,N} => false
3313 // {P,P} & superc!=subc => false
3314 _prim_same_path, // {P,P} & superc==subc => true
3315 _prim_1_path, // {N,P} => false
3316 _ref_subtype_path, // {N,N} & subtype check wins => true
3317 _both_ref_path, // {N,N} & subtype check loses => false
3318 PATH_LIMIT
3319 };
3321 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3322 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3323 record_for_igvn(region);
3325 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3326 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3327 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3329 // First null-check both mirrors and load each mirror's klass metaobject.
3330 int which_arg;
3331 for (which_arg = 0; which_arg <= 1; which_arg++) {
3332 Node* arg = args[which_arg];
3333 _sp += nargs; // set original stack for use by uncommon_trap
3334 arg = do_null_check(arg, T_OBJECT);
3335 _sp -= nargs;
3336 if (stopped()) break;
3337 args[which_arg] = _gvn.transform(arg);
3339 Node* p = basic_plus_adr(arg, class_klass_offset);
3340 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3341 klasses[which_arg] = _gvn.transform(kls);
3342 }
3344 // Having loaded both klasses, test each for null.
3345 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3346 for (which_arg = 0; which_arg <= 1; which_arg++) {
3347 Node* kls = klasses[which_arg];
3348 Node* null_ctl = top();
3349 _sp += nargs; // set original stack for use by uncommon_trap
3350 kls = null_check_oop(kls, &null_ctl, never_see_null);
3351 _sp -= nargs;
3352 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3353 region->init_req(prim_path, null_ctl);
3354 if (stopped()) break;
3355 klasses[which_arg] = kls;
3356 }
3358 if (!stopped()) {
3359 // now we have two reference types, in klasses[0..1]
3360 Node* subk = klasses[1]; // the argument to isAssignableFrom
3361 Node* superk = klasses[0]; // the receiver
3362 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3363 // now we have a successful reference subtype check
3364 region->set_req(_ref_subtype_path, control());
3365 }
3367 // If both operands are primitive (both klasses null), then
3368 // we must return true when they are identical primitives.
3369 // It is convenient to test this after the first null klass check.
3370 set_control(region->in(_prim_0_path)); // go back to first null check
3371 if (!stopped()) {
3372 // Since superc is primitive, make a guard for the superc==subc case.
3373 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3374 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3375 generate_guard(bol_eq, region, PROB_FAIR);
3376 if (region->req() == PATH_LIMIT+1) {
3377 // A guard was added. If the added guard is taken, superc==subc.
3378 region->swap_edges(PATH_LIMIT, _prim_same_path);
3379 region->del_req(PATH_LIMIT);
3380 }
3381 region->set_req(_prim_0_path, control()); // Not equal after all.
3382 }
3384 // these are the only paths that produce 'true':
3385 phi->set_req(_prim_same_path, intcon(1));
3386 phi->set_req(_ref_subtype_path, intcon(1));
3388 // pull together the cases:
3389 assert(region->req() == PATH_LIMIT, "sane region");
3390 for (uint i = 1; i < region->req(); i++) {
3391 Node* ctl = region->in(i);
3392 if (ctl == NULL || ctl == top()) {
3393 region->set_req(i, top());
3394 phi ->set_req(i, top());
3395 } else if (phi->in(i) == NULL) {
3396 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3397 }
3398 }
3400 set_control(_gvn.transform(region));
3401 push(_gvn.transform(phi));
3403 return true;
3404 }
3406 //---------------------generate_array_guard_common------------------------
3407 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3408 bool obj_array, bool not_array) {
3409 // If obj_array/non_array==false/false:
3410 // Branch around if the given klass is in fact an array (either obj or prim).
3411 // If obj_array/non_array==false/true:
3412 // Branch around if the given klass is not an array klass of any kind.
3413 // If obj_array/non_array==true/true:
3414 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3415 // If obj_array/non_array==true/false:
3416 // Branch around if the kls is an oop array (Object[] or subtype)
3417 //
3418 // Like generate_guard, adds a new path onto the region.
3419 jint layout_con = 0;
3420 Node* layout_val = get_layout_helper(kls, layout_con);
3421 if (layout_val == NULL) {
3422 bool query = (obj_array
3423 ? Klass::layout_helper_is_objArray(layout_con)
3424 : Klass::layout_helper_is_javaArray(layout_con));
3425 if (query == not_array) {
3426 return NULL; // never a branch
3427 } else { // always a branch
3428 Node* always_branch = control();
3429 if (region != NULL)
3430 region->add_req(always_branch);
3431 set_control(top());
3432 return always_branch;
3433 }
3434 }
3435 // Now test the correct condition.
3436 jint nval = (obj_array
3437 ? ((jint)Klass::_lh_array_tag_type_value
3438 << Klass::_lh_array_tag_shift)
3439 : Klass::_lh_neutral_value);
3440 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3441 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3442 // invert the test if we are looking for a non-array
3443 if (not_array) btest = BoolTest(btest).negate();
3444 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3445 return generate_fair_guard(bol, region);
3446 }
3449 //-----------------------inline_native_newArray--------------------------
3450 bool LibraryCallKit::inline_native_newArray() {
3451 int nargs = 2;
3452 Node* mirror = argument(0);
3453 Node* count_val = argument(1);
3455 _sp += nargs; // set original stack for use by uncommon_trap
3456 mirror = do_null_check(mirror, T_OBJECT);
3457 _sp -= nargs;
3458 // If mirror or obj is dead, only null-path is taken.
3459 if (stopped()) return true;
3461 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3462 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3463 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3464 TypeInstPtr::NOTNULL);
3465 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3466 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3467 TypePtr::BOTTOM);
3469 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3470 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3471 nargs,
3472 result_reg, _slow_path);
3473 Node* normal_ctl = control();
3474 Node* no_array_ctl = result_reg->in(_slow_path);
3476 // Generate code for the slow case. We make a call to newArray().
3477 set_control(no_array_ctl);
3478 if (!stopped()) {
3479 // Either the input type is void.class, or else the
3480 // array klass has not yet been cached. Either the
3481 // ensuing call will throw an exception, or else it
3482 // will cache the array klass for next time.
3483 PreserveJVMState pjvms(this);
3484 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3485 Node* slow_result = set_results_for_java_call(slow_call);
3486 // this->control() comes from set_results_for_java_call
3487 result_reg->set_req(_slow_path, control());
3488 result_val->set_req(_slow_path, slow_result);
3489 result_io ->set_req(_slow_path, i_o());
3490 result_mem->set_req(_slow_path, reset_memory());
3491 }
3493 set_control(normal_ctl);
3494 if (!stopped()) {
3495 // Normal case: The array type has been cached in the java.lang.Class.
3496 // The following call works fine even if the array type is polymorphic.
3497 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3498 Node* obj = new_array(klass_node, count_val, nargs);
3499 result_reg->init_req(_normal_path, control());
3500 result_val->init_req(_normal_path, obj);
3501 result_io ->init_req(_normal_path, i_o());
3502 result_mem->init_req(_normal_path, reset_memory());
3503 }
3505 // Return the combined state.
3506 set_i_o( _gvn.transform(result_io) );
3507 set_all_memory( _gvn.transform(result_mem) );
3508 push_result(result_reg, result_val);
3509 C->set_has_split_ifs(true); // Has chance for split-if optimization
3511 return true;
3512 }
3514 //----------------------inline_native_getLength--------------------------
3515 bool LibraryCallKit::inline_native_getLength() {
3516 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3518 int nargs = 1;
3519 Node* array = argument(0);
3521 _sp += nargs; // set original stack for use by uncommon_trap
3522 array = do_null_check(array, T_OBJECT);
3523 _sp -= nargs;
3525 // If array is dead, only null-path is taken.
3526 if (stopped()) return true;
3528 // Deoptimize if it is a non-array.
3529 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3531 if (non_array != NULL) {
3532 PreserveJVMState pjvms(this);
3533 set_control(non_array);
3534 _sp += nargs; // push the arguments back on the stack
3535 uncommon_trap(Deoptimization::Reason_intrinsic,
3536 Deoptimization::Action_maybe_recompile);
3537 }
3539 // If control is dead, only non-array-path is taken.
3540 if (stopped()) return true;
3542 // The works fine even if the array type is polymorphic.
3543 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3544 push( load_array_length(array) );
3546 C->set_has_split_ifs(true); // Has chance for split-if optimization
3548 return true;
3549 }
3551 //------------------------inline_array_copyOf----------------------------
3552 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3553 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3555 // Restore the stack and pop off the arguments.
3556 int nargs = 3 + (is_copyOfRange? 1: 0);
3557 Node* original = argument(0);
3558 Node* start = is_copyOfRange? argument(1): intcon(0);
3559 Node* end = is_copyOfRange? argument(2): argument(1);
3560 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3562 Node* newcopy;
3564 //set the original stack and the reexecute bit for the interpreter to reexecute
3565 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3566 { PreserveReexecuteState preexecs(this);
3567 _sp += nargs;
3568 jvms()->set_should_reexecute(true);
3570 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3571 original = do_null_check(original, T_OBJECT);
3573 // Check if a null path was taken unconditionally.
3574 if (stopped()) return true;
3576 Node* orig_length = load_array_length(original);
3578 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3579 NULL, 0);
3580 klass_node = do_null_check(klass_node, T_OBJECT);
3582 RegionNode* bailout = new (C, 1) RegionNode(1);
3583 record_for_igvn(bailout);
3585 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3586 // Bail out if that is so.
3587 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3588 if (not_objArray != NULL) {
3589 // Improve the klass node's type from the new optimistic assumption:
3590 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3591 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3592 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3593 cast->init_req(0, control());
3594 klass_node = _gvn.transform(cast);
3595 }
3597 // Bail out if either start or end is negative.
3598 generate_negative_guard(start, bailout, &start);
3599 generate_negative_guard(end, bailout, &end);
3601 Node* length = end;
3602 if (_gvn.type(start) != TypeInt::ZERO) {
3603 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3604 }
3606 // Bail out if length is negative.
3607 // ...Not needed, since the new_array will throw the right exception.
3608 //generate_negative_guard(length, bailout, &length);
3610 if (bailout->req() > 1) {
3611 PreserveJVMState pjvms(this);
3612 set_control( _gvn.transform(bailout) );
3613 uncommon_trap(Deoptimization::Reason_intrinsic,
3614 Deoptimization::Action_maybe_recompile);
3615 }
3617 if (!stopped()) {
3619 // How many elements will we copy from the original?
3620 // The answer is MinI(orig_length - start, length).
3621 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3622 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3624 newcopy = new_array(klass_node, length, 0);
3626 // Generate a direct call to the right arraycopy function(s).
3627 // We know the copy is disjoint but we might not know if the
3628 // oop stores need checking.
3629 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3630 // This will fail a store-check if x contains any non-nulls.
3631 bool disjoint_bases = true;
3632 bool length_never_negative = true;
3633 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3634 original, start, newcopy, intcon(0), moved,
3635 disjoint_bases, length_never_negative);
3636 }
3637 } //original reexecute and sp are set back here
3639 if(!stopped()) {
3640 push(newcopy);
3641 }
3643 C->set_has_split_ifs(true); // Has chance for split-if optimization
3645 return true;
3646 }
3649 //----------------------generate_virtual_guard---------------------------
3650 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3651 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3652 RegionNode* slow_region) {
3653 ciMethod* method = callee();
3654 int vtable_index = method->vtable_index();
3655 // Get the methodOop out of the appropriate vtable entry.
3656 int entry_offset = (instanceKlass::vtable_start_offset() +
3657 vtable_index*vtableEntry::size()) * wordSize +
3658 vtableEntry::method_offset_in_bytes();
3659 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3660 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3662 // Compare the target method with the expected method (e.g., Object.hashCode).
3663 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3665 Node* native_call = makecon(native_call_addr);
3666 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3667 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3669 return generate_slow_guard(test_native, slow_region);
3670 }
3672 //-----------------------generate_method_call----------------------------
3673 // Use generate_method_call to make a slow-call to the real
3674 // method if the fast path fails. An alternative would be to
3675 // use a stub like OptoRuntime::slow_arraycopy_Java.
3676 // This only works for expanding the current library call,
3677 // not another intrinsic. (E.g., don't use this for making an
3678 // arraycopy call inside of the copyOf intrinsic.)
3679 CallJavaNode*
3680 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3681 // When compiling the intrinsic method itself, do not use this technique.
3682 guarantee(callee() != C->method(), "cannot make slow-call to self");
3684 ciMethod* method = callee();
3685 // ensure the JVMS we have will be correct for this call
3686 guarantee(method_id == method->intrinsic_id(), "must match");
3688 const TypeFunc* tf = TypeFunc::make(method);
3689 int tfdc = tf->domain()->cnt();
3690 CallJavaNode* slow_call;
3691 if (is_static) {
3692 assert(!is_virtual, "");
3693 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3694 SharedRuntime::get_resolve_static_call_stub(),
3695 method, bci());
3696 } else if (is_virtual) {
3697 null_check_receiver(method);
3698 int vtable_index = methodOopDesc::invalid_vtable_index;
3699 if (UseInlineCaches) {
3700 // Suppress the vtable call
3701 } else {
3702 // hashCode and clone are not a miranda methods,
3703 // so the vtable index is fixed.
3704 // No need to use the linkResolver to get it.
3705 vtable_index = method->vtable_index();
3706 }
3707 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3708 SharedRuntime::get_resolve_virtual_call_stub(),
3709 method, vtable_index, bci());
3710 } else { // neither virtual nor static: opt_virtual
3711 null_check_receiver(method);
3712 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3713 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3714 method, bci());
3715 slow_call->set_optimized_virtual(true);
3716 }
3717 set_arguments_for_java_call(slow_call);
3718 set_edges_for_java_call(slow_call);
3719 return slow_call;
3720 }
3723 //------------------------------inline_native_hashcode--------------------
3724 // Build special case code for calls to hashCode on an object.
3725 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3726 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3727 assert(!(is_virtual && is_static), "either virtual, special, or static");
3729 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3731 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3732 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3733 TypeInt::INT);
3734 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3735 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3736 TypePtr::BOTTOM);
3737 Node* obj = NULL;
3738 if (!is_static) {
3739 // Check for hashing null object
3740 obj = null_check_receiver(callee());
3741 if (stopped()) return true; // unconditionally null
3742 result_reg->init_req(_null_path, top());
3743 result_val->init_req(_null_path, top());
3744 } else {
3745 // Do a null check, and return zero if null.
3746 // System.identityHashCode(null) == 0
3747 obj = argument(0);
3748 Node* null_ctl = top();
3749 obj = null_check_oop(obj, &null_ctl);
3750 result_reg->init_req(_null_path, null_ctl);
3751 result_val->init_req(_null_path, _gvn.intcon(0));
3752 }
3754 // Unconditionally null? Then return right away.
3755 if (stopped()) {
3756 set_control( result_reg->in(_null_path) );
3757 if (!stopped())
3758 push( result_val ->in(_null_path) );
3759 return true;
3760 }
3762 // After null check, get the object's klass.
3763 Node* obj_klass = load_object_klass(obj);
3765 // This call may be virtual (invokevirtual) or bound (invokespecial).
3766 // For each case we generate slightly different code.
3768 // We only go to the fast case code if we pass a number of guards. The
3769 // paths which do not pass are accumulated in the slow_region.
3770 RegionNode* slow_region = new (C, 1) RegionNode(1);
3771 record_for_igvn(slow_region);
3773 // If this is a virtual call, we generate a funny guard. We pull out
3774 // the vtable entry corresponding to hashCode() from the target object.
3775 // If the target method which we are calling happens to be the native
3776 // Object hashCode() method, we pass the guard. We do not need this
3777 // guard for non-virtual calls -- the caller is known to be the native
3778 // Object hashCode().
3779 if (is_virtual) {
3780 generate_virtual_guard(obj_klass, slow_region);
3781 }
3783 // Get the header out of the object, use LoadMarkNode when available
3784 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3785 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3787 // Test the header to see if it is unlocked.
3788 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3789 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3790 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3791 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3792 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3794 generate_slow_guard(test_unlocked, slow_region);
3796 // Get the hash value and check to see that it has been properly assigned.
3797 // We depend on hash_mask being at most 32 bits and avoid the use of
3798 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3799 // vm: see markOop.hpp.
3800 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3801 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3802 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3803 // This hack lets the hash bits live anywhere in the mark object now, as long
3804 // as the shift drops the relevant bits into the low 32 bits. Note that
3805 // Java spec says that HashCode is an int so there's no point in capturing
3806 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3807 hshifted_header = ConvX2I(hshifted_header);
3808 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3810 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3811 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3812 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3814 generate_slow_guard(test_assigned, slow_region);
3816 Node* init_mem = reset_memory();
3817 // fill in the rest of the null path:
3818 result_io ->init_req(_null_path, i_o());
3819 result_mem->init_req(_null_path, init_mem);
3821 result_val->init_req(_fast_path, hash_val);
3822 result_reg->init_req(_fast_path, control());
3823 result_io ->init_req(_fast_path, i_o());
3824 result_mem->init_req(_fast_path, init_mem);
3826 // Generate code for the slow case. We make a call to hashCode().
3827 set_control(_gvn.transform(slow_region));
3828 if (!stopped()) {
3829 // No need for PreserveJVMState, because we're using up the present state.
3830 set_all_memory(init_mem);
3831 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3832 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3833 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3834 Node* slow_result = set_results_for_java_call(slow_call);
3835 // this->control() comes from set_results_for_java_call
3836 result_reg->init_req(_slow_path, control());
3837 result_val->init_req(_slow_path, slow_result);
3838 result_io ->set_req(_slow_path, i_o());
3839 result_mem ->set_req(_slow_path, reset_memory());
3840 }
3842 // Return the combined state.
3843 set_i_o( _gvn.transform(result_io) );
3844 set_all_memory( _gvn.transform(result_mem) );
3845 push_result(result_reg, result_val);
3847 return true;
3848 }
3850 //---------------------------inline_native_getClass----------------------------
3851 // Build special case code for calls to getClass on an object.
3852 bool LibraryCallKit::inline_native_getClass() {
3853 Node* obj = null_check_receiver(callee());
3854 if (stopped()) return true;
3855 push( load_mirror_from_klass(load_object_klass(obj)) );
3856 return true;
3857 }
3859 //-----------------inline_native_Reflection_getCallerClass---------------------
3860 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3861 //
3862 // NOTE that this code must perform the same logic as
3863 // vframeStream::security_get_caller_frame in that it must skip
3864 // Method.invoke() and auxiliary frames.
3869 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3870 ciMethod* method = callee();
3872 #ifndef PRODUCT
3873 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3874 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3875 }
3876 #endif
3878 debug_only(int saved_sp = _sp);
3880 // Argument words: (int depth)
3881 int nargs = 1;
3883 _sp += nargs;
3884 Node* caller_depth_node = pop();
3886 assert(saved_sp == _sp, "must have correct argument count");
3888 // The depth value must be a constant in order for the runtime call
3889 // to be eliminated.
3890 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3891 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3892 #ifndef PRODUCT
3893 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3894 tty->print_cr(" Bailing out because caller depth was not a constant");
3895 }
3896 #endif
3897 return false;
3898 }
3899 // Note that the JVM state at this point does not include the
3900 // getCallerClass() frame which we are trying to inline. The
3901 // semantics of getCallerClass(), however, are that the "first"
3902 // frame is the getCallerClass() frame, so we subtract one from the
3903 // requested depth before continuing. We don't inline requests of
3904 // getCallerClass(0).
3905 int caller_depth = caller_depth_type->get_con() - 1;
3906 if (caller_depth < 0) {
3907 #ifndef PRODUCT
3908 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3909 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3910 }
3911 #endif
3912 return false;
3913 }
3915 if (!jvms()->has_method()) {
3916 #ifndef PRODUCT
3917 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3918 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3919 }
3920 #endif
3921 return false;
3922 }
3923 int _depth = jvms()->depth(); // cache call chain depth
3925 // Walk back up the JVM state to find the caller at the required
3926 // depth. NOTE that this code must perform the same logic as
3927 // vframeStream::security_get_caller_frame in that it must skip
3928 // Method.invoke() and auxiliary frames. Note also that depth is
3929 // 1-based (1 is the bottom of the inlining).
3930 int inlining_depth = _depth;
3931 JVMState* caller_jvms = NULL;
3933 if (inlining_depth > 0) {
3934 caller_jvms = jvms();
3935 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3936 do {
3937 // The following if-tests should be performed in this order
3938 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3939 // Skip a Method.invoke() or auxiliary frame
3940 } else if (caller_depth > 0) {
3941 // Skip real frame
3942 --caller_depth;
3943 } else {
3944 // We're done: reached desired caller after skipping.
3945 break;
3946 }
3947 caller_jvms = caller_jvms->caller();
3948 --inlining_depth;
3949 } while (inlining_depth > 0);
3950 }
3952 if (inlining_depth == 0) {
3953 #ifndef PRODUCT
3954 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3955 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3956 tty->print_cr(" JVM state at this point:");
3957 for (int i = _depth; i >= 1; i--) {
3958 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3959 }
3960 }
3961 #endif
3962 return false; // Reached end of inlining
3963 }
3965 // Acquire method holder as java.lang.Class
3966 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3967 ciInstance* caller_mirror = caller_klass->java_mirror();
3968 // Push this as a constant
3969 push(makecon(TypeInstPtr::make(caller_mirror)));
3970 #ifndef PRODUCT
3971 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3972 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);
3973 tty->print_cr(" JVM state at this point:");
3974 for (int i = _depth; i >= 1; i--) {
3975 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3976 }
3977 }
3978 #endif
3979 return true;
3980 }
3982 // Helper routine for above
3983 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3984 ciMethod* method = jvms->method();
3986 // Is this the Method.invoke method itself?
3987 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3988 return true;
3990 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3991 ciKlass* k = method->holder();
3992 if (k->is_instance_klass()) {
3993 ciInstanceKlass* ik = k->as_instance_klass();
3994 for (; ik != NULL; ik = ik->super()) {
3995 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3996 ik == env()->find_system_klass(ik->name())) {
3997 return true;
3998 }
3999 }
4000 }
4001 else if (method->is_method_handle_adapter()) {
4002 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
4003 return true;
4004 }
4006 return false;
4007 }
4009 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
4010 // inline_native_AtomicLong_attemptUpdate() but it has no way of
4011 // computing it since there is no lookup field by name function in the
4012 // CI interface. This is computed and set by inline_native_AtomicLong_get().
4013 // Using a static variable here is safe even if we have multiple compilation
4014 // threads because the offset is constant. At worst the same offset will be
4015 // computed and stored multiple
4017 bool LibraryCallKit::inline_native_AtomicLong_get() {
4018 // Restore the stack and pop off the argument
4019 _sp+=1;
4020 Node *obj = pop();
4022 // get the offset of the "value" field. Since the CI interfaces
4023 // does not provide a way to look up a field by name, we scan the bytecodes
4024 // to get the field index. We expect the first 2 instructions of the method
4025 // to be:
4026 // 0 aload_0
4027 // 1 getfield "value"
4028 ciMethod* method = callee();
4029 if (value_field_offset == -1)
4030 {
4031 ciField* value_field;
4032 ciBytecodeStream iter(method);
4033 Bytecodes::Code bc = iter.next();
4035 if ((bc != Bytecodes::_aload_0) &&
4036 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
4037 return false;
4038 bc = iter.next();
4039 if (bc != Bytecodes::_getfield)
4040 return false;
4041 bool ignore;
4042 value_field = iter.get_field(ignore);
4043 value_field_offset = value_field->offset_in_bytes();
4044 }
4046 // Null check without removing any arguments.
4047 _sp++;
4048 obj = do_null_check(obj, T_OBJECT);
4049 _sp--;
4050 // Check for locking null object
4051 if (stopped()) return true;
4053 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
4054 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
4055 int alias_idx = C->get_alias_index(adr_type);
4057 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
4059 push_pair(result);
4061 return true;
4062 }
4064 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
4065 // Restore the stack and pop off the arguments
4066 _sp+=5;
4067 Node *newVal = pop_pair();
4068 Node *oldVal = pop_pair();
4069 Node *obj = pop();
4071 // we need the offset of the "value" field which was computed when
4072 // inlining the get() method. Give up if we don't have it.
4073 if (value_field_offset == -1)
4074 return false;
4076 // Null check without removing any arguments.
4077 _sp+=5;
4078 obj = do_null_check(obj, T_OBJECT);
4079 _sp-=5;
4080 // Check for locking null object
4081 if (stopped()) return true;
4083 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
4084 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
4085 int alias_idx = C->get_alias_index(adr_type);
4087 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
4088 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
4089 set_memory(store_proj, alias_idx);
4090 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
4092 Node *result;
4093 // CMove node is not used to be able fold a possible check code
4094 // after attemptUpdate() call. This code could be transformed
4095 // into CMove node by loop optimizations.
4096 {
4097 RegionNode *r = new (C, 3) RegionNode(3);
4098 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4100 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4101 Node *iftrue = opt_iff(r, iff);
4102 r->init_req(1, iftrue);
4103 result->init_req(1, intcon(1));
4104 result->init_req(2, intcon(0));
4106 set_control(_gvn.transform(r));
4107 record_for_igvn(r);
4109 C->set_has_split_ifs(true); // Has chance for split-if optimization
4110 }
4112 push(_gvn.transform(result));
4113 return true;
4114 }
4116 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4117 // restore the arguments
4118 _sp += arg_size();
4120 switch (id) {
4121 case vmIntrinsics::_floatToRawIntBits:
4122 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4123 break;
4125 case vmIntrinsics::_intBitsToFloat:
4126 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4127 break;
4129 case vmIntrinsics::_doubleToRawLongBits:
4130 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4131 break;
4133 case vmIntrinsics::_longBitsToDouble:
4134 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4135 break;
4137 case vmIntrinsics::_doubleToLongBits: {
4138 Node* value = pop_pair();
4140 // two paths (plus control) merge in a wood
4141 RegionNode *r = new (C, 3) RegionNode(3);
4142 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4144 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4145 // Build the boolean node
4146 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4148 // Branch either way.
4149 // NaN case is less traveled, which makes all the difference.
4150 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4151 Node *opt_isnan = _gvn.transform(ifisnan);
4152 assert( opt_isnan->is_If(), "Expect an IfNode");
4153 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4154 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4156 set_control(iftrue);
4158 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4159 Node *slow_result = longcon(nan_bits); // return NaN
4160 phi->init_req(1, _gvn.transform( slow_result ));
4161 r->init_req(1, iftrue);
4163 // Else fall through
4164 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4165 set_control(iffalse);
4167 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4168 r->init_req(2, iffalse);
4170 // Post merge
4171 set_control(_gvn.transform(r));
4172 record_for_igvn(r);
4174 Node* result = _gvn.transform(phi);
4175 assert(result->bottom_type()->isa_long(), "must be");
4176 push_pair(result);
4178 C->set_has_split_ifs(true); // Has chance for split-if optimization
4180 break;
4181 }
4183 case vmIntrinsics::_floatToIntBits: {
4184 Node* value = pop();
4186 // two paths (plus control) merge in a wood
4187 RegionNode *r = new (C, 3) RegionNode(3);
4188 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4190 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4191 // Build the boolean node
4192 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4194 // Branch either way.
4195 // NaN case is less traveled, which makes all the difference.
4196 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4197 Node *opt_isnan = _gvn.transform(ifisnan);
4198 assert( opt_isnan->is_If(), "Expect an IfNode");
4199 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4200 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4202 set_control(iftrue);
4204 static const jint nan_bits = 0x7fc00000;
4205 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4206 phi->init_req(1, _gvn.transform( slow_result ));
4207 r->init_req(1, iftrue);
4209 // Else fall through
4210 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4211 set_control(iffalse);
4213 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4214 r->init_req(2, iffalse);
4216 // Post merge
4217 set_control(_gvn.transform(r));
4218 record_for_igvn(r);
4220 Node* result = _gvn.transform(phi);
4221 assert(result->bottom_type()->isa_int(), "must be");
4222 push(result);
4224 C->set_has_split_ifs(true); // Has chance for split-if optimization
4226 break;
4227 }
4229 default:
4230 ShouldNotReachHere();
4231 }
4233 return true;
4234 }
4236 #ifdef _LP64
4237 #define XTOP ,top() /*additional argument*/
4238 #else //_LP64
4239 #define XTOP /*no additional argument*/
4240 #endif //_LP64
4242 //----------------------inline_unsafe_copyMemory-------------------------
4243 bool LibraryCallKit::inline_unsafe_copyMemory() {
4244 if (callee()->is_static()) return false; // caller must have the capability!
4245 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4246 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4247 null_check_receiver(callee()); // check then ignore argument(0)
4248 if (stopped()) return true;
4250 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4252 Node* src_ptr = argument(1);
4253 Node* src_off = ConvL2X(argument(2));
4254 assert(argument(3)->is_top(), "2nd half of long");
4255 Node* dst_ptr = argument(4);
4256 Node* dst_off = ConvL2X(argument(5));
4257 assert(argument(6)->is_top(), "2nd half of long");
4258 Node* size = ConvL2X(argument(7));
4259 assert(argument(8)->is_top(), "2nd half of long");
4261 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4262 "fieldOffset must be byte-scaled");
4264 Node* src = make_unsafe_address(src_ptr, src_off);
4265 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4267 // Conservatively insert a memory barrier on all memory slices.
4268 // Do not let writes of the copy source or destination float below the copy.
4269 insert_mem_bar(Op_MemBarCPUOrder);
4271 // Call it. Note that the length argument is not scaled.
4272 make_runtime_call(RC_LEAF|RC_NO_FP,
4273 OptoRuntime::fast_arraycopy_Type(),
4274 StubRoutines::unsafe_arraycopy(),
4275 "unsafe_arraycopy",
4276 TypeRawPtr::BOTTOM,
4277 src, dst, size XTOP);
4279 // Do not let reads of the copy destination float above the copy.
4280 insert_mem_bar(Op_MemBarCPUOrder);
4282 return true;
4283 }
4285 //------------------------clone_coping-----------------------------------
4286 // Helper function for inline_native_clone.
4287 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4288 assert(obj_size != NULL, "");
4289 Node* raw_obj = alloc_obj->in(1);
4290 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4292 AllocateNode* alloc = NULL;
4293 if (ReduceBulkZeroing) {
4294 // We will be completely responsible for initializing this object -
4295 // mark Initialize node as complete.
4296 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4297 // The object was just allocated - there should be no any stores!
4298 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4299 // Mark as complete_with_arraycopy so that on AllocateNode
4300 // expansion, we know this AllocateNode is initialized by an array
4301 // copy and a StoreStore barrier exists after the array copy.
4302 alloc->initialization()->set_complete_with_arraycopy();
4303 }
4305 // Copy the fastest available way.
4306 // TODO: generate fields copies for small objects instead.
4307 Node* src = obj;
4308 Node* dest = alloc_obj;
4309 Node* size = _gvn.transform(obj_size);
4311 // Exclude the header but include array length to copy by 8 bytes words.
4312 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4313 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4314 instanceOopDesc::base_offset_in_bytes();
4315 // base_off:
4316 // 8 - 32-bit VM
4317 // 12 - 64-bit VM, compressed oops
4318 // 16 - 64-bit VM, normal oops
4319 if (base_off % BytesPerLong != 0) {
4320 assert(UseCompressedOops, "");
4321 if (is_array) {
4322 // Exclude length to copy by 8 bytes words.
4323 base_off += sizeof(int);
4324 } else {
4325 // Include klass to copy by 8 bytes words.
4326 base_off = instanceOopDesc::klass_offset_in_bytes();
4327 }
4328 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4329 }
4330 src = basic_plus_adr(src, base_off);
4331 dest = basic_plus_adr(dest, base_off);
4333 // Compute the length also, if needed:
4334 Node* countx = size;
4335 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4336 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4338 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4339 bool disjoint_bases = true;
4340 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4341 src, NULL, dest, NULL, countx,
4342 /*dest_uninitialized*/true);
4344 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4345 if (card_mark) {
4346 assert(!is_array, "");
4347 // Put in store barrier for any and all oops we are sticking
4348 // into this object. (We could avoid this if we could prove
4349 // that the object type contains no oop fields at all.)
4350 Node* no_particular_value = NULL;
4351 Node* no_particular_field = NULL;
4352 int raw_adr_idx = Compile::AliasIdxRaw;
4353 post_barrier(control(),
4354 memory(raw_adr_type),
4355 alloc_obj,
4356 no_particular_field,
4357 raw_adr_idx,
4358 no_particular_value,
4359 T_OBJECT,
4360 false);
4361 }
4363 // Do not let reads from the cloned object float above the arraycopy.
4364 if (alloc != NULL) {
4365 // Do not let stores that initialize this object be reordered with
4366 // a subsequent store that would make this object accessible by
4367 // other threads.
4368 // Record what AllocateNode this StoreStore protects so that
4369 // escape analysis can go from the MemBarStoreStoreNode to the
4370 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4371 // based on the escape status of the AllocateNode.
4372 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4373 } else {
4374 insert_mem_bar(Op_MemBarCPUOrder);
4375 }
4376 }
4378 //------------------------inline_native_clone----------------------------
4379 // Here are the simple edge cases:
4380 // null receiver => normal trap
4381 // virtual and clone was overridden => slow path to out-of-line clone
4382 // not cloneable or finalizer => slow path to out-of-line Object.clone
4383 //
4384 // The general case has two steps, allocation and copying.
4385 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4386 //
4387 // Copying also has two cases, oop arrays and everything else.
4388 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4389 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4390 //
4391 // These steps fold up nicely if and when the cloned object's klass
4392 // can be sharply typed as an object array, a type array, or an instance.
4393 //
4394 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4395 int nargs = 1;
4396 PhiNode* result_val;
4398 //set the original stack and the reexecute bit for the interpreter to reexecute
4399 //the bytecode that invokes Object.clone if deoptimization happens
4400 { PreserveReexecuteState preexecs(this);
4401 jvms()->set_should_reexecute(true);
4403 //null_check_receiver will adjust _sp (push and pop)
4404 Node* obj = null_check_receiver(callee());
4405 if (stopped()) return true;
4407 _sp += nargs;
4409 Node* obj_klass = load_object_klass(obj);
4410 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4411 const TypeOopPtr* toop = ((tklass != NULL)
4412 ? tklass->as_instance_type()
4413 : TypeInstPtr::NOTNULL);
4415 // Conservatively insert a memory barrier on all memory slices.
4416 // Do not let writes into the original float below the clone.
4417 insert_mem_bar(Op_MemBarCPUOrder);
4419 // paths into result_reg:
4420 enum {
4421 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4422 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4423 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4424 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4425 PATH_LIMIT
4426 };
4427 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4428 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4429 TypeInstPtr::NOTNULL);
4430 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4431 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4432 TypePtr::BOTTOM);
4433 record_for_igvn(result_reg);
4435 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4436 int raw_adr_idx = Compile::AliasIdxRaw;
4438 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4439 if (array_ctl != NULL) {
4440 // It's an array.
4441 PreserveJVMState pjvms(this);
4442 set_control(array_ctl);
4443 Node* obj_length = load_array_length(obj);
4444 Node* obj_size = NULL;
4445 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4447 if (!use_ReduceInitialCardMarks()) {
4448 // If it is an oop array, it requires very special treatment,
4449 // because card marking is required on each card of the array.
4450 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4451 if (is_obja != NULL) {
4452 PreserveJVMState pjvms2(this);
4453 set_control(is_obja);
4454 // Generate a direct call to the right arraycopy function(s).
4455 bool disjoint_bases = true;
4456 bool length_never_negative = true;
4457 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4458 obj, intcon(0), alloc_obj, intcon(0),
4459 obj_length,
4460 disjoint_bases, length_never_negative);
4461 result_reg->init_req(_objArray_path, control());
4462 result_val->init_req(_objArray_path, alloc_obj);
4463 result_i_o ->set_req(_objArray_path, i_o());
4464 result_mem ->set_req(_objArray_path, reset_memory());
4465 }
4466 }
4467 // Otherwise, there are no card marks to worry about.
4468 // (We can dispense with card marks if we know the allocation
4469 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4470 // causes the non-eden paths to take compensating steps to
4471 // simulate a fresh allocation, so that no further
4472 // card marks are required in compiled code to initialize
4473 // the object.)
4475 if (!stopped()) {
4476 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4478 // Present the results of the copy.
4479 result_reg->init_req(_array_path, control());
4480 result_val->init_req(_array_path, alloc_obj);
4481 result_i_o ->set_req(_array_path, i_o());
4482 result_mem ->set_req(_array_path, reset_memory());
4483 }
4484 }
4486 // We only go to the instance fast case code if we pass a number of guards.
4487 // The paths which do not pass are accumulated in the slow_region.
4488 RegionNode* slow_region = new (C, 1) RegionNode(1);
4489 record_for_igvn(slow_region);
4490 if (!stopped()) {
4491 // It's an instance (we did array above). Make the slow-path tests.
4492 // If this is a virtual call, we generate a funny guard. We grab
4493 // the vtable entry corresponding to clone() from the target object.
4494 // If the target method which we are calling happens to be the
4495 // Object clone() method, we pass the guard. We do not need this
4496 // guard for non-virtual calls; the caller is known to be the native
4497 // Object clone().
4498 if (is_virtual) {
4499 generate_virtual_guard(obj_klass, slow_region);
4500 }
4502 // The object must be cloneable and must not have a finalizer.
4503 // Both of these conditions may be checked in a single test.
4504 // We could optimize the cloneable test further, but we don't care.
4505 generate_access_flags_guard(obj_klass,
4506 // Test both conditions:
4507 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4508 // Must be cloneable but not finalizer:
4509 JVM_ACC_IS_CLONEABLE,
4510 slow_region);
4511 }
4513 if (!stopped()) {
4514 // It's an instance, and it passed the slow-path tests.
4515 PreserveJVMState pjvms(this);
4516 Node* obj_size = NULL;
4517 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4519 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4521 // Present the results of the slow call.
4522 result_reg->init_req(_instance_path, control());
4523 result_val->init_req(_instance_path, alloc_obj);
4524 result_i_o ->set_req(_instance_path, i_o());
4525 result_mem ->set_req(_instance_path, reset_memory());
4526 }
4528 // Generate code for the slow case. We make a call to clone().
4529 set_control(_gvn.transform(slow_region));
4530 if (!stopped()) {
4531 PreserveJVMState pjvms(this);
4532 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4533 Node* slow_result = set_results_for_java_call(slow_call);
4534 // this->control() comes from set_results_for_java_call
4535 result_reg->init_req(_slow_path, control());
4536 result_val->init_req(_slow_path, slow_result);
4537 result_i_o ->set_req(_slow_path, i_o());
4538 result_mem ->set_req(_slow_path, reset_memory());
4539 }
4541 // Return the combined state.
4542 set_control( _gvn.transform(result_reg) );
4543 set_i_o( _gvn.transform(result_i_o) );
4544 set_all_memory( _gvn.transform(result_mem) );
4545 } //original reexecute and sp are set back here
4547 push(_gvn.transform(result_val));
4549 return true;
4550 }
4552 //------------------------------basictype2arraycopy----------------------------
4553 address LibraryCallKit::basictype2arraycopy(BasicType t,
4554 Node* src_offset,
4555 Node* dest_offset,
4556 bool disjoint_bases,
4557 const char* &name,
4558 bool dest_uninitialized) {
4559 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4560 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4562 bool aligned = false;
4563 bool disjoint = disjoint_bases;
4565 // if the offsets are the same, we can treat the memory regions as
4566 // disjoint, because either the memory regions are in different arrays,
4567 // or they are identical (which we can treat as disjoint.) We can also
4568 // treat a copy with a destination index less that the source index
4569 // as disjoint since a low->high copy will work correctly in this case.
4570 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4571 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4572 // both indices are constants
4573 int s_offs = src_offset_inttype->get_con();
4574 int d_offs = dest_offset_inttype->get_con();
4575 int element_size = type2aelembytes(t);
4576 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4577 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4578 if (s_offs >= d_offs) disjoint = true;
4579 } else if (src_offset == dest_offset && src_offset != NULL) {
4580 // This can occur if the offsets are identical non-constants.
4581 disjoint = true;
4582 }
4584 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4585 }
4588 //------------------------------inline_arraycopy-----------------------
4589 bool LibraryCallKit::inline_arraycopy() {
4590 // Restore the stack and pop off the arguments.
4591 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4592 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4594 Node *src = argument(0);
4595 Node *src_offset = argument(1);
4596 Node *dest = argument(2);
4597 Node *dest_offset = argument(3);
4598 Node *length = argument(4);
4600 // Compile time checks. If any of these checks cannot be verified at compile time,
4601 // we do not make a fast path for this call. Instead, we let the call remain as it
4602 // is. The checks we choose to mandate at compile time are:
4603 //
4604 // (1) src and dest are arrays.
4605 const Type* src_type = src->Value(&_gvn);
4606 const Type* dest_type = dest->Value(&_gvn);
4607 const TypeAryPtr* top_src = src_type->isa_aryptr();
4608 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4609 if (top_src == NULL || top_src->klass() == NULL ||
4610 top_dest == NULL || top_dest->klass() == NULL) {
4611 // Conservatively insert a memory barrier on all memory slices.
4612 // Do not let writes into the source float below the arraycopy.
4613 insert_mem_bar(Op_MemBarCPUOrder);
4615 // Call StubRoutines::generic_arraycopy stub.
4616 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4617 src, src_offset, dest, dest_offset, length);
4619 // Do not let reads from the destination float above the arraycopy.
4620 // Since we cannot type the arrays, we don't know which slices
4621 // might be affected. We could restrict this barrier only to those
4622 // memory slices which pertain to array elements--but don't bother.
4623 if (!InsertMemBarAfterArraycopy)
4624 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4625 insert_mem_bar(Op_MemBarCPUOrder);
4626 return true;
4627 }
4629 // (2) src and dest arrays must have elements of the same BasicType
4630 // Figure out the size and type of the elements we will be copying.
4631 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4632 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4633 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4634 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4636 if (src_elem != dest_elem || dest_elem == T_VOID) {
4637 // The component types are not the same or are not recognized. Punt.
4638 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4639 generate_slow_arraycopy(TypePtr::BOTTOM,
4640 src, src_offset, dest, dest_offset, length,
4641 /*dest_uninitialized*/false);
4642 return true;
4643 }
4645 //---------------------------------------------------------------------------
4646 // We will make a fast path for this call to arraycopy.
4648 // We have the following tests left to perform:
4649 //
4650 // (3) src and dest must not be null.
4651 // (4) src_offset must not be negative.
4652 // (5) dest_offset must not be negative.
4653 // (6) length must not be negative.
4654 // (7) src_offset + length must not exceed length of src.
4655 // (8) dest_offset + length must not exceed length of dest.
4656 // (9) each element of an oop array must be assignable
4658 RegionNode* slow_region = new (C, 1) RegionNode(1);
4659 record_for_igvn(slow_region);
4661 // (3) operands must not be null
4662 // We currently perform our null checks with the do_null_check routine.
4663 // This means that the null exceptions will be reported in the caller
4664 // rather than (correctly) reported inside of the native arraycopy call.
4665 // This should be corrected, given time. We do our null check with the
4666 // stack pointer restored.
4667 _sp += nargs;
4668 src = do_null_check(src, T_ARRAY);
4669 dest = do_null_check(dest, T_ARRAY);
4670 _sp -= nargs;
4672 // (4) src_offset must not be negative.
4673 generate_negative_guard(src_offset, slow_region);
4675 // (5) dest_offset must not be negative.
4676 generate_negative_guard(dest_offset, slow_region);
4678 // (6) length must not be negative (moved to generate_arraycopy()).
4679 // generate_negative_guard(length, slow_region);
4681 // (7) src_offset + length must not exceed length of src.
4682 generate_limit_guard(src_offset, length,
4683 load_array_length(src),
4684 slow_region);
4686 // (8) dest_offset + length must not exceed length of dest.
4687 generate_limit_guard(dest_offset, length,
4688 load_array_length(dest),
4689 slow_region);
4691 // (9) each element of an oop array must be assignable
4692 // The generate_arraycopy subroutine checks this.
4694 // This is where the memory effects are placed:
4695 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4696 generate_arraycopy(adr_type, dest_elem,
4697 src, src_offset, dest, dest_offset, length,
4698 false, false, slow_region);
4700 return true;
4701 }
4703 //-----------------------------generate_arraycopy----------------------
4704 // Generate an optimized call to arraycopy.
4705 // Caller must guard against non-arrays.
4706 // Caller must determine a common array basic-type for both arrays.
4707 // Caller must validate offsets against array bounds.
4708 // The slow_region has already collected guard failure paths
4709 // (such as out of bounds length or non-conformable array types).
4710 // The generated code has this shape, in general:
4711 //
4712 // if (length == 0) return // via zero_path
4713 // slowval = -1
4714 // if (types unknown) {
4715 // slowval = call generic copy loop
4716 // if (slowval == 0) return // via checked_path
4717 // } else if (indexes in bounds) {
4718 // if ((is object array) && !(array type check)) {
4719 // slowval = call checked copy loop
4720 // if (slowval == 0) return // via checked_path
4721 // } else {
4722 // call bulk copy loop
4723 // return // via fast_path
4724 // }
4725 // }
4726 // // adjust params for remaining work:
4727 // if (slowval != -1) {
4728 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4729 // }
4730 // slow_region:
4731 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4732 // return // via slow_call_path
4733 //
4734 // This routine is used from several intrinsics: System.arraycopy,
4735 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4736 //
4737 void
4738 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4739 BasicType basic_elem_type,
4740 Node* src, Node* src_offset,
4741 Node* dest, Node* dest_offset,
4742 Node* copy_length,
4743 bool disjoint_bases,
4744 bool length_never_negative,
4745 RegionNode* slow_region) {
4747 if (slow_region == NULL) {
4748 slow_region = new(C,1) RegionNode(1);
4749 record_for_igvn(slow_region);
4750 }
4752 Node* original_dest = dest;
4753 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4754 bool dest_uninitialized = false;
4756 // See if this is the initialization of a newly-allocated array.
4757 // If so, we will take responsibility here for initializing it to zero.
4758 // (Note: Because tightly_coupled_allocation performs checks on the
4759 // out-edges of the dest, we need to avoid making derived pointers
4760 // from it until we have checked its uses.)
4761 if (ReduceBulkZeroing
4762 && !ZeroTLAB // pointless if already zeroed
4763 && basic_elem_type != T_CONFLICT // avoid corner case
4764 && !src->eqv_uncast(dest)
4765 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4766 != NULL)
4767 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4768 && alloc->maybe_set_complete(&_gvn)) {
4769 // "You break it, you buy it."
4770 InitializeNode* init = alloc->initialization();
4771 assert(init->is_complete(), "we just did this");
4772 init->set_complete_with_arraycopy();
4773 assert(dest->is_CheckCastPP(), "sanity");
4774 assert(dest->in(0)->in(0) == init, "dest pinned");
4775 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4776 // From this point on, every exit path is responsible for
4777 // initializing any non-copied parts of the object to zero.
4778 // Also, if this flag is set we make sure that arraycopy interacts properly
4779 // with G1, eliding pre-barriers. See CR 6627983.
4780 dest_uninitialized = true;
4781 } else {
4782 // No zeroing elimination here.
4783 alloc = NULL;
4784 //original_dest = dest;
4785 //dest_uninitialized = false;
4786 }
4788 // Results are placed here:
4789 enum { fast_path = 1, // normal void-returning assembly stub
4790 checked_path = 2, // special assembly stub with cleanup
4791 slow_call_path = 3, // something went wrong; call the VM
4792 zero_path = 4, // bypass when length of copy is zero
4793 bcopy_path = 5, // copy primitive array by 64-bit blocks
4794 PATH_LIMIT = 6
4795 };
4796 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4797 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4798 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4799 record_for_igvn(result_region);
4800 _gvn.set_type_bottom(result_i_o);
4801 _gvn.set_type_bottom(result_memory);
4802 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4804 // The slow_control path:
4805 Node* slow_control;
4806 Node* slow_i_o = i_o();
4807 Node* slow_mem = memory(adr_type);
4808 debug_only(slow_control = (Node*) badAddress);
4810 // Checked control path:
4811 Node* checked_control = top();
4812 Node* checked_mem = NULL;
4813 Node* checked_i_o = NULL;
4814 Node* checked_value = NULL;
4816 if (basic_elem_type == T_CONFLICT) {
4817 assert(!dest_uninitialized, "");
4818 Node* cv = generate_generic_arraycopy(adr_type,
4819 src, src_offset, dest, dest_offset,
4820 copy_length, dest_uninitialized);
4821 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4822 checked_control = control();
4823 checked_i_o = i_o();
4824 checked_mem = memory(adr_type);
4825 checked_value = cv;
4826 set_control(top()); // no fast path
4827 }
4829 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4830 if (not_pos != NULL) {
4831 PreserveJVMState pjvms(this);
4832 set_control(not_pos);
4834 // (6) length must not be negative.
4835 if (!length_never_negative) {
4836 generate_negative_guard(copy_length, slow_region);
4837 }
4839 // copy_length is 0.
4840 if (!stopped() && dest_uninitialized) {
4841 Node* dest_length = alloc->in(AllocateNode::ALength);
4842 if (copy_length->eqv_uncast(dest_length)
4843 || _gvn.find_int_con(dest_length, 1) <= 0) {
4844 // There is no zeroing to do. No need for a secondary raw memory barrier.
4845 } else {
4846 // Clear the whole thing since there are no source elements to copy.
4847 generate_clear_array(adr_type, dest, basic_elem_type,
4848 intcon(0), NULL,
4849 alloc->in(AllocateNode::AllocSize));
4850 // Use a secondary InitializeNode as raw memory barrier.
4851 // Currently it is needed only on this path since other
4852 // paths have stub or runtime calls as raw memory barriers.
4853 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4854 Compile::AliasIdxRaw,
4855 top())->as_Initialize();
4856 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4857 }
4858 }
4860 // Present the results of the fast call.
4861 result_region->init_req(zero_path, control());
4862 result_i_o ->init_req(zero_path, i_o());
4863 result_memory->init_req(zero_path, memory(adr_type));
4864 }
4866 if (!stopped() && dest_uninitialized) {
4867 // We have to initialize the *uncopied* part of the array to zero.
4868 // The copy destination is the slice dest[off..off+len]. The other slices
4869 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4870 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4871 Node* dest_length = alloc->in(AllocateNode::ALength);
4872 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4873 copy_length) );
4875 // If there is a head section that needs zeroing, do it now.
4876 if (find_int_con(dest_offset, -1) != 0) {
4877 generate_clear_array(adr_type, dest, basic_elem_type,
4878 intcon(0), dest_offset,
4879 NULL);
4880 }
4882 // Next, perform a dynamic check on the tail length.
4883 // It is often zero, and we can win big if we prove this.
4884 // There are two wins: Avoid generating the ClearArray
4885 // with its attendant messy index arithmetic, and upgrade
4886 // the copy to a more hardware-friendly word size of 64 bits.
4887 Node* tail_ctl = NULL;
4888 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4889 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4890 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4891 tail_ctl = generate_slow_guard(bol_lt, NULL);
4892 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4893 }
4895 // At this point, let's assume there is no tail.
4896 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4897 // There is no tail. Try an upgrade to a 64-bit copy.
4898 bool didit = false;
4899 { PreserveJVMState pjvms(this);
4900 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4901 src, src_offset, dest, dest_offset,
4902 dest_size, dest_uninitialized);
4903 if (didit) {
4904 // Present the results of the block-copying fast call.
4905 result_region->init_req(bcopy_path, control());
4906 result_i_o ->init_req(bcopy_path, i_o());
4907 result_memory->init_req(bcopy_path, memory(adr_type));
4908 }
4909 }
4910 if (didit)
4911 set_control(top()); // no regular fast path
4912 }
4914 // Clear the tail, if any.
4915 if (tail_ctl != NULL) {
4916 Node* notail_ctl = stopped() ? NULL : control();
4917 set_control(tail_ctl);
4918 if (notail_ctl == NULL) {
4919 generate_clear_array(adr_type, dest, basic_elem_type,
4920 dest_tail, NULL,
4921 dest_size);
4922 } else {
4923 // Make a local merge.
4924 Node* done_ctl = new(C,3) RegionNode(3);
4925 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4926 done_ctl->init_req(1, notail_ctl);
4927 done_mem->init_req(1, memory(adr_type));
4928 generate_clear_array(adr_type, dest, basic_elem_type,
4929 dest_tail, NULL,
4930 dest_size);
4931 done_ctl->init_req(2, control());
4932 done_mem->init_req(2, memory(adr_type));
4933 set_control( _gvn.transform(done_ctl) );
4934 set_memory( _gvn.transform(done_mem), adr_type );
4935 }
4936 }
4937 }
4939 BasicType copy_type = basic_elem_type;
4940 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4941 if (!stopped() && copy_type == T_OBJECT) {
4942 // If src and dest have compatible element types, we can copy bits.
4943 // Types S[] and D[] are compatible if D is a supertype of S.
4944 //
4945 // If they are not, we will use checked_oop_disjoint_arraycopy,
4946 // which performs a fast optimistic per-oop check, and backs off
4947 // further to JVM_ArrayCopy on the first per-oop check that fails.
4948 // (Actually, we don't move raw bits only; the GC requires card marks.)
4950 // Get the klassOop for both src and dest
4951 Node* src_klass = load_object_klass(src);
4952 Node* dest_klass = load_object_klass(dest);
4954 // Generate the subtype check.
4955 // This might fold up statically, or then again it might not.
4956 //
4957 // Non-static example: Copying List<String>.elements to a new String[].
4958 // The backing store for a List<String> is always an Object[],
4959 // but its elements are always type String, if the generic types
4960 // are correct at the source level.
4961 //
4962 // Test S[] against D[], not S against D, because (probably)
4963 // the secondary supertype cache is less busy for S[] than S.
4964 // This usually only matters when D is an interface.
4965 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4966 // Plug failing path into checked_oop_disjoint_arraycopy
4967 if (not_subtype_ctrl != top()) {
4968 PreserveJVMState pjvms(this);
4969 set_control(not_subtype_ctrl);
4970 // (At this point we can assume disjoint_bases, since types differ.)
4971 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4972 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4973 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4974 Node* dest_elem_klass = _gvn.transform(n1);
4975 Node* cv = generate_checkcast_arraycopy(adr_type,
4976 dest_elem_klass,
4977 src, src_offset, dest, dest_offset,
4978 ConvI2X(copy_length), dest_uninitialized);
4979 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4980 checked_control = control();
4981 checked_i_o = i_o();
4982 checked_mem = memory(adr_type);
4983 checked_value = cv;
4984 }
4985 // At this point we know we do not need type checks on oop stores.
4987 // Let's see if we need card marks:
4988 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4989 // If we do not need card marks, copy using the jint or jlong stub.
4990 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4991 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4992 "sizes agree");
4993 }
4994 }
4996 if (!stopped()) {
4997 // Generate the fast path, if possible.
4998 PreserveJVMState pjvms(this);
4999 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
5000 src, src_offset, dest, dest_offset,
5001 ConvI2X(copy_length), dest_uninitialized);
5003 // Present the results of the fast call.
5004 result_region->init_req(fast_path, control());
5005 result_i_o ->init_req(fast_path, i_o());
5006 result_memory->init_req(fast_path, memory(adr_type));
5007 }
5009 // Here are all the slow paths up to this point, in one bundle:
5010 slow_control = top();
5011 if (slow_region != NULL)
5012 slow_control = _gvn.transform(slow_region);
5013 debug_only(slow_region = (RegionNode*)badAddress);
5015 set_control(checked_control);
5016 if (!stopped()) {
5017 // Clean up after the checked call.
5018 // The returned value is either 0 or -1^K,
5019 // where K = number of partially transferred array elements.
5020 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
5021 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
5022 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
5024 // If it is 0, we are done, so transfer to the end.
5025 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
5026 result_region->init_req(checked_path, checks_done);
5027 result_i_o ->init_req(checked_path, checked_i_o);
5028 result_memory->init_req(checked_path, checked_mem);
5030 // If it is not zero, merge into the slow call.
5031 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
5032 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
5033 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
5034 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
5035 record_for_igvn(slow_reg2);
5036 slow_reg2 ->init_req(1, slow_control);
5037 slow_i_o2 ->init_req(1, slow_i_o);
5038 slow_mem2 ->init_req(1, slow_mem);
5039 slow_reg2 ->init_req(2, control());
5040 slow_i_o2 ->init_req(2, checked_i_o);
5041 slow_mem2 ->init_req(2, checked_mem);
5043 slow_control = _gvn.transform(slow_reg2);
5044 slow_i_o = _gvn.transform(slow_i_o2);
5045 slow_mem = _gvn.transform(slow_mem2);
5047 if (alloc != NULL) {
5048 // We'll restart from the very beginning, after zeroing the whole thing.
5049 // This can cause double writes, but that's OK since dest is brand new.
5050 // So we ignore the low 31 bits of the value returned from the stub.
5051 } else {
5052 // We must continue the copy exactly where it failed, or else
5053 // another thread might see the wrong number of writes to dest.
5054 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
5055 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
5056 slow_offset->init_req(1, intcon(0));
5057 slow_offset->init_req(2, checked_offset);
5058 slow_offset = _gvn.transform(slow_offset);
5060 // Adjust the arguments by the conditionally incoming offset.
5061 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
5062 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
5063 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
5065 // Tweak the node variables to adjust the code produced below:
5066 src_offset = src_off_plus;
5067 dest_offset = dest_off_plus;
5068 copy_length = length_minus;
5069 }
5070 }
5072 set_control(slow_control);
5073 if (!stopped()) {
5074 // Generate the slow path, if needed.
5075 PreserveJVMState pjvms(this); // replace_in_map may trash the map
5077 set_memory(slow_mem, adr_type);
5078 set_i_o(slow_i_o);
5080 if (dest_uninitialized) {
5081 generate_clear_array(adr_type, dest, basic_elem_type,
5082 intcon(0), NULL,
5083 alloc->in(AllocateNode::AllocSize));
5084 }
5086 generate_slow_arraycopy(adr_type,
5087 src, src_offset, dest, dest_offset,
5088 copy_length, /*dest_uninitialized*/false);
5090 result_region->init_req(slow_call_path, control());
5091 result_i_o ->init_req(slow_call_path, i_o());
5092 result_memory->init_req(slow_call_path, memory(adr_type));
5093 }
5095 // Remove unused edges.
5096 for (uint i = 1; i < result_region->req(); i++) {
5097 if (result_region->in(i) == NULL)
5098 result_region->init_req(i, top());
5099 }
5101 // Finished; return the combined state.
5102 set_control( _gvn.transform(result_region) );
5103 set_i_o( _gvn.transform(result_i_o) );
5104 set_memory( _gvn.transform(result_memory), adr_type );
5106 // The memory edges above are precise in order to model effects around
5107 // array copies accurately to allow value numbering of field loads around
5108 // arraycopy. Such field loads, both before and after, are common in Java
5109 // collections and similar classes involving header/array data structures.
5110 //
5111 // But with low number of register or when some registers are used or killed
5112 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5113 // The next memory barrier is added to avoid it. If the arraycopy can be
5114 // optimized away (which it can, sometimes) then we can manually remove
5115 // the membar also.
5116 //
5117 // Do not let reads from the cloned object float above the arraycopy.
5118 if (alloc != NULL) {
5119 // Do not let stores that initialize this object be reordered with
5120 // a subsequent store that would make this object accessible by
5121 // other threads.
5122 // Record what AllocateNode this StoreStore protects so that
5123 // escape analysis can go from the MemBarStoreStoreNode to the
5124 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5125 // based on the escape status of the AllocateNode.
5126 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5127 } else if (InsertMemBarAfterArraycopy)
5128 insert_mem_bar(Op_MemBarCPUOrder);
5129 }
5132 // Helper function which determines if an arraycopy immediately follows
5133 // an allocation, with no intervening tests or other escapes for the object.
5134 AllocateArrayNode*
5135 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5136 RegionNode* slow_region) {
5137 if (stopped()) return NULL; // no fast path
5138 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5140 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5141 if (alloc == NULL) return NULL;
5143 Node* rawmem = memory(Compile::AliasIdxRaw);
5144 // Is the allocation's memory state untouched?
5145 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5146 // Bail out if there have been raw-memory effects since the allocation.
5147 // (Example: There might have been a call or safepoint.)
5148 return NULL;
5149 }
5150 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5151 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5152 return NULL;
5153 }
5155 // There must be no unexpected observers of this allocation.
5156 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5157 Node* obs = ptr->fast_out(i);
5158 if (obs != this->map()) {
5159 return NULL;
5160 }
5161 }
5163 // This arraycopy must unconditionally follow the allocation of the ptr.
5164 Node* alloc_ctl = ptr->in(0);
5165 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5167 Node* ctl = control();
5168 while (ctl != alloc_ctl) {
5169 // There may be guards which feed into the slow_region.
5170 // Any other control flow means that we might not get a chance
5171 // to finish initializing the allocated object.
5172 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5173 IfNode* iff = ctl->in(0)->as_If();
5174 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5175 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5176 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5177 ctl = iff->in(0); // This test feeds the known slow_region.
5178 continue;
5179 }
5180 // One more try: Various low-level checks bottom out in
5181 // uncommon traps. If the debug-info of the trap omits
5182 // any reference to the allocation, as we've already
5183 // observed, then there can be no objection to the trap.
5184 bool found_trap = false;
5185 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5186 Node* obs = not_ctl->fast_out(j);
5187 if (obs->in(0) == not_ctl && obs->is_Call() &&
5188 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5189 found_trap = true; break;
5190 }
5191 }
5192 if (found_trap) {
5193 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5194 continue;
5195 }
5196 }
5197 return NULL;
5198 }
5200 // If we get this far, we have an allocation which immediately
5201 // precedes the arraycopy, and we can take over zeroing the new object.
5202 // The arraycopy will finish the initialization, and provide
5203 // a new control state to which we will anchor the destination pointer.
5205 return alloc;
5206 }
5208 // Helper for initialization of arrays, creating a ClearArray.
5209 // It writes zero bits in [start..end), within the body of an array object.
5210 // The memory effects are all chained onto the 'adr_type' alias category.
5211 //
5212 // Since the object is otherwise uninitialized, we are free
5213 // to put a little "slop" around the edges of the cleared area,
5214 // as long as it does not go back into the array's header,
5215 // or beyond the array end within the heap.
5216 //
5217 // The lower edge can be rounded down to the nearest jint and the
5218 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5219 //
5220 // Arguments:
5221 // adr_type memory slice where writes are generated
5222 // dest oop of the destination array
5223 // basic_elem_type element type of the destination
5224 // slice_idx array index of first element to store
5225 // slice_len number of elements to store (or NULL)
5226 // dest_size total size in bytes of the array object
5227 //
5228 // Exactly one of slice_len or dest_size must be non-NULL.
5229 // If dest_size is non-NULL, zeroing extends to the end of the object.
5230 // If slice_len is non-NULL, the slice_idx value must be a constant.
5231 void
5232 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5233 Node* dest,
5234 BasicType basic_elem_type,
5235 Node* slice_idx,
5236 Node* slice_len,
5237 Node* dest_size) {
5238 // one or the other but not both of slice_len and dest_size:
5239 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5240 if (slice_len == NULL) slice_len = top();
5241 if (dest_size == NULL) dest_size = top();
5243 // operate on this memory slice:
5244 Node* mem = memory(adr_type); // memory slice to operate on
5246 // scaling and rounding of indexes:
5247 int scale = exact_log2(type2aelembytes(basic_elem_type));
5248 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5249 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5250 int bump_bit = (-1 << scale) & BytesPerInt;
5252 // determine constant starts and ends
5253 const intptr_t BIG_NEG = -128;
5254 assert(BIG_NEG + 2*abase < 0, "neg enough");
5255 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5256 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5257 if (slice_len_con == 0) {
5258 return; // nothing to do here
5259 }
5260 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5261 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5262 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5263 assert(end_con < 0, "not two cons");
5264 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5265 BytesPerLong);
5266 }
5268 if (start_con >= 0 && end_con >= 0) {
5269 // Constant start and end. Simple.
5270 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5271 start_con, end_con, &_gvn);
5272 } else if (start_con >= 0 && dest_size != top()) {
5273 // Constant start, pre-rounded end after the tail of the array.
5274 Node* end = dest_size;
5275 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5276 start_con, end, &_gvn);
5277 } else if (start_con >= 0 && slice_len != top()) {
5278 // Constant start, non-constant end. End needs rounding up.
5279 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5280 intptr_t end_base = abase + (slice_idx_con << scale);
5281 int end_round = (-1 << scale) & (BytesPerLong - 1);
5282 Node* end = ConvI2X(slice_len);
5283 if (scale != 0)
5284 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5285 end_base += end_round;
5286 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5287 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5288 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5289 start_con, end, &_gvn);
5290 } else if (start_con < 0 && dest_size != top()) {
5291 // Non-constant start, pre-rounded end after the tail of the array.
5292 // This is almost certainly a "round-to-end" operation.
5293 Node* start = slice_idx;
5294 start = ConvI2X(start);
5295 if (scale != 0)
5296 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5297 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5298 if ((bump_bit | clear_low) != 0) {
5299 int to_clear = (bump_bit | clear_low);
5300 // Align up mod 8, then store a jint zero unconditionally
5301 // just before the mod-8 boundary.
5302 if (((abase + bump_bit) & ~to_clear) - bump_bit
5303 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5304 bump_bit = 0;
5305 assert((abase & to_clear) == 0, "array base must be long-aligned");
5306 } else {
5307 // Bump 'start' up to (or past) the next jint boundary:
5308 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5309 assert((abase & clear_low) == 0, "array base must be int-aligned");
5310 }
5311 // Round bumped 'start' down to jlong boundary in body of array.
5312 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5313 if (bump_bit != 0) {
5314 // Store a zero to the immediately preceding jint:
5315 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5316 Node* p1 = basic_plus_adr(dest, x1);
5317 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5318 mem = _gvn.transform(mem);
5319 }
5320 }
5321 Node* end = dest_size; // pre-rounded
5322 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5323 start, end, &_gvn);
5324 } else {
5325 // Non-constant start, unrounded non-constant end.
5326 // (Nobody zeroes a random midsection of an array using this routine.)
5327 ShouldNotReachHere(); // fix caller
5328 }
5330 // Done.
5331 set_memory(mem, adr_type);
5332 }
5335 bool
5336 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5337 BasicType basic_elem_type,
5338 AllocateNode* alloc,
5339 Node* src, Node* src_offset,
5340 Node* dest, Node* dest_offset,
5341 Node* dest_size, bool dest_uninitialized) {
5342 // See if there is an advantage from block transfer.
5343 int scale = exact_log2(type2aelembytes(basic_elem_type));
5344 if (scale >= LogBytesPerLong)
5345 return false; // it is already a block transfer
5347 // Look at the alignment of the starting offsets.
5348 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5350 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5351 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5352 if (src_off_con < 0 || dest_off_con < 0)
5353 // At present, we can only understand constants.
5354 return false;
5356 intptr_t src_off = abase + (src_off_con << scale);
5357 intptr_t dest_off = abase + (dest_off_con << scale);
5359 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5360 // Non-aligned; too bad.
5361 // One more chance: Pick off an initial 32-bit word.
5362 // This is a common case, since abase can be odd mod 8.
5363 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5364 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5365 Node* sptr = basic_plus_adr(src, src_off);
5366 Node* dptr = basic_plus_adr(dest, dest_off);
5367 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5368 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5369 src_off += BytesPerInt;
5370 dest_off += BytesPerInt;
5371 } else {
5372 return false;
5373 }
5374 }
5375 assert(src_off % BytesPerLong == 0, "");
5376 assert(dest_off % BytesPerLong == 0, "");
5378 // Do this copy by giant steps.
5379 Node* sptr = basic_plus_adr(src, src_off);
5380 Node* dptr = basic_plus_adr(dest, dest_off);
5381 Node* countx = dest_size;
5382 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5383 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5385 bool disjoint_bases = true; // since alloc != NULL
5386 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5387 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5389 return true;
5390 }
5393 // Helper function; generates code for the slow case.
5394 // We make a call to a runtime method which emulates the native method,
5395 // but without the native wrapper overhead.
5396 void
5397 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5398 Node* src, Node* src_offset,
5399 Node* dest, Node* dest_offset,
5400 Node* copy_length, bool dest_uninitialized) {
5401 assert(!dest_uninitialized, "Invariant");
5402 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5403 OptoRuntime::slow_arraycopy_Type(),
5404 OptoRuntime::slow_arraycopy_Java(),
5405 "slow_arraycopy", adr_type,
5406 src, src_offset, dest, dest_offset,
5407 copy_length);
5409 // Handle exceptions thrown by this fellow:
5410 make_slow_call_ex(call, env()->Throwable_klass(), false);
5411 }
5413 // Helper function; generates code for cases requiring runtime checks.
5414 Node*
5415 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5416 Node* dest_elem_klass,
5417 Node* src, Node* src_offset,
5418 Node* dest, Node* dest_offset,
5419 Node* copy_length, bool dest_uninitialized) {
5420 if (stopped()) return NULL;
5422 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5423 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5424 return NULL;
5425 }
5427 // Pick out the parameters required to perform a store-check
5428 // for the target array. This is an optimistic check. It will
5429 // look in each non-null element's class, at the desired klass's
5430 // super_check_offset, for the desired klass.
5431 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5432 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5433 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5434 Node* check_offset = ConvI2X(_gvn.transform(n3));
5435 Node* check_value = dest_elem_klass;
5437 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5438 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5440 // (We know the arrays are never conjoint, because their types differ.)
5441 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5442 OptoRuntime::checkcast_arraycopy_Type(),
5443 copyfunc_addr, "checkcast_arraycopy", adr_type,
5444 // five arguments, of which two are
5445 // intptr_t (jlong in LP64)
5446 src_start, dest_start,
5447 copy_length XTOP,
5448 check_offset XTOP,
5449 check_value);
5451 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5452 }
5455 // Helper function; generates code for cases requiring runtime checks.
5456 Node*
5457 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5458 Node* src, Node* src_offset,
5459 Node* dest, Node* dest_offset,
5460 Node* copy_length, bool dest_uninitialized) {
5461 assert(!dest_uninitialized, "Invariant");
5462 if (stopped()) return NULL;
5463 address copyfunc_addr = StubRoutines::generic_arraycopy();
5464 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5465 return NULL;
5466 }
5468 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5469 OptoRuntime::generic_arraycopy_Type(),
5470 copyfunc_addr, "generic_arraycopy", adr_type,
5471 src, src_offset, dest, dest_offset, copy_length);
5473 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5474 }
5476 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5477 void
5478 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5479 BasicType basic_elem_type,
5480 bool disjoint_bases,
5481 Node* src, Node* src_offset,
5482 Node* dest, Node* dest_offset,
5483 Node* copy_length, bool dest_uninitialized) {
5484 if (stopped()) return; // nothing to do
5486 Node* src_start = src;
5487 Node* dest_start = dest;
5488 if (src_offset != NULL || dest_offset != NULL) {
5489 assert(src_offset != NULL && dest_offset != NULL, "");
5490 src_start = array_element_address(src, src_offset, basic_elem_type);
5491 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5492 }
5494 // Figure out which arraycopy runtime method to call.
5495 const char* copyfunc_name = "arraycopy";
5496 address copyfunc_addr =
5497 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5498 disjoint_bases, copyfunc_name, dest_uninitialized);
5500 // Call it. Note that the count_ix value is not scaled to a byte-size.
5501 make_runtime_call(RC_LEAF|RC_NO_FP,
5502 OptoRuntime::fast_arraycopy_Type(),
5503 copyfunc_addr, copyfunc_name, adr_type,
5504 src_start, dest_start, copy_length XTOP);
5505 }
5507 //----------------------------inline_reference_get----------------------------
5509 bool LibraryCallKit::inline_reference_get() {
5510 const int nargs = 1; // self
5512 guarantee(java_lang_ref_Reference::referent_offset > 0,
5513 "should have already been set");
5515 int referent_offset = java_lang_ref_Reference::referent_offset;
5517 // Restore the stack and pop off the argument
5518 _sp += nargs;
5519 Node *reference_obj = pop();
5521 // Null check on self without removing any arguments.
5522 _sp += nargs;
5523 reference_obj = do_null_check(reference_obj, T_OBJECT);
5524 _sp -= nargs;;
5526 if (stopped()) return true;
5528 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5530 ciInstanceKlass* klass = env()->Object_klass();
5531 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5533 Node* no_ctrl = NULL;
5534 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5536 // Use the pre-barrier to record the value in the referent field
5537 pre_barrier(false /* do_load */,
5538 control(),
5539 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5540 result /* pre_val */,
5541 T_OBJECT);
5543 push(result);
5544 return true;
5545 }