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::_bitCount_i:
341 case vmIntrinsics::_bitCount_l:
342 #ifdef SPARC
343 // On sparc these depend on the popcount instruction
344 // On x86 these intrinsics always exist but may use bsr
345 case vmIntrinsics::_numberOfLeadingZeros_i:
346 case vmIntrinsics::_numberOfLeadingZeros_l:
347 case vmIntrinsics::_numberOfTrailingZeros_i:
348 case vmIntrinsics::_numberOfTrailingZeros_l:
349 #endif
350 if (!UsePopCountInstruction) return NULL;
351 break;
352
353 case vmIntrinsics::_Reference_get:
354 // It is only when G1 is enabled that we absolutely
355 // need to use the intrinsic version of Reference.get()
356 // so that the value in the referent field, if necessary,
357 // can be registered by the pre-barrier code.
358 if (!UseG1GC) return NULL;
359 break;
360
361 default:
362 assert(id <= vmIntrinsics::LAST_COMPILER_INLINE, "caller responsibility");
363 assert(id != vmIntrinsics::_Object_init && id != vmIntrinsics::_invoke, "enum out of order?");
364 break;
365 }
366
367 // -XX:-InlineClassNatives disables natives from the Class class.
368 // The flag applies to all reflective calls, notably Array.newArray
369 // (visible to Java programmers as Array.newInstance).
370 if (m->holder()->name() == ciSymbol::java_lang_Class() ||
371 m->holder()->name() == ciSymbol::java_lang_reflect_Array()) {
372 if (!InlineClassNatives) return NULL;
373 }
374
375 // -XX:-InlineThreadNatives disables natives from the Thread class.
376 if (m->holder()->name() == ciSymbol::java_lang_Thread()) {
377 if (!InlineThreadNatives) return NULL;
378 }
379
380 // -XX:-InlineMathNatives disables natives from the Math,Float and Double classes.
381 if (m->holder()->name() == ciSymbol::java_lang_Math() ||
382 m->holder()->name() == ciSymbol::java_lang_Float() ||
383 m->holder()->name() == ciSymbol::java_lang_Double()) {
384 if (!InlineMathNatives) return NULL;
385 }
386
387 // -XX:-InlineUnsafeOps disables natives from the Unsafe class.
388 if (m->holder()->name() == ciSymbol::sun_misc_Unsafe()) {
389 if (!InlineUnsafeOps) return NULL;
390 }
391
392 return new LibraryIntrinsic(m, is_virtual, (vmIntrinsics::ID) id);
393 }
394
395 //----------------------register_library_intrinsics-----------------------
396 // Initialize this file's data structures, for each Compile instance.
397 void Compile::register_library_intrinsics() {
398 // Nothing to do here.
399 }
400
401 JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
402 LibraryCallKit kit(jvms, this);
403 Compile* C = kit.C;
404 int nodes = C->unique();
405 #ifndef PRODUCT
406 if ((PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) && Verbose) {
407 char buf[1000];
408 const char* str = vmIntrinsics::short_name_as_C_string(intrinsic_id(), buf, sizeof(buf));
409 tty->print_cr("Intrinsic %s", str);
410 }
411 #endif
412
413 if (kit.try_to_inline()) {
414 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
415 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), is_virtual() ? "(intrinsic, virtual)" : "(intrinsic)");
416 }
417 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_worked);
418 if (C->log()) {
419 C->log()->elem("intrinsic id='%s'%s nodes='%d'",
420 vmIntrinsics::name_at(intrinsic_id()),
421 (is_virtual() ? " virtual='1'" : ""),
422 C->unique() - nodes);
423 }
424 return kit.transfer_exceptions_into_jvms();
425 }
426
427 // The intrinsic bailed out
428 if (PrintIntrinsics || PrintInlining NOT_PRODUCT( || PrintOptoInlining) ) {
429 const char* msg = is_virtual() ? "failed to inline (intrinsic, virtual)" : "failed to inline (intrinsic)";
430 if (jvms->has_method()) {
431 // Not a root compile.
432 CompileTask::print_inlining(kit.callee(), jvms->depth() - 1, kit.bci(), msg);
433 } else {
434 // Root compile
435 tty->print("Did not generate intrinsic %s%s at bci:%d in",
436 vmIntrinsics::name_at(intrinsic_id()),
437 (is_virtual() ? " (virtual)" : ""), kit.bci());
438 }
439 }
440 C->gather_intrinsic_statistics(intrinsic_id(), is_virtual(), Compile::_intrinsic_failed);
441 return NULL;
442 }
443
444 bool LibraryCallKit::try_to_inline() {
445 // Handle symbolic names for otherwise undistinguished boolean switches:
446 const bool is_store = true;
447 const bool is_native_ptr = true;
448 const bool is_static = true;
449
450 if (!jvms()->has_method()) {
451 // Root JVMState has a null method.
452 assert(map()->memory()->Opcode() == Op_Parm, "");
453 // Insert the memory aliasing node
454 set_all_memory(reset_memory());
455 }
456 assert(merged_memory(), "");
457
458 switch (intrinsic_id()) {
459 case vmIntrinsics::_hashCode:
460 return inline_native_hashcode(intrinsic()->is_virtual(), !is_static);
461 case vmIntrinsics::_identityHashCode:
462 return inline_native_hashcode(/*!virtual*/ false, is_static);
463 case vmIntrinsics::_getClass:
464 return inline_native_getClass();
465
466 case vmIntrinsics::_dsin:
467 case vmIntrinsics::_dcos:
468 case vmIntrinsics::_dtan:
469 case vmIntrinsics::_dabs:
470 case vmIntrinsics::_datan2:
471 case vmIntrinsics::_dsqrt:
472 case vmIntrinsics::_dexp:
473 case vmIntrinsics::_dlog:
474 case vmIntrinsics::_dlog10:
475 case vmIntrinsics::_dpow:
476 return inline_math_native(intrinsic_id());
477
478 case vmIntrinsics::_min:
479 case vmIntrinsics::_max:
480 return inline_min_max(intrinsic_id());
481
482 case vmIntrinsics::_arraycopy:
483 return inline_arraycopy();
484
485 case vmIntrinsics::_compareTo:
486 return inline_string_compareTo();
487 case vmIntrinsics::_indexOf:
488 return inline_string_indexOf();
489 case vmIntrinsics::_equals:
490 return inline_string_equals();
491
492 case vmIntrinsics::_getObject:
493 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, false);
494 case vmIntrinsics::_getBoolean:
495 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, false);
496 case vmIntrinsics::_getByte:
497 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, false);
498 case vmIntrinsics::_getShort:
499 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, false);
500 case vmIntrinsics::_getChar:
501 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, false);
502 case vmIntrinsics::_getInt:
503 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, false);
504 case vmIntrinsics::_getLong:
505 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, false);
506 case vmIntrinsics::_getFloat:
507 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, false);
508 case vmIntrinsics::_getDouble:
509 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, false);
510
511 case vmIntrinsics::_putObject:
512 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, false);
513 case vmIntrinsics::_putBoolean:
514 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, false);
515 case vmIntrinsics::_putByte:
516 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, false);
517 case vmIntrinsics::_putShort:
518 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, false);
519 case vmIntrinsics::_putChar:
520 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, false);
521 case vmIntrinsics::_putInt:
522 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, false);
523 case vmIntrinsics::_putLong:
524 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, false);
525 case vmIntrinsics::_putFloat:
526 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, false);
527 case vmIntrinsics::_putDouble:
528 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, false);
529
530 case vmIntrinsics::_getByte_raw:
531 return inline_unsafe_access(is_native_ptr, !is_store, T_BYTE, false);
532 case vmIntrinsics::_getShort_raw:
533 return inline_unsafe_access(is_native_ptr, !is_store, T_SHORT, false);
534 case vmIntrinsics::_getChar_raw:
535 return inline_unsafe_access(is_native_ptr, !is_store, T_CHAR, false);
536 case vmIntrinsics::_getInt_raw:
537 return inline_unsafe_access(is_native_ptr, !is_store, T_INT, false);
538 case vmIntrinsics::_getLong_raw:
539 return inline_unsafe_access(is_native_ptr, !is_store, T_LONG, false);
540 case vmIntrinsics::_getFloat_raw:
541 return inline_unsafe_access(is_native_ptr, !is_store, T_FLOAT, false);
542 case vmIntrinsics::_getDouble_raw:
543 return inline_unsafe_access(is_native_ptr, !is_store, T_DOUBLE, false);
544 case vmIntrinsics::_getAddress_raw:
545 return inline_unsafe_access(is_native_ptr, !is_store, T_ADDRESS, false);
546
547 case vmIntrinsics::_putByte_raw:
548 return inline_unsafe_access(is_native_ptr, is_store, T_BYTE, false);
549 case vmIntrinsics::_putShort_raw:
550 return inline_unsafe_access(is_native_ptr, is_store, T_SHORT, false);
551 case vmIntrinsics::_putChar_raw:
552 return inline_unsafe_access(is_native_ptr, is_store, T_CHAR, false);
553 case vmIntrinsics::_putInt_raw:
554 return inline_unsafe_access(is_native_ptr, is_store, T_INT, false);
555 case vmIntrinsics::_putLong_raw:
556 return inline_unsafe_access(is_native_ptr, is_store, T_LONG, false);
557 case vmIntrinsics::_putFloat_raw:
558 return inline_unsafe_access(is_native_ptr, is_store, T_FLOAT, false);
559 case vmIntrinsics::_putDouble_raw:
560 return inline_unsafe_access(is_native_ptr, is_store, T_DOUBLE, false);
561 case vmIntrinsics::_putAddress_raw:
562 return inline_unsafe_access(is_native_ptr, is_store, T_ADDRESS, false);
563
564 case vmIntrinsics::_getObjectVolatile:
565 return inline_unsafe_access(!is_native_ptr, !is_store, T_OBJECT, true);
566 case vmIntrinsics::_getBooleanVolatile:
567 return inline_unsafe_access(!is_native_ptr, !is_store, T_BOOLEAN, true);
568 case vmIntrinsics::_getByteVolatile:
569 return inline_unsafe_access(!is_native_ptr, !is_store, T_BYTE, true);
570 case vmIntrinsics::_getShortVolatile:
571 return inline_unsafe_access(!is_native_ptr, !is_store, T_SHORT, true);
572 case vmIntrinsics::_getCharVolatile:
573 return inline_unsafe_access(!is_native_ptr, !is_store, T_CHAR, true);
574 case vmIntrinsics::_getIntVolatile:
575 return inline_unsafe_access(!is_native_ptr, !is_store, T_INT, true);
576 case vmIntrinsics::_getLongVolatile:
577 return inline_unsafe_access(!is_native_ptr, !is_store, T_LONG, true);
578 case vmIntrinsics::_getFloatVolatile:
579 return inline_unsafe_access(!is_native_ptr, !is_store, T_FLOAT, true);
580 case vmIntrinsics::_getDoubleVolatile:
581 return inline_unsafe_access(!is_native_ptr, !is_store, T_DOUBLE, true);
582
583 case vmIntrinsics::_putObjectVolatile:
584 return inline_unsafe_access(!is_native_ptr, is_store, T_OBJECT, true);
585 case vmIntrinsics::_putBooleanVolatile:
586 return inline_unsafe_access(!is_native_ptr, is_store, T_BOOLEAN, true);
587 case vmIntrinsics::_putByteVolatile:
588 return inline_unsafe_access(!is_native_ptr, is_store, T_BYTE, true);
589 case vmIntrinsics::_putShortVolatile:
590 return inline_unsafe_access(!is_native_ptr, is_store, T_SHORT, true);
591 case vmIntrinsics::_putCharVolatile:
592 return inline_unsafe_access(!is_native_ptr, is_store, T_CHAR, true);
593 case vmIntrinsics::_putIntVolatile:
594 return inline_unsafe_access(!is_native_ptr, is_store, T_INT, true);
595 case vmIntrinsics::_putLongVolatile:
596 return inline_unsafe_access(!is_native_ptr, is_store, T_LONG, true);
597 case vmIntrinsics::_putFloatVolatile:
598 return inline_unsafe_access(!is_native_ptr, is_store, T_FLOAT, true);
599 case vmIntrinsics::_putDoubleVolatile:
600 return inline_unsafe_access(!is_native_ptr, is_store, T_DOUBLE, true);
601
602 case vmIntrinsics::_prefetchRead:
603 return inline_unsafe_prefetch(!is_native_ptr, !is_store, !is_static);
604 case vmIntrinsics::_prefetchWrite:
605 return inline_unsafe_prefetch(!is_native_ptr, is_store, !is_static);
606 case vmIntrinsics::_prefetchReadStatic:
607 return inline_unsafe_prefetch(!is_native_ptr, !is_store, is_static);
608 case vmIntrinsics::_prefetchWriteStatic:
609 return inline_unsafe_prefetch(!is_native_ptr, is_store, is_static);
610
611 case vmIntrinsics::_compareAndSwapObject:
612 return inline_unsafe_CAS(T_OBJECT);
613 case vmIntrinsics::_compareAndSwapInt:
614 return inline_unsafe_CAS(T_INT);
615 case vmIntrinsics::_compareAndSwapLong:
616 return inline_unsafe_CAS(T_LONG);
617
618 case vmIntrinsics::_putOrderedObject:
619 return inline_unsafe_ordered_store(T_OBJECT);
620 case vmIntrinsics::_putOrderedInt:
621 return inline_unsafe_ordered_store(T_INT);
622 case vmIntrinsics::_putOrderedLong:
623 return inline_unsafe_ordered_store(T_LONG);
624
625 case vmIntrinsics::_currentThread:
626 return inline_native_currentThread();
627 case vmIntrinsics::_isInterrupted:
628 return inline_native_isInterrupted();
629
630 case vmIntrinsics::_currentTimeMillis:
631 return inline_native_time_funcs(false);
632 case vmIntrinsics::_nanoTime:
633 return inline_native_time_funcs(true);
634 case vmIntrinsics::_allocateInstance:
635 return inline_unsafe_allocate();
636 case vmIntrinsics::_copyMemory:
637 return inline_unsafe_copyMemory();
638 case vmIntrinsics::_newArray:
639 return inline_native_newArray();
640 case vmIntrinsics::_getLength:
641 return inline_native_getLength();
642 case vmIntrinsics::_copyOf:
643 return inline_array_copyOf(false);
644 case vmIntrinsics::_copyOfRange:
645 return inline_array_copyOf(true);
646 case vmIntrinsics::_equalsC:
647 return inline_array_equals();
648 case vmIntrinsics::_clone:
649 return inline_native_clone(intrinsic()->is_virtual());
650
651 case vmIntrinsics::_isAssignableFrom:
652 return inline_native_subtype_check();
653
654 case vmIntrinsics::_isInstance:
655 case vmIntrinsics::_getModifiers:
656 case vmIntrinsics::_isInterface:
657 case vmIntrinsics::_isArray:
658 case vmIntrinsics::_isPrimitive:
659 case vmIntrinsics::_getSuperclass:
660 case vmIntrinsics::_getComponentType:
661 case vmIntrinsics::_getClassAccessFlags:
662 return inline_native_Class_query(intrinsic_id());
663
664 case vmIntrinsics::_floatToRawIntBits:
665 case vmIntrinsics::_floatToIntBits:
666 case vmIntrinsics::_intBitsToFloat:
667 case vmIntrinsics::_doubleToRawLongBits:
668 case vmIntrinsics::_doubleToLongBits:
669 case vmIntrinsics::_longBitsToDouble:
670 return inline_fp_conversions(intrinsic_id());
671
672 case vmIntrinsics::_numberOfLeadingZeros_i:
673 case vmIntrinsics::_numberOfLeadingZeros_l:
674 return inline_numberOfLeadingZeros(intrinsic_id());
675
676 case vmIntrinsics::_numberOfTrailingZeros_i:
677 case vmIntrinsics::_numberOfTrailingZeros_l:
678 return inline_numberOfTrailingZeros(intrinsic_id());
679
680 case vmIntrinsics::_bitCount_i:
681 case vmIntrinsics::_bitCount_l:
682 return inline_bitCount(intrinsic_id());
683
684 case vmIntrinsics::_reverseBytes_i:
685 case vmIntrinsics::_reverseBytes_l:
686 case vmIntrinsics::_reverseBytes_s:
687 case vmIntrinsics::_reverseBytes_c:
688 return inline_reverseBytes((vmIntrinsics::ID) intrinsic_id());
689
690 case vmIntrinsics::_get_AtomicLong:
691 return inline_native_AtomicLong_get();
692 case vmIntrinsics::_attemptUpdate:
693 return inline_native_AtomicLong_attemptUpdate();
694
695 case vmIntrinsics::_getCallerClass:
696 return inline_native_Reflection_getCallerClass();
697
698 case vmIntrinsics::_Reference_get:
699 return inline_reference_get();
700
701 default:
702 // If you get here, it may be that someone has added a new intrinsic
703 // to the list in vmSymbols.hpp without implementing it here.
704 #ifndef PRODUCT
705 if ((PrintMiscellaneous && (Verbose || WizardMode)) || PrintOpto) {
706 tty->print_cr("*** Warning: Unimplemented intrinsic %s(%d)",
707 vmIntrinsics::name_at(intrinsic_id()), intrinsic_id());
708 }
709 #endif
710 return false;
711 }
712 }
713
714 //------------------------------push_result------------------------------
715 // Helper function for finishing intrinsics.
716 void LibraryCallKit::push_result(RegionNode* region, PhiNode* value) {
717 record_for_igvn(region);
718 set_control(_gvn.transform(region));
719 BasicType value_type = value->type()->basic_type();
720 push_node(value_type, _gvn.transform(value));
721 }
722
723 //------------------------------generate_guard---------------------------
724 // Helper function for generating guarded fast-slow graph structures.
725 // The given 'test', if true, guards a slow path. If the test fails
726 // then a fast path can be taken. (We generally hope it fails.)
727 // In all cases, GraphKit::control() is updated to the fast path.
728 // The returned value represents the control for the slow path.
729 // The return value is never 'top'; it is either a valid control
730 // or NULL if it is obvious that the slow path can never be taken.
731 // Also, if region and the slow control are not NULL, the slow edge
732 // is appended to the region.
733 Node* LibraryCallKit::generate_guard(Node* test, RegionNode* region, float true_prob) {
734 if (stopped()) {
735 // Already short circuited.
736 return NULL;
737 }
738
739 // Build an if node and its projections.
740 // If test is true we take the slow path, which we assume is uncommon.
741 if (_gvn.type(test) == TypeInt::ZERO) {
742 // The slow branch is never taken. No need to build this guard.
743 return NULL;
744 }
745
746 IfNode* iff = create_and_map_if(control(), test, true_prob, COUNT_UNKNOWN);
747
748 Node* if_slow = _gvn.transform( new (C, 1) IfTrueNode(iff) );
749 if (if_slow == top()) {
750 // The slow branch is never taken. No need to build this guard.
751 return NULL;
752 }
753
754 if (region != NULL)
755 region->add_req(if_slow);
756
757 Node* if_fast = _gvn.transform( new (C, 1) IfFalseNode(iff) );
758 set_control(if_fast);
759
760 return if_slow;
761 }
762
763 inline Node* LibraryCallKit::generate_slow_guard(Node* test, RegionNode* region) {
764 return generate_guard(test, region, PROB_UNLIKELY_MAG(3));
765 }
766 inline Node* LibraryCallKit::generate_fair_guard(Node* test, RegionNode* region) {
767 return generate_guard(test, region, PROB_FAIR);
768 }
769
770 inline Node* LibraryCallKit::generate_negative_guard(Node* index, RegionNode* region,
771 Node* *pos_index) {
772 if (stopped())
773 return NULL; // already stopped
774 if (_gvn.type(index)->higher_equal(TypeInt::POS)) // [0,maxint]
775 return NULL; // index is already adequately typed
776 Node* cmp_lt = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
777 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
778 Node* is_neg = generate_guard(bol_lt, region, PROB_MIN);
779 if (is_neg != NULL && pos_index != NULL) {
780 // Emulate effect of Parse::adjust_map_after_if.
781 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS);
782 ccast->set_req(0, control());
783 (*pos_index) = _gvn.transform(ccast);
784 }
785 return is_neg;
786 }
787
788 inline Node* LibraryCallKit::generate_nonpositive_guard(Node* index, bool never_negative,
789 Node* *pos_index) {
790 if (stopped())
791 return NULL; // already stopped
792 if (_gvn.type(index)->higher_equal(TypeInt::POS1)) // [1,maxint]
793 return NULL; // index is already adequately typed
794 Node* cmp_le = _gvn.transform( new (C, 3) CmpINode(index, intcon(0)) );
795 BoolTest::mask le_or_eq = (never_negative ? BoolTest::eq : BoolTest::le);
796 Node* bol_le = _gvn.transform( new (C, 2) BoolNode(cmp_le, le_or_eq) );
797 Node* is_notp = generate_guard(bol_le, NULL, PROB_MIN);
798 if (is_notp != NULL && pos_index != NULL) {
799 // Emulate effect of Parse::adjust_map_after_if.
800 Node* ccast = new (C, 2) CastIINode(index, TypeInt::POS1);
801 ccast->set_req(0, control());
802 (*pos_index) = _gvn.transform(ccast);
803 }
804 return is_notp;
805 }
806
807 // Make sure that 'position' is a valid limit index, in [0..length].
808 // There are two equivalent plans for checking this:
809 // A. (offset + copyLength) unsigned<= arrayLength
810 // B. offset <= (arrayLength - copyLength)
811 // We require that all of the values above, except for the sum and
812 // difference, are already known to be non-negative.
813 // Plan A is robust in the face of overflow, if offset and copyLength
814 // are both hugely positive.
815 //
816 // Plan B is less direct and intuitive, but it does not overflow at
817 // all, since the difference of two non-negatives is always
818 // representable. Whenever Java methods must perform the equivalent
819 // check they generally use Plan B instead of Plan A.
820 // For the moment we use Plan A.
821 inline Node* LibraryCallKit::generate_limit_guard(Node* offset,
822 Node* subseq_length,
823 Node* array_length,
824 RegionNode* region) {
825 if (stopped())
826 return NULL; // already stopped
827 bool zero_offset = _gvn.type(offset) == TypeInt::ZERO;
828 if (zero_offset && subseq_length->eqv_uncast(array_length))
829 return NULL; // common case of whole-array copy
830 Node* last = subseq_length;
831 if (!zero_offset) // last += offset
832 last = _gvn.transform( new (C, 3) AddINode(last, offset));
833 Node* cmp_lt = _gvn.transform( new (C, 3) CmpUNode(array_length, last) );
834 Node* bol_lt = _gvn.transform( new (C, 2) BoolNode(cmp_lt, BoolTest::lt) );
835 Node* is_over = generate_guard(bol_lt, region, PROB_MIN);
836 return is_over;
837 }
838
839
840 //--------------------------generate_current_thread--------------------
841 Node* LibraryCallKit::generate_current_thread(Node* &tls_output) {
842 ciKlass* thread_klass = env()->Thread_klass();
843 const Type* thread_type = TypeOopPtr::make_from_klass(thread_klass)->cast_to_ptr_type(TypePtr::NotNull);
844 Node* thread = _gvn.transform(new (C, 1) ThreadLocalNode());
845 Node* p = basic_plus_adr(top()/*!oop*/, thread, in_bytes(JavaThread::threadObj_offset()));
846 Node* threadObj = make_load(NULL, p, thread_type, T_OBJECT);
847 tls_output = thread;
848 return threadObj;
849 }
850
851
852 //------------------------------make_string_method_node------------------------
853 // Helper method for String intrinsic finctions.
854 Node* LibraryCallKit::make_string_method_node(int opcode, Node* str1, Node* cnt1, Node* str2, Node* cnt2) {
855 const int value_offset = java_lang_String::value_offset_in_bytes();
856 const int count_offset = java_lang_String::count_offset_in_bytes();
857 const int offset_offset = java_lang_String::offset_offset_in_bytes();
858
859 Node* no_ctrl = NULL;
860
861 ciInstanceKlass* klass = env()->String_klass();
862 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
863
864 const TypeAryPtr* value_type =
865 TypeAryPtr::make(TypePtr::NotNull,
866 TypeAry::make(TypeInt::CHAR,TypeInt::POS),
867 ciTypeArrayKlass::make(T_CHAR), true, 0);
868
869 // Get start addr of string and substring
870 Node* str1_valuea = basic_plus_adr(str1, str1, value_offset);
871 Node* str1_value = make_load(no_ctrl, str1_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
872 Node* str1_offseta = basic_plus_adr(str1, str1, offset_offset);
873 Node* str1_offset = make_load(no_ctrl, str1_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
874 Node* str1_start = array_element_address(str1_value, str1_offset, T_CHAR);
875
876 Node* str2_valuea = basic_plus_adr(str2, str2, value_offset);
877 Node* str2_value = make_load(no_ctrl, str2_valuea, value_type, T_OBJECT, string_type->add_offset(value_offset));
878 Node* str2_offseta = basic_plus_adr(str2, str2, offset_offset);
879 Node* str2_offset = make_load(no_ctrl, str2_offseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
880 Node* str2_start = array_element_address(str2_value, str2_offset, T_CHAR);
881
882 Node* result = NULL;
883 switch (opcode) {
884 case Op_StrIndexOf:
885 result = new (C, 6) StrIndexOfNode(control(), memory(TypeAryPtr::CHARS),
886 str1_start, cnt1, str2_start, cnt2);
887 break;
888 case Op_StrComp:
889 result = new (C, 6) StrCompNode(control(), memory(TypeAryPtr::CHARS),
890 str1_start, cnt1, str2_start, cnt2);
891 break;
892 case Op_StrEquals:
893 result = new (C, 5) StrEqualsNode(control(), memory(TypeAryPtr::CHARS),
894 str1_start, str2_start, cnt1);
895 break;
896 default:
897 ShouldNotReachHere();
898 return NULL;
899 }
900
901 // All these intrinsics have checks.
902 C->set_has_split_ifs(true); // Has chance for split-if optimization
903
904 return _gvn.transform(result);
905 }
906
907 //------------------------------inline_string_compareTo------------------------
908 bool LibraryCallKit::inline_string_compareTo() {
909
910 if (!Matcher::has_match_rule(Op_StrComp)) return false;
911
912 const int value_offset = java_lang_String::value_offset_in_bytes();
913 const int count_offset = java_lang_String::count_offset_in_bytes();
914 const int offset_offset = java_lang_String::offset_offset_in_bytes();
915
916 _sp += 2;
917 Node *argument = pop(); // pop non-receiver first: it was pushed second
918 Node *receiver = pop();
919
920 // Null check on self without removing any arguments. The argument
921 // null check technically happens in the wrong place, which can lead to
922 // invalid stack traces when string compare is inlined into a method
923 // which handles NullPointerExceptions.
924 _sp += 2;
925 receiver = do_null_check(receiver, T_OBJECT);
926 argument = do_null_check(argument, T_OBJECT);
927 _sp -= 2;
928 if (stopped()) {
929 return true;
930 }
931
932 ciInstanceKlass* klass = env()->String_klass();
933 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
934 Node* no_ctrl = NULL;
935
936 // Get counts for string and argument
937 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
938 Node* receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
939
940 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
941 Node* argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
942
943 Node* compare = make_string_method_node(Op_StrComp, receiver, receiver_cnt, argument, argument_cnt);
944 push(compare);
945 return true;
946 }
947
948 //------------------------------inline_string_equals------------------------
949 bool LibraryCallKit::inline_string_equals() {
950
951 if (!Matcher::has_match_rule(Op_StrEquals)) return false;
952
953 const int value_offset = java_lang_String::value_offset_in_bytes();
954 const int count_offset = java_lang_String::count_offset_in_bytes();
955 const int offset_offset = java_lang_String::offset_offset_in_bytes();
956
957 int nargs = 2;
958 _sp += nargs;
959 Node* argument = pop(); // pop non-receiver first: it was pushed second
960 Node* receiver = pop();
961
962 // Null check on self without removing any arguments. The argument
963 // null check technically happens in the wrong place, which can lead to
964 // invalid stack traces when string compare is inlined into a method
965 // which handles NullPointerExceptions.
966 _sp += nargs;
967 receiver = do_null_check(receiver, T_OBJECT);
968 //should not do null check for argument for String.equals(), because spec
969 //allows to specify NULL as argument.
970 _sp -= nargs;
971
972 if (stopped()) {
973 return true;
974 }
975
976 // paths (plus control) merge
977 RegionNode* region = new (C, 5) RegionNode(5);
978 Node* phi = new (C, 5) PhiNode(region, TypeInt::BOOL);
979
980 // does source == target string?
981 Node* cmp = _gvn.transform(new (C, 3) CmpPNode(receiver, argument));
982 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::eq));
983
984 Node* if_eq = generate_slow_guard(bol, NULL);
985 if (if_eq != NULL) {
986 // receiver == argument
987 phi->init_req(2, intcon(1));
988 region->init_req(2, if_eq);
989 }
990
991 // get String klass for instanceOf
992 ciInstanceKlass* klass = env()->String_klass();
993
994 if (!stopped()) {
995 _sp += nargs; // gen_instanceof might do an uncommon trap
996 Node* inst = gen_instanceof(argument, makecon(TypeKlassPtr::make(klass)));
997 _sp -= nargs;
998 Node* cmp = _gvn.transform(new (C, 3) CmpINode(inst, intcon(1)));
999 Node* bol = _gvn.transform(new (C, 2) BoolNode(cmp, BoolTest::ne));
1000
1001 Node* inst_false = generate_guard(bol, NULL, PROB_MIN);
1002 //instanceOf == true, fallthrough
1003
1004 if (inst_false != NULL) {
1005 phi->init_req(3, intcon(0));
1006 region->init_req(3, inst_false);
1007 }
1008 }
1009
1010 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1011
1012 Node* no_ctrl = NULL;
1013 Node* receiver_cnt;
1014 Node* argument_cnt;
1015
1016 if (!stopped()) {
1017 // Properly cast the argument to String
1018 argument = _gvn.transform(new (C, 2) CheckCastPPNode(control(), argument, string_type));
1019 // This path is taken only when argument's type is String:NotNull.
1020 argument = cast_not_null(argument, false);
1021
1022 // Get counts for string and argument
1023 Node* receiver_cnta = basic_plus_adr(receiver, receiver, count_offset);
1024 receiver_cnt = make_load(no_ctrl, receiver_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1025
1026 Node* argument_cnta = basic_plus_adr(argument, argument, count_offset);
1027 argument_cnt = make_load(no_ctrl, argument_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1028
1029 // Check for receiver count != argument count
1030 Node* cmp = _gvn.transform( new(C, 3) CmpINode(receiver_cnt, argument_cnt) );
1031 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::ne) );
1032 Node* if_ne = generate_slow_guard(bol, NULL);
1033 if (if_ne != NULL) {
1034 phi->init_req(4, intcon(0));
1035 region->init_req(4, if_ne);
1036 }
1037 }
1038
1039 // Check for count == 0 is done by mach node StrEquals.
1040
1041 if (!stopped()) {
1042 Node* equals = make_string_method_node(Op_StrEquals, receiver, receiver_cnt, argument, argument_cnt);
1043 phi->init_req(1, equals);
1044 region->init_req(1, control());
1045 }
1046
1047 // post merge
1048 set_control(_gvn.transform(region));
1049 record_for_igvn(region);
1050
1051 push(_gvn.transform(phi));
1052
1053 return true;
1054 }
1055
1056 //------------------------------inline_array_equals----------------------------
1057 bool LibraryCallKit::inline_array_equals() {
1058
1059 if (!Matcher::has_match_rule(Op_AryEq)) return false;
1060
1061 _sp += 2;
1062 Node *argument2 = pop();
1063 Node *argument1 = pop();
1064
1065 Node* equals =
1066 _gvn.transform(new (C, 4) AryEqNode(control(), memory(TypeAryPtr::CHARS),
1067 argument1, argument2) );
1068 push(equals);
1069 return true;
1070 }
1071
1072 // Java version of String.indexOf(constant string)
1073 // class StringDecl {
1074 // StringDecl(char[] ca) {
1075 // offset = 0;
1076 // count = ca.length;
1077 // value = ca;
1078 // }
1079 // int offset;
1080 // int count;
1081 // char[] value;
1082 // }
1083 //
1084 // static int string_indexOf_J(StringDecl string_object, char[] target_object,
1085 // int targetOffset, int cache_i, int md2) {
1086 // int cache = cache_i;
1087 // int sourceOffset = string_object.offset;
1088 // int sourceCount = string_object.count;
1089 // int targetCount = target_object.length;
1090 //
1091 // int targetCountLess1 = targetCount - 1;
1092 // int sourceEnd = sourceOffset + sourceCount - targetCountLess1;
1093 //
1094 // char[] source = string_object.value;
1095 // char[] target = target_object;
1096 // int lastChar = target[targetCountLess1];
1097 //
1098 // outer_loop:
1099 // for (int i = sourceOffset; i < sourceEnd; ) {
1100 // int src = source[i + targetCountLess1];
1101 // if (src == lastChar) {
1102 // // With random strings and a 4-character alphabet,
1103 // // reverse matching at this point sets up 0.8% fewer
1104 // // frames, but (paradoxically) makes 0.3% more probes.
1105 // // Since those probes are nearer the lastChar probe,
1106 // // there is may be a net D$ win with reverse matching.
1107 // // But, reversing loop inhibits unroll of inner loop
1108 // // for unknown reason. So, does running outer loop from
1109 // // (sourceOffset - targetCountLess1) to (sourceOffset + sourceCount)
1110 // for (int j = 0; j < targetCountLess1; j++) {
1111 // if (target[targetOffset + j] != source[i+j]) {
1112 // if ((cache & (1 << source[i+j])) == 0) {
1113 // if (md2 < j+1) {
1114 // i += j+1;
1115 // continue outer_loop;
1116 // }
1117 // }
1118 // i += md2;
1119 // continue outer_loop;
1120 // }
1121 // }
1122 // return i - sourceOffset;
1123 // }
1124 // if ((cache & (1 << src)) == 0) {
1125 // i += targetCountLess1;
1126 // } // using "i += targetCount;" and an "else i++;" causes a jump to jump.
1127 // i++;
1128 // }
1129 // return -1;
1130 // }
1131
1132 //------------------------------string_indexOf------------------------
1133 Node* LibraryCallKit::string_indexOf(Node* string_object, ciTypeArray* target_array, jint targetOffset_i,
1134 jint cache_i, jint md2_i) {
1135
1136 Node* no_ctrl = NULL;
1137 float likely = PROB_LIKELY(0.9);
1138 float unlikely = PROB_UNLIKELY(0.9);
1139
1140 const int nargs = 2; // number of arguments to push back for uncommon trap in predicate
1141
1142 const int value_offset = java_lang_String::value_offset_in_bytes();
1143 const int count_offset = java_lang_String::count_offset_in_bytes();
1144 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1145
1146 ciInstanceKlass* klass = env()->String_klass();
1147 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(klass);
1148 const TypeAryPtr* source_type = TypeAryPtr::make(TypePtr::NotNull, TypeAry::make(TypeInt::CHAR,TypeInt::POS), ciTypeArrayKlass::make(T_CHAR), true, 0);
1149
1150 Node* sourceOffseta = basic_plus_adr(string_object, string_object, offset_offset);
1151 Node* sourceOffset = make_load(no_ctrl, sourceOffseta, TypeInt::INT, T_INT, string_type->add_offset(offset_offset));
1152 Node* sourceCounta = basic_plus_adr(string_object, string_object, count_offset);
1153 Node* sourceCount = make_load(no_ctrl, sourceCounta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1154 Node* sourcea = basic_plus_adr(string_object, string_object, value_offset);
1155 Node* source = make_load(no_ctrl, sourcea, source_type, T_OBJECT, string_type->add_offset(value_offset));
1156
1157 Node* target = _gvn.transform( makecon(TypeOopPtr::make_from_constant(target_array, true)) );
1158 jint target_length = target_array->length();
1159 const TypeAry* target_array_type = TypeAry::make(TypeInt::CHAR, TypeInt::make(0, target_length, Type::WidenMin));
1160 const TypeAryPtr* target_type = TypeAryPtr::make(TypePtr::BotPTR, target_array_type, target_array->klass(), true, Type::OffsetBot);
1161
1162 IdealKit kit(this, false, true);
1163 #define __ kit.
1164 Node* zero = __ ConI(0);
1165 Node* one = __ ConI(1);
1166 Node* cache = __ ConI(cache_i);
1167 Node* md2 = __ ConI(md2_i);
1168 Node* lastChar = __ ConI(target_array->char_at(target_length - 1));
1169 Node* targetCount = __ ConI(target_length);
1170 Node* targetCountLess1 = __ ConI(target_length - 1);
1171 Node* targetOffset = __ ConI(targetOffset_i);
1172 Node* sourceEnd = __ SubI(__ AddI(sourceOffset, sourceCount), targetCountLess1);
1173
1174 IdealVariable rtn(kit), i(kit), j(kit); __ declarations_done();
1175 Node* outer_loop = __ make_label(2 /* goto */);
1176 Node* return_ = __ make_label(1);
1177
1178 __ set(rtn,__ ConI(-1));
1179 __ loop(this, nargs, i, sourceOffset, BoolTest::lt, sourceEnd); {
1180 Node* i2 = __ AddI(__ value(i), targetCountLess1);
1181 // pin to prohibit loading of "next iteration" value which may SEGV (rare)
1182 Node* src = load_array_element(__ ctrl(), source, i2, TypeAryPtr::CHARS);
1183 __ if_then(src, BoolTest::eq, lastChar, unlikely); {
1184 __ loop(this, nargs, j, zero, BoolTest::lt, targetCountLess1); {
1185 Node* tpj = __ AddI(targetOffset, __ value(j));
1186 Node* targ = load_array_element(no_ctrl, target, tpj, target_type);
1187 Node* ipj = __ AddI(__ value(i), __ value(j));
1188 Node* src2 = load_array_element(no_ctrl, source, ipj, TypeAryPtr::CHARS);
1189 __ if_then(targ, BoolTest::ne, src2); {
1190 __ if_then(__ AndI(cache, __ LShiftI(one, src2)), BoolTest::eq, zero); {
1191 __ if_then(md2, BoolTest::lt, __ AddI(__ value(j), one)); {
1192 __ increment(i, __ AddI(__ value(j), one));
1193 __ goto_(outer_loop);
1194 } __ end_if(); __ dead(j);
1195 }__ end_if(); __ dead(j);
1196 __ increment(i, md2);
1197 __ goto_(outer_loop);
1198 }__ end_if();
1199 __ increment(j, one);
1200 }__ end_loop(); __ dead(j);
1201 __ set(rtn, __ SubI(__ value(i), sourceOffset)); __ dead(i);
1202 __ goto_(return_);
1203 }__ end_if();
1204 __ if_then(__ AndI(cache, __ LShiftI(one, src)), BoolTest::eq, zero, likely); {
1205 __ increment(i, targetCountLess1);
1206 }__ end_if();
1207 __ increment(i, one);
1208 __ bind(outer_loop);
1209 }__ end_loop(); __ dead(i);
1210 __ bind(return_);
1211
1212 // Final sync IdealKit and GraphKit.
1213 final_sync(kit);
1214 Node* result = __ value(rtn);
1215 #undef __
1216 C->set_has_loops(true);
1217 return result;
1218 }
1219
1220 //------------------------------inline_string_indexOf------------------------
1221 bool LibraryCallKit::inline_string_indexOf() {
1222
1223 const int value_offset = java_lang_String::value_offset_in_bytes();
1224 const int count_offset = java_lang_String::count_offset_in_bytes();
1225 const int offset_offset = java_lang_String::offset_offset_in_bytes();
1226
1227 _sp += 2;
1228 Node *argument = pop(); // pop non-receiver first: it was pushed second
1229 Node *receiver = pop();
1230
1231 Node* result;
1232 // Disable the use of pcmpestri until it can be guaranteed that
1233 // the load doesn't cross into the uncommited space.
1234 if (Matcher::has_match_rule(Op_StrIndexOf) &&
1235 UseSSE42Intrinsics) {
1236 // Generate SSE4.2 version of indexOf
1237 // We currently only have match rules that use SSE4.2
1238
1239 // Null check on self without removing any arguments. The argument
1240 // null check technically happens in the wrong place, which can lead to
1241 // invalid stack traces when string compare is inlined into a method
1242 // which handles NullPointerExceptions.
1243 _sp += 2;
1244 receiver = do_null_check(receiver, T_OBJECT);
1245 argument = do_null_check(argument, T_OBJECT);
1246 _sp -= 2;
1247
1248 if (stopped()) {
1249 return true;
1250 }
1251
1252 ciInstanceKlass* str_klass = env()->String_klass();
1253 const TypeOopPtr* string_type = TypeOopPtr::make_from_klass(str_klass);
1254
1255 // Make the merge point
1256 RegionNode* result_rgn = new (C, 4) RegionNode(4);
1257 Node* result_phi = new (C, 4) PhiNode(result_rgn, TypeInt::INT);
1258 Node* no_ctrl = NULL;
1259
1260 // Get counts for string and substr
1261 Node* source_cnta = basic_plus_adr(receiver, receiver, count_offset);
1262 Node* source_cnt = make_load(no_ctrl, source_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1263
1264 Node* substr_cnta = basic_plus_adr(argument, argument, count_offset);
1265 Node* substr_cnt = make_load(no_ctrl, substr_cnta, TypeInt::INT, T_INT, string_type->add_offset(count_offset));
1266
1267 // Check for substr count > string count
1268 Node* cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, source_cnt) );
1269 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::gt) );
1270 Node* if_gt = generate_slow_guard(bol, NULL);
1271 if (if_gt != NULL) {
1272 result_phi->init_req(2, intcon(-1));
1273 result_rgn->init_req(2, if_gt);
1274 }
1275
1276 if (!stopped()) {
1277 // Check for substr count == 0
1278 cmp = _gvn.transform( new(C, 3) CmpINode(substr_cnt, intcon(0)) );
1279 bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
1280 Node* if_zero = generate_slow_guard(bol, NULL);
1281 if (if_zero != NULL) {
1282 result_phi->init_req(3, intcon(0));
1283 result_rgn->init_req(3, if_zero);
1284 }
1285 }
1286
1287 if (!stopped()) {
1288 result = make_string_method_node(Op_StrIndexOf, receiver, source_cnt, argument, substr_cnt);
1289 result_phi->init_req(1, result);
1290 result_rgn->init_req(1, control());
1291 }
1292 set_control(_gvn.transform(result_rgn));
1293 record_for_igvn(result_rgn);
1294 result = _gvn.transform(result_phi);
1295
1296 } else { // Use LibraryCallKit::string_indexOf
1297 // don't intrinsify if argument isn't a constant string.
1298 if (!argument->is_Con()) {
1299 return false;
1300 }
1301 const TypeOopPtr* str_type = _gvn.type(argument)->isa_oopptr();
1302 if (str_type == NULL) {
1303 return false;
1304 }
1305 ciInstanceKlass* klass = env()->String_klass();
1306 ciObject* str_const = str_type->const_oop();
1307 if (str_const == NULL || str_const->klass() != klass) {
1308 return false;
1309 }
1310 ciInstance* str = str_const->as_instance();
1311 assert(str != NULL, "must be instance");
1312
1313 ciObject* v = str->field_value_by_offset(value_offset).as_object();
1314 int o = str->field_value_by_offset(offset_offset).as_int();
1315 int c = str->field_value_by_offset(count_offset).as_int();
1316 ciTypeArray* pat = v->as_type_array(); // pattern (argument) character array
1317
1318 // constant strings have no offset and count == length which
1319 // simplifies the resulting code somewhat so lets optimize for that.
1320 if (o != 0 || c != pat->length()) {
1321 return false;
1322 }
1323
1324 // Null check on self without removing any arguments. The argument
1325 // null check technically happens in the wrong place, which can lead to
1326 // invalid stack traces when string compare is inlined into a method
1327 // which handles NullPointerExceptions.
1328 _sp += 2;
1329 receiver = do_null_check(receiver, T_OBJECT);
1330 // No null check on the argument is needed since it's a constant String oop.
1331 _sp -= 2;
1332 if (stopped()) {
1333 return true;
1334 }
1335
1336 // The null string as a pattern always returns 0 (match at beginning of string)
1337 if (c == 0) {
1338 push(intcon(0));
1339 return true;
1340 }
1341
1342 // Generate default indexOf
1343 jchar lastChar = pat->char_at(o + (c - 1));
1344 int cache = 0;
1345 int i;
1346 for (i = 0; i < c - 1; i++) {
1347 assert(i < pat->length(), "out of range");
1348 cache |= (1 << (pat->char_at(o + i) & (sizeof(cache) * BitsPerByte - 1)));
1349 }
1350
1351 int md2 = c;
1352 for (i = 0; i < c - 1; i++) {
1353 assert(i < pat->length(), "out of range");
1354 if (pat->char_at(o + i) == lastChar) {
1355 md2 = (c - 1) - i;
1356 }
1357 }
1358
1359 result = string_indexOf(receiver, pat, o, cache, md2);
1360 }
1361
1362 push(result);
1363 return true;
1364 }
1365
1366 //--------------------------pop_math_arg--------------------------------
1367 // Pop a double argument to a math function from the stack
1368 // rounding it if necessary.
1369 Node * LibraryCallKit::pop_math_arg() {
1370 Node *arg = pop_pair();
1371 if( Matcher::strict_fp_requires_explicit_rounding && UseSSE<=1 )
1372 arg = _gvn.transform( new (C, 2) RoundDoubleNode(0, arg) );
1373 return arg;
1374 }
1375
1376 //------------------------------inline_trig----------------------------------
1377 // Inline sin/cos/tan instructions, if possible. If rounding is required, do
1378 // argument reduction which will turn into a fast/slow diamond.
1379 bool LibraryCallKit::inline_trig(vmIntrinsics::ID id) {
1380 _sp += arg_size(); // restore stack pointer
1381 Node* arg = pop_math_arg();
1382 Node* trig = NULL;
1383
1384 switch (id) {
1385 case vmIntrinsics::_dsin:
1386 trig = _gvn.transform((Node*)new (C, 2) SinDNode(arg));
1387 break;
1388 case vmIntrinsics::_dcos:
1389 trig = _gvn.transform((Node*)new (C, 2) CosDNode(arg));
1390 break;
1391 case vmIntrinsics::_dtan:
1392 trig = _gvn.transform((Node*)new (C, 2) TanDNode(arg));
1393 break;
1394 default:
1395 assert(false, "bad intrinsic was passed in");
1396 return false;
1397 }
1398
1399 // Rounding required? Check for argument reduction!
1400 if( Matcher::strict_fp_requires_explicit_rounding ) {
1401
1402 static const double pi_4 = 0.7853981633974483;
1403 static const double neg_pi_4 = -0.7853981633974483;
1404 // pi/2 in 80-bit extended precision
1405 // static const unsigned char pi_2_bits_x[] = {0x35,0xc2,0x68,0x21,0xa2,0xda,0x0f,0xc9,0xff,0x3f,0x00,0x00,0x00,0x00,0x00,0x00};
1406 // -pi/2 in 80-bit extended precision
1407 // 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};
1408 // Cutoff value for using this argument reduction technique
1409 //static const double pi_2_minus_epsilon = 1.564660403643354;
1410 //static const double neg_pi_2_plus_epsilon = -1.564660403643354;
1411
1412 // Pseudocode for sin:
1413 // if (x <= Math.PI / 4.0) {
1414 // if (x >= -Math.PI / 4.0) return fsin(x);
1415 // if (x >= -Math.PI / 2.0) return -fcos(x + Math.PI / 2.0);
1416 // } else {
1417 // if (x <= Math.PI / 2.0) return fcos(x - Math.PI / 2.0);
1418 // }
1419 // return StrictMath.sin(x);
1420
1421 // Pseudocode for cos:
1422 // if (x <= Math.PI / 4.0) {
1423 // if (x >= -Math.PI / 4.0) return fcos(x);
1424 // if (x >= -Math.PI / 2.0) return fsin(x + Math.PI / 2.0);
1425 // } else {
1426 // if (x <= Math.PI / 2.0) return -fsin(x - Math.PI / 2.0);
1427 // }
1428 // return StrictMath.cos(x);
1429
1430 // Actually, sticking in an 80-bit Intel value into C2 will be tough; it
1431 // requires a special machine instruction to load it. Instead we'll try
1432 // the 'easy' case. If we really need the extra range +/- PI/2 we'll
1433 // probably do the math inside the SIN encoding.
1434
1435 // Make the merge point
1436 RegionNode *r = new (C, 3) RegionNode(3);
1437 Node *phi = new (C, 3) PhiNode(r,Type::DOUBLE);
1438
1439 // Flatten arg so we need only 1 test
1440 Node *abs = _gvn.transform(new (C, 2) AbsDNode(arg));
1441 // Node for PI/4 constant
1442 Node *pi4 = makecon(TypeD::make(pi_4));
1443 // Check PI/4 : abs(arg)
1444 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(pi4,abs));
1445 // Check: If PI/4 < abs(arg) then go slow
1446 Node *bol = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::lt ) );
1447 // Branch either way
1448 IfNode *iff = create_and_xform_if(control(),bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
1449 set_control(opt_iff(r,iff));
1450
1451 // Set fast path result
1452 phi->init_req(2,trig);
1453
1454 // Slow path - non-blocking leaf call
1455 Node* call = NULL;
1456 switch (id) {
1457 case vmIntrinsics::_dsin:
1458 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1459 CAST_FROM_FN_PTR(address, SharedRuntime::dsin),
1460 "Sin", NULL, arg, top());
1461 break;
1462 case vmIntrinsics::_dcos:
1463 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1464 CAST_FROM_FN_PTR(address, SharedRuntime::dcos),
1465 "Cos", NULL, arg, top());
1466 break;
1467 case vmIntrinsics::_dtan:
1468 call = make_runtime_call(RC_LEAF, OptoRuntime::Math_D_D_Type(),
1469 CAST_FROM_FN_PTR(address, SharedRuntime::dtan),
1470 "Tan", NULL, arg, top());
1471 break;
1472 }
1473 assert(control()->in(0) == call, "");
1474 Node* slow_result = _gvn.transform(new (C, 1) ProjNode(call,TypeFunc::Parms));
1475 r->init_req(1,control());
1476 phi->init_req(1,slow_result);
1477
1478 // Post-merge
1479 set_control(_gvn.transform(r));
1480 record_for_igvn(r);
1481 trig = _gvn.transform(phi);
1482
1483 C->set_has_split_ifs(true); // Has chance for split-if optimization
1484 }
1485 // Push result back on JVM stack
1486 push_pair(trig);
1487 return true;
1488 }
1489
1490 //------------------------------inline_sqrt-------------------------------------
1491 // Inline square root instruction, if possible.
1492 bool LibraryCallKit::inline_sqrt(vmIntrinsics::ID id) {
1493 assert(id == vmIntrinsics::_dsqrt, "Not square root");
1494 _sp += arg_size(); // restore stack pointer
1495 push_pair(_gvn.transform(new (C, 2) SqrtDNode(0, pop_math_arg())));
1496 return true;
1497 }
1498
1499 //------------------------------inline_abs-------------------------------------
1500 // Inline absolute value instruction, if possible.
1501 bool LibraryCallKit::inline_abs(vmIntrinsics::ID id) {
1502 assert(id == vmIntrinsics::_dabs, "Not absolute value");
1503 _sp += arg_size(); // restore stack pointer
1504 push_pair(_gvn.transform(new (C, 2) AbsDNode(pop_math_arg())));
1505 return true;
1506 }
1507
1508 //------------------------------inline_exp-------------------------------------
1509 // Inline exp instructions, if possible. The Intel hardware only misses
1510 // really odd corner cases (+/- Infinity). Just uncommon-trap them.
1511 bool LibraryCallKit::inline_exp(vmIntrinsics::ID id) {
1512 assert(id == vmIntrinsics::_dexp, "Not exp");
1513
1514 // If this inlining ever returned NaN in the past, we do not intrinsify it
1515 // every again. NaN results requires StrictMath.exp handling.
1516 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1517
1518 // Do not intrinsify on older platforms which lack cmove.
1519 if (ConditionalMoveLimit == 0) return false;
1520
1521 _sp += arg_size(); // restore stack pointer
1522 Node *x = pop_math_arg();
1523 Node *result = _gvn.transform(new (C, 2) ExpDNode(0,x));
1524
1525 //-------------------
1526 //result=(result.isNaN())? StrictMath::exp():result;
1527 // Check: If isNaN() by checking result!=result? then go to Strict Math
1528 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1529 // Build the boolean node
1530 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1531
1532 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1533 // End the current control-flow path
1534 push_pair(x);
1535 // Math.exp intrinsic returned a NaN, which requires StrictMath.exp
1536 // to handle. Recompile without intrinsifying Math.exp
1537 uncommon_trap(Deoptimization::Reason_intrinsic,
1538 Deoptimization::Action_make_not_entrant);
1539 }
1540
1541 C->set_has_split_ifs(true); // Has chance for split-if optimization
1542
1543 push_pair(result);
1544
1545 return true;
1546 }
1547
1548 //------------------------------inline_pow-------------------------------------
1549 // Inline power instructions, if possible.
1550 bool LibraryCallKit::inline_pow(vmIntrinsics::ID id) {
1551 assert(id == vmIntrinsics::_dpow, "Not pow");
1552
1553 // If this inlining ever returned NaN in the past, we do not intrinsify it
1554 // every again. NaN results requires StrictMath.pow handling.
1555 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
1556
1557 // Do not intrinsify on older platforms which lack cmove.
1558 if (ConditionalMoveLimit == 0) return false;
1559
1560 // Pseudocode for pow
1561 // if (x <= 0.0) {
1562 // if ((double)((int)y)==y) { // if y is int
1563 // result = ((1&(int)y)==0)?-DPow(abs(x), y):DPow(abs(x), y)
1564 // } else {
1565 // result = NaN;
1566 // }
1567 // } else {
1568 // result = DPow(x,y);
1569 // }
1570 // if (result != result)? {
1571 // uncommon_trap();
1572 // }
1573 // return result;
1574
1575 _sp += arg_size(); // restore stack pointer
1576 Node* y = pop_math_arg();
1577 Node* x = pop_math_arg();
1578
1579 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, x, y) );
1580
1581 // Short form: if not top-level (i.e., Math.pow but inlining Math.pow
1582 // inside of something) then skip the fancy tests and just check for
1583 // NaN result.
1584 Node *result = NULL;
1585 if( jvms()->depth() >= 1 ) {
1586 result = fast_result;
1587 } else {
1588
1589 // Set the merge point for If node with condition of (x <= 0.0)
1590 // There are four possible paths to region node and phi node
1591 RegionNode *r = new (C, 4) RegionNode(4);
1592 Node *phi = new (C, 4) PhiNode(r, Type::DOUBLE);
1593
1594 // Build the first if node: if (x <= 0.0)
1595 // Node for 0 constant
1596 Node *zeronode = makecon(TypeD::ZERO);
1597 // Check x:0
1598 Node *cmp = _gvn.transform(new (C, 3) CmpDNode(x, zeronode));
1599 // Check: If (x<=0) then go complex path
1600 Node *bol1 = _gvn.transform( new (C, 2) BoolNode( cmp, BoolTest::le ) );
1601 // Branch either way
1602 IfNode *if1 = create_and_xform_if(control(),bol1, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1603 Node *opt_test = _gvn.transform(if1);
1604 //assert( opt_test->is_If(), "Expect an IfNode");
1605 IfNode *opt_if1 = (IfNode*)opt_test;
1606 // Fast path taken; set region slot 3
1607 Node *fast_taken = _gvn.transform( new (C, 1) IfFalseNode(opt_if1) );
1608 r->init_req(3,fast_taken); // Capture fast-control
1609
1610 // Fast path not-taken, i.e. slow path
1611 Node *complex_path = _gvn.transform( new (C, 1) IfTrueNode(opt_if1) );
1612
1613 // Set fast path result
1614 Node *fast_result = _gvn.transform( new (C, 3) PowDNode(0, y, x) );
1615 phi->init_req(3, fast_result);
1616
1617 // Complex path
1618 // Build the second if node (if y is int)
1619 // Node for (int)y
1620 Node *inty = _gvn.transform( new (C, 2) ConvD2INode(y));
1621 // Node for (double)((int) y)
1622 Node *doubleinty= _gvn.transform( new (C, 2) ConvI2DNode(inty));
1623 // Check (double)((int) y) : y
1624 Node *cmpinty= _gvn.transform(new (C, 3) CmpDNode(doubleinty, y));
1625 // Check if (y isn't int) then go to slow path
1626
1627 Node *bol2 = _gvn.transform( new (C, 2) BoolNode( cmpinty, BoolTest::ne ) );
1628 // Branch either way
1629 IfNode *if2 = create_and_xform_if(complex_path,bol2, PROB_STATIC_INFREQUENT, COUNT_UNKNOWN);
1630 Node *slow_path = opt_iff(r,if2); // Set region path 2
1631
1632 // Calculate DPow(abs(x), y)*(1 & (int)y)
1633 // Node for constant 1
1634 Node *conone = intcon(1);
1635 // 1& (int)y
1636 Node *signnode= _gvn.transform( new (C, 3) AndINode(conone, inty) );
1637 // zero node
1638 Node *conzero = intcon(0);
1639 // Check (1&(int)y)==0?
1640 Node *cmpeq1 = _gvn.transform(new (C, 3) CmpINode(signnode, conzero));
1641 // Check if (1&(int)y)!=0?, if so the result is negative
1642 Node *bol3 = _gvn.transform( new (C, 2) BoolNode( cmpeq1, BoolTest::ne ) );
1643 // abs(x)
1644 Node *absx=_gvn.transform( new (C, 2) AbsDNode(x));
1645 // abs(x)^y
1646 Node *absxpowy = _gvn.transform( new (C, 3) PowDNode(0, y, absx) );
1647 // -abs(x)^y
1648 Node *negabsxpowy = _gvn.transform(new (C, 2) NegDNode (absxpowy));
1649 // (1&(int)y)==1?-DPow(abs(x), y):DPow(abs(x), y)
1650 Node *signresult = _gvn.transform( CMoveNode::make(C, NULL, bol3, absxpowy, negabsxpowy, Type::DOUBLE));
1651 // Set complex path fast result
1652 phi->init_req(2, signresult);
1653
1654 static const jlong nan_bits = CONST64(0x7ff8000000000000);
1655 Node *slow_result = makecon(TypeD::make(*(double*)&nan_bits)); // return NaN
1656 r->init_req(1,slow_path);
1657 phi->init_req(1,slow_result);
1658
1659 // Post merge
1660 set_control(_gvn.transform(r));
1661 record_for_igvn(r);
1662 result=_gvn.transform(phi);
1663 }
1664
1665 //-------------------
1666 //result=(result.isNaN())? uncommon_trap():result;
1667 // Check: If isNaN() by checking result!=result? then go to Strict Math
1668 Node* cmpisnan = _gvn.transform(new (C, 3) CmpDNode(result,result));
1669 // Build the boolean node
1670 Node* bolisnum = _gvn.transform( new (C, 2) BoolNode(cmpisnan, BoolTest::eq) );
1671
1672 { BuildCutout unless(this, bolisnum, PROB_STATIC_FREQUENT);
1673 // End the current control-flow path
1674 push_pair(x);
1675 push_pair(y);
1676 // Math.pow intrinsic returned a NaN, which requires StrictMath.pow
1677 // to handle. Recompile without intrinsifying Math.pow.
1678 uncommon_trap(Deoptimization::Reason_intrinsic,
1679 Deoptimization::Action_make_not_entrant);
1680 }
1681
1682 C->set_has_split_ifs(true); // Has chance for split-if optimization
1683
1684 push_pair(result);
1685
1686 return true;
1687 }
1688
1689 //------------------------------inline_trans-------------------------------------
1690 // Inline transcendental instructions, if possible. The Intel hardware gets
1691 // these right, no funny corner cases missed.
1692 bool LibraryCallKit::inline_trans(vmIntrinsics::ID id) {
1693 _sp += arg_size(); // restore stack pointer
1694 Node* arg = pop_math_arg();
1695 Node* trans = NULL;
1696
1697 switch (id) {
1698 case vmIntrinsics::_dlog:
1699 trans = _gvn.transform((Node*)new (C, 2) LogDNode(arg));
1700 break;
1701 case vmIntrinsics::_dlog10:
1702 trans = _gvn.transform((Node*)new (C, 2) Log10DNode(arg));
1703 break;
1704 default:
1705 assert(false, "bad intrinsic was passed in");
1706 return false;
1707 }
1708
1709 // Push result back on JVM stack
1710 push_pair(trans);
1711 return true;
1712 }
1713
1714 //------------------------------runtime_math-----------------------------
1715 bool LibraryCallKit::runtime_math(const TypeFunc* call_type, address funcAddr, const char* funcName) {
1716 Node* a = NULL;
1717 Node* b = NULL;
1718
1719 assert(call_type == OptoRuntime::Math_DD_D_Type() || call_type == OptoRuntime::Math_D_D_Type(),
1720 "must be (DD)D or (D)D type");
1721
1722 // Inputs
1723 _sp += arg_size(); // restore stack pointer
1724 if (call_type == OptoRuntime::Math_DD_D_Type()) {
1725 b = pop_math_arg();
1726 }
1727 a = pop_math_arg();
1728
1729 const TypePtr* no_memory_effects = NULL;
1730 Node* trig = make_runtime_call(RC_LEAF, call_type, funcAddr, funcName,
1731 no_memory_effects,
1732 a, top(), b, b ? top() : NULL);
1733 Node* value = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+0));
1734 #ifdef ASSERT
1735 Node* value_top = _gvn.transform(new (C, 1) ProjNode(trig, TypeFunc::Parms+1));
1736 assert(value_top == top(), "second value must be top");
1737 #endif
1738
1739 push_pair(value);
1740 return true;
1741 }
1742
1743 //------------------------------inline_math_native-----------------------------
1744 bool LibraryCallKit::inline_math_native(vmIntrinsics::ID id) {
1745 switch (id) {
1746 // These intrinsics are not properly supported on all hardware
1747 case vmIntrinsics::_dcos: return Matcher::has_match_rule(Op_CosD) ? inline_trig(id) :
1748 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dcos), "COS");
1749 case vmIntrinsics::_dsin: return Matcher::has_match_rule(Op_SinD) ? inline_trig(id) :
1750 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dsin), "SIN");
1751 case vmIntrinsics::_dtan: return Matcher::has_match_rule(Op_TanD) ? inline_trig(id) :
1752 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dtan), "TAN");
1753
1754 case vmIntrinsics::_dlog: return Matcher::has_match_rule(Op_LogD) ? inline_trans(id) :
1755 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog), "LOG");
1756 case vmIntrinsics::_dlog10: return Matcher::has_match_rule(Op_Log10D) ? inline_trans(id) :
1757 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dlog10), "LOG10");
1758
1759 // These intrinsics are supported on all hardware
1760 case vmIntrinsics::_dsqrt: return Matcher::has_match_rule(Op_SqrtD) ? inline_sqrt(id) : false;
1761 case vmIntrinsics::_dabs: return Matcher::has_match_rule(Op_AbsD) ? inline_abs(id) : false;
1762
1763 // These intrinsics don't work on X86. The ad implementation doesn't
1764 // handle NaN's properly. Instead of returning infinity, the ad
1765 // implementation returns a NaN on overflow. See bug: 6304089
1766 // Once the ad implementations are fixed, change the code below
1767 // to match the intrinsics above
1768
1769 case vmIntrinsics::_dexp: return
1770 runtime_math(OptoRuntime::Math_D_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dexp), "EXP");
1771 case vmIntrinsics::_dpow: return
1772 runtime_math(OptoRuntime::Math_DD_D_Type(), CAST_FROM_FN_PTR(address, SharedRuntime::dpow), "POW");
1773
1774 // These intrinsics are not yet correctly implemented
1775 case vmIntrinsics::_datan2:
1776 return false;
1777
1778 default:
1779 ShouldNotReachHere();
1780 return false;
1781 }
1782 }
1783
1784 static bool is_simple_name(Node* n) {
1785 return (n->req() == 1 // constant
1786 || (n->is_Type() && n->as_Type()->type()->singleton())
1787 || n->is_Proj() // parameter or return value
1788 || n->is_Phi() // local of some sort
1789 );
1790 }
1791
1792 //----------------------------inline_min_max-----------------------------------
1793 bool LibraryCallKit::inline_min_max(vmIntrinsics::ID id) {
1794 push(generate_min_max(id, argument(0), argument(1)));
1795
1796 return true;
1797 }
1798
1799 Node*
1800 LibraryCallKit::generate_min_max(vmIntrinsics::ID id, Node* x0, Node* y0) {
1801 // These are the candidate return value:
1802 Node* xvalue = x0;
1803 Node* yvalue = y0;
1804
1805 if (xvalue == yvalue) {
1806 return xvalue;
1807 }
1808
1809 bool want_max = (id == vmIntrinsics::_max);
1810
1811 const TypeInt* txvalue = _gvn.type(xvalue)->isa_int();
1812 const TypeInt* tyvalue = _gvn.type(yvalue)->isa_int();
1813 if (txvalue == NULL || tyvalue == NULL) return top();
1814 // This is not really necessary, but it is consistent with a
1815 // hypothetical MaxINode::Value method:
1816 int widen = MAX2(txvalue->_widen, tyvalue->_widen);
1817
1818 // %%% This folding logic should (ideally) be in a different place.
1819 // Some should be inside IfNode, and there to be a more reliable
1820 // transformation of ?: style patterns into cmoves. We also want
1821 // more powerful optimizations around cmove and min/max.
1822
1823 // Try to find a dominating comparison of these guys.
1824 // It can simplify the index computation for Arrays.copyOf
1825 // and similar uses of System.arraycopy.
1826 // First, compute the normalized version of CmpI(x, y).
1827 int cmp_op = Op_CmpI;
1828 Node* xkey = xvalue;
1829 Node* ykey = yvalue;
1830 Node* ideal_cmpxy = _gvn.transform( new(C, 3) CmpINode(xkey, ykey) );
1831 if (ideal_cmpxy->is_Cmp()) {
1832 // E.g., if we have CmpI(length - offset, count),
1833 // it might idealize to CmpI(length, count + offset)
1834 cmp_op = ideal_cmpxy->Opcode();
1835 xkey = ideal_cmpxy->in(1);
1836 ykey = ideal_cmpxy->in(2);
1837 }
1838
1839 // Start by locating any relevant comparisons.
1840 Node* start_from = (xkey->outcnt() < ykey->outcnt()) ? xkey : ykey;
1841 Node* cmpxy = NULL;
1842 Node* cmpyx = NULL;
1843 for (DUIterator_Fast kmax, k = start_from->fast_outs(kmax); k < kmax; k++) {
1844 Node* cmp = start_from->fast_out(k);
1845 if (cmp->outcnt() > 0 && // must have prior uses
1846 cmp->in(0) == NULL && // must be context-independent
1847 cmp->Opcode() == cmp_op) { // right kind of compare
1848 if (cmp->in(1) == xkey && cmp->in(2) == ykey) cmpxy = cmp;
1849 if (cmp->in(1) == ykey && cmp->in(2) == xkey) cmpyx = cmp;
1850 }
1851 }
1852
1853 const int NCMPS = 2;
1854 Node* cmps[NCMPS] = { cmpxy, cmpyx };
1855 int cmpn;
1856 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1857 if (cmps[cmpn] != NULL) break; // find a result
1858 }
1859 if (cmpn < NCMPS) {
1860 // Look for a dominating test that tells us the min and max.
1861 int depth = 0; // Limit search depth for speed
1862 Node* dom = control();
1863 for (; dom != NULL; dom = IfNode::up_one_dom(dom, true)) {
1864 if (++depth >= 100) break;
1865 Node* ifproj = dom;
1866 if (!ifproj->is_Proj()) continue;
1867 Node* iff = ifproj->in(0);
1868 if (!iff->is_If()) continue;
1869 Node* bol = iff->in(1);
1870 if (!bol->is_Bool()) continue;
1871 Node* cmp = bol->in(1);
1872 if (cmp == NULL) continue;
1873 for (cmpn = 0; cmpn < NCMPS; cmpn++)
1874 if (cmps[cmpn] == cmp) break;
1875 if (cmpn == NCMPS) continue;
1876 BoolTest::mask btest = bol->as_Bool()->_test._test;
1877 if (ifproj->is_IfFalse()) btest = BoolTest(btest).negate();
1878 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1879 // At this point, we know that 'x btest y' is true.
1880 switch (btest) {
1881 case BoolTest::eq:
1882 // They are proven equal, so we can collapse the min/max.
1883 // Either value is the answer. Choose the simpler.
1884 if (is_simple_name(yvalue) && !is_simple_name(xvalue))
1885 return yvalue;
1886 return xvalue;
1887 case BoolTest::lt: // x < y
1888 case BoolTest::le: // x <= y
1889 return (want_max ? yvalue : xvalue);
1890 case BoolTest::gt: // x > y
1891 case BoolTest::ge: // x >= y
1892 return (want_max ? xvalue : yvalue);
1893 }
1894 }
1895 }
1896
1897 // We failed to find a dominating test.
1898 // Let's pick a test that might GVN with prior tests.
1899 Node* best_bol = NULL;
1900 BoolTest::mask best_btest = BoolTest::illegal;
1901 for (cmpn = 0; cmpn < NCMPS; cmpn++) {
1902 Node* cmp = cmps[cmpn];
1903 if (cmp == NULL) continue;
1904 for (DUIterator_Fast jmax, j = cmp->fast_outs(jmax); j < jmax; j++) {
1905 Node* bol = cmp->fast_out(j);
1906 if (!bol->is_Bool()) continue;
1907 BoolTest::mask btest = bol->as_Bool()->_test._test;
1908 if (btest == BoolTest::eq || btest == BoolTest::ne) continue;
1909 if (cmp->in(1) == ykey) btest = BoolTest(btest).commute();
1910 if (bol->outcnt() > (best_bol == NULL ? 0 : best_bol->outcnt())) {
1911 best_bol = bol->as_Bool();
1912 best_btest = btest;
1913 }
1914 }
1915 }
1916
1917 Node* answer_if_true = NULL;
1918 Node* answer_if_false = NULL;
1919 switch (best_btest) {
1920 default:
1921 if (cmpxy == NULL)
1922 cmpxy = ideal_cmpxy;
1923 best_bol = _gvn.transform( new(C, 2) BoolNode(cmpxy, BoolTest::lt) );
1924 // and fall through:
1925 case BoolTest::lt: // x < y
1926 case BoolTest::le: // x <= y
1927 answer_if_true = (want_max ? yvalue : xvalue);
1928 answer_if_false = (want_max ? xvalue : yvalue);
1929 break;
1930 case BoolTest::gt: // x > y
1931 case BoolTest::ge: // x >= y
1932 answer_if_true = (want_max ? xvalue : yvalue);
1933 answer_if_false = (want_max ? yvalue : xvalue);
1934 break;
1935 }
1936
1937 jint hi, lo;
1938 if (want_max) {
1939 // We can sharpen the minimum.
1940 hi = MAX2(txvalue->_hi, tyvalue->_hi);
1941 lo = MAX2(txvalue->_lo, tyvalue->_lo);
1942 } else {
1943 // We can sharpen the maximum.
1944 hi = MIN2(txvalue->_hi, tyvalue->_hi);
1945 lo = MIN2(txvalue->_lo, tyvalue->_lo);
1946 }
1947
1948 // Use a flow-free graph structure, to avoid creating excess control edges
1949 // which could hinder other optimizations.
1950 // Since Math.min/max is often used with arraycopy, we want
1951 // tightly_coupled_allocation to be able to see beyond min/max expressions.
1952 Node* cmov = CMoveNode::make(C, NULL, best_bol,
1953 answer_if_false, answer_if_true,
1954 TypeInt::make(lo, hi, widen));
1955
1956 return _gvn.transform(cmov);
1957
1958 /*
1959 // This is not as desirable as it may seem, since Min and Max
1960 // nodes do not have a full set of optimizations.
1961 // And they would interfere, anyway, with 'if' optimizations
1962 // and with CMoveI canonical forms.
1963 switch (id) {
1964 case vmIntrinsics::_min:
1965 result_val = _gvn.transform(new (C, 3) MinINode(x,y)); break;
1966 case vmIntrinsics::_max:
1967 result_val = _gvn.transform(new (C, 3) MaxINode(x,y)); break;
1968 default:
1969 ShouldNotReachHere();
1970 }
1971 */
1972 }
1973
1974 inline int
1975 LibraryCallKit::classify_unsafe_addr(Node* &base, Node* &offset) {
1976 const TypePtr* base_type = TypePtr::NULL_PTR;
1977 if (base != NULL) base_type = _gvn.type(base)->isa_ptr();
1978 if (base_type == NULL) {
1979 // Unknown type.
1980 return Type::AnyPtr;
1981 } else if (base_type == TypePtr::NULL_PTR) {
1982 // Since this is a NULL+long form, we have to switch to a rawptr.
1983 base = _gvn.transform( new (C, 2) CastX2PNode(offset) );
1984 offset = MakeConX(0);
1985 return Type::RawPtr;
1986 } else if (base_type->base() == Type::RawPtr) {
1987 return Type::RawPtr;
1988 } else if (base_type->isa_oopptr()) {
1989 // Base is never null => always a heap address.
1990 if (base_type->ptr() == TypePtr::NotNull) {
1991 return Type::OopPtr;
1992 }
1993 // Offset is small => always a heap address.
1994 const TypeX* offset_type = _gvn.type(offset)->isa_intptr_t();
1995 if (offset_type != NULL &&
1996 base_type->offset() == 0 && // (should always be?)
1997 offset_type->_lo >= 0 &&
1998 !MacroAssembler::needs_explicit_null_check(offset_type->_hi)) {
1999 return Type::OopPtr;
2000 }
2001 // Otherwise, it might either be oop+off or NULL+addr.
2002 return Type::AnyPtr;
2003 } else {
2004 // No information:
2005 return Type::AnyPtr;
2006 }
2007 }
2008
2009 inline Node* LibraryCallKit::make_unsafe_address(Node* base, Node* offset) {
2010 int kind = classify_unsafe_addr(base, offset);
2011 if (kind == Type::RawPtr) {
2012 return basic_plus_adr(top(), base, offset);
2013 } else {
2014 return basic_plus_adr(base, offset);
2015 }
2016 }
2017
2018 //-------------------inline_numberOfLeadingZeros_int/long-----------------------
2019 // inline int Integer.numberOfLeadingZeros(int)
2020 // inline int Long.numberOfLeadingZeros(long)
2021 bool LibraryCallKit::inline_numberOfLeadingZeros(vmIntrinsics::ID id) {
2022 assert(id == vmIntrinsics::_numberOfLeadingZeros_i || id == vmIntrinsics::_numberOfLeadingZeros_l, "not numberOfLeadingZeros");
2023 if (id == vmIntrinsics::_numberOfLeadingZeros_i && !Matcher::match_rule_supported(Op_CountLeadingZerosI)) return false;
2024 if (id == vmIntrinsics::_numberOfLeadingZeros_l && !Matcher::match_rule_supported(Op_CountLeadingZerosL)) return false;
2025 _sp += arg_size(); // restore stack pointer
2026 switch (id) {
2027 case vmIntrinsics::_numberOfLeadingZeros_i:
2028 push(_gvn.transform(new (C, 2) CountLeadingZerosINode(pop())));
2029 break;
2030 case vmIntrinsics::_numberOfLeadingZeros_l:
2031 push(_gvn.transform(new (C, 2) CountLeadingZerosLNode(pop_pair())));
2032 break;
2033 default:
2034 ShouldNotReachHere();
2035 }
2036 return true;
2037 }
2038
2039 //-------------------inline_numberOfTrailingZeros_int/long----------------------
2040 // inline int Integer.numberOfTrailingZeros(int)
2041 // inline int Long.numberOfTrailingZeros(long)
2042 bool LibraryCallKit::inline_numberOfTrailingZeros(vmIntrinsics::ID id) {
2043 assert(id == vmIntrinsics::_numberOfTrailingZeros_i || id == vmIntrinsics::_numberOfTrailingZeros_l, "not numberOfTrailingZeros");
2044 if (id == vmIntrinsics::_numberOfTrailingZeros_i && !Matcher::match_rule_supported(Op_CountTrailingZerosI)) return false;
2045 if (id == vmIntrinsics::_numberOfTrailingZeros_l && !Matcher::match_rule_supported(Op_CountTrailingZerosL)) return false;
2046 _sp += arg_size(); // restore stack pointer
2047 switch (id) {
2048 case vmIntrinsics::_numberOfTrailingZeros_i:
2049 push(_gvn.transform(new (C, 2) CountTrailingZerosINode(pop())));
2050 break;
2051 case vmIntrinsics::_numberOfTrailingZeros_l:
2052 push(_gvn.transform(new (C, 2) CountTrailingZerosLNode(pop_pair())));
2053 break;
2054 default:
2055 ShouldNotReachHere();
2056 }
2057 return true;
2058 }
2059
2060 //----------------------------inline_bitCount_int/long-----------------------
2061 // inline int Integer.bitCount(int)
2062 // inline int Long.bitCount(long)
2063 bool LibraryCallKit::inline_bitCount(vmIntrinsics::ID id) {
2064 assert(id == vmIntrinsics::_bitCount_i || id == vmIntrinsics::_bitCount_l, "not bitCount");
2065 if (id == vmIntrinsics::_bitCount_i && !Matcher::has_match_rule(Op_PopCountI)) return false;
2066 if (id == vmIntrinsics::_bitCount_l && !Matcher::has_match_rule(Op_PopCountL)) return false;
2067 _sp += arg_size(); // restore stack pointer
2068 switch (id) {
2069 case vmIntrinsics::_bitCount_i:
2070 push(_gvn.transform(new (C, 2) PopCountINode(pop())));
2071 break;
2072 case vmIntrinsics::_bitCount_l:
2073 push(_gvn.transform(new (C, 2) PopCountLNode(pop_pair())));
2074 break;
2075 default:
2076 ShouldNotReachHere();
2077 }
2078 return true;
2079 }
2080
2081 //----------------------------inline_reverseBytes_int/long/char/short-------------------
2082 // inline Integer.reverseBytes(int)
2083 // inline Long.reverseBytes(long)
2084 // inline Character.reverseBytes(char)
2085 // inline Short.reverseBytes(short)
2086 bool LibraryCallKit::inline_reverseBytes(vmIntrinsics::ID id) {
2087 assert(id == vmIntrinsics::_reverseBytes_i || id == vmIntrinsics::_reverseBytes_l ||
2088 id == vmIntrinsics::_reverseBytes_c || id == vmIntrinsics::_reverseBytes_s,
2089 "not reverse Bytes");
2090 if (id == vmIntrinsics::_reverseBytes_i && !Matcher::has_match_rule(Op_ReverseBytesI)) return false;
2091 if (id == vmIntrinsics::_reverseBytes_l && !Matcher::has_match_rule(Op_ReverseBytesL)) return false;
2092 if (id == vmIntrinsics::_reverseBytes_c && !Matcher::has_match_rule(Op_ReverseBytesUS)) return false;
2093 if (id == vmIntrinsics::_reverseBytes_s && !Matcher::has_match_rule(Op_ReverseBytesS)) return false;
2094 _sp += arg_size(); // restore stack pointer
2095 switch (id) {
2096 case vmIntrinsics::_reverseBytes_i:
2097 push(_gvn.transform(new (C, 2) ReverseBytesINode(0, pop())));
2098 break;
2099 case vmIntrinsics::_reverseBytes_l:
2100 push_pair(_gvn.transform(new (C, 2) ReverseBytesLNode(0, pop_pair())));
2101 break;
2102 case vmIntrinsics::_reverseBytes_c:
2103 push(_gvn.transform(new (C, 2) ReverseBytesUSNode(0, pop())));
2104 break;
2105 case vmIntrinsics::_reverseBytes_s:
2106 push(_gvn.transform(new (C, 2) ReverseBytesSNode(0, pop())));
2107 break;
2108 default:
2109 ;
2110 }
2111 return true;
2112 }
2113
2114 //----------------------------inline_unsafe_access----------------------------
2115
2116 const static BasicType T_ADDRESS_HOLDER = T_LONG;
2117
2118 // Helper that guards and inserts a G1 pre-barrier.
2119 void LibraryCallKit::insert_g1_pre_barrier(Node* base_oop, Node* offset, Node* pre_val) {
2120 assert(UseG1GC, "should not call this otherwise");
2121
2122 // We could be accessing the referent field of a reference object. If so, when G1
2123 // is enabled, we need to log the value in the referent field in an SATB buffer.
2124 // This routine performs some compile time filters and generates suitable
2125 // runtime filters that guard the pre-barrier code.
2126
2127 // Some compile time checks.
2128
2129 // If offset is a constant, is it java_lang_ref_Reference::_reference_offset?
2130 const TypeX* otype = offset->find_intptr_t_type();
2131 if (otype != NULL && otype->is_con() &&
2132 otype->get_con() != java_lang_ref_Reference::referent_offset) {
2133 // Constant offset but not the reference_offset so just return
2134 return;
2135 }
2136
2137 // We only need to generate the runtime guards for instances.
2138 const TypeOopPtr* btype = base_oop->bottom_type()->isa_oopptr();
2139 if (btype != NULL) {
2140 if (btype->isa_aryptr()) {
2141 // Array type so nothing to do
2142 return;
2143 }
2144
2145 const TypeInstPtr* itype = btype->isa_instptr();
2146 if (itype != NULL) {
2147 // Can the klass of base_oop be statically determined
2148 // to be _not_ a sub-class of Reference?
2149 ciKlass* klass = itype->klass();
2150 if (klass->is_subtype_of(env()->Reference_klass()) &&
2151 !env()->Reference_klass()->is_subtype_of(klass)) {
2152 return;
2153 }
2154 }
2155 }
2156
2157 // The compile time filters did not reject base_oop/offset so
2158 // we need to generate the following runtime filters
2159 //
2160 // if (offset == java_lang_ref_Reference::_reference_offset) {
2161 // if (base != null) {
2162 // if (instance_of(base, java.lang.ref.Reference)) {
2163 // pre_barrier(_, pre_val, ...);
2164 // }
2165 // }
2166 // }
2167
2168 float likely = PROB_LIKELY(0.999);
2169 float unlikely = PROB_UNLIKELY(0.999);
2170
2171 IdealKit ideal(this);
2172 #define __ ideal.
2173
2174 Node* referent_off = __ ConX(java_lang_ref_Reference::referent_offset);
2175
2176 __ if_then(offset, BoolTest::eq, referent_off, unlikely); {
2177 __ if_then(base_oop, BoolTest::ne, null(), likely); {
2178
2179 // Update graphKit memory and control from IdealKit.
2180 sync_kit(ideal);
2181
2182 Node* ref_klass_con = makecon(TypeKlassPtr::make(env()->Reference_klass()));
2183 Node* is_instof = gen_instanceof(base_oop, ref_klass_con);
2184
2185 // Update IdealKit memory and control from graphKit.
2186 __ sync_kit(this);
2187
2188 Node* one = __ ConI(1);
2189
2190 __ if_then(is_instof, BoolTest::eq, one, unlikely); {
2191
2192 // Update graphKit from IdeakKit.
2193 sync_kit(ideal);
2194
2195 // Use the pre-barrier to record the value in the referent field
2196 pre_barrier(false /* do_load */,
2197 __ ctrl(),
2198 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
2199 pre_val /* pre_val */,
2200 T_OBJECT);
2201
2202 // Update IdealKit from graphKit.
2203 __ sync_kit(this);
2204
2205 } __ end_if(); // _ref_type != ref_none
2206 } __ end_if(); // base != NULL
2207 } __ end_if(); // offset == referent_offset
2208
2209 // Final sync IdealKit and GraphKit.
2210 final_sync(ideal);
2211 #undef __
2212 }
2213
2214
2215 // Interpret Unsafe.fieldOffset cookies correctly:
2216 extern jlong Unsafe_field_offset_to_byte_offset(jlong field_offset);
2217
2218 bool LibraryCallKit::inline_unsafe_access(bool is_native_ptr, bool is_store, BasicType type, bool is_volatile) {
2219 if (callee()->is_static()) return false; // caller must have the capability!
2220
2221 #ifndef PRODUCT
2222 {
2223 ResourceMark rm;
2224 // Check the signatures.
2225 ciSignature* sig = signature();
2226 #ifdef ASSERT
2227 if (!is_store) {
2228 // Object getObject(Object base, int/long offset), etc.
2229 BasicType rtype = sig->return_type()->basic_type();
2230 if (rtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::getAddress_name())
2231 rtype = T_ADDRESS; // it is really a C void*
2232 assert(rtype == type, "getter must return the expected value");
2233 if (!is_native_ptr) {
2234 assert(sig->count() == 2, "oop getter has 2 arguments");
2235 assert(sig->type_at(0)->basic_type() == T_OBJECT, "getter base is object");
2236 assert(sig->type_at(1)->basic_type() == T_LONG, "getter offset is correct");
2237 } else {
2238 assert(sig->count() == 1, "native getter has 1 argument");
2239 assert(sig->type_at(0)->basic_type() == T_LONG, "getter base is long");
2240 }
2241 } else {
2242 // void putObject(Object base, int/long offset, Object x), etc.
2243 assert(sig->return_type()->basic_type() == T_VOID, "putter must not return a value");
2244 if (!is_native_ptr) {
2245 assert(sig->count() == 3, "oop putter has 3 arguments");
2246 assert(sig->type_at(0)->basic_type() == T_OBJECT, "putter base is object");
2247 assert(sig->type_at(1)->basic_type() == T_LONG, "putter offset is correct");
2248 } else {
2249 assert(sig->count() == 2, "native putter has 2 arguments");
2250 assert(sig->type_at(0)->basic_type() == T_LONG, "putter base is long");
2251 }
2252 BasicType vtype = sig->type_at(sig->count()-1)->basic_type();
2253 if (vtype == T_ADDRESS_HOLDER && callee()->name() == ciSymbol::putAddress_name())
2254 vtype = T_ADDRESS; // it is really a C void*
2255 assert(vtype == type, "putter must accept the expected value");
2256 }
2257 #endif // ASSERT
2258 }
2259 #endif //PRODUCT
2260
2261 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2262
2263 int type_words = type2size[ (type == T_ADDRESS) ? T_LONG : type ];
2264
2265 // Argument words: "this" plus (oop/offset) or (lo/hi) args plus maybe 1 or 2 value words
2266 int nargs = 1 + (is_native_ptr ? 2 : 3) + (is_store ? type_words : 0);
2267
2268 debug_only(int saved_sp = _sp);
2269 _sp += nargs;
2270
2271 Node* val;
2272 debug_only(val = (Node*)(uintptr_t)-1);
2273
2274
2275 if (is_store) {
2276 // Get the value being stored. (Pop it first; it was pushed last.)
2277 switch (type) {
2278 case T_DOUBLE:
2279 case T_LONG:
2280 case T_ADDRESS:
2281 val = pop_pair();
2282 break;
2283 default:
2284 val = pop();
2285 }
2286 }
2287
2288 // Build address expression. See the code in inline_unsafe_prefetch.
2289 Node *adr;
2290 Node *heap_base_oop = top();
2291 Node* offset = top();
2292
2293 if (!is_native_ptr) {
2294 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2295 offset = pop_pair();
2296 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2297 Node* base = pop();
2298 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2299 // to be plain byte offsets, which are also the same as those accepted
2300 // by oopDesc::field_base.
2301 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2302 "fieldOffset must be byte-scaled");
2303 // 32-bit machines ignore the high half!
2304 offset = ConvL2X(offset);
2305 adr = make_unsafe_address(base, offset);
2306 heap_base_oop = base;
2307 } else {
2308 Node* ptr = pop_pair();
2309 // Adjust Java long to machine word:
2310 ptr = ConvL2X(ptr);
2311 adr = make_unsafe_address(NULL, ptr);
2312 }
2313
2314 // Pop receiver last: it was pushed first.
2315 Node *receiver = pop();
2316
2317 assert(saved_sp == _sp, "must have correct argument count");
2318
2319 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2320
2321 // First guess at the value type.
2322 const Type *value_type = Type::get_const_basic_type(type);
2323
2324 // Try to categorize the address. If it comes up as TypeJavaPtr::BOTTOM,
2325 // there was not enough information to nail it down.
2326 Compile::AliasType* alias_type = C->alias_type(adr_type);
2327 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2328
2329 // We will need memory barriers unless we can determine a unique
2330 // alias category for this reference. (Note: If for some reason
2331 // the barriers get omitted and the unsafe reference begins to "pollute"
2332 // the alias analysis of the rest of the graph, either Compile::can_alias
2333 // or Compile::must_alias will throw a diagnostic assert.)
2334 bool need_mem_bar = (alias_type->adr_type() == TypeOopPtr::BOTTOM);
2335
2336 // If we are reading the value of the referent field of a Reference
2337 // object (either by using Unsafe directly or through reflection)
2338 // then, if G1 is enabled, we need to record the referent in an
2339 // SATB log buffer using the pre-barrier mechanism.
2340 bool need_read_barrier = UseG1GC && !is_native_ptr && !is_store &&
2341 offset != top() && heap_base_oop != top();
2342
2343 if (!is_store && type == T_OBJECT) {
2344 // Attempt to infer a sharper value type from the offset and base type.
2345 ciKlass* sharpened_klass = NULL;
2346
2347 // See if it is an instance field, with an object type.
2348 if (alias_type->field() != NULL) {
2349 assert(!is_native_ptr, "native pointer op cannot use a java address");
2350 if (alias_type->field()->type()->is_klass()) {
2351 sharpened_klass = alias_type->field()->type()->as_klass();
2352 }
2353 }
2354
2355 // See if it is a narrow oop array.
2356 if (adr_type->isa_aryptr()) {
2357 if (adr_type->offset() >= objArrayOopDesc::base_offset_in_bytes()) {
2358 const TypeOopPtr *elem_type = adr_type->is_aryptr()->elem()->isa_oopptr();
2359 if (elem_type != NULL) {
2360 sharpened_klass = elem_type->klass();
2361 }
2362 }
2363 }
2364
2365 if (sharpened_klass != NULL) {
2366 const TypeOopPtr* tjp = TypeOopPtr::make_from_klass(sharpened_klass);
2367
2368 // Sharpen the value type.
2369 value_type = tjp;
2370
2371 #ifndef PRODUCT
2372 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
2373 tty->print(" from base type: "); adr_type->dump();
2374 tty->print(" sharpened value: "); value_type->dump();
2375 }
2376 #endif
2377 }
2378 }
2379
2380 // Null check on self without removing any arguments. The argument
2381 // null check technically happens in the wrong place, which can lead to
2382 // invalid stack traces when the primitive is inlined into a method
2383 // which handles NullPointerExceptions.
2384 _sp += nargs;
2385 do_null_check(receiver, T_OBJECT);
2386 _sp -= nargs;
2387 if (stopped()) {
2388 return true;
2389 }
2390 // Heap pointers get a null-check from the interpreter,
2391 // as a courtesy. However, this is not guaranteed by Unsafe,
2392 // and it is not possible to fully distinguish unintended nulls
2393 // from intended ones in this API.
2394
2395 if (is_volatile) {
2396 // We need to emit leading and trailing CPU membars (see below) in
2397 // addition to memory membars when is_volatile. This is a little
2398 // too strong, but avoids the need to insert per-alias-type
2399 // volatile membars (for stores; compare Parse::do_put_xxx), which
2400 // we cannot do effectively here because we probably only have a
2401 // rough approximation of type.
2402 need_mem_bar = true;
2403 // For Stores, place a memory ordering barrier now.
2404 if (is_store)
2405 insert_mem_bar(Op_MemBarRelease);
2406 }
2407
2408 // Memory barrier to prevent normal and 'unsafe' accesses from
2409 // bypassing each other. Happens after null checks, so the
2410 // exception paths do not take memory state from the memory barrier,
2411 // so there's no problems making a strong assert about mixing users
2412 // of safe & unsafe memory. Otherwise fails in a CTW of rt.jar
2413 // around 5701, class sun/reflect/UnsafeBooleanFieldAccessorImpl.
2414 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2415
2416 if (!is_store) {
2417 Node* p = make_load(control(), adr, value_type, type, adr_type, is_volatile);
2418 // load value and push onto stack
2419 switch (type) {
2420 case T_BOOLEAN:
2421 case T_CHAR:
2422 case T_BYTE:
2423 case T_SHORT:
2424 case T_INT:
2425 case T_FLOAT:
2426 push(p);
2427 break;
2428 case T_OBJECT:
2429 if (need_read_barrier) {
2430 insert_g1_pre_barrier(heap_base_oop, offset, p);
2431 }
2432 push(p);
2433 break;
2434 case T_ADDRESS:
2435 // Cast to an int type.
2436 p = _gvn.transform( new (C, 2) CastP2XNode(NULL,p) );
2437 p = ConvX2L(p);
2438 push_pair(p);
2439 break;
2440 case T_DOUBLE:
2441 case T_LONG:
2442 push_pair( p );
2443 break;
2444 default: ShouldNotReachHere();
2445 }
2446 } else {
2447 // place effect of store into memory
2448 switch (type) {
2449 case T_DOUBLE:
2450 val = dstore_rounding(val);
2451 break;
2452 case T_ADDRESS:
2453 // Repackage the long as a pointer.
2454 val = ConvL2X(val);
2455 val = _gvn.transform( new (C, 2) CastX2PNode(val) );
2456 break;
2457 }
2458
2459 if (type != T_OBJECT ) {
2460 (void) store_to_memory(control(), adr, val, type, adr_type, is_volatile);
2461 } else {
2462 // Possibly an oop being stored to Java heap or native memory
2463 if (!TypePtr::NULL_PTR->higher_equal(_gvn.type(heap_base_oop))) {
2464 // oop to Java heap.
2465 (void) store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2466 } else {
2467 // We can't tell at compile time if we are storing in the Java heap or outside
2468 // of it. So we need to emit code to conditionally do the proper type of
2469 // store.
2470
2471 IdealKit ideal(this);
2472 #define __ ideal.
2473 // QQQ who knows what probability is here??
2474 __ if_then(heap_base_oop, BoolTest::ne, null(), PROB_UNLIKELY(0.999)); {
2475 // Sync IdealKit and graphKit.
2476 sync_kit(ideal);
2477 Node* st = store_oop_to_unknown(control(), heap_base_oop, adr, adr_type, val, type);
2478 // Update IdealKit memory.
2479 __ sync_kit(this);
2480 } __ else_(); {
2481 __ store(__ ctrl(), adr, val, type, alias_type->index(), is_volatile);
2482 } __ end_if();
2483 // Final sync IdealKit and GraphKit.
2484 final_sync(ideal);
2485 #undef __
2486 }
2487 }
2488 }
2489
2490 if (is_volatile) {
2491 if (!is_store)
2492 insert_mem_bar(Op_MemBarAcquire);
2493 else
2494 insert_mem_bar(Op_MemBarVolatile);
2495 }
2496
2497 if (need_mem_bar) insert_mem_bar(Op_MemBarCPUOrder);
2498
2499 return true;
2500 }
2501
2502 //----------------------------inline_unsafe_prefetch----------------------------
2503
2504 bool LibraryCallKit::inline_unsafe_prefetch(bool is_native_ptr, bool is_store, bool is_static) {
2505 #ifndef PRODUCT
2506 {
2507 ResourceMark rm;
2508 // Check the signatures.
2509 ciSignature* sig = signature();
2510 #ifdef ASSERT
2511 // Object getObject(Object base, int/long offset), etc.
2512 BasicType rtype = sig->return_type()->basic_type();
2513 if (!is_native_ptr) {
2514 assert(sig->count() == 2, "oop prefetch has 2 arguments");
2515 assert(sig->type_at(0)->basic_type() == T_OBJECT, "prefetch base is object");
2516 assert(sig->type_at(1)->basic_type() == T_LONG, "prefetcha offset is correct");
2517 } else {
2518 assert(sig->count() == 1, "native prefetch has 1 argument");
2519 assert(sig->type_at(0)->basic_type() == T_LONG, "prefetch base is long");
2520 }
2521 #endif // ASSERT
2522 }
2523 #endif // !PRODUCT
2524
2525 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2526
2527 // Argument words: "this" if not static, plus (oop/offset) or (lo/hi) args
2528 int nargs = (is_static ? 0 : 1) + (is_native_ptr ? 2 : 3);
2529
2530 debug_only(int saved_sp = _sp);
2531 _sp += nargs;
2532
2533 // Build address expression. See the code in inline_unsafe_access.
2534 Node *adr;
2535 if (!is_native_ptr) {
2536 // The offset is a value produced by Unsafe.staticFieldOffset or Unsafe.objectFieldOffset
2537 Node* offset = pop_pair();
2538 // The base is either a Java object or a value produced by Unsafe.staticFieldBase
2539 Node* base = pop();
2540 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2541 // to be plain byte offsets, which are also the same as those accepted
2542 // by oopDesc::field_base.
2543 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
2544 "fieldOffset must be byte-scaled");
2545 // 32-bit machines ignore the high half!
2546 offset = ConvL2X(offset);
2547 adr = make_unsafe_address(base, offset);
2548 } else {
2549 Node* ptr = pop_pair();
2550 // Adjust Java long to machine word:
2551 ptr = ConvL2X(ptr);
2552 adr = make_unsafe_address(NULL, ptr);
2553 }
2554
2555 if (is_static) {
2556 assert(saved_sp == _sp, "must have correct argument count");
2557 } else {
2558 // Pop receiver last: it was pushed first.
2559 Node *receiver = pop();
2560 assert(saved_sp == _sp, "must have correct argument count");
2561
2562 // Null check on self without removing any arguments. The argument
2563 // null check technically happens in the wrong place, which can lead to
2564 // invalid stack traces when the primitive is inlined into a method
2565 // which handles NullPointerExceptions.
2566 _sp += nargs;
2567 do_null_check(receiver, T_OBJECT);
2568 _sp -= nargs;
2569 if (stopped()) {
2570 return true;
2571 }
2572 }
2573
2574 // Generate the read or write prefetch
2575 Node *prefetch;
2576 if (is_store) {
2577 prefetch = new (C, 3) PrefetchWriteNode(i_o(), adr);
2578 } else {
2579 prefetch = new (C, 3) PrefetchReadNode(i_o(), adr);
2580 }
2581 prefetch->init_req(0, control());
2582 set_i_o(_gvn.transform(prefetch));
2583
2584 return true;
2585 }
2586
2587 //----------------------------inline_unsafe_CAS----------------------------
2588
2589 bool LibraryCallKit::inline_unsafe_CAS(BasicType type) {
2590 // This basic scheme here is the same as inline_unsafe_access, but
2591 // differs in enough details that combining them would make the code
2592 // overly confusing. (This is a true fact! I originally combined
2593 // them, but even I was confused by it!) As much code/comments as
2594 // possible are retained from inline_unsafe_access though to make
2595 // the correspondences clearer. - dl
2596
2597 if (callee()->is_static()) return false; // caller must have the capability!
2598
2599 #ifndef PRODUCT
2600 {
2601 ResourceMark rm;
2602 // Check the signatures.
2603 ciSignature* sig = signature();
2604 #ifdef ASSERT
2605 BasicType rtype = sig->return_type()->basic_type();
2606 assert(rtype == T_BOOLEAN, "CAS must return boolean");
2607 assert(sig->count() == 4, "CAS has 4 arguments");
2608 assert(sig->type_at(0)->basic_type() == T_OBJECT, "CAS base is object");
2609 assert(sig->type_at(1)->basic_type() == T_LONG, "CAS offset is long");
2610 #endif // ASSERT
2611 }
2612 #endif //PRODUCT
2613
2614 // number of stack slots per value argument (1 or 2)
2615 int type_words = type2size[type];
2616
2617 // Cannot inline wide CAS on machines that don't support it natively
2618 if (type2aelembytes(type) > BytesPerInt && !VM_Version::supports_cx8())
2619 return false;
2620
2621 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2622
2623 // Argument words: "this" plus oop plus offset plus oldvalue plus newvalue;
2624 int nargs = 1 + 1 + 2 + type_words + type_words;
2625
2626 // pop arguments: newval, oldval, offset, base, and receiver
2627 debug_only(int saved_sp = _sp);
2628 _sp += nargs;
2629 Node* newval = (type_words == 1) ? pop() : pop_pair();
2630 Node* oldval = (type_words == 1) ? pop() : pop_pair();
2631 Node *offset = pop_pair();
2632 Node *base = pop();
2633 Node *receiver = pop();
2634 assert(saved_sp == _sp, "must have correct argument count");
2635
2636 // Null check receiver.
2637 _sp += nargs;
2638 do_null_check(receiver, T_OBJECT);
2639 _sp -= nargs;
2640 if (stopped()) {
2641 return true;
2642 }
2643
2644 // Build field offset expression.
2645 // We currently rely on the cookies produced by Unsafe.xxxFieldOffset
2646 // to be plain byte offsets, which are also the same as those accepted
2647 // by oopDesc::field_base.
2648 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2649 // 32-bit machines ignore the high half of long offsets
2650 offset = ConvL2X(offset);
2651 Node* adr = make_unsafe_address(base, offset);
2652 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2653
2654 // (Unlike inline_unsafe_access, there seems no point in trying
2655 // to refine types. Just use the coarse types here.
2656 const Type *value_type = Type::get_const_basic_type(type);
2657 Compile::AliasType* alias_type = C->alias_type(adr_type);
2658 assert(alias_type->index() != Compile::AliasIdxBot, "no bare pointers here");
2659 int alias_idx = C->get_alias_index(adr_type);
2660
2661 // Memory-model-wise, a CAS acts like a little synchronized block,
2662 // so needs barriers on each side. These don't translate into
2663 // actual barriers on most machines, but we still need rest of
2664 // compiler to respect ordering.
2665
2666 insert_mem_bar(Op_MemBarRelease);
2667 insert_mem_bar(Op_MemBarCPUOrder);
2668
2669 // 4984716: MemBars must be inserted before this
2670 // memory node in order to avoid a false
2671 // dependency which will confuse the scheduler.
2672 Node *mem = memory(alias_idx);
2673
2674 // For now, we handle only those cases that actually exist: ints,
2675 // longs, and Object. Adding others should be straightforward.
2676 Node* cas;
2677 switch(type) {
2678 case T_INT:
2679 cas = _gvn.transform(new (C, 5) CompareAndSwapINode(control(), mem, adr, newval, oldval));
2680 break;
2681 case T_LONG:
2682 cas = _gvn.transform(new (C, 5) CompareAndSwapLNode(control(), mem, adr, newval, oldval));
2683 break;
2684 case T_OBJECT:
2685 // Transformation of a value which could be NULL pointer (CastPP #NULL)
2686 // could be delayed during Parse (for example, in adjust_map_after_if()).
2687 // Execute transformation here to avoid barrier generation in such case.
2688 if (_gvn.type(newval) == TypePtr::NULL_PTR)
2689 newval = _gvn.makecon(TypePtr::NULL_PTR);
2690
2691 // Reference stores need a store barrier.
2692 // (They don't if CAS fails, but it isn't worth checking.)
2693 pre_barrier(true /* do_load*/,
2694 control(), base, adr, alias_idx, newval, value_type->make_oopptr(),
2695 NULL /* pre_val*/,
2696 T_OBJECT);
2697 #ifdef _LP64
2698 if (adr->bottom_type()->is_ptr_to_narrowoop()) {
2699 Node *newval_enc = _gvn.transform(new (C, 2) EncodePNode(newval, newval->bottom_type()->make_narrowoop()));
2700 Node *oldval_enc = _gvn.transform(new (C, 2) EncodePNode(oldval, oldval->bottom_type()->make_narrowoop()));
2701 cas = _gvn.transform(new (C, 5) CompareAndSwapNNode(control(), mem, adr,
2702 newval_enc, oldval_enc));
2703 } else
2704 #endif
2705 {
2706 cas = _gvn.transform(new (C, 5) CompareAndSwapPNode(control(), mem, adr, newval, oldval));
2707 }
2708 post_barrier(control(), cas, base, adr, alias_idx, newval, T_OBJECT, true);
2709 break;
2710 default:
2711 ShouldNotReachHere();
2712 break;
2713 }
2714
2715 // SCMemProjNodes represent the memory state of CAS. Their main
2716 // role is to prevent CAS nodes from being optimized away when their
2717 // results aren't used.
2718 Node* proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
2719 set_memory(proj, alias_idx);
2720
2721 // Add the trailing membar surrounding the access
2722 insert_mem_bar(Op_MemBarCPUOrder);
2723 insert_mem_bar(Op_MemBarAcquire);
2724
2725 push(cas);
2726 return true;
2727 }
2728
2729 bool LibraryCallKit::inline_unsafe_ordered_store(BasicType type) {
2730 // This is another variant of inline_unsafe_access, differing in
2731 // that it always issues store-store ("release") barrier and ensures
2732 // store-atomicity (which only matters for "long").
2733
2734 if (callee()->is_static()) return false; // caller must have the capability!
2735
2736 #ifndef PRODUCT
2737 {
2738 ResourceMark rm;
2739 // Check the signatures.
2740 ciSignature* sig = signature();
2741 #ifdef ASSERT
2742 BasicType rtype = sig->return_type()->basic_type();
2743 assert(rtype == T_VOID, "must return void");
2744 assert(sig->count() == 3, "has 3 arguments");
2745 assert(sig->type_at(0)->basic_type() == T_OBJECT, "base is object");
2746 assert(sig->type_at(1)->basic_type() == T_LONG, "offset is long");
2747 #endif // ASSERT
2748 }
2749 #endif //PRODUCT
2750
2751 // number of stack slots per value argument (1 or 2)
2752 int type_words = type2size[type];
2753
2754 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
2755
2756 // Argument words: "this" plus oop plus offset plus value;
2757 int nargs = 1 + 1 + 2 + type_words;
2758
2759 // pop arguments: val, offset, base, and receiver
2760 debug_only(int saved_sp = _sp);
2761 _sp += nargs;
2762 Node* val = (type_words == 1) ? pop() : pop_pair();
2763 Node *offset = pop_pair();
2764 Node *base = pop();
2765 Node *receiver = pop();
2766 assert(saved_sp == _sp, "must have correct argument count");
2767
2768 // Null check receiver.
2769 _sp += nargs;
2770 do_null_check(receiver, T_OBJECT);
2771 _sp -= nargs;
2772 if (stopped()) {
2773 return true;
2774 }
2775
2776 // Build field offset expression.
2777 assert(Unsafe_field_offset_to_byte_offset(11) == 11, "fieldOffset must be byte-scaled");
2778 // 32-bit machines ignore the high half of long offsets
2779 offset = ConvL2X(offset);
2780 Node* adr = make_unsafe_address(base, offset);
2781 const TypePtr *adr_type = _gvn.type(adr)->isa_ptr();
2782 const Type *value_type = Type::get_const_basic_type(type);
2783 Compile::AliasType* alias_type = C->alias_type(adr_type);
2784
2785 insert_mem_bar(Op_MemBarRelease);
2786 insert_mem_bar(Op_MemBarCPUOrder);
2787 // Ensure that the store is atomic for longs:
2788 bool require_atomic_access = true;
2789 Node* store;
2790 if (type == T_OBJECT) // reference stores need a store barrier.
2791 store = store_oop_to_unknown(control(), base, adr, adr_type, val, type);
2792 else {
2793 store = store_to_memory(control(), adr, val, type, adr_type, require_atomic_access);
2794 }
2795 insert_mem_bar(Op_MemBarCPUOrder);
2796 return true;
2797 }
2798
2799 bool LibraryCallKit::inline_unsafe_allocate() {
2800 if (callee()->is_static()) return false; // caller must have the capability!
2801 int nargs = 1 + 1;
2802 assert(signature()->size() == nargs-1, "alloc has 1 argument");
2803 null_check_receiver(callee()); // check then ignore argument(0)
2804 _sp += nargs; // set original stack for use by uncommon_trap
2805 Node* cls = do_null_check(argument(1), T_OBJECT);
2806 _sp -= nargs;
2807 if (stopped()) return true;
2808
2809 Node* kls = load_klass_from_mirror(cls, false, nargs, NULL, 0);
2810 _sp += nargs; // set original stack for use by uncommon_trap
2811 kls = do_null_check(kls, T_OBJECT);
2812 _sp -= nargs;
2813 if (stopped()) return true; // argument was like int.class
2814
2815 // Note: The argument might still be an illegal value like
2816 // Serializable.class or Object[].class. The runtime will handle it.
2817 // But we must make an explicit check for initialization.
2818 Node* insp = basic_plus_adr(kls, in_bytes(instanceKlass::init_state_offset()));
2819 // Use T_BOOLEAN for instanceKlass::_init_state so the compiler
2820 // can generate code to load it as unsigned byte.
2821 Node* inst = make_load(NULL, insp, TypeInt::UBYTE, T_BOOLEAN);
2822 Node* bits = intcon(instanceKlass::fully_initialized);
2823 Node* test = _gvn.transform( new (C, 3) SubINode(inst, bits) );
2824 // The 'test' is non-zero if we need to take a slow path.
2825
2826 Node* obj = new_instance(kls, test);
2827 push(obj);
2828
2829 return true;
2830 }
2831
2832 //------------------------inline_native_time_funcs--------------
2833 // inline code for System.currentTimeMillis() and System.nanoTime()
2834 // these have the same type and signature
2835 bool LibraryCallKit::inline_native_time_funcs(bool isNano) {
2836 address funcAddr = isNano ? CAST_FROM_FN_PTR(address, os::javaTimeNanos) :
2837 CAST_FROM_FN_PTR(address, os::javaTimeMillis);
2838 const char * funcName = isNano ? "nanoTime" : "currentTimeMillis";
2839 const TypeFunc *tf = OptoRuntime::current_time_millis_Type();
2840 const TypePtr* no_memory_effects = NULL;
2841 Node* time = make_runtime_call(RC_LEAF, tf, funcAddr, funcName, no_memory_effects);
2842 Node* value = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms+0));
2843 #ifdef ASSERT
2844 Node* value_top = _gvn.transform(new (C, 1) ProjNode(time, TypeFunc::Parms + 1));
2845 assert(value_top == top(), "second value must be top");
2846 #endif
2847 push_pair(value);
2848 return true;
2849 }
2850
2851 //------------------------inline_native_currentThread------------------
2852 bool LibraryCallKit::inline_native_currentThread() {
2853 Node* junk = NULL;
2854 push(generate_current_thread(junk));
2855 return true;
2856 }
2857
2858 //------------------------inline_native_isInterrupted------------------
2859 bool LibraryCallKit::inline_native_isInterrupted() {
2860 const int nargs = 1+1; // receiver + boolean
2861 assert(nargs == arg_size(), "sanity");
2862 // Add a fast path to t.isInterrupted(clear_int):
2863 // (t == Thread.current() && (!TLS._osthread._interrupted || !clear_int))
2864 // ? TLS._osthread._interrupted : /*slow path:*/ t.isInterrupted(clear_int)
2865 // So, in the common case that the interrupt bit is false,
2866 // we avoid making a call into the VM. Even if the interrupt bit
2867 // is true, if the clear_int argument is false, we avoid the VM call.
2868 // However, if the receiver is not currentThread, we must call the VM,
2869 // because there must be some locking done around the operation.
2870
2871 // We only go to the fast case code if we pass two guards.
2872 // Paths which do not pass are accumulated in the slow_region.
2873 RegionNode* slow_region = new (C, 1) RegionNode(1);
2874 record_for_igvn(slow_region);
2875 RegionNode* result_rgn = new (C, 4) RegionNode(1+3); // fast1, fast2, slow
2876 PhiNode* result_val = new (C, 4) PhiNode(result_rgn, TypeInt::BOOL);
2877 enum { no_int_result_path = 1,
2878 no_clear_result_path = 2,
2879 slow_result_path = 3
2880 };
2881
2882 // (a) Receiving thread must be the current thread.
2883 Node* rec_thr = argument(0);
2884 Node* tls_ptr = NULL;
2885 Node* cur_thr = generate_current_thread(tls_ptr);
2886 Node* cmp_thr = _gvn.transform( new (C, 3) CmpPNode(cur_thr, rec_thr) );
2887 Node* bol_thr = _gvn.transform( new (C, 2) BoolNode(cmp_thr, BoolTest::ne) );
2888
2889 bool known_current_thread = (_gvn.type(bol_thr) == TypeInt::ZERO);
2890 if (!known_current_thread)
2891 generate_slow_guard(bol_thr, slow_region);
2892
2893 // (b) Interrupt bit on TLS must be false.
2894 Node* p = basic_plus_adr(top()/*!oop*/, tls_ptr, in_bytes(JavaThread::osthread_offset()));
2895 Node* osthread = make_load(NULL, p, TypeRawPtr::NOTNULL, T_ADDRESS);
2896 p = basic_plus_adr(top()/*!oop*/, osthread, in_bytes(OSThread::interrupted_offset()));
2897 // Set the control input on the field _interrupted read to prevent it floating up.
2898 Node* int_bit = make_load(control(), p, TypeInt::BOOL, T_INT);
2899 Node* cmp_bit = _gvn.transform( new (C, 3) CmpINode(int_bit, intcon(0)) );
2900 Node* bol_bit = _gvn.transform( new (C, 2) BoolNode(cmp_bit, BoolTest::ne) );
2901
2902 IfNode* iff_bit = create_and_map_if(control(), bol_bit, PROB_UNLIKELY_MAG(3), COUNT_UNKNOWN);
2903
2904 // First fast path: if (!TLS._interrupted) return false;
2905 Node* false_bit = _gvn.transform( new (C, 1) IfFalseNode(iff_bit) );
2906 result_rgn->init_req(no_int_result_path, false_bit);
2907 result_val->init_req(no_int_result_path, intcon(0));
2908
2909 // drop through to next case
2910 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_bit)) );
2911
2912 // (c) Or, if interrupt bit is set and clear_int is false, use 2nd fast path.
2913 Node* clr_arg = argument(1);
2914 Node* cmp_arg = _gvn.transform( new (C, 3) CmpINode(clr_arg, intcon(0)) );
2915 Node* bol_arg = _gvn.transform( new (C, 2) BoolNode(cmp_arg, BoolTest::ne) );
2916 IfNode* iff_arg = create_and_map_if(control(), bol_arg, PROB_FAIR, COUNT_UNKNOWN);
2917
2918 // Second fast path: ... else if (!clear_int) return true;
2919 Node* false_arg = _gvn.transform( new (C, 1) IfFalseNode(iff_arg) );
2920 result_rgn->init_req(no_clear_result_path, false_arg);
2921 result_val->init_req(no_clear_result_path, intcon(1));
2922
2923 // drop through to next case
2924 set_control( _gvn.transform(new (C, 1) IfTrueNode(iff_arg)) );
2925
2926 // (d) Otherwise, go to the slow path.
2927 slow_region->add_req(control());
2928 set_control( _gvn.transform(slow_region) );
2929
2930 if (stopped()) {
2931 // There is no slow path.
2932 result_rgn->init_req(slow_result_path, top());
2933 result_val->init_req(slow_result_path, top());
2934 } else {
2935 // non-virtual because it is a private non-static
2936 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_isInterrupted);
2937
2938 Node* slow_val = set_results_for_java_call(slow_call);
2939 // this->control() comes from set_results_for_java_call
2940
2941 // If we know that the result of the slow call will be true, tell the optimizer!
2942 if (known_current_thread) slow_val = intcon(1);
2943
2944 Node* fast_io = slow_call->in(TypeFunc::I_O);
2945 Node* fast_mem = slow_call->in(TypeFunc::Memory);
2946 // These two phis are pre-filled with copies of of the fast IO and Memory
2947 Node* io_phi = PhiNode::make(result_rgn, fast_io, Type::ABIO);
2948 Node* mem_phi = PhiNode::make(result_rgn, fast_mem, Type::MEMORY, TypePtr::BOTTOM);
2949
2950 result_rgn->init_req(slow_result_path, control());
2951 io_phi ->init_req(slow_result_path, i_o());
2952 mem_phi ->init_req(slow_result_path, reset_memory());
2953 result_val->init_req(slow_result_path, slow_val);
2954
2955 set_all_memory( _gvn.transform(mem_phi) );
2956 set_i_o( _gvn.transform(io_phi) );
2957 }
2958
2959 push_result(result_rgn, result_val);
2960 C->set_has_split_ifs(true); // Has chance for split-if optimization
2961
2962 return true;
2963 }
2964
2965 //---------------------------load_mirror_from_klass----------------------------
2966 // Given a klass oop, load its java mirror (a java.lang.Class oop).
2967 Node* LibraryCallKit::load_mirror_from_klass(Node* klass) {
2968 Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
2969 return make_load(NULL, p, TypeInstPtr::MIRROR, T_OBJECT);
2970 }
2971
2972 //-----------------------load_klass_from_mirror_common-------------------------
2973 // Given a java mirror (a java.lang.Class oop), load its corresponding klass oop.
2974 // Test the klass oop for null (signifying a primitive Class like Integer.TYPE),
2975 // and branch to the given path on the region.
2976 // If never_see_null, take an uncommon trap on null, so we can optimistically
2977 // compile for the non-null case.
2978 // If the region is NULL, force never_see_null = true.
2979 Node* LibraryCallKit::load_klass_from_mirror_common(Node* mirror,
2980 bool never_see_null,
2981 int nargs,
2982 RegionNode* region,
2983 int null_path,
2984 int offset) {
2985 if (region == NULL) never_see_null = true;
2986 Node* p = basic_plus_adr(mirror, offset);
2987 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
2988 Node* kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, kls_type) );
2989 _sp += nargs; // any deopt will start just before call to enclosing method
2990 Node* null_ctl = top();
2991 kls = null_check_oop(kls, &null_ctl, never_see_null);
2992 if (region != NULL) {
2993 // Set region->in(null_path) if the mirror is a primitive (e.g, int.class).
2994 region->init_req(null_path, null_ctl);
2995 } else {
2996 assert(null_ctl == top(), "no loose ends");
2997 }
2998 _sp -= nargs;
2999 return kls;
3000 }
3001
3002 //--------------------(inline_native_Class_query helpers)---------------------
3003 // Use this for JVM_ACC_INTERFACE, JVM_ACC_IS_CLONEABLE, JVM_ACC_HAS_FINALIZER.
3004 // Fall through if (mods & mask) == bits, take the guard otherwise.
3005 Node* LibraryCallKit::generate_access_flags_guard(Node* kls, int modifier_mask, int modifier_bits, RegionNode* region) {
3006 // Branch around if the given klass has the given modifier bit set.
3007 // Like generate_guard, adds a new path onto the region.
3008 Node* modp = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3009 Node* mods = make_load(NULL, modp, TypeInt::INT, T_INT);
3010 Node* mask = intcon(modifier_mask);
3011 Node* bits = intcon(modifier_bits);
3012 Node* mbit = _gvn.transform( new (C, 3) AndINode(mods, mask) );
3013 Node* cmp = _gvn.transform( new (C, 3) CmpINode(mbit, bits) );
3014 Node* bol = _gvn.transform( new (C, 2) BoolNode(cmp, BoolTest::ne) );
3015 return generate_fair_guard(bol, region);
3016 }
3017 Node* LibraryCallKit::generate_interface_guard(Node* kls, RegionNode* region) {
3018 return generate_access_flags_guard(kls, JVM_ACC_INTERFACE, 0, region);
3019 }
3020
3021 //-------------------------inline_native_Class_query-------------------
3022 bool LibraryCallKit::inline_native_Class_query(vmIntrinsics::ID id) {
3023 int nargs = 1+0; // just the Class mirror, in most cases
3024 const Type* return_type = TypeInt::BOOL;
3025 Node* prim_return_value = top(); // what happens if it's a primitive class?
3026 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3027 bool expect_prim = false; // most of these guys expect to work on refs
3028
3029 enum { _normal_path = 1, _prim_path = 2, PATH_LIMIT };
3030
3031 switch (id) {
3032 case vmIntrinsics::_isInstance:
3033 nargs = 1+1; // the Class mirror, plus the object getting queried about
3034 // nothing is an instance of a primitive type
3035 prim_return_value = intcon(0);
3036 break;
3037 case vmIntrinsics::_getModifiers:
3038 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3039 assert(is_power_of_2((int)JVM_ACC_WRITTEN_FLAGS+1), "change next line");
3040 return_type = TypeInt::make(0, JVM_ACC_WRITTEN_FLAGS, Type::WidenMin);
3041 break;
3042 case vmIntrinsics::_isInterface:
3043 prim_return_value = intcon(0);
3044 break;
3045 case vmIntrinsics::_isArray:
3046 prim_return_value = intcon(0);
3047 expect_prim = true; // cf. ObjectStreamClass.getClassSignature
3048 break;
3049 case vmIntrinsics::_isPrimitive:
3050 prim_return_value = intcon(1);
3051 expect_prim = true; // obviously
3052 break;
3053 case vmIntrinsics::_getSuperclass:
3054 prim_return_value = null();
3055 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3056 break;
3057 case vmIntrinsics::_getComponentType:
3058 prim_return_value = null();
3059 return_type = TypeInstPtr::MIRROR->cast_to_ptr_type(TypePtr::BotPTR);
3060 break;
3061 case vmIntrinsics::_getClassAccessFlags:
3062 prim_return_value = intcon(JVM_ACC_ABSTRACT | JVM_ACC_FINAL | JVM_ACC_PUBLIC);
3063 return_type = TypeInt::INT; // not bool! 6297094
3064 break;
3065 default:
3066 ShouldNotReachHere();
3067 }
3068
3069 Node* mirror = argument(0);
3070 Node* obj = (nargs <= 1)? top(): argument(1);
3071
3072 const TypeInstPtr* mirror_con = _gvn.type(mirror)->isa_instptr();
3073 if (mirror_con == NULL) return false; // cannot happen?
3074
3075 #ifndef PRODUCT
3076 if (PrintIntrinsics || PrintInlining || PrintOptoInlining) {
3077 ciType* k = mirror_con->java_mirror_type();
3078 if (k) {
3079 tty->print("Inlining %s on constant Class ", vmIntrinsics::name_at(intrinsic_id()));
3080 k->print_name();
3081 tty->cr();
3082 }
3083 }
3084 #endif
3085
3086 // Null-check the mirror, and the mirror's klass ptr (in case it is a primitive).
3087 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3088 record_for_igvn(region);
3089 PhiNode* phi = new (C, PATH_LIMIT) PhiNode(region, return_type);
3090
3091 // The mirror will never be null of Reflection.getClassAccessFlags, however
3092 // it may be null for Class.isInstance or Class.getModifiers. Throw a NPE
3093 // if it is. See bug 4774291.
3094
3095 // For Reflection.getClassAccessFlags(), the null check occurs in
3096 // the wrong place; see inline_unsafe_access(), above, for a similar
3097 // situation.
3098 _sp += nargs; // set original stack for use by uncommon_trap
3099 mirror = do_null_check(mirror, T_OBJECT);
3100 _sp -= nargs;
3101 // If mirror or obj is dead, only null-path is taken.
3102 if (stopped()) return true;
3103
3104 if (expect_prim) never_see_null = false; // expect nulls (meaning prims)
3105
3106 // Now load the mirror's klass metaobject, and null-check it.
3107 // Side-effects region with the control path if the klass is null.
3108 Node* kls = load_klass_from_mirror(mirror, never_see_null, nargs,
3109 region, _prim_path);
3110 // If kls is null, we have a primitive mirror.
3111 phi->init_req(_prim_path, prim_return_value);
3112 if (stopped()) { push_result(region, phi); return true; }
3113
3114 Node* p; // handy temp
3115 Node* null_ctl;
3116
3117 // Now that we have the non-null klass, we can perform the real query.
3118 // For constant classes, the query will constant-fold in LoadNode::Value.
3119 Node* query_value = top();
3120 switch (id) {
3121 case vmIntrinsics::_isInstance:
3122 // nothing is an instance of a primitive type
3123 _sp += nargs; // gen_instanceof might do an uncommon trap
3124 query_value = gen_instanceof(obj, kls);
3125 _sp -= nargs;
3126 break;
3127
3128 case vmIntrinsics::_getModifiers:
3129 p = basic_plus_adr(kls, in_bytes(Klass::modifier_flags_offset()));
3130 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3131 break;
3132
3133 case vmIntrinsics::_isInterface:
3134 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3135 if (generate_interface_guard(kls, region) != NULL)
3136 // A guard was added. If the guard is taken, it was an interface.
3137 phi->add_req(intcon(1));
3138 // If we fall through, it's a plain class.
3139 query_value = intcon(0);
3140 break;
3141
3142 case vmIntrinsics::_isArray:
3143 // (To verify this code sequence, check the asserts in JVM_IsArrayClass.)
3144 if (generate_array_guard(kls, region) != NULL)
3145 // A guard was added. If the guard is taken, it was an array.
3146 phi->add_req(intcon(1));
3147 // If we fall through, it's a plain class.
3148 query_value = intcon(0);
3149 break;
3150
3151 case vmIntrinsics::_isPrimitive:
3152 query_value = intcon(0); // "normal" path produces false
3153 break;
3154
3155 case vmIntrinsics::_getSuperclass:
3156 // The rules here are somewhat unfortunate, but we can still do better
3157 // with random logic than with a JNI call.
3158 // Interfaces store null or Object as _super, but must report null.
3159 // Arrays store an intermediate super as _super, but must report Object.
3160 // Other types can report the actual _super.
3161 // (To verify this code sequence, check the asserts in JVM_IsInterface.)
3162 if (generate_interface_guard(kls, region) != NULL)
3163 // A guard was added. If the guard is taken, it was an interface.
3164 phi->add_req(null());
3165 if (generate_array_guard(kls, region) != NULL)
3166 // A guard was added. If the guard is taken, it was an array.
3167 phi->add_req(makecon(TypeInstPtr::make(env()->Object_klass()->java_mirror())));
3168 // If we fall through, it's a plain class. Get its _super.
3169 p = basic_plus_adr(kls, in_bytes(Klass::super_offset()));
3170 kls = _gvn.transform( LoadKlassNode::make(_gvn, immutable_memory(), p, TypeRawPtr::BOTTOM, TypeKlassPtr::OBJECT_OR_NULL) );
3171 null_ctl = top();
3172 kls = null_check_oop(kls, &null_ctl);
3173 if (null_ctl != top()) {
3174 // If the guard is taken, Object.superClass is null (both klass and mirror).
3175 region->add_req(null_ctl);
3176 phi ->add_req(null());
3177 }
3178 if (!stopped()) {
3179 query_value = load_mirror_from_klass(kls);
3180 }
3181 break;
3182
3183 case vmIntrinsics::_getComponentType:
3184 if (generate_array_guard(kls, region) != NULL) {
3185 // Be sure to pin the oop load to the guard edge just created:
3186 Node* is_array_ctrl = region->in(region->req()-1);
3187 Node* cma = basic_plus_adr(kls, in_bytes(arrayKlass::component_mirror_offset()));
3188 Node* cmo = make_load(is_array_ctrl, cma, TypeInstPtr::MIRROR, T_OBJECT);
3189 phi->add_req(cmo);
3190 }
3191 query_value = null(); // non-array case is null
3192 break;
3193
3194 case vmIntrinsics::_getClassAccessFlags:
3195 p = basic_plus_adr(kls, in_bytes(Klass::access_flags_offset()));
3196 query_value = make_load(NULL, p, TypeInt::INT, T_INT);
3197 break;
3198
3199 default:
3200 ShouldNotReachHere();
3201 }
3202
3203 // Fall-through is the normal case of a query to a real class.
3204 phi->init_req(1, query_value);
3205 region->init_req(1, control());
3206
3207 push_result(region, phi);
3208 C->set_has_split_ifs(true); // Has chance for split-if optimization
3209
3210 return true;
3211 }
3212
3213 //--------------------------inline_native_subtype_check------------------------
3214 // This intrinsic takes the JNI calls out of the heart of
3215 // UnsafeFieldAccessorImpl.set, which improves Field.set, readObject, etc.
3216 bool LibraryCallKit::inline_native_subtype_check() {
3217 int nargs = 1+1; // the Class mirror, plus the other class getting examined
3218
3219 // Pull both arguments off the stack.
3220 Node* args[2]; // two java.lang.Class mirrors: superc, subc
3221 args[0] = argument(0);
3222 args[1] = argument(1);
3223 Node* klasses[2]; // corresponding Klasses: superk, subk
3224 klasses[0] = klasses[1] = top();
3225
3226 enum {
3227 // A full decision tree on {superc is prim, subc is prim}:
3228 _prim_0_path = 1, // {P,N} => false
3229 // {P,P} & superc!=subc => false
3230 _prim_same_path, // {P,P} & superc==subc => true
3231 _prim_1_path, // {N,P} => false
3232 _ref_subtype_path, // {N,N} & subtype check wins => true
3233 _both_ref_path, // {N,N} & subtype check loses => false
3234 PATH_LIMIT
3235 };
3236
3237 RegionNode* region = new (C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3238 Node* phi = new (C, PATH_LIMIT) PhiNode(region, TypeInt::BOOL);
3239 record_for_igvn(region);
3240
3241 const TypePtr* adr_type = TypeRawPtr::BOTTOM; // memory type of loads
3242 const TypeKlassPtr* kls_type = TypeKlassPtr::OBJECT_OR_NULL;
3243 int class_klass_offset = java_lang_Class::klass_offset_in_bytes();
3244
3245 // First null-check both mirrors and load each mirror's klass metaobject.
3246 int which_arg;
3247 for (which_arg = 0; which_arg <= 1; which_arg++) {
3248 Node* arg = args[which_arg];
3249 _sp += nargs; // set original stack for use by uncommon_trap
3250 arg = do_null_check(arg, T_OBJECT);
3251 _sp -= nargs;
3252 if (stopped()) break;
3253 args[which_arg] = _gvn.transform(arg);
3254
3255 Node* p = basic_plus_adr(arg, class_klass_offset);
3256 Node* kls = LoadKlassNode::make(_gvn, immutable_memory(), p, adr_type, kls_type);
3257 klasses[which_arg] = _gvn.transform(kls);
3258 }
3259
3260 // Having loaded both klasses, test each for null.
3261 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3262 for (which_arg = 0; which_arg <= 1; which_arg++) {
3263 Node* kls = klasses[which_arg];
3264 Node* null_ctl = top();
3265 _sp += nargs; // set original stack for use by uncommon_trap
3266 kls = null_check_oop(kls, &null_ctl, never_see_null);
3267 _sp -= nargs;
3268 int prim_path = (which_arg == 0 ? _prim_0_path : _prim_1_path);
3269 region->init_req(prim_path, null_ctl);
3270 if (stopped()) break;
3271 klasses[which_arg] = kls;
3272 }
3273
3274 if (!stopped()) {
3275 // now we have two reference types, in klasses[0..1]
3276 Node* subk = klasses[1]; // the argument to isAssignableFrom
3277 Node* superk = klasses[0]; // the receiver
3278 region->set_req(_both_ref_path, gen_subtype_check(subk, superk));
3279 // now we have a successful reference subtype check
3280 region->set_req(_ref_subtype_path, control());
3281 }
3282
3283 // If both operands are primitive (both klasses null), then
3284 // we must return true when they are identical primitives.
3285 // It is convenient to test this after the first null klass check.
3286 set_control(region->in(_prim_0_path)); // go back to first null check
3287 if (!stopped()) {
3288 // Since superc is primitive, make a guard for the superc==subc case.
3289 Node* cmp_eq = _gvn.transform( new (C, 3) CmpPNode(args[0], args[1]) );
3290 Node* bol_eq = _gvn.transform( new (C, 2) BoolNode(cmp_eq, BoolTest::eq) );
3291 generate_guard(bol_eq, region, PROB_FAIR);
3292 if (region->req() == PATH_LIMIT+1) {
3293 // A guard was added. If the added guard is taken, superc==subc.
3294 region->swap_edges(PATH_LIMIT, _prim_same_path);
3295 region->del_req(PATH_LIMIT);
3296 }
3297 region->set_req(_prim_0_path, control()); // Not equal after all.
3298 }
3299
3300 // these are the only paths that produce 'true':
3301 phi->set_req(_prim_same_path, intcon(1));
3302 phi->set_req(_ref_subtype_path, intcon(1));
3303
3304 // pull together the cases:
3305 assert(region->req() == PATH_LIMIT, "sane region");
3306 for (uint i = 1; i < region->req(); i++) {
3307 Node* ctl = region->in(i);
3308 if (ctl == NULL || ctl == top()) {
3309 region->set_req(i, top());
3310 phi ->set_req(i, top());
3311 } else if (phi->in(i) == NULL) {
3312 phi->set_req(i, intcon(0)); // all other paths produce 'false'
3313 }
3314 }
3315
3316 set_control(_gvn.transform(region));
3317 push(_gvn.transform(phi));
3318
3319 return true;
3320 }
3321
3322 //---------------------generate_array_guard_common------------------------
3323 Node* LibraryCallKit::generate_array_guard_common(Node* kls, RegionNode* region,
3324 bool obj_array, bool not_array) {
3325 // If obj_array/non_array==false/false:
3326 // Branch around if the given klass is in fact an array (either obj or prim).
3327 // If obj_array/non_array==false/true:
3328 // Branch around if the given klass is not an array klass of any kind.
3329 // If obj_array/non_array==true/true:
3330 // Branch around if the kls is not an oop array (kls is int[], String, etc.)
3331 // If obj_array/non_array==true/false:
3332 // Branch around if the kls is an oop array (Object[] or subtype)
3333 //
3334 // Like generate_guard, adds a new path onto the region.
3335 jint layout_con = 0;
3336 Node* layout_val = get_layout_helper(kls, layout_con);
3337 if (layout_val == NULL) {
3338 bool query = (obj_array
3339 ? Klass::layout_helper_is_objArray(layout_con)
3340 : Klass::layout_helper_is_javaArray(layout_con));
3341 if (query == not_array) {
3342 return NULL; // never a branch
3343 } else { // always a branch
3344 Node* always_branch = control();
3345 if (region != NULL)
3346 region->add_req(always_branch);
3347 set_control(top());
3348 return always_branch;
3349 }
3350 }
3351 // Now test the correct condition.
3352 jint nval = (obj_array
3353 ? ((jint)Klass::_lh_array_tag_type_value
3354 << Klass::_lh_array_tag_shift)
3355 : Klass::_lh_neutral_value);
3356 Node* cmp = _gvn.transform( new(C, 3) CmpINode(layout_val, intcon(nval)) );
3357 BoolTest::mask btest = BoolTest::lt; // correct for testing is_[obj]array
3358 // invert the test if we are looking for a non-array
3359 if (not_array) btest = BoolTest(btest).negate();
3360 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, btest) );
3361 return generate_fair_guard(bol, region);
3362 }
3363
3364
3365 //-----------------------inline_native_newArray--------------------------
3366 bool LibraryCallKit::inline_native_newArray() {
3367 int nargs = 2;
3368 Node* mirror = argument(0);
3369 Node* count_val = argument(1);
3370
3371 _sp += nargs; // set original stack for use by uncommon_trap
3372 mirror = do_null_check(mirror, T_OBJECT);
3373 _sp -= nargs;
3374 // If mirror or obj is dead, only null-path is taken.
3375 if (stopped()) return true;
3376
3377 enum { _normal_path = 1, _slow_path = 2, PATH_LIMIT };
3378 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3379 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3380 TypeInstPtr::NOTNULL);
3381 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3382 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3383 TypePtr::BOTTOM);
3384
3385 bool never_see_null = !too_many_traps(Deoptimization::Reason_null_check);
3386 Node* klass_node = load_array_klass_from_mirror(mirror, never_see_null,
3387 nargs,
3388 result_reg, _slow_path);
3389 Node* normal_ctl = control();
3390 Node* no_array_ctl = result_reg->in(_slow_path);
3391
3392 // Generate code for the slow case. We make a call to newArray().
3393 set_control(no_array_ctl);
3394 if (!stopped()) {
3395 // Either the input type is void.class, or else the
3396 // array klass has not yet been cached. Either the
3397 // ensuing call will throw an exception, or else it
3398 // will cache the array klass for next time.
3399 PreserveJVMState pjvms(this);
3400 CallJavaNode* slow_call = generate_method_call_static(vmIntrinsics::_newArray);
3401 Node* slow_result = set_results_for_java_call(slow_call);
3402 // this->control() comes from set_results_for_java_call
3403 result_reg->set_req(_slow_path, control());
3404 result_val->set_req(_slow_path, slow_result);
3405 result_io ->set_req(_slow_path, i_o());
3406 result_mem->set_req(_slow_path, reset_memory());
3407 }
3408
3409 set_control(normal_ctl);
3410 if (!stopped()) {
3411 // Normal case: The array type has been cached in the java.lang.Class.
3412 // The following call works fine even if the array type is polymorphic.
3413 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3414 Node* obj = new_array(klass_node, count_val, nargs);
3415 result_reg->init_req(_normal_path, control());
3416 result_val->init_req(_normal_path, obj);
3417 result_io ->init_req(_normal_path, i_o());
3418 result_mem->init_req(_normal_path, reset_memory());
3419 }
3420
3421 // Return the combined state.
3422 set_i_o( _gvn.transform(result_io) );
3423 set_all_memory( _gvn.transform(result_mem) );
3424 push_result(result_reg, result_val);
3425 C->set_has_split_ifs(true); // Has chance for split-if optimization
3426
3427 return true;
3428 }
3429
3430 //----------------------inline_native_getLength--------------------------
3431 bool LibraryCallKit::inline_native_getLength() {
3432 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3433
3434 int nargs = 1;
3435 Node* array = argument(0);
3436
3437 _sp += nargs; // set original stack for use by uncommon_trap
3438 array = do_null_check(array, T_OBJECT);
3439 _sp -= nargs;
3440
3441 // If array is dead, only null-path is taken.
3442 if (stopped()) return true;
3443
3444 // Deoptimize if it is a non-array.
3445 Node* non_array = generate_non_array_guard(load_object_klass(array), NULL);
3446
3447 if (non_array != NULL) {
3448 PreserveJVMState pjvms(this);
3449 set_control(non_array);
3450 _sp += nargs; // push the arguments back on the stack
3451 uncommon_trap(Deoptimization::Reason_intrinsic,
3452 Deoptimization::Action_maybe_recompile);
3453 }
3454
3455 // If control is dead, only non-array-path is taken.
3456 if (stopped()) return true;
3457
3458 // The works fine even if the array type is polymorphic.
3459 // It could be a dynamic mix of int[], boolean[], Object[], etc.
3460 push( load_array_length(array) );
3461
3462 C->set_has_split_ifs(true); // Has chance for split-if optimization
3463
3464 return true;
3465 }
3466
3467 //------------------------inline_array_copyOf----------------------------
3468 bool LibraryCallKit::inline_array_copyOf(bool is_copyOfRange) {
3469 if (too_many_traps(Deoptimization::Reason_intrinsic)) return false;
3470
3471 // Restore the stack and pop off the arguments.
3472 int nargs = 3 + (is_copyOfRange? 1: 0);
3473 Node* original = argument(0);
3474 Node* start = is_copyOfRange? argument(1): intcon(0);
3475 Node* end = is_copyOfRange? argument(2): argument(1);
3476 Node* array_type_mirror = is_copyOfRange? argument(3): argument(2);
3477
3478 Node* newcopy;
3479
3480 //set the original stack and the reexecute bit for the interpreter to reexecute
3481 //the bytecode that invokes Arrays.copyOf if deoptimization happens
3482 { PreserveReexecuteState preexecs(this);
3483 _sp += nargs;
3484 jvms()->set_should_reexecute(true);
3485
3486 array_type_mirror = do_null_check(array_type_mirror, T_OBJECT);
3487 original = do_null_check(original, T_OBJECT);
3488
3489 // Check if a null path was taken unconditionally.
3490 if (stopped()) return true;
3491
3492 Node* orig_length = load_array_length(original);
3493
3494 Node* klass_node = load_klass_from_mirror(array_type_mirror, false, 0,
3495 NULL, 0);
3496 klass_node = do_null_check(klass_node, T_OBJECT);
3497
3498 RegionNode* bailout = new (C, 1) RegionNode(1);
3499 record_for_igvn(bailout);
3500
3501 // Despite the generic type of Arrays.copyOf, the mirror might be int, int[], etc.
3502 // Bail out if that is so.
3503 Node* not_objArray = generate_non_objArray_guard(klass_node, bailout);
3504 if (not_objArray != NULL) {
3505 // Improve the klass node's type from the new optimistic assumption:
3506 ciKlass* ak = ciArrayKlass::make(env()->Object_klass());
3507 const Type* akls = TypeKlassPtr::make(TypePtr::NotNull, ak, 0/*offset*/);
3508 Node* cast = new (C, 2) CastPPNode(klass_node, akls);
3509 cast->init_req(0, control());
3510 klass_node = _gvn.transform(cast);
3511 }
3512
3513 // Bail out if either start or end is negative.
3514 generate_negative_guard(start, bailout, &start);
3515 generate_negative_guard(end, bailout, &end);
3516
3517 Node* length = end;
3518 if (_gvn.type(start) != TypeInt::ZERO) {
3519 length = _gvn.transform( new (C, 3) SubINode(end, start) );
3520 }
3521
3522 // Bail out if length is negative.
3523 // ...Not needed, since the new_array will throw the right exception.
3524 //generate_negative_guard(length, bailout, &length);
3525
3526 if (bailout->req() > 1) {
3527 PreserveJVMState pjvms(this);
3528 set_control( _gvn.transform(bailout) );
3529 uncommon_trap(Deoptimization::Reason_intrinsic,
3530 Deoptimization::Action_maybe_recompile);
3531 }
3532
3533 if (!stopped()) {
3534
3535 // How many elements will we copy from the original?
3536 // The answer is MinI(orig_length - start, length).
3537 Node* orig_tail = _gvn.transform( new(C, 3) SubINode(orig_length, start) );
3538 Node* moved = generate_min_max(vmIntrinsics::_min, orig_tail, length);
3539
3540 newcopy = new_array(klass_node, length, 0);
3541
3542 // Generate a direct call to the right arraycopy function(s).
3543 // We know the copy is disjoint but we might not know if the
3544 // oop stores need checking.
3545 // Extreme case: Arrays.copyOf((Integer[])x, 10, String[].class).
3546 // This will fail a store-check if x contains any non-nulls.
3547 bool disjoint_bases = true;
3548 bool length_never_negative = true;
3549 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
3550 original, start, newcopy, intcon(0), moved,
3551 disjoint_bases, length_never_negative);
3552 }
3553 } //original reexecute and sp are set back here
3554
3555 if(!stopped()) {
3556 push(newcopy);
3557 }
3558
3559 C->set_has_split_ifs(true); // Has chance for split-if optimization
3560
3561 return true;
3562 }
3563
3564
3565 //----------------------generate_virtual_guard---------------------------
3566 // Helper for hashCode and clone. Peeks inside the vtable to avoid a call.
3567 Node* LibraryCallKit::generate_virtual_guard(Node* obj_klass,
3568 RegionNode* slow_region) {
3569 ciMethod* method = callee();
3570 int vtable_index = method->vtable_index();
3571 // Get the methodOop out of the appropriate vtable entry.
3572 int entry_offset = (instanceKlass::vtable_start_offset() +
3573 vtable_index*vtableEntry::size()) * wordSize +
3574 vtableEntry::method_offset_in_bytes();
3575 Node* entry_addr = basic_plus_adr(obj_klass, entry_offset);
3576 Node* target_call = make_load(NULL, entry_addr, TypeInstPtr::NOTNULL, T_OBJECT);
3577
3578 // Compare the target method with the expected method (e.g., Object.hashCode).
3579 const TypeInstPtr* native_call_addr = TypeInstPtr::make(method);
3580
3581 Node* native_call = makecon(native_call_addr);
3582 Node* chk_native = _gvn.transform( new(C, 3) CmpPNode(target_call, native_call) );
3583 Node* test_native = _gvn.transform( new(C, 2) BoolNode(chk_native, BoolTest::ne) );
3584
3585 return generate_slow_guard(test_native, slow_region);
3586 }
3587
3588 //-----------------------generate_method_call----------------------------
3589 // Use generate_method_call to make a slow-call to the real
3590 // method if the fast path fails. An alternative would be to
3591 // use a stub like OptoRuntime::slow_arraycopy_Java.
3592 // This only works for expanding the current library call,
3593 // not another intrinsic. (E.g., don't use this for making an
3594 // arraycopy call inside of the copyOf intrinsic.)
3595 CallJavaNode*
3596 LibraryCallKit::generate_method_call(vmIntrinsics::ID method_id, bool is_virtual, bool is_static) {
3597 // When compiling the intrinsic method itself, do not use this technique.
3598 guarantee(callee() != C->method(), "cannot make slow-call to self");
3599
3600 ciMethod* method = callee();
3601 // ensure the JVMS we have will be correct for this call
3602 guarantee(method_id == method->intrinsic_id(), "must match");
3603
3604 const TypeFunc* tf = TypeFunc::make(method);
3605 int tfdc = tf->domain()->cnt();
3606 CallJavaNode* slow_call;
3607 if (is_static) {
3608 assert(!is_virtual, "");
3609 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3610 SharedRuntime::get_resolve_static_call_stub(),
3611 method, bci());
3612 } else if (is_virtual) {
3613 null_check_receiver(method);
3614 int vtable_index = methodOopDesc::invalid_vtable_index;
3615 if (UseInlineCaches) {
3616 // Suppress the vtable call
3617 } else {
3618 // hashCode and clone are not a miranda methods,
3619 // so the vtable index is fixed.
3620 // No need to use the linkResolver to get it.
3621 vtable_index = method->vtable_index();
3622 }
3623 slow_call = new(C, tfdc) CallDynamicJavaNode(tf,
3624 SharedRuntime::get_resolve_virtual_call_stub(),
3625 method, vtable_index, bci());
3626 } else { // neither virtual nor static: opt_virtual
3627 null_check_receiver(method);
3628 slow_call = new(C, tfdc) CallStaticJavaNode(tf,
3629 SharedRuntime::get_resolve_opt_virtual_call_stub(),
3630 method, bci());
3631 slow_call->set_optimized_virtual(true);
3632 }
3633 set_arguments_for_java_call(slow_call);
3634 set_edges_for_java_call(slow_call);
3635 return slow_call;
3636 }
3637
3638
3639 //------------------------------inline_native_hashcode--------------------
3640 // Build special case code for calls to hashCode on an object.
3641 bool LibraryCallKit::inline_native_hashcode(bool is_virtual, bool is_static) {
3642 assert(is_static == callee()->is_static(), "correct intrinsic selection");
3643 assert(!(is_virtual && is_static), "either virtual, special, or static");
3644
3645 enum { _slow_path = 1, _fast_path, _null_path, PATH_LIMIT };
3646
3647 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
3648 PhiNode* result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
3649 TypeInt::INT);
3650 PhiNode* result_io = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
3651 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
3652 TypePtr::BOTTOM);
3653 Node* obj = NULL;
3654 if (!is_static) {
3655 // Check for hashing null object
3656 obj = null_check_receiver(callee());
3657 if (stopped()) return true; // unconditionally null
3658 result_reg->init_req(_null_path, top());
3659 result_val->init_req(_null_path, top());
3660 } else {
3661 // Do a null check, and return zero if null.
3662 // System.identityHashCode(null) == 0
3663 obj = argument(0);
3664 Node* null_ctl = top();
3665 obj = null_check_oop(obj, &null_ctl);
3666 result_reg->init_req(_null_path, null_ctl);
3667 result_val->init_req(_null_path, _gvn.intcon(0));
3668 }
3669
3670 // Unconditionally null? Then return right away.
3671 if (stopped()) {
3672 set_control( result_reg->in(_null_path) );
3673 if (!stopped())
3674 push( result_val ->in(_null_path) );
3675 return true;
3676 }
3677
3678 // After null check, get the object's klass.
3679 Node* obj_klass = load_object_klass(obj);
3680
3681 // This call may be virtual (invokevirtual) or bound (invokespecial).
3682 // For each case we generate slightly different code.
3683
3684 // We only go to the fast case code if we pass a number of guards. The
3685 // paths which do not pass are accumulated in the slow_region.
3686 RegionNode* slow_region = new (C, 1) RegionNode(1);
3687 record_for_igvn(slow_region);
3688
3689 // If this is a virtual call, we generate a funny guard. We pull out
3690 // the vtable entry corresponding to hashCode() from the target object.
3691 // If the target method which we are calling happens to be the native
3692 // Object hashCode() method, we pass the guard. We do not need this
3693 // guard for non-virtual calls -- the caller is known to be the native
3694 // Object hashCode().
3695 if (is_virtual) {
3696 generate_virtual_guard(obj_klass, slow_region);
3697 }
3698
3699 // Get the header out of the object, use LoadMarkNode when available
3700 Node* header_addr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3701 Node* header = make_load(control(), header_addr, TypeX_X, TypeX_X->basic_type());
3702
3703 // Test the header to see if it is unlocked.
3704 Node *lock_mask = _gvn.MakeConX(markOopDesc::biased_lock_mask_in_place);
3705 Node *lmasked_header = _gvn.transform( new (C, 3) AndXNode(header, lock_mask) );
3706 Node *unlocked_val = _gvn.MakeConX(markOopDesc::unlocked_value);
3707 Node *chk_unlocked = _gvn.transform( new (C, 3) CmpXNode( lmasked_header, unlocked_val));
3708 Node *test_unlocked = _gvn.transform( new (C, 2) BoolNode( chk_unlocked, BoolTest::ne) );
3709
3710 generate_slow_guard(test_unlocked, slow_region);
3711
3712 // Get the hash value and check to see that it has been properly assigned.
3713 // We depend on hash_mask being at most 32 bits and avoid the use of
3714 // hash_mask_in_place because it could be larger than 32 bits in a 64-bit
3715 // vm: see markOop.hpp.
3716 Node *hash_mask = _gvn.intcon(markOopDesc::hash_mask);
3717 Node *hash_shift = _gvn.intcon(markOopDesc::hash_shift);
3718 Node *hshifted_header= _gvn.transform( new (C, 3) URShiftXNode(header, hash_shift) );
3719 // This hack lets the hash bits live anywhere in the mark object now, as long
3720 // as the shift drops the relevant bits into the low 32 bits. Note that
3721 // Java spec says that HashCode is an int so there's no point in capturing
3722 // an 'X'-sized hashcode (32 in 32-bit build or 64 in 64-bit build).
3723 hshifted_header = ConvX2I(hshifted_header);
3724 Node *hash_val = _gvn.transform( new (C, 3) AndINode(hshifted_header, hash_mask) );
3725
3726 Node *no_hash_val = _gvn.intcon(markOopDesc::no_hash);
3727 Node *chk_assigned = _gvn.transform( new (C, 3) CmpINode( hash_val, no_hash_val));
3728 Node *test_assigned = _gvn.transform( new (C, 2) BoolNode( chk_assigned, BoolTest::eq) );
3729
3730 generate_slow_guard(test_assigned, slow_region);
3731
3732 Node* init_mem = reset_memory();
3733 // fill in the rest of the null path:
3734 result_io ->init_req(_null_path, i_o());
3735 result_mem->init_req(_null_path, init_mem);
3736
3737 result_val->init_req(_fast_path, hash_val);
3738 result_reg->init_req(_fast_path, control());
3739 result_io ->init_req(_fast_path, i_o());
3740 result_mem->init_req(_fast_path, init_mem);
3741
3742 // Generate code for the slow case. We make a call to hashCode().
3743 set_control(_gvn.transform(slow_region));
3744 if (!stopped()) {
3745 // No need for PreserveJVMState, because we're using up the present state.
3746 set_all_memory(init_mem);
3747 vmIntrinsics::ID hashCode_id = vmIntrinsics::_hashCode;
3748 if (is_static) hashCode_id = vmIntrinsics::_identityHashCode;
3749 CallJavaNode* slow_call = generate_method_call(hashCode_id, is_virtual, is_static);
3750 Node* slow_result = set_results_for_java_call(slow_call);
3751 // this->control() comes from set_results_for_java_call
3752 result_reg->init_req(_slow_path, control());
3753 result_val->init_req(_slow_path, slow_result);
3754 result_io ->set_req(_slow_path, i_o());
3755 result_mem ->set_req(_slow_path, reset_memory());
3756 }
3757
3758 // Return the combined state.
3759 set_i_o( _gvn.transform(result_io) );
3760 set_all_memory( _gvn.transform(result_mem) );
3761 push_result(result_reg, result_val);
3762
3763 return true;
3764 }
3765
3766 //---------------------------inline_native_getClass----------------------------
3767 // Build special case code for calls to getClass on an object.
3768 bool LibraryCallKit::inline_native_getClass() {
3769 Node* obj = null_check_receiver(callee());
3770 if (stopped()) return true;
3771 push( load_mirror_from_klass(load_object_klass(obj)) );
3772 return true;
3773 }
3774
3775 //-----------------inline_native_Reflection_getCallerClass---------------------
3776 // In the presence of deep enough inlining, getCallerClass() becomes a no-op.
3777 //
3778 // NOTE that this code must perform the same logic as
3779 // vframeStream::security_get_caller_frame in that it must skip
3780 // Method.invoke() and auxiliary frames.
3781
3782
3783
3784
3785 bool LibraryCallKit::inline_native_Reflection_getCallerClass() {
3786 ciMethod* method = callee();
3787
3788 #ifndef PRODUCT
3789 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3790 tty->print_cr("Attempting to inline sun.reflect.Reflection.getCallerClass");
3791 }
3792 #endif
3793
3794 debug_only(int saved_sp = _sp);
3795
3796 // Argument words: (int depth)
3797 int nargs = 1;
3798
3799 _sp += nargs;
3800 Node* caller_depth_node = pop();
3801
3802 assert(saved_sp == _sp, "must have correct argument count");
3803
3804 // The depth value must be a constant in order for the runtime call
3805 // to be eliminated.
3806 const TypeInt* caller_depth_type = _gvn.type(caller_depth_node)->isa_int();
3807 if (caller_depth_type == NULL || !caller_depth_type->is_con()) {
3808 #ifndef PRODUCT
3809 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3810 tty->print_cr(" Bailing out because caller depth was not a constant");
3811 }
3812 #endif
3813 return false;
3814 }
3815 // Note that the JVM state at this point does not include the
3816 // getCallerClass() frame which we are trying to inline. The
3817 // semantics of getCallerClass(), however, are that the "first"
3818 // frame is the getCallerClass() frame, so we subtract one from the
3819 // requested depth before continuing. We don't inline requests of
3820 // getCallerClass(0).
3821 int caller_depth = caller_depth_type->get_con() - 1;
3822 if (caller_depth < 0) {
3823 #ifndef PRODUCT
3824 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3825 tty->print_cr(" Bailing out because caller depth was %d", caller_depth);
3826 }
3827 #endif
3828 return false;
3829 }
3830
3831 if (!jvms()->has_method()) {
3832 #ifndef PRODUCT
3833 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3834 tty->print_cr(" Bailing out because intrinsic was inlined at top level");
3835 }
3836 #endif
3837 return false;
3838 }
3839 int _depth = jvms()->depth(); // cache call chain depth
3840
3841 // Walk back up the JVM state to find the caller at the required
3842 // depth. NOTE that this code must perform the same logic as
3843 // vframeStream::security_get_caller_frame in that it must skip
3844 // Method.invoke() and auxiliary frames. Note also that depth is
3845 // 1-based (1 is the bottom of the inlining).
3846 int inlining_depth = _depth;
3847 JVMState* caller_jvms = NULL;
3848
3849 if (inlining_depth > 0) {
3850 caller_jvms = jvms();
3851 assert(caller_jvms = jvms()->of_depth(inlining_depth), "inlining_depth == our depth");
3852 do {
3853 // The following if-tests should be performed in this order
3854 if (is_method_invoke_or_aux_frame(caller_jvms)) {
3855 // Skip a Method.invoke() or auxiliary frame
3856 } else if (caller_depth > 0) {
3857 // Skip real frame
3858 --caller_depth;
3859 } else {
3860 // We're done: reached desired caller after skipping.
3861 break;
3862 }
3863 caller_jvms = caller_jvms->caller();
3864 --inlining_depth;
3865 } while (inlining_depth > 0);
3866 }
3867
3868 if (inlining_depth == 0) {
3869 #ifndef PRODUCT
3870 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3871 tty->print_cr(" Bailing out because caller depth (%d) exceeded inlining depth (%d)", caller_depth_type->get_con(), _depth);
3872 tty->print_cr(" JVM state at this point:");
3873 for (int i = _depth; i >= 1; i--) {
3874 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3875 }
3876 }
3877 #endif
3878 return false; // Reached end of inlining
3879 }
3880
3881 // Acquire method holder as java.lang.Class
3882 ciInstanceKlass* caller_klass = caller_jvms->method()->holder();
3883 ciInstance* caller_mirror = caller_klass->java_mirror();
3884 // Push this as a constant
3885 push(makecon(TypeInstPtr::make(caller_mirror)));
3886 #ifndef PRODUCT
3887 if ((PrintIntrinsics || PrintInlining || PrintOptoInlining) && Verbose) {
3888 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);
3889 tty->print_cr(" JVM state at this point:");
3890 for (int i = _depth; i >= 1; i--) {
3891 tty->print_cr(" %d) %s", i, jvms()->of_depth(i)->method()->name()->as_utf8());
3892 }
3893 }
3894 #endif
3895 return true;
3896 }
3897
3898 // Helper routine for above
3899 bool LibraryCallKit::is_method_invoke_or_aux_frame(JVMState* jvms) {
3900 ciMethod* method = jvms->method();
3901
3902 // Is this the Method.invoke method itself?
3903 if (method->intrinsic_id() == vmIntrinsics::_invoke)
3904 return true;
3905
3906 // Is this a helper, defined somewhere underneath MethodAccessorImpl.
3907 ciKlass* k = method->holder();
3908 if (k->is_instance_klass()) {
3909 ciInstanceKlass* ik = k->as_instance_klass();
3910 for (; ik != NULL; ik = ik->super()) {
3911 if (ik->name() == ciSymbol::sun_reflect_MethodAccessorImpl() &&
3912 ik == env()->find_system_klass(ik->name())) {
3913 return true;
3914 }
3915 }
3916 }
3917 else if (method->is_method_handle_adapter()) {
3918 // This is an internal adapter frame from the MethodHandleCompiler -- skip it
3919 return true;
3920 }
3921
3922 return false;
3923 }
3924
3925 static int value_field_offset = -1; // offset of the "value" field of AtomicLongCSImpl. This is needed by
3926 // inline_native_AtomicLong_attemptUpdate() but it has no way of
3927 // computing it since there is no lookup field by name function in the
3928 // CI interface. This is computed and set by inline_native_AtomicLong_get().
3929 // Using a static variable here is safe even if we have multiple compilation
3930 // threads because the offset is constant. At worst the same offset will be
3931 // computed and stored multiple
3932
3933 bool LibraryCallKit::inline_native_AtomicLong_get() {
3934 // Restore the stack and pop off the argument
3935 _sp+=1;
3936 Node *obj = pop();
3937
3938 // get the offset of the "value" field. Since the CI interfaces
3939 // does not provide a way to look up a field by name, we scan the bytecodes
3940 // to get the field index. We expect the first 2 instructions of the method
3941 // to be:
3942 // 0 aload_0
3943 // 1 getfield "value"
3944 ciMethod* method = callee();
3945 if (value_field_offset == -1)
3946 {
3947 ciField* value_field;
3948 ciBytecodeStream iter(method);
3949 Bytecodes::Code bc = iter.next();
3950
3951 if ((bc != Bytecodes::_aload_0) &&
3952 ((bc != Bytecodes::_aload) || (iter.get_index() != 0)))
3953 return false;
3954 bc = iter.next();
3955 if (bc != Bytecodes::_getfield)
3956 return false;
3957 bool ignore;
3958 value_field = iter.get_field(ignore);
3959 value_field_offset = value_field->offset_in_bytes();
3960 }
3961
3962 // Null check without removing any arguments.
3963 _sp++;
3964 obj = do_null_check(obj, T_OBJECT);
3965 _sp--;
3966 // Check for locking null object
3967 if (stopped()) return true;
3968
3969 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
3970 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
3971 int alias_idx = C->get_alias_index(adr_type);
3972
3973 Node *result = _gvn.transform(new (C, 3) LoadLLockedNode(control(), memory(alias_idx), adr));
3974
3975 push_pair(result);
3976
3977 return true;
3978 }
3979
3980 bool LibraryCallKit::inline_native_AtomicLong_attemptUpdate() {
3981 // Restore the stack and pop off the arguments
3982 _sp+=5;
3983 Node *newVal = pop_pair();
3984 Node *oldVal = pop_pair();
3985 Node *obj = pop();
3986
3987 // we need the offset of the "value" field which was computed when
3988 // inlining the get() method. Give up if we don't have it.
3989 if (value_field_offset == -1)
3990 return false;
3991
3992 // Null check without removing any arguments.
3993 _sp+=5;
3994 obj = do_null_check(obj, T_OBJECT);
3995 _sp-=5;
3996 // Check for locking null object
3997 if (stopped()) return true;
3998
3999 Node *adr = basic_plus_adr(obj, obj, value_field_offset);
4000 const TypePtr *adr_type = _gvn.type(adr)->is_ptr();
4001 int alias_idx = C->get_alias_index(adr_type);
4002
4003 Node *cas = _gvn.transform(new (C, 5) StoreLConditionalNode(control(), memory(alias_idx), adr, newVal, oldVal));
4004 Node *store_proj = _gvn.transform( new (C, 1) SCMemProjNode(cas));
4005 set_memory(store_proj, alias_idx);
4006 Node *bol = _gvn.transform( new (C, 2) BoolNode( cas, BoolTest::eq ) );
4007
4008 Node *result;
4009 // CMove node is not used to be able fold a possible check code
4010 // after attemptUpdate() call. This code could be transformed
4011 // into CMove node by loop optimizations.
4012 {
4013 RegionNode *r = new (C, 3) RegionNode(3);
4014 result = new (C, 3) PhiNode(r, TypeInt::BOOL);
4015
4016 Node *iff = create_and_xform_if(control(), bol, PROB_FAIR, COUNT_UNKNOWN);
4017 Node *iftrue = opt_iff(r, iff);
4018 r->init_req(1, iftrue);
4019 result->init_req(1, intcon(1));
4020 result->init_req(2, intcon(0));
4021
4022 set_control(_gvn.transform(r));
4023 record_for_igvn(r);
4024
4025 C->set_has_split_ifs(true); // Has chance for split-if optimization
4026 }
4027
4028 push(_gvn.transform(result));
4029 return true;
4030 }
4031
4032 bool LibraryCallKit::inline_fp_conversions(vmIntrinsics::ID id) {
4033 // restore the arguments
4034 _sp += arg_size();
4035
4036 switch (id) {
4037 case vmIntrinsics::_floatToRawIntBits:
4038 push(_gvn.transform( new (C, 2) MoveF2INode(pop())));
4039 break;
4040
4041 case vmIntrinsics::_intBitsToFloat:
4042 push(_gvn.transform( new (C, 2) MoveI2FNode(pop())));
4043 break;
4044
4045 case vmIntrinsics::_doubleToRawLongBits:
4046 push_pair(_gvn.transform( new (C, 2) MoveD2LNode(pop_pair())));
4047 break;
4048
4049 case vmIntrinsics::_longBitsToDouble:
4050 push_pair(_gvn.transform( new (C, 2) MoveL2DNode(pop_pair())));
4051 break;
4052
4053 case vmIntrinsics::_doubleToLongBits: {
4054 Node* value = pop_pair();
4055
4056 // two paths (plus control) merge in a wood
4057 RegionNode *r = new (C, 3) RegionNode(3);
4058 Node *phi = new (C, 3) PhiNode(r, TypeLong::LONG);
4059
4060 Node *cmpisnan = _gvn.transform( new (C, 3) CmpDNode(value, value));
4061 // Build the boolean node
4062 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4063
4064 // Branch either way.
4065 // NaN case is less traveled, which makes all the difference.
4066 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4067 Node *opt_isnan = _gvn.transform(ifisnan);
4068 assert( opt_isnan->is_If(), "Expect an IfNode");
4069 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4070 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4071
4072 set_control(iftrue);
4073
4074 static const jlong nan_bits = CONST64(0x7ff8000000000000);
4075 Node *slow_result = longcon(nan_bits); // return NaN
4076 phi->init_req(1, _gvn.transform( slow_result ));
4077 r->init_req(1, iftrue);
4078
4079 // Else fall through
4080 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4081 set_control(iffalse);
4082
4083 phi->init_req(2, _gvn.transform( new (C, 2) MoveD2LNode(value)));
4084 r->init_req(2, iffalse);
4085
4086 // Post merge
4087 set_control(_gvn.transform(r));
4088 record_for_igvn(r);
4089
4090 Node* result = _gvn.transform(phi);
4091 assert(result->bottom_type()->isa_long(), "must be");
4092 push_pair(result);
4093
4094 C->set_has_split_ifs(true); // Has chance for split-if optimization
4095
4096 break;
4097 }
4098
4099 case vmIntrinsics::_floatToIntBits: {
4100 Node* value = pop();
4101
4102 // two paths (plus control) merge in a wood
4103 RegionNode *r = new (C, 3) RegionNode(3);
4104 Node *phi = new (C, 3) PhiNode(r, TypeInt::INT);
4105
4106 Node *cmpisnan = _gvn.transform( new (C, 3) CmpFNode(value, value));
4107 // Build the boolean node
4108 Node *bolisnan = _gvn.transform( new (C, 2) BoolNode( cmpisnan, BoolTest::ne ) );
4109
4110 // Branch either way.
4111 // NaN case is less traveled, which makes all the difference.
4112 IfNode *ifisnan = create_and_xform_if(control(), bolisnan, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
4113 Node *opt_isnan = _gvn.transform(ifisnan);
4114 assert( opt_isnan->is_If(), "Expect an IfNode");
4115 IfNode *opt_ifisnan = (IfNode*)opt_isnan;
4116 Node *iftrue = _gvn.transform( new (C, 1) IfTrueNode(opt_ifisnan) );
4117
4118 set_control(iftrue);
4119
4120 static const jint nan_bits = 0x7fc00000;
4121 Node *slow_result = makecon(TypeInt::make(nan_bits)); // return NaN
4122 phi->init_req(1, _gvn.transform( slow_result ));
4123 r->init_req(1, iftrue);
4124
4125 // Else fall through
4126 Node *iffalse = _gvn.transform( new (C, 1) IfFalseNode(opt_ifisnan) );
4127 set_control(iffalse);
4128
4129 phi->init_req(2, _gvn.transform( new (C, 2) MoveF2INode(value)));
4130 r->init_req(2, iffalse);
4131
4132 // Post merge
4133 set_control(_gvn.transform(r));
4134 record_for_igvn(r);
4135
4136 Node* result = _gvn.transform(phi);
4137 assert(result->bottom_type()->isa_int(), "must be");
4138 push(result);
4139
4140 C->set_has_split_ifs(true); // Has chance for split-if optimization
4141
4142 break;
4143 }
4144
4145 default:
4146 ShouldNotReachHere();
4147 }
4148
4149 return true;
4150 }
4151
4152 #ifdef _LP64
4153 #define XTOP ,top() /*additional argument*/
4154 #else //_LP64
4155 #define XTOP /*no additional argument*/
4156 #endif //_LP64
4157
4158 //----------------------inline_unsafe_copyMemory-------------------------
4159 bool LibraryCallKit::inline_unsafe_copyMemory() {
4160 if (callee()->is_static()) return false; // caller must have the capability!
4161 int nargs = 1 + 5 + 3; // 5 args: (src: ptr,off, dst: ptr,off, size)
4162 assert(signature()->size() == nargs-1, "copy has 5 arguments");
4163 null_check_receiver(callee()); // check then ignore argument(0)
4164 if (stopped()) return true;
4165
4166 C->set_has_unsafe_access(true); // Mark eventual nmethod as "unsafe".
4167
4168 Node* src_ptr = argument(1);
4169 Node* src_off = ConvL2X(argument(2));
4170 assert(argument(3)->is_top(), "2nd half of long");
4171 Node* dst_ptr = argument(4);
4172 Node* dst_off = ConvL2X(argument(5));
4173 assert(argument(6)->is_top(), "2nd half of long");
4174 Node* size = ConvL2X(argument(7));
4175 assert(argument(8)->is_top(), "2nd half of long");
4176
4177 assert(Unsafe_field_offset_to_byte_offset(11) == 11,
4178 "fieldOffset must be byte-scaled");
4179
4180 Node* src = make_unsafe_address(src_ptr, src_off);
4181 Node* dst = make_unsafe_address(dst_ptr, dst_off);
4182
4183 // Conservatively insert a memory barrier on all memory slices.
4184 // Do not let writes of the copy source or destination float below the copy.
4185 insert_mem_bar(Op_MemBarCPUOrder);
4186
4187 // Call it. Note that the length argument is not scaled.
4188 make_runtime_call(RC_LEAF|RC_NO_FP,
4189 OptoRuntime::fast_arraycopy_Type(),
4190 StubRoutines::unsafe_arraycopy(),
4191 "unsafe_arraycopy",
4192 TypeRawPtr::BOTTOM,
4193 src, dst, size XTOP);
4194
4195 // Do not let reads of the copy destination float above the copy.
4196 insert_mem_bar(Op_MemBarCPUOrder);
4197
4198 return true;
4199 }
4200
4201 //------------------------clone_coping-----------------------------------
4202 // Helper function for inline_native_clone.
4203 void LibraryCallKit::copy_to_clone(Node* obj, Node* alloc_obj, Node* obj_size, bool is_array, bool card_mark) {
4204 assert(obj_size != NULL, "");
4205 Node* raw_obj = alloc_obj->in(1);
4206 assert(alloc_obj->is_CheckCastPP() && raw_obj->is_Proj() && raw_obj->in(0)->is_Allocate(), "");
4207
4208 AllocateNode* alloc = NULL;
4209 if (ReduceBulkZeroing) {
4210 // We will be completely responsible for initializing this object -
4211 // mark Initialize node as complete.
4212 alloc = AllocateNode::Ideal_allocation(alloc_obj, &_gvn);
4213 // The object was just allocated - there should be no any stores!
4214 guarantee(alloc != NULL && alloc->maybe_set_complete(&_gvn), "");
4215 // Mark as complete_with_arraycopy so that on AllocateNode
4216 // expansion, we know this AllocateNode is initialized by an array
4217 // copy and a StoreStore barrier exists after the array copy.
4218 alloc->initialization()->set_complete_with_arraycopy();
4219 }
4220
4221 // Copy the fastest available way.
4222 // TODO: generate fields copies for small objects instead.
4223 Node* src = obj;
4224 Node* dest = alloc_obj;
4225 Node* size = _gvn.transform(obj_size);
4226
4227 // Exclude the header but include array length to copy by 8 bytes words.
4228 // Can't use base_offset_in_bytes(bt) since basic type is unknown.
4229 int base_off = is_array ? arrayOopDesc::length_offset_in_bytes() :
4230 instanceOopDesc::base_offset_in_bytes();
4231 // base_off:
4232 // 8 - 32-bit VM
4233 // 12 - 64-bit VM, compressed oops
4234 // 16 - 64-bit VM, normal oops
4235 if (base_off % BytesPerLong != 0) {
4236 assert(UseCompressedOops, "");
4237 if (is_array) {
4238 // Exclude length to copy by 8 bytes words.
4239 base_off += sizeof(int);
4240 } else {
4241 // Include klass to copy by 8 bytes words.
4242 base_off = instanceOopDesc::klass_offset_in_bytes();
4243 }
4244 assert(base_off % BytesPerLong == 0, "expect 8 bytes alignment");
4245 }
4246 src = basic_plus_adr(src, base_off);
4247 dest = basic_plus_adr(dest, base_off);
4248
4249 // Compute the length also, if needed:
4250 Node* countx = size;
4251 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(base_off)) );
4252 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong) ));
4253
4254 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4255 bool disjoint_bases = true;
4256 generate_unchecked_arraycopy(raw_adr_type, T_LONG, disjoint_bases,
4257 src, NULL, dest, NULL, countx,
4258 /*dest_uninitialized*/true);
4259
4260 // If necessary, emit some card marks afterwards. (Non-arrays only.)
4261 if (card_mark) {
4262 assert(!is_array, "");
4263 // Put in store barrier for any and all oops we are sticking
4264 // into this object. (We could avoid this if we could prove
4265 // that the object type contains no oop fields at all.)
4266 Node* no_particular_value = NULL;
4267 Node* no_particular_field = NULL;
4268 int raw_adr_idx = Compile::AliasIdxRaw;
4269 post_barrier(control(),
4270 memory(raw_adr_type),
4271 alloc_obj,
4272 no_particular_field,
4273 raw_adr_idx,
4274 no_particular_value,
4275 T_OBJECT,
4276 false);
4277 }
4278
4279 // Do not let reads from the cloned object float above the arraycopy.
4280 if (alloc != NULL) {
4281 // Do not let stores that initialize this object be reordered with
4282 // a subsequent store that would make this object accessible by
4283 // other threads.
4284 // Record what AllocateNode this StoreStore protects so that
4285 // escape analysis can go from the MemBarStoreStoreNode to the
4286 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
4287 // based on the escape status of the AllocateNode.
4288 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
4289 } else {
4290 insert_mem_bar(Op_MemBarCPUOrder);
4291 }
4292 }
4293
4294 //------------------------inline_native_clone----------------------------
4295 // Here are the simple edge cases:
4296 // null receiver => normal trap
4297 // virtual and clone was overridden => slow path to out-of-line clone
4298 // not cloneable or finalizer => slow path to out-of-line Object.clone
4299 //
4300 // The general case has two steps, allocation and copying.
4301 // Allocation has two cases, and uses GraphKit::new_instance or new_array.
4302 //
4303 // Copying also has two cases, oop arrays and everything else.
4304 // Oop arrays use arrayof_oop_arraycopy (same as System.arraycopy).
4305 // Everything else uses the tight inline loop supplied by CopyArrayNode.
4306 //
4307 // These steps fold up nicely if and when the cloned object's klass
4308 // can be sharply typed as an object array, a type array, or an instance.
4309 //
4310 bool LibraryCallKit::inline_native_clone(bool is_virtual) {
4311 int nargs = 1;
4312 PhiNode* result_val;
4313
4314 //set the original stack and the reexecute bit for the interpreter to reexecute
4315 //the bytecode that invokes Object.clone if deoptimization happens
4316 { PreserveReexecuteState preexecs(this);
4317 jvms()->set_should_reexecute(true);
4318
4319 //null_check_receiver will adjust _sp (push and pop)
4320 Node* obj = null_check_receiver(callee());
4321 if (stopped()) return true;
4322
4323 _sp += nargs;
4324
4325 Node* obj_klass = load_object_klass(obj);
4326 const TypeKlassPtr* tklass = _gvn.type(obj_klass)->isa_klassptr();
4327 const TypeOopPtr* toop = ((tklass != NULL)
4328 ? tklass->as_instance_type()
4329 : TypeInstPtr::NOTNULL);
4330
4331 // Conservatively insert a memory barrier on all memory slices.
4332 // Do not let writes into the original float below the clone.
4333 insert_mem_bar(Op_MemBarCPUOrder);
4334
4335 // paths into result_reg:
4336 enum {
4337 _slow_path = 1, // out-of-line call to clone method (virtual or not)
4338 _objArray_path, // plain array allocation, plus arrayof_oop_arraycopy
4339 _array_path, // plain array allocation, plus arrayof_long_arraycopy
4340 _instance_path, // plain instance allocation, plus arrayof_long_arraycopy
4341 PATH_LIMIT
4342 };
4343 RegionNode* result_reg = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4344 result_val = new(C, PATH_LIMIT) PhiNode(result_reg,
4345 TypeInstPtr::NOTNULL);
4346 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_reg, Type::ABIO);
4347 PhiNode* result_mem = new(C, PATH_LIMIT) PhiNode(result_reg, Type::MEMORY,
4348 TypePtr::BOTTOM);
4349 record_for_igvn(result_reg);
4350
4351 const TypePtr* raw_adr_type = TypeRawPtr::BOTTOM;
4352 int raw_adr_idx = Compile::AliasIdxRaw;
4353
4354 Node* array_ctl = generate_array_guard(obj_klass, (RegionNode*)NULL);
4355 if (array_ctl != NULL) {
4356 // It's an array.
4357 PreserveJVMState pjvms(this);
4358 set_control(array_ctl);
4359 Node* obj_length = load_array_length(obj);
4360 Node* obj_size = NULL;
4361 Node* alloc_obj = new_array(obj_klass, obj_length, 0, &obj_size);
4362
4363 if (!use_ReduceInitialCardMarks()) {
4364 // If it is an oop array, it requires very special treatment,
4365 // because card marking is required on each card of the array.
4366 Node* is_obja = generate_objArray_guard(obj_klass, (RegionNode*)NULL);
4367 if (is_obja != NULL) {
4368 PreserveJVMState pjvms2(this);
4369 set_control(is_obja);
4370 // Generate a direct call to the right arraycopy function(s).
4371 bool disjoint_bases = true;
4372 bool length_never_negative = true;
4373 generate_arraycopy(TypeAryPtr::OOPS, T_OBJECT,
4374 obj, intcon(0), alloc_obj, intcon(0),
4375 obj_length,
4376 disjoint_bases, length_never_negative);
4377 result_reg->init_req(_objArray_path, control());
4378 result_val->init_req(_objArray_path, alloc_obj);
4379 result_i_o ->set_req(_objArray_path, i_o());
4380 result_mem ->set_req(_objArray_path, reset_memory());
4381 }
4382 }
4383 // Otherwise, there are no card marks to worry about.
4384 // (We can dispense with card marks if we know the allocation
4385 // comes out of eden (TLAB)... In fact, ReduceInitialCardMarks
4386 // causes the non-eden paths to take compensating steps to
4387 // simulate a fresh allocation, so that no further
4388 // card marks are required in compiled code to initialize
4389 // the object.)
4390
4391 if (!stopped()) {
4392 copy_to_clone(obj, alloc_obj, obj_size, true, false);
4393
4394 // Present the results of the copy.
4395 result_reg->init_req(_array_path, control());
4396 result_val->init_req(_array_path, alloc_obj);
4397 result_i_o ->set_req(_array_path, i_o());
4398 result_mem ->set_req(_array_path, reset_memory());
4399 }
4400 }
4401
4402 // We only go to the instance fast case code if we pass a number of guards.
4403 // The paths which do not pass are accumulated in the slow_region.
4404 RegionNode* slow_region = new (C, 1) RegionNode(1);
4405 record_for_igvn(slow_region);
4406 if (!stopped()) {
4407 // It's an instance (we did array above). Make the slow-path tests.
4408 // If this is a virtual call, we generate a funny guard. We grab
4409 // the vtable entry corresponding to clone() from the target object.
4410 // If the target method which we are calling happens to be the
4411 // Object clone() method, we pass the guard. We do not need this
4412 // guard for non-virtual calls; the caller is known to be the native
4413 // Object clone().
4414 if (is_virtual) {
4415 generate_virtual_guard(obj_klass, slow_region);
4416 }
4417
4418 // The object must be cloneable and must not have a finalizer.
4419 // Both of these conditions may be checked in a single test.
4420 // We could optimize the cloneable test further, but we don't care.
4421 generate_access_flags_guard(obj_klass,
4422 // Test both conditions:
4423 JVM_ACC_IS_CLONEABLE | JVM_ACC_HAS_FINALIZER,
4424 // Must be cloneable but not finalizer:
4425 JVM_ACC_IS_CLONEABLE,
4426 slow_region);
4427 }
4428
4429 if (!stopped()) {
4430 // It's an instance, and it passed the slow-path tests.
4431 PreserveJVMState pjvms(this);
4432 Node* obj_size = NULL;
4433 Node* alloc_obj = new_instance(obj_klass, NULL, &obj_size);
4434
4435 copy_to_clone(obj, alloc_obj, obj_size, false, !use_ReduceInitialCardMarks());
4436
4437 // Present the results of the slow call.
4438 result_reg->init_req(_instance_path, control());
4439 result_val->init_req(_instance_path, alloc_obj);
4440 result_i_o ->set_req(_instance_path, i_o());
4441 result_mem ->set_req(_instance_path, reset_memory());
4442 }
4443
4444 // Generate code for the slow case. We make a call to clone().
4445 set_control(_gvn.transform(slow_region));
4446 if (!stopped()) {
4447 PreserveJVMState pjvms(this);
4448 CallJavaNode* slow_call = generate_method_call(vmIntrinsics::_clone, is_virtual);
4449 Node* slow_result = set_results_for_java_call(slow_call);
4450 // this->control() comes from set_results_for_java_call
4451 result_reg->init_req(_slow_path, control());
4452 result_val->init_req(_slow_path, slow_result);
4453 result_i_o ->set_req(_slow_path, i_o());
4454 result_mem ->set_req(_slow_path, reset_memory());
4455 }
4456
4457 // Return the combined state.
4458 set_control( _gvn.transform(result_reg) );
4459 set_i_o( _gvn.transform(result_i_o) );
4460 set_all_memory( _gvn.transform(result_mem) );
4461 } //original reexecute and sp are set back here
4462
4463 push(_gvn.transform(result_val));
4464
4465 return true;
4466 }
4467
4468 //------------------------------basictype2arraycopy----------------------------
4469 address LibraryCallKit::basictype2arraycopy(BasicType t,
4470 Node* src_offset,
4471 Node* dest_offset,
4472 bool disjoint_bases,
4473 const char* &name,
4474 bool dest_uninitialized) {
4475 const TypeInt* src_offset_inttype = gvn().find_int_type(src_offset);;
4476 const TypeInt* dest_offset_inttype = gvn().find_int_type(dest_offset);;
4477
4478 bool aligned = false;
4479 bool disjoint = disjoint_bases;
4480
4481 // if the offsets are the same, we can treat the memory regions as
4482 // disjoint, because either the memory regions are in different arrays,
4483 // or they are identical (which we can treat as disjoint.) We can also
4484 // treat a copy with a destination index less that the source index
4485 // as disjoint since a low->high copy will work correctly in this case.
4486 if (src_offset_inttype != NULL && src_offset_inttype->is_con() &&
4487 dest_offset_inttype != NULL && dest_offset_inttype->is_con()) {
4488 // both indices are constants
4489 int s_offs = src_offset_inttype->get_con();
4490 int d_offs = dest_offset_inttype->get_con();
4491 int element_size = type2aelembytes(t);
4492 aligned = ((arrayOopDesc::base_offset_in_bytes(t) + s_offs * element_size) % HeapWordSize == 0) &&
4493 ((arrayOopDesc::base_offset_in_bytes(t) + d_offs * element_size) % HeapWordSize == 0);
4494 if (s_offs >= d_offs) disjoint = true;
4495 } else if (src_offset == dest_offset && src_offset != NULL) {
4496 // This can occur if the offsets are identical non-constants.
4497 disjoint = true;
4498 }
4499
4500 return StubRoutines::select_arraycopy_function(t, aligned, disjoint, name, dest_uninitialized);
4501 }
4502
4503
4504 //------------------------------inline_arraycopy-----------------------
4505 bool LibraryCallKit::inline_arraycopy() {
4506 // Restore the stack and pop off the arguments.
4507 int nargs = 5; // 2 oops, 3 ints, no size_t or long
4508 assert(callee()->signature()->size() == nargs, "copy has 5 arguments");
4509
4510 Node *src = argument(0);
4511 Node *src_offset = argument(1);
4512 Node *dest = argument(2);
4513 Node *dest_offset = argument(3);
4514 Node *length = argument(4);
4515
4516 // Compile time checks. If any of these checks cannot be verified at compile time,
4517 // we do not make a fast path for this call. Instead, we let the call remain as it
4518 // is. The checks we choose to mandate at compile time are:
4519 //
4520 // (1) src and dest are arrays.
4521 const Type* src_type = src->Value(&_gvn);
4522 const Type* dest_type = dest->Value(&_gvn);
4523 const TypeAryPtr* top_src = src_type->isa_aryptr();
4524 const TypeAryPtr* top_dest = dest_type->isa_aryptr();
4525 if (top_src == NULL || top_src->klass() == NULL ||
4526 top_dest == NULL || top_dest->klass() == NULL) {
4527 // Conservatively insert a memory barrier on all memory slices.
4528 // Do not let writes into the source float below the arraycopy.
4529 insert_mem_bar(Op_MemBarCPUOrder);
4530
4531 // Call StubRoutines::generic_arraycopy stub.
4532 generate_arraycopy(TypeRawPtr::BOTTOM, T_CONFLICT,
4533 src, src_offset, dest, dest_offset, length);
4534
4535 // Do not let reads from the destination float above the arraycopy.
4536 // Since we cannot type the arrays, we don't know which slices
4537 // might be affected. We could restrict this barrier only to those
4538 // memory slices which pertain to array elements--but don't bother.
4539 if (!InsertMemBarAfterArraycopy)
4540 // (If InsertMemBarAfterArraycopy, there is already one in place.)
4541 insert_mem_bar(Op_MemBarCPUOrder);
4542 return true;
4543 }
4544
4545 // (2) src and dest arrays must have elements of the same BasicType
4546 // Figure out the size and type of the elements we will be copying.
4547 BasicType src_elem = top_src->klass()->as_array_klass()->element_type()->basic_type();
4548 BasicType dest_elem = top_dest->klass()->as_array_klass()->element_type()->basic_type();
4549 if (src_elem == T_ARRAY) src_elem = T_OBJECT;
4550 if (dest_elem == T_ARRAY) dest_elem = T_OBJECT;
4551
4552 if (src_elem != dest_elem || dest_elem == T_VOID) {
4553 // The component types are not the same or are not recognized. Punt.
4554 // (But, avoid the native method wrapper to JVM_ArrayCopy.)
4555 generate_slow_arraycopy(TypePtr::BOTTOM,
4556 src, src_offset, dest, dest_offset, length,
4557 /*dest_uninitialized*/false);
4558 return true;
4559 }
4560
4561 //---------------------------------------------------------------------------
4562 // We will make a fast path for this call to arraycopy.
4563
4564 // We have the following tests left to perform:
4565 //
4566 // (3) src and dest must not be null.
4567 // (4) src_offset must not be negative.
4568 // (5) dest_offset must not be negative.
4569 // (6) length must not be negative.
4570 // (7) src_offset + length must not exceed length of src.
4571 // (8) dest_offset + length must not exceed length of dest.
4572 // (9) each element of an oop array must be assignable
4573
4574 RegionNode* slow_region = new (C, 1) RegionNode(1);
4575 record_for_igvn(slow_region);
4576
4577 // (3) operands must not be null
4578 // We currently perform our null checks with the do_null_check routine.
4579 // This means that the null exceptions will be reported in the caller
4580 // rather than (correctly) reported inside of the native arraycopy call.
4581 // This should be corrected, given time. We do our null check with the
4582 // stack pointer restored.
4583 _sp += nargs;
4584 src = do_null_check(src, T_ARRAY);
4585 dest = do_null_check(dest, T_ARRAY);
4586 _sp -= nargs;
4587
4588 // (4) src_offset must not be negative.
4589 generate_negative_guard(src_offset, slow_region);
4590
4591 // (5) dest_offset must not be negative.
4592 generate_negative_guard(dest_offset, slow_region);
4593
4594 // (6) length must not be negative (moved to generate_arraycopy()).
4595 // generate_negative_guard(length, slow_region);
4596
4597 // (7) src_offset + length must not exceed length of src.
4598 generate_limit_guard(src_offset, length,
4599 load_array_length(src),
4600 slow_region);
4601
4602 // (8) dest_offset + length must not exceed length of dest.
4603 generate_limit_guard(dest_offset, length,
4604 load_array_length(dest),
4605 slow_region);
4606
4607 // (9) each element of an oop array must be assignable
4608 // The generate_arraycopy subroutine checks this.
4609
4610 // This is where the memory effects are placed:
4611 const TypePtr* adr_type = TypeAryPtr::get_array_body_type(dest_elem);
4612 generate_arraycopy(adr_type, dest_elem,
4613 src, src_offset, dest, dest_offset, length,
4614 false, false, slow_region);
4615
4616 return true;
4617 }
4618
4619 //-----------------------------generate_arraycopy----------------------
4620 // Generate an optimized call to arraycopy.
4621 // Caller must guard against non-arrays.
4622 // Caller must determine a common array basic-type for both arrays.
4623 // Caller must validate offsets against array bounds.
4624 // The slow_region has already collected guard failure paths
4625 // (such as out of bounds length or non-conformable array types).
4626 // The generated code has this shape, in general:
4627 //
4628 // if (length == 0) return // via zero_path
4629 // slowval = -1
4630 // if (types unknown) {
4631 // slowval = call generic copy loop
4632 // if (slowval == 0) return // via checked_path
4633 // } else if (indexes in bounds) {
4634 // if ((is object array) && !(array type check)) {
4635 // slowval = call checked copy loop
4636 // if (slowval == 0) return // via checked_path
4637 // } else {
4638 // call bulk copy loop
4639 // return // via fast_path
4640 // }
4641 // }
4642 // // adjust params for remaining work:
4643 // if (slowval != -1) {
4644 // n = -1^slowval; src_offset += n; dest_offset += n; length -= n
4645 // }
4646 // slow_region:
4647 // call slow arraycopy(src, src_offset, dest, dest_offset, length)
4648 // return // via slow_call_path
4649 //
4650 // This routine is used from several intrinsics: System.arraycopy,
4651 // Object.clone (the array subcase), and Arrays.copyOf[Range].
4652 //
4653 void
4654 LibraryCallKit::generate_arraycopy(const TypePtr* adr_type,
4655 BasicType basic_elem_type,
4656 Node* src, Node* src_offset,
4657 Node* dest, Node* dest_offset,
4658 Node* copy_length,
4659 bool disjoint_bases,
4660 bool length_never_negative,
4661 RegionNode* slow_region) {
4662
4663 if (slow_region == NULL) {
4664 slow_region = new(C,1) RegionNode(1);
4665 record_for_igvn(slow_region);
4666 }
4667
4668 Node* original_dest = dest;
4669 AllocateArrayNode* alloc = NULL; // used for zeroing, if needed
4670 bool dest_uninitialized = false;
4671
4672 // See if this is the initialization of a newly-allocated array.
4673 // If so, we will take responsibility here for initializing it to zero.
4674 // (Note: Because tightly_coupled_allocation performs checks on the
4675 // out-edges of the dest, we need to avoid making derived pointers
4676 // from it until we have checked its uses.)
4677 if (ReduceBulkZeroing
4678 && !ZeroTLAB // pointless if already zeroed
4679 && basic_elem_type != T_CONFLICT // avoid corner case
4680 && !src->eqv_uncast(dest)
4681 && ((alloc = tightly_coupled_allocation(dest, slow_region))
4682 != NULL)
4683 && _gvn.find_int_con(alloc->in(AllocateNode::ALength), 1) > 0
4684 && alloc->maybe_set_complete(&_gvn)) {
4685 // "You break it, you buy it."
4686 InitializeNode* init = alloc->initialization();
4687 assert(init->is_complete(), "we just did this");
4688 init->set_complete_with_arraycopy();
4689 assert(dest->is_CheckCastPP(), "sanity");
4690 assert(dest->in(0)->in(0) == init, "dest pinned");
4691 adr_type = TypeRawPtr::BOTTOM; // all initializations are into raw memory
4692 // From this point on, every exit path is responsible for
4693 // initializing any non-copied parts of the object to zero.
4694 // Also, if this flag is set we make sure that arraycopy interacts properly
4695 // with G1, eliding pre-barriers. See CR 6627983.
4696 dest_uninitialized = true;
4697 } else {
4698 // No zeroing elimination here.
4699 alloc = NULL;
4700 //original_dest = dest;
4701 //dest_uninitialized = false;
4702 }
4703
4704 // Results are placed here:
4705 enum { fast_path = 1, // normal void-returning assembly stub
4706 checked_path = 2, // special assembly stub with cleanup
4707 slow_call_path = 3, // something went wrong; call the VM
4708 zero_path = 4, // bypass when length of copy is zero
4709 bcopy_path = 5, // copy primitive array by 64-bit blocks
4710 PATH_LIMIT = 6
4711 };
4712 RegionNode* result_region = new(C, PATH_LIMIT) RegionNode(PATH_LIMIT);
4713 PhiNode* result_i_o = new(C, PATH_LIMIT) PhiNode(result_region, Type::ABIO);
4714 PhiNode* result_memory = new(C, PATH_LIMIT) PhiNode(result_region, Type::MEMORY, adr_type);
4715 record_for_igvn(result_region);
4716 _gvn.set_type_bottom(result_i_o);
4717 _gvn.set_type_bottom(result_memory);
4718 assert(adr_type != TypePtr::BOTTOM, "must be RawMem or a T[] slice");
4719
4720 // The slow_control path:
4721 Node* slow_control;
4722 Node* slow_i_o = i_o();
4723 Node* slow_mem = memory(adr_type);
4724 debug_only(slow_control = (Node*) badAddress);
4725
4726 // Checked control path:
4727 Node* checked_control = top();
4728 Node* checked_mem = NULL;
4729 Node* checked_i_o = NULL;
4730 Node* checked_value = NULL;
4731
4732 if (basic_elem_type == T_CONFLICT) {
4733 assert(!dest_uninitialized, "");
4734 Node* cv = generate_generic_arraycopy(adr_type,
4735 src, src_offset, dest, dest_offset,
4736 copy_length, dest_uninitialized);
4737 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4738 checked_control = control();
4739 checked_i_o = i_o();
4740 checked_mem = memory(adr_type);
4741 checked_value = cv;
4742 set_control(top()); // no fast path
4743 }
4744
4745 Node* not_pos = generate_nonpositive_guard(copy_length, length_never_negative);
4746 if (not_pos != NULL) {
4747 PreserveJVMState pjvms(this);
4748 set_control(not_pos);
4749
4750 // (6) length must not be negative.
4751 if (!length_never_negative) {
4752 generate_negative_guard(copy_length, slow_region);
4753 }
4754
4755 // copy_length is 0.
4756 if (!stopped() && dest_uninitialized) {
4757 Node* dest_length = alloc->in(AllocateNode::ALength);
4758 if (copy_length->eqv_uncast(dest_length)
4759 || _gvn.find_int_con(dest_length, 1) <= 0) {
4760 // There is no zeroing to do. No need for a secondary raw memory barrier.
4761 } else {
4762 // Clear the whole thing since there are no source elements to copy.
4763 generate_clear_array(adr_type, dest, basic_elem_type,
4764 intcon(0), NULL,
4765 alloc->in(AllocateNode::AllocSize));
4766 // Use a secondary InitializeNode as raw memory barrier.
4767 // Currently it is needed only on this path since other
4768 // paths have stub or runtime calls as raw memory barriers.
4769 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize,
4770 Compile::AliasIdxRaw,
4771 top())->as_Initialize();
4772 init->set_complete(&_gvn); // (there is no corresponding AllocateNode)
4773 }
4774 }
4775
4776 // Present the results of the fast call.
4777 result_region->init_req(zero_path, control());
4778 result_i_o ->init_req(zero_path, i_o());
4779 result_memory->init_req(zero_path, memory(adr_type));
4780 }
4781
4782 if (!stopped() && dest_uninitialized) {
4783 // We have to initialize the *uncopied* part of the array to zero.
4784 // The copy destination is the slice dest[off..off+len]. The other slices
4785 // are dest_head = dest[0..off] and dest_tail = dest[off+len..dest.length].
4786 Node* dest_size = alloc->in(AllocateNode::AllocSize);
4787 Node* dest_length = alloc->in(AllocateNode::ALength);
4788 Node* dest_tail = _gvn.transform( new(C,3) AddINode(dest_offset,
4789 copy_length) );
4790
4791 // If there is a head section that needs zeroing, do it now.
4792 if (find_int_con(dest_offset, -1) != 0) {
4793 generate_clear_array(adr_type, dest, basic_elem_type,
4794 intcon(0), dest_offset,
4795 NULL);
4796 }
4797
4798 // Next, perform a dynamic check on the tail length.
4799 // It is often zero, and we can win big if we prove this.
4800 // There are two wins: Avoid generating the ClearArray
4801 // with its attendant messy index arithmetic, and upgrade
4802 // the copy to a more hardware-friendly word size of 64 bits.
4803 Node* tail_ctl = NULL;
4804 if (!stopped() && !dest_tail->eqv_uncast(dest_length)) {
4805 Node* cmp_lt = _gvn.transform( new(C,3) CmpINode(dest_tail, dest_length) );
4806 Node* bol_lt = _gvn.transform( new(C,2) BoolNode(cmp_lt, BoolTest::lt) );
4807 tail_ctl = generate_slow_guard(bol_lt, NULL);
4808 assert(tail_ctl != NULL || !stopped(), "must be an outcome");
4809 }
4810
4811 // At this point, let's assume there is no tail.
4812 if (!stopped() && alloc != NULL && basic_elem_type != T_OBJECT) {
4813 // There is no tail. Try an upgrade to a 64-bit copy.
4814 bool didit = false;
4815 { PreserveJVMState pjvms(this);
4816 didit = generate_block_arraycopy(adr_type, basic_elem_type, alloc,
4817 src, src_offset, dest, dest_offset,
4818 dest_size, dest_uninitialized);
4819 if (didit) {
4820 // Present the results of the block-copying fast call.
4821 result_region->init_req(bcopy_path, control());
4822 result_i_o ->init_req(bcopy_path, i_o());
4823 result_memory->init_req(bcopy_path, memory(adr_type));
4824 }
4825 }
4826 if (didit)
4827 set_control(top()); // no regular fast path
4828 }
4829
4830 // Clear the tail, if any.
4831 if (tail_ctl != NULL) {
4832 Node* notail_ctl = stopped() ? NULL : control();
4833 set_control(tail_ctl);
4834 if (notail_ctl == NULL) {
4835 generate_clear_array(adr_type, dest, basic_elem_type,
4836 dest_tail, NULL,
4837 dest_size);
4838 } else {
4839 // Make a local merge.
4840 Node* done_ctl = new(C,3) RegionNode(3);
4841 Node* done_mem = new(C,3) PhiNode(done_ctl, Type::MEMORY, adr_type);
4842 done_ctl->init_req(1, notail_ctl);
4843 done_mem->init_req(1, memory(adr_type));
4844 generate_clear_array(adr_type, dest, basic_elem_type,
4845 dest_tail, NULL,
4846 dest_size);
4847 done_ctl->init_req(2, control());
4848 done_mem->init_req(2, memory(adr_type));
4849 set_control( _gvn.transform(done_ctl) );
4850 set_memory( _gvn.transform(done_mem), adr_type );
4851 }
4852 }
4853 }
4854
4855 BasicType copy_type = basic_elem_type;
4856 assert(basic_elem_type != T_ARRAY, "caller must fix this");
4857 if (!stopped() && copy_type == T_OBJECT) {
4858 // If src and dest have compatible element types, we can copy bits.
4859 // Types S[] and D[] are compatible if D is a supertype of S.
4860 //
4861 // If they are not, we will use checked_oop_disjoint_arraycopy,
4862 // which performs a fast optimistic per-oop check, and backs off
4863 // further to JVM_ArrayCopy on the first per-oop check that fails.
4864 // (Actually, we don't move raw bits only; the GC requires card marks.)
4865
4866 // Get the klassOop for both src and dest
4867 Node* src_klass = load_object_klass(src);
4868 Node* dest_klass = load_object_klass(dest);
4869
4870 // Generate the subtype check.
4871 // This might fold up statically, or then again it might not.
4872 //
4873 // Non-static example: Copying List<String>.elements to a new String[].
4874 // The backing store for a List<String> is always an Object[],
4875 // but its elements are always type String, if the generic types
4876 // are correct at the source level.
4877 //
4878 // Test S[] against D[], not S against D, because (probably)
4879 // the secondary supertype cache is less busy for S[] than S.
4880 // This usually only matters when D is an interface.
4881 Node* not_subtype_ctrl = gen_subtype_check(src_klass, dest_klass);
4882 // Plug failing path into checked_oop_disjoint_arraycopy
4883 if (not_subtype_ctrl != top()) {
4884 PreserveJVMState pjvms(this);
4885 set_control(not_subtype_ctrl);
4886 // (At this point we can assume disjoint_bases, since types differ.)
4887 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
4888 Node* p1 = basic_plus_adr(dest_klass, ek_offset);
4889 Node* n1 = LoadKlassNode::make(_gvn, immutable_memory(), p1, TypeRawPtr::BOTTOM);
4890 Node* dest_elem_klass = _gvn.transform(n1);
4891 Node* cv = generate_checkcast_arraycopy(adr_type,
4892 dest_elem_klass,
4893 src, src_offset, dest, dest_offset,
4894 ConvI2X(copy_length), dest_uninitialized);
4895 if (cv == NULL) cv = intcon(-1); // failure (no stub available)
4896 checked_control = control();
4897 checked_i_o = i_o();
4898 checked_mem = memory(adr_type);
4899 checked_value = cv;
4900 }
4901 // At this point we know we do not need type checks on oop stores.
4902
4903 // Let's see if we need card marks:
4904 if (alloc != NULL && use_ReduceInitialCardMarks()) {
4905 // If we do not need card marks, copy using the jint or jlong stub.
4906 copy_type = LP64_ONLY(UseCompressedOops ? T_INT : T_LONG) NOT_LP64(T_INT);
4907 assert(type2aelembytes(basic_elem_type) == type2aelembytes(copy_type),
4908 "sizes agree");
4909 }
4910 }
4911
4912 if (!stopped()) {
4913 // Generate the fast path, if possible.
4914 PreserveJVMState pjvms(this);
4915 generate_unchecked_arraycopy(adr_type, copy_type, disjoint_bases,
4916 src, src_offset, dest, dest_offset,
4917 ConvI2X(copy_length), dest_uninitialized);
4918
4919 // Present the results of the fast call.
4920 result_region->init_req(fast_path, control());
4921 result_i_o ->init_req(fast_path, i_o());
4922 result_memory->init_req(fast_path, memory(adr_type));
4923 }
4924
4925 // Here are all the slow paths up to this point, in one bundle:
4926 slow_control = top();
4927 if (slow_region != NULL)
4928 slow_control = _gvn.transform(slow_region);
4929 debug_only(slow_region = (RegionNode*)badAddress);
4930
4931 set_control(checked_control);
4932 if (!stopped()) {
4933 // Clean up after the checked call.
4934 // The returned value is either 0 or -1^K,
4935 // where K = number of partially transferred array elements.
4936 Node* cmp = _gvn.transform( new(C, 3) CmpINode(checked_value, intcon(0)) );
4937 Node* bol = _gvn.transform( new(C, 2) BoolNode(cmp, BoolTest::eq) );
4938 IfNode* iff = create_and_map_if(control(), bol, PROB_MAX, COUNT_UNKNOWN);
4939
4940 // If it is 0, we are done, so transfer to the end.
4941 Node* checks_done = _gvn.transform( new(C, 1) IfTrueNode(iff) );
4942 result_region->init_req(checked_path, checks_done);
4943 result_i_o ->init_req(checked_path, checked_i_o);
4944 result_memory->init_req(checked_path, checked_mem);
4945
4946 // If it is not zero, merge into the slow call.
4947 set_control( _gvn.transform( new(C, 1) IfFalseNode(iff) ));
4948 RegionNode* slow_reg2 = new(C, 3) RegionNode(3);
4949 PhiNode* slow_i_o2 = new(C, 3) PhiNode(slow_reg2, Type::ABIO);
4950 PhiNode* slow_mem2 = new(C, 3) PhiNode(slow_reg2, Type::MEMORY, adr_type);
4951 record_for_igvn(slow_reg2);
4952 slow_reg2 ->init_req(1, slow_control);
4953 slow_i_o2 ->init_req(1, slow_i_o);
4954 slow_mem2 ->init_req(1, slow_mem);
4955 slow_reg2 ->init_req(2, control());
4956 slow_i_o2 ->init_req(2, checked_i_o);
4957 slow_mem2 ->init_req(2, checked_mem);
4958
4959 slow_control = _gvn.transform(slow_reg2);
4960 slow_i_o = _gvn.transform(slow_i_o2);
4961 slow_mem = _gvn.transform(slow_mem2);
4962
4963 if (alloc != NULL) {
4964 // We'll restart from the very beginning, after zeroing the whole thing.
4965 // This can cause double writes, but that's OK since dest is brand new.
4966 // So we ignore the low 31 bits of the value returned from the stub.
4967 } else {
4968 // We must continue the copy exactly where it failed, or else
4969 // another thread might see the wrong number of writes to dest.
4970 Node* checked_offset = _gvn.transform( new(C, 3) XorINode(checked_value, intcon(-1)) );
4971 Node* slow_offset = new(C, 3) PhiNode(slow_reg2, TypeInt::INT);
4972 slow_offset->init_req(1, intcon(0));
4973 slow_offset->init_req(2, checked_offset);
4974 slow_offset = _gvn.transform(slow_offset);
4975
4976 // Adjust the arguments by the conditionally incoming offset.
4977 Node* src_off_plus = _gvn.transform( new(C, 3) AddINode(src_offset, slow_offset) );
4978 Node* dest_off_plus = _gvn.transform( new(C, 3) AddINode(dest_offset, slow_offset) );
4979 Node* length_minus = _gvn.transform( new(C, 3) SubINode(copy_length, slow_offset) );
4980
4981 // Tweak the node variables to adjust the code produced below:
4982 src_offset = src_off_plus;
4983 dest_offset = dest_off_plus;
4984 copy_length = length_minus;
4985 }
4986 }
4987
4988 set_control(slow_control);
4989 if (!stopped()) {
4990 // Generate the slow path, if needed.
4991 PreserveJVMState pjvms(this); // replace_in_map may trash the map
4992
4993 set_memory(slow_mem, adr_type);
4994 set_i_o(slow_i_o);
4995
4996 if (dest_uninitialized) {
4997 generate_clear_array(adr_type, dest, basic_elem_type,
4998 intcon(0), NULL,
4999 alloc->in(AllocateNode::AllocSize));
5000 }
5001
5002 generate_slow_arraycopy(adr_type,
5003 src, src_offset, dest, dest_offset,
5004 copy_length, /*dest_uninitialized*/false);
5005
5006 result_region->init_req(slow_call_path, control());
5007 result_i_o ->init_req(slow_call_path, i_o());
5008 result_memory->init_req(slow_call_path, memory(adr_type));
5009 }
5010
5011 // Remove unused edges.
5012 for (uint i = 1; i < result_region->req(); i++) {
5013 if (result_region->in(i) == NULL)
5014 result_region->init_req(i, top());
5015 }
5016
5017 // Finished; return the combined state.
5018 set_control( _gvn.transform(result_region) );
5019 set_i_o( _gvn.transform(result_i_o) );
5020 set_memory( _gvn.transform(result_memory), adr_type );
5021
5022 // The memory edges above are precise in order to model effects around
5023 // array copies accurately to allow value numbering of field loads around
5024 // arraycopy. Such field loads, both before and after, are common in Java
5025 // collections and similar classes involving header/array data structures.
5026 //
5027 // But with low number of register or when some registers are used or killed
5028 // by arraycopy calls it causes registers spilling on stack. See 6544710.
5029 // The next memory barrier is added to avoid it. If the arraycopy can be
5030 // optimized away (which it can, sometimes) then we can manually remove
5031 // the membar also.
5032 //
5033 // Do not let reads from the cloned object float above the arraycopy.
5034 if (alloc != NULL) {
5035 // Do not let stores that initialize this object be reordered with
5036 // a subsequent store that would make this object accessible by
5037 // other threads.
5038 // Record what AllocateNode this StoreStore protects so that
5039 // escape analysis can go from the MemBarStoreStoreNode to the
5040 // AllocateNode and eliminate the MemBarStoreStoreNode if possible
5041 // based on the escape status of the AllocateNode.
5042 insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
5043 } else if (InsertMemBarAfterArraycopy)
5044 insert_mem_bar(Op_MemBarCPUOrder);
5045 }
5046
5047
5048 // Helper function which determines if an arraycopy immediately follows
5049 // an allocation, with no intervening tests or other escapes for the object.
5050 AllocateArrayNode*
5051 LibraryCallKit::tightly_coupled_allocation(Node* ptr,
5052 RegionNode* slow_region) {
5053 if (stopped()) return NULL; // no fast path
5054 if (C->AliasLevel() == 0) return NULL; // no MergeMems around
5055
5056 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(ptr, &_gvn);
5057 if (alloc == NULL) return NULL;
5058
5059 Node* rawmem = memory(Compile::AliasIdxRaw);
5060 // Is the allocation's memory state untouched?
5061 if (!(rawmem->is_Proj() && rawmem->in(0)->is_Initialize())) {
5062 // Bail out if there have been raw-memory effects since the allocation.
5063 // (Example: There might have been a call or safepoint.)
5064 return NULL;
5065 }
5066 rawmem = rawmem->in(0)->as_Initialize()->memory(Compile::AliasIdxRaw);
5067 if (!(rawmem->is_Proj() && rawmem->in(0) == alloc)) {
5068 return NULL;
5069 }
5070
5071 // There must be no unexpected observers of this allocation.
5072 for (DUIterator_Fast imax, i = ptr->fast_outs(imax); i < imax; i++) {
5073 Node* obs = ptr->fast_out(i);
5074 if (obs != this->map()) {
5075 return NULL;
5076 }
5077 }
5078
5079 // This arraycopy must unconditionally follow the allocation of the ptr.
5080 Node* alloc_ctl = ptr->in(0);
5081 assert(just_allocated_object(alloc_ctl) == ptr, "most recent allo");
5082
5083 Node* ctl = control();
5084 while (ctl != alloc_ctl) {
5085 // There may be guards which feed into the slow_region.
5086 // Any other control flow means that we might not get a chance
5087 // to finish initializing the allocated object.
5088 if ((ctl->is_IfFalse() || ctl->is_IfTrue()) && ctl->in(0)->is_If()) {
5089 IfNode* iff = ctl->in(0)->as_If();
5090 Node* not_ctl = iff->proj_out(1 - ctl->as_Proj()->_con);
5091 assert(not_ctl != NULL && not_ctl != ctl, "found alternate");
5092 if (slow_region != NULL && slow_region->find_edge(not_ctl) >= 1) {
5093 ctl = iff->in(0); // This test feeds the known slow_region.
5094 continue;
5095 }
5096 // One more try: Various low-level checks bottom out in
5097 // uncommon traps. If the debug-info of the trap omits
5098 // any reference to the allocation, as we've already
5099 // observed, then there can be no objection to the trap.
5100 bool found_trap = false;
5101 for (DUIterator_Fast jmax, j = not_ctl->fast_outs(jmax); j < jmax; j++) {
5102 Node* obs = not_ctl->fast_out(j);
5103 if (obs->in(0) == not_ctl && obs->is_Call() &&
5104 (obs->as_Call()->entry_point() == SharedRuntime::uncommon_trap_blob()->entry_point())) {
5105 found_trap = true; break;
5106 }
5107 }
5108 if (found_trap) {
5109 ctl = iff->in(0); // This test feeds a harmless uncommon trap.
5110 continue;
5111 }
5112 }
5113 return NULL;
5114 }
5115
5116 // If we get this far, we have an allocation which immediately
5117 // precedes the arraycopy, and we can take over zeroing the new object.
5118 // The arraycopy will finish the initialization, and provide
5119 // a new control state to which we will anchor the destination pointer.
5120
5121 return alloc;
5122 }
5123
5124 // Helper for initialization of arrays, creating a ClearArray.
5125 // It writes zero bits in [start..end), within the body of an array object.
5126 // The memory effects are all chained onto the 'adr_type' alias category.
5127 //
5128 // Since the object is otherwise uninitialized, we are free
5129 // to put a little "slop" around the edges of the cleared area,
5130 // as long as it does not go back into the array's header,
5131 // or beyond the array end within the heap.
5132 //
5133 // The lower edge can be rounded down to the nearest jint and the
5134 // upper edge can be rounded up to the nearest MinObjAlignmentInBytes.
5135 //
5136 // Arguments:
5137 // adr_type memory slice where writes are generated
5138 // dest oop of the destination array
5139 // basic_elem_type element type of the destination
5140 // slice_idx array index of first element to store
5141 // slice_len number of elements to store (or NULL)
5142 // dest_size total size in bytes of the array object
5143 //
5144 // Exactly one of slice_len or dest_size must be non-NULL.
5145 // If dest_size is non-NULL, zeroing extends to the end of the object.
5146 // If slice_len is non-NULL, the slice_idx value must be a constant.
5147 void
5148 LibraryCallKit::generate_clear_array(const TypePtr* adr_type,
5149 Node* dest,
5150 BasicType basic_elem_type,
5151 Node* slice_idx,
5152 Node* slice_len,
5153 Node* dest_size) {
5154 // one or the other but not both of slice_len and dest_size:
5155 assert((slice_len != NULL? 1: 0) + (dest_size != NULL? 1: 0) == 1, "");
5156 if (slice_len == NULL) slice_len = top();
5157 if (dest_size == NULL) dest_size = top();
5158
5159 // operate on this memory slice:
5160 Node* mem = memory(adr_type); // memory slice to operate on
5161
5162 // scaling and rounding of indexes:
5163 int scale = exact_log2(type2aelembytes(basic_elem_type));
5164 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5165 int clear_low = (-1 << scale) & (BytesPerInt - 1);
5166 int bump_bit = (-1 << scale) & BytesPerInt;
5167
5168 // determine constant starts and ends
5169 const intptr_t BIG_NEG = -128;
5170 assert(BIG_NEG + 2*abase < 0, "neg enough");
5171 intptr_t slice_idx_con = (intptr_t) find_int_con(slice_idx, BIG_NEG);
5172 intptr_t slice_len_con = (intptr_t) find_int_con(slice_len, BIG_NEG);
5173 if (slice_len_con == 0) {
5174 return; // nothing to do here
5175 }
5176 intptr_t start_con = (abase + (slice_idx_con << scale)) & ~clear_low;
5177 intptr_t end_con = find_intptr_t_con(dest_size, -1);
5178 if (slice_idx_con >= 0 && slice_len_con >= 0) {
5179 assert(end_con < 0, "not two cons");
5180 end_con = round_to(abase + ((slice_idx_con + slice_len_con) << scale),
5181 BytesPerLong);
5182 }
5183
5184 if (start_con >= 0 && end_con >= 0) {
5185 // Constant start and end. Simple.
5186 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5187 start_con, end_con, &_gvn);
5188 } else if (start_con >= 0 && dest_size != top()) {
5189 // Constant start, pre-rounded end after the tail of the array.
5190 Node* end = dest_size;
5191 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5192 start_con, end, &_gvn);
5193 } else if (start_con >= 0 && slice_len != top()) {
5194 // Constant start, non-constant end. End needs rounding up.
5195 // End offset = round_up(abase + ((slice_idx_con + slice_len) << scale), 8)
5196 intptr_t end_base = abase + (slice_idx_con << scale);
5197 int end_round = (-1 << scale) & (BytesPerLong - 1);
5198 Node* end = ConvI2X(slice_len);
5199 if (scale != 0)
5200 end = _gvn.transform( new(C,3) LShiftXNode(end, intcon(scale) ));
5201 end_base += end_round;
5202 end = _gvn.transform( new(C,3) AddXNode(end, MakeConX(end_base)) );
5203 end = _gvn.transform( new(C,3) AndXNode(end, MakeConX(~end_round)) );
5204 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5205 start_con, end, &_gvn);
5206 } else if (start_con < 0 && dest_size != top()) {
5207 // Non-constant start, pre-rounded end after the tail of the array.
5208 // This is almost certainly a "round-to-end" operation.
5209 Node* start = slice_idx;
5210 start = ConvI2X(start);
5211 if (scale != 0)
5212 start = _gvn.transform( new(C,3) LShiftXNode( start, intcon(scale) ));
5213 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(abase)) );
5214 if ((bump_bit | clear_low) != 0) {
5215 int to_clear = (bump_bit | clear_low);
5216 // Align up mod 8, then store a jint zero unconditionally
5217 // just before the mod-8 boundary.
5218 if (((abase + bump_bit) & ~to_clear) - bump_bit
5219 < arrayOopDesc::length_offset_in_bytes() + BytesPerInt) {
5220 bump_bit = 0;
5221 assert((abase & to_clear) == 0, "array base must be long-aligned");
5222 } else {
5223 // Bump 'start' up to (or past) the next jint boundary:
5224 start = _gvn.transform( new(C,3) AddXNode(start, MakeConX(bump_bit)) );
5225 assert((abase & clear_low) == 0, "array base must be int-aligned");
5226 }
5227 // Round bumped 'start' down to jlong boundary in body of array.
5228 start = _gvn.transform( new(C,3) AndXNode(start, MakeConX(~to_clear)) );
5229 if (bump_bit != 0) {
5230 // Store a zero to the immediately preceding jint:
5231 Node* x1 = _gvn.transform( new(C,3) AddXNode(start, MakeConX(-bump_bit)) );
5232 Node* p1 = basic_plus_adr(dest, x1);
5233 mem = StoreNode::make(_gvn, control(), mem, p1, adr_type, intcon(0), T_INT);
5234 mem = _gvn.transform(mem);
5235 }
5236 }
5237 Node* end = dest_size; // pre-rounded
5238 mem = ClearArrayNode::clear_memory(control(), mem, dest,
5239 start, end, &_gvn);
5240 } else {
5241 // Non-constant start, unrounded non-constant end.
5242 // (Nobody zeroes a random midsection of an array using this routine.)
5243 ShouldNotReachHere(); // fix caller
5244 }
5245
5246 // Done.
5247 set_memory(mem, adr_type);
5248 }
5249
5250
5251 bool
5252 LibraryCallKit::generate_block_arraycopy(const TypePtr* adr_type,
5253 BasicType basic_elem_type,
5254 AllocateNode* alloc,
5255 Node* src, Node* src_offset,
5256 Node* dest, Node* dest_offset,
5257 Node* dest_size, bool dest_uninitialized) {
5258 // See if there is an advantage from block transfer.
5259 int scale = exact_log2(type2aelembytes(basic_elem_type));
5260 if (scale >= LogBytesPerLong)
5261 return false; // it is already a block transfer
5262
5263 // Look at the alignment of the starting offsets.
5264 int abase = arrayOopDesc::base_offset_in_bytes(basic_elem_type);
5265
5266 intptr_t src_off_con = (intptr_t) find_int_con(src_offset, -1);
5267 intptr_t dest_off_con = (intptr_t) find_int_con(dest_offset, -1);
5268 if (src_off_con < 0 || dest_off_con < 0)
5269 // At present, we can only understand constants.
5270 return false;
5271
5272 intptr_t src_off = abase + (src_off_con << scale);
5273 intptr_t dest_off = abase + (dest_off_con << scale);
5274
5275 if (((src_off | dest_off) & (BytesPerLong-1)) != 0) {
5276 // Non-aligned; too bad.
5277 // One more chance: Pick off an initial 32-bit word.
5278 // This is a common case, since abase can be odd mod 8.
5279 if (((src_off | dest_off) & (BytesPerLong-1)) == BytesPerInt &&
5280 ((src_off ^ dest_off) & (BytesPerLong-1)) == 0) {
5281 Node* sptr = basic_plus_adr(src, src_off);
5282 Node* dptr = basic_plus_adr(dest, dest_off);
5283 Node* sval = make_load(control(), sptr, TypeInt::INT, T_INT, adr_type);
5284 store_to_memory(control(), dptr, sval, T_INT, adr_type);
5285 src_off += BytesPerInt;
5286 dest_off += BytesPerInt;
5287 } else {
5288 return false;
5289 }
5290 }
5291 assert(src_off % BytesPerLong == 0, "");
5292 assert(dest_off % BytesPerLong == 0, "");
5293
5294 // Do this copy by giant steps.
5295 Node* sptr = basic_plus_adr(src, src_off);
5296 Node* dptr = basic_plus_adr(dest, dest_off);
5297 Node* countx = dest_size;
5298 countx = _gvn.transform( new (C, 3) SubXNode(countx, MakeConX(dest_off)) );
5299 countx = _gvn.transform( new (C, 3) URShiftXNode(countx, intcon(LogBytesPerLong)) );
5300
5301 bool disjoint_bases = true; // since alloc != NULL
5302 generate_unchecked_arraycopy(adr_type, T_LONG, disjoint_bases,
5303 sptr, NULL, dptr, NULL, countx, dest_uninitialized);
5304
5305 return true;
5306 }
5307
5308
5309 // Helper function; generates code for the slow case.
5310 // We make a call to a runtime method which emulates the native method,
5311 // but without the native wrapper overhead.
5312 void
5313 LibraryCallKit::generate_slow_arraycopy(const TypePtr* adr_type,
5314 Node* src, Node* src_offset,
5315 Node* dest, Node* dest_offset,
5316 Node* copy_length, bool dest_uninitialized) {
5317 assert(!dest_uninitialized, "Invariant");
5318 Node* call = make_runtime_call(RC_NO_LEAF | RC_UNCOMMON,
5319 OptoRuntime::slow_arraycopy_Type(),
5320 OptoRuntime::slow_arraycopy_Java(),
5321 "slow_arraycopy", adr_type,
5322 src, src_offset, dest, dest_offset,
5323 copy_length);
5324
5325 // Handle exceptions thrown by this fellow:
5326 make_slow_call_ex(call, env()->Throwable_klass(), false);
5327 }
5328
5329 // Helper function; generates code for cases requiring runtime checks.
5330 Node*
5331 LibraryCallKit::generate_checkcast_arraycopy(const TypePtr* adr_type,
5332 Node* dest_elem_klass,
5333 Node* src, Node* src_offset,
5334 Node* dest, Node* dest_offset,
5335 Node* copy_length, bool dest_uninitialized) {
5336 if (stopped()) return NULL;
5337
5338 address copyfunc_addr = StubRoutines::checkcast_arraycopy(dest_uninitialized);
5339 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5340 return NULL;
5341 }
5342
5343 // Pick out the parameters required to perform a store-check
5344 // for the target array. This is an optimistic check. It will
5345 // look in each non-null element's class, at the desired klass's
5346 // super_check_offset, for the desired klass.
5347 int sco_offset = in_bytes(Klass::super_check_offset_offset());
5348 Node* p3 = basic_plus_adr(dest_elem_klass, sco_offset);
5349 Node* n3 = new(C, 3) LoadINode(NULL, memory(p3), p3, _gvn.type(p3)->is_ptr());
5350 Node* check_offset = ConvI2X(_gvn.transform(n3));
5351 Node* check_value = dest_elem_klass;
5352
5353 Node* src_start = array_element_address(src, src_offset, T_OBJECT);
5354 Node* dest_start = array_element_address(dest, dest_offset, T_OBJECT);
5355
5356 // (We know the arrays are never conjoint, because their types differ.)
5357 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5358 OptoRuntime::checkcast_arraycopy_Type(),
5359 copyfunc_addr, "checkcast_arraycopy", adr_type,
5360 // five arguments, of which two are
5361 // intptr_t (jlong in LP64)
5362 src_start, dest_start,
5363 copy_length XTOP,
5364 check_offset XTOP,
5365 check_value);
5366
5367 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5368 }
5369
5370
5371 // Helper function; generates code for cases requiring runtime checks.
5372 Node*
5373 LibraryCallKit::generate_generic_arraycopy(const TypePtr* adr_type,
5374 Node* src, Node* src_offset,
5375 Node* dest, Node* dest_offset,
5376 Node* copy_length, bool dest_uninitialized) {
5377 assert(!dest_uninitialized, "Invariant");
5378 if (stopped()) return NULL;
5379 address copyfunc_addr = StubRoutines::generic_arraycopy();
5380 if (copyfunc_addr == NULL) { // Stub was not generated, go slow path.
5381 return NULL;
5382 }
5383
5384 Node* call = make_runtime_call(RC_LEAF|RC_NO_FP,
5385 OptoRuntime::generic_arraycopy_Type(),
5386 copyfunc_addr, "generic_arraycopy", adr_type,
5387 src, src_offset, dest, dest_offset, copy_length);
5388
5389 return _gvn.transform(new (C, 1) ProjNode(call, TypeFunc::Parms));
5390 }
5391
5392 // Helper function; generates the fast out-of-line call to an arraycopy stub.
5393 void
5394 LibraryCallKit::generate_unchecked_arraycopy(const TypePtr* adr_type,
5395 BasicType basic_elem_type,
5396 bool disjoint_bases,
5397 Node* src, Node* src_offset,
5398 Node* dest, Node* dest_offset,
5399 Node* copy_length, bool dest_uninitialized) {
5400 if (stopped()) return; // nothing to do
5401
5402 Node* src_start = src;
5403 Node* dest_start = dest;
5404 if (src_offset != NULL || dest_offset != NULL) {
5405 assert(src_offset != NULL && dest_offset != NULL, "");
5406 src_start = array_element_address(src, src_offset, basic_elem_type);
5407 dest_start = array_element_address(dest, dest_offset, basic_elem_type);
5408 }
5409
5410 // Figure out which arraycopy runtime method to call.
5411 const char* copyfunc_name = "arraycopy";
5412 address copyfunc_addr =
5413 basictype2arraycopy(basic_elem_type, src_offset, dest_offset,
5414 disjoint_bases, copyfunc_name, dest_uninitialized);
5415
5416 // Call it. Note that the count_ix value is not scaled to a byte-size.
5417 make_runtime_call(RC_LEAF|RC_NO_FP,
5418 OptoRuntime::fast_arraycopy_Type(),
5419 copyfunc_addr, copyfunc_name, adr_type,
5420 src_start, dest_start, copy_length XTOP);
5421 }
5422
5423 //----------------------------inline_reference_get----------------------------
5424
5425 bool LibraryCallKit::inline_reference_get() {
5426 const int nargs = 1; // self
5427
5428 guarantee(java_lang_ref_Reference::referent_offset > 0,
5429 "should have already been set");
5430
5431 int referent_offset = java_lang_ref_Reference::referent_offset;
5432
5433 // Restore the stack and pop off the argument
5434 _sp += nargs;
5435 Node *reference_obj = pop();
5436
5437 // Null check on self without removing any arguments.
5438 _sp += nargs;
5439 reference_obj = do_null_check(reference_obj, T_OBJECT);
5440 _sp -= nargs;;
5441
5442 if (stopped()) return true;
5443
5444 Node *adr = basic_plus_adr(reference_obj, reference_obj, referent_offset);
5445
5446 ciInstanceKlass* klass = env()->Object_klass();
5447 const TypeOopPtr* object_type = TypeOopPtr::make_from_klass(klass);
5448
5449 Node* no_ctrl = NULL;
5450 Node *result = make_load(no_ctrl, adr, object_type, T_OBJECT);
5451
5452 // Use the pre-barrier to record the value in the referent field
5453 pre_barrier(false /* do_load */,
5454 control(),
5455 NULL /* obj */, NULL /* adr */, max_juint /* alias_idx */, NULL /* val */, NULL /* val_type */,
5456 result /* pre_val */,
5457 T_OBJECT);
5458
5459 push(result);
5460 return true;
5461 }
5462