1 /*
2 * Copyright (c) 1999, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/systemDictionary.hpp"
27 #include "classfile/vmSymbols.hpp"
28 #include "compiler/compileBroker.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "oops/objArrayKlass.hpp"
31 #include "opto/addnode.hpp"
32 #include "opto/callGenerator.hpp"
33 #include "opto/cfgnode.hpp"
34 #include "opto/idealKit.hpp"
35 #include "opto/mulnode.hpp"
36 #include "opto/parse.hpp"
37 #include "opto/runtime.hpp"
38 #include "opto/subnode.hpp"
39 #include "prims/nativeLookup.hpp"
40 #include "runtime/sharedRuntime.hpp"
41
42 class LibraryIntrinsic : public InlineCallGenerator {
43 // Extend the set of intrinsics known to the runtime:
44 public:
45 private:
46 bool _is_virtual;
47 vmIntrinsics::ID _intrinsic_id;
48
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 };
61
62
63 // Local helper class for LibraryIntrinsic:
64 class LibraryCallKit : public GraphKit {
65 private:
66 LibraryIntrinsic* _intrinsic; // the library intrinsic being called
67
68 public:
69 LibraryCallKit(JVMState* caller, LibraryIntrinsic* intrinsic)
70 : GraphKit(caller),
71 _intrinsic(intrinsic)
72 {
73 }
74
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(); }
82
83 bool try_to_inline();
84
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 }
149
150 Node* make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2);
151 bool inline_string_compareTo();
152 bool inline_string_indexOf();
153 Node* string_indexOf(Node* string_object, ciTypeArray* target_array, jint offset, jint cache_i, jint md2_i);
154 bool inline_string_equals();
155 Node* pop_math_arg();
156 bool runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName);
157 bool inline_math_native(vmIntrinsics::ID id);
158 bool inline_trig(vmIntrinsics::ID id);
159 bool inline_trans(vmIntrinsics::ID id);
160 bool inline_abs(vmIntrinsics::ID id);
161 bool inline_sqrt(vmIntrinsics::ID id);
162 bool inline_pow(vmIntrinsics::ID id);
163 bool inline_exp(vmIntrinsics::ID id);
164 bool inline_min_max(vmIntrinsics::ID id);
165 Node* generate_min_max(vmIntrinsics::ID id, Node* x, Node* y);
166 // This returns Type::AnyPtr, RawPtr, or OopPtr.
167 int classify_unsafe_addr(Node* &base, Node* &offset);
168 Node* make_unsafe_address(Node* base, Node* offset);
169 // Helper for inline_unsafe_access.
170 // Generates the guards that check whether the result of
171 // Unsafe.getObject should be recorded in an SATB log buffer.
172 void insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val);
173 bool inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile);
174 bool inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static);
175 bool inline_unsafe_allocate();
176 bool inline_unsafe_copyMemory();
177 bool inline_native_currentThread();
178 bool inline_native_time_funcs(bool isNano);
179 bool inline_native_isInterrupted();
180 bool inline_native_Class_query(vmIntrinsics::ID id);
181 bool inline_native_subtype_check();
182
183 bool inline_native_newArray();
184 bool inline_native_getLength();
185 bool inline_array_copyOf(bool is_copyOfRange);
186 bool inline_array_equals();
187 void copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark);
188 bool inline_native_clone(bool is_virtual);
189 bool inline_native_Reflection_getCallerClass();
190 bool inline_native_AtomicLong_get();
191 bool inline_native_AtomicLong_attemptUpdate();
192 bool is_method_invoke_or_aux_frame(JVMState* jvms);
193 // Helper function for inlining native object hash method
194 bool inline_native_hashcode(bool is_virtual, bool is_static);
195 bool inline_native_getClass();
196
197 // Helper functions for inlining arraycopy
198 bool inline_arraycopy();
199 void generate_arraycopy(const TypePtr* adr_type,
200 BasicType basic_elem_type,
201 Node* src, Node* src_offset,
202 Node* dest, Node* dest_offset,
203 Node* copy_length,
204 bool disjoint_bases = false,
205 bool length_never_negative = false,
206 RegionNode* slow_region = NULL);
207 AllocateArrayNode* tightly_coupled_allocation(Node* ptr,
208 RegionNode* slow_region);
209 void generate_clear_array(const TypePtr* adr_type,
210 Node* dest,
211 BasicType basic_elem_type,
212 Node* slice_off,
213 Node* slice_len,
214 Node* slice_end);
215 bool generate_block_arraycopy(const TypePtr* adr_type,
216 BasicType basic_elem_type,
217 AllocateNode* alloc,
218 Node* src, Node* src_offset,
219 Node* dest, Node* dest_offset,
220 Node* dest_size, bool dest_uninitialized);
221 void generate_slow_arraycopy(const TypePtr* adr_type,
222 Node* src, Node* src_offset,
223 Node* dest, Node* dest_offset,
224 Node* copy_length, bool dest_uninitialized);
225 Node* generate_checkcast_arraycopy(const TypePtr* adr_type,
226 Node* dest_elem_klass,
227 Node* src, Node* src_offset,
228 Node* dest, Node* dest_offset,
229 Node* copy_length, bool dest_uninitialized);
230 Node* generate_generic_arraycopy(const TypePtr* adr_type,
231 Node* src, Node* src_offset,
232 Node* dest, Node* dest_offset,
233 Node* copy_length, bool dest_uninitialized);
234 void generate_unchecked_arraycopy(const TypePtr* adr_type,
235 BasicType basic_elem_type,
236 bool disjoint_bases,
237 Node* src, Node* src_offset,
238 Node* dest, Node* dest_offset,
239 Node* copy_length, bool dest_uninitialized);
240 bool inline_unsafe_CAS(BasicType type);
241 bool inline_unsafe_ordered_store(BasicType type);
242 bool inline_fp_conversions(vmIntrinsics::ID id);
243 bool inline_numberOfLeadingZeros(vmIntrinsics::ID id);
244 bool inline_numberOfTrailingZeros(vmIntrinsics::ID id);
245 bool inline_bitCount(vmIntrinsics::ID id);
246 bool inline_reverseBytes(vmIntrinsics::ID id);
247
248 bool inline_reference_get();
249 };
250
251
252 //---------------------------make_vm_intrinsic----------------------------
253 CallGenerator* Compile::make_vm_intrinsic(ciMethod* m, bool is_virtual) {
254 vmIntrinsics::ID id = m->intrinsic_id();
255 assert(id != vmIntrinsics::_none, "must be a VM intrinsic");
256
257 if (DisableIntrinsic[0] != '\0'
258 && strstr(DisableIntrinsic, vmIntrinsics::name_at(id)) != NULL) {
259 // disabled by a user request on the command line:
260 // example: -XX:DisableIntrinsic=_hashCode,_getClass
261 return NULL;
262 }
263
264 if (!m->is_loaded()) {
265 // do not attempt to inline unloaded methods
266 return NULL;
267 }
268
269 // Only a few intrinsics implement a virtual dispatch.
270 // They are expensive calls which are also frequently overridden.
271 if (is_virtual) {
272 switch (id) {
273 case vmIntrinsics::_hashCode:
274 case vmIntrinsics::_clone:
275 // OK, Object.hashCode and Object.clone intrinsics come in both flavors
276 break;
277 default:
278 return NULL;
279 }
280 }
281
282 // -XX:-InlineNatives disables nearly all intrinsics:
283 if (!InlineNatives) {
284 switch (id) {
285 case vmIntrinsics::_indexOf:
286 case vmIntrinsics::_compareTo:
287 case vmIntrinsics::_equals:
288 case vmIntrinsics::_equalsC:
289 break; // InlineNatives does not control String.compareTo
290 default:
291 return NULL;
292 }
293 }
294
295 switch (id) {
296 case vmIntrinsics::_compareTo:
297 if (!SpecialStringCompareTo) return NULL;
298 break;
299 case vmIntrinsics::_indexOf:
300 if (!SpecialStringIndexOf) return NULL;
301 break;
302 case vmIntrinsics::_equals:
303 if (!SpecialStringEquals) return NULL;
304 break;
305 case vmIntrinsics::_equalsC:
306 if (!SpecialArraysEquals) return NULL;
307 break;
308 case vmIntrinsics::_arraycopy:
309 if (!InlineArrayCopy) return NULL;
310 break;
311 case vmIntrinsics::_copyMemory:
312 if (StubRoutines::unsafe_arraycopy() == NULL) return NULL;
313 if (!InlineArrayCopy) return NULL;
314 break;
315 case vmIntrinsics::_hashCode:
316 if (!InlineObjectHash) return NULL;
317 break;
318 case vmIntrinsics::_clone:
319 case vmIntrinsics::_copyOf:
320 case vmIntrinsics::_copyOfRange:
321 if (!InlineObjectCopy) return NULL;
322 // These also use the arraycopy intrinsic mechanism:
323 if (!InlineArrayCopy) return NULL;
324 break;
325 case vmIntrinsics::_checkIndex:
326 // We do not intrinsify this. The optimizer does fine with it.
327 return NULL;
328
329 case vmIntrinsics::_get_AtomicLong:
330 case vmIntrinsics::_attemptUpdate:
331 if (!InlineAtomicLong) return NULL;
332 break;
333
334 case vmIntrinsics::_getCallerClass:
335 if (!UseNewReflection) return NULL;
336 if (!InlineReflectionGetCallerClass) return NULL;
337 if (!JDK_Version::is_gte_jdk14x_version()) return NULL;
338 break;
339
340 case vmIntrinsics::_Reference_get:
341 // It is only when G1 is enabled that we absolutely
342 // need to use the intrinsic version of Reference.get()
343 // so that the value in the referent field, if necessary,
344 // can be registered by the pre-barrier code.
345 if (!UseG1GC) return NULL;
346 break;
347
348 default:
349 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
350 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
351 break;
352 }
353
354 // -XX:-InlineClassNatives disables natives from the Class class.
355 // The flag applies to all reflective calls, notably Array.newArray
356 // (visible to Java programmers as Array.newInstance).
357 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
358 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
359 if (!InlineClassNatives) return NULL;
360 }
361
362 // -XX:-InlineThreadNatives disables natives from the Thread class.
363 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
364 if (!InlineThreadNatives) return NULL;
365 }
366
367 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
368 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
369 m->holder()->name() == ciSymbol::java_lang_Float() ||
370 m->holder()->name() == ciSymbol::java_lang_Double()) {
371 if (!InlineMathNatives) return NULL;
372 }
373
374 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
375 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
376 if (!InlineUnsafeOps) return NULL;
377 }
378
379 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
380 }
381
382 //----------------------register_library_intrinsics-----------------------
383 // Initialize this file's data structures, for each Compile instance.
384 void Compile::register_library_intrinsics() {
385 // Nothing to do here.
386 }
387
388 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
389 LibraryCallKit kit(jvms, this);
390 Compile* C = kit.C;
391 int nodes = C->unique();
392 #ifndef PRODUCT
393 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
394 char buf[1000];
395 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
396 tty->print_cr("Intrinsic %s", str);
397 }
398 #endif
399
400 if (kit.try_to_inline()) {
401 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
402 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
403 }
404 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
405 if (C->log()) {
406 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
407 vmIntrinsics::name_at(intrinsic_id()),
408 (is_virtual() ? " virtual='1'" : ""),
409 C->unique() - nodes);
410 }
411 return kit.transfer_exceptions_into_jvms();
412 }
413
414 // The intrinsic bailed out
415 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
416 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
417 if (jvms->has_method()) {
418 // Not a root compile.
419 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
420 } else {
421 // Root compile
422 tty->print("Did not generate intrinsic %s%s at bci:%d in",
423 vmIntrinsics::name_at(intrinsic_id()),
424 (is_virtual() ? " (virtual)" : ""), kit.bci());
425 }
426 }
427 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
428 return NULL;
429 }
430
431 bool LibraryCallKit::try_to_inline() {
432 // Handle symbolic names for otherwise undistinguished boolean switches:
433 const bool is_store = true;
434 const bool is_native_ptr = true;
435 const bool is_static = true;
436
437 if (!jvms()->has_method()) {
438 // Root JVMState has a null method.
439 assert(map()->memory()->Opcode() == Op_Parm, "");
440 // Insert the memory aliasing node
441 set_all_memory(reset_memory());
442 }
443 assert(merged_memory(), "");
444
445 switch (intrinsic_id()) {
446 case vmIntrinsics::_hashCode:
447 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
448 case vmIntrinsics::_identityHashCode:
449 return inline_native_hashcode(/*!virtual*/ false, is_static);
450 case vmIntrinsics::_getClass:
451 return inline_native_getClass();
452
453 case vmIntrinsics::_dsin:
454 case vmIntrinsics::_dcos:
455 case vmIntrinsics::_dtan:
456 case vmIntrinsics::_dabs:
457 case vmIntrinsics::_datan2:
458 case vmIntrinsics::_dsqrt:
459 case vmIntrinsics::_dexp:
460 case vmIntrinsics::_dlog:
461 case vmIntrinsics::_dlog10:
462 case vmIntrinsics::_dpow:
463 return inline_math_native(intrinsic_id());
464
465 case vmIntrinsics::_min:
466 case vmIntrinsics::_max:
467 return inline_min_max(intrinsic_id());
468
469 case vmIntrinsics::_arraycopy:
470 return inline_arraycopy();
471
472 case vmIntrinsics::_compareTo:
473 return inline_string_compareTo();
474 case vmIntrinsics::_indexOf:
475 return inline_string_indexOf();
476 case vmIntrinsics::_equals:
477 return inline_string_equals();
478
479 case vmIntrinsics::_getObject:
480 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
481 case vmIntrinsics::_getBoolean:
482 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
483 case vmIntrinsics::_getByte:
484 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
485 case vmIntrinsics::_getShort:
486 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
487 case vmIntrinsics::_getChar:
488 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
489 case vmIntrinsics::_getInt:
490 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
491 case vmIntrinsics::_getLong:
492 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
493 case vmIntrinsics::_getFloat:
494 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
495 case vmIntrinsics::_getDouble:
496 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
497
498 case vmIntrinsics::_putObject:
499 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
500 case vmIntrinsics::_putBoolean:
501 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
502 case vmIntrinsics::_putByte:
503 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
504 case vmIntrinsics::_putShort:
505 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
506 case vmIntrinsics::_putChar:
507 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
508 case vmIntrinsics::_putInt:
509 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
510 case vmIntrinsics::_putLong:
511 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
512 case vmIntrinsics::_putFloat:
513 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
514 case vmIntrinsics::_putDouble:
515 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
516
517 case vmIntrinsics::_getByte_raw:
518 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
519 case vmIntrinsics::_getShort_raw:
520 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
521 case vmIntrinsics::_getChar_raw:
522 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
523 case vmIntrinsics::_getInt_raw:
524 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
525 case vmIntrinsics::_getLong_raw:
526 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
527 case vmIntrinsics::_getFloat_raw:
528 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
529 case vmIntrinsics::_getDouble_raw:
530 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
531 case vmIntrinsics::_getAddress_raw:
532 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
533
534 case vmIntrinsics::_putByte_raw:
535 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
536 case vmIntrinsics::_putShort_raw:
537 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
538 case vmIntrinsics::_putChar_raw:
539 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
540 case vmIntrinsics::_putInt_raw:
541 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
542 case vmIntrinsics::_putLong_raw:
543 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
544 case vmIntrinsics::_putFloat_raw:
545 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
546 case vmIntrinsics::_putDouble_raw:
547 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
548 case vmIntrinsics::_putAddress_raw:
549 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
550
551 case vmIntrinsics::_getObjectVolatile:
552 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
553 case vmIntrinsics::_getBooleanVolatile:
554 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
555 case vmIntrinsics::_getByteVolatile:
556 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
557 case vmIntrinsics::_getShortVolatile:
558 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
559 case vmIntrinsics::_getCharVolatile:
560 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
561 case vmIntrinsics::_getIntVolatile:
562 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
563 case vmIntrinsics::_getLongVolatile:
564 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
565 case vmIntrinsics::_getFloatVolatile:
566 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
567 case vmIntrinsics::_getDoubleVolatile:
568 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
569
570 case vmIntrinsics::_putObjectVolatile:
571 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
572 case vmIntrinsics::_putBooleanVolatile:
573 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
574 case vmIntrinsics::_putByteVolatile:
575 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
576 case vmIntrinsics::_putShortVolatile:
577 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
578 case vmIntrinsics::_putCharVolatile:
579 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
580 case vmIntrinsics::_putIntVolatile:
581 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
582 case vmIntrinsics::_putLongVolatile:
583 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
584 case vmIntrinsics::_putFloatVolatile:
585 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
586 case vmIntrinsics::_putDoubleVolatile:
587 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
588
589 case vmIntrinsics::_prefetchRead:
590 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
591 case vmIntrinsics::_prefetchWrite:
592 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
593 case vmIntrinsics::_prefetchReadStatic:
594 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
595 case vmIntrinsics::_prefetchWriteStatic:
596 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
597
598 case vmIntrinsics::_compareAndSwapObject:
599 return inline_unsafe_CAS(T_OBJECT);
600 case vmIntrinsics::_compareAndSwapInt:
601 return inline_unsafe_CAS(T_INT);
602 case vmIntrinsics::_compareAndSwapLong:
603 return inline_unsafe_CAS(T_LONG);
604
605 case vmIntrinsics::_putOrderedObject:
606 return inline_unsafe_ordered_store(T_OBJECT);
607 case vmIntrinsics::_putOrderedInt:
608 return inline_unsafe_ordered_store(T_INT);
609 case vmIntrinsics::_putOrderedLong:
610 return inline_unsafe_ordered_store(T_LONG);
611
612 case vmIntrinsics::_currentThread:
613 return inline_native_currentThread();
614 case vmIntrinsics::_isInterrupted:
615 return inline_native_isInterrupted();
616
617 case vmIntrinsics::_currentTimeMillis:
618 return inline_native_time_funcs(false);
619 case vmIntrinsics::_nanoTime:
620 return inline_native_time_funcs(true);
621 case vmIntrinsics::_allocateInstance:
622 return inline_unsafe_allocate();
623 case vmIntrinsics::_copyMemory:
624 return inline_unsafe_copyMemory();
625 case vmIntrinsics::_newArray:
626 return inline_native_newArray();
627 case vmIntrinsics::_getLength:
628 return inline_native_getLength();
629 case vmIntrinsics::_copyOf:
630 return inline_array_copyOf(false);
631 case vmIntrinsics::_copyOfRange:
632 return inline_array_copyOf(true);
633 case vmIntrinsics::_equalsC:
634 return inline_array_equals();
635 case vmIntrinsics::_clone:
636 return inline_native_clone(intrinsic()->is_virtual());
637
638 case vmIntrinsics::_isAssignableFrom:
639 return inline_native_subtype_check();
640
641 case vmIntrinsics::_isInstance:
642 case vmIntrinsics::_getModifiers:
643 case vmIntrinsics::_isInterface:
644 case vmIntrinsics::_isArray:
645 case vmIntrinsics::_isPrimitive:
646 case vmIntrinsics::_getSuperclass:
647 case vmIntrinsics::_getComponentType:
648 case vmIntrinsics::_getClassAccessFlags:
649 return inline_native_Class_query(intrinsic_id());
650
651 case vmIntrinsics::_floatToRawIntBits:
652 case vmIntrinsics::_floatToIntBits:
653 case vmIntrinsics::_intBitsToFloat:
654 case vmIntrinsics::_doubleToRawLongBits:
655 case vmIntrinsics::_doubleToLongBits:
656 case vmIntrinsics::_longBitsToDouble:
657 return inline_fp_conversions(intrinsic_id());
658
659 case vmIntrinsics::_numberOfLeadingZeros_i:
660 case vmIntrinsics::_numberOfLeadingZeros_l:
661 return inline_numberOfLeadingZeros(intrinsic_id());
662
663 case vmIntrinsics::_numberOfTrailingZeros_i:
664 case vmIntrinsics::_numberOfTrailingZeros_l:
665 return inline_numberOfTrailingZeros(intrinsic_id());
666
667 case vmIntrinsics::_bitCount_i:
668 case vmIntrinsics::_bitCount_l:
669 return inline_bitCount(intrinsic_id());
670
671 case vmIntrinsics::_reverseBytes_i:
672 case vmIntrinsics::_reverseBytes_l:
673 case vmIntrinsics::_reverseBytes_s:
674 case vmIntrinsics::_reverseBytes_c:
675 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
676
677 case vmIntrinsics::_get_AtomicLong:
678 return inline_native_AtomicLong_get();
679 case vmIntrinsics::_attemptUpdate:
680 return inline_native_AtomicLong_attemptUpdate();
681
682 case vmIntrinsics::_getCallerClass:
683 return inline_native_Reflection_getCallerClass();
684
685 case vmIntrinsics::_Reference_get:
686 return inline_reference_get();
687
688 default:
689 // If you get here, it may be that someone has added a new intrinsic
690 // to the list in vmSymbols.hpp without implementing it here.
691 #ifndef PRODUCT
692 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
693 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
694 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
695 }
696 #endif
697 return false;
698 }
699 }
700
701 //------------------------------push_result------------------------------
702 // Helper function for finishing intrinsics.
703 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
704 record_for_igvn(region);
705 set_control(_gvn.transform(region));
706 BasicType value_type = value->type()->basic_type();
707 push_node(value_type, _gvn.transform(value));
708 }
709
710 //------------------------------generate_guard---------------------------
711 // Helper function for generating guarded fast-slow graph structures.
712 // The given 'test', if true, guards a slow path. If the test fails
713 // then a fast path can be taken. (We generally hope it fails.)
714 // In all cases, GraphKit::control() is updated to the fast path.
715 // The returned value represents the control for the slow path.
716 // The return value is never 'top'; it is either a valid control
717 // or NULL if it is obvious that the slow path can never be taken.
718 // Also, if region and the slow control are not NULL, the slow edge
719 // is appended to the region.
720 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
721 if (stopped()) {
722 // Already short circuited.
723 return NULL;
724 }
725
726 // Build an if node and its projections.
727 // If test is true we take the slow path, which we assume is uncommon.
728 if (_gvn.type(test) == TypeInt::ZERO) {
729 // The slow branch is never taken. No need to build this guard.
730 return NULL;
731 }
732
733 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
734
735 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
736 if (if_slow == top()) {
737 // The slow branch is never taken. No need to build this guard.
738 return NULL;
739 }
740
741 if (region != NULL)
742 region->add_req(if_slow);
743
744 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
745 set_control(if_fast);
746
747 return if_slow;
748 }
749
750 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
751 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
752 }
753 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
754 return generate_guard(test, region, PROB_FAIR);
755 }
756
757 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
758 Node* *pos_index) {
759 if (stopped())
760 return NULL; // already stopped
761 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
762 return NULL; // index is already adequately typed
763 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
764 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
765 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
766 if (is_neg != NULL && pos_index != NULL) {
767 // Emulate effect of Parse::adjust_map_after_if.
768 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
769 ccast->set_req(0, control());
770 (*pos_index) = _gvn.transform(ccast);
771 }
772 return is_neg;
773 }
774
775 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
776 Node* *pos_index) {
777 if (stopped())
778 return NULL; // already stopped
779 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
780 return NULL; // index is already adequately typed
781 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
782 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
783 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
784 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
785 if (is_notp != NULL && pos_index != NULL) {
786 // Emulate effect of Parse::adjust_map_after_if.
787 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
788 ccast->set_req(0, control());
789 (*pos_index) = _gvn.transform(ccast);
790 }
791 return is_notp;
792 }
793
794 // Make sure that 'position' is a valid limit index, in [0..length].
795 // There are two equivalent plans for checking this:
796 // A. (offset + copyLength) unsigned<= arrayLength
797 // B. offset <= (arrayLength - copyLength)
798 // We require that all of the values above, except for the sum and
799 // difference, are already known to be non-negative.
800 // Plan A is robust in the face of overflow, if offset and copyLength
801 // are both hugely positive.
802 //
803 // Plan B is less direct and intuitive, but it does not overflow at
804 // all, since the difference of two non-negatives is always
805 // representable. Whenever Java methods must perform the equivalent
806 // check they generally use Plan B instead of Plan A.
807 // For the moment we use Plan A.
808 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
809 Node* subseq_length,
810 Node* array_length,
811 RegionNode* region) {
812 if (stopped())
813 return NULL; // already stopped
814 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
815 if (zero_offset && subseq_length->eqv_uncast(array_length))
816 return NULL; // common case of whole-array copy
817 Node* last = subseq_length;
818 if (!zero_offset) // last += offset
819 last = _gvn.transform( new (C, 3) AddINode(last, offset));
820 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
821 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
822 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
823 return is_over;
824 }
825
826
827 //--------------------------generate_current_thread--------------------
828 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
829 ciKlass* thread_klass = env()->Thread_klass();
830 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
831 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
832 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
833 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
834 tls_output = thread;
835 return threadObj;
836 }
837
838
839 //------------------------------make_string_method_node------------------------
840 // Helper method for String intrinsic finctions.
841 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
842 const int value_offset = java_lang_String::value_offset_in_bytes();
843 const int count_offset = java_lang_String::count_offset_in_bytes();
844 const int offset_offset = java_lang_String::offset_offset_in_bytes();
845
846 Node* no_ctrl = NULL;
847
848 ciInstanceKlass* klass = env()->String_klass();
849 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
850
851 const TypeAryPtr* value_type =
852 TypeAryPtr::make(TypePtr::NotNull,
853 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
854 ciTypeArrayKlass::make(T_CHAR), true, 0);
855
856 // Get start addr of string and substring
857 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
858 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
859 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
860 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
861 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
862
863 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
864 Node* str2_value = make_load(no_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
865 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
866 Node* str2_offset = make_load(no_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
867 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
868
869 Node* result = NULL;
870 switch (opcode) {
871 case Op_StrIndexOf:
872 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
873 str1_start, cnt1, str2_start, cnt2);
874 break;
875 case Op_StrComp:
876 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
877 str1_start, cnt1, str2_start, cnt2);
878 break;
879 case Op_StrEquals:
880 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
881 str1_start, str2_start, cnt1);
882 break;
883 default:
884 ShouldNotReachHere();
885 return NULL;
886 }
887
888 // All these intrinsics have checks.
889 C->set_has_split_ifs(true); // Has chance for split-if optimization
890
891 return _gvn.transform(result);
892 }
893
894 //------------------------------inline_string_compareTo------------------------
895 bool LibraryCallKit::inline_string_compareTo() {
896
897 if (!Matcher::has_match_rule(Op_StrComp)) return false;
898
899 const int value_offset = java_lang_String::value_offset_in_bytes();
900 const int count_offset = java_lang_String::count_offset_in_bytes();
901 const int offset_offset = java_lang_String::offset_offset_in_bytes();
902
903 _sp += 2;
904 Node *argument = pop(); // pop non-receiver first: it was pushed second
905 Node *receiver = pop();
906
907 // Null check on self without removing any arguments. The argument
908 // null check technically happens in the wrong place, which can lead to
909 // invalid stack traces when string compare is inlined into a method
910 // which handles NullPointerExceptions.
911 _sp += 2;
912 receiver = do_null_check(receiver, T_OBJECT);
913 argument = do_null_check(argument, T_OBJECT);
914 _sp -= 2;
915 if (stopped()) {
916 return true;
917 }
918
919 ciInstanceKlass* klass = env()->String_klass();
920 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
921 Node* no_ctrl = NULL;
922
923 // Get counts for string and argument
924 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
925 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
926
927 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
928 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
929
930 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
931 push(compare);
932 return true;
933 }
934
935 //------------------------------inline_string_equals------------------------
936 bool LibraryCallKit::inline_string_equals() {
937
938 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
939
940 const int value_offset = java_lang_String::value_offset_in_bytes();
941 const int count_offset = java_lang_String::count_offset_in_bytes();
942 const int offset_offset = java_lang_String::offset_offset_in_bytes();
943
944 int nargs = 2;
945 _sp += nargs;
946 Node* argument = pop(); // pop non-receiver first: it was pushed second
947 Node* receiver = pop();
948
949 // Null check on self without removing any arguments. The argument
950 // null check technically happens in the wrong place, which can lead to
951 // invalid stack traces when string compare is inlined into a method
952 // which handles NullPointerExceptions.
953 _sp += nargs;
954 receiver = do_null_check(receiver, T_OBJECT);
955 //should not do null check for argument for String.equals(), because spec
956 //allows to specify NULL as argument.
957 _sp -= nargs;
958
959 if (stopped()) {
960 return true;
961 }
962
963 // paths (plus control) merge
964 RegionNode* region = new (C, 5) RegionNode(5);
965 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
966
967 // does source == target string?
968 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
969 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
970
971 Node* if_eq = generate_slow_guard(bol, NULL);
972 if (if_eq != NULL) {
973 // receiver == argument
974 phi->init_req(2, intcon(1));
975 region->init_req(2, if_eq);
976 }
977
978 // get String klass for instanceOf
979 ciInstanceKlass* klass = env()->String_klass();
980
981 if (!stopped()) {
982 _sp += nargs; // gen_instanceof might do an uncommon trap
983 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
984 _sp -= nargs;
985 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
986 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
987
988 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
989 //instanceOf == true, fallthrough
990
991 if (inst_false != NULL) {
992 phi->init_req(3, intcon(0));
993 region->init_req(3, inst_false);
994 }
995 }
996
997 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
998
999 Node* no_ctrl = NULL;
1000 Node* receiver_cnt;
1001 Node* argument_cnt;
1002
1003 if (!stopped()) {
1004 // Properly cast the argument to String
1005 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1006 // This path is taken only when argument's type is String:NotNull.
1007 argument = cast_not_null(argument, false);
1008
1009 // Get counts for string and argument
1010 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
1011 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1012
1013 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
1014 argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1015
1016 // Check for receiver count != argument count
1017 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1018 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1019 Node* if_ne = generate_slow_guard(bol, NULL);
1020 if (if_ne != NULL) {
1021 phi->init_req(4, intcon(0));
1022 region->init_req(4, if_ne);
1023 }
1024 }
1025
1026 // Check for count == 0 is done by mach node StrEquals.
1027
1028 if (!stopped()) {
1029 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1030 phi->init_req(1, equals);
1031 region->init_req(1, control());
1032 }
1033
1034 // post merge
1035 set_control(_gvn.transform(region));
1036 record_for_igvn(region);
1037
1038 push(_gvn.transform(phi));
1039
1040 return true;
1041 }
1042
1043 //------------------------------inline_array_equals----------------------------
1044 bool LibraryCallKit::inline_array_equals() {
1045
1046 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1047
1048 _sp += 2;
1049 Node *argument2 = pop();
1050 Node *argument1 = pop();
1051
1052 Node* equals =
1053 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1054 argument1, argument2) );
1055 push(equals);
1056 return true;
1057 }
1058
1059 // Java version of String.indexOf(constant string)
1060 // class StringDecl {
1061 // StringDecl(char[] ca) {
1062 // offset = 0;
1063 // count = ca.length;
1064 // value = ca;
1065 // }
1066 // int offset;
1067 // int count;
1068 // char[] value;
1069 // }
1070 //
1071 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1072 // int targetOffset, int cache_i, int md2) {
1073 // int cache = cache_i;
1074 // int sourceOffset = string_object.offset;
1075 // int sourceCount = string_object.count;
1076 // int targetCount = target_object.length;
1077 //
1078 // int targetCountLess1 = targetCount - 1;
1079 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1080 //
1081 // char[] source = string_object.value;
1082 // char[] target = target_object;
1083 // int lastChar = target[targetCountLess1];
1084 //
1085 // outer_loop:
1086 // for (int i = sourceOffset; i < sourceEnd; ) {
1087 // int src = source[i + targetCountLess1];
1088 // if (src == lastChar) {
1089 // // With random strings and a 4-character alphabet,
1090 // // reverse matching at this point sets up 0.8% fewer
1091 // // frames, but (paradoxically) makes 0.3% more probes.
1092 // // Since those probes are nearer the lastChar probe,
1093 // // there is may be a net D$ win with reverse matching.
1094 // // But, reversing loop inhibits unroll of inner loop
1095 // // for unknown reason. So, does running outer loop from
1096 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1097 // for (int j = 0; j < targetCountLess1; j++) {
1098 // if (target[targetOffset + j] != source[i+j]) {
1099 // if ((cache & (1 << source[i+j])) == 0) {
1100 // if (md2 < j+1) {
1101 // i += j+1;
1102 // continue outer_loop;
1103 // }
1104 // }
1105 // i += md2;
1106 // continue outer_loop;
1107 // }
1108 // }
1109 // return i - sourceOffset;
1110 // }
1111 // if ((cache & (1 << src)) == 0) {
1112 // i += targetCountLess1;
1113 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1114 // i++;
1115 // }
1116 // return -1;
1117 // }
1118
1119 //------------------------------string_indexOf------------------------
1120 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1121 jint cache_i, jint md2_i) {
1122
1123 Node* no_ctrl = NULL;
1124 float likely = PROB_LIKELY(0.9);
1125 float unlikely = PROB_UNLIKELY(0.9);
1126
1127 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1128
1129 const int value_offset = java_lang_String::value_offset_in_bytes();
1130 const int count_offset = java_lang_String::count_offset_in_bytes();
1131 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1132
1133 ciInstanceKlass* klass = env()->String_klass();
1134 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1135 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1136
1137 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1138 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1139 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1140 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1141 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1142 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1143
1144 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1145 jint target_length = target_array->length();
1146 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1147 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1148
1149 IdealKit kit(this, false, true);
1150 #define __ kit.
1151 Node* zero = __ ConI(0);
1152 Node* one = __ ConI(1);
1153 Node* cache = __ ConI(cache_i);
1154 Node* md2 = __ ConI(md2_i);
1155 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1156 Node* targetCount = __ ConI(target_length);
1157 Node* targetCountLess1 = __ ConI(target_length - 1);
1158 Node* targetOffset = __ ConI(targetOffset_i);
1159 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1160
1161 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1162 Node* outer_loop = __ make_label(2 /* goto */);
1163 Node* return_ = __ make_label(1);
1164
1165 __ set(rtn,__ ConI(-1));
1166 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1167 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1168 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1169 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1170 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1171 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1172 Node* tpj = __ AddI(targetOffset, __ value(j));
1173 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1174 Node* ipj = __ AddI(__ value(i), __ value(j));
1175 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1176 __ if_then(targ, BoolTest::ne, src2); {
1177 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1178 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1179 __ increment(i, __ AddI(__ value(j), one));
1180 __ goto_(outer_loop);
1181 } __ end_if(); __ dead(j);
1182 }__ end_if(); __ dead(j);
1183 __ increment(i, md2);
1184 __ goto_(outer_loop);
1185 }__ end_if();
1186 __ increment(j, one);
1187 }__ end_loop(); __ dead(j);
1188 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1189 __ goto_(return_);
1190 }__ end_if();
1191 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1192 __ increment(i, targetCountLess1);
1193 }__ end_if();
1194 __ increment(i, one);
1195 __ bind(outer_loop);
1196 }__ end_loop(); __ dead(i);
1197 __ bind(return_);
1198
1199 // Final sync IdealKit and GraphKit.
1200 final_sync(kit);
1201 Node* result = __ value(rtn);
1202 #undef __
1203 C->set_has_loops(true);
1204 return result;
1205 }
1206
1207 //------------------------------inline_string_indexOf------------------------
1208 bool LibraryCallKit::inline_string_indexOf() {
1209
1210 const int value_offset = java_lang_String::value_offset_in_bytes();
1211 const int count_offset = java_lang_String::count_offset_in_bytes();
1212 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1213
1214 _sp += 2;
1215 Node *argument = pop(); // pop non-receiver first: it was pushed second
1216 Node *receiver = pop();
1217
1218 Node* result;
1219 // Disable the use of pcmpestri until it can be guaranteed that
1220 // the load doesn't cross into the uncommited space.
1221 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1222 UseSSE42Intrinsics) {
1223 // Generate SSE4.2 version of indexOf
1224 // We currently only have match rules that use SSE4.2
1225
1226 // Null check on self without removing any arguments. The argument
1227 // null check technically happens in the wrong place, which can lead to
1228 // invalid stack traces when string compare is inlined into a method
1229 // which handles NullPointerExceptions.
1230 _sp += 2;
1231 receiver = do_null_check(receiver, T_OBJECT);
1232 argument = do_null_check(argument, T_OBJECT);
1233 _sp -= 2;
1234
1235 if (stopped()) {
1236 return true;
1237 }
1238
1239 ciInstanceKlass* str_klass = env()->String_klass();
1240 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1241
1242 // Make the merge point
1243 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1244 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1245 Node* no_ctrl = NULL;
1246
1247 // Get counts for string and substr
1248 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1249 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1250
1251 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1252 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1253
1254 // Check for substr count > string count
1255 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1256 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1257 Node* if_gt = generate_slow_guard(bol, NULL);
1258 if (if_gt != NULL) {
1259 result_phi->init_req(2, intcon(-1));
1260 result_rgn->init_req(2, if_gt);
1261 }
1262
1263 if (!stopped()) {
1264 // Check for substr count == 0
1265 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1266 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1267 Node* if_zero = generate_slow_guard(bol, NULL);
1268 if (if_zero != NULL) {
1269 result_phi->init_req(3, intcon(0));
1270 result_rgn->init_req(3, if_zero);
1271 }
1272 }
1273
1274 if (!stopped()) {
1275 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1276 result_phi->init_req(1, result);
1277 result_rgn->init_req(1, control());
1278 }
1279 set_control(_gvn.transform(result_rgn));
1280 record_for_igvn(result_rgn);
1281 result = _gvn.transform(result_phi);
1282
1283 } else { // Use LibraryCallKit::string_indexOf
1284 // don't intrinsify if argument isn't a constant string.
1285 if (!argument->is_Con()) {
1286 return false;
1287 }
1288 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1289 if (str_type == NULL) {
1290 return false;
1291 }
1292 ciInstanceKlass* klass = env()->String_klass();
1293 ciObject* str_const = str_type->const_oop();
1294 if (str_const == NULL || str_const->klass() != klass) {
1295 return false;
1296 }
1297 ciInstance* str = str_const->as_instance();
1298 assert(str != NULL, "must be instance");
1299
1300 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1301 int o = str->field_value_by_offset(offset_offset).as_int();
1302 int c = str->field_value_by_offset(count_offset).as_int();
1303 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1304
1305 // constant strings have no offset and count == length which
1306 // simplifies the resulting code somewhat so lets optimize for that.
1307 if (o != 0 || c != pat->length()) {
1308 return false;
1309 }
1310
1311 // Null check on self without removing any arguments. The argument
1312 // null check technically happens in the wrong place, which can lead to
1313 // invalid stack traces when string compare is inlined into a method
1314 // which handles NullPointerExceptions.
1315 _sp += 2;
1316 receiver = do_null_check(receiver, T_OBJECT);
1317 // No null check on the argument is needed since it's a constant String oop.
1318 _sp -= 2;
1319 if (stopped()) {
1320 return true;
1321 }
1322
1323 // The null string as a pattern always returns 0 (match at beginning of string)
1324 if (c == 0) {
1325 push(intcon(0));
1326 return true;
1327 }
1328
1329 // Generate default indexOf
1330 jchar lastChar = pat->char_at(o + (c - 1));
1331 int cache = 0;
1332 int i;
1333 for (i = 0; i < c - 1; i++) {
1334 assert(i < pat->length(), "out of range");
1335 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1336 }
1337
1338 int md2 = c;
1339 for (i = 0; i < c - 1; i++) {
1340 assert(i < pat->length(), "out of range");
1341 if (pat->char_at(o + i) == lastChar) {
1342 md2 = (c - 1) - i;
1343 }
1344 }
1345
1346 result = string_indexOf(receiver, pat, o, cache, md2);
1347 }
1348
1349 push(result);
1350 return true;
1351 }
1352
1353 //--------------------------pop_math_arg--------------------------------
1354 // Pop a double argument to a math function from the stack
1355 // rounding it if necessary.
1356 Node * LibraryCallKit::pop_math_arg() {
1357 Node *arg = pop_pair();
1358 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1359 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1360 return arg;
1361 }
1362
1363 //------------------------------inline_trig----------------------------------
1364 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1365 // argument reduction which will turn into a fast/slow diamond.
1366 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1367 _sp += arg_size(); // restore stack pointer
1368 Node* arg = pop_math_arg();
1369 Node* trig = NULL;
1370
1371 switch (id) {
1372 case vmIntrinsics::_dsin:
1373 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1374 break;
1375 case vmIntrinsics::_dcos:
1376 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1377 break;
1378 case vmIntrinsics::_dtan:
1379 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1380 break;
1381 default:
1382 assert(false, "bad intrinsic was passed in");
1383 return false;
1384 }
1385
1386 // Rounding required? Check for argument reduction!
1387 if( Matcher::strict_fp_requires_explicit_rounding ) {
1388
1389 static const double pi_4 = 0.7853981633974483;
1390 static const double neg_pi_4 = -0.7853981633974483;
1391 // pi/2 in 80-bit extended precision
1392 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1393 // -pi/2 in 80-bit extended precision
1394 // 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};
1395 // Cutoff value for using this argument reduction technique
1396 //static const double pi_2_minus_epsilon = 1.564660403643354;
1397 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1398
1399 // Pseudocode for sin:
1400 // if (x <= Math.PI / 4.0) {
1401 // if (x >= -Math.PI / 4.0) return fsin(x);
1402 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1403 // } else {
1404 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1405 // }
1406 // return StrictMath.sin(x);
1407
1408 // Pseudocode for cos:
1409 // if (x <= Math.PI / 4.0) {
1410 // if (x >= -Math.PI / 4.0) return fcos(x);
1411 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1412 // } else {
1413 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1414 // }
1415 // return StrictMath.cos(x);
1416
1417 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1418 // requires a special machine instruction to load it. Instead we'll try
1419 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1420 // probably do the math inside the SIN encoding.
1421
1422 // Make the merge point
1423 RegionNode *r = new (C, 3) RegionNode(3);
1424 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1425
1426 // Flatten arg so we need only 1 test
1427 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1428 // Node for PI/4 constant
1429 Node *pi4 = makecon(TypeD::make(pi_4));
1430 // Check PI/4 : abs(arg)
1431 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1432 // Check: If PI/4 < abs(arg) then go slow
1433 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1434 // Branch either way
1435 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1436 set_control(opt_iff(r,iff));
1437
1438 // Set fast path result
1439 phi->init_req(2,trig);
1440
1441 // Slow path - non-blocking leaf call
1442 Node* call = NULL;
1443 switch (id) {
1444 case vmIntrinsics::_dsin:
1445 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1446 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1447 "Sin", NULL, arg, top());
1448 break;
1449 case vmIntrinsics::_dcos:
1450 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1451 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1452 "Cos", NULL, arg, top());
1453 break;
1454 case vmIntrinsics::_dtan:
1455 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1456 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1457 "Tan", NULL, arg, top());
1458 break;
1459 }
1460 assert(control()->in(0) == call, "");
1461 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1462 r->init_req(1,control());
1463 phi->init_req(1,slow_result);
1464
1465 // Post-merge
1466 set_control(_gvn.transform(r));
1467 record_for_igvn(r);
1468 trig = _gvn.transform(phi);
1469
1470 C->set_has_split_ifs(true); // Has chance for split-if optimization
1471 }
1472 // Push result back on JVM stack
1473 push_pair(trig);
1474 return true;
1475 }
1476
1477 //------------------------------inline_sqrt-------------------------------------
1478 // Inline square root instruction, if possible.
1479 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1480 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1481 _sp += arg_size(); // restore stack pointer
1482 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1483 return true;
1484 }
1485
1486 //------------------------------inline_abs-------------------------------------
1487 // Inline absolute value instruction, if possible.
1488 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1489 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1490 _sp += arg_size(); // restore stack pointer
1491 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1492 return true;
1493 }
1494
1495 //------------------------------inline_exp-------------------------------------
1496 // Inline exp instructions, if possible. The Intel hardware only misses
1497 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1498 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1499 assert(id == vmIntrinsics::_dexp, "Not exp");
1500
1501 // If this inlining ever returned NaN in the past, we do not intrinsify it
1502 // every again. NaN results requires StrictMath.exp handling.
1503 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1504
1505 // Do not intrinsify on older platforms which lack cmove.
1506 if (ConditionalMoveLimit == 0) return false;
1507
1508 _sp += arg_size(); // restore stack pointer
1509 Node *x = pop_math_arg();
1510 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1511
1512 //-------------------
1513 //result=(result.isNaN())? StrictMath::exp():result;
1514 // Check: If isNaN() by checking result!=result? then go to Strict Math
1515 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1516 // Build the boolean node
1517 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1518
1519 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1520 // End the current control-flow path
1521 push_pair(x);
1522 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1523 // to handle. Recompile without intrinsifying Math.exp
1524 uncommon_trap(Deoptimization::Reason_intrinsic,
1525 Deoptimization::Action_make_not_entrant);
1526 }
1527
1528 C->set_has_split_ifs(true); // Has chance for split-if optimization
1529
1530 push_pair(result);
1531
1532 return true;
1533 }
1534
1535 //------------------------------inline_pow-------------------------------------
1536 // Inline power instructions, if possible.
1537 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1538 assert(id == vmIntrinsics::_dpow, "Not pow");
1539
1540 // If this inlining ever returned NaN in the past, we do not intrinsify it
1541 // every again. NaN results requires StrictMath.pow handling.
1542 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1543
1544 // Do not intrinsify on older platforms which lack cmove.
1545 if (ConditionalMoveLimit == 0) return false;
1546
1547 // Pseudocode for pow
1548 // if (x <= 0.0) {
1549 // if ((double)((int)y)==y) { // if y is int
1550 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1551 // } else {
1552 // result = NaN;
1553 // }
1554 // } else {
1555 // result = DPow(x,y);
1556 // }
1557 // if (result != result)? {
1558 // uncommon_trap();
1559 // }
1560 // return result;
1561
1562 _sp += arg_size(); // restore stack pointer
1563 Node* y = pop_math_arg();
1564 Node* x = pop_math_arg();
1565
1566 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1567
1568 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1569 // inside of something) then skip the fancy tests and just check for
1570 // NaN result.
1571 Node *result = NULL;
1572 if( jvms()->depth() >= 1 ) {
1573 result = fast_result;
1574 } else {
1575
1576 // Set the merge point for If node with condition of (x <= 0.0)
1577 // There are four possible paths to region node and phi node
1578 RegionNode *r = new (C, 4) RegionNode(4);
1579 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1580
1581 // Build the first if node: if (x <= 0.0)
1582 // Node for 0 constant
1583 Node *zeronode = makecon(TypeD::ZERO);
1584 // Check x:0
1585 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1586 // Check: If (x<=0) then go complex path
1587 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1588 // Branch either way
1589 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1590 Node *opt_test = _gvn.transform(if1);
1591 //assert( opt_test->is_If(), "Expect an IfNode");
1592 IfNode *opt_if1 = (IfNode*)opt_test;
1593 // Fast path taken; set region slot 3
1594 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1595 r->init_req(3,fast_taken); // Capture fast-control
1596
1597 // Fast path not-taken, i.e. slow path
1598 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1599
1600 // Set fast path result
1601 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1602 phi->init_req(3, fast_result);
1603
1604 // Complex path
1605 // Build the second if node (if y is int)
1606 // Node for (int)y
1607 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1608 // Node for (double)((int) y)
1609 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1610 // Check (double)((int) y) : y
1611 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1612 // Check if (y isn't int) then go to slow path
1613
1614 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1615 // Branch either way
1616 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1617 Node *slow_path = opt_iff(r,if2); // Set region path 2
1618
1619 // Calculate DPow(abs(x), y)*(1 & (int)y)
1620 // Node for constant 1
1621 Node *conone = intcon(1);
1622 // 1& (int)y
1623 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1624 // zero node
1625 Node *conzero = intcon(0);
1626 // Check (1&(int)y)==0?
1627 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1628 // Check if (1&(int)y)!=0?, if so the result is negative
1629 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1630 // abs(x)
1631 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1632 // abs(x)^y
1633 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1634 // -abs(x)^y
1635 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1636 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1637 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1638 // Set complex path fast result
1639 phi->init_req(2, signresult);
1640
1641 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1642 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1643 r->init_req(1,slow_path);
1644 phi->init_req(1,slow_result);
1645
1646 // Post merge
1647 set_control(_gvn.transform(r));
1648 record_for_igvn(r);
1649 result=_gvn.transform(phi);
1650 }
1651
1652 //-------------------
1653 //result=(result.isNaN())? uncommon_trap():result;
1654 // Check: If isNaN() by checking result!=result? then go to Strict Math
1655 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1656 // Build the boolean node
1657 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1658
1659 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1660 // End the current control-flow path
1661 push_pair(x);
1662 push_pair(y);
1663 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1664 // to handle. Recompile without intrinsifying Math.pow.
1665 uncommon_trap(Deoptimization::Reason_intrinsic,
1666 Deoptimization::Action_make_not_entrant);
1667 }
1668
1669 C->set_has_split_ifs(true); // Has chance for split-if optimization
1670
1671 push_pair(result);
1672
1673 return true;
1674 }
1675
1676 //------------------------------inline_trans-------------------------------------
1677 // Inline transcendental instructions, if possible. The Intel hardware gets
1678 // these right, no funny corner cases missed.
1679 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1680 _sp += arg_size(); // restore stack pointer
1681 Node* arg = pop_math_arg();
1682 Node* trans = NULL;
1683
1684 switch (id) {
1685 case vmIntrinsics::_dlog:
1686 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1687 break;
1688 case vmIntrinsics::_dlog10:
1689 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1690 break;
1691 default:
1692 assert(false, "bad intrinsic was passed in");
1693 return false;
1694 }
1695
1696 // Push result back on JVM stack
1697 push_pair(trans);
1698 return true;
1699 }
1700
1701 //------------------------------runtime_math-----------------------------
1702 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1703 Node* a = NULL;
1704 Node* b = NULL;
1705
1706 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1707 "must be (DD)D or (D)D type");
1708
1709 // Inputs
1710 _sp += arg_size(); // restore stack pointer
1711 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1712 b = pop_math_arg();
1713 }
1714 a = pop_math_arg();
1715
1716 const TypePtr* no_memory_effects = NULL;
1717 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1718 no_memory_effects,
1719 a, top(), b, b ? top() : NULL);
1720 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1721 #ifdef ASSERT
1722 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1723 assert(value_top == top(), "second value must be top");
1724 #endif
1725
1726 push_pair(value);
1727 return true;
1728 }
1729
1730 //------------------------------inline_math_native-----------------------------
1731 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1732 switch (id) {
1733 // These intrinsics are not properly supported on all hardware
1734 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1735 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1736 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1737 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1738 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1739 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1740
1741 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1742 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1743 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1744 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1745
1746 // These intrinsics are supported on all hardware
1747 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1748 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1749
1750 // These intrinsics don't work on X86. The ad implementation doesn't
1751 // handle NaN's properly. Instead of returning infinity, the ad
1752 // implementation returns a NaN on overflow. See bug: 6304089
1753 // Once the ad implementations are fixed, change the code below
1754 // to match the intrinsics above
1755
1756 case vmIntrinsics::_dexp: return
1757 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1758 case vmIntrinsics::_dpow: return
1759 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1760
1761 // These intrinsics are not yet correctly implemented
1762 case vmIntrinsics::_datan2:
1763 return false;
1764
1765 default:
1766 ShouldNotReachHere();
1767 return false;
1768 }
1769 }
1770
1771 static bool is_simple_name(Node* n) {
1772 return (n->req() == 1 // constant
1773 || (n->is_Type() && n->as_Type()->type()->singleton())
1774 || n->is_Proj() // parameter or return value
1775 || n->is_Phi() // local of some sort
1776 );
1777 }
1778
1779 //----------------------------inline_min_max-----------------------------------
1780 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1781 push(generate_min_max(id, argument(0), argument(1)));
1782
1783 return true;
1784 }
1785
1786 Node*
1787 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1788 // These are the candidate return value:
1789 Node* xvalue = x0;
1790 Node* yvalue = y0;
1791
1792 if (xvalue == yvalue) {
1793 return xvalue;
1794 }
1795
1796 bool want_max = (id == vmIntrinsics::_max);
1797
1798 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1799 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1800 if (txvalue == NULL || tyvalue == NULL) return top();
1801 // This is not really necessary, but it is consistent with a
1802 // hypothetical MaxINode::Value method:
1803 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1804
1805 // %%% This folding logic should (ideally) be in a different place.
1806 // Some should be inside IfNode, and there to be a more reliable
1807 // transformation of ?: style patterns into cmoves. We also want
1808 // more powerful optimizations around cmove and min/max.
1809
1810 // Try to find a dominating comparison of these guys.
1811 // It can simplify the index computation for Arrays.copyOf
1812 // and similar uses of System.arraycopy.
1813 // First, compute the normalized version of CmpI(x, y).
1814 int cmp_op = Op_CmpI;
1815 Node* xkey = xvalue;
1816 Node* ykey = yvalue;
1817 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1818 if (ideal_cmpxy->is_Cmp()) {
1819 // E.g., if we have CmpI(length - offset, count),
1820 // it might idealize to CmpI(length, count + offset)
1821 cmp_op = ideal_cmpxy->Opcode();
1822 xkey = ideal_cmpxy->in(1);
1823 ykey = ideal_cmpxy->in(2);
1824 }
1825
1826 // Start by locating any relevant comparisons.
1827 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1828 Node* cmpxy = NULL;
1829 Node* cmpyx = NULL;
1830 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1831 Node* cmp = start_from->fast_out(k);
1832 if (cmp->outcnt() > 0 && // must have prior uses
1833 cmp->in(0) == NULL && // must be context-independent
1834 cmp->Opcode() == cmp_op) { // right kind of compare
1835 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1836 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1837 }
1838 }
1839
1840 const int NCMPS = 2;
1841 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1842 int cmpn;
1843 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1844 if (cmps[cmpn] != NULL) break; // find a result
1845 }
1846 if (cmpn < NCMPS) {
1847 // Look for a dominating test that tells us the min and max.
1848 int depth = 0; // Limit search depth for speed
1849 Node* dom = control();
1850 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1851 if (++depth >= 100) break;
1852 Node* ifproj = dom;
1853 if (!ifproj->is_Proj()) continue;
1854 Node* iff = ifproj->in(0);
1855 if (!iff->is_If()) continue;
1856 Node* bol = iff->in(1);
1857 if (!bol->is_Bool()) continue;
1858 Node* cmp = bol->in(1);
1859 if (cmp == NULL) continue;
1860 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1861 if (cmps[cmpn] == cmp) break;
1862 if (cmpn == NCMPS) continue;
1863 BoolTest::mask btest = bol->as_Bool()->_test._test;
1864 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1865 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1866 // At this point, we know that 'x btest y' is true.
1867 switch (btest) {
1868 case BoolTest::eq:
1869 // They are proven equal, so we can collapse the min/max.
1870 // Either value is the answer. Choose the simpler.
1871 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1872 return yvalue;
1873 return xvalue;
1874 case BoolTest::lt: // x < y
1875 case BoolTest::le: // x <= y
1876 return (want_max ? yvalue : xvalue);
1877 case BoolTest::gt: // x > y
1878 case BoolTest::ge: // x >= y
1879 return (want_max ? xvalue : yvalue);
1880 }
1881 }
1882 }
1883
1884 // We failed to find a dominating test.
1885 // Let's pick a test that might GVN with prior tests.
1886 Node* best_bol = NULL;
1887 BoolTest::mask best_btest = BoolTest::illegal;
1888 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1889 Node* cmp = cmps[cmpn];
1890 if (cmp == NULL) continue;
1891 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1892 Node* bol = cmp->fast_out(j);
1893 if (!bol->is_Bool()) continue;
1894 BoolTest::mask btest = bol->as_Bool()->_test._test;
1895 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1896 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1897 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1898 best_bol = bol->as_Bool();
1899 best_btest = btest;
1900 }
1901 }
1902 }
1903
1904 Node* answer_if_true = NULL;
1905 Node* answer_if_false = NULL;
1906 switch (best_btest) {
1907 default:
1908 if (cmpxy == NULL)
1909 cmpxy = ideal_cmpxy;
1910 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1911 // and fall through:
1912 case BoolTest::lt: // x < y
1913 case BoolTest::le: // x <= y
1914 answer_if_true = (want_max ? yvalue : xvalue);
1915 answer_if_false = (want_max ? xvalue : yvalue);
1916 break;
1917 case BoolTest::gt: // x > y
1918 case BoolTest::ge: // x >= y
1919 answer_if_true = (want_max ? xvalue : yvalue);
1920 answer_if_false = (want_max ? yvalue : xvalue);
1921 break;
1922 }
1923
1924 jint hi, lo;
1925 if (want_max) {
1926 // We can sharpen the minimum.
1927 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1928 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1929 } else {
1930 // We can sharpen the maximum.
1931 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1932 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1933 }
1934
1935 // Use a flow-free graph structure, to avoid creating excess control edges
1936 // which could hinder other optimizations.
1937 // Since Math.min/max is often used with arraycopy, we want
1938 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1939 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1940 answer_if_false, answer_if_true,
1941 TypeInt::make(lo, hi, widen));
1942
1943 return _gvn.transform(cmov);
1944
1945 /*
1946 // This is not as desirable as it may seem, since Min and Max
1947 // nodes do not have a full set of optimizations.
1948 // And they would interfere, anyway, with 'if' optimizations
1949 // and with CMoveI canonical forms.
1950 switch (id) {
1951 case vmIntrinsics::_min:
1952 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1953 case vmIntrinsics::_max:
1954 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1955 default:
1956 ShouldNotReachHere();
1957 }
1958 */
1959 }
1960
1961 inline int
1962 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1963 const TypePtr* base_type = TypePtr::NULL_PTR;
1964 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1965 if (base_type == NULL) {
1966 // Unknown type.
1967 return Type::AnyPtr;
1968 } else if (base_type == TypePtr::NULL_PTR) {
1969 // Since this is a NULL+long form, we have to switch to a rawptr.
1970 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1971 offset = MakeConX(0);
1972 return Type::RawPtr;
1973 } else if (base_type->base() == Type::RawPtr) {
1974 return Type::RawPtr;
1975 } else if (base_type->isa_oopptr()) {
1976 // Base is never null => always a heap address.
1977 if (base_type->ptr() == TypePtr::NotNull) {
1978 return Type::OopPtr;
1979 }
1980 // Offset is small => always a heap address.
1981 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1982 if (offset_type != NULL &&
1983 base_type->offset() == 0 && // (should always be?)
1984 offset_type->_lo >= 0 &&
1985 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1986 return Type::OopPtr;
1987 }
1988 // Otherwise, it might either be oop+off or NULL+addr.
1989 return Type::AnyPtr;
1990 } else {
1991 // No information:
1992 return Type::AnyPtr;
1993 }
1994 }
1995
1996 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
1997 int kind = classify_unsafe_addr(base, offset);
1998 if (kind == Type::RawPtr) {
1999 return basic_plus_adr(top(), base, offset);
2000 } else {
2001 return basic_plus_adr(base, offset);
2002 }
2003 }
2004
2005 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2006 // inline int Integer.numberOfLeadingZeros(int)
2007 // inline int Long.numberOfLeadingZeros(long)
2008 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2009 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2010 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2011 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2012 _sp += arg_size(); // restore stack pointer
2013 switch (id) {
2014 case vmIntrinsics::_numberOfLeadingZeros_i:
2015 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2016 break;
2017 case vmIntrinsics::_numberOfLeadingZeros_l:
2018 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2019 break;
2020 default:
2021 ShouldNotReachHere();
2022 }
2023 return true;
2024 }
2025
2026 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2027 // inline int Integer.numberOfTrailingZeros(int)
2028 // inline int Long.numberOfTrailingZeros(long)
2029 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2030 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2031 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2032 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2033 _sp += arg_size(); // restore stack pointer
2034 switch (id) {
2035 case vmIntrinsics::_numberOfTrailingZeros_i:
2036 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2037 break;
2038 case vmIntrinsics::_numberOfTrailingZeros_l:
2039 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2040 break;
2041 default:
2042 ShouldNotReachHere();
2043 }
2044 return true;
2045 }
2046
2047 //----------------------------inline_bitCount_int/long-----------------------
2048 // inline int Integer.bitCount(int)
2049 // inline int Long.bitCount(long)
2050 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2051 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2052 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2053 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2054 _sp += arg_size(); // restore stack pointer
2055 switch (id) {
2056 case vmIntrinsics::_bitCount_i:
2057 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2058 break;
2059 case vmIntrinsics::_bitCount_l:
2060 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2061 break;
2062 default:
2063 ShouldNotReachHere();
2064 }
2065 return true;
2066 }
2067
2068 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2069 // inline Integer.reverseBytes(int)
2070 // inline Long.reverseBytes(long)
2071 // inline Character.reverseBytes(char)
2072 // inline Short.reverseBytes(short)
2073 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2074 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2075 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2076 "not reverse Bytes");
2077 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2078 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2079 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2080 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2081 _sp += arg_size(); // restore stack pointer
2082 switch (id) {
2083 case vmIntrinsics::_reverseBytes_i:
2084 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2085 break;
2086 case vmIntrinsics::_reverseBytes_l:
2087 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2088 break;
2089 case vmIntrinsics::_reverseBytes_c:
2090 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2091 break;
2092 case vmIntrinsics::_reverseBytes_s:
2093 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2094 break;
2095 default:
2096 ;
2097 }
2098 return true;
2099 }
2100
2101 //----------------------------inline_unsafe_access----------------------------
2102
2103 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2104
2105 // Helper that guards and inserts a G1 pre-barrier.
2106 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2107 assert(UseG1GC, "should not call this otherwise");
2108
2109 // We could be accessing the referent field of a reference object. If so, when G1
2110 // is enabled, we need to log the value in the referent field in an SATB buffer.
2111 // This routine performs some compile time filters and generates suitable
2112 // runtime filters that guard the pre-barrier code.
2113
2114 // Some compile time checks.
2115
2116 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2117 const TypeX* otype = offset->find_intptr_t_type();
2118 if (otype != NULL && otype->is_con() &&
2119 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2120 // Constant offset but not the reference_offset so just return
2121 return;
2122 }
2123
2124 // We only need to generate the runtime guards for instances.
2125 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2126 if (btype != NULL) {
2127 if (btype->isa_aryptr()) {
2128 // Array type so nothing to do
2129 return;
2130 }
2131
2132 const TypeInstPtr* itype = btype->isa_instptr();
2133 if (itype != NULL) {
2134 // Can the klass of base_oop be statically determined
2135 // to be _not_ a sub-class of Reference?
2136 ciKlass* klass = itype->klass();
2137 if (klass->is_subtype_of(env()->Reference_klass()) &&
2138 !env()->Reference_klass()->is_subtype_of(klass)) {
2139 return;
2140 }
2141 }
2142 }
2143
2144 // The compile time filters did not reject base_oop/offset so
2145 // we need to generate the following runtime filters
2146 //
2147 // if (offset == java_lang_ref_Reference::_reference_offset) {
2148 // if (base != null) {
2149 // if (instance_of(base, java.lang.ref.Reference)) {
2150 // pre_barrier(_, pre_val, ...);
2151 // }
2152 // }
2153 // }
2154
2155 float likely = PROB_LIKELY(0.999);
2156 float unlikely = PROB_UNLIKELY(0.999);
2157
2158 IdealKit ideal(this);
2159 #define __ ideal.
2160
2161 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2162
2163 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2164 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2165
2166 // Update graphKit memory and control from IdealKit.
2167 sync_kit(ideal);
2168
2169 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2170 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2171
2172 // Update IdealKit memory and control from graphKit.
2173 __ sync_kit(this);
2174
2175 Node* one = __ ConI(1);
2176
2177 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2178
2179 // Update graphKit from IdeakKit.
2180 sync_kit(ideal);
2181
2182 // Use the pre-barrier to record the value in the referent field
2183 pre_barrier(false /* do_load */,
2184 __ ctrl(),
2185 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2186 pre_val /* pre_val */,
2187 T_OBJECT);
2188
2189 // Update IdealKit from graphKit.
2190 __ sync_kit(this);
2191
2192 } __ end_if(); // _ref_type != ref_none
2193 } __ end_if(); // base != NULL
2194 } __ end_if(); // offset == referent_offset
2195
2196 // Final sync IdealKit and GraphKit.
2197 final_sync(ideal);
2198 #undef __
2199 }
2200
2201
2202 // Interpret Unsafe.fieldOffset cookies correctly:
2203 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2204
2205 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2206 if (callee()->is_static()) return false; // caller must have the capability!
2207
2208 #ifndef PRODUCT
2209 {
2210 ResourceMark rm;
2211 // Check the signatures.
2212 ciSignature* sig = signature();
2213 #ifdef ASSERT
2214 if (!is_store) {
2215 // Object getObject(Object base, int/long offset), etc.
2216 BasicType rtype = sig->return_type()->basic_type();
2217 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2218 rtype = T_ADDRESS; // it is really a C void*
2219 assert(rtype == type, "getter must return the expected value");
2220 if (!is_native_ptr) {
2221 assert(sig->count() == 2, "oop getter has 2 arguments");
2222 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2223 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2224 } else {
2225 assert(sig->count() == 1, "native getter has 1 argument");
2226 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2227 }
2228 } else {
2229 // void putObject(Object base, int/long offset, Object x), etc.
2230 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2231 if (!is_native_ptr) {
2232 assert(sig->count() == 3, "oop putter has 3 arguments");
2233 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2234 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2235 } else {
2236 assert(sig->count() == 2, "native putter has 2 arguments");
2237 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2238 }
2239 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2240 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2241 vtype = T_ADDRESS; // it is really a C void*
2242 assert(vtype == type, "putter must accept the expected value");
2243 }
2244 #endif // ASSERT
2245 }
2246 #endif //PRODUCT
2247
2248 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2249
2250 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2251
2252 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2253 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2254
2255 debug_only(int saved_sp = _sp);
2256 _sp += nargs;
2257
2258 Node* val;
2259 debug_only(val = (Node*)(uintptr_t)-1);
2260
2261
2262 if (is_store) {
2263 // Get the value being stored. (Pop it first; it was pushed last.)
2264 switch (type) {
2265 case T_DOUBLE:
2266 case T_LONG:
2267 case T_ADDRESS:
2268 val = pop_pair();
2269 break;
2270 default:
2271 val = pop();
2272 }
2273 }
2274
2275 // Build address expression. See the code in inline_unsafe_prefetch.
2276 Node *adr;
2277 Node *heap_base_oop = top();
2278 Node* offset = top();
2279
2280 if (!is_native_ptr) {
2281 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2282 offset = pop_pair();
2283 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2284 Node* base = pop();
2285 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2286 // to be plain byte offsets, which are also the same as those accepted
2287 // by oopDesc::field_base.
2288 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2289 "fieldOffset must be byte-scaled");
2290 // 32-bit machines ignore the high half!
2291 offset = ConvL2X(offset);
2292 adr = make_unsafe_address(base, offset);
2293 heap_base_oop = base;
2294 } else {
2295 Node* ptr = pop_pair();
2296 // Adjust Java long to machine word:
2297 ptr = ConvL2X(ptr);
2298 adr = make_unsafe_address(NULL, ptr);
2299 }
2300
2301 // Pop receiver last: it was pushed first.
2302 Node *receiver = pop();
2303
2304 assert(saved_sp == _sp, "must have correct argument count");
2305
2306 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2307
2308 // First guess at the value type.
2309 const Type *value_type = Type::get_const_basic_type(type);
2310
2311 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2312 // there was not enough information to nail it down.
2313 Compile::AliasType* alias_type = C->alias_type(adr_type);
2314 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2315
2316 // We will need memory barriers unless we can determine a unique
2317 // alias category for this reference. (Note: If for some reason
2318 // the barriers get omitted and the unsafe reference begins to "pollute"
2319 // the alias analysis of the rest of the graph, either Compile::can_alias
2320 // or Compile::must_alias will throw a diagnostic assert.)
2321 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2322
2323 // If we are reading the value of the referent field of a Reference
2324 // object (either by using Unsafe directly or through reflection)
2325 // then, if G1 is enabled, we need to record the referent in an
2326 // SATB log buffer using the pre-barrier mechanism.
2327 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2328 offset != top() && heap_base_oop != top();
2329
2330 if (!is_store && type == T_OBJECT) {
2331 // Attempt to infer a sharper value type from the offset and base type.
2332 ciKlass* sharpened_klass = NULL;
2333
2334 // See if it is an instance field, with an object type.
2335 if (alias_type->field() != NULL) {
2336 assert(!is_native_ptr, "native pointer op cannot use a java address");
2337 if (alias_type->field()->type()->is_klass()) {
2338 sharpened_klass = alias_type->field()->type()->as_klass();
2339 }
2340 }
2341
2342 // See if it is a narrow oop array.
2343 if (adr_type->isa_aryptr()) {
2344 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2345 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2346 if (elem_type != NULL) {
2347 sharpened_klass = elem_type->klass();
2348 }
2349 }
2350 }
2351
2352 if (sharpened_klass != NULL) {
2353 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2354
2355 // Sharpen the value type.
2356 value_type = tjp;
2357
2358 #ifndef PRODUCT
2359 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2360 tty->print(" from base type: "); adr_type->dump();
2361 tty->print(" sharpened value: "); value_type->dump();
2362 }
2363 #endif
2364 }
2365 }
2366
2367 // Null check on self without removing any arguments. The argument
2368 // null check technically happens in the wrong place, which can lead to
2369 // invalid stack traces when the primitive is inlined into a method
2370 // which handles NullPointerExceptions.
2371 _sp += nargs;
2372 do_null_check(receiver, T_OBJECT);
2373 _sp -= nargs;
2374 if (stopped()) {
2375 return true;
2376 }
2377 // Heap pointers get a null-check from the interpreter,
2378 // as a courtesy. However, this is not guaranteed by Unsafe,
2379 // and it is not possible to fully distinguish unintended nulls
2380 // from intended ones in this API.
2381
2382 if (is_volatile) {
2383 // We need to emit leading and trailing CPU membars (see below) in
2384 // addition to memory membars when is_volatile. This is a little
2385 // too strong, but avoids the need to insert per-alias-type
2386 // volatile membars (for stores; compare Parse::do_put_xxx), which
2387 // we cannot do effectively here because we probably only have a
2388 // rough approximation of type.
2389 need_mem_bar = true;
2390 // For Stores, place a memory ordering barrier now.
2391 if (is_store)
2392 insert_mem_bar(Op_MemBarRelease);
2393 }
2394
2395 // Memory barrier to prevent normal and 'unsafe' accesses from
2396 // bypassing each other. Happens after null checks, so the
2397 // exception paths do not take memory state from the memory barrier,
2398 // so there's no problems making a strong assert about mixing users
2399 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2400 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2401 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2402
2403 if (!is_store) {
2404 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2405 // load value and push onto stack
2406 switch (type) {
2407 case T_BOOLEAN:
2408 case T_CHAR:
2409 case T_BYTE:
2410 case T_SHORT:
2411 case T_INT:
2412 case T_FLOAT:
2413 push(p);
2414 break;
2415 case T_OBJECT:
2416 if (need_read_barrier) {
2417 insert_g1_pre_barrier(heap_base_oop, offset, p);
2418 }
2419 push(p);
2420 break;
2421 case T_ADDRESS:
2422 // Cast to an int type.
2423 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2424 p = ConvX2L(p);
2425 push_pair(p);
2426 break;
2427 case T_DOUBLE:
2428 case T_LONG:
2429 push_pair( p );
2430 break;
2431 default: ShouldNotReachHere();
2432 }
2433 } else {
2434 // place effect of store into memory
2435 switch (type) {
2436 case T_DOUBLE:
2437 val = dstore_rounding(val);
2438 break;
2439 case T_ADDRESS:
2440 // Repackage the long as a pointer.
2441 val = ConvL2X(val);
2442 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2443 break;
2444 }
2445
2446 if (type != T_OBJECT ) {
2447 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2448 } else {
2449 // Possibly an oop being stored to Java heap or native memory
2450 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2451 // oop to Java heap.
2452 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2453 } else {
2454 // We can't tell at compile time if we are storing in the Java heap or outside
2455 // of it. So we need to emit code to conditionally do the proper type of
2456 // store.
2457
2458 IdealKit ideal(this);
2459 #define __ ideal.
2460 // QQQ who knows what probability is here??
2461 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2462 // Sync IdealKit and graphKit.
2463 sync_kit(ideal);
2464 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2465 // Update IdealKit memory.
2466 __ sync_kit(this);
2467 } __ else_(); {
2468 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2469 } __ end_if();
2470 // Final sync IdealKit and GraphKit.
2471 final_sync(ideal);
2472 #undef __
2473 }
2474 }
2475 }
2476
2477 if (is_volatile) {
2478 if (!is_store)
2479 insert_mem_bar(Op_MemBarAcquire);
2480 else
2481 insert_mem_bar(Op_MemBarVolatile);
2482 }
2483
2484 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2485
2486 return true;
2487 }
2488
2489 //----------------------------inline_unsafe_prefetch----------------------------
2490
2491 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2492 #ifndef PRODUCT
2493 {
2494 ResourceMark rm;
2495 // Check the signatures.
2496 ciSignature* sig = signature();
2497 #ifdef ASSERT
2498 // Object getObject(Object base, int/long offset), etc.
2499 BasicType rtype = sig->return_type()->basic_type();
2500 if (!is_native_ptr) {
2501 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2502 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2503 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2504 } else {
2505 assert(sig->count() == 1, "native prefetch has 1 argument");
2506 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2507 }
2508 #endif // ASSERT
2509 }
2510 #endif // !PRODUCT
2511
2512 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2513
2514 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2515 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2516
2517 debug_only(int saved_sp = _sp);
2518 _sp += nargs;
2519
2520 // Build address expression. See the code in inline_unsafe_access.
2521 Node *adr;
2522 if (!is_native_ptr) {
2523 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2524 Node* offset = pop_pair();
2525 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2526 Node* base = pop();
2527 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2528 // to be plain byte offsets, which are also the same as those accepted
2529 // by oopDesc::field_base.
2530 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2531 "fieldOffset must be byte-scaled");
2532 // 32-bit machines ignore the high half!
2533 offset = ConvL2X(offset);
2534 adr = make_unsafe_address(base, offset);
2535 } else {
2536 Node* ptr = pop_pair();
2537 // Adjust Java long to machine word:
2538 ptr = ConvL2X(ptr);
2539 adr = make_unsafe_address(NULL, ptr);
2540 }
2541
2542 if (is_static) {
2543 assert(saved_sp == _sp, "must have correct argument count");
2544 } else {
2545 // Pop receiver last: it was pushed first.
2546 Node *receiver = pop();
2547 assert(saved_sp == _sp, "must have correct argument count");
2548
2549 // Null check on self without removing any arguments. The argument
2550 // null check technically happens in the wrong place, which can lead to
2551 // invalid stack traces when the primitive is inlined into a method
2552 // which handles NullPointerExceptions.
2553 _sp += nargs;
2554 do_null_check(receiver, T_OBJECT);
2555 _sp -= nargs;
2556 if (stopped()) {
2557 return true;
2558 }
2559 }
2560
2561 // Generate the read or write prefetch
2562 Node *prefetch;
2563 if (is_store) {
2564 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2565 } else {
2566 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2567 }
2568 prefetch->init_req(0, control());
2569 set_i_o(_gvn.transform(prefetch));
2570
2571 return true;
2572 }
2573
2574 //----------------------------inline_unsafe_CAS----------------------------
2575
2576 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2577 // This basic scheme here is the same as inline_unsafe_access, but
2578 // differs in enough details that combining them would make the code
2579 // overly confusing. (This is a true fact! I originally combined
2580 // them, but even I was confused by it!) As much code/comments as
2581 // possible are retained from inline_unsafe_access though to make
2582 // the correspondences clearer. - dl
2583
2584 if (callee()->is_static()) return false; // caller must have the capability!
2585
2586 #ifndef PRODUCT
2587 {
2588 ResourceMark rm;
2589 // Check the signatures.
2590 ciSignature* sig = signature();
2591 #ifdef ASSERT
2592 BasicType rtype = sig->return_type()->basic_type();
2593 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2594 assert(sig->count() == 4, "CAS has 4 arguments");
2595 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2596 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2597 #endif // ASSERT
2598 }
2599 #endif //PRODUCT
2600
2601 // number of stack slots per value argument (1 or 2)
2602 int type_words = type2size[type];
2603
2604 // Cannot inline wide CAS on machines that don't support it natively
2605 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2606 return false;
2607
2608 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2609
2610 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2611 int nargs = 1 + 1 + 2 + type_words + type_words;
2612
2613 // pop arguments: newval, oldval, offset, base, and receiver
2614 debug_only(int saved_sp = _sp);
2615 _sp += nargs;
2616 Node* newval = (type_words == 1) ? pop() : pop_pair();
2617 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2618 Node *offset = pop_pair();
2619 Node *base = pop();
2620 Node *receiver = pop();
2621 assert(saved_sp == _sp, "must have correct argument count");
2622
2623 // Null check receiver.
2624 _sp += nargs;
2625 do_null_check(receiver, T_OBJECT);
2626 _sp -= nargs;
2627 if (stopped()) {
2628 return true;
2629 }
2630
2631 // Build field offset expression.
2632 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2633 // to be plain byte offsets, which are also the same as those accepted
2634 // by oopDesc::field_base.
2635 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2636 // 32-bit machines ignore the high half of long offsets
2637 offset = ConvL2X(offset);
2638 Node* adr = make_unsafe_address(base, offset);
2639 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2640
2641 // (Unlike inline_unsafe_access, there seems no point in trying
2642 // to refine types. Just use the coarse types here.
2643 const Type *value_type = Type::get_const_basic_type(type);
2644 Compile::AliasType* alias_type = C->alias_type(adr_type);
2645 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2646 int alias_idx = C->get_alias_index(adr_type);
2647
2648 // Memory-model-wise, a CAS acts like a little synchronized block,
2649 // so needs barriers on each side. These don't translate into
2650 // actual barriers on most machines, but we still need rest of
2651 // compiler to respect ordering.
2652
2653 insert_mem_bar(Op_MemBarRelease);
2654 insert_mem_bar(Op_MemBarCPUOrder);
2655
2656 // 4984716: MemBars must be inserted before this
2657 // memory node in order to avoid a false
2658 // dependency which will confuse the scheduler.
2659 Node *mem = memory(alias_idx);
2660
2661 // For now, we handle only those cases that actually exist: ints,
2662 // longs, and Object. Adding others should be straightforward.
2663 Node* cas;
2664 switch(type) {
2665 case T_INT:
2666 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2667 break;
2668 case T_LONG:
2669 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2670 break;
2671 case T_OBJECT:
2672 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2673 // could be delayed during Parse (for example, in adjust_map_after_if()).
2674 // Execute transformation here to avoid barrier generation in such case.
2675 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2676 newval = _gvn.makecon(TypePtr::NULL_PTR);
2677
2678 // Reference stores need a store barrier.
2679 // (They don't if CAS fails, but it isn't worth checking.)
2680 pre_barrier(true /* do_load*/,
2681 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2682 NULL /* pre_val*/,
2683 T_OBJECT);
2684 #ifdef _LP64
2685 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2686 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2687 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2688 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2689 newval_enc, oldval_enc));
2690 } else
2691 #endif
2692 {
2693 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2694 }
2695 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2696 break;
2697 default:
2698 ShouldNotReachHere();
2699 break;
2700 }
2701
2702 // SCMemProjNodes represent the memory state of CAS. Their main
2703 // role is to prevent CAS nodes from being optimized away when their
2704 // results aren't used.
2705 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2706 set_memory(proj, alias_idx);
2707
2708 // Add the trailing membar surrounding the access
2709 insert_mem_bar(Op_MemBarCPUOrder);
2710 insert_mem_bar(Op_MemBarAcquire);
2711
2712 push(cas);
2713 return true;
2714 }
2715
2716 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2717 // This is another variant of inline_unsafe_access, differing in
2718 // that it always issues store-store ("release") barrier and ensures
2719 // store-atomicity (which only matters for "long").
2720
2721 if (callee()->is_static()) return false; // caller must have the capability!
2722
2723 #ifndef PRODUCT
2724 {
2725 ResourceMark rm;
2726 // Check the signatures.
2727 ciSignature* sig = signature();
2728 #ifdef ASSERT
2729 BasicType rtype = sig->return_type()->basic_type();
2730 assert(rtype == T_VOID, "must return void");
2731 assert(sig->count() == 3, "has 3 arguments");
2732 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2733 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2734 #endif // ASSERT
2735 }
2736 #endif //PRODUCT
2737
2738 // number of stack slots per value argument (1 or 2)
2739 int type_words = type2size[type];
2740
2741 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2742
2743 // Argument words: "this" plus oop plus offset plus value;
2744 int nargs = 1 + 1 + 2 + type_words;
2745
2746 // pop arguments: val, offset, base, and receiver
2747 debug_only(int saved_sp = _sp);
2748 _sp += nargs;
2749 Node* val = (type_words == 1) ? pop() : pop_pair();
2750 Node *offset = pop_pair();
2751 Node *base = pop();
2752 Node *receiver = pop();
2753 assert(saved_sp == _sp, "must have correct argument count");
2754
2755 // Null check receiver.
2756 _sp += nargs;
2757 do_null_check(receiver, T_OBJECT);
2758 _sp -= nargs;
2759 if (stopped()) {
2760 return true;
2761 }
2762
2763 // Build field offset expression.
2764 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2765 // 32-bit machines ignore the high half of long offsets
2766 offset = ConvL2X(offset);
2767 Node* adr = make_unsafe_address(base, offset);
2768 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2769 const Type *value_type = Type::get_const_basic_type(type);
2770 Compile::AliasType* alias_type = C->alias_type(adr_type);
2771
2772 insert_mem_bar(Op_MemBarRelease);
2773 insert_mem_bar(Op_MemBarCPUOrder);
2774 // Ensure that the store is atomic for longs:
2775 bool require_atomic_access = true;
2776 Node* store;
2777 if (type == T_OBJECT) // reference stores need a store barrier.
2778 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2779 else {
2780 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2781 }
2782 insert_mem_bar(Op_MemBarCPUOrder);
2783 return true;
2784 }
2785
2786 bool LibraryCallKit::inline_unsafe_allocate() {
2787 if (callee()->is_static()) return false; // caller must have the capability!
2788 int nargs = 1 + 1;
2789 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2790 null_check_receiver(callee()); // check then ignore argument(0)
2791 _sp += nargs; // set original stack for use by uncommon_trap
2792 Node* cls = do_null_check(argument(1), T_OBJECT);
2793 _sp -= nargs;
2794 if (stopped()) return true;
2795
2796 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2797 _sp += nargs; // set original stack for use by uncommon_trap
2798 kls = do_null_check(kls, T_OBJECT);
2799 _sp -= nargs;
2800 if (stopped()) return true; // argument was like int.class
2801
2802 // Note: The argument might still be an illegal value like
2803 // Serializable.class or Object[].class. The runtime will handle it.
2804 // But we must make an explicit check for initialization.
2805 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2806 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2807 // can generate code to load it as unsigned byte.
2808 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2809 Node* bits = intcon(instanceKlass::fully_initialized);
2810 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2811 // The 'test' is non-zero if we need to take a slow path.
2812
2813 Node* obj = new_instance(kls, test);
2814 push(obj);
2815
2816 return true;
2817 }
2818
2819 //------------------------inline_native_time_funcs--------------
2820 // inline code for System.currentTimeMillis() and System.nanoTime()
2821 // these have the same type and signature
2822 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2823 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2824 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2825 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2826 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2827 const TypePtr* no_memory_effects = NULL;
2828 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2829 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2830 #ifdef ASSERT
2831 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2832 assert(value_top == top(), "second value must be top");
2833 #endif
2834 push_pair(value);
2835 return true;
2836 }
2837
2838 //------------------------inline_native_currentThread------------------
2839 bool LibraryCallKit::inline_native_currentThread() {
2840 Node* junk = NULL;
2841 push(generate_current_thread(junk));
2842 return true;
2843 }
2844
2845 //------------------------inline_native_isInterrupted------------------
2846 bool LibraryCallKit::inline_native_isInterrupted() {
2847 const int nargs = 1+1; // receiver + boolean
2848 assert(nargs == arg_size(), "sanity");
2849 // Add a fast path to t.isInterrupted(clear_int):
2850 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2851 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2852 // So, in the common case that the interrupt bit is false,
2853 // we avoid making a call into the VM. Even if the interrupt bit
2854 // is true, if the clear_int argument is false, we avoid the VM call.
2855 // However, if the receiver is not currentThread, we must call the VM,
2856 // because there must be some locking done around the operation.
2857
2858 // We only go to the fast case code if we pass two guards.
2859 // Paths which do not pass are accumulated in the slow_region.
2860 RegionNode* slow_region = new (C, 1) RegionNode(1);
2861 record_for_igvn(slow_region);
2862 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2863 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2864 enum { no_int_result_path = 1,
2865 no_clear_result_path = 2,
2866 slow_result_path = 3
2867 };
2868
2869 // (a) Receiving thread must be the current thread.
2870 Node* rec_thr = argument(0);
2871 Node* tls_ptr = NULL;
2872 Node* cur_thr = generate_current_thread(tls_ptr);
2873 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2874 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2875
2876 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2877 if (!known_current_thread)
2878 generate_slow_guard(bol_thr, slow_region);
2879
2880 // (b) Interrupt bit on TLS must be false.
2881 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2882 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2883 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2884 // Set the control input on the field _interrupted read to prevent it floating up.
2885 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2886 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2887 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2888
2889 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2890
2891 // First fast path: if (!TLS._interrupted) return false;
2892 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2893 result_rgn->init_req(no_int_result_path, false_bit);
2894 result_val->init_req(no_int_result_path, intcon(0));
2895
2896 // drop through to next case
2897 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2898
2899 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2900 Node* clr_arg = argument(1);
2901 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2902 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2903 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2904
2905 // Second fast path: ... else if (!clear_int) return true;
2906 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2907 result_rgn->init_req(no_clear_result_path, false_arg);
2908 result_val->init_req(no_clear_result_path, intcon(1));
2909
2910 // drop through to next case
2911 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2912
2913 // (d) Otherwise, go to the slow path.
2914 slow_region->add_req(control());
2915 set_control( _gvn.transform(slow_region) );
2916
2917 if (stopped()) {
2918 // There is no slow path.
2919 result_rgn->init_req(slow_result_path, top());
2920 result_val->init_req(slow_result_path, top());
2921 } else {
2922 // non-virtual because it is a private non-static
2923 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2924
2925 Node* slow_val = set_results_for_java_call(slow_call);
2926 // this->control() comes from set_results_for_java_call
2927
2928 // If we know that the result of the slow call will be true, tell the optimizer!
2929 if (known_current_thread) slow_val = intcon(1);
2930
2931 Node* fast_io = slow_call->in(TypeFunc::I_O);
2932 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2933 // These two phis are pre-filled with copies of of the fast IO and Memory
2934 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2935 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2936
2937 result_rgn->init_req(slow_result_path, control());
2938 io_phi ->init_req(slow_result_path, i_o());
2939 mem_phi ->init_req(slow_result_path, reset_memory());
2940 result_val->init_req(slow_result_path, slow_val);
2941
2942 set_all_memory( _gvn.transform(mem_phi) );
2943 set_i_o( _gvn.transform(io_phi) );
2944 }
2945
2946 push_result(result_rgn, result_val);
2947 C->set_has_split_ifs(true); // Has chance for split-if optimization
2948
2949 return true;
2950 }
2951
2952 //---------------------------load_mirror_from_klass----------------------------
2953 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2954 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2955 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
2956 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2957 }
2958
2959 //-----------------------load_klass_from_mirror_common-------------------------
2960 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2961 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2962 // and branch to the given path on the region.
2963 // If never_see_null, take an uncommon trap on null, so we can optimistically
2964 // compile for the non-null case.
2965 // If the region is NULL, force never_see_null = true.
2966 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2967 bool never_see_null,
2968 int nargs,
2969 RegionNode* region,
2970 int null_path,
2971 int offset) {
2972 if (region == NULL) never_see_null = true;
2973 Node* p = basic_plus_adr(mirror, offset);
2974 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2975 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2976 _sp += nargs; // any deopt will start just before call to enclosing method
2977 Node* null_ctl = top();
2978 kls = null_check_oop(kls, &null_ctl, never_see_null);
2979 if (region != NULL) {
2980 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2981 region->init_req(null_path, null_ctl);
2982 } else {
2983 assert(null_ctl == top(), "no loose ends");
2984 }
2985 _sp -= nargs;
2986 return kls;
2987 }
2988
2989 //--------------------(inline_native_Class_query helpers)---------------------
2990 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
2991 // Fall through if (mods & mask) == bits, take the guard otherwise.
2992 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
2993 // Branch around if the given klass has the given modifier bit set.
2994 // Like generate_guard, adds a new path onto the region.
2995 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
2996 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
2997 Node* mask = intcon(modifier_mask);
2998 Node* bits = intcon(modifier_bits);
2999 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3000 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3001 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3002 return generate_fair_guard(bol, region);
3003 }
3004 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3005 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3006 }
3007
3008 //-------------------------inline_native_Class_query-------------------
3009 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3010 int nargs = 1+0; // just the Class mirror, in most cases
3011 const Type* return_type = TypeInt::BOOL;
3012 Node* prim_return_value = top(); // what happens if it's a primitive class?
3013 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3014 bool expect_prim = false; // most of these guys expect to work on refs
3015
3016 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3017
3018 switch (id) {
3019 case vmIntrinsics::_isInstance:
3020 nargs = 1+1; // the Class mirror, plus the object getting queried about
3021 // nothing is an instance of a primitive type
3022 prim_return_value = intcon(0);
3023 break;
3024 case vmIntrinsics::_getModifiers:
3025 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3026 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3027 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3028 break;
3029 case vmIntrinsics::_isInterface:
3030 prim_return_value = intcon(0);
3031 break;
3032 case vmIntrinsics::_isArray:
3033 prim_return_value = intcon(0);
3034 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3035 break;
3036 case vmIntrinsics::_isPrimitive:
3037 prim_return_value = intcon(1);
3038 expect_prim = true; // obviously
3039 break;
3040 case vmIntrinsics::_getSuperclass:
3041 prim_return_value = null();
3042 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3043 break;
3044 case vmIntrinsics::_getComponentType:
3045 prim_return_value = null();
3046 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3047 break;
3048 case vmIntrinsics::_getClassAccessFlags:
3049 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3050 return_type = TypeInt::INT; // not bool! 6297094
3051 break;
3052 default:
3053 ShouldNotReachHere();
3054 }
3055
3056 Node* mirror = argument(0);
3057 Node* obj = (nargs <= 1)? top(): argument(1);
3058
3059 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3060 if (mirror_con == NULL) return false; // cannot happen?
3061
3062 #ifndef PRODUCT
3063 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3064 ciType* k = mirror_con->java_mirror_type();
3065 if (k) {
3066 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3067 k->print_name();
3068 tty->cr();
3069 }
3070 }
3071 #endif
3072
3073 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3074 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3075 record_for_igvn(region);
3076 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3077
3078 // The mirror will never be null of Reflection.getClassAccessFlags, however
3079 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3080 // if it is. See bug 4774291.
3081
3082 // For Reflection.getClassAccessFlags(), the null check occurs in
3083 // the wrong place; see inline_unsafe_access(), above, for a similar
3084 // situation.
3085 _sp += nargs; // set original stack for use by uncommon_trap
3086 mirror = do_null_check(mirror, T_OBJECT);
3087 _sp -= nargs;
3088 // If mirror or obj is dead, only null-path is taken.
3089 if (stopped()) return true;
3090
3091 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3092
3093 // Now load the mirror's klass metaobject, and null-check it.
3094 // Side-effects region with the control path if the klass is null.
3095 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3096 region, _prim_path);
3097 // If kls is null, we have a primitive mirror.
3098 phi->init_req(_prim_path, prim_return_value);
3099 if (stopped()) { push_result(region, phi); return true; }
3100
3101 Node* p; // handy temp
3102 Node* null_ctl;
3103
3104 // Now that we have the non-null klass, we can perform the real query.
3105 // For constant classes, the query will constant-fold in LoadNode::Value.
3106 Node* query_value = top();
3107 switch (id) {
3108 case vmIntrinsics::_isInstance:
3109 // nothing is an instance of a primitive type
3110 _sp += nargs; // gen_instanceof might do an uncommon trap
3111 query_value = gen_instanceof(obj, kls);
3112 _sp -= nargs;
3113 break;
3114
3115 case vmIntrinsics::_getModifiers:
3116 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3117 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3118 break;
3119
3120 case vmIntrinsics::_isInterface:
3121 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3122 if (generate_interface_guard(kls, region) != NULL)
3123 // A guard was added. If the guard is taken, it was an interface.
3124 phi->add_req(intcon(1));
3125 // If we fall through, it's a plain class.
3126 query_value = intcon(0);
3127 break;
3128
3129 case vmIntrinsics::_isArray:
3130 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3131 if (generate_array_guard(kls, region) != NULL)
3132 // A guard was added. If the guard is taken, it was an array.
3133 phi->add_req(intcon(1));
3134 // If we fall through, it's a plain class.
3135 query_value = intcon(0);
3136 break;
3137
3138 case vmIntrinsics::_isPrimitive:
3139 query_value = intcon(0); // "normal" path produces false
3140 break;
3141
3142 case vmIntrinsics::_getSuperclass:
3143 // The rules here are somewhat unfortunate, but we can still do better
3144 // with random logic than with a JNI call.
3145 // Interfaces store null or Object as _super, but must report null.
3146 // Arrays store an intermediate super as _super, but must report Object.
3147 // Other types can report the actual _super.
3148 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3149 if (generate_interface_guard(kls, region) != NULL)
3150 // A guard was added. If the guard is taken, it was an interface.
3151 phi->add_req(null());
3152 if (generate_array_guard(kls, region) != NULL)
3153 // A guard was added. If the guard is taken, it was an array.
3154 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3155 // If we fall through, it's a plain class. Get its _super.
3156 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3157 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3158 null_ctl = top();
3159 kls = null_check_oop(kls, &null_ctl);
3160 if (null_ctl != top()) {
3161 // If the guard is taken, Object.superClass is null (both klass and mirror).
3162 region->add_req(null_ctl);
3163 phi ->add_req(null());
3164 }
3165 if (!stopped()) {
3166 query_value = load_mirror_from_klass(kls);
3167 }
3168 break;
3169
3170 case vmIntrinsics::_getComponentType:
3171 if (generate_array_guard(kls, region) != NULL) {
3172 // Be sure to pin the oop load to the guard edge just created:
3173 Node* is_array_ctrl = region->in(region->req()-1);
3174 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3175 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3176 phi->add_req(cmo);
3177 }
3178 query_value = null(); // non-array case is null
3179 break;
3180
3181 case vmIntrinsics::_getClassAccessFlags:
3182 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3183 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3184 break;
3185
3186 default:
3187 ShouldNotReachHere();
3188 }
3189
3190 // Fall-through is the normal case of a query to a real class.
3191 phi->init_req(1, query_value);
3192 region->init_req(1, control());
3193
3194 push_result(region, phi);
3195 C->set_has_split_ifs(true); // Has chance for split-if optimization
3196
3197 return true;
3198 }
3199
3200 //--------------------------inline_native_subtype_check------------------------
3201 // This intrinsic takes the JNI calls out of the heart of
3202 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3203 bool LibraryCallKit::inline_native_subtype_check() {
3204 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3205
3206 // Pull both arguments off the stack.
3207 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3208 args[0] = argument(0);
3209 args[1] = argument(1);
3210 Node* klasses[2]; // corresponding Klasses: superk, subk
3211 klasses[0] = klasses[1] = top();
3212
3213 enum {
3214 // A full decision tree on {superc is prim, subc is prim}:
3215 _prim_0_path = 1, // {P,N} => false
3216 // {P,P} & superc!=subc => false
3217 _prim_same_path, // {P,P} & superc==subc => true
3218 _prim_1_path, // {N,P} => false
3219 _ref_subtype_path, // {N,N} & subtype check wins => true
3220 _both_ref_path, // {N,N} & subtype check loses => false
3221 PATH_LIMIT
3222 };
3223
3224 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3225 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3226 record_for_igvn(region);
3227
3228 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3229 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3230 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3231
3232 // First null-check both mirrors and load each mirror's klass metaobject.
3233 int which_arg;
3234 for (which_arg = 0; which_arg <= 1; which_arg++) {
3235 Node* arg = args[which_arg];
3236 _sp += nargs; // set original stack for use by uncommon_trap
3237 arg = do_null_check(arg, T_OBJECT);
3238 _sp -= nargs;
3239 if (stopped()) break;
3240 args[which_arg] = _gvn.transform(arg);
3241
3242 Node* p = basic_plus_adr(arg, class_klass_offset);
3243 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3244 klasses[which_arg] = _gvn.transform(kls);
3245 }
3246
3247 // Having loaded both klasses, test each for null.
3248 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3249 for (which_arg = 0; which_arg <= 1; which_arg++) {
3250 Node* kls = klasses[which_arg];
3251 Node* null_ctl = top();
3252 _sp += nargs; // set original stack for use by uncommon_trap
3253 kls = null_check_oop(kls, &null_ctl, never_see_null);
3254 _sp -= nargs;
3255 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3256 region->init_req(prim_path, null_ctl);
3257 if (stopped()) break;
3258 klasses[which_arg] = kls;
3259 }
3260
3261 if (!stopped()) {
3262 // now we have two reference types, in klasses[0..1]
3263 Node* subk = klasses[1]; // the argument to isAssignableFrom
3264 Node* superk = klasses[0]; // the receiver
3265 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3266 // now we have a successful reference subtype check
3267 region->set_req(_ref_subtype_path, control());
3268 }
3269
3270 // If both operands are primitive (both klasses null), then
3271 // we must return true when they are identical primitives.
3272 // It is convenient to test this after the first null klass check.
3273 set_control(region->in(_prim_0_path)); // go back to first null check
3274 if (!stopped()) {
3275 // Since superc is primitive, make a guard for the superc==subc case.
3276 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3277 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3278 generate_guard(bol_eq, region, PROB_FAIR);
3279 if (region->req() == PATH_LIMIT+1) {
3280 // A guard was added. If the added guard is taken, superc==subc.
3281 region->swap_edges(PATH_LIMIT, _prim_same_path);
3282 region->del_req(PATH_LIMIT);
3283 }
3284 region->set_req(_prim_0_path, control()); // Not equal after all.
3285 }
3286
3287 // these are the only paths that produce 'true':
3288 phi->set_req(_prim_same_path, intcon(1));
3289 phi->set_req(_ref_subtype_path, intcon(1));
3290
3291 // pull together the cases:
3292 assert(region->req() == PATH_LIMIT, "sane region");
3293 for (uint i = 1; i < region->req(); i++) {
3294 Node* ctl = region->in(i);
3295 if (ctl == NULL || ctl == top()) {
3296 region->set_req(i, top());
3297 phi ->set_req(i, top());
3298 } else if (phi->in(i) == NULL) {
3299 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3300 }
3301 }
3302
3303 set_control(_gvn.transform(region));
3304 push(_gvn.transform(phi));
3305
3306 return true;
3307 }
3308
3309 //---------------------generate_array_guard_common------------------------
3310 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3311 bool obj_array, bool not_array) {
3312 // If obj_array/non_array==false/false:
3313 // Branch around if the given klass is in fact an array (either obj or prim).
3314 // If obj_array/non_array==false/true:
3315 // Branch around if the given klass is not an array klass of any kind.
3316 // If obj_array/non_array==true/true:
3317 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3318 // If obj_array/non_array==true/false:
3319 // Branch around if the kls is an oop array (Object[] or subtype)
3320 //
3321 // Like generate_guard, adds a new path onto the region.
3322 jint layout_con = 0;
3323 Node* layout_val = get_layout_helper(kls, layout_con);
3324 if (layout_val == NULL) {
3325 bool query = (obj_array
3326 ? Klass::layout_helper_is_objArray(layout_con)
3327 : Klass::layout_helper_is_javaArray(layout_con));
3328 if (query == not_array) {
3329 return NULL; // never a branch
3330 } else { // always a branch
3331 Node* always_branch = control();
3332 if (region != NULL)
3333 region->add_req(always_branch);
3334 set_control(top());
3335 return always_branch;
3336 }
3337 }
3338 // Now test the correct condition.
3339 jint nval = (obj_array
3340 ? ((jint)Klass::_lh_array_tag_type_value
3341 << Klass::_lh_array_tag_shift)
3342 : Klass::_lh_neutral_value);
3343 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3344 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3345 // invert the test if we are looking for a non-array
3346 if (not_array) btest = BoolTest(btest).negate();
3347 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3348 return generate_fair_guard(bol, region);
3349 }
3350
3351
3352 //-----------------------inline_native_newArray--------------------------
3353 bool LibraryCallKit::inline_native_newArray() {
3354 int nargs = 2;
3355 Node* mirror = argument(0);
3356 Node* count_val = argument(1);
3357
3358 _sp += nargs; // set original stack for use by uncommon_trap
3359 mirror = do_null_check(mirror, T_OBJECT);
3360 _sp -= nargs;
3361 // If mirror or obj is dead, only null-path is taken.
3362 if (stopped()) return true;
3363
3364 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3365 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3366 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3367 TypeInstPtr::NOTNULL);
3368 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3369 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3370 TypePtr::BOTTOM);
3371
3372 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3373 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3374 nargs,
3375 result_reg, _slow_path);
3376 Node* normal_ctl = control();
3377 Node* no_array_ctl = result_reg->in(_slow_path);
3378
3379 // Generate code for the slow case. We make a call to newArray().
3380 set_control(no_array_ctl);
3381 if (!stopped()) {
3382 // Either the input type is void.class, or else the
3383 // array klass has not yet been cached. Either the
3384 // ensuing call will throw an exception, or else it
3385 // will cache the array klass for next time.
3386 PreserveJVMState pjvms(this);
3387 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3388 Node* slow_result = set_results_for_java_call(slow_call);
3389 // this->control() comes from set_results_for_java_call
3390 result_reg->set_req(_slow_path, control());
3391 result_val->set_req(_slow_path, slow_result);
3392 result_io ->set_req(_slow_path, i_o());
3393 result_mem->set_req(_slow_path, reset_memory());
3394 }
3395
3396 set_control(normal_ctl);
3397 if (!stopped()) {
3398 // Normal case: The array type has been cached in the java.lang.Class.
3399 // The following call works fine even if the array type is polymorphic.
3400 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3401 Node* obj = new_array(klass_node, count_val, nargs);
3402 result_reg->init_req(_normal_path, control());
3403 result_val->init_req(_normal_path, obj);
3404 result_io ->init_req(_normal_path, i_o());
3405 result_mem->init_req(_normal_path, reset_memory());
3406 }
3407
3408 // Return the combined state.
3409 set_i_o( _gvn.transform(result_io) );
3410 set_all_memory( _gvn.transform(result_mem) );
3411 push_result(result_reg, result_val);
3412 C->set_has_split_ifs(true); // Has chance for split-if optimization
3413
3414 return true;
3415 }
3416
3417 //----------------------inline_native_getLength--------------------------
3418 bool LibraryCallKit::inline_native_getLength() {
3419 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3420
3421 int nargs = 1;
3422 Node* array = argument(0);
3423
3424 _sp += nargs; // set original stack for use by uncommon_trap
3425 array = do_null_check(array, T_OBJECT);
3426 _sp -= nargs;
3427
3428 // If array is dead, only null-path is taken.
3429 if (stopped()) return true;
3430
3431 // Deoptimize if it is a non-array.
3432 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3433
3434 if (non_array != NULL) {
3435 PreserveJVMState pjvms(this);
3436 set_control(non_array);
3437 _sp += nargs; // push the arguments back on the stack
3438 uncommon_trap(Deoptimization::Reason_intrinsic,
3439 Deoptimization::Action_maybe_recompile);
3440 }
3441
3442 // If control is dead, only non-array-path is taken.
3443 if (stopped()) return true;
3444
3445 // The works fine even if the array type is polymorphic.
3446 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3447 push( load_array_length(array) );
3448
3449 C->set_has_split_ifs(true); // Has chance for split-if optimization
3450
3451 return true;
3452 }
3453
3454 //------------------------inline_array_copyOf----------------------------
3455 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3456 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3457
3458 // Restore the stack and pop off the arguments.
3459 int nargs = 3 + (is_copyOfRange? 1: 0);
3460 Node* original = argument(0);
3461 Node* start = is_copyOfRange? argument(1): intcon(0);
3462 Node* end = is_copyOfRange? argument(2): argument(1);
3463 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3464
3465 Node* newcopy;
3466
3467 //set the original stack and the reexecute bit for the interpreter to reexecute
3468 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3469 { PreserveReexecuteState preexecs(this);
3470 _sp += nargs;
3471 jvms()->set_should_reexecute(true);
3472
3473 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3474 original = do_null_check(original, T_OBJECT);
3475
3476 // Check if a null path was taken unconditionally.
3477 if (stopped()) return true;
3478
3479 Node* orig_length = load_array_length(original);
3480
3481 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3482 NULL, 0);
3483 klass_node = do_null_check(klass_node, T_OBJECT);
3484
3485 RegionNode* bailout = new (C, 1) RegionNode(1);
3486 record_for_igvn(bailout);
3487
3488 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3489 // Bail out if that is so.
3490 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3491 if (not_objArray != NULL) {
3492 // Improve the klass node's type from the new optimistic assumption:
3493 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3494 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3495 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3496 cast->init_req(0, control());
3497 klass_node = _gvn.transform(cast);
3498 }
3499
3500 // Bail out if either start or end is negative.
3501 generate_negative_guard(start, bailout, &start);
3502 generate_negative_guard(end, bailout, &end);
3503
3504 Node* length = end;
3505 if (_gvn.type(start) != TypeInt::ZERO) {
3506 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3507 }
3508
3509 // Bail out if length is negative.
3510 // ...Not needed, since the new_array will throw the right exception.
3511 //generate_negative_guard(length, bailout, &length);
3512
3513 if (bailout->req() > 1) {
3514 PreserveJVMState pjvms(this);
3515 set_control( _gvn.transform(bailout) );
3516 uncommon_trap(Deoptimization::Reason_intrinsic,
3517 Deoptimization::Action_maybe_recompile);
3518 }
3519
3520 if (!stopped()) {
3521
3522 // How many elements will we copy from the original?
3523 // The answer is MinI(orig_length - start, length).
3524 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3525 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3526
3527 newcopy = new_array(klass_node, length, 0);
3528
3529 // Generate a direct call to the right arraycopy function(s).
3530 // We know the copy is disjoint but we might not know if the
3531 // oop stores need checking.
3532 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3533 // This will fail a store-check if x contains any non-nulls.
3534 bool disjoint_bases = true;
3535 bool length_never_negative = true;
3536 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3537 original, start, newcopy, intcon(0), moved,
3538 disjoint_bases, length_never_negative);
3539 }
3540 } //original reexecute and sp are set back here
3541
3542 if(!stopped()) {
3543 push(newcopy);
3544 }
3545
3546 C->set_has_split_ifs(true); // Has chance for split-if optimization
3547
3548 return true;
3549 }
3550
3551
3552 //----------------------generate_virtual_guard---------------------------
3553 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3554 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3555 RegionNode* slow_region) {
3556 ciMethod* method = callee();
3557 int vtable_index = method->vtable_index();
3558 // Get the methodOop out of the appropriate vtable entry.
3559 int entry_offset = (instanceKlass::vtable_start_offset() +
3560 vtable_index*vtableEntry::size()) * wordSize +
3561 vtableEntry::method_offset_in_bytes();
3562 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3563 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3564
3565 // Compare the target method with the expected method (e.g., Object.hashCode).
3566 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3567
3568 Node* native_call = makecon(native_call_addr);
3569 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3570 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3571
3572 return generate_slow_guard(test_native, slow_region);
3573 }
3574
3575 //-----------------------generate_method_call----------------------------
3576 // Use generate_method_call to make a slow-call to the real
3577 // method if the fast path fails. An alternative would be to
3578 // use a stub like OptoRuntime::slow_arraycopy_Java.
3579 // This only works for expanding the current library call,
3580 // not another intrinsic. (E.g., don't use this for making an
3581 // arraycopy call inside of the copyOf intrinsic.)
3582 CallJavaNode*
3583 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3584 // When compiling the intrinsic method itself, do not use this technique.
3585 guarantee(callee() != C->method(), "cannot make slow-call to self");
3586
3587 ciMethod* method = callee();
3588 // ensure the JVMS we have will be correct for this call
3589 guarantee(method_id == method->intrinsic_id(), "must match");
3590
3591 const TypeFunc* tf = TypeFunc::make(method);
3592 int tfdc = tf->domain()->cnt();
3593 CallJavaNode* slow_call;
3594 if (is_static) {
3595 assert(!is_virtual, "");
3596 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3597 SharedRuntime::get_resolve_static_call_stub(),
3598 method, bci());
3599 } else if (is_virtual) {
3600 null_check_receiver(method);
3601 int vtable_index = methodOopDesc::invalid_vtable_index;
3602 if (UseInlineCaches) {
3603 // Suppress the vtable call
3604 } else {
3605 // hashCode and clone are not a miranda methods,
3606 // so the vtable index is fixed.
3607 // No need to use the linkResolver to get it.
3608 vtable_index = method->vtable_index();
3609 }
3610 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3611 SharedRuntime::get_resolve_virtual_call_stub(),
3612 method, vtable_index, bci());
3613 } else { // neither virtual nor static: opt_virtual
3614 null_check_receiver(method);
3615 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3616 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3617 method, bci());
3618 slow_call->set_optimized_virtual(true);
3619 }
3620 set_arguments_for_java_call(slow_call);
3621 set_edges_for_java_call(slow_call);
3622 return slow_call;
3623 }
3624
3625
3626 //------------------------------inline_native_hashcode--------------------
3627 // Build special case code for calls to hashCode on an object.
3628 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3629 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3630 assert(!(is_virtual && is_static), "either virtual, special, or static");
3631
3632 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3633
3634 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3635 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3636 TypeInt::INT);
3637 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3638 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3639 TypePtr::BOTTOM);
3640 Node* obj = NULL;
3641 if (!is_static) {
3642 // Check for hashing null object
3643 obj = null_check_receiver(callee());
3644 if (stopped()) return true; // unconditionally null
3645 result_reg->init_req(_null_path, top());
3646 result_val->init_req(_null_path, top());
3647 } else {
3648 // Do a null check, and return zero if null.
3649 // System.identityHashCode(null) == 0
3650 obj = argument(0);
3651 Node* null_ctl = top();
3652 obj = null_check_oop(obj, &null_ctl);
3653 result_reg->init_req(_null_path, null_ctl);
3654 result_val->init_req(_null_path, _gvn.intcon(0));
3655 }
3656
3657 // Unconditionally null? Then return right away.
3658 if (stopped()) {
3659 set_control( result_reg->in(_null_path) );
3660 if (!stopped())
3661 push( result_val ->in(_null_path) );
3662 return true;
3663 }
3664
3665 // After null check, get the object's klass.
3666 Node* obj_klass = load_object_klass(obj);
3667
3668 // This call may be virtual (invokevirtual) or bound (invokespecial).
3669 // For each case we generate slightly different code.
3670
3671 // We only go to the fast case code if we pass a number of guards. The
3672 // paths which do not pass are accumulated in the slow_region.
3673 RegionNode* slow_region = new (C, 1) RegionNode(1);
3674 record_for_igvn(slow_region);
3675
3676 // If this is a virtual call, we generate a funny guard. We pull out
3677 // the vtable entry corresponding to hashCode() from the target object.
3678 // If the target method which we are calling happens to be the native
3679 // Object hashCode() method, we pass the guard. We do not need this
3680 // guard for non-virtual calls -- the caller is known to be the native
3681 // Object hashCode().
3682 if (is_virtual) {
3683 generate_virtual_guard(obj_klass, slow_region);
3684 }
3685
3686 // Get the header out of the object, use LoadMarkNode when available
3687 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3688 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3689
3690 // Test the header to see if it is unlocked.
3691 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3692 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3693 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3694 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3695 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3696
3697 generate_slow_guard(test_unlocked, slow_region);
3698
3699 // Get the hash value and check to see that it has been properly assigned.
3700 // We depend on hash_mask being at most 32 bits and avoid the use of
3701 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3702 // vm: see markOop.hpp.
3703 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3704 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3705 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3706 // This hack lets the hash bits live anywhere in the mark object now, as long
3707 // as the shift drops the relevant bits into the low 32 bits. Note that
3708 // Java spec says that HashCode is an int so there's no point in capturing
3709 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3710 hshifted_header = ConvX2I(hshifted_header);
3711 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3712
3713 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3714 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3715 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3716
3717 generate_slow_guard(test_assigned, slow_region);
3718
3719 Node* init_mem = reset_memory();
3720 // fill in the rest of the null path:
3721 result_io ->init_req(_null_path, i_o());
3722 result_mem->init_req(_null_path, init_mem);
3723
3724 result_val->init_req(_fast_path, hash_val);
3725 result_reg->init_req(_fast_path, control());
3726 result_io ->init_req(_fast_path, i_o());
3727 result_mem->init_req(_fast_path, init_mem);
3728
3729 // Generate code for the slow case. We make a call to hashCode().
3730 set_control(_gvn.transform(slow_region));
3731 if (!stopped()) {
3732 // No need for PreserveJVMState, because we're using up the present state.
3733 set_all_memory(init_mem);
3734 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3735 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3736 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3737 Node* slow_result = set_results_for_java_call(slow_call);
3738 // this->control() comes from set_results_for_java_call
3739 result_reg->init_req(_slow_path, control());
3740 result_val->init_req(_slow_path, slow_result);
3741 result_io ->set_req(_slow_path, i_o());
3742 result_mem ->set_req(_slow_path, reset_memory());
3743 }
3744
3745 // Return the combined state.
3746 set_i_o( _gvn.transform(result_io) );
3747 set_all_memory( _gvn.transform(result_mem) );
3748 push_result(result_reg, result_val);
3749
3750 return true;
3751 }
3752
3753 //---------------------------inline_native_getClass----------------------------
3754 // Build special case code for calls to getClass on an object.
3755 bool LibraryCallKit::inline_native_getClass() {
3756 Node* obj = null_check_receiver(callee());
3757 if (stopped()) return true;
3758 push( load_mirror_from_klass(load_object_klass(obj)) );
3759 return true;
3760 }
3761
3762 //-----------------inline_native_Reflection_getCallerClass---------------------
3763 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3764 //
3765 // NOTE that this code must perform the same logic as
3766 // vframeStream::security_get_caller_frame in that it must skip
3767 // Method.invoke() and auxiliary frames.
3768
3769
3770
3771
3772 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3773 ciMethod* method = callee();
3774
3775 #ifndef PRODUCT
3776 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3777 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3778 }
3779 #endif
3780
3781 debug_only(int saved_sp = _sp);
3782
3783 // Argument words: (int depth)
3784 int nargs = 1;
3785
3786 _sp += nargs;
3787 Node* caller_depth_node = pop();
3788
3789 assert(saved_sp == _sp, "must have correct argument count");
3790
3791 // The depth value must be a constant in order for the runtime call
3792 // to be eliminated.
3793 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3794 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3795 #ifndef PRODUCT
3796 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3797 tty->print_cr(" Bailing out because caller depth was not a constant");
3798 }
3799 #endif
3800 return false;
3801 }
3802 // Note that the JVM state at this point does not include the
3803 // getCallerClass() frame which we are trying to inline. The
3804 // semantics of getCallerClass(), however, are that the "first"
3805 // frame is the getCallerClass() frame, so we subtract one from the
3806 // requested depth before continuing. We don't inline requests of
3807 // getCallerClass(0).
3808 int caller_depth = caller_depth_type->get_con() - 1;
3809 if (caller_depth < 0) {
3810 #ifndef PRODUCT
3811 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3812 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3813 }
3814 #endif
3815 return false;
3816 }
3817
3818 if (!jvms()->has_method()) {
3819 #ifndef PRODUCT
3820 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3821 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3822 }
3823 #endif
3824 return false;
3825 }
3826 int _depth = jvms()->depth(); // cache call chain depth
3827
3828 // Walk back up the JVM state to find the caller at the required
3829 // depth. NOTE that this code must perform the same logic as
3830 // vframeStream::security_get_caller_frame in that it must skip
3831 // Method.invoke() and auxiliary frames. Note also that depth is
3832 // 1-based (1 is the bottom of the inlining).
3833 int inlining_depth = _depth;
3834 JVMState* caller_jvms = NULL;
3835
3836 if (inlining_depth > 0) {
3837 caller_jvms = jvms();
3838 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3839 do {
3840 // The following if-tests should be performed in this order
3841 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3842 // Skip a Method.invoke() or auxiliary frame
3843 } else if (caller_depth > 0) {
3844 // Skip real frame
3845 --caller_depth;
3846 } else {
3847 // We're done: reached desired caller after skipping.
3848 break;
3849 }
3850 caller_jvms = caller_jvms->caller();
3851 --inlining_depth;
3852 } while (inlining_depth > 0);
3853 }
3854
3855 if (inlining_depth == 0) {
3856 #ifndef PRODUCT
3857 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3858 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3859 tty->print_cr(" JVM state at this point:");
3860 for (int i = _depth; i >= 1; i--) {
3861 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3862 }
3863 }
3864 #endif
3865 return false; // Reached end of inlining
3866 }
3867
3868 // Acquire method holder as java.lang.Class
3869 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3870 ciInstance* caller_mirror = caller_klass->java_mirror();
3871 // Push this as a constant
3872 push(makecon(TypeInstPtr::make(caller_mirror)));
3873 #ifndef PRODUCT
3874 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3875 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);
3876 tty->print_cr(" JVM state at this point:");
3877 for (int i = _depth; i >= 1; i--) {
3878 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3879 }
3880 }
3881 #endif
3882 return true;
3883 }
3884
3885 // Helper routine for above
3886 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3887 ciMethod* method = jvms->method();
3888
3889 // Is this the Method.invoke method itself?
3890 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3891 return true;
3892
3893 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3894 ciKlass* k = method->holder();
3895 if (k->is_instance_klass()) {
3896 ciInstanceKlass* ik = k->as_instance_klass();
3897 for (; ik != NULL; ik = ik->super()) {
3898 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3899 ik == env()->find_system_klass(ik->name())) {
3900 return true;
3901 }
3902 }
3903 }
3904 else if (method->is_method_handle_adapter()) {
3905 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3906 return true;
3907 }
3908
3909 return false;
3910 }
3911
3912 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3913 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3914 // computing it since there is no lookup field by name function in the
3915 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3916 // Using a static variable here is safe even if we have multiple compilation
3917 // threads because the offset is constant. At worst the same offset will be
3918 // computed and stored multiple
3919
3920 bool LibraryCallKit::inline_native_AtomicLong_get() {
3921 // Restore the stack and pop off the argument
3922 _sp+=1;
3923 Node *obj = pop();
3924
3925 // get the offset of the "value" field. Since the CI interfaces
3926 // does not provide a way to look up a field by name, we scan the bytecodes
3927 // to get the field index. We expect the first 2 instructions of the method
3928 // to be:
3929 // 0 aload_0
3930 // 1 getfield "value"
3931 ciMethod* method = callee();
3932 if (value_field_offset == -1)
3933 {
3934 ciField* value_field;
3935 ciBytecodeStream iter(method);
3936 Bytecodes::Code bc = iter.next();
3937
3938 if ((bc != Bytecodes::_aload_0) &&
3939 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3940 return false;
3941 bc = iter.next();
3942 if (bc != Bytecodes::_getfield)
3943 return false;
3944 bool ignore;
3945 value_field = iter.get_field(ignore);
3946 value_field_offset = value_field->offset_in_bytes();
3947 }
3948
3949 // Null check without removing any arguments.
3950 _sp++;
3951 obj = do_null_check(obj, T_OBJECT);
3952 _sp--;
3953 // Check for locking null object
3954 if (stopped()) return true;
3955
3956 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3957 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3958 int alias_idx = C->get_alias_index(adr_type);
3959
3960 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3961
3962 push_pair(result);
3963
3964 return true;
3965 }
3966
3967 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3968 // Restore the stack and pop off the arguments
3969 _sp+=5;
3970 Node *newVal = pop_pair();
3971 Node *oldVal = pop_pair();
3972 Node *obj = pop();
3973
3974 // we need the offset of the "value" field which was computed when
3975 // inlining the get() method. Give up if we don't have it.
3976 if (value_field_offset == -1)
3977 return false;
3978
3979 // Null check without removing any arguments.
3980 _sp+=5;
3981 obj = do_null_check(obj, T_OBJECT);
3982 _sp-=5;
3983 // Check for locking null object
3984 if (stopped()) return true;
3985
3986 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3987 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3988 int alias_idx = C->get_alias_index(adr_type);
3989
3990 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
3991 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
3992 set_memory(store_proj, alias_idx);
3993 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
3994
3995 Node *result;
3996 // CMove node is not used to be able fold a possible check code
3997 // after attemptUpdate() call. This code could be transformed
3998 // into CMove node by loop optimizations.
3999 {
4000 RegionNode *r = new (C, 3) RegionNode(3);
4001 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4002
4003 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4004 Node *iftrue = opt_iff(r, iff);
4005 r->init_req(1, iftrue);
4006 result->init_req(1, intcon(1));
4007 result->init_req(2, intcon(0));
4008
4009 set_control(_gvn.transform(r));
4010 record_for_igvn(r);
4011
4012 C->set_has_split_ifs(true); // Has chance for split-if optimization
4013 }
4014
4015 push(_gvn.transform(result));
4016 return true;
4017 }
4018
4019 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4020 // restore the arguments
4021 _sp += arg_size();
4022
4023 switch (id) {
4024 case vmIntrinsics::_floatToRawIntBits:
4025 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4026 break;
4027
4028 case vmIntrinsics::_intBitsToFloat:
4029 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4030 break;
4031
4032 case vmIntrinsics::_doubleToRawLongBits:
4033 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4034 break;
4035
4036 case vmIntrinsics::_longBitsToDouble:
4037 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4038 break;
4039
4040 case vmIntrinsics::_doubleToLongBits: {
4041 Node* value = pop_pair();
4042
4043 // two paths (plus control) merge in a wood
4044 RegionNode *r = new (C, 3) RegionNode(3);
4045 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4046
4047 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4048 // Build the boolean node
4049 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4050
4051 // Branch either way.
4052 // NaN case is less traveled, which makes all the difference.
4053 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4054 Node *opt_isnan = _gvn.transform(ifisnan);
4055 assert( opt_isnan->is_If(), "Expect an IfNode");
4056 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4057 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4058
4059 set_control(iftrue);
4060
4061 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4062 Node *slow_result = longcon(nan_bits); // return NaN
4063 phi->init_req(1, _gvn.transform( slow_result ));
4064 r->init_req(1, iftrue);
4065
4066 // Else fall through
4067 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4068 set_control(iffalse);
4069
4070 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4071 r->init_req(2, iffalse);
4072
4073 // Post merge
4074 set_control(_gvn.transform(r));
4075 record_for_igvn(r);
4076
4077 Node* result = _gvn.transform(phi);
4078 assert(result->bottom_type()->isa_long(), "must be");
4079 push_pair(result);
4080
4081 C->set_has_split_ifs(true); // Has chance for split-if optimization
4082
4083 break;
4084 }
4085
4086 case vmIntrinsics::_floatToIntBits: {
4087 Node* value = pop();
4088
4089 // two paths (plus control) merge in a wood
4090 RegionNode *r = new (C, 3) RegionNode(3);
4091 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4092
4093 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4094 // Build the boolean node
4095 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4096
4097 // Branch either way.
4098 // NaN case is less traveled, which makes all the difference.
4099 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4100 Node *opt_isnan = _gvn.transform(ifisnan);
4101 assert( opt_isnan->is_If(), "Expect an IfNode");
4102 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4103 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4104
4105 set_control(iftrue);
4106
4107 static const jint nan_bits = 0x7fc00000;
4108 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4109 phi->init_req(1, _gvn.transform( slow_result ));
4110 r->init_req(1, iftrue);
4111
4112 // Else fall through
4113 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4114 set_control(iffalse);
4115
4116 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4117 r->init_req(2, iffalse);
4118
4119 // Post merge
4120 set_control(_gvn.transform(r));
4121 record_for_igvn(r);
4122
4123 Node* result = _gvn.transform(phi);
4124 assert(result->bottom_type()->isa_int(), "must be");
4125 push(result);
4126
4127 C->set_has_split_ifs(true); // Has chance for split-if optimization
4128
4129 break;
4130 }
4131
4132 default:
4133 ShouldNotReachHere();
4134 }
4135
4136 return true;
4137 }
4138
4139 #ifdef _LP64
4140 #define XTOP ,top() /*additional argument*/
4141 #else //_LP64
4142 #define XTOP /*no additional argument*/
4143 #endif //_LP64
4144
4145 //----------------------inline_unsafe_copyMemory-------------------------
4146 bool LibraryCallKit::inline_unsafe_copyMemory() {
4147 if (callee()->is_static()) return false; // caller must have the capability!
4148 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4149 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4150 null_check_receiver(callee()); // check then ignore argument(0)
4151 if (stopped()) return true;
4152
4153 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4154
4155 Node* src_ptr = argument(1);
4156 Node* src_off = ConvL2X(argument(2));
4157 assert(argument(3)->is_top(), "2nd half of long");
4158 Node* dst_ptr = argument(4);
4159 Node* dst_off = ConvL2X(argument(5));
4160 assert(argument(6)->is_top(), "2nd half of long");
4161 Node* size = ConvL2X(argument(7));
4162 assert(argument(8)->is_top(), "2nd half of long");
4163
4164 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4165 "fieldOffset must be byte-scaled");
4166
4167 Node* src = make_unsafe_address(src_ptr, src_off);
4168 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4169
4170 // Conservatively insert a memory barrier on all memory slices.
4171 // Do not let writes of the copy source or destination float below the copy.
4172 insert_mem_bar(Op_MemBarCPUOrder);
4173
4174 // Call it. Note that the length argument is not scaled.
4175 make_runtime_call(RC_LEAF|RC_NO_FP,
4176 OptoRuntime::fast_arraycopy_Type(),
4177 StubRoutines::unsafe_arraycopy(),
4178 "unsafe_arraycopy",
4179 TypeRawPtr::BOTTOM,
4180 src, dst, size XTOP);
4181
4182 // Do not let reads of the copy destination float above the copy.
4183 insert_mem_bar(Op_MemBarCPUOrder);
4184
4185 return true;
4186 }
4187
4188 //------------------------clone_coping-----------------------------------
4189 // Helper function for inline_native_clone.
4190 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4191 assert(obj_size != NULL, "");
4192 Node* raw_obj = alloc_obj->in(1);
4193 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4194
4195 AllocateNode* alloc = NULL;
4196 if (ReduceBulkZeroing) {
4197 // We will be completely responsible for initializing this object -
4198 // mark Initialize node as complete.
4199 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4200 // The object was just allocated - there should be no any stores!
4201 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4202 // Mark as complete_with_arraycopy so that on AllocateNode
4203 // expansion, we know this AllocateNode is initialized by an array
4204 // copy and a StoreStore barrier exists after the array copy.
4205 alloc->initialization()->set_complete_with_arraycopy();
4206 }
4207
4208 // Copy the fastest available way.
4209 // TODO: generate fields copies for small objects instead.
4210 Node* src = obj;
4211 Node* dest = alloc_obj;
4212 Node* size = _gvn.transform(obj_size);
4213
4214 // Exclude the header but include array length to copy by 8 bytes words.
4215 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4216 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4217 instanceOopDesc::base_offset_in_bytes();
4218 // base_off:
4219 // 8 - 32-bit VM
4220 // 12 - 64-bit VM, compressed oops
4221 // 16 - 64-bit VM, normal oops
4222 if (base_off % BytesPerLong != 0) {
4223 assert(UseCompressedOops, "");
4224 if (is_array) {
4225 // Exclude length to copy by 8 bytes words.
4226 base_off += sizeof(int);
4227 } else {
4228 // Include klass to copy by 8 bytes words.
4229 base_off = instanceOopDesc::klass_offset_in_bytes();
4230 }
4231 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4232 }
4233 src = basic_plus_adr(src, base_off);
4234 dest = basic_plus_adr(dest, base_off);
4235
4236 // Compute the length also, if needed:
4237 Node* countx = size;
4238 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4239 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4240
4241 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4242 bool disjoint_bases = true;
4243 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4244 src, NULL, dest, NULL, countx,
4245 /*dest_uninitialized*/true);
4246
4247 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4248 if (card_mark) {
4249 assert(!is_array, "");
4250 // Put in store barrier for any and all oops we are sticking
4251 // into this object. (We could avoid this if we could prove
4252 // that the object type contains no oop fields at all.)
4253 Node* no_particular_value = NULL;
4254 Node* no_particular_field = NULL;
4255 int raw_adr_idx = Compile::AliasIdxRaw;
4256 post_barrier(control(),
4257 memory(raw_adr_type),
4258 alloc_obj,
4259 no_particular_field,
4260 raw_adr_idx,
4261 no_particular_value,
4262 T_OBJECT,
4263 false);
4264 }
4265
4266 // Do not let reads from the cloned object float above the arraycopy.
4267 if (alloc != NULL) {
4268 // Do not let stores that initialize this object be reordered with
4269 // a subsequent store that would make this object accessible by
4270 // other threads.
4271 // Record what AllocateNode this StoreStore protects so that
4272 // escape analysis can go from the MemBarStoreStoreNode to the
4273 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4274 // based on the escape status of the AllocateNode.
4275 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4276 } else {
4277 insert_mem_bar(Op_MemBarCPUOrder);
4278 }
4279 }
4280
4281 //------------------------inline_native_clone----------------------------
4282 // Here are the simple edge cases:
4283 // null receiver => normal trap
4284 // virtual and clone was overridden => slow path to out-of-line clone
4285 // not cloneable or finalizer => slow path to out-of-line Object.clone
4286 //
4287 // The general case has two steps, allocation and copying.
4288 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4289 //
4290 // Copying also has two cases, oop arrays and everything else.
4291 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4292 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4293 //
4294 // These steps fold up nicely if and when the cloned object's klass
4295 // can be sharply typed as an object array, a type array, or an instance.
4296 //
4297 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4298 int nargs = 1;
4299 PhiNode* result_val;
4300
4301 //set the original stack and the reexecute bit for the interpreter to reexecute
4302 //the bytecode that invokes Object.clone if deoptimization happens
4303 { PreserveReexecuteState preexecs(this);
4304 jvms()->set_should_reexecute(true);
4305
4306 //null_check_receiver will adjust _sp (push and pop)
4307 Node* obj = null_check_receiver(callee());
4308 if (stopped()) return true;
4309
4310 _sp += nargs;
4311
4312 Node* obj_klass = load_object_klass(obj);
4313 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4314 const TypeOopPtr* toop = ((tklass != NULL)
4315 ? tklass->as_instance_type()
4316 : TypeInstPtr::NOTNULL);
4317
4318 // Conservatively insert a memory barrier on all memory slices.
4319 // Do not let writes into the original float below the clone.
4320 insert_mem_bar(Op_MemBarCPUOrder);
4321
4322 // paths into result_reg:
4323 enum {
4324 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4325 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4326 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4327 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4328 PATH_LIMIT
4329 };
4330 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4331 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4332 TypeInstPtr::NOTNULL);
4333 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4334 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4335 TypePtr::BOTTOM);
4336 record_for_igvn(result_reg);
4337
4338 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4339 int raw_adr_idx = Compile::AliasIdxRaw;
4340
4341 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4342 if (array_ctl != NULL) {
4343 // It's an array.
4344 PreserveJVMState pjvms(this);
4345 set_control(array_ctl);
4346 Node* obj_length = load_array_length(obj);
4347 Node* obj_size = NULL;
4348 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4349
4350 if (!use_ReduceInitialCardMarks()) {
4351 // If it is an oop array, it requires very special treatment,
4352 // because card marking is required on each card of the array.
4353 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4354 if (is_obja != NULL) {
4355 PreserveJVMState pjvms2(this);
4356 set_control(is_obja);
4357 // Generate a direct call to the right arraycopy function(s).
4358 bool disjoint_bases = true;
4359 bool length_never_negative = true;
4360 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4361 obj, intcon(0), alloc_obj, intcon(0),
4362 obj_length,
4363 disjoint_bases, length_never_negative);
4364 result_reg->init_req(_objArray_path, control());
4365 result_val->init_req(_objArray_path, alloc_obj);
4366 result_i_o ->set_req(_objArray_path, i_o());
4367 result_mem ->set_req(_objArray_path, reset_memory());
4368 }
4369 }
4370 // Otherwise, there are no card marks to worry about.
4371 // (We can dispense with card marks if we know the allocation
4372 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4373 // causes the non-eden paths to take compensating steps to
4374 // simulate a fresh allocation, so that no further
4375 // card marks are required in compiled code to initialize
4376 // the object.)
4377
4378 if (!stopped()) {
4379 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4380
4381 // Present the results of the copy.
4382 result_reg->init_req(_array_path, control());
4383 result_val->init_req(_array_path, alloc_obj);
4384 result_i_o ->set_req(_array_path, i_o());
4385 result_mem ->set_req(_array_path, reset_memory());
4386 }
4387 }
4388
4389 // We only go to the instance fast case code if we pass a number of guards.
4390 // The paths which do not pass are accumulated in the slow_region.
4391 RegionNode* slow_region = new (C, 1) RegionNode(1);
4392 record_for_igvn(slow_region);
4393 if (!stopped()) {
4394 // It's an instance (we did array above). Make the slow-path tests.
4395 // If this is a virtual call, we generate a funny guard. We grab
4396 // the vtable entry corresponding to clone() from the target object.
4397 // If the target method which we are calling happens to be the
4398 // Object clone() method, we pass the guard. We do not need this
4399 // guard for non-virtual calls; the caller is known to be the native
4400 // Object clone().
4401 if (is_virtual) {
4402 generate_virtual_guard(obj_klass, slow_region);
4403 }
4404
4405 // The object must be cloneable and must not have a finalizer.
4406 // Both of these conditions may be checked in a single test.
4407 // We could optimize the cloneable test further, but we don't care.
4408 generate_access_flags_guard(obj_klass,
4409 // Test both conditions:
4410 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4411 // Must be cloneable but not finalizer:
4412 JVM_ACC_IS_CLONEABLE,
4413 slow_region);
4414 }
4415
4416 if (!stopped()) {
4417 // It's an instance, and it passed the slow-path tests.
4418 PreserveJVMState pjvms(this);
4419 Node* obj_size = NULL;
4420 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4421
4422 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4423
4424 // Present the results of the slow call.
4425 result_reg->init_req(_instance_path, control());
4426 result_val->init_req(_instance_path, alloc_obj);
4427 result_i_o ->set_req(_instance_path, i_o());
4428 result_mem ->set_req(_instance_path, reset_memory());
4429 }
4430
4431 // Generate code for the slow case. We make a call to clone().
4432 set_control(_gvn.transform(slow_region));
4433 if (!stopped()) {
4434 PreserveJVMState pjvms(this);
4435 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4436 Node* slow_result = set_results_for_java_call(slow_call);
4437 // this->control() comes from set_results_for_java_call
4438 result_reg->init_req(_slow_path, control());
4439 result_val->init_req(_slow_path, slow_result);
4440 result_i_o ->set_req(_slow_path, i_o());
4441 result_mem ->set_req(_slow_path, reset_memory());
4442 }
4443
4444 // Return the combined state.
4445 set_control( _gvn.transform(result_reg) );
4446 set_i_o( _gvn.transform(result_i_o) );
4447 set_all_memory( _gvn.transform(result_mem) );
4448 } //original reexecute and sp are set back here
4449
4450 push(_gvn.transform(result_val));
4451
4452 return true;
4453 }
4454
4455 //------------------------------basictype2arraycopy----------------------------
4456 address LibraryCallKit::basictype2arraycopy(BasicType t,
4457 Node* src_offset,
4458 Node* dest_offset,
4459 bool disjoint_bases,
4460 const char* &name,
4461 bool dest_uninitialized) {
4462 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4463 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4464
4465 bool aligned = false;
4466 bool disjoint = disjoint_bases;
4467
4468 // if the offsets are the same, we can treat the memory regions as
4469 // disjoint, because either the memory regions are in different arrays,
4470 // or they are identical (which we can treat as disjoint.) We can also
4471 // treat a copy with a destination index less that the source index
4472 // as disjoint since a low->high copy will work correctly in this case.
4473 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4474 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4475 // both indices are constants
4476 int s_offs = src_offset_inttype->get_con();
4477 int d_offs = dest_offset_inttype->get_con();
4478 int element_size = type2aelembytes(t);
4479 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4480 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4481 if (s_offs >= d_offs) disjoint = true;
4482 } else if (src_offset == dest_offset && src_offset != NULL) {
4483 // This can occur if the offsets are identical non-constants.
4484 disjoint = true;
4485 }
4486
4487 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4488 }
4489
4490
4491 //------------------------------inline_arraycopy-----------------------
4492 bool LibraryCallKit::inline_arraycopy() {
4493 // Restore the stack and pop off the arguments.
4494 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4495 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4496
4497 Node *src = argument(0);
4498 Node *src_offset = argument(1);
4499 Node *dest = argument(2);
4500 Node *dest_offset = argument(3);
4501 Node *length = argument(4);
4502
4503 // Compile time checks. If any of these checks cannot be verified at compile time,
4504 // we do not make a fast path for this call. Instead, we let the call remain as it
4505 // is. The checks we choose to mandate at compile time are:
4506 //
4507 // (1) src and dest are arrays.
4508 const Type* src_type = src->Value(&_gvn);
4509 const Type* dest_type = dest->Value(&_gvn);
4510 const TypeAryPtr* top_src = src_type->isa_aryptr();
4511 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4512 if (top_src == NULL || top_src->klass() == NULL ||
4513 top_dest == NULL || top_dest->klass() == NULL) {
4514 // Conservatively insert a memory barrier on all memory slices.
4515 // Do not let writes into the source float below the arraycopy.
4516 insert_mem_bar(Op_MemBarCPUOrder);
4517
4518 // Call StubRoutines::generic_arraycopy stub.
4519 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4520 src, src_offset, dest, dest_offset, length);
4521
4522 // Do not let reads from the destination float above the arraycopy.
4523 // Since we cannot type the arrays, we don't know which slices
4524 // might be affected. We could restrict this barrier only to those
4525 // memory slices which pertain to array elements--but don't bother.
4526 if (!InsertMemBarAfterArraycopy)
4527 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4528 insert_mem_bar(Op_MemBarCPUOrder);
4529 return true;
4530 }
4531
4532 // (2) src and dest arrays must have elements of the same BasicType
4533 // Figure out the size and type of the elements we will be copying.
4534 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4535 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4536 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4537 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4538
4539 if (src_elem != dest_elem || dest_elem == T_VOID) {
4540 // The component types are not the same or are not recognized. Punt.
4541 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4542 generate_slow_arraycopy(TypePtr::BOTTOM,
4543 src, src_offset, dest, dest_offset, length,
4544 /*dest_uninitialized*/false);
4545 return true;
4546 }
4547
4548 //---------------------------------------------------------------------------
4549 // We will make a fast path for this call to arraycopy.
4550
4551 // We have the following tests left to perform:
4552 //
4553 // (3) src and dest must not be null.
4554 // (4) src_offset must not be negative.
4555 // (5) dest_offset must not be negative.
4556 // (6) length must not be negative.
4557 // (7) src_offset + length must not exceed length of src.
4558 // (8) dest_offset + length must not exceed length of dest.
4559 // (9) each element of an oop array must be assignable
4560
4561 RegionNode* slow_region = new (C, 1) RegionNode(1);
4562 record_for_igvn(slow_region);
4563
4564 // (3) operands must not be null
4565 // We currently perform our null checks with the do_null_check routine.
4566 // This means that the null exceptions will be reported in the caller
4567 // rather than (correctly) reported inside of the native arraycopy call.
4568 // This should be corrected, given time. We do our null check with the
4569 // stack pointer restored.
4570 _sp += nargs;
4571 src = do_null_check(src, T_ARRAY);
4572 dest = do_null_check(dest, T_ARRAY);
4573 _sp -= nargs;
4574
4575 // (4) src_offset must not be negative.
4576 generate_negative_guard(src_offset, slow_region);
4577
4578 // (5) dest_offset must not be negative.
4579 generate_negative_guard(dest_offset, slow_region);
4580
4581 // (6) length must not be negative (moved to generate_arraycopy()).
4582 // generate_negative_guard(length, slow_region);
4583
4584 // (7) src_offset + length must not exceed length of src.
4585 generate_limit_guard(src_offset, length,
4586 load_array_length(src),
4587 slow_region);
4588
4589 // (8) dest_offset + length must not exceed length of dest.
4590 generate_limit_guard(dest_offset, length,
4591 load_array_length(dest),
4592 slow_region);
4593
4594 // (9) each element of an oop array must be assignable
4595 // The generate_arraycopy subroutine checks this.
4596
4597 // This is where the memory effects are placed:
4598 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4599 generate_arraycopy(adr_type, dest_elem,
4600 src, src_offset, dest, dest_offset, length,
4601 false, false, slow_region);
4602
4603 return true;
4604 }
4605
4606 //-----------------------------generate_arraycopy----------------------
4607 // Generate an optimized call to arraycopy.
4608 // Caller must guard against non-arrays.
4609 // Caller must determine a common array basic-type for both arrays.
4610 // Caller must validate offsets against array bounds.
4611 // The slow_region has already collected guard failure paths
4612 // (such as out of bounds length or non-conformable array types).
4613 // The generated code has this shape, in general:
4614 //
4615 // if (length == 0) return // via zero_path
4616 // slowval = -1
4617 // if (types unknown) {
4618 // slowval = call generic copy loop
4619 // if (slowval == 0) return // via checked_path
4620 // } else if (indexes in bounds) {
4621 // if ((is object array) && !(array type check)) {
4622 // slowval = call checked copy loop
4623 // if (slowval == 0) return // via checked_path
4624 // } else {
4625 // call bulk copy loop
4626 // return // via fast_path
4627 // }
4628 // }
4629 // // adjust params for remaining work:
4630 // if (slowval != -1) {
4631 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4632 // }
4633 // slow_region:
4634 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4635 // return // via slow_call_path
4636 //
4637 // This routine is used from several intrinsics: System.arraycopy,
4638 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4639 //
4640 void
4641 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4642 BasicType basic_elem_type,
4643 Node* src, Node* src_offset,
4644 Node* dest, Node* dest_offset,
4645 Node* copy_length,
4646 bool disjoint_bases,
4647 bool length_never_negative,
4648 RegionNode* slow_region) {
4649
4650 if (slow_region == NULL) {
4651 slow_region = new(C,1) RegionNode(1);
4652 record_for_igvn(slow_region);
4653 }
4654
4655 Node* original_dest = dest;
4656 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4657 bool dest_uninitialized = false;
4658
4659 // See if this is the initialization of a newly-allocated array.
4660 // If so, we will take responsibility here for initializing it to zero.
4661 // (Note: Because tightly_coupled_allocation performs checks on the
4662 // out-edges of the dest, we need to avoid making derived pointers
4663 // from it until we have checked its uses.)
4664 if (ReduceBulkZeroing
4665 && !ZeroTLAB // pointless if already zeroed
4666 && basic_elem_type != T_CONFLICT // avoid corner case
4667 && !src->eqv_uncast(dest)
4668 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4669 != NULL)
4670 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4671 && alloc->maybe_set_complete(&_gvn)) {
4672 // "You break it, you buy it."
4673 InitializeNode* init = alloc->initialization();
4674 assert(init->is_complete(), "we just did this");
4675 init->set_complete_with_arraycopy();
4676 assert(dest->is_CheckCastPP(), "sanity");
4677 assert(dest->in(0)->in(0) == init, "dest pinned");
4678 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4679 // From this point on, every exit path is responsible for
4680 // initializing any non-copied parts of the object to zero.
4681 // Also, if this flag is set we make sure that arraycopy interacts properly
4682 // with G1, eliding pre-barriers. See CR 6627983.
4683 dest_uninitialized = true;
4684 } else {
4685 // No zeroing elimination here.
4686 alloc = NULL;
4687 //original_dest = dest;
4688 //dest_uninitialized = false;
4689 }
4690
4691 // Results are placed here:
4692 enum { fast_path = 1, // normal void-returning assembly stub
4693 checked_path = 2, // special assembly stub with cleanup
4694 slow_call_path = 3, // something went wrong; call the VM
4695 zero_path = 4, // bypass when length of copy is zero
4696 bcopy_path = 5, // copy primitive array by 64-bit blocks
4697 PATH_LIMIT = 6
4698 };
4699 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4700 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4701 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4702 record_for_igvn(result_region);
4703 _gvn.set_type_bottom(result_i_o);
4704 _gvn.set_type_bottom(result_memory);
4705 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4706
4707 // The slow_control path:
4708 Node* slow_control;
4709 Node* slow_i_o = i_o();
4710 Node* slow_mem = memory(adr_type);
4711 debug_only(slow_control = (Node*) badAddress);
4712
4713 // Checked control path:
4714 Node* checked_control = top();
4715 Node* checked_mem = NULL;
4716 Node* checked_i_o = NULL;
4717 Node* checked_value = NULL;
4718
4719 if (basic_elem_type == T_CONFLICT) {
4720 assert(!dest_uninitialized, "");
4721 Node* cv = generate_generic_arraycopy(adr_type,
4722 src, src_offset, dest, dest_offset,
4723 copy_length, dest_uninitialized);
4724 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4725 checked_control = control();
4726 checked_i_o = i_o();
4727 checked_mem = memory(adr_type);
4728 checked_value = cv;
4729 set_control(top()); // no fast path
4730 }
4731
4732 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4733 if (not_pos != NULL) {
4734 PreserveJVMState pjvms(this);
4735 set_control(not_pos);
4736
4737 // (6) length must not be negative.
4738 if (!length_never_negative) {
4739 generate_negative_guard(copy_length, slow_region);
4740 }
4741
4742 // copy_length is 0.
4743 if (!stopped() && dest_uninitialized) {
4744 Node* dest_length = alloc->in(AllocateNode::ALength);
4745 if (copy_length->eqv_uncast(dest_length)
4746 || _gvn.find_int_con(dest_length, 1) <= 0) {
4747 // There is no zeroing to do. No need for a secondary raw memory barrier.
4748 } else {
4749 // Clear the whole thing since there are no source elements to copy.
4750 generate_clear_array(adr_type, dest, basic_elem_type,
4751 intcon(0), NULL,
4752 alloc->in(AllocateNode::AllocSize));
4753 // Use a secondary InitializeNode as raw memory barrier.
4754 // Currently it is needed only on this path since other
4755 // paths have stub or runtime calls as raw memory barriers.
4756 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4757 Compile::AliasIdxRaw,
4758 top())->as_Initialize();
4759 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4760 }
4761 }
4762
4763 // Present the results of the fast call.
4764 result_region->init_req(zero_path, control());
4765 result_i_o ->init_req(zero_path, i_o());
4766 result_memory->init_req(zero_path, memory(adr_type));
4767 }
4768
4769 if (!stopped() && dest_uninitialized) {
4770 // We have to initialize the *uncopied* part of the array to zero.
4771 // The copy destination is the slice dest[off..off+len]. The other slices
4772 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4773 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4774 Node* dest_length = alloc->in(AllocateNode::ALength);
4775 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4776 copy_length) );
4777
4778 // If there is a head section that needs zeroing, do it now.
4779 if (find_int_con(dest_offset, -1) != 0) {
4780 generate_clear_array(adr_type, dest, basic_elem_type,
4781 intcon(0), dest_offset,
4782 NULL);
4783 }
4784
4785 // Next, perform a dynamic check on the tail length.
4786 // It is often zero, and we can win big if we prove this.
4787 // There are two wins: Avoid generating the ClearArray
4788 // with its attendant messy index arithmetic, and upgrade
4789 // the copy to a more hardware-friendly word size of 64 bits.
4790 Node* tail_ctl = NULL;
4791 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4792 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4793 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4794 tail_ctl = generate_slow_guard(bol_lt, NULL);
4795 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4796 }
4797
4798 // At this point, let's assume there is no tail.
4799 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4800 // There is no tail. Try an upgrade to a 64-bit copy.
4801 bool didit = false;
4802 { PreserveJVMState pjvms(this);
4803 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4804 src, src_offset, dest, dest_offset,
4805 dest_size, dest_uninitialized);
4806 if (didit) {
4807 // Present the results of the block-copying fast call.
4808 result_region->init_req(bcopy_path, control());
4809 result_i_o ->init_req(bcopy_path, i_o());
4810 result_memory->init_req(bcopy_path, memory(adr_type));
4811 }
4812 }
4813 if (didit)
4814 set_control(top()); // no regular fast path
4815 }
4816
4817 // Clear the tail, if any.
4818 if (tail_ctl != NULL) {
4819 Node* notail_ctl = stopped() ? NULL : control();
4820 set_control(tail_ctl);
4821 if (notail_ctl == NULL) {
4822 generate_clear_array(adr_type, dest, basic_elem_type,
4823 dest_tail, NULL,
4824 dest_size);
4825 } else {
4826 // Make a local merge.
4827 Node* done_ctl = new(C,3) RegionNode(3);
4828 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4829 done_ctl->init_req(1, notail_ctl);
4830 done_mem->init_req(1, memory(adr_type));
4831 generate_clear_array(adr_type, dest, basic_elem_type,
4832 dest_tail, NULL,
4833 dest_size);
4834 done_ctl->init_req(2, control());
4835 done_mem->init_req(2, memory(adr_type));
4836 set_control( _gvn.transform(done_ctl) );
4837 set_memory( _gvn.transform(done_mem), adr_type );
4838 }
4839 }
4840 }
4841
4842 BasicType copy_type = basic_elem_type;
4843 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4844 if (!stopped() && copy_type == T_OBJECT) {
4845 // If src and dest have compatible element types, we can copy bits.
4846 // Types S[] and D[] are compatible if D is a supertype of S.
4847 //
4848 // If they are not, we will use checked_oop_disjoint_arraycopy,
4849 // which performs a fast optimistic per-oop check, and backs off
4850 // further to JVM_ArrayCopy on the first per-oop check that fails.
4851 // (Actually, we don't move raw bits only; the GC requires card marks.)
4852
4853 // Get the klassOop for both src and dest
4854 Node* src_klass = load_object_klass(src);
4855 Node* dest_klass = load_object_klass(dest);
4856
4857 // Generate the subtype check.
4858 // This might fold up statically, or then again it might not.
4859 //
4860 // Non-static example: Copying List<String>.elements to a new String[].
4861 // The backing store for a List<String> is always an Object[],
4862 // but its elements are always type String, if the generic types
4863 // are correct at the source level.
4864 //
4865 // Test S[] against D[], not S against D, because (probably)
4866 // the secondary supertype cache is less busy for S[] than S.
4867 // This usually only matters when D is an interface.
4868 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4869 // Plug failing path into checked_oop_disjoint_arraycopy
4870 if (not_subtype_ctrl != top()) {
4871 PreserveJVMState pjvms(this);
4872 set_control(not_subtype_ctrl);
4873 // (At this point we can assume disjoint_bases, since types differ.)
4874 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4875 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4876 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4877 Node* dest_elem_klass = _gvn.transform(n1);
4878 Node* cv = generate_checkcast_arraycopy(adr_type,
4879 dest_elem_klass,
4880 src, src_offset, dest, dest_offset,
4881 ConvI2X(copy_length), dest_uninitialized);
4882 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4883 checked_control = control();
4884 checked_i_o = i_o();
4885 checked_mem = memory(adr_type);
4886 checked_value = cv;
4887 }
4888 // At this point we know we do not need type checks on oop stores.
4889
4890 // Let's see if we need card marks:
4891 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4892 // If we do not need card marks, copy using the jint or jlong stub.
4893 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4894 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4895 "sizes agree");
4896 }
4897 }
4898
4899 if (!stopped()) {
4900 // Generate the fast path, if possible.
4901 PreserveJVMState pjvms(this);
4902 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4903 src, src_offset, dest, dest_offset,
4904 ConvI2X(copy_length), dest_uninitialized);
4905
4906 // Present the results of the fast call.
4907 result_region->init_req(fast_path, control());
4908 result_i_o ->init_req(fast_path, i_o());
4909 result_memory->init_req(fast_path, memory(adr_type));
4910 }
4911
4912 // Here are all the slow paths up to this point, in one bundle:
4913 slow_control = top();
4914 if (slow_region != NULL)
4915 slow_control = _gvn.transform(slow_region);
4916 debug_only(slow_region = (RegionNode*)badAddress);
4917
4918 set_control(checked_control);
4919 if (!stopped()) {
4920 // Clean up after the checked call.
4921 // The returned value is either 0 or -1^K,
4922 // where K = number of partially transferred array elements.
4923 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4924 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4925 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4926
4927 // If it is 0, we are done, so transfer to the end.
4928 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4929 result_region->init_req(checked_path, checks_done);
4930 result_i_o ->init_req(checked_path, checked_i_o);
4931 result_memory->init_req(checked_path, checked_mem);
4932
4933 // If it is not zero, merge into the slow call.
4934 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4935 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4936 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4937 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4938 record_for_igvn(slow_reg2);
4939 slow_reg2 ->init_req(1, slow_control);
4940 slow_i_o2 ->init_req(1, slow_i_o);
4941 slow_mem2 ->init_req(1, slow_mem);
4942 slow_reg2 ->init_req(2, control());
4943 slow_i_o2 ->init_req(2, checked_i_o);
4944 slow_mem2 ->init_req(2, checked_mem);
4945
4946 slow_control = _gvn.transform(slow_reg2);
4947 slow_i_o = _gvn.transform(slow_i_o2);
4948 slow_mem = _gvn.transform(slow_mem2);
4949
4950 if (alloc != NULL) {
4951 // We'll restart from the very beginning, after zeroing the whole thing.
4952 // This can cause double writes, but that's OK since dest is brand new.
4953 // So we ignore the low 31 bits of the value returned from the stub.
4954 } else {
4955 // We must continue the copy exactly where it failed, or else
4956 // another thread might see the wrong number of writes to dest.
4957 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4958 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4959 slow_offset->init_req(1, intcon(0));
4960 slow_offset->init_req(2, checked_offset);
4961 slow_offset = _gvn.transform(slow_offset);
4962
4963 // Adjust the arguments by the conditionally incoming offset.
4964 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4965 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4966 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4967
4968 // Tweak the node variables to adjust the code produced below:
4969 src_offset = src_off_plus;
4970 dest_offset = dest_off_plus;
4971 copy_length = length_minus;
4972 }
4973 }
4974
4975 set_control(slow_control);
4976 if (!stopped()) {
4977 // Generate the slow path, if needed.
4978 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4979
4980 set_memory(slow_mem, adr_type);
4981 set_i_o(slow_i_o);
4982
4983 if (dest_uninitialized) {
4984 generate_clear_array(adr_type, dest, basic_elem_type,
4985 intcon(0), NULL,
4986 alloc->in(AllocateNode::AllocSize));
4987 }
4988
4989 generate_slow_arraycopy(adr_type,
4990 src, src_offset, dest, dest_offset,
4991 copy_length, /*dest_uninitialized*/false);
4992
4993 result_region->init_req(slow_call_path, control());
4994 result_i_o ->init_req(slow_call_path, i_o());
4995 result_memory->init_req(slow_call_path, memory(adr_type));
4996 }
4997
4998 // Remove unused edges.
4999 for (uint i = 1; i < result_region->req(); i++) {
5000 if (result_region->in(i) == NULL)
5001 result_region->init_req(i, top());
5002 }
5003
5004 // Finished; return the combined state.
5005 set_control( _gvn.transform(result_region) );
5006 set_i_o( _gvn.transform(result_i_o) );
5007 set_memory( _gvn.transform(result_memory), adr_type );
5008
5009 // The memory edges above are precise in order to model effects around
5010 // array copies accurately to allow value numbering of field loads around
5011 // arraycopy. Such field loads, both before and after, are common in Java
5012 // collections and similar classes involving header/array data structures.
5013 //
5014 // But with low number of register or when some registers are used or killed
5015 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5016 // The next memory barrier is added to avoid it. If the arraycopy can be
5017 // optimized away (which it can, sometimes) then we can manually remove
5018 // the membar also.
5019 //
5020 // Do not let reads from the cloned object float above the arraycopy.
5021 if (alloc != NULL) {
5022 // Do not let stores that initialize this object be reordered with
5023 // a subsequent store that would make this object accessible by
5024 // other threads.
5025 // Record what AllocateNode this StoreStore protects so that
5026 // escape analysis can go from the MemBarStoreStoreNode to the
5027 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5028 // based on the escape status of the AllocateNode.
5029 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5030 } else if (InsertMemBarAfterArraycopy)
5031 insert_mem_bar(Op_MemBarCPUOrder);
5032 }
5033
5034
5035 // Helper function which determines if an arraycopy immediately follows
5036 // an allocation, with no intervening tests or other escapes for the object.
5037 AllocateArrayNode*
5038 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5039 RegionNode* slow_region) {
5040 if (stopped()) return NULL; // no fast path
5041 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5042
5043 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5044 if (alloc == NULL) return NULL;
5045
5046 Node* rawmem = memory(Compile::AliasIdxRaw);
5047 // Is the allocation's memory state untouched?
5048 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5049 // Bail out if there have been raw-memory effects since the allocation.
5050 // (Example: There might have been a call or safepoint.)
5051 return NULL;
5052 }
5053 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5054 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5055 return NULL;
5056 }
5057
5058 // There must be no unexpected observers of this allocation.
5059 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5060 Node* obs = ptr->fast_out(i);
5061 if (obs != this->map()) {
5062 return NULL;
5063 }
5064 }
5065
5066 // This arraycopy must unconditionally follow the allocation of the ptr.
5067 Node* alloc_ctl = ptr->in(0);
5068 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5069
5070 Node* ctl = control();
5071 while (ctl != alloc_ctl) {
5072 // There may be guards which feed into the slow_region.
5073 // Any other control flow means that we might not get a chance
5074 // to finish initializing the allocated object.
5075 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5076 IfNode* iff = ctl->in(0)->as_If();
5077 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5078 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5079 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5080 ctl = iff->in(0); // This test feeds the known slow_region.
5081 continue;
5082 }
5083 // One more try: Various low-level checks bottom out in
5084 // uncommon traps. If the debug-info of the trap omits
5085 // any reference to the allocation, as we've already
5086 // observed, then there can be no objection to the trap.
5087 bool found_trap = false;
5088 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5089 Node* obs = not_ctl->fast_out(j);
5090 if (obs->in(0) == not_ctl && obs->is_Call() &&
5091 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5092 found_trap = true; break;
5093 }
5094 }
5095 if (found_trap) {
5096 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5097 continue;
5098 }
5099 }
5100 return NULL;
5101 }
5102
5103 // If we get this far, we have an allocation which immediately
5104 // precedes the arraycopy, and we can take over zeroing the new object.
5105 // The arraycopy will finish the initialization, and provide
5106 // a new control state to which we will anchor the destination pointer.
5107
5108 return alloc;
5109 }
5110
5111 // Helper for initialization of arrays, creating a ClearArray.
5112 // It writes zero bits in [start..end), within the body of an array object.
5113 // The memory effects are all chained onto the 'adr_type' alias category.
5114 //
5115 // Since the object is otherwise uninitialized, we are free
5116 // to put a little "slop" around the edges of the cleared area,
5117 // as long as it does not go back into the array's header,
5118 // or beyond the array end within the heap.
5119 //
5120 // The lower edge can be rounded down to the nearest jint and the
5121 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5122 //
5123 // Arguments:
5124 // adr_type memory slice where writes are generated
5125 // dest oop of the destination array
5126 // basic_elem_type element type of the destination
5127 // slice_idx array index of first element to store
5128 // slice_len number of elements to store (or NULL)
5129 // dest_size total size in bytes of the array object
5130 //
5131 // Exactly one of slice_len or dest_size must be non-NULL.
5132 // If dest_size is non-NULL, zeroing extends to the end of the object.
5133 // If slice_len is non-NULL, the slice_idx value must be a constant.
5134 void
5135 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5136 Node* dest,
5137 BasicType basic_elem_type,
5138 Node* slice_idx,
5139 Node* slice_len,
5140 Node* dest_size) {
5141 // one or the other but not both of slice_len and dest_size:
5142 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5143 if (slice_len == NULL) slice_len = top();
5144 if (dest_size == NULL) dest_size = top();
5145
5146 // operate on this memory slice:
5147 Node* mem = memory(adr_type); // memory slice to operate on
5148
5149 // scaling and rounding of indexes:
5150 int scale = exact_log2(type2aelembytes(basic_elem_type));
5151 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5152 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5153 int bump_bit = (-1 << scale) & BytesPerInt;
5154
5155 // determine constant starts and ends
5156 const intptr_t BIG_NEG = -128;
5157 assert(BIG_NEG + 2*abase < 0, "neg enough");
5158 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5159 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5160 if (slice_len_con == 0) {
5161 return; // nothing to do here
5162 }
5163 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5164 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5165 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5166 assert(end_con < 0, "not two cons");
5167 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5168 BytesPerLong);
5169 }
5170
5171 if (start_con >= 0 && end_con >= 0) {
5172 // Constant start and end. Simple.
5173 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5174 start_con, end_con, &_gvn);
5175 } else if (start_con >= 0 && dest_size != top()) {
5176 // Constant start, pre-rounded end after the tail of the array.
5177 Node* end = dest_size;
5178 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5179 start_con, end, &_gvn);
5180 } else if (start_con >= 0 && slice_len != top()) {
5181 // Constant start, non-constant end. End needs rounding up.
5182 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5183 intptr_t end_base = abase + (slice_idx_con << scale);
5184 int end_round = (-1 << scale) & (BytesPerLong - 1);
5185 Node* end = ConvI2X(slice_len);
5186 if (scale != 0)
5187 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5188 end_base += end_round;
5189 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5190 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5191 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5192 start_con, end, &_gvn);
5193 } else if (start_con < 0 && dest_size != top()) {
5194 // Non-constant start, pre-rounded end after the tail of the array.
5195 // This is almost certainly a "round-to-end" operation.
5196 Node* start = slice_idx;
5197 start = ConvI2X(start);
5198 if (scale != 0)
5199 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5200 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5201 if ((bump_bit | clear_low) != 0) {
5202 int to_clear = (bump_bit | clear_low);
5203 // Align up mod 8, then store a jint zero unconditionally
5204 // just before the mod-8 boundary.
5205 if (((abase + bump_bit) & ~to_clear) - bump_bit
5206 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5207 bump_bit = 0;
5208 assert((abase & to_clear) == 0, "array base must be long-aligned");
5209 } else {
5210 // Bump 'start' up to (or past) the next jint boundary:
5211 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5212 assert((abase & clear_low) == 0, "array base must be int-aligned");
5213 }
5214 // Round bumped 'start' down to jlong boundary in body of array.
5215 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5216 if (bump_bit != 0) {
5217 // Store a zero to the immediately preceding jint:
5218 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5219 Node* p1 = basic_plus_adr(dest, x1);
5220 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5221 mem = _gvn.transform(mem);
5222 }
5223 }
5224 Node* end = dest_size; // pre-rounded
5225 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5226 start, end, &_gvn);
5227 } else {
5228 // Non-constant start, unrounded non-constant end.
5229 // (Nobody zeroes a random midsection of an array using this routine.)
5230 ShouldNotReachHere(); // fix caller
5231 }
5232
5233 // Done.
5234 set_memory(mem, adr_type);
5235 }
5236
5237
5238 bool
5239 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5240 BasicType basic_elem_type,
5241 AllocateNode* alloc,
5242 Node* src, Node* src_offset,
5243 Node* dest, Node* dest_offset,
5244 Node* dest_size, bool dest_uninitialized) {
5245 // See if there is an advantage from block transfer.
5246 int scale = exact_log2(type2aelembytes(basic_elem_type));
5247 if (scale >= LogBytesPerLong)
5248 return false; // it is already a block transfer
5249
5250 // Look at the alignment of the starting offsets.
5251 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5252
5253 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5254 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5255 if (src_off_con < 0 || dest_off_con < 0)
5256 // At present, we can only understand constants.
5257 return false;
5258
5259 intptr_t src_off = abase + (src_off_con << scale);
5260 intptr_t dest_off = abase + (dest_off_con << scale);
5261
5262 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5263 // Non-aligned; too bad.
5264 // One more chance: Pick off an initial 32-bit word.
5265 // This is a common case, since abase can be odd mod 8.
5266 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5267 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5268 Node* sptr = basic_plus_adr(src, src_off);
5269 Node* dptr = basic_plus_adr(dest, dest_off);
5270 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5271 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5272 src_off += BytesPerInt;
5273 dest_off += BytesPerInt;
5274 } else {
5275 return false;
5276 }
5277 }
5278 assert(src_off % BytesPerLong == 0, "");
5279 assert(dest_off % BytesPerLong == 0, "");
5280
5281 // Do this copy by giant steps.
5282 Node* sptr = basic_plus_adr(src, src_off);
5283 Node* dptr = basic_plus_adr(dest, dest_off);
5284 Node* countx = dest_size;
5285 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5286 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5287
5288 bool disjoint_bases = true; // since alloc != NULL
5289 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5290 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5291
5292 return true;
5293 }
5294
5295
5296 // Helper function; generates code for the slow case.
5297 // We make a call to a runtime method which emulates the native method,
5298 // but without the native wrapper overhead.
5299 void
5300 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5301 Node* src, Node* src_offset,
5302 Node* dest, Node* dest_offset,
5303 Node* copy_length, bool dest_uninitialized) {
5304 assert(!dest_uninitialized, "Invariant");
5305 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5306 OptoRuntime::slow_arraycopy_Type(),
5307 OptoRuntime::slow_arraycopy_Java(),
5308 "slow_arraycopy", adr_type,
5309 src, src_offset, dest, dest_offset,
5310 copy_length);
5311
5312 // Handle exceptions thrown by this fellow:
5313 make_slow_call_ex(call, env()->Throwable_klass(), false);
5314 }
5315
5316 // Helper function; generates code for cases requiring runtime checks.
5317 Node*
5318 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5319 Node* dest_elem_klass,
5320 Node* src, Node* src_offset,
5321 Node* dest, Node* dest_offset,
5322 Node* copy_length, bool dest_uninitialized) {
5323 if (stopped()) return NULL;
5324
5325 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5326 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5327 return NULL;
5328 }
5329
5330 // Pick out the parameters required to perform a store-check
5331 // for the target array. This is an optimistic check. It will
5332 // look in each non-null element's class, at the desired klass's
5333 // super_check_offset, for the desired klass.
5334 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5335 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5336 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5337 Node* check_offset = ConvI2X(_gvn.transform(n3));
5338 Node* check_value = dest_elem_klass;
5339
5340 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5341 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5342
5343 // (We know the arrays are never conjoint, because their types differ.)
5344 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5345 OptoRuntime::checkcast_arraycopy_Type(),
5346 copyfunc_addr, "checkcast_arraycopy", adr_type,
5347 // five arguments, of which two are
5348 // intptr_t (jlong in LP64)
5349 src_start, dest_start,
5350 copy_length XTOP,
5351 check_offset XTOP,
5352 check_value);
5353
5354 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5355 }
5356
5357
5358 // Helper function; generates code for cases requiring runtime checks.
5359 Node*
5360 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5361 Node* src, Node* src_offset,
5362 Node* dest, Node* dest_offset,
5363 Node* copy_length, bool dest_uninitialized) {
5364 assert(!dest_uninitialized, "Invariant");
5365 if (stopped()) return NULL;
5366 address copyfunc_addr = StubRoutines::generic_arraycopy();
5367 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5368 return NULL;
5369 }
5370
5371 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5372 OptoRuntime::generic_arraycopy_Type(),
5373 copyfunc_addr, "generic_arraycopy", adr_type,
5374 src, src_offset, dest, dest_offset, copy_length);
5375
5376 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5377 }
5378
5379 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5380 void
5381 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5382 BasicType basic_elem_type,
5383 bool disjoint_bases,
5384 Node* src, Node* src_offset,
5385 Node* dest, Node* dest_offset,
5386 Node* copy_length, bool dest_uninitialized) {
5387 if (stopped()) return; // nothing to do
5388
5389 Node* src_start = src;
5390 Node* dest_start = dest;
5391 if (src_offset != NULL || dest_offset != NULL) {
5392 assert(src_offset != NULL && dest_offset != NULL, "");
5393 src_start = array_element_address(src, src_offset, basic_elem_type);
5394 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5395 }
5396
5397 // Figure out which arraycopy runtime method to call.
5398 const char* copyfunc_name = "arraycopy";
5399 address copyfunc_addr =
5400 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5401 disjoint_bases, copyfunc_name, dest_uninitialized);
5402
5403 // Call it. Note that the count_ix value is not scaled to a byte-size.
5404 make_runtime_call(RC_LEAF|RC_NO_FP,
5405 OptoRuntime::fast_arraycopy_Type(),
5406 copyfunc_addr, copyfunc_name, adr_type,
5407 src_start, dest_start, copy_length XTOP);
5408 }
5409
5410 //----------------------------inline_reference_get----------------------------
5411
5412 bool LibraryCallKit::inline_reference_get() {
5413 const int nargs = 1; // self
5414
5415 guarantee(java_lang_ref_Reference::referent_offset > 0,
5416 "should have already been set");
5417
5418 int referent_offset = java_lang_ref_Reference::referent_offset;
5419
5420 // Restore the stack and pop off the argument
5421 _sp += nargs;
5422 Node *reference_obj = pop();
5423
5424 // Null check on self without removing any arguments.
5425 _sp += nargs;
5426 reference_obj = do_null_check(reference_obj, T_OBJECT);
5427 _sp -= nargs;;
5428
5429 if (stopped()) return true;
5430
5431 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5432
5433 ciInstanceKlass* klass = env()->Object_klass();
5434 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5435
5436 Node* no_ctrl = NULL;
5437 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5438
5439 // Use the pre-barrier to record the value in the referent field
5440 pre_barrier(false /* do_load */,
5441 control(),
5442 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5443 result /* pre_val */,
5444 T_OBJECT);
5445
5446 push(result);
5447 return true;
5448 }
5449